rcutorture: Make grace-period kthread report match RCU flavor being tested
[linux-2.6-block.git] / kernel / rcu / tasks.h
CommitLineData
eacd6f04
PM
1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * Task-based RCU implementations.
4 *
5 * Copyright (C) 2020 Paul E. McKenney
6 */
7
8fd8ca38 8#ifdef CONFIG_TASKS_RCU_GENERIC
5873b8a9
PM
9
10////////////////////////////////////////////////////////////////////////
11//
12// Generic data structures.
13
14struct rcu_tasks;
15typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp);
e4fe5dd6
PM
16typedef void (*pregp_func_t)(void);
17typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop);
9796e1ae 18typedef void (*postscan_func_t)(struct list_head *hop);
e4fe5dd6 19typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp);
af051ca4 20typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
eacd6f04 21
07e10515
PM
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.
5873b8a9 29 * @gp_func: This flavor's grace-period-wait function.
af051ca4 30 * @gp_state: Grace period's most recent state transition (debugging).
4fe192df 31 * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
2393a613 32 * @init_fract: Initial backoff sleep interval.
af051ca4
PM
33 * @gp_jiffies: Time of last @gp_state transition.
34 * @gp_start: Most recent grace-period start in jiffies.
238dbce3
PM
35 * @n_gps: Number of grace periods completed since boot.
36 * @n_ipis: Number of IPIs sent to encourage grace periods to end.
7e0669c3 37 * @n_ipis_fails: Number of IPI-send failures.
e4fe5dd6
PM
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).
5873b8a9 43 * @call_func: This flavor's call_rcu()-equivalent function.
c97d12a6
PM
44 * @name: This flavor's textual name.
45 * @kname: This flavor's kthread name.
07e10515
PM
46 */
47struct 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;
af051ca4 52 int gp_state;
4fe192df 53 int gp_sleep;
2393a613 54 int init_fract;
af051ca4 55 unsigned long gp_jiffies;
88092d0c 56 unsigned long gp_start;
238dbce3
PM
57 unsigned long n_gps;
58 unsigned long n_ipis;
7e0669c3 59 unsigned long n_ipis_fails;
07e10515 60 struct task_struct *kthread_ptr;
5873b8a9 61 rcu_tasks_gp_func_t gp_func;
e4fe5dd6
PM
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;
5873b8a9 67 call_rcu_func_t call_func;
c97d12a6
PM
68 char *name;
69 char *kname;
07e10515
PM
70};
71
c97d12a6
PM
72#define DEFINE_RCU_TASKS(rt_name, gp, call, n) \
73static struct rcu_tasks rt_name = \
07e10515 74{ \
c97d12a6
PM
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), \
5873b8a9
PM
78 .gp_func = gp, \
79 .call_func = call, \
c97d12a6
PM
80 .name = n, \
81 .kname = #rt_name, \
07e10515
PM
82}
83
eacd6f04
PM
84/* Track exiting tasks in order to allow them to be waited for. */
85DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
86
b0afa0f0 87/* Avoid IPIing CPUs early in the grace period. */
574de876 88#define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
b0afa0f0
PM
89static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
90module_param(rcu_task_ipi_delay, int, 0644);
91
eacd6f04
PM
92/* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
93#define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
94static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
95module_param(rcu_task_stall_timeout, int, 0644);
96
af051ca4
PM
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
8344496e 110#ifndef CONFIG_TINY_RCU
af051ca4
PM
111static 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};
8344496e 125#endif /* #ifndef CONFIG_TINY_RCU */
af051ca4 126
5873b8a9
PM
127////////////////////////////////////////////////////////////////////////
128//
129// Generic code.
130
af051ca4
PM
131/* Record grace-period phase and time. */
132static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
133{
134 rtp->gp_state = newstate;
135 rtp->gp_jiffies = jiffies;
136}
137
8344496e 138#ifndef CONFIG_TINY_RCU
af051ca4
PM
139/* Return state name. */
140static 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}
8344496e 149#endif /* #ifndef CONFIG_TINY_RCU */
af051ca4 150
5873b8a9
PM
151// Enqueue a callback for the specified flavor of Tasks RCU.
152static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
153 struct rcu_tasks *rtp)
eacd6f04
PM
154{
155 unsigned long flags;
156 bool needwake;
157
158 rhp->next = NULL;
159 rhp->func = func;
07e10515
PM
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);
eacd6f04 165 /* We can't create the thread unless interrupts are enabled. */
07e10515
PM
166 if (needwake && READ_ONCE(rtp->kthread_ptr))
167 wake_up(&rtp->cbs_wq);
eacd6f04 168}
eacd6f04 169
5873b8a9
PM
170// Wait for a grace period for the specified flavor of Tasks RCU.
171static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
eacd6f04
PM
172{
173 /* Complain if the scheduler has not started. */
174 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
175 "synchronize_rcu_tasks called too soon");
176
177 /* Wait for the grace period. */
5873b8a9 178 wait_rcu_gp(rtp->call_func);
eacd6f04
PM
179}
180
181/* RCU-tasks kthread that detects grace periods and invokes callbacks. */
182static int __noreturn rcu_tasks_kthread(void *arg)
183{
184 unsigned long flags;
eacd6f04
PM
185 struct rcu_head *list;
186 struct rcu_head *next;
07e10515 187 struct rcu_tasks *rtp = arg;
eacd6f04
PM
188
189 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
190 housekeeping_affine(current, HK_FLAG_RCU);
07e10515 191 WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
eacd6f04
PM
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
201 /* Pick up any new callbacks. */
07e10515 202 raw_spin_lock_irqsave(&rtp->cbs_lock, flags);
43766c3e 203 smp_mb__after_spinlock(); // Order updates vs. GP.
07e10515
PM
204 list = rtp->cbs_head;
205 rtp->cbs_head = NULL;
206 rtp->cbs_tail = &rtp->cbs_head;
207 raw_spin_unlock_irqrestore(&rtp->cbs_lock, flags);
eacd6f04
PM
208
209 /* If there were none, wait a bit and start over. */
210 if (!list) {
07e10515
PM
211 wait_event_interruptible(rtp->cbs_wq,
212 READ_ONCE(rtp->cbs_head));
213 if (!rtp->cbs_head) {
eacd6f04 214 WARN_ON(signal_pending(current));
af051ca4 215 set_tasks_gp_state(rtp, RTGS_WAIT_WAIT_CBS);
ea6eed9f 216 schedule_timeout_idle(HZ/10);
eacd6f04
PM
217 }
218 continue;
219 }
220
5873b8a9 221 // Wait for one grace period.
af051ca4 222 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
88092d0c 223 rtp->gp_start = jiffies;
5873b8a9 224 rtp->gp_func(rtp);
238dbce3 225 rtp->n_gps++;
eacd6f04
PM
226
227 /* Invoke the callbacks. */
af051ca4 228 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
eacd6f04
PM
229 while (list) {
230 next = list->next;
231 local_bh_disable();
232 list->func(list);
233 local_bh_enable();
234 list = next;
235 cond_resched();
236 }
237 /* Paranoid sleep to keep this from entering a tight loop */
4fe192df 238 schedule_timeout_idle(rtp->gp_sleep);
af051ca4
PM
239
240 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
eacd6f04
PM
241 }
242}
243
5873b8a9
PM
244/* Spawn RCU-tasks grace-period kthread, e.g., at core_initcall() time. */
245static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
eacd6f04
PM
246{
247 struct task_struct *t;
248
c97d12a6
PM
249 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
250 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
5873b8a9 251 return;
eacd6f04 252 smp_mb(); /* Ensure others see full kthread. */
eacd6f04 253}
eacd6f04 254
eacd6f04
PM
255#ifndef CONFIG_TINY_RCU
256
257/*
258 * Print any non-default Tasks RCU settings.
259 */
260static void __init rcu_tasks_bootup_oddness(void)
261{
d5f177d3 262#if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
eacd6f04
PM
263 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
264 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
d5f177d3
PM
265#endif /* #ifdef CONFIG_TASKS_RCU */
266#ifdef CONFIG_TASKS_RCU
267 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
eacd6f04 268#endif /* #ifdef CONFIG_TASKS_RCU */
c84aad76
PM
269#ifdef CONFIG_TASKS_RUDE_RCU
270 pr_info("\tRude variant of Tasks RCU enabled.\n");
271#endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
d5f177d3
PM
272#ifdef CONFIG_TASKS_TRACE_RCU
273 pr_info("\tTracing variant of Tasks RCU enabled.\n");
274#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
eacd6f04
PM
275}
276
277#endif /* #ifndef CONFIG_TINY_RCU */
5873b8a9 278
8344496e 279#ifndef CONFIG_TINY_RCU
e21408ce
PM
280/* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
281static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
282{
7e0669c3 283 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
e21408ce 284 rtp->kname,
7e0669c3 285 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
af051ca4 286 jiffies - data_race(rtp->gp_jiffies),
7e0669c3
PM
287 data_race(rtp->n_gps),
288 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
e21408ce
PM
289 ".k"[!!data_race(rtp->kthread_ptr)],
290 ".C"[!!data_race(rtp->cbs_head)],
291 s);
292}
27c0f144 293#endif // #ifndef CONFIG_TINY_RCU
e21408ce 294
25246fc8
PM
295static void exit_tasks_rcu_finish_trace(struct task_struct *t);
296
297#if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
5873b8a9 298
d01aa263
PM
299////////////////////////////////////////////////////////////////////////
300//
301// Shared code between task-list-scanning variants of Tasks RCU.
302
303/* Wait for one RCU-tasks grace period. */
304static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
305{
306 struct task_struct *g, *t;
307 unsigned long lastreport;
308 LIST_HEAD(holdouts);
309 int fract;
310
af051ca4 311 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
d01aa263
PM
312 rtp->pregp_func();
313
314 /*
315 * There were callbacks, so we need to wait for an RCU-tasks
316 * grace period. Start off by scanning the task list for tasks
317 * that are not already voluntarily blocked. Mark these tasks
318 * and make a list of them in holdouts.
319 */
af051ca4 320 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
d01aa263
PM
321 rcu_read_lock();
322 for_each_process_thread(g, t)
323 rtp->pertask_func(t, &holdouts);
324 rcu_read_unlock();
325
af051ca4 326 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
9796e1ae 327 rtp->postscan_func(&holdouts);
d01aa263
PM
328
329 /*
330 * Each pass through the following loop scans the list of holdout
331 * tasks, removing any that are no longer holdouts. When the list
332 * is empty, we are done.
333 */
334 lastreport = jiffies;
335
2393a613
PM
336 // Start off with initial wait and slowly back off to 1 HZ wait.
337 fract = rtp->init_fract;
338 if (fract > HZ)
339 fract = HZ;
d01aa263 340
77dc1741 341 while (!list_empty(&holdouts)) {
d01aa263
PM
342 bool firstreport;
343 bool needreport;
344 int rtst;
345
d01aa263 346 /* Slowly back off waiting for holdouts */
af051ca4 347 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
ea6eed9f 348 schedule_timeout_idle(HZ/fract);
d01aa263
PM
349
350 if (fract > 1)
351 fract--;
352
353 rtst = READ_ONCE(rcu_task_stall_timeout);
354 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
355 if (needreport)
356 lastreport = jiffies;
357 firstreport = true;
358 WARN_ON(signal_pending(current));
af051ca4 359 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
d01aa263
PM
360 rtp->holdouts_func(&holdouts, needreport, &firstreport);
361 }
362
af051ca4
PM
363 set_tasks_gp_state(rtp, RTGS_POST_GP);
364 rtp->postgp_func(rtp);
d01aa263
PM
365}
366
25246fc8
PM
367#endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
368
369#ifdef CONFIG_TASKS_RCU
370
5873b8a9
PM
371////////////////////////////////////////////////////////////////////////
372//
373// Simple variant of RCU whose quiescent states are voluntary context
374// switch, cond_resched_rcu_qs(), user-space execution, and idle.
375// As such, grace periods can take one good long time. There are no
376// read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
377// because this implementation is intended to get the system into a safe
378// state for some of the manipulations involved in tracing and the like.
379// Finally, this implementation does not support high call_rcu_tasks()
380// rates from multiple CPUs. If this is required, per-CPU callback lists
381// will be needed.
382
e4fe5dd6
PM
383/* Pre-grace-period preparation. */
384static void rcu_tasks_pregp_step(void)
385{
386 /*
387 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
388 * to complete. Invoking synchronize_rcu() suffices because all
389 * these transitions occur with interrupts disabled. Without this
390 * synchronize_rcu(), a read-side critical section that started
391 * before the grace period might be incorrectly seen as having
392 * started after the grace period.
393 *
394 * This synchronize_rcu() also dispenses with the need for a
395 * memory barrier on the first store to t->rcu_tasks_holdout,
396 * as it forces the store to happen after the beginning of the
397 * grace period.
398 */
399 synchronize_rcu();
400}
401
402/* Per-task initial processing. */
403static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
404{
405 if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
406 get_task_struct(t);
407 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
408 WRITE_ONCE(t->rcu_tasks_holdout, true);
409 list_add(&t->rcu_tasks_holdout_list, hop);
410 }
411}
412
413/* Processing between scanning taskslist and draining the holdout list. */
04a3c5aa 414static void rcu_tasks_postscan(struct list_head *hop)
e4fe5dd6
PM
415{
416 /*
417 * Wait for tasks that are in the process of exiting. This
418 * does only part of the job, ensuring that all tasks that were
419 * previously exiting reach the point where they have disabled
420 * preemption, allowing the later synchronize_rcu() to finish
421 * the job.
422 */
423 synchronize_srcu(&tasks_rcu_exit_srcu);
424}
425
5873b8a9
PM
426/* See if tasks are still holding out, complain if so. */
427static void check_holdout_task(struct task_struct *t,
428 bool needreport, bool *firstreport)
429{
430 int cpu;
431
432 if (!READ_ONCE(t->rcu_tasks_holdout) ||
433 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
434 !READ_ONCE(t->on_rq) ||
435 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
436 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
437 WRITE_ONCE(t->rcu_tasks_holdout, false);
438 list_del_init(&t->rcu_tasks_holdout_list);
439 put_task_struct(t);
440 return;
441 }
442 rcu_request_urgent_qs_task(t);
443 if (!needreport)
444 return;
445 if (*firstreport) {
446 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
447 *firstreport = false;
448 }
449 cpu = task_cpu(t);
450 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
451 t, ".I"[is_idle_task(t)],
452 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
453 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
454 t->rcu_tasks_idle_cpu, cpu);
455 sched_show_task(t);
456}
457
e4fe5dd6
PM
458/* Scan the holdout lists for tasks no longer holding out. */
459static void check_all_holdout_tasks(struct list_head *hop,
460 bool needreport, bool *firstreport)
461{
462 struct task_struct *t, *t1;
463
464 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
465 check_holdout_task(t, needreport, firstreport);
466 cond_resched();
467 }
468}
469
470/* Finish off the Tasks-RCU grace period. */
af051ca4 471static void rcu_tasks_postgp(struct rcu_tasks *rtp)
e4fe5dd6
PM
472{
473 /*
474 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
475 * memory barriers prior to them in the schedule() path, memory
476 * reordering on other CPUs could cause their RCU-tasks read-side
477 * critical sections to extend past the end of the grace period.
478 * However, because these ->nvcsw updates are carried out with
479 * interrupts disabled, we can use synchronize_rcu() to force the
480 * needed ordering on all such CPUs.
481 *
482 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
483 * accesses to be within the grace period, avoiding the need for
484 * memory barriers for ->rcu_tasks_holdout accesses.
485 *
486 * In addition, this synchronize_rcu() waits for exiting tasks
487 * to complete their final preempt_disable() region of execution,
488 * cleaning up after the synchronize_srcu() above.
489 */
490 synchronize_rcu();
491}
492
5873b8a9 493void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
c97d12a6 494DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
5873b8a9
PM
495
496/**
497 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
498 * @rhp: structure to be used for queueing the RCU updates.
499 * @func: actual callback function to be invoked after the grace period
500 *
501 * The callback function will be invoked some time after a full grace
502 * period elapses, in other words after all currently executing RCU
503 * read-side critical sections have completed. call_rcu_tasks() assumes
504 * that the read-side critical sections end at a voluntary context
505 * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle,
506 * or transition to usermode execution. As such, there are no read-side
507 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
508 * this primitive is intended to determine that all tasks have passed
509 * through a safe state, not so much for data-strcuture synchronization.
510 *
511 * See the description of call_rcu() for more detailed information on
512 * memory ordering guarantees.
513 */
514void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
515{
516 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
517}
518EXPORT_SYMBOL_GPL(call_rcu_tasks);
519
520/**
521 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
522 *
523 * Control will return to the caller some time after a full rcu-tasks
524 * grace period has elapsed, in other words after all currently
525 * executing rcu-tasks read-side critical sections have elapsed. These
526 * read-side critical sections are delimited by calls to schedule(),
527 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
528 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
529 *
530 * This is a very specialized primitive, intended only for a few uses in
531 * tracing and other situations requiring manipulation of function
532 * preambles and profiling hooks. The synchronize_rcu_tasks() function
533 * is not (yet) intended for heavy use from multiple CPUs.
534 *
535 * See the description of synchronize_rcu() for more detailed information
536 * on memory ordering guarantees.
537 */
538void synchronize_rcu_tasks(void)
539{
540 synchronize_rcu_tasks_generic(&rcu_tasks);
541}
542EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
543
544/**
545 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
546 *
547 * Although the current implementation is guaranteed to wait, it is not
548 * obligated to, for example, if there are no pending callbacks.
549 */
550void rcu_barrier_tasks(void)
551{
552 /* There is only one callback queue, so this is easy. ;-) */
553 synchronize_rcu_tasks();
554}
555EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
556
557static int __init rcu_spawn_tasks_kthread(void)
558{
4fe192df 559 rcu_tasks.gp_sleep = HZ / 10;
2393a613 560 rcu_tasks.init_fract = 10;
e4fe5dd6
PM
561 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
562 rcu_tasks.pertask_func = rcu_tasks_pertask;
563 rcu_tasks.postscan_func = rcu_tasks_postscan;
564 rcu_tasks.holdouts_func = check_all_holdout_tasks;
565 rcu_tasks.postgp_func = rcu_tasks_postgp;
5873b8a9
PM
566 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
567 return 0;
568}
569core_initcall(rcu_spawn_tasks_kthread);
570
27c0f144
PM
571#if !defined(CONFIG_TINY_RCU)
572void show_rcu_tasks_classic_gp_kthread(void)
e21408ce
PM
573{
574 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
575}
27c0f144
PM
576EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
577#endif // !defined(CONFIG_TINY_RCU)
e21408ce 578
25246fc8
PM
579/* Do the srcu_read_lock() for the above synchronize_srcu(). */
580void 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(). */
588void 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
e21408ce 598#else /* #ifdef CONFIG_TASKS_RCU */
25246fc8
PM
599void exit_tasks_rcu_start(void) { }
600void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
e21408ce 601#endif /* #else #ifdef CONFIG_TASKS_RCU */
c84aad76
PM
602
603#ifdef CONFIG_TASKS_RUDE_RCU
604
605////////////////////////////////////////////////////////////////////////
606//
607// "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
608// passing an empty function to schedule_on_each_cpu(). This approach
609// provides an asynchronous call_rcu_tasks_rude() API and batching
610// of concurrent calls to the synchronous synchronize_rcu_rude() API.
611// This sends IPIs far and wide and induces otherwise unnecessary context
612// switches on all online CPUs, whether idle or not.
613
614// Empty function to allow workqueues to force a context switch.
615static void rcu_tasks_be_rude(struct work_struct *work)
616{
617}
618
619// Wait for one rude RCU-tasks grace period.
620static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
621{
238dbce3 622 rtp->n_ipis += cpumask_weight(cpu_online_mask);
c84aad76
PM
623 schedule_on_each_cpu(rcu_tasks_be_rude);
624}
625
626void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
c97d12a6
PM
627DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
628 "RCU Tasks Rude");
c84aad76
PM
629
630/**
631 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
632 * @rhp: structure to be used for queueing the RCU updates.
633 * @func: actual callback function to be invoked after the grace period
634 *
635 * The callback function will be invoked some time after a full grace
636 * period elapses, in other words after all currently executing RCU
637 * read-side critical sections have completed. call_rcu_tasks_rude()
638 * assumes that the read-side critical sections end at context switch,
639 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
640 * there are no read-side primitives analogous to rcu_read_lock() and
641 * rcu_read_unlock() because this primitive is intended to determine
642 * that all tasks have passed through a safe state, not so much for
643 * data-strcuture synchronization.
644 *
645 * See the description of call_rcu() for more detailed information on
646 * memory ordering guarantees.
647 */
648void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
649{
650 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
651}
652EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
653
654/**
655 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
656 *
657 * Control will return to the caller some time after a rude rcu-tasks
658 * grace period has elapsed, in other words after all currently
659 * executing rcu-tasks read-side critical sections have elapsed. These
660 * read-side critical sections are delimited by calls to schedule(),
661 * cond_resched_tasks_rcu_qs(), userspace execution, and (in theory,
662 * anyway) cond_resched().
663 *
664 * This is a very specialized primitive, intended only for a few uses in
665 * tracing and other situations requiring manipulation of function preambles
666 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
667 * (yet) intended for heavy use from multiple CPUs.
668 *
669 * See the description of synchronize_rcu() for more detailed information
670 * on memory ordering guarantees.
671 */
672void synchronize_rcu_tasks_rude(void)
673{
674 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
675}
676EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
677
678/**
679 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
680 *
681 * Although the current implementation is guaranteed to wait, it is not
682 * obligated to, for example, if there are no pending callbacks.
683 */
684void rcu_barrier_tasks_rude(void)
685{
686 /* There is only one callback queue, so this is easy. ;-) */
687 synchronize_rcu_tasks_rude();
688}
689EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
690
691static int __init rcu_spawn_tasks_rude_kthread(void)
692{
4fe192df 693 rcu_tasks_rude.gp_sleep = HZ / 10;
c84aad76
PM
694 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
695 return 0;
696}
697core_initcall(rcu_spawn_tasks_rude_kthread);
698
27c0f144
PM
699#if !defined(CONFIG_TINY_RCU)
700void show_rcu_tasks_rude_gp_kthread(void)
e21408ce
PM
701{
702 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
703}
27c0f144
PM
704EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
705#endif // !defined(CONFIG_TINY_RCU)
706#endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
d5f177d3
PM
707
708////////////////////////////////////////////////////////////////////////
709//
710// Tracing variant of Tasks RCU. This variant is designed to be used
711// to protect tracing hooks, including those of BPF. This variant
712// therefore:
713//
714// 1. Has explicit read-side markers to allow finite grace periods
715// in the face of in-kernel loops for PREEMPT=n builds.
716//
717// 2. Protects code in the idle loop, exception entry/exit, and
718// CPU-hotplug code paths, similar to the capabilities of SRCU.
719//
720// 3. Avoids expensive read-side instruction, having overhead similar
721// to that of Preemptible RCU.
722//
723// There are of course downsides. The grace-period code can send IPIs to
724// CPUs, even when those CPUs are in the idle loop or in nohz_full userspace.
725// It is necessary to scan the full tasklist, much as for Tasks RCU. There
726// is a single callback queue guarded by a single lock, again, much as for
727// Tasks RCU. If needed, these downsides can be at least partially remedied.
728//
729// Perhaps most important, this variant of RCU does not affect the vanilla
730// flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
731// readers can operate from idle, offline, and exception entry/exit in no
732// way allows rcu_preempt and rcu_sched readers to also do so.
733
734// The lockdep state must be outside of #ifdef to be useful.
735#ifdef CONFIG_DEBUG_LOCK_ALLOC
736static struct lock_class_key rcu_lock_trace_key;
737struct lockdep_map rcu_trace_lock_map =
738 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
739EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
740#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
741
742#ifdef CONFIG_TASKS_TRACE_RCU
743
30d8aa51
PM
744static atomic_t trc_n_readers_need_end; // Number of waited-for readers.
745static DECLARE_WAIT_QUEUE_HEAD(trc_wait); // List of holdout tasks.
d5f177d3
PM
746
747// Record outstanding IPIs to each CPU. No point in sending two...
748static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
749
40471509
PM
750// The number of detections of task quiescent state relying on
751// heavyweight readers executing explicit memory barriers.
6731da9e
PM
752static unsigned long n_heavy_reader_attempts;
753static unsigned long n_heavy_reader_updates;
754static unsigned long n_heavy_reader_ofl_updates;
40471509 755
b0afa0f0
PM
756void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
757DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
758 "RCU Tasks Trace");
759
b38f57c1
PM
760/*
761 * This irq_work handler allows rcu_read_unlock_trace() to be invoked
762 * while the scheduler locks are held.
763 */
764static void rcu_read_unlock_iw(struct irq_work *iwp)
765{
766 wake_up(&trc_wait);
767}
768static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw);
769
d5f177d3 770/* If we are the last reader, wake up the grace-period kthread. */
276c4104 771void rcu_read_unlock_trace_special(struct task_struct *t, int nesting)
d5f177d3 772{
276c4104
PM
773 int nq = t->trc_reader_special.b.need_qs;
774
9ae58d7b
PM
775 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) &&
776 t->trc_reader_special.b.need_mb)
276c4104
PM
777 smp_mb(); // Pairs with update-side barriers.
778 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
779 if (nq)
780 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
781 WRITE_ONCE(t->trc_reader_nesting, nesting);
782 if (nq && atomic_dec_and_test(&trc_n_readers_need_end))
b38f57c1 783 irq_work_queue(&rcu_tasks_trace_iw);
d5f177d3
PM
784}
785EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
786
787/* Add a task to the holdout list, if it is not already on the list. */
788static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
789{
790 if (list_empty(&t->trc_holdout_list)) {
791 get_task_struct(t);
792 list_add(&t->trc_holdout_list, bhp);
793 }
794}
795
796/* Remove a task from the holdout list, if it is in fact present. */
797static void trc_del_holdout(struct task_struct *t)
798{
799 if (!list_empty(&t->trc_holdout_list)) {
800 list_del_init(&t->trc_holdout_list);
801 put_task_struct(t);
802 }
803}
804
805/* IPI handler to check task state. */
806static void trc_read_check_handler(void *t_in)
807{
808 struct task_struct *t = current;
809 struct task_struct *texp = t_in;
810
811 // If the task is no longer running on this CPU, leave.
812 if (unlikely(texp != t)) {
813 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
814 wake_up(&trc_wait);
815 goto reset_ipi; // Already on holdout list, so will check later.
816 }
817
818 // If the task is not in a read-side critical section, and
819 // if this is the last reader, awaken the grace-period kthread.
820 if (likely(!t->trc_reader_nesting)) {
821 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
822 wake_up(&trc_wait);
823 // Mark as checked after decrement to avoid false
824 // positives on the above WARN_ON_ONCE().
825 WRITE_ONCE(t->trc_reader_checked, true);
826 goto reset_ipi;
827 }
ba3a86e4
PM
828 // If we are racing with an rcu_read_unlock_trace(), try again later.
829 if (unlikely(t->trc_reader_nesting < 0)) {
830 if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end)))
831 wake_up(&trc_wait);
832 goto reset_ipi;
833 }
d5f177d3
PM
834 WRITE_ONCE(t->trc_reader_checked, true);
835
836 // Get here if the task is in a read-side critical section. Set
837 // its state so that it will awaken the grace-period kthread upon
838 // exit from that critical section.
276c4104
PM
839 WARN_ON_ONCE(t->trc_reader_special.b.need_qs);
840 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
d5f177d3
PM
841
842reset_ipi:
843 // Allow future IPIs to be sent on CPU and for task.
844 // Also order this IPI handler against any later manipulations of
845 // the intended task.
846 smp_store_release(&per_cpu(trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
847 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
848}
849
850/* Callback function for scheduler to check locked-down task. */
851static bool trc_inspect_reader(struct task_struct *t, void *arg)
852{
7d0c9c50
PM
853 int cpu = task_cpu(t);
854 bool in_qs = false;
7e3b70e0 855 bool ofl = cpu_is_offline(cpu);
7d0c9c50
PM
856
857 if (task_curr(t)) {
30d8aa51 858 WARN_ON_ONCE(ofl && !is_idle_task(t));
7e3b70e0 859
7d0c9c50 860 // If no chance of heavyweight readers, do it the hard way.
7e3b70e0 861 if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
7d0c9c50
PM
862 return false;
863
864 // If heavyweight readers are enabled on the remote task,
865 // we can inspect its state despite its currently running.
866 // However, we cannot safely change its state.
40471509 867 n_heavy_reader_attempts++;
7e3b70e0
PM
868 if (!ofl && // Check for "running" idle tasks on offline CPUs.
869 !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
7d0c9c50 870 return false; // No quiescent state, do it the hard way.
40471509 871 n_heavy_reader_updates++;
edf3775f
PM
872 if (ofl)
873 n_heavy_reader_ofl_updates++;
7d0c9c50
PM
874 in_qs = true;
875 } else {
876 in_qs = likely(!t->trc_reader_nesting);
877 }
d5f177d3
PM
878
879 // Mark as checked. Because this is called from the grace-period
880 // kthread, also remove the task from the holdout list.
881 t->trc_reader_checked = true;
882 trc_del_holdout(t);
883
7d0c9c50
PM
884 if (in_qs)
885 return true; // Already in quiescent state, done!!!
886
887 // The task is in a read-side critical section, so set up its
888 // state so that it will awaken the grace-period kthread upon exit
889 // from that critical section.
890 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
891 WARN_ON_ONCE(t->trc_reader_special.b.need_qs);
892 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
d5f177d3
PM
893 return true;
894}
895
896/* Attempt to extract the state for the specified task. */
897static void trc_wait_for_one_reader(struct task_struct *t,
898 struct list_head *bhp)
899{
900 int cpu;
901
902 // If a previous IPI is still in flight, let it complete.
903 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
904 return;
905
906 // The current task had better be in a quiescent state.
907 if (t == current) {
908 t->trc_reader_checked = true;
909 trc_del_holdout(t);
910 WARN_ON_ONCE(t->trc_reader_nesting);
911 return;
912 }
913
914 // Attempt to nail down the task for inspection.
915 get_task_struct(t);
916 if (try_invoke_on_locked_down_task(t, trc_inspect_reader, NULL)) {
917 put_task_struct(t);
918 return;
919 }
920 put_task_struct(t);
921
922 // If currently running, send an IPI, either way, add to list.
923 trc_add_holdout(t, bhp);
574de876
PM
924 if (task_curr(t) &&
925 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
d5f177d3
PM
926 // The task is currently running, so try IPIing it.
927 cpu = task_cpu(t);
928
929 // If there is already an IPI outstanding, let it happen.
930 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
931 return;
932
933 atomic_inc(&trc_n_readers_need_end);
934 per_cpu(trc_ipi_to_cpu, cpu) = true;
935 t->trc_ipi_to_cpu = cpu;
238dbce3 936 rcu_tasks_trace.n_ipis++;
d5f177d3
PM
937 if (smp_call_function_single(cpu,
938 trc_read_check_handler, t, 0)) {
939 // Just in case there is some other reason for
940 // failure than the target CPU being offline.
7e0669c3 941 rcu_tasks_trace.n_ipis_fails++;
d5f177d3
PM
942 per_cpu(trc_ipi_to_cpu, cpu) = false;
943 t->trc_ipi_to_cpu = cpu;
944 if (atomic_dec_and_test(&trc_n_readers_need_end)) {
945 WARN_ON_ONCE(1);
946 wake_up(&trc_wait);
947 }
948 }
949 }
950}
951
952/* Initialize for a new RCU-tasks-trace grace period. */
953static void rcu_tasks_trace_pregp_step(void)
954{
955 int cpu;
956
d5f177d3
PM
957 // Allow for fast-acting IPIs.
958 atomic_set(&trc_n_readers_need_end, 1);
959
960 // There shouldn't be any old IPIs, but...
961 for_each_possible_cpu(cpu)
962 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
81b4a7bc
PM
963
964 // Disable CPU hotplug across the tasklist scan.
965 // This also waits for all readers in CPU-hotplug code paths.
966 cpus_read_lock();
d5f177d3
PM
967}
968
969/* Do first-round processing for the specified task. */
970static void rcu_tasks_trace_pertask(struct task_struct *t,
971 struct list_head *hop)
972{
276c4104 973 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
43766c3e 974 WRITE_ONCE(t->trc_reader_checked, false);
d5f177d3
PM
975 t->trc_ipi_to_cpu = -1;
976 trc_wait_for_one_reader(t, hop);
977}
978
9796e1ae
PM
979/*
980 * Do intermediate processing between task and holdout scans and
981 * pick up the idle tasks.
982 */
983static void rcu_tasks_trace_postscan(struct list_head *hop)
d5f177d3 984{
9796e1ae
PM
985 int cpu;
986
987 for_each_possible_cpu(cpu)
988 rcu_tasks_trace_pertask(idle_task(cpu), hop);
989
81b4a7bc
PM
990 // Re-enable CPU hotplug now that the tasklist scan has completed.
991 cpus_read_unlock();
992
d5f177d3
PM
993 // Wait for late-stage exiting tasks to finish exiting.
994 // These might have passed the call to exit_tasks_rcu_finish().
995 synchronize_rcu();
996 // Any tasks that exit after this point will set ->trc_reader_checked.
997}
998
4593e772
PM
999/* Show the state of a task stalling the current RCU tasks trace GP. */
1000static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1001{
1002 int cpu;
1003
1004 if (*firstreport) {
1005 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1006 *firstreport = false;
1007 }
1008 // FIXME: This should attempt to use try_invoke_on_nonrunning_task().
1009 cpu = task_cpu(t);
1010 pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n",
1011 t->pid,
1012 ".I"[READ_ONCE(t->trc_ipi_to_cpu) > 0],
1013 ".i"[is_idle_task(t)],
1014 ".N"[cpu > 0 && tick_nohz_full_cpu(cpu)],
1015 t->trc_reader_nesting,
276c4104 1016 " N"[!!t->trc_reader_special.b.need_qs],
4593e772
PM
1017 cpu);
1018 sched_show_task(t);
1019}
1020
1021/* List stalled IPIs for RCU tasks trace. */
1022static void show_stalled_ipi_trace(void)
1023{
1024 int cpu;
1025
1026 for_each_possible_cpu(cpu)
1027 if (per_cpu(trc_ipi_to_cpu, cpu))
1028 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1029}
1030
d5f177d3
PM
1031/* Do one scan of the holdout list. */
1032static void check_all_holdout_tasks_trace(struct list_head *hop,
4593e772 1033 bool needreport, bool *firstreport)
d5f177d3
PM
1034{
1035 struct task_struct *g, *t;
1036
81b4a7bc
PM
1037 // Disable CPU hotplug across the holdout list scan.
1038 cpus_read_lock();
1039
d5f177d3
PM
1040 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1041 // If safe and needed, try to check the current task.
1042 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1043 !READ_ONCE(t->trc_reader_checked))
1044 trc_wait_for_one_reader(t, hop);
1045
1046 // If check succeeded, remove this task from the list.
1047 if (READ_ONCE(t->trc_reader_checked))
1048 trc_del_holdout(t);
4593e772
PM
1049 else if (needreport)
1050 show_stalled_task_trace(t, firstreport);
1051 }
81b4a7bc
PM
1052
1053 // Re-enable CPU hotplug now that the holdout list scan has completed.
1054 cpus_read_unlock();
1055
4593e772
PM
1056 if (needreport) {
1057 if (firstreport)
1058 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1059 show_stalled_ipi_trace();
d5f177d3
PM
1060 }
1061}
1062
1063/* Wait for grace period to complete and provide ordering. */
af051ca4 1064static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
d5f177d3 1065{
4593e772
PM
1066 bool firstreport;
1067 struct task_struct *g, *t;
1068 LIST_HEAD(holdouts);
1069 long ret;
1070
d5f177d3
PM
1071 // Remove the safety count.
1072 smp_mb__before_atomic(); // Order vs. earlier atomics
1073 atomic_dec(&trc_n_readers_need_end);
1074 smp_mb__after_atomic(); // Order vs. later atomics
1075
1076 // Wait for readers.
af051ca4 1077 set_tasks_gp_state(rtp, RTGS_WAIT_READERS);
4593e772
PM
1078 for (;;) {
1079 ret = wait_event_idle_exclusive_timeout(
1080 trc_wait,
1081 atomic_read(&trc_n_readers_need_end) == 0,
1082 READ_ONCE(rcu_task_stall_timeout));
1083 if (ret)
1084 break; // Count reached zero.
af051ca4 1085 // Stall warning time, so make a list of the offenders.
f747c7e1 1086 rcu_read_lock();
4593e772 1087 for_each_process_thread(g, t)
276c4104 1088 if (READ_ONCE(t->trc_reader_special.b.need_qs))
4593e772 1089 trc_add_holdout(t, &holdouts);
f747c7e1 1090 rcu_read_unlock();
4593e772 1091 firstreport = true;
592031cc
PM
1092 list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) {
1093 if (READ_ONCE(t->trc_reader_special.b.need_qs))
4593e772 1094 show_stalled_task_trace(t, &firstreport);
592031cc
PM
1095 trc_del_holdout(t); // Release task_struct reference.
1096 }
4593e772
PM
1097 if (firstreport)
1098 pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n");
1099 show_stalled_ipi_trace();
1100 pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end));
1101 }
d5f177d3 1102 smp_mb(); // Caller's code must be ordered after wakeup.
43766c3e 1103 // Pairs with pretty much every ordering primitive.
d5f177d3
PM
1104}
1105
1106/* Report any needed quiescent state for this exiting task. */
25246fc8 1107static void exit_tasks_rcu_finish_trace(struct task_struct *t)
d5f177d3
PM
1108{
1109 WRITE_ONCE(t->trc_reader_checked, true);
1110 WARN_ON_ONCE(t->trc_reader_nesting);
1111 WRITE_ONCE(t->trc_reader_nesting, 0);
276c4104
PM
1112 if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)))
1113 rcu_read_unlock_trace_special(t, 0);
d5f177d3
PM
1114}
1115
d5f177d3
PM
1116/**
1117 * call_rcu_tasks_trace() - Queue a callback trace 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_trace()
1124 * assumes that the read-side critical sections end at context switch,
1125 * cond_resched_rcu_qs(), or transition to usermode execution. As such,
1126 * there are no read-side primitives analogous to rcu_read_lock() and
1127 * rcu_read_unlock() because this primitive is intended to determine
1128 * that all tasks have passed through a safe state, not so much for
1129 * data-strcuture synchronization.
1130 *
1131 * See the description of call_rcu() for more detailed information on
1132 * memory ordering guarantees.
1133 */
1134void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1135{
1136 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1137}
1138EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1139
1140/**
1141 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1142 *
1143 * Control will return to the caller some time after a trace rcu-tasks
c7dcf810
PM
1144 * grace period has elapsed, in other words after all currently executing
1145 * rcu-tasks read-side critical sections have elapsed. These read-side
1146 * critical sections are delimited by calls to rcu_read_lock_trace()
1147 * and rcu_read_unlock_trace().
d5f177d3
PM
1148 *
1149 * This is a very specialized primitive, intended only for a few uses in
1150 * tracing and other situations requiring manipulation of function preambles
1151 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1152 * (yet) intended for heavy use from multiple CPUs.
1153 *
1154 * See the description of synchronize_rcu() for more detailed information
1155 * on memory ordering guarantees.
1156 */
1157void synchronize_rcu_tasks_trace(void)
1158{
1159 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1160 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1161}
1162EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1163
1164/**
1165 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() 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 */
1170void rcu_barrier_tasks_trace(void)
1171{
1172 /* There is only one callback queue, so this is easy. ;-) */
1173 synchronize_rcu_tasks_trace();
1174}
1175EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1176
1177static int __init rcu_spawn_tasks_trace_kthread(void)
1178{
2393a613 1179 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
4fe192df 1180 rcu_tasks_trace.gp_sleep = HZ / 10;
2393a613
PM
1181 rcu_tasks_trace.init_fract = 10;
1182 } else {
4fe192df
PM
1183 rcu_tasks_trace.gp_sleep = HZ / 200;
1184 if (rcu_tasks_trace.gp_sleep <= 0)
1185 rcu_tasks_trace.gp_sleep = 1;
2393a613
PM
1186 rcu_tasks_trace.init_fract = HZ / 5;
1187 if (rcu_tasks_trace.init_fract <= 0)
1188 rcu_tasks_trace.init_fract = 1;
1189 }
d5f177d3
PM
1190 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1191 rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask;
1192 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1193 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1194 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1195 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1196 return 0;
1197}
1198core_initcall(rcu_spawn_tasks_trace_kthread);
1199
27c0f144
PM
1200#if !defined(CONFIG_TINY_RCU)
1201void show_rcu_tasks_trace_gp_kthread(void)
e21408ce 1202{
40471509 1203 char buf[64];
e21408ce 1204
edf3775f
PM
1205 sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end),
1206 data_race(n_heavy_reader_ofl_updates),
40471509
PM
1207 data_race(n_heavy_reader_updates),
1208 data_race(n_heavy_reader_attempts));
e21408ce
PM
1209 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1210}
27c0f144
PM
1211EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1212#endif // !defined(CONFIG_TINY_RCU)
e21408ce 1213
d5f177d3 1214#else /* #ifdef CONFIG_TASKS_TRACE_RCU */
25246fc8 1215static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
d5f177d3 1216#endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
8fd8ca38 1217
8344496e 1218#ifndef CONFIG_TINY_RCU
e21408ce
PM
1219void show_rcu_tasks_gp_kthreads(void)
1220{
1221 show_rcu_tasks_classic_gp_kthread();
1222 show_rcu_tasks_rude_gp_kthread();
1223 show_rcu_tasks_trace_gp_kthread();
1224}
8344496e 1225#endif /* #ifndef CONFIG_TINY_RCU */
e21408ce 1226
8fd8ca38
PM
1227#else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1228static inline void rcu_tasks_bootup_oddness(void) {}
e21408ce 1229void show_rcu_tasks_gp_kthreads(void) {}
8fd8ca38 1230#endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */