xref: /openbmc/linux/kernel/rcu/tasks.h (revision bbaf1ff0)
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  * @rtp_n_lock_retries: Rough lock-contention statistic.
29  * @rtp_work: Work queue for invoking callbacks.
30  * @rtp_irq_work: IRQ work queue for deferred wakeups.
31  * @barrier_q_head: RCU callback for barrier operation.
32  * @rtp_blkd_tasks: List of tasks blocked as readers.
33  * @cpu: CPU number corresponding to this entry.
34  * @rtpp: Pointer to the rcu_tasks structure.
35  */
36 struct rcu_tasks_percpu {
37 	struct rcu_segcblist cblist;
38 	raw_spinlock_t __private lock;
39 	unsigned long rtp_jiffies;
40 	unsigned long rtp_n_lock_retries;
41 	struct work_struct rtp_work;
42 	struct irq_work rtp_irq_work;
43 	struct rcu_head barrier_q_head;
44 	struct list_head rtp_blkd_tasks;
45 	int cpu;
46 	struct rcu_tasks *rtpp;
47 };
48 
49 /**
50  * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
51  * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
52  * @cbs_gbl_lock: Lock protecting callback list.
53  * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone.
54  * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
55  * @gp_func: This flavor's grace-period-wait function.
56  * @gp_state: Grace period's most recent state transition (debugging).
57  * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
58  * @init_fract: Initial backoff sleep interval.
59  * @gp_jiffies: Time of last @gp_state transition.
60  * @gp_start: Most recent grace-period start in jiffies.
61  * @tasks_gp_seq: Number of grace periods completed since boot.
62  * @n_ipis: Number of IPIs sent to encourage grace periods to end.
63  * @n_ipis_fails: Number of IPI-send failures.
64  * @pregp_func: This flavor's pre-grace-period function (optional).
65  * @pertask_func: This flavor's per-task scan function (optional).
66  * @postscan_func: This flavor's post-task scan function (optional).
67  * @holdouts_func: This flavor's holdout-list scan function (optional).
68  * @postgp_func: This flavor's post-grace-period function (optional).
69  * @call_func: This flavor's call_rcu()-equivalent function.
70  * @rtpcpu: This flavor's rcu_tasks_percpu structure.
71  * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
72  * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
73  * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
74  * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
75  * @barrier_q_mutex: Serialize barrier operations.
76  * @barrier_q_count: Number of queues being waited on.
77  * @barrier_q_completion: Barrier wait/wakeup mechanism.
78  * @barrier_q_seq: Sequence number for barrier operations.
79  * @name: This flavor's textual name.
80  * @kname: This flavor's kthread name.
81  */
82 struct rcu_tasks {
83 	struct rcuwait cbs_wait;
84 	raw_spinlock_t cbs_gbl_lock;
85 	struct mutex tasks_gp_mutex;
86 	int gp_state;
87 	int gp_sleep;
88 	int init_fract;
89 	unsigned long gp_jiffies;
90 	unsigned long gp_start;
91 	unsigned long tasks_gp_seq;
92 	unsigned long n_ipis;
93 	unsigned long n_ipis_fails;
94 	struct task_struct *kthread_ptr;
95 	rcu_tasks_gp_func_t gp_func;
96 	pregp_func_t pregp_func;
97 	pertask_func_t pertask_func;
98 	postscan_func_t postscan_func;
99 	holdouts_func_t holdouts_func;
100 	postgp_func_t postgp_func;
101 	call_rcu_func_t call_func;
102 	struct rcu_tasks_percpu __percpu *rtpcpu;
103 	int percpu_enqueue_shift;
104 	int percpu_enqueue_lim;
105 	int percpu_dequeue_lim;
106 	unsigned long percpu_dequeue_gpseq;
107 	struct mutex barrier_q_mutex;
108 	atomic_t barrier_q_count;
109 	struct completion barrier_q_completion;
110 	unsigned long barrier_q_seq;
111 	char *name;
112 	char *kname;
113 };
114 
115 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
116 
117 #define DEFINE_RCU_TASKS(rt_name, gp, call, n)						\
118 static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = {			\
119 	.lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock),		\
120 	.rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup),			\
121 };											\
122 static struct rcu_tasks rt_name =							\
123 {											\
124 	.cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait),				\
125 	.cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock),			\
126 	.tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex),			\
127 	.gp_func = gp,									\
128 	.call_func = call,								\
129 	.rtpcpu = &rt_name ## __percpu,							\
130 	.name = n,									\
131 	.percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS),				\
132 	.percpu_enqueue_lim = 1,							\
133 	.percpu_dequeue_lim = 1,							\
134 	.barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex),		\
135 	.barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT,				\
136 	.kname = #rt_name,								\
137 }
138 
139 #ifdef CONFIG_TASKS_RCU
140 /* Track exiting tasks in order to allow them to be waited for. */
141 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
142 #endif
143 
144 #ifdef CONFIG_TASKS_RCU
145 /* Report delay in synchronize_srcu() completion in rcu_tasks_postscan(). */
146 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused);
147 static DEFINE_TIMER(tasks_rcu_exit_srcu_stall_timer, tasks_rcu_exit_srcu_stall);
148 #endif
149 
150 /* Avoid IPIing CPUs early in the grace period. */
151 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
152 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
153 module_param(rcu_task_ipi_delay, int, 0644);
154 
155 /* Control stall timeouts.  Disable with <= 0, otherwise jiffies till stall. */
156 #define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
157 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
158 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
159 module_param(rcu_task_stall_timeout, int, 0644);
160 #define RCU_TASK_STALL_INFO (HZ * 10)
161 static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
162 module_param(rcu_task_stall_info, int, 0644);
163 static int rcu_task_stall_info_mult __read_mostly = 3;
164 module_param(rcu_task_stall_info_mult, int, 0444);
165 
166 static int rcu_task_enqueue_lim __read_mostly = -1;
167 module_param(rcu_task_enqueue_lim, int, 0444);
168 
169 static bool rcu_task_cb_adjust;
170 static int rcu_task_contend_lim __read_mostly = 100;
171 module_param(rcu_task_contend_lim, int, 0444);
172 static int rcu_task_collapse_lim __read_mostly = 10;
173 module_param(rcu_task_collapse_lim, int, 0444);
174 
175 /* RCU tasks grace-period state for debugging. */
176 #define RTGS_INIT		 0
177 #define RTGS_WAIT_WAIT_CBS	 1
178 #define RTGS_WAIT_GP		 2
179 #define RTGS_PRE_WAIT_GP	 3
180 #define RTGS_SCAN_TASKLIST	 4
181 #define RTGS_POST_SCAN_TASKLIST	 5
182 #define RTGS_WAIT_SCAN_HOLDOUTS	 6
183 #define RTGS_SCAN_HOLDOUTS	 7
184 #define RTGS_POST_GP		 8
185 #define RTGS_WAIT_READERS	 9
186 #define RTGS_INVOKE_CBS		10
187 #define RTGS_WAIT_CBS		11
188 #ifndef CONFIG_TINY_RCU
189 static const char * const rcu_tasks_gp_state_names[] = {
190 	"RTGS_INIT",
191 	"RTGS_WAIT_WAIT_CBS",
192 	"RTGS_WAIT_GP",
193 	"RTGS_PRE_WAIT_GP",
194 	"RTGS_SCAN_TASKLIST",
195 	"RTGS_POST_SCAN_TASKLIST",
196 	"RTGS_WAIT_SCAN_HOLDOUTS",
197 	"RTGS_SCAN_HOLDOUTS",
198 	"RTGS_POST_GP",
199 	"RTGS_WAIT_READERS",
200 	"RTGS_INVOKE_CBS",
201 	"RTGS_WAIT_CBS",
202 };
203 #endif /* #ifndef CONFIG_TINY_RCU */
204 
205 ////////////////////////////////////////////////////////////////////////
206 //
207 // Generic code.
208 
209 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
210 
211 /* Record grace-period phase and time. */
212 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
213 {
214 	rtp->gp_state = newstate;
215 	rtp->gp_jiffies = jiffies;
216 }
217 
218 #ifndef CONFIG_TINY_RCU
219 /* Return state name. */
220 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
221 {
222 	int i = data_race(rtp->gp_state); // Let KCSAN detect update races
223 	int j = READ_ONCE(i); // Prevent the compiler from reading twice
224 
225 	if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
226 		return "???";
227 	return rcu_tasks_gp_state_names[j];
228 }
229 #endif /* #ifndef CONFIG_TINY_RCU */
230 
231 // Initialize per-CPU callback lists for the specified flavor of
232 // Tasks RCU.
233 static void cblist_init_generic(struct rcu_tasks *rtp)
234 {
235 	int cpu;
236 	unsigned long flags;
237 	int lim;
238 	int shift;
239 
240 	raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
241 	if (rcu_task_enqueue_lim < 0) {
242 		rcu_task_enqueue_lim = 1;
243 		rcu_task_cb_adjust = true;
244 	} else if (rcu_task_enqueue_lim == 0) {
245 		rcu_task_enqueue_lim = 1;
246 	}
247 	lim = rcu_task_enqueue_lim;
248 
249 	if (lim > nr_cpu_ids)
250 		lim = nr_cpu_ids;
251 	shift = ilog2(nr_cpu_ids / lim);
252 	if (((nr_cpu_ids - 1) >> shift) >= lim)
253 		shift++;
254 	WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
255 	WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
256 	smp_store_release(&rtp->percpu_enqueue_lim, lim);
257 	for_each_possible_cpu(cpu) {
258 		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
259 
260 		WARN_ON_ONCE(!rtpcp);
261 		if (cpu)
262 			raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
263 		raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
264 		if (rcu_segcblist_empty(&rtpcp->cblist))
265 			rcu_segcblist_init(&rtpcp->cblist);
266 		INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
267 		rtpcp->cpu = cpu;
268 		rtpcp->rtpp = rtp;
269 		if (!rtpcp->rtp_blkd_tasks.next)
270 			INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
271 		raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
272 	}
273 	raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
274 
275 	pr_info("%s: Setting shift to %d and lim to %d rcu_task_cb_adjust=%d.\n", rtp->name,
276 			data_race(rtp->percpu_enqueue_shift), data_race(rtp->percpu_enqueue_lim), rcu_task_cb_adjust);
277 }
278 
279 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
280 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
281 {
282 	struct rcu_tasks *rtp;
283 	struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
284 
285 	rtp = rtpcp->rtpp;
286 	rcuwait_wake_up(&rtp->cbs_wait);
287 }
288 
289 // Enqueue a callback for the specified flavor of Tasks RCU.
290 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
291 				   struct rcu_tasks *rtp)
292 {
293 	int chosen_cpu;
294 	unsigned long flags;
295 	int ideal_cpu;
296 	unsigned long j;
297 	bool needadjust = false;
298 	bool needwake;
299 	struct rcu_tasks_percpu *rtpcp;
300 
301 	rhp->next = NULL;
302 	rhp->func = func;
303 	local_irq_save(flags);
304 	rcu_read_lock();
305 	ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
306 	chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
307 	rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
308 	if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
309 		raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
310 		j = jiffies;
311 		if (rtpcp->rtp_jiffies != j) {
312 			rtpcp->rtp_jiffies = j;
313 			rtpcp->rtp_n_lock_retries = 0;
314 		}
315 		if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
316 		    READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
317 			needadjust = true;  // Defer adjustment to avoid deadlock.
318 	}
319 	if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) {
320 		raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
321 		cblist_init_generic(rtp);
322 		raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
323 	}
324 	needwake = rcu_segcblist_empty(&rtpcp->cblist);
325 	rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
326 	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
327 	if (unlikely(needadjust)) {
328 		raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
329 		if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
330 			WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
331 			WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
332 			smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
333 			pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
334 		}
335 		raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
336 	}
337 	rcu_read_unlock();
338 	/* We can't create the thread unless interrupts are enabled. */
339 	if (needwake && READ_ONCE(rtp->kthread_ptr))
340 		irq_work_queue(&rtpcp->rtp_irq_work);
341 }
342 
343 // RCU callback function for rcu_barrier_tasks_generic().
344 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
345 {
346 	struct rcu_tasks *rtp;
347 	struct rcu_tasks_percpu *rtpcp;
348 
349 	rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
350 	rtp = rtpcp->rtpp;
351 	if (atomic_dec_and_test(&rtp->barrier_q_count))
352 		complete(&rtp->barrier_q_completion);
353 }
354 
355 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
356 // Operates in a manner similar to rcu_barrier().
357 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
358 {
359 	int cpu;
360 	unsigned long flags;
361 	struct rcu_tasks_percpu *rtpcp;
362 	unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
363 
364 	mutex_lock(&rtp->barrier_q_mutex);
365 	if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
366 		smp_mb();
367 		mutex_unlock(&rtp->barrier_q_mutex);
368 		return;
369 	}
370 	rcu_seq_start(&rtp->barrier_q_seq);
371 	init_completion(&rtp->barrier_q_completion);
372 	atomic_set(&rtp->barrier_q_count, 2);
373 	for_each_possible_cpu(cpu) {
374 		if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
375 			break;
376 		rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
377 		rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
378 		raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
379 		if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
380 			atomic_inc(&rtp->barrier_q_count);
381 		raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
382 	}
383 	if (atomic_sub_and_test(2, &rtp->barrier_q_count))
384 		complete(&rtp->barrier_q_completion);
385 	wait_for_completion(&rtp->barrier_q_completion);
386 	rcu_seq_end(&rtp->barrier_q_seq);
387 	mutex_unlock(&rtp->barrier_q_mutex);
388 }
389 
390 // Advance callbacks and indicate whether either a grace period or
391 // callback invocation is needed.
392 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
393 {
394 	int cpu;
395 	unsigned long flags;
396 	bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
397 	long n;
398 	long ncbs = 0;
399 	long ncbsnz = 0;
400 	int needgpcb = 0;
401 
402 	for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
403 		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
404 
405 		/* Advance and accelerate any new callbacks. */
406 		if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
407 			continue;
408 		raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
409 		// Should we shrink down to a single callback queue?
410 		n = rcu_segcblist_n_cbs(&rtpcp->cblist);
411 		if (n) {
412 			ncbs += n;
413 			if (cpu > 0)
414 				ncbsnz += n;
415 		}
416 		rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
417 		(void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
418 		if (rcu_segcblist_pend_cbs(&rtpcp->cblist))
419 			needgpcb |= 0x3;
420 		if (!rcu_segcblist_empty(&rtpcp->cblist))
421 			needgpcb |= 0x1;
422 		raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
423 	}
424 
425 	// Shrink down to a single callback queue if appropriate.
426 	// This is done in two stages: (1) If there are no more than
427 	// rcu_task_collapse_lim callbacks on CPU 0 and none on any other
428 	// CPU, limit enqueueing to CPU 0.  (2) After an RCU grace period,
429 	// if there has not been an increase in callbacks, limit dequeuing
430 	// to CPU 0.  Note the matching RCU read-side critical section in
431 	// call_rcu_tasks_generic().
432 	if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
433 		raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
434 		if (rtp->percpu_enqueue_lim > 1) {
435 			WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
436 			smp_store_release(&rtp->percpu_enqueue_lim, 1);
437 			rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
438 			gpdone = false;
439 			pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
440 		}
441 		raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
442 	}
443 	if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
444 		raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
445 		if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
446 			WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
447 			pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
448 		}
449 		if (rtp->percpu_dequeue_lim == 1) {
450 			for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) {
451 				struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
452 
453 				WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
454 			}
455 		}
456 		raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
457 	}
458 
459 	return needgpcb;
460 }
461 
462 // Advance callbacks and invoke any that are ready.
463 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
464 {
465 	int cpu;
466 	int cpunext;
467 	int cpuwq;
468 	unsigned long flags;
469 	int len;
470 	struct rcu_head *rhp;
471 	struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
472 	struct rcu_tasks_percpu *rtpcp_next;
473 
474 	cpu = rtpcp->cpu;
475 	cpunext = cpu * 2 + 1;
476 	if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
477 		rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
478 		cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
479 		queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
480 		cpunext++;
481 		if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
482 			rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
483 			cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
484 			queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
485 		}
486 	}
487 
488 	if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
489 		return;
490 	raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
491 	rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
492 	rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
493 	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
494 	len = rcl.len;
495 	for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
496 		local_bh_disable();
497 		rhp->func(rhp);
498 		local_bh_enable();
499 		cond_resched();
500 	}
501 	raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
502 	rcu_segcblist_add_len(&rtpcp->cblist, -len);
503 	(void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
504 	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
505 }
506 
507 // Workqueue flood to advance callbacks and invoke any that are ready.
508 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
509 {
510 	struct rcu_tasks *rtp;
511 	struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
512 
513 	rtp = rtpcp->rtpp;
514 	rcu_tasks_invoke_cbs(rtp, rtpcp);
515 }
516 
517 // Wait for one grace period.
518 static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
519 {
520 	int needgpcb;
521 
522 	mutex_lock(&rtp->tasks_gp_mutex);
523 
524 	// If there were none, wait a bit and start over.
525 	if (unlikely(midboot)) {
526 		needgpcb = 0x2;
527 	} else {
528 		set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
529 		rcuwait_wait_event(&rtp->cbs_wait,
530 				   (needgpcb = rcu_tasks_need_gpcb(rtp)),
531 				   TASK_IDLE);
532 	}
533 
534 	if (needgpcb & 0x2) {
535 		// Wait for one grace period.
536 		set_tasks_gp_state(rtp, RTGS_WAIT_GP);
537 		rtp->gp_start = jiffies;
538 		rcu_seq_start(&rtp->tasks_gp_seq);
539 		rtp->gp_func(rtp);
540 		rcu_seq_end(&rtp->tasks_gp_seq);
541 	}
542 
543 	// Invoke callbacks.
544 	set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
545 	rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
546 	mutex_unlock(&rtp->tasks_gp_mutex);
547 }
548 
549 // RCU-tasks kthread that detects grace periods and invokes callbacks.
550 static int __noreturn rcu_tasks_kthread(void *arg)
551 {
552 	struct rcu_tasks *rtp = arg;
553 
554 	/* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
555 	housekeeping_affine(current, HK_TYPE_RCU);
556 	WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
557 
558 	/*
559 	 * Each pass through the following loop makes one check for
560 	 * newly arrived callbacks, and, if there are some, waits for
561 	 * one RCU-tasks grace period and then invokes the callbacks.
562 	 * This loop is terminated by the system going down.  ;-)
563 	 */
564 	for (;;) {
565 		// Wait for one grace period and invoke any callbacks
566 		// that are ready.
567 		rcu_tasks_one_gp(rtp, false);
568 
569 		// Paranoid sleep to keep this from entering a tight loop.
570 		schedule_timeout_idle(rtp->gp_sleep);
571 	}
572 }
573 
574 // Wait for a grace period for the specified flavor of Tasks RCU.
575 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
576 {
577 	/* Complain if the scheduler has not started.  */
578 	if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
579 			 "synchronize_%s() called too soon", rtp->name))
580 		return;
581 
582 	// If the grace-period kthread is running, use it.
583 	if (READ_ONCE(rtp->kthread_ptr)) {
584 		wait_rcu_gp(rtp->call_func);
585 		return;
586 	}
587 	rcu_tasks_one_gp(rtp, true);
588 }
589 
590 /* Spawn RCU-tasks grace-period kthread. */
591 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
592 {
593 	struct task_struct *t;
594 
595 	t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
596 	if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
597 		return;
598 	smp_mb(); /* Ensure others see full kthread. */
599 }
600 
601 #ifndef CONFIG_TINY_RCU
602 
603 /*
604  * Print any non-default Tasks RCU settings.
605  */
606 static void __init rcu_tasks_bootup_oddness(void)
607 {
608 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
609 	int rtsimc;
610 
611 	if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
612 		pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
613 	rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
614 	if (rtsimc != rcu_task_stall_info_mult) {
615 		pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
616 		rcu_task_stall_info_mult = rtsimc;
617 	}
618 #endif /* #ifdef CONFIG_TASKS_RCU */
619 #ifdef CONFIG_TASKS_RCU
620 	pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
621 #endif /* #ifdef CONFIG_TASKS_RCU */
622 #ifdef CONFIG_TASKS_RUDE_RCU
623 	pr_info("\tRude variant of Tasks RCU enabled.\n");
624 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
625 #ifdef CONFIG_TASKS_TRACE_RCU
626 	pr_info("\tTracing variant of Tasks RCU enabled.\n");
627 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
628 }
629 
630 #endif /* #ifndef CONFIG_TINY_RCU */
631 
632 #ifndef CONFIG_TINY_RCU
633 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
634 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
635 {
636 	int cpu;
637 	bool havecbs = false;
638 
639 	for_each_possible_cpu(cpu) {
640 		struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
641 
642 		if (!data_race(rcu_segcblist_empty(&rtpcp->cblist))) {
643 			havecbs = true;
644 			break;
645 		}
646 	}
647 	pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
648 		rtp->kname,
649 		tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
650 		jiffies - data_race(rtp->gp_jiffies),
651 		data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
652 		data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
653 		".k"[!!data_race(rtp->kthread_ptr)],
654 		".C"[havecbs],
655 		s);
656 }
657 #endif // #ifndef CONFIG_TINY_RCU
658 
659 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
660 
661 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
662 
663 ////////////////////////////////////////////////////////////////////////
664 //
665 // Shared code between task-list-scanning variants of Tasks RCU.
666 
667 /* Wait for one RCU-tasks grace period. */
668 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
669 {
670 	struct task_struct *g;
671 	int fract;
672 	LIST_HEAD(holdouts);
673 	unsigned long j;
674 	unsigned long lastinfo;
675 	unsigned long lastreport;
676 	bool reported = false;
677 	int rtsi;
678 	struct task_struct *t;
679 
680 	set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
681 	rtp->pregp_func(&holdouts);
682 
683 	/*
684 	 * There were callbacks, so we need to wait for an RCU-tasks
685 	 * grace period.  Start off by scanning the task list for tasks
686 	 * that are not already voluntarily blocked.  Mark these tasks
687 	 * and make a list of them in holdouts.
688 	 */
689 	set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
690 	if (rtp->pertask_func) {
691 		rcu_read_lock();
692 		for_each_process_thread(g, t)
693 			rtp->pertask_func(t, &holdouts);
694 		rcu_read_unlock();
695 	}
696 
697 	set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
698 	rtp->postscan_func(&holdouts);
699 
700 	/*
701 	 * Each pass through the following loop scans the list of holdout
702 	 * tasks, removing any that are no longer holdouts.  When the list
703 	 * is empty, we are done.
704 	 */
705 	lastreport = jiffies;
706 	lastinfo = lastreport;
707 	rtsi = READ_ONCE(rcu_task_stall_info);
708 
709 	// Start off with initial wait and slowly back off to 1 HZ wait.
710 	fract = rtp->init_fract;
711 
712 	while (!list_empty(&holdouts)) {
713 		ktime_t exp;
714 		bool firstreport;
715 		bool needreport;
716 		int rtst;
717 
718 		// Slowly back off waiting for holdouts
719 		set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
720 		if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
721 			schedule_timeout_idle(fract);
722 		} else {
723 			exp = jiffies_to_nsecs(fract);
724 			__set_current_state(TASK_IDLE);
725 			schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
726 		}
727 
728 		if (fract < HZ)
729 			fract++;
730 
731 		rtst = READ_ONCE(rcu_task_stall_timeout);
732 		needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
733 		if (needreport) {
734 			lastreport = jiffies;
735 			reported = true;
736 		}
737 		firstreport = true;
738 		WARN_ON(signal_pending(current));
739 		set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
740 		rtp->holdouts_func(&holdouts, needreport, &firstreport);
741 
742 		// Print pre-stall informational messages if needed.
743 		j = jiffies;
744 		if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
745 			lastinfo = j;
746 			rtsi = rtsi * rcu_task_stall_info_mult;
747 			pr_info("%s: %s grace period number %lu (since boot) is %lu jiffies old.\n",
748 				__func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
749 		}
750 	}
751 
752 	set_tasks_gp_state(rtp, RTGS_POST_GP);
753 	rtp->postgp_func(rtp);
754 }
755 
756 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
757 
758 #ifdef CONFIG_TASKS_RCU
759 
760 ////////////////////////////////////////////////////////////////////////
761 //
762 // Simple variant of RCU whose quiescent states are voluntary context
763 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
764 // As such, grace periods can take one good long time.  There are no
765 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
766 // because this implementation is intended to get the system into a safe
767 // state for some of the manipulations involved in tracing and the like.
768 // Finally, this implementation does not support high call_rcu_tasks()
769 // rates from multiple CPUs.  If this is required, per-CPU callback lists
770 // will be needed.
771 //
772 // The implementation uses rcu_tasks_wait_gp(), which relies on function
773 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_kthread()
774 // function sets these function pointers up so that rcu_tasks_wait_gp()
775 // invokes these functions in this order:
776 //
777 // rcu_tasks_pregp_step():
778 //	Invokes synchronize_rcu() in order to wait for all in-flight
779 //	t->on_rq and t->nvcsw transitions to complete.	This works because
780 //	all such transitions are carried out with interrupts disabled.
781 // rcu_tasks_pertask(), invoked on every non-idle task:
782 //	For every runnable non-idle task other than the current one, use
783 //	get_task_struct() to pin down that task, snapshot that task's
784 //	number of voluntary context switches, and add that task to the
785 //	holdout list.
786 // rcu_tasks_postscan():
787 //	Invoke synchronize_srcu() to ensure that all tasks that were
788 //	in the process of exiting (and which thus might not know to
789 //	synchronize with this RCU Tasks grace period) have completed
790 //	exiting.
791 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
792 //	Scans the holdout list, attempting to identify a quiescent state
793 //	for each task on the list.  If there is a quiescent state, the
794 //	corresponding task is removed from the holdout list.
795 // rcu_tasks_postgp():
796 //	Invokes synchronize_rcu() in order to ensure that all prior
797 //	t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
798 //	to have happened before the end of this RCU Tasks grace period.
799 //	Again, this works because all such transitions are carried out
800 //	with interrupts disabled.
801 //
802 // For each exiting task, the exit_tasks_rcu_start() and
803 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
804 // read-side critical sections waited for by rcu_tasks_postscan().
805 //
806 // Pre-grace-period update-side code is ordered before the grace
807 // via the raw_spin_lock.*rcu_node().  Pre-grace-period read-side code
808 // is ordered before the grace period via synchronize_rcu() call in
809 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
810 // disabling.
811 
812 /* Pre-grace-period preparation. */
813 static void rcu_tasks_pregp_step(struct list_head *hop)
814 {
815 	/*
816 	 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
817 	 * to complete.  Invoking synchronize_rcu() suffices because all
818 	 * these transitions occur with interrupts disabled.  Without this
819 	 * synchronize_rcu(), a read-side critical section that started
820 	 * before the grace period might be incorrectly seen as having
821 	 * started after the grace period.
822 	 *
823 	 * This synchronize_rcu() also dispenses with the need for a
824 	 * memory barrier on the first store to t->rcu_tasks_holdout,
825 	 * as it forces the store to happen after the beginning of the
826 	 * grace period.
827 	 */
828 	synchronize_rcu();
829 }
830 
831 /* Per-task initial processing. */
832 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
833 {
834 	if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
835 		get_task_struct(t);
836 		t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
837 		WRITE_ONCE(t->rcu_tasks_holdout, true);
838 		list_add(&t->rcu_tasks_holdout_list, hop);
839 	}
840 }
841 
842 /* Processing between scanning taskslist and draining the holdout list. */
843 static void rcu_tasks_postscan(struct list_head *hop)
844 {
845 	int rtsi = READ_ONCE(rcu_task_stall_info);
846 
847 	if (!IS_ENABLED(CONFIG_TINY_RCU)) {
848 		tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
849 		add_timer(&tasks_rcu_exit_srcu_stall_timer);
850 	}
851 
852 	/*
853 	 * Exiting tasks may escape the tasklist scan. Those are vulnerable
854 	 * until their final schedule() with TASK_DEAD state. To cope with
855 	 * this, divide the fragile exit path part in two intersecting
856 	 * read side critical sections:
857 	 *
858 	 * 1) An _SRCU_ read side starting before calling exit_notify(),
859 	 *    which may remove the task from the tasklist, and ending after
860 	 *    the final preempt_disable() call in do_exit().
861 	 *
862 	 * 2) An _RCU_ read side starting with the final preempt_disable()
863 	 *    call in do_exit() and ending with the final call to schedule()
864 	 *    with TASK_DEAD state.
865 	 *
866 	 * This handles the part 1). And postgp will handle part 2) with a
867 	 * call to synchronize_rcu().
868 	 */
869 	synchronize_srcu(&tasks_rcu_exit_srcu);
870 
871 	if (!IS_ENABLED(CONFIG_TINY_RCU))
872 		del_timer_sync(&tasks_rcu_exit_srcu_stall_timer);
873 }
874 
875 /* See if tasks are still holding out, complain if so. */
876 static void check_holdout_task(struct task_struct *t,
877 			       bool needreport, bool *firstreport)
878 {
879 	int cpu;
880 
881 	if (!READ_ONCE(t->rcu_tasks_holdout) ||
882 	    t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
883 	    !READ_ONCE(t->on_rq) ||
884 	    (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
885 	     !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
886 		WRITE_ONCE(t->rcu_tasks_holdout, false);
887 		list_del_init(&t->rcu_tasks_holdout_list);
888 		put_task_struct(t);
889 		return;
890 	}
891 	rcu_request_urgent_qs_task(t);
892 	if (!needreport)
893 		return;
894 	if (*firstreport) {
895 		pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
896 		*firstreport = false;
897 	}
898 	cpu = task_cpu(t);
899 	pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
900 		 t, ".I"[is_idle_task(t)],
901 		 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
902 		 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
903 		 t->rcu_tasks_idle_cpu, cpu);
904 	sched_show_task(t);
905 }
906 
907 /* Scan the holdout lists for tasks no longer holding out. */
908 static void check_all_holdout_tasks(struct list_head *hop,
909 				    bool needreport, bool *firstreport)
910 {
911 	struct task_struct *t, *t1;
912 
913 	list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
914 		check_holdout_task(t, needreport, firstreport);
915 		cond_resched();
916 	}
917 }
918 
919 /* Finish off the Tasks-RCU grace period. */
920 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
921 {
922 	/*
923 	 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
924 	 * memory barriers prior to them in the schedule() path, memory
925 	 * reordering on other CPUs could cause their RCU-tasks read-side
926 	 * critical sections to extend past the end of the grace period.
927 	 * However, because these ->nvcsw updates are carried out with
928 	 * interrupts disabled, we can use synchronize_rcu() to force the
929 	 * needed ordering on all such CPUs.
930 	 *
931 	 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
932 	 * accesses to be within the grace period, avoiding the need for
933 	 * memory barriers for ->rcu_tasks_holdout accesses.
934 	 *
935 	 * In addition, this synchronize_rcu() waits for exiting tasks
936 	 * to complete their final preempt_disable() region of execution,
937 	 * cleaning up after synchronize_srcu(&tasks_rcu_exit_srcu),
938 	 * enforcing the whole region before tasklist removal until
939 	 * the final schedule() with TASK_DEAD state to be an RCU TASKS
940 	 * read side critical section.
941 	 */
942 	synchronize_rcu();
943 }
944 
945 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
946 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
947 
948 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused)
949 {
950 #ifndef CONFIG_TINY_RCU
951 	int rtsi;
952 
953 	rtsi = READ_ONCE(rcu_task_stall_info);
954 	pr_info("%s: %s grace period number %lu (since boot) gp_state: %s is %lu jiffies old.\n",
955 		__func__, rcu_tasks.kname, rcu_tasks.tasks_gp_seq,
956 		tasks_gp_state_getname(&rcu_tasks), jiffies - rcu_tasks.gp_jiffies);
957 	pr_info("Please check any exiting tasks stuck between calls to exit_tasks_rcu_start() and exit_tasks_rcu_finish()\n");
958 	tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
959 	add_timer(&tasks_rcu_exit_srcu_stall_timer);
960 #endif // #ifndef CONFIG_TINY_RCU
961 }
962 
963 /**
964  * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
965  * @rhp: structure to be used for queueing the RCU updates.
966  * @func: actual callback function to be invoked after the grace period
967  *
968  * The callback function will be invoked some time after a full grace
969  * period elapses, in other words after all currently executing RCU
970  * read-side critical sections have completed. call_rcu_tasks() assumes
971  * that the read-side critical sections end at a voluntary context
972  * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
973  * or transition to usermode execution.  As such, there are no read-side
974  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
975  * this primitive is intended to determine that all tasks have passed
976  * through a safe state, not so much for data-structure synchronization.
977  *
978  * See the description of call_rcu() for more detailed information on
979  * memory ordering guarantees.
980  */
981 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
982 {
983 	call_rcu_tasks_generic(rhp, func, &rcu_tasks);
984 }
985 EXPORT_SYMBOL_GPL(call_rcu_tasks);
986 
987 /**
988  * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
989  *
990  * Control will return to the caller some time after a full rcu-tasks
991  * grace period has elapsed, in other words after all currently
992  * executing rcu-tasks read-side critical sections have elapsed.  These
993  * read-side critical sections are delimited by calls to schedule(),
994  * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
995  * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
996  *
997  * This is a very specialized primitive, intended only for a few uses in
998  * tracing and other situations requiring manipulation of function
999  * preambles and profiling hooks.  The synchronize_rcu_tasks() function
1000  * is not (yet) intended for heavy use from multiple CPUs.
1001  *
1002  * See the description of synchronize_rcu() for more detailed information
1003  * on memory ordering guarantees.
1004  */
1005 void synchronize_rcu_tasks(void)
1006 {
1007 	synchronize_rcu_tasks_generic(&rcu_tasks);
1008 }
1009 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
1010 
1011 /**
1012  * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
1013  *
1014  * Although the current implementation is guaranteed to wait, it is not
1015  * obligated to, for example, if there are no pending callbacks.
1016  */
1017 void rcu_barrier_tasks(void)
1018 {
1019 	rcu_barrier_tasks_generic(&rcu_tasks);
1020 }
1021 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
1022 
1023 static int __init rcu_spawn_tasks_kthread(void)
1024 {
1025 	cblist_init_generic(&rcu_tasks);
1026 	rcu_tasks.gp_sleep = HZ / 10;
1027 	rcu_tasks.init_fract = HZ / 10;
1028 	rcu_tasks.pregp_func = rcu_tasks_pregp_step;
1029 	rcu_tasks.pertask_func = rcu_tasks_pertask;
1030 	rcu_tasks.postscan_func = rcu_tasks_postscan;
1031 	rcu_tasks.holdouts_func = check_all_holdout_tasks;
1032 	rcu_tasks.postgp_func = rcu_tasks_postgp;
1033 	rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1034 	return 0;
1035 }
1036 
1037 #if !defined(CONFIG_TINY_RCU)
1038 void show_rcu_tasks_classic_gp_kthread(void)
1039 {
1040 	show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1041 }
1042 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1043 #endif // !defined(CONFIG_TINY_RCU)
1044 
1045 /*
1046  * Contribute to protect against tasklist scan blind spot while the
1047  * task is exiting and may be removed from the tasklist. See
1048  * corresponding synchronize_srcu() for further details.
1049  */
1050 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
1051 {
1052 	current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
1053 }
1054 
1055 /*
1056  * Contribute to protect against tasklist scan blind spot while the
1057  * task is exiting and may be removed from the tasklist. See
1058  * corresponding synchronize_srcu() for further details.
1059  */
1060 void exit_tasks_rcu_stop(void) __releases(&tasks_rcu_exit_srcu)
1061 {
1062 	struct task_struct *t = current;
1063 
1064 	__srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
1065 }
1066 
1067 /*
1068  * Contribute to protect against tasklist scan blind spot while the
1069  * task is exiting and may be removed from the tasklist. See
1070  * corresponding synchronize_srcu() for further details.
1071  */
1072 void exit_tasks_rcu_finish(void)
1073 {
1074 	exit_tasks_rcu_stop();
1075 	exit_tasks_rcu_finish_trace(current);
1076 }
1077 
1078 #else /* #ifdef CONFIG_TASKS_RCU */
1079 void exit_tasks_rcu_start(void) { }
1080 void exit_tasks_rcu_stop(void) { }
1081 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1082 #endif /* #else #ifdef CONFIG_TASKS_RCU */
1083 
1084 #ifdef CONFIG_TASKS_RUDE_RCU
1085 
1086 ////////////////////////////////////////////////////////////////////////
1087 //
1088 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
1089 // passing an empty function to schedule_on_each_cpu().  This approach
1090 // provides an asynchronous call_rcu_tasks_rude() API and batching of
1091 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
1092 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
1093 // and induces otherwise unnecessary context switches on all online CPUs,
1094 // whether idle or not.
1095 //
1096 // Callback handling is provided by the rcu_tasks_kthread() function.
1097 //
1098 // Ordering is provided by the scheduler's context-switch code.
1099 
1100 // Empty function to allow workqueues to force a context switch.
1101 static void rcu_tasks_be_rude(struct work_struct *work)
1102 {
1103 }
1104 
1105 // Wait for one rude RCU-tasks grace period.
1106 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1107 {
1108 	rtp->n_ipis += cpumask_weight(cpu_online_mask);
1109 	schedule_on_each_cpu(rcu_tasks_be_rude);
1110 }
1111 
1112 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1113 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1114 		 "RCU Tasks Rude");
1115 
1116 /**
1117  * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1118  * @rhp: structure to be used for queueing the RCU updates.
1119  * @func: actual callback function to be invoked after the grace period
1120  *
1121  * The callback function will be invoked some time after a full grace
1122  * period elapses, in other words after all currently executing RCU
1123  * read-side critical sections have completed. call_rcu_tasks_rude()
1124  * assumes that the read-side critical sections end at context switch,
1125  * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1126  * usermode execution is schedulable). As such, there are no read-side
1127  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1128  * this primitive is intended to determine that all tasks have passed
1129  * through a safe state, not so much for data-structure synchronization.
1130  *
1131  * See the description of call_rcu() for more detailed information on
1132  * memory ordering guarantees.
1133  */
1134 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1135 {
1136 	call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1137 }
1138 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1139 
1140 /**
1141  * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1142  *
1143  * Control will return to the caller some time after a rude rcu-tasks
1144  * grace period has elapsed, in other words after all currently
1145  * executing rcu-tasks read-side critical sections have elapsed.  These
1146  * read-side critical sections are delimited by calls to schedule(),
1147  * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1148  * context), and (in theory, anyway) cond_resched().
1149  *
1150  * This is a very specialized primitive, intended only for a few uses in
1151  * tracing and other situations requiring manipulation of function preambles
1152  * and profiling hooks.  The synchronize_rcu_tasks_rude() function is not
1153  * (yet) intended for heavy use from multiple CPUs.
1154  *
1155  * See the description of synchronize_rcu() for more detailed information
1156  * on memory ordering guarantees.
1157  */
1158 void synchronize_rcu_tasks_rude(void)
1159 {
1160 	synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1161 }
1162 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1163 
1164 /**
1165  * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1166  *
1167  * Although the current implementation is guaranteed to wait, it is not
1168  * obligated to, for example, if there are no pending callbacks.
1169  */
1170 void rcu_barrier_tasks_rude(void)
1171 {
1172 	rcu_barrier_tasks_generic(&rcu_tasks_rude);
1173 }
1174 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1175 
1176 static int __init rcu_spawn_tasks_rude_kthread(void)
1177 {
1178 	cblist_init_generic(&rcu_tasks_rude);
1179 	rcu_tasks_rude.gp_sleep = HZ / 10;
1180 	rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1181 	return 0;
1182 }
1183 
1184 #if !defined(CONFIG_TINY_RCU)
1185 void show_rcu_tasks_rude_gp_kthread(void)
1186 {
1187 	show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1188 }
1189 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1190 #endif // !defined(CONFIG_TINY_RCU)
1191 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1192 
1193 ////////////////////////////////////////////////////////////////////////
1194 //
1195 // Tracing variant of Tasks RCU.  This variant is designed to be used
1196 // to protect tracing hooks, including those of BPF.  This variant
1197 // therefore:
1198 //
1199 // 1.	Has explicit read-side markers to allow finite grace periods
1200 //	in the face of in-kernel loops for PREEMPT=n builds.
1201 //
1202 // 2.	Protects code in the idle loop, exception entry/exit, and
1203 //	CPU-hotplug code paths, similar to the capabilities of SRCU.
1204 //
1205 // 3.	Avoids expensive read-side instructions, having overhead similar
1206 //	to that of Preemptible RCU.
1207 //
1208 // There are of course downsides.  For example, the grace-period code
1209 // can send IPIs to CPUs, even when those CPUs are in the idle loop or
1210 // in nohz_full userspace.  If needed, these downsides can be at least
1211 // partially remedied.
1212 //
1213 // Perhaps most important, this variant of RCU does not affect the vanilla
1214 // flavors, rcu_preempt and rcu_sched.  The fact that RCU Tasks Trace
1215 // readers can operate from idle, offline, and exception entry/exit in no
1216 // way allows rcu_preempt and rcu_sched readers to also do so.
1217 //
1218 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1219 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_trace_kthread()
1220 // function sets these function pointers up so that rcu_tasks_wait_gp()
1221 // invokes these functions in this order:
1222 //
1223 // rcu_tasks_trace_pregp_step():
1224 //	Disables CPU hotplug, adds all currently executing tasks to the
1225 //	holdout list, then checks the state of all tasks that blocked
1226 //	or were preempted within their current RCU Tasks Trace read-side
1227 //	critical section, adding them to the holdout list if appropriate.
1228 //	Finally, this function re-enables CPU hotplug.
1229 // The ->pertask_func() pointer is NULL, so there is no per-task processing.
1230 // rcu_tasks_trace_postscan():
1231 //	Invokes synchronize_rcu() to wait for late-stage exiting tasks
1232 //	to finish exiting.
1233 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1234 //	Scans the holdout list, attempting to identify a quiescent state
1235 //	for each task on the list.  If there is a quiescent state, the
1236 //	corresponding task is removed from the holdout list.  Once this
1237 //	list is empty, the grace period has completed.
1238 // rcu_tasks_trace_postgp():
1239 //	Provides the needed full memory barrier and does debug checks.
1240 //
1241 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1242 //
1243 // Pre-grace-period update-side code is ordered before the grace period
1244 // via the ->cbs_lock and barriers in rcu_tasks_kthread().  Pre-grace-period
1245 // read-side code is ordered before the grace period by atomic operations
1246 // on .b.need_qs flag of each task involved in this process, or by scheduler
1247 // context-switch ordering (for locked-down non-running readers).
1248 
1249 // The lockdep state must be outside of #ifdef to be useful.
1250 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1251 static struct lock_class_key rcu_lock_trace_key;
1252 struct lockdep_map rcu_trace_lock_map =
1253 	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1254 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1255 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1256 
1257 #ifdef CONFIG_TASKS_TRACE_RCU
1258 
1259 // Record outstanding IPIs to each CPU.  No point in sending two...
1260 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1261 
1262 // The number of detections of task quiescent state relying on
1263 // heavyweight readers executing explicit memory barriers.
1264 static unsigned long n_heavy_reader_attempts;
1265 static unsigned long n_heavy_reader_updates;
1266 static unsigned long n_heavy_reader_ofl_updates;
1267 static unsigned long n_trc_holdouts;
1268 
1269 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1270 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1271 		 "RCU Tasks Trace");
1272 
1273 /* Load from ->trc_reader_special.b.need_qs with proper ordering. */
1274 static u8 rcu_ld_need_qs(struct task_struct *t)
1275 {
1276 	smp_mb(); // Enforce full grace-period ordering.
1277 	return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1278 }
1279 
1280 /* Store to ->trc_reader_special.b.need_qs with proper ordering. */
1281 static void rcu_st_need_qs(struct task_struct *t, u8 v)
1282 {
1283 	smp_store_release(&t->trc_reader_special.b.need_qs, v);
1284 	smp_mb(); // Enforce full grace-period ordering.
1285 }
1286 
1287 /*
1288  * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1289  * the four-byte operand-size restriction of some platforms.
1290  * Returns the old value, which is often ignored.
1291  */
1292 u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1293 {
1294 	union rcu_special ret;
1295 	union rcu_special trs_old = READ_ONCE(t->trc_reader_special);
1296 	union rcu_special trs_new = trs_old;
1297 
1298 	if (trs_old.b.need_qs != old)
1299 		return trs_old.b.need_qs;
1300 	trs_new.b.need_qs = new;
1301 	ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s);
1302 	return ret.b.need_qs;
1303 }
1304 EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1305 
1306 /*
1307  * If we are the last reader, signal the grace-period kthread.
1308  * Also remove from the per-CPU list of blocked tasks.
1309  */
1310 void rcu_read_unlock_trace_special(struct task_struct *t)
1311 {
1312 	unsigned long flags;
1313 	struct rcu_tasks_percpu *rtpcp;
1314 	union rcu_special trs;
1315 
1316 	// Open-coded full-word version of rcu_ld_need_qs().
1317 	smp_mb(); // Enforce full grace-period ordering.
1318 	trs = smp_load_acquire(&t->trc_reader_special);
1319 
1320 	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1321 		smp_mb(); // Pairs with update-side barriers.
1322 	// Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1323 	if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1324 		u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1325 						       TRC_NEED_QS_CHECKED);
1326 
1327 		WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1328 	}
1329 	if (trs.b.blocked) {
1330 		rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1331 		raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1332 		list_del_init(&t->trc_blkd_node);
1333 		WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1334 		raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1335 	}
1336 	WRITE_ONCE(t->trc_reader_nesting, 0);
1337 }
1338 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1339 
1340 /* Add a newly blocked reader task to its CPU's list. */
1341 void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1342 {
1343 	unsigned long flags;
1344 	struct rcu_tasks_percpu *rtpcp;
1345 
1346 	local_irq_save(flags);
1347 	rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1348 	raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1349 	t->trc_blkd_cpu = smp_processor_id();
1350 	if (!rtpcp->rtp_blkd_tasks.next)
1351 		INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1352 	list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1353 	WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1354 	raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1355 }
1356 EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1357 
1358 /* Add a task to the holdout list, if it is not already on the list. */
1359 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1360 {
1361 	if (list_empty(&t->trc_holdout_list)) {
1362 		get_task_struct(t);
1363 		list_add(&t->trc_holdout_list, bhp);
1364 		n_trc_holdouts++;
1365 	}
1366 }
1367 
1368 /* Remove a task from the holdout list, if it is in fact present. */
1369 static void trc_del_holdout(struct task_struct *t)
1370 {
1371 	if (!list_empty(&t->trc_holdout_list)) {
1372 		list_del_init(&t->trc_holdout_list);
1373 		put_task_struct(t);
1374 		n_trc_holdouts--;
1375 	}
1376 }
1377 
1378 /* IPI handler to check task state. */
1379 static void trc_read_check_handler(void *t_in)
1380 {
1381 	int nesting;
1382 	struct task_struct *t = current;
1383 	struct task_struct *texp = t_in;
1384 
1385 	// If the task is no longer running on this CPU, leave.
1386 	if (unlikely(texp != t))
1387 		goto reset_ipi; // Already on holdout list, so will check later.
1388 
1389 	// If the task is not in a read-side critical section, and
1390 	// if this is the last reader, awaken the grace-period kthread.
1391 	nesting = READ_ONCE(t->trc_reader_nesting);
1392 	if (likely(!nesting)) {
1393 		rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1394 		goto reset_ipi;
1395 	}
1396 	// If we are racing with an rcu_read_unlock_trace(), try again later.
1397 	if (unlikely(nesting < 0))
1398 		goto reset_ipi;
1399 
1400 	// Get here if the task is in a read-side critical section.
1401 	// Set its state so that it will update state for the grace-period
1402 	// kthread upon exit from that critical section.
1403 	rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1404 
1405 reset_ipi:
1406 	// Allow future IPIs to be sent on CPU and for task.
1407 	// Also order this IPI handler against any later manipulations of
1408 	// the intended task.
1409 	smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1410 	smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1411 }
1412 
1413 /* Callback function for scheduler to check locked-down task.  */
1414 static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1415 {
1416 	struct list_head *bhp = bhp_in;
1417 	int cpu = task_cpu(t);
1418 	int nesting;
1419 	bool ofl = cpu_is_offline(cpu);
1420 
1421 	if (task_curr(t) && !ofl) {
1422 		// If no chance of heavyweight readers, do it the hard way.
1423 		if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1424 			return -EINVAL;
1425 
1426 		// If heavyweight readers are enabled on the remote task,
1427 		// we can inspect its state despite its currently running.
1428 		// However, we cannot safely change its state.
1429 		n_heavy_reader_attempts++;
1430 		// Check for "running" idle tasks on offline CPUs.
1431 		if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1432 			return -EINVAL; // No quiescent state, do it the hard way.
1433 		n_heavy_reader_updates++;
1434 		nesting = 0;
1435 	} else {
1436 		// The task is not running, so C-language access is safe.
1437 		nesting = t->trc_reader_nesting;
1438 		WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
1439 		if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1440 			n_heavy_reader_ofl_updates++;
1441 	}
1442 
1443 	// If not exiting a read-side critical section, mark as checked
1444 	// so that the grace-period kthread will remove it from the
1445 	// holdout list.
1446 	if (!nesting) {
1447 		rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1448 		return 0;  // In QS, so done.
1449 	}
1450 	if (nesting < 0)
1451 		return -EINVAL; // Reader transitioning, try again later.
1452 
1453 	// The task is in a read-side critical section, so set up its
1454 	// state so that it will update state upon exit from that critical
1455 	// section.
1456 	if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1457 		trc_add_holdout(t, bhp);
1458 	return 0;
1459 }
1460 
1461 /* Attempt to extract the state for the specified task. */
1462 static void trc_wait_for_one_reader(struct task_struct *t,
1463 				    struct list_head *bhp)
1464 {
1465 	int cpu;
1466 
1467 	// If a previous IPI is still in flight, let it complete.
1468 	if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1469 		return;
1470 
1471 	// The current task had better be in a quiescent state.
1472 	if (t == current) {
1473 		rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1474 		WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1475 		return;
1476 	}
1477 
1478 	// Attempt to nail down the task for inspection.
1479 	get_task_struct(t);
1480 	if (!task_call_func(t, trc_inspect_reader, bhp)) {
1481 		put_task_struct(t);
1482 		return;
1483 	}
1484 	put_task_struct(t);
1485 
1486 	// If this task is not yet on the holdout list, then we are in
1487 	// an RCU read-side critical section.  Otherwise, the invocation of
1488 	// trc_add_holdout() that added it to the list did the necessary
1489 	// get_task_struct().  Either way, the task cannot be freed out
1490 	// from under this code.
1491 
1492 	// If currently running, send an IPI, either way, add to list.
1493 	trc_add_holdout(t, bhp);
1494 	if (task_curr(t) &&
1495 	    time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1496 		// The task is currently running, so try IPIing it.
1497 		cpu = task_cpu(t);
1498 
1499 		// If there is already an IPI outstanding, let it happen.
1500 		if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1501 			return;
1502 
1503 		per_cpu(trc_ipi_to_cpu, cpu) = true;
1504 		t->trc_ipi_to_cpu = cpu;
1505 		rcu_tasks_trace.n_ipis++;
1506 		if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1507 			// Just in case there is some other reason for
1508 			// failure than the target CPU being offline.
1509 			WARN_ONCE(1, "%s():  smp_call_function_single() failed for CPU: %d\n",
1510 				  __func__, cpu);
1511 			rcu_tasks_trace.n_ipis_fails++;
1512 			per_cpu(trc_ipi_to_cpu, cpu) = false;
1513 			t->trc_ipi_to_cpu = -1;
1514 		}
1515 	}
1516 }
1517 
1518 /*
1519  * Initialize for first-round processing for the specified task.
1520  * Return false if task is NULL or already taken care of, true otherwise.
1521  */
1522 static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1523 {
1524 	// During early boot when there is only the one boot CPU, there
1525 	// is no idle task for the other CPUs.	Also, the grace-period
1526 	// kthread is always in a quiescent state.  In addition, just return
1527 	// if this task is already on the list.
1528 	if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1529 		return false;
1530 
1531 	rcu_st_need_qs(t, 0);
1532 	t->trc_ipi_to_cpu = -1;
1533 	return true;
1534 }
1535 
1536 /* Do first-round processing for the specified task. */
1537 static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1538 {
1539 	if (rcu_tasks_trace_pertask_prep(t, true))
1540 		trc_wait_for_one_reader(t, hop);
1541 }
1542 
1543 /* Initialize for a new RCU-tasks-trace grace period. */
1544 static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1545 {
1546 	LIST_HEAD(blkd_tasks);
1547 	int cpu;
1548 	unsigned long flags;
1549 	struct rcu_tasks_percpu *rtpcp;
1550 	struct task_struct *t;
1551 
1552 	// There shouldn't be any old IPIs, but...
1553 	for_each_possible_cpu(cpu)
1554 		WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1555 
1556 	// Disable CPU hotplug across the CPU scan for the benefit of
1557 	// any IPIs that might be needed.  This also waits for all readers
1558 	// in CPU-hotplug code paths.
1559 	cpus_read_lock();
1560 
1561 	// These rcu_tasks_trace_pertask_prep() calls are serialized to
1562 	// allow safe access to the hop list.
1563 	for_each_online_cpu(cpu) {
1564 		rcu_read_lock();
1565 		t = cpu_curr_snapshot(cpu);
1566 		if (rcu_tasks_trace_pertask_prep(t, true))
1567 			trc_add_holdout(t, hop);
1568 		rcu_read_unlock();
1569 		cond_resched_tasks_rcu_qs();
1570 	}
1571 
1572 	// Only after all running tasks have been accounted for is it
1573 	// safe to take care of the tasks that have blocked within their
1574 	// current RCU tasks trace read-side critical section.
1575 	for_each_possible_cpu(cpu) {
1576 		rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1577 		raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1578 		list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1579 		while (!list_empty(&blkd_tasks)) {
1580 			rcu_read_lock();
1581 			t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1582 			list_del_init(&t->trc_blkd_node);
1583 			list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1584 			raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1585 			rcu_tasks_trace_pertask(t, hop);
1586 			rcu_read_unlock();
1587 			raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1588 		}
1589 		raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1590 		cond_resched_tasks_rcu_qs();
1591 	}
1592 
1593 	// Re-enable CPU hotplug now that the holdout list is populated.
1594 	cpus_read_unlock();
1595 }
1596 
1597 /*
1598  * Do intermediate processing between task and holdout scans.
1599  */
1600 static void rcu_tasks_trace_postscan(struct list_head *hop)
1601 {
1602 	// Wait for late-stage exiting tasks to finish exiting.
1603 	// These might have passed the call to exit_tasks_rcu_finish().
1604 
1605 	// If you remove the following line, update rcu_trace_implies_rcu_gp()!!!
1606 	synchronize_rcu();
1607 	// Any tasks that exit after this point will set
1608 	// TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1609 }
1610 
1611 /* Communicate task state back to the RCU tasks trace stall warning request. */
1612 struct trc_stall_chk_rdr {
1613 	int nesting;
1614 	int ipi_to_cpu;
1615 	u8 needqs;
1616 };
1617 
1618 static int trc_check_slow_task(struct task_struct *t, void *arg)
1619 {
1620 	struct trc_stall_chk_rdr *trc_rdrp = arg;
1621 
1622 	if (task_curr(t) && cpu_online(task_cpu(t)))
1623 		return false; // It is running, so decline to inspect it.
1624 	trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1625 	trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1626 	trc_rdrp->needqs = rcu_ld_need_qs(t);
1627 	return true;
1628 }
1629 
1630 /* Show the state of a task stalling the current RCU tasks trace GP. */
1631 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1632 {
1633 	int cpu;
1634 	struct trc_stall_chk_rdr trc_rdr;
1635 	bool is_idle_tsk = is_idle_task(t);
1636 
1637 	if (*firstreport) {
1638 		pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1639 		*firstreport = false;
1640 	}
1641 	cpu = task_cpu(t);
1642 	if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1643 		pr_alert("P%d: %c%c\n",
1644 			 t->pid,
1645 			 ".I"[t->trc_ipi_to_cpu >= 0],
1646 			 ".i"[is_idle_tsk]);
1647 	else
1648 		pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1649 			 t->pid,
1650 			 ".I"[trc_rdr.ipi_to_cpu >= 0],
1651 			 ".i"[is_idle_tsk],
1652 			 ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1653 			 ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1654 			 trc_rdr.nesting,
1655 			 " !CN"[trc_rdr.needqs & 0x3],
1656 			 " ?"[trc_rdr.needqs > 0x3],
1657 			 cpu, cpu_online(cpu) ? "" : "(offline)");
1658 	sched_show_task(t);
1659 }
1660 
1661 /* List stalled IPIs for RCU tasks trace. */
1662 static void show_stalled_ipi_trace(void)
1663 {
1664 	int cpu;
1665 
1666 	for_each_possible_cpu(cpu)
1667 		if (per_cpu(trc_ipi_to_cpu, cpu))
1668 			pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1669 }
1670 
1671 /* Do one scan of the holdout list. */
1672 static void check_all_holdout_tasks_trace(struct list_head *hop,
1673 					  bool needreport, bool *firstreport)
1674 {
1675 	struct task_struct *g, *t;
1676 
1677 	// Disable CPU hotplug across the holdout list scan for IPIs.
1678 	cpus_read_lock();
1679 
1680 	list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1681 		// If safe and needed, try to check the current task.
1682 		if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1683 		    !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1684 			trc_wait_for_one_reader(t, hop);
1685 
1686 		// If check succeeded, remove this task from the list.
1687 		if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1688 		    rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1689 			trc_del_holdout(t);
1690 		else if (needreport)
1691 			show_stalled_task_trace(t, firstreport);
1692 		cond_resched_tasks_rcu_qs();
1693 	}
1694 
1695 	// Re-enable CPU hotplug now that the holdout list scan has completed.
1696 	cpus_read_unlock();
1697 
1698 	if (needreport) {
1699 		if (*firstreport)
1700 			pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1701 		show_stalled_ipi_trace();
1702 	}
1703 }
1704 
1705 static void rcu_tasks_trace_empty_fn(void *unused)
1706 {
1707 }
1708 
1709 /* Wait for grace period to complete and provide ordering. */
1710 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1711 {
1712 	int cpu;
1713 
1714 	// Wait for any lingering IPI handlers to complete.  Note that
1715 	// if a CPU has gone offline or transitioned to userspace in the
1716 	// meantime, all IPI handlers should have been drained beforehand.
1717 	// Yes, this assumes that CPUs process IPIs in order.  If that ever
1718 	// changes, there will need to be a recheck and/or timed wait.
1719 	for_each_online_cpu(cpu)
1720 		if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1721 			smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1722 
1723 	smp_mb(); // Caller's code must be ordered after wakeup.
1724 		  // Pairs with pretty much every ordering primitive.
1725 }
1726 
1727 /* Report any needed quiescent state for this exiting task. */
1728 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1729 {
1730 	union rcu_special trs = READ_ONCE(t->trc_reader_special);
1731 
1732 	rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1733 	WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1734 	if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1735 		rcu_read_unlock_trace_special(t);
1736 	else
1737 		WRITE_ONCE(t->trc_reader_nesting, 0);
1738 }
1739 
1740 /**
1741  * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1742  * @rhp: structure to be used for queueing the RCU updates.
1743  * @func: actual callback function to be invoked after the grace period
1744  *
1745  * The callback function will be invoked some time after a trace rcu-tasks
1746  * grace period elapses, in other words after all currently executing
1747  * trace rcu-tasks read-side critical sections have completed. These
1748  * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1749  * and rcu_read_unlock_trace().
1750  *
1751  * See the description of call_rcu() for more detailed information on
1752  * memory ordering guarantees.
1753  */
1754 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1755 {
1756 	call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1757 }
1758 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1759 
1760 /**
1761  * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1762  *
1763  * Control will return to the caller some time after a trace rcu-tasks
1764  * grace period has elapsed, in other words after all currently executing
1765  * trace rcu-tasks read-side critical sections have elapsed. These read-side
1766  * critical sections are delimited by calls to rcu_read_lock_trace()
1767  * and rcu_read_unlock_trace().
1768  *
1769  * This is a very specialized primitive, intended only for a few uses in
1770  * tracing and other situations requiring manipulation of function preambles
1771  * and profiling hooks.  The synchronize_rcu_tasks_trace() function is not
1772  * (yet) intended for heavy use from multiple CPUs.
1773  *
1774  * See the description of synchronize_rcu() for more detailed information
1775  * on memory ordering guarantees.
1776  */
1777 void synchronize_rcu_tasks_trace(void)
1778 {
1779 	RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1780 	synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1781 }
1782 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1783 
1784 /**
1785  * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1786  *
1787  * Although the current implementation is guaranteed to wait, it is not
1788  * obligated to, for example, if there are no pending callbacks.
1789  */
1790 void rcu_barrier_tasks_trace(void)
1791 {
1792 	rcu_barrier_tasks_generic(&rcu_tasks_trace);
1793 }
1794 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1795 
1796 static int __init rcu_spawn_tasks_trace_kthread(void)
1797 {
1798 	cblist_init_generic(&rcu_tasks_trace);
1799 	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1800 		rcu_tasks_trace.gp_sleep = HZ / 10;
1801 		rcu_tasks_trace.init_fract = HZ / 10;
1802 	} else {
1803 		rcu_tasks_trace.gp_sleep = HZ / 200;
1804 		if (rcu_tasks_trace.gp_sleep <= 0)
1805 			rcu_tasks_trace.gp_sleep = 1;
1806 		rcu_tasks_trace.init_fract = HZ / 200;
1807 		if (rcu_tasks_trace.init_fract <= 0)
1808 			rcu_tasks_trace.init_fract = 1;
1809 	}
1810 	rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1811 	rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1812 	rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1813 	rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1814 	rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1815 	return 0;
1816 }
1817 
1818 #if !defined(CONFIG_TINY_RCU)
1819 void show_rcu_tasks_trace_gp_kthread(void)
1820 {
1821 	char buf[64];
1822 
1823 	sprintf(buf, "N%lu h:%lu/%lu/%lu",
1824 		data_race(n_trc_holdouts),
1825 		data_race(n_heavy_reader_ofl_updates),
1826 		data_race(n_heavy_reader_updates),
1827 		data_race(n_heavy_reader_attempts));
1828 	show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1829 }
1830 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1831 #endif // !defined(CONFIG_TINY_RCU)
1832 
1833 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1834 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1835 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1836 
1837 #ifndef CONFIG_TINY_RCU
1838 void show_rcu_tasks_gp_kthreads(void)
1839 {
1840 	show_rcu_tasks_classic_gp_kthread();
1841 	show_rcu_tasks_rude_gp_kthread();
1842 	show_rcu_tasks_trace_gp_kthread();
1843 }
1844 #endif /* #ifndef CONFIG_TINY_RCU */
1845 
1846 #ifdef CONFIG_PROVE_RCU
1847 struct rcu_tasks_test_desc {
1848 	struct rcu_head rh;
1849 	const char *name;
1850 	bool notrun;
1851 	unsigned long runstart;
1852 };
1853 
1854 static struct rcu_tasks_test_desc tests[] = {
1855 	{
1856 		.name = "call_rcu_tasks()",
1857 		/* If not defined, the test is skipped. */
1858 		.notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1859 	},
1860 	{
1861 		.name = "call_rcu_tasks_rude()",
1862 		/* If not defined, the test is skipped. */
1863 		.notrun = IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1864 	},
1865 	{
1866 		.name = "call_rcu_tasks_trace()",
1867 		/* If not defined, the test is skipped. */
1868 		.notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1869 	}
1870 };
1871 
1872 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1873 {
1874 	struct rcu_tasks_test_desc *rttd =
1875 		container_of(rhp, struct rcu_tasks_test_desc, rh);
1876 
1877 	pr_info("Callback from %s invoked.\n", rttd->name);
1878 
1879 	rttd->notrun = false;
1880 }
1881 
1882 static void rcu_tasks_initiate_self_tests(void)
1883 {
1884 	pr_info("Running RCU-tasks wait API self tests\n");
1885 #ifdef CONFIG_TASKS_RCU
1886 	tests[0].runstart = jiffies;
1887 	synchronize_rcu_tasks();
1888 	call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1889 #endif
1890 
1891 #ifdef CONFIG_TASKS_RUDE_RCU
1892 	tests[1].runstart = jiffies;
1893 	synchronize_rcu_tasks_rude();
1894 	call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1895 #endif
1896 
1897 #ifdef CONFIG_TASKS_TRACE_RCU
1898 	tests[2].runstart = jiffies;
1899 	synchronize_rcu_tasks_trace();
1900 	call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1901 #endif
1902 }
1903 
1904 /*
1905  * Return:  0 - test passed
1906  *	    1 - test failed, but have not timed out yet
1907  *	   -1 - test failed and timed out
1908  */
1909 static int rcu_tasks_verify_self_tests(void)
1910 {
1911 	int ret = 0;
1912 	int i;
1913 	unsigned long bst = rcu_task_stall_timeout;
1914 
1915 	if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
1916 		bst = RCU_TASK_BOOT_STALL_TIMEOUT;
1917 	for (i = 0; i < ARRAY_SIZE(tests); i++) {
1918 		while (tests[i].notrun) {		// still hanging.
1919 			if (time_after(jiffies, tests[i].runstart + bst)) {
1920 				pr_err("%s has failed boot-time tests.\n", tests[i].name);
1921 				ret = -1;
1922 				break;
1923 			}
1924 			ret = 1;
1925 			break;
1926 		}
1927 	}
1928 	WARN_ON(ret < 0);
1929 
1930 	return ret;
1931 }
1932 
1933 /*
1934  * Repeat the rcu_tasks_verify_self_tests() call once every second until the
1935  * test passes or has timed out.
1936  */
1937 static struct delayed_work rcu_tasks_verify_work;
1938 static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
1939 {
1940 	int ret = rcu_tasks_verify_self_tests();
1941 
1942 	if (ret <= 0)
1943 		return;
1944 
1945 	/* Test fails but not timed out yet, reschedule another check */
1946 	schedule_delayed_work(&rcu_tasks_verify_work, HZ);
1947 }
1948 
1949 static int rcu_tasks_verify_schedule_work(void)
1950 {
1951 	INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
1952 	rcu_tasks_verify_work_fn(NULL);
1953 	return 0;
1954 }
1955 late_initcall(rcu_tasks_verify_schedule_work);
1956 #else /* #ifdef CONFIG_PROVE_RCU */
1957 static void rcu_tasks_initiate_self_tests(void) { }
1958 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1959 
1960 void __init rcu_init_tasks_generic(void)
1961 {
1962 #ifdef CONFIG_TASKS_RCU
1963 	rcu_spawn_tasks_kthread();
1964 #endif
1965 
1966 #ifdef CONFIG_TASKS_RUDE_RCU
1967 	rcu_spawn_tasks_rude_kthread();
1968 #endif
1969 
1970 #ifdef CONFIG_TASKS_TRACE_RCU
1971 	rcu_spawn_tasks_trace_kthread();
1972 #endif
1973 
1974 	// Run the self-tests.
1975 	rcu_tasks_initiate_self_tests();
1976 }
1977 
1978 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1979 static inline void rcu_tasks_bootup_oddness(void) {}
1980 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */
1981