• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
10 ////////////////////////////////////////////////////////////////////////
11 //
12 // Generic data structures.
13 
14 struct rcu_tasks;
15 typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp);
16 typedef void (*pregp_func_t)(void);
17 typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop);
18 typedef void (*postscan_func_t)(struct list_head *hop);
19 typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp);
20 typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
21 
22 /**
23  * Definition for a Tasks-RCU-like mechanism.
24  * @cbs_head: Head of callback list.
25  * @cbs_tail: Tail pointer for callback list.
26  * @cbs_wq: Wait queue allowning new callback to get kthread's attention.
27  * @cbs_lock: Lock protecting callback list.
28  * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
29  * @gp_func: This flavor's grace-period-wait function.
30  * @gp_state: Grace period's most recent state transition (debugging).
31  * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
32  * @init_fract: Initial backoff sleep interval.
33  * @gp_jiffies: Time of last @gp_state transition.
34  * @gp_start: Most recent grace-period start in jiffies.
35  * @n_gps: Number of grace periods completed since boot.
36  * @n_ipis: Number of IPIs sent to encourage grace periods to end.
37  * @n_ipis_fails: Number of IPI-send failures.
38  * @pregp_func: This flavor's pre-grace-period function (optional).
39  * @pertask_func: This flavor's per-task scan function (optional).
40  * @postscan_func: This flavor's post-task scan function (optional).
41  * @holdout_func: This flavor's holdout-list scan function (optional).
42  * @postgp_func: This flavor's post-grace-period function (optional).
43  * @call_func: This flavor's call_rcu()-equivalent function.
44  * @name: This flavor's textual name.
45  * @kname: This flavor's kthread name.
46  */
47 struct rcu_tasks {
48 	struct rcu_head *cbs_head;
49 	struct rcu_head **cbs_tail;
50 	struct wait_queue_head cbs_wq;
51 	raw_spinlock_t cbs_lock;
52 	int gp_state;
53 	int gp_sleep;
54 	int init_fract;
55 	unsigned long gp_jiffies;
56 	unsigned long gp_start;
57 	unsigned long n_gps;
58 	unsigned long n_ipis;
59 	unsigned long n_ipis_fails;
60 	struct task_struct *kthread_ptr;
61 	rcu_tasks_gp_func_t gp_func;
62 	pregp_func_t pregp_func;
63 	pertask_func_t pertask_func;
64 	postscan_func_t postscan_func;
65 	holdouts_func_t holdouts_func;
66 	postgp_func_t postgp_func;
67 	call_rcu_func_t call_func;
68 	char *name;
69 	char *kname;
70 };
71 
72 #define DEFINE_RCU_TASKS(rt_name, gp, call, n)				\
73 static struct rcu_tasks rt_name =					\
74 {									\
75 	.cbs_tail = &rt_name.cbs_head,					\
76 	.cbs_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rt_name.cbs_wq),	\
77 	.cbs_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_lock),		\
78 	.gp_func = gp,							\
79 	.call_func = call,						\
80 	.name = n,							\
81 	.kname = #rt_name,						\
82 }
83 
84 /* Track exiting tasks in order to allow them to be waited for. */
85 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
86 
87 /* Avoid IPIing CPUs early in the grace period. */
88 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
89 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
90 module_param(rcu_task_ipi_delay, int, 0644);
91 
92 /* Control stall timeouts.  Disable with <= 0, otherwise jiffies till stall. */
93 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
94 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
95 module_param(rcu_task_stall_timeout, int, 0644);
96 
97 /* RCU tasks grace-period state for debugging. */
98 #define RTGS_INIT		 0
99 #define RTGS_WAIT_WAIT_CBS	 1
100 #define RTGS_WAIT_GP		 2
101 #define RTGS_PRE_WAIT_GP	 3
102 #define RTGS_SCAN_TASKLIST	 4
103 #define RTGS_POST_SCAN_TASKLIST	 5
104 #define RTGS_WAIT_SCAN_HOLDOUTS	 6
105 #define RTGS_SCAN_HOLDOUTS	 7
106 #define RTGS_POST_GP		 8
107 #define RTGS_WAIT_READERS	 9
108 #define RTGS_INVOKE_CBS		10
109 #define RTGS_WAIT_CBS		11
110 #ifndef CONFIG_TINY_RCU
111 static const char * const rcu_tasks_gp_state_names[] = {
112 	"RTGS_INIT",
113 	"RTGS_WAIT_WAIT_CBS",
114 	"RTGS_WAIT_GP",
115 	"RTGS_PRE_WAIT_GP",
116 	"RTGS_SCAN_TASKLIST",
117 	"RTGS_POST_SCAN_TASKLIST",
118 	"RTGS_WAIT_SCAN_HOLDOUTS",
119 	"RTGS_SCAN_HOLDOUTS",
120 	"RTGS_POST_GP",
121 	"RTGS_WAIT_READERS",
122 	"RTGS_INVOKE_CBS",
123 	"RTGS_WAIT_CBS",
124 };
125 #endif /* #ifndef CONFIG_TINY_RCU */
126 
127 ////////////////////////////////////////////////////////////////////////
128 //
129 // Generic code.
130 
131 /* Record grace-period phase and time. */
set_tasks_gp_state(struct rcu_tasks * rtp,int newstate)132 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
133 {
134 	rtp->gp_state = newstate;
135 	rtp->gp_jiffies = jiffies;
136 }
137 
138 #ifndef CONFIG_TINY_RCU
139 /* Return state name. */
tasks_gp_state_getname(struct rcu_tasks * rtp)140 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
141 {
142 	int i = data_race(rtp->gp_state); // Let KCSAN detect update races
143 	int j = READ_ONCE(i); // Prevent the compiler from reading twice
144 
145 	if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
146 		return "???";
147 	return rcu_tasks_gp_state_names[j];
148 }
149 #endif /* #ifndef CONFIG_TINY_RCU */
150 
151 // Enqueue a callback for the specified flavor of Tasks RCU.
call_rcu_tasks_generic(struct rcu_head * rhp,rcu_callback_t func,struct rcu_tasks * rtp)152 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
153 				   struct rcu_tasks *rtp)
154 {
155 	unsigned long flags;
156 	bool needwake;
157 
158 	rhp->next = NULL;
159 	rhp->func = func;
160 	raw_spin_lock_irqsave(&rtp->cbs_lock, flags);
161 	needwake = !rtp->cbs_head;
162 	WRITE_ONCE(*rtp->cbs_tail, rhp);
163 	rtp->cbs_tail = &rhp->next;
164 	raw_spin_unlock_irqrestore(&rtp->cbs_lock, flags);
165 	/* We can't create the thread unless interrupts are enabled. */
166 	if (needwake && READ_ONCE(rtp->kthread_ptr))
167 		wake_up(&rtp->cbs_wq);
168 }
169 
170 // Wait for a grace period for the specified flavor of Tasks RCU.
synchronize_rcu_tasks_generic(struct rcu_tasks * rtp)171 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
172 {
173 	/* Complain if the scheduler has not started.  */
174 	WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
175 			 "synchronize_rcu_tasks called too soon");
176 
177 	/* Wait for the grace period. */
178 	wait_rcu_gp(rtp->call_func);
179 }
180 
181 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
rcu_tasks_kthread(void * arg)182 static int __noreturn rcu_tasks_kthread(void *arg)
183 {
184 	unsigned long flags;
185 	struct rcu_head *list;
186 	struct rcu_head *next;
187 	struct rcu_tasks *rtp = arg;
188 
189 	/* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
190 	housekeeping_affine(current, HK_FLAG_RCU);
191 	WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
192 
193 	/*
194 	 * Each pass through the following loop makes one check for
195 	 * newly arrived callbacks, and, if there are some, waits for
196 	 * one RCU-tasks grace period and then invokes the callbacks.
197 	 * This loop is terminated by the system going down.  ;-)
198 	 */
199 	for (;;) {
200 		set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
201 
202 		/* Pick up any new callbacks. */
203 		raw_spin_lock_irqsave(&rtp->cbs_lock, flags);
204 		smp_mb__after_spinlock(); // Order updates vs. GP.
205 		list = rtp->cbs_head;
206 		rtp->cbs_head = NULL;
207 		rtp->cbs_tail = &rtp->cbs_head;
208 		raw_spin_unlock_irqrestore(&rtp->cbs_lock, flags);
209 
210 		/* If there were none, wait a bit and start over. */
211 		if (!list) {
212 			wait_event_interruptible(rtp->cbs_wq,
213 						 READ_ONCE(rtp->cbs_head));
214 			if (!rtp->cbs_head) {
215 				WARN_ON(signal_pending(current));
216 				set_tasks_gp_state(rtp, RTGS_WAIT_WAIT_CBS);
217 				schedule_timeout_idle(HZ/10);
218 			}
219 			continue;
220 		}
221 
222 		// Wait for one grace period.
223 		set_tasks_gp_state(rtp, RTGS_WAIT_GP);
224 		rtp->gp_start = jiffies;
225 		rtp->gp_func(rtp);
226 		rtp->n_gps++;
227 
228 		/* Invoke the callbacks. */
229 		set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
230 		while (list) {
231 			next = list->next;
232 			local_bh_disable();
233 			list->func(list);
234 			local_bh_enable();
235 			list = next;
236 			cond_resched();
237 		}
238 		/* Paranoid sleep to keep this from entering a tight loop */
239 		schedule_timeout_idle(rtp->gp_sleep);
240 	}
241 }
242 
243 /* Spawn RCU-tasks grace-period kthread. */
rcu_spawn_tasks_kthread_generic(struct rcu_tasks * rtp)244 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
245 {
246 	struct task_struct *t;
247 
248 	t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
249 	if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
250 		return;
251 	smp_mb(); /* Ensure others see full kthread. */
252 }
253 
254 #ifndef CONFIG_TINY_RCU
255 
256 /*
257  * Print any non-default Tasks RCU settings.
258  */
rcu_tasks_bootup_oddness(void)259 static void __init rcu_tasks_bootup_oddness(void)
260 {
261 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
262 	if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
263 		pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
264 #endif /* #ifdef CONFIG_TASKS_RCU */
265 #ifdef CONFIG_TASKS_RCU
266 	pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
267 #endif /* #ifdef CONFIG_TASKS_RCU */
268 #ifdef CONFIG_TASKS_RUDE_RCU
269 	pr_info("\tRude variant of Tasks RCU enabled.\n");
270 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
271 #ifdef CONFIG_TASKS_TRACE_RCU
272 	pr_info("\tTracing variant of Tasks RCU enabled.\n");
273 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
274 }
275 
276 #endif /* #ifndef CONFIG_TINY_RCU */
277 
278 #ifndef CONFIG_TINY_RCU
279 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
show_rcu_tasks_generic_gp_kthread(struct rcu_tasks * rtp,char * s)280 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
281 {
282 	pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
283 		rtp->kname,
284 		tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
285 		jiffies - data_race(rtp->gp_jiffies),
286 		data_race(rtp->n_gps),
287 		data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
288 		".k"[!!data_race(rtp->kthread_ptr)],
289 		".C"[!!data_race(rtp->cbs_head)],
290 		s);
291 }
292 #endif /* #ifndef CONFIG_TINY_RCU */
293 
294 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
295 
296 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
297 
298 ////////////////////////////////////////////////////////////////////////
299 //
300 // Shared code between task-list-scanning variants of Tasks RCU.
301 
302 /* Wait for one RCU-tasks grace period. */
rcu_tasks_wait_gp(struct rcu_tasks * rtp)303 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
304 {
305 	struct task_struct *g, *t;
306 	unsigned long lastreport;
307 	LIST_HEAD(holdouts);
308 	int fract;
309 
310 	set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
311 	rtp->pregp_func();
312 
313 	/*
314 	 * There were callbacks, so we need to wait for an RCU-tasks
315 	 * grace period.  Start off by scanning the task list for tasks
316 	 * that are not already voluntarily blocked.  Mark these tasks
317 	 * and make a list of them in holdouts.
318 	 */
319 	set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
320 	rcu_read_lock();
321 	for_each_process_thread(g, t)
322 		rtp->pertask_func(t, &holdouts);
323 	rcu_read_unlock();
324 
325 	set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
326 	rtp->postscan_func(&holdouts);
327 
328 	/*
329 	 * Each pass through the following loop scans the list of holdout
330 	 * tasks, removing any that are no longer holdouts.  When the list
331 	 * is empty, we are done.
332 	 */
333 	lastreport = jiffies;
334 
335 	// Start off with initial wait and slowly back off to 1 HZ wait.
336 	fract = rtp->init_fract;
337 	if (fract > HZ)
338 		fract = HZ;
339 
340 	for (;;) {
341 		bool firstreport;
342 		bool needreport;
343 		int rtst;
344 
345 		if (list_empty(&holdouts))
346 			break;
347 
348 		/* Slowly back off waiting for holdouts */
349 		set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
350 		schedule_timeout_idle(HZ/fract);
351 
352 		if (fract > 1)
353 			fract--;
354 
355 		rtst = READ_ONCE(rcu_task_stall_timeout);
356 		needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
357 		if (needreport)
358 			lastreport = jiffies;
359 		firstreport = true;
360 		WARN_ON(signal_pending(current));
361 		set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
362 		rtp->holdouts_func(&holdouts, needreport, &firstreport);
363 	}
364 
365 	set_tasks_gp_state(rtp, RTGS_POST_GP);
366 	rtp->postgp_func(rtp);
367 }
368 
369 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
370 
371 #ifdef CONFIG_TASKS_RCU
372 
373 ////////////////////////////////////////////////////////////////////////
374 //
375 // Simple variant of RCU whose quiescent states are voluntary context
376 // switch, cond_resched_rcu_qs(), user-space execution, and idle.
377 // As such, grace periods can take one good long time.  There are no
378 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
379 // because this implementation is intended to get the system into a safe
380 // state for some of the manipulations involved in tracing and the like.
381 // Finally, this implementation does not support high call_rcu_tasks()
382 // rates from multiple CPUs.  If this is required, per-CPU callback lists
383 // will be needed.
384 
385 /* Pre-grace-period preparation. */
rcu_tasks_pregp_step(void)386 static void rcu_tasks_pregp_step(void)
387 {
388 	/*
389 	 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
390 	 * to complete.  Invoking synchronize_rcu() suffices because all
391 	 * these transitions occur with interrupts disabled.  Without this
392 	 * synchronize_rcu(), a read-side critical section that started
393 	 * before the grace period might be incorrectly seen as having
394 	 * started after the grace period.
395 	 *
396 	 * This synchronize_rcu() also dispenses with the need for a
397 	 * memory barrier on the first store to t->rcu_tasks_holdout,
398 	 * as it forces the store to happen after the beginning of the
399 	 * grace period.
400 	 */
401 	synchronize_rcu();
402 }
403 
404 /* Per-task initial processing. */
rcu_tasks_pertask(struct task_struct * t,struct list_head * hop)405 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
406 {
407 	if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
408 		get_task_struct(t);
409 		t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
410 		WRITE_ONCE(t->rcu_tasks_holdout, true);
411 		list_add(&t->rcu_tasks_holdout_list, hop);
412 	}
413 }
414 
415 /* Processing between scanning taskslist and draining the holdout list. */
rcu_tasks_postscan(struct list_head * hop)416 static void rcu_tasks_postscan(struct list_head *hop)
417 {
418 	/*
419 	 * Wait for tasks that are in the process of exiting.  This
420 	 * does only part of the job, ensuring that all tasks that were
421 	 * previously exiting reach the point where they have disabled
422 	 * preemption, allowing the later synchronize_rcu() to finish
423 	 * the job.
424 	 */
425 	synchronize_srcu(&tasks_rcu_exit_srcu);
426 }
427 
428 /* See if tasks are still holding out, complain if so. */
check_holdout_task(struct task_struct * t,bool needreport,bool * firstreport)429 static void check_holdout_task(struct task_struct *t,
430 			       bool needreport, bool *firstreport)
431 {
432 	int cpu;
433 
434 	if (!READ_ONCE(t->rcu_tasks_holdout) ||
435 	    t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
436 	    !READ_ONCE(t->on_rq) ||
437 	    (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
438 	     !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
439 		WRITE_ONCE(t->rcu_tasks_holdout, false);
440 		list_del_init(&t->rcu_tasks_holdout_list);
441 		put_task_struct(t);
442 		return;
443 	}
444 	rcu_request_urgent_qs_task(t);
445 	if (!needreport)
446 		return;
447 	if (*firstreport) {
448 		pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
449 		*firstreport = false;
450 	}
451 	cpu = task_cpu(t);
452 	pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
453 		 t, ".I"[is_idle_task(t)],
454 		 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
455 		 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
456 		 t->rcu_tasks_idle_cpu, cpu);
457 	sched_show_task(t);
458 }
459 
460 /* Scan the holdout lists for tasks no longer holding out. */
check_all_holdout_tasks(struct list_head * hop,bool needreport,bool * firstreport)461 static void check_all_holdout_tasks(struct list_head *hop,
462 				    bool needreport, bool *firstreport)
463 {
464 	struct task_struct *t, *t1;
465 
466 	list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
467 		check_holdout_task(t, needreport, firstreport);
468 		cond_resched();
469 	}
470 }
471 
472 /* Finish off the Tasks-RCU grace period. */
rcu_tasks_postgp(struct rcu_tasks * rtp)473 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
474 {
475 	/*
476 	 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
477 	 * memory barriers prior to them in the schedule() path, memory
478 	 * reordering on other CPUs could cause their RCU-tasks read-side
479 	 * critical sections to extend past the end of the grace period.
480 	 * However, because these ->nvcsw updates are carried out with
481 	 * interrupts disabled, we can use synchronize_rcu() to force the
482 	 * needed ordering on all such CPUs.
483 	 *
484 	 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
485 	 * accesses to be within the grace period, avoiding the need for
486 	 * memory barriers for ->rcu_tasks_holdout accesses.
487 	 *
488 	 * In addition, this synchronize_rcu() waits for exiting tasks
489 	 * to complete their final preempt_disable() region of execution,
490 	 * cleaning up after the synchronize_srcu() above.
491 	 */
492 	synchronize_rcu();
493 }
494 
495 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
496 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
497 
498 /**
499  * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
500  * @rhp: structure to be used for queueing the RCU updates.
501  * @func: actual callback function to be invoked after the grace period
502  *
503  * The callback function will be invoked some time after a full grace
504  * period elapses, in other words after all currently executing RCU
505  * read-side critical sections have completed. call_rcu_tasks() assumes
506  * that the read-side critical sections end at a voluntary context
507  * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
508  * or transition to usermode execution.  As such, there are no read-side
509  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
510  * this primitive is intended to determine that all tasks have passed
511  * through a safe state, not so much for data-strcuture synchronization.
512  *
513  * See the description of call_rcu() for more detailed information on
514  * memory ordering guarantees.
515  */
call_rcu_tasks(struct rcu_head * rhp,rcu_callback_t func)516 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
517 {
518 	call_rcu_tasks_generic(rhp, func, &rcu_tasks);
519 }
520 EXPORT_SYMBOL_GPL(call_rcu_tasks);
521 
522 /**
523  * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
524  *
525  * Control will return to the caller some time after a full rcu-tasks
526  * grace period has elapsed, in other words after all currently
527  * executing rcu-tasks read-side critical sections have elapsed.  These
528  * read-side critical sections are delimited by calls to schedule(),
529  * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
530  * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
531  *
532  * This is a very specialized primitive, intended only for a few uses in
533  * tracing and other situations requiring manipulation of function
534  * preambles and profiling hooks.  The synchronize_rcu_tasks() function
535  * is not (yet) intended for heavy use from multiple CPUs.
536  *
537  * See the description of synchronize_rcu() for more detailed information
538  * on memory ordering guarantees.
539  */
synchronize_rcu_tasks(void)540 void synchronize_rcu_tasks(void)
541 {
542 	synchronize_rcu_tasks_generic(&rcu_tasks);
543 }
544 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
545 
546 /**
547  * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
548  *
549  * Although the current implementation is guaranteed to wait, it is not
550  * obligated to, for example, if there are no pending callbacks.
551  */
rcu_barrier_tasks(void)552 void rcu_barrier_tasks(void)
553 {
554 	/* There is only one callback queue, so this is easy.  ;-) */
555 	synchronize_rcu_tasks();
556 }
557 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
558 
rcu_spawn_tasks_kthread(void)559 static int __init rcu_spawn_tasks_kthread(void)
560 {
561 	rcu_tasks.gp_sleep = HZ / 10;
562 	rcu_tasks.init_fract = 10;
563 	rcu_tasks.pregp_func = rcu_tasks_pregp_step;
564 	rcu_tasks.pertask_func = rcu_tasks_pertask;
565 	rcu_tasks.postscan_func = rcu_tasks_postscan;
566 	rcu_tasks.holdouts_func = check_all_holdout_tasks;
567 	rcu_tasks.postgp_func = rcu_tasks_postgp;
568 	rcu_spawn_tasks_kthread_generic(&rcu_tasks);
569 	return 0;
570 }
571 
572 #ifndef CONFIG_TINY_RCU
show_rcu_tasks_classic_gp_kthread(void)573 static void show_rcu_tasks_classic_gp_kthread(void)
574 {
575 	show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
576 }
577 #endif /* #ifndef CONFIG_TINY_RCU */
578 
579 /* Do the srcu_read_lock() for the above synchronize_srcu().  */
exit_tasks_rcu_start(void)580 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
581 {
582 	preempt_disable();
583 	current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
584 	preempt_enable();
585 }
586 
587 /* Do the srcu_read_unlock() for the above synchronize_srcu().  */
exit_tasks_rcu_finish(void)588 void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu)
589 {
590 	struct task_struct *t = current;
591 
592 	preempt_disable();
593 	__srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
594 	preempt_enable();
595 	exit_tasks_rcu_finish_trace(t);
596 }
597 
598 #else /* #ifdef CONFIG_TASKS_RCU */
show_rcu_tasks_classic_gp_kthread(void)599 static inline void show_rcu_tasks_classic_gp_kthread(void) { }
exit_tasks_rcu_start(void)600 void exit_tasks_rcu_start(void) { }
exit_tasks_rcu_finish(void)601 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
602 #endif /* #else #ifdef CONFIG_TASKS_RCU */
603 
604 #ifdef CONFIG_TASKS_RUDE_RCU
605 
606 ////////////////////////////////////////////////////////////////////////
607 //
608 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
609 // passing an empty function to schedule_on_each_cpu().  This approach
610 // provides an asynchronous call_rcu_tasks_rude() API and batching
611 // of concurrent calls to the synchronous synchronize_rcu_rude() API.
612 // This sends IPIs far and wide and induces otherwise unnecessary context
613 // switches on all online CPUs, whether idle or not.
614 
615 // Empty function to allow workqueues to force a context switch.
rcu_tasks_be_rude(struct work_struct * work)616 static void rcu_tasks_be_rude(struct work_struct *work)
617 {
618 }
619 
620 // Wait for one rude RCU-tasks grace period.
rcu_tasks_rude_wait_gp(struct rcu_tasks * rtp)621 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
622 {
623 	if (num_online_cpus() <= 1)
624 		return;	// Fastpath for only one CPU.
625 
626 	rtp->n_ipis += cpumask_weight(cpu_online_mask);
627 	schedule_on_each_cpu(rcu_tasks_be_rude);
628 }
629 
630 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
631 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
632 		 "RCU Tasks Rude");
633 
634 /**
635  * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
636  * @rhp: structure to be used for queueing the RCU updates.
637  * @func: actual callback function to be invoked after the grace period
638  *
639  * The callback function will be invoked some time after a full grace
640  * period elapses, in other words after all currently executing RCU
641  * read-side critical sections have completed. call_rcu_tasks_rude()
642  * assumes that the read-side critical sections end at context switch,
643  * cond_resched_rcu_qs(), or transition to usermode execution.  As such,
644  * there are no read-side primitives analogous to rcu_read_lock() and
645  * rcu_read_unlock() because this primitive is intended to determine
646  * that all tasks have passed through a safe state, not so much for
647  * data-strcuture synchronization.
648  *
649  * See the description of call_rcu() for more detailed information on
650  * memory ordering guarantees.
651  */
call_rcu_tasks_rude(struct rcu_head * rhp,rcu_callback_t func)652 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
653 {
654 	call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
655 }
656 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
657 
658 /**
659  * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
660  *
661  * Control will return to the caller some time after a rude rcu-tasks
662  * grace period has elapsed, in other words after all currently
663  * executing rcu-tasks read-side critical sections have elapsed.  These
664  * read-side critical sections are delimited by calls to schedule(),
665  * cond_resched_tasks_rcu_qs(), userspace execution, and (in theory,
666  * anyway) cond_resched().
667  *
668  * This is a very specialized primitive, intended only for a few uses in
669  * tracing and other situations requiring manipulation of function preambles
670  * and profiling hooks.  The synchronize_rcu_tasks_rude() function is not
671  * (yet) intended for heavy use from multiple CPUs.
672  *
673  * See the description of synchronize_rcu() for more detailed information
674  * on memory ordering guarantees.
675  */
synchronize_rcu_tasks_rude(void)676 void synchronize_rcu_tasks_rude(void)
677 {
678 	synchronize_rcu_tasks_generic(&rcu_tasks_rude);
679 }
680 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
681 
682 /**
683  * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
684  *
685  * Although the current implementation is guaranteed to wait, it is not
686  * obligated to, for example, if there are no pending callbacks.
687  */
rcu_barrier_tasks_rude(void)688 void rcu_barrier_tasks_rude(void)
689 {
690 	/* There is only one callback queue, so this is easy.  ;-) */
691 	synchronize_rcu_tasks_rude();
692 }
693 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
694 
rcu_spawn_tasks_rude_kthread(void)695 static int __init rcu_spawn_tasks_rude_kthread(void)
696 {
697 	rcu_tasks_rude.gp_sleep = HZ / 10;
698 	rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
699 	return 0;
700 }
701 
702 #ifndef CONFIG_TINY_RCU
show_rcu_tasks_rude_gp_kthread(void)703 static void show_rcu_tasks_rude_gp_kthread(void)
704 {
705 	show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
706 }
707 #endif /* #ifndef CONFIG_TINY_RCU */
708 
709 #else /* #ifdef CONFIG_TASKS_RUDE_RCU */
show_rcu_tasks_rude_gp_kthread(void)710 static void show_rcu_tasks_rude_gp_kthread(void) {}
711 #endif /* #else #ifdef CONFIG_TASKS_RUDE_RCU */
712 
713 ////////////////////////////////////////////////////////////////////////
714 //
715 // Tracing variant of Tasks RCU.  This variant is designed to be used
716 // to protect tracing hooks, including those of BPF.  This variant
717 // therefore:
718 //
719 // 1.	Has explicit read-side markers to allow finite grace periods
720 //	in the face of in-kernel loops for PREEMPT=n builds.
721 //
722 // 2.	Protects code in the idle loop, exception entry/exit, and
723 //	CPU-hotplug code paths, similar to the capabilities of SRCU.
724 //
725 // 3.	Avoids expensive read-side instruction, having overhead similar
726 //	to that of Preemptible RCU.
727 //
728 // There are of course downsides.  The grace-period code can send IPIs to
729 // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace.
730 // It is necessary to scan the full tasklist, much as for Tasks RCU.  There
731 // is a single callback queue guarded by a single lock, again, much as for
732 // Tasks RCU.  If needed, these downsides can be at least partially remedied.
733 //
734 // Perhaps most important, this variant of RCU does not affect the vanilla
735 // flavors, rcu_preempt and rcu_sched.  The fact that RCU Tasks Trace
736 // readers can operate from idle, offline, and exception entry/exit in no
737 // way allows rcu_preempt and rcu_sched readers to also do so.
738 
739 // The lockdep state must be outside of #ifdef to be useful.
740 #ifdef CONFIG_DEBUG_LOCK_ALLOC
741 static struct lock_class_key rcu_lock_trace_key;
742 struct lockdep_map rcu_trace_lock_map =
743 	STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
744 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
745 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
746 
747 #ifdef CONFIG_TASKS_TRACE_RCU
748 
749 static atomic_t trc_n_readers_need_end;		// Number of waited-for readers.
750 static DECLARE_WAIT_QUEUE_HEAD(trc_wait);	// List of holdout tasks.
751 
752 // Record outstanding IPIs to each CPU.  No point in sending two...
753 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
754 
755 // The number of detections of task quiescent state relying on
756 // heavyweight readers executing explicit memory barriers.
757 static unsigned long n_heavy_reader_attempts;
758 static unsigned long n_heavy_reader_updates;
759 static unsigned long n_heavy_reader_ofl_updates;
760 
761 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
762 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
763 		 "RCU Tasks Trace");
764 
765 /*
766  * This irq_work handler allows rcu_read_unlock_trace() to be invoked
767  * while the scheduler locks are held.
768  */
rcu_read_unlock_iw(struct irq_work * iwp)769 static void rcu_read_unlock_iw(struct irq_work *iwp)
770 {
771 	wake_up(&trc_wait);
772 }
773 static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw);
774 
775 /* If we are the last reader, wake up the grace-period kthread. */
rcu_read_unlock_trace_special(struct task_struct * t,int nesting)776 void rcu_read_unlock_trace_special(struct task_struct *t, int nesting)
777 {
778 	int nq = t->trc_reader_special.b.need_qs;
779 
780 	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) &&
781 	    t->trc_reader_special.b.need_mb)
782 		smp_mb(); // Pairs with update-side barriers.
783 	// Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
784 	if (nq)
785 		WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
786 	WRITE_ONCE(t->trc_reader_nesting, nesting);
787 	if (nq && atomic_dec_and_test(&trc_n_readers_need_end))
788 		irq_work_queue(&rcu_tasks_trace_iw);
789 }
790 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
791 
792 /* Add a task to the holdout list, if it is not already on the list. */
trc_add_holdout(struct task_struct * t,struct list_head * bhp)793 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
794 {
795 	if (list_empty(&t->trc_holdout_list)) {
796 		get_task_struct(t);
797 		list_add(&t->trc_holdout_list, bhp);
798 	}
799 }
800 
801 /* Remove a task from the holdout list, if it is in fact present. */
trc_del_holdout(struct task_struct * t)802 static void trc_del_holdout(struct task_struct *t)
803 {
804 	if (!list_empty(&t->trc_holdout_list)) {
805 		list_del_init(&t->trc_holdout_list);
806 		put_task_struct(t);
807 	}
808 }
809 
810 /* IPI handler to check task state. */
trc_read_check_handler(void * t_in)811 static void trc_read_check_handler(void *t_in)
812 {
813 	struct task_struct *t = current;
814 	struct task_struct *texp = t_in;
815 
816 	// If the task is no longer running on this CPU, leave.
817 	if (unlikely(texp != t)) {
818 		if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
819 			wake_up(&trc_wait);
820 		goto reset_ipi; // Already on holdout list, so will check later.
821 	}
822 
823 	// If the task is not in a read-side critical section, and
824 	// if this is the last reader, awaken the grace-period kthread.
825 	if (likely(!t->trc_reader_nesting)) {
826 		if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
827 			wake_up(&trc_wait);
828 		// Mark as checked after decrement to avoid false
829 		// positives on the above WARN_ON_ONCE().
830 		WRITE_ONCE(t->trc_reader_checked, true);
831 		goto reset_ipi;
832 	}
833 	// If we are racing with an rcu_read_unlock_trace(), try again later.
834 	if (unlikely(t->trc_reader_nesting < 0)) {
835 		if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
836 			wake_up(&trc_wait);
837 		goto reset_ipi;
838 	}
839 	WRITE_ONCE(t->trc_reader_checked, true);
840 
841 	// Get here if the task is in a read-side critical section.  Set
842 	// its state so that it will awaken the grace-period kthread upon
843 	// exit from that critical section.
844 	WARN_ON_ONCE(t->trc_reader_special.b.need_qs);
845 	WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
846 
847 reset_ipi:
848 	// Allow future IPIs to be sent on CPU and for task.
849 	// Also order this IPI handler against any later manipulations of
850 	// the intended task.
851 	smp_store_release(&per_cpu(trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
852 	smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
853 }
854 
855 /* Callback function for scheduler to check locked-down task.  */
trc_inspect_reader(struct task_struct * t,void * arg)856 static bool trc_inspect_reader(struct task_struct *t, void *arg)
857 {
858 	int cpu = task_cpu(t);
859 	bool in_qs = false;
860 	bool ofl = cpu_is_offline(cpu);
861 
862 	if (task_curr(t)) {
863 		WARN_ON_ONCE(ofl && !is_idle_task(t));
864 
865 		// If no chance of heavyweight readers, do it the hard way.
866 		if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
867 			return false;
868 
869 		// If heavyweight readers are enabled on the remote task,
870 		// we can inspect its state despite its currently running.
871 		// However, we cannot safely change its state.
872 		n_heavy_reader_attempts++;
873 		if (!ofl && // Check for "running" idle tasks on offline CPUs.
874 		    !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
875 			return false; // No quiescent state, do it the hard way.
876 		n_heavy_reader_updates++;
877 		if (ofl)
878 			n_heavy_reader_ofl_updates++;
879 		in_qs = true;
880 	} else {
881 		in_qs = likely(!t->trc_reader_nesting);
882 	}
883 
884 	// Mark as checked so that the grace-period kthread will
885 	// remove it from the holdout list.
886 	t->trc_reader_checked = true;
887 
888 	if (in_qs)
889 		return true;  // Already in quiescent state, done!!!
890 
891 	// The task is in a read-side critical section, so set up its
892 	// state so that it will awaken the grace-period kthread upon exit
893 	// from that critical section.
894 	atomic_inc(&trc_n_readers_need_end); // One more to wait on.
895 	WARN_ON_ONCE(t->trc_reader_special.b.need_qs);
896 	WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
897 	return true;
898 }
899 
900 /* Attempt to extract the state for the specified task. */
trc_wait_for_one_reader(struct task_struct * t,struct list_head * bhp)901 static void trc_wait_for_one_reader(struct task_struct *t,
902 				    struct list_head *bhp)
903 {
904 	int cpu;
905 
906 	// If a previous IPI is still in flight, let it complete.
907 	if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
908 		return;
909 
910 	// The current task had better be in a quiescent state.
911 	if (t == current) {
912 		t->trc_reader_checked = true;
913 		WARN_ON_ONCE(t->trc_reader_nesting);
914 		return;
915 	}
916 
917 	// Attempt to nail down the task for inspection.
918 	get_task_struct(t);
919 	if (try_invoke_on_locked_down_task(t, trc_inspect_reader, NULL)) {
920 		put_task_struct(t);
921 		return;
922 	}
923 	put_task_struct(t);
924 
925 	// If currently running, send an IPI, either way, add to list.
926 	trc_add_holdout(t, bhp);
927 	if (task_curr(t) &&
928 	    time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
929 		// The task is currently running, so try IPIing it.
930 		cpu = task_cpu(t);
931 
932 		// If there is already an IPI outstanding, let it happen.
933 		if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
934 			return;
935 
936 		atomic_inc(&trc_n_readers_need_end);
937 		per_cpu(trc_ipi_to_cpu, cpu) = true;
938 		t->trc_ipi_to_cpu = cpu;
939 		rcu_tasks_trace.n_ipis++;
940 		if (smp_call_function_single(cpu,
941 					     trc_read_check_handler, t, 0)) {
942 			// Just in case there is some other reason for
943 			// failure than the target CPU being offline.
944 			rcu_tasks_trace.n_ipis_fails++;
945 			per_cpu(trc_ipi_to_cpu, cpu) = false;
946 			t->trc_ipi_to_cpu = cpu;
947 			if (atomic_dec_and_test(&trc_n_readers_need_end)) {
948 				WARN_ON_ONCE(1);
949 				wake_up(&trc_wait);
950 			}
951 		}
952 	}
953 }
954 
955 /* Initialize for a new RCU-tasks-trace grace period. */
rcu_tasks_trace_pregp_step(void)956 static void rcu_tasks_trace_pregp_step(void)
957 {
958 	int cpu;
959 
960 	// Allow for fast-acting IPIs.
961 	atomic_set(&trc_n_readers_need_end, 1);
962 
963 	// There shouldn't be any old IPIs, but...
964 	for_each_possible_cpu(cpu)
965 		WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
966 
967 	// Disable CPU hotplug across the tasklist scan.
968 	// This also waits for all readers in CPU-hotplug code paths.
969 	cpus_read_lock();
970 }
971 
972 /* Do first-round processing for the specified task. */
rcu_tasks_trace_pertask(struct task_struct * t,struct list_head * hop)973 static void rcu_tasks_trace_pertask(struct task_struct *t,
974 				    struct list_head *hop)
975 {
976 	// During early boot when there is only the one boot CPU, there
977 	// is no idle task for the other CPUs. Just return.
978 	if (unlikely(t == NULL))
979 		return;
980 
981 	WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
982 	WRITE_ONCE(t->trc_reader_checked, false);
983 	t->trc_ipi_to_cpu = -1;
984 	trc_wait_for_one_reader(t, hop);
985 }
986 
987 /*
988  * Do intermediate processing between task and holdout scans and
989  * pick up the idle tasks.
990  */
rcu_tasks_trace_postscan(struct list_head * hop)991 static void rcu_tasks_trace_postscan(struct list_head *hop)
992 {
993 	int cpu;
994 
995 	for_each_possible_cpu(cpu)
996 		rcu_tasks_trace_pertask(idle_task(cpu), hop);
997 
998 	// Re-enable CPU hotplug now that the tasklist scan has completed.
999 	cpus_read_unlock();
1000 
1001 	// Wait for late-stage exiting tasks to finish exiting.
1002 	// These might have passed the call to exit_tasks_rcu_finish().
1003 	synchronize_rcu();
1004 	// Any tasks that exit after this point will set ->trc_reader_checked.
1005 }
1006 
1007 /* Show the state of a task stalling the current RCU tasks trace GP. */
show_stalled_task_trace(struct task_struct * t,bool * firstreport)1008 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1009 {
1010 	int cpu;
1011 
1012 	if (*firstreport) {
1013 		pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1014 		*firstreport = false;
1015 	}
1016 	// FIXME: This should attempt to use try_invoke_on_nonrunning_task().
1017 	cpu = task_cpu(t);
1018 	pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n",
1019 		 t->pid,
1020 		 ".I"[READ_ONCE(t->trc_ipi_to_cpu) > 0],
1021 		 ".i"[is_idle_task(t)],
1022 		 ".N"[cpu > 0 && tick_nohz_full_cpu(cpu)],
1023 		 t->trc_reader_nesting,
1024 		 " N"[!!t->trc_reader_special.b.need_qs],
1025 		 cpu);
1026 	sched_show_task(t);
1027 }
1028 
1029 /* List stalled IPIs for RCU tasks trace. */
show_stalled_ipi_trace(void)1030 static void show_stalled_ipi_trace(void)
1031 {
1032 	int cpu;
1033 
1034 	for_each_possible_cpu(cpu)
1035 		if (per_cpu(trc_ipi_to_cpu, cpu))
1036 			pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1037 }
1038 
1039 /* Do one scan of the holdout list. */
check_all_holdout_tasks_trace(struct list_head * hop,bool needreport,bool * firstreport)1040 static void check_all_holdout_tasks_trace(struct list_head *hop,
1041 					  bool needreport, bool *firstreport)
1042 {
1043 	struct task_struct *g, *t;
1044 
1045 	// Disable CPU hotplug across the holdout list scan.
1046 	cpus_read_lock();
1047 
1048 	list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1049 		// If safe and needed, try to check the current task.
1050 		if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1051 		    !READ_ONCE(t->trc_reader_checked))
1052 			trc_wait_for_one_reader(t, hop);
1053 
1054 		// If check succeeded, remove this task from the list.
1055 		if (READ_ONCE(t->trc_reader_checked))
1056 			trc_del_holdout(t);
1057 		else if (needreport)
1058 			show_stalled_task_trace(t, firstreport);
1059 	}
1060 
1061 	// Re-enable CPU hotplug now that the holdout list scan has completed.
1062 	cpus_read_unlock();
1063 
1064 	if (needreport) {
1065 		if (firstreport)
1066 			pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1067 		show_stalled_ipi_trace();
1068 	}
1069 }
1070 
1071 /* Wait for grace period to complete and provide ordering. */
rcu_tasks_trace_postgp(struct rcu_tasks * rtp)1072 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1073 {
1074 	bool firstreport;
1075 	struct task_struct *g, *t;
1076 	LIST_HEAD(holdouts);
1077 	long ret;
1078 
1079 	// Remove the safety count.
1080 	smp_mb__before_atomic();  // Order vs. earlier atomics
1081 	atomic_dec(&trc_n_readers_need_end);
1082 	smp_mb__after_atomic();  // Order vs. later atomics
1083 
1084 	// Wait for readers.
1085 	set_tasks_gp_state(rtp, RTGS_WAIT_READERS);
1086 	for (;;) {
1087 		ret = wait_event_idle_exclusive_timeout(
1088 				trc_wait,
1089 				atomic_read(&trc_n_readers_need_end) == 0,
1090 				READ_ONCE(rcu_task_stall_timeout));
1091 		if (ret)
1092 			break;  // Count reached zero.
1093 		// Stall warning time, so make a list of the offenders.
1094 		rcu_read_lock();
1095 		for_each_process_thread(g, t)
1096 			if (READ_ONCE(t->trc_reader_special.b.need_qs))
1097 				trc_add_holdout(t, &holdouts);
1098 		rcu_read_unlock();
1099 		firstreport = true;
1100 		list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) {
1101 			if (READ_ONCE(t->trc_reader_special.b.need_qs))
1102 				show_stalled_task_trace(t, &firstreport);
1103 			trc_del_holdout(t); // Release task_struct reference.
1104 		}
1105 		if (firstreport)
1106 			pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n");
1107 		show_stalled_ipi_trace();
1108 		pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end));
1109 	}
1110 	smp_mb(); // Caller's code must be ordered after wakeup.
1111 		  // Pairs with pretty much every ordering primitive.
1112 }
1113 
1114 /* Report any needed quiescent state for this exiting task. */
exit_tasks_rcu_finish_trace(struct task_struct * t)1115 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1116 {
1117 	WRITE_ONCE(t->trc_reader_checked, true);
1118 	WARN_ON_ONCE(t->trc_reader_nesting);
1119 	WRITE_ONCE(t->trc_reader_nesting, 0);
1120 	if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)))
1121 		rcu_read_unlock_trace_special(t, 0);
1122 }
1123 
1124 /**
1125  * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1126  * @rhp: structure to be used for queueing the RCU updates.
1127  * @func: actual callback function to be invoked after the grace period
1128  *
1129  * The callback function will be invoked some time after a full grace
1130  * period elapses, in other words after all currently executing RCU
1131  * read-side critical sections have completed. call_rcu_tasks_trace()
1132  * assumes that the read-side critical sections end at context switch,
1133  * cond_resched_rcu_qs(), or transition to usermode execution.  As such,
1134  * there are no read-side primitives analogous to rcu_read_lock() and
1135  * rcu_read_unlock() because this primitive is intended to determine
1136  * that all tasks have passed through a safe state, not so much for
1137  * data-strcuture synchronization.
1138  *
1139  * See the description of call_rcu() for more detailed information on
1140  * memory ordering guarantees.
1141  */
call_rcu_tasks_trace(struct rcu_head * rhp,rcu_callback_t func)1142 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1143 {
1144 	call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1145 }
1146 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1147 
1148 /**
1149  * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1150  *
1151  * Control will return to the caller some time after a trace rcu-tasks
1152  * grace period has elapsed, in other words after all currently executing
1153  * rcu-tasks read-side critical sections have elapsed.  These read-side
1154  * critical sections are delimited by calls to rcu_read_lock_trace()
1155  * and rcu_read_unlock_trace().
1156  *
1157  * This is a very specialized primitive, intended only for a few uses in
1158  * tracing and other situations requiring manipulation of function preambles
1159  * and profiling hooks.  The synchronize_rcu_tasks_trace() function is not
1160  * (yet) intended for heavy use from multiple CPUs.
1161  *
1162  * See the description of synchronize_rcu() for more detailed information
1163  * on memory ordering guarantees.
1164  */
synchronize_rcu_tasks_trace(void)1165 void synchronize_rcu_tasks_trace(void)
1166 {
1167 	RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1168 	synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1169 }
1170 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1171 
1172 /**
1173  * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1174  *
1175  * Although the current implementation is guaranteed to wait, it is not
1176  * obligated to, for example, if there are no pending callbacks.
1177  */
rcu_barrier_tasks_trace(void)1178 void rcu_barrier_tasks_trace(void)
1179 {
1180 	/* There is only one callback queue, so this is easy.  ;-) */
1181 	synchronize_rcu_tasks_trace();
1182 }
1183 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1184 
rcu_spawn_tasks_trace_kthread(void)1185 static int __init rcu_spawn_tasks_trace_kthread(void)
1186 {
1187 	if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1188 		rcu_tasks_trace.gp_sleep = HZ / 10;
1189 		rcu_tasks_trace.init_fract = 10;
1190 	} else {
1191 		rcu_tasks_trace.gp_sleep = HZ / 200;
1192 		if (rcu_tasks_trace.gp_sleep <= 0)
1193 			rcu_tasks_trace.gp_sleep = 1;
1194 		rcu_tasks_trace.init_fract = HZ / 5;
1195 		if (rcu_tasks_trace.init_fract <= 0)
1196 			rcu_tasks_trace.init_fract = 1;
1197 	}
1198 	rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1199 	rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask;
1200 	rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1201 	rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1202 	rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1203 	rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1204 	return 0;
1205 }
1206 
1207 #ifndef CONFIG_TINY_RCU
show_rcu_tasks_trace_gp_kthread(void)1208 static void show_rcu_tasks_trace_gp_kthread(void)
1209 {
1210 	char buf[64];
1211 
1212 	sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end),
1213 		data_race(n_heavy_reader_ofl_updates),
1214 		data_race(n_heavy_reader_updates),
1215 		data_race(n_heavy_reader_attempts));
1216 	show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1217 }
1218 #endif /* #ifndef CONFIG_TINY_RCU */
1219 
1220 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
exit_tasks_rcu_finish_trace(struct task_struct * t)1221 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
show_rcu_tasks_trace_gp_kthread(void)1222 static inline void show_rcu_tasks_trace_gp_kthread(void) {}
1223 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1224 
1225 #ifndef CONFIG_TINY_RCU
show_rcu_tasks_gp_kthreads(void)1226 void show_rcu_tasks_gp_kthreads(void)
1227 {
1228 	show_rcu_tasks_classic_gp_kthread();
1229 	show_rcu_tasks_rude_gp_kthread();
1230 	show_rcu_tasks_trace_gp_kthread();
1231 }
1232 #endif /* #ifndef CONFIG_TINY_RCU */
1233 
rcu_init_tasks_generic(void)1234 void __init rcu_init_tasks_generic(void)
1235 {
1236 #ifdef CONFIG_TASKS_RCU
1237 	rcu_spawn_tasks_kthread();
1238 #endif
1239 
1240 #ifdef CONFIG_TASKS_RUDE_RCU
1241 	rcu_spawn_tasks_rude_kthread();
1242 #endif
1243 
1244 #ifdef CONFIG_TASKS_TRACE_RCU
1245 	rcu_spawn_tasks_trace_kthread();
1246 #endif
1247 }
1248 
1249 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
rcu_tasks_bootup_oddness(void)1250 static inline void rcu_tasks_bootup_oddness(void) {}
show_rcu_tasks_gp_kthreads(void)1251 void show_rcu_tasks_gp_kthreads(void) {}
1252 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */
1253