Commit | Line | Data |
---|---|---|
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 | ||
14 | struct rcu_tasks; | |
15 | typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp); | |
e4fe5dd6 PM |
16 | typedef void (*pregp_func_t)(void); |
17 | typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop); | |
9796e1ae | 18 | typedef void (*postscan_func_t)(struct list_head *hop); |
e4fe5dd6 | 19 | typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp); |
af051ca4 | 20 | typedef 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). |
2393a613 | 31 | * @init_fract: Initial backoff sleep interval. |
af051ca4 PM |
32 | * @gp_jiffies: Time of last @gp_state transition. |
33 | * @gp_start: Most recent grace-period start in jiffies. | |
238dbce3 PM |
34 | * @n_gps: Number of grace periods completed since boot. |
35 | * @n_ipis: Number of IPIs sent to encourage grace periods to end. | |
7e0669c3 | 36 | * @n_ipis_fails: Number of IPI-send failures. |
e4fe5dd6 PM |
37 | * @pregp_func: This flavor's pre-grace-period function (optional). |
38 | * @pertask_func: This flavor's per-task scan function (optional). | |
39 | * @postscan_func: This flavor's post-task scan function (optional). | |
40 | * @holdout_func: This flavor's holdout-list scan function (optional). | |
41 | * @postgp_func: This flavor's post-grace-period function (optional). | |
5873b8a9 | 42 | * @call_func: This flavor's call_rcu()-equivalent function. |
c97d12a6 PM |
43 | * @name: This flavor's textual name. |
44 | * @kname: This flavor's kthread name. | |
07e10515 PM |
45 | */ |
46 | struct rcu_tasks { | |
47 | struct rcu_head *cbs_head; | |
48 | struct rcu_head **cbs_tail; | |
49 | struct wait_queue_head cbs_wq; | |
50 | raw_spinlock_t cbs_lock; | |
af051ca4 | 51 | int gp_state; |
2393a613 | 52 | int init_fract; |
af051ca4 | 53 | unsigned long gp_jiffies; |
88092d0c | 54 | unsigned long gp_start; |
238dbce3 PM |
55 | unsigned long n_gps; |
56 | unsigned long n_ipis; | |
7e0669c3 | 57 | unsigned long n_ipis_fails; |
07e10515 | 58 | struct task_struct *kthread_ptr; |
5873b8a9 | 59 | rcu_tasks_gp_func_t gp_func; |
e4fe5dd6 PM |
60 | pregp_func_t pregp_func; |
61 | pertask_func_t pertask_func; | |
62 | postscan_func_t postscan_func; | |
63 | holdouts_func_t holdouts_func; | |
64 | postgp_func_t postgp_func; | |
5873b8a9 | 65 | call_rcu_func_t call_func; |
c97d12a6 PM |
66 | char *name; |
67 | char *kname; | |
07e10515 PM |
68 | }; |
69 | ||
c97d12a6 PM |
70 | #define DEFINE_RCU_TASKS(rt_name, gp, call, n) \ |
71 | static struct rcu_tasks rt_name = \ | |
07e10515 | 72 | { \ |
c97d12a6 PM |
73 | .cbs_tail = &rt_name.cbs_head, \ |
74 | .cbs_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rt_name.cbs_wq), \ | |
75 | .cbs_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_lock), \ | |
5873b8a9 PM |
76 | .gp_func = gp, \ |
77 | .call_func = call, \ | |
c97d12a6 PM |
78 | .name = n, \ |
79 | .kname = #rt_name, \ | |
07e10515 PM |
80 | } |
81 | ||
eacd6f04 PM |
82 | /* Track exiting tasks in order to allow them to be waited for. */ |
83 | DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu); | |
84 | ||
b0afa0f0 | 85 | /* Avoid IPIing CPUs early in the grace period. */ |
574de876 | 86 | #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0) |
b0afa0f0 PM |
87 | static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY; |
88 | module_param(rcu_task_ipi_delay, int, 0644); | |
89 | ||
eacd6f04 PM |
90 | /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */ |
91 | #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10) | |
92 | static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT; | |
93 | module_param(rcu_task_stall_timeout, int, 0644); | |
94 | ||
af051ca4 PM |
95 | /* RCU tasks grace-period state for debugging. */ |
96 | #define RTGS_INIT 0 | |
97 | #define RTGS_WAIT_WAIT_CBS 1 | |
98 | #define RTGS_WAIT_GP 2 | |
99 | #define RTGS_PRE_WAIT_GP 3 | |
100 | #define RTGS_SCAN_TASKLIST 4 | |
101 | #define RTGS_POST_SCAN_TASKLIST 5 | |
102 | #define RTGS_WAIT_SCAN_HOLDOUTS 6 | |
103 | #define RTGS_SCAN_HOLDOUTS 7 | |
104 | #define RTGS_POST_GP 8 | |
105 | #define RTGS_WAIT_READERS 9 | |
106 | #define RTGS_INVOKE_CBS 10 | |
107 | #define RTGS_WAIT_CBS 11 | |
8344496e | 108 | #ifndef CONFIG_TINY_RCU |
af051ca4 PM |
109 | static const char * const rcu_tasks_gp_state_names[] = { |
110 | "RTGS_INIT", | |
111 | "RTGS_WAIT_WAIT_CBS", | |
112 | "RTGS_WAIT_GP", | |
113 | "RTGS_PRE_WAIT_GP", | |
114 | "RTGS_SCAN_TASKLIST", | |
115 | "RTGS_POST_SCAN_TASKLIST", | |
116 | "RTGS_WAIT_SCAN_HOLDOUTS", | |
117 | "RTGS_SCAN_HOLDOUTS", | |
118 | "RTGS_POST_GP", | |
119 | "RTGS_WAIT_READERS", | |
120 | "RTGS_INVOKE_CBS", | |
121 | "RTGS_WAIT_CBS", | |
122 | }; | |
8344496e | 123 | #endif /* #ifndef CONFIG_TINY_RCU */ |
af051ca4 | 124 | |
5873b8a9 PM |
125 | //////////////////////////////////////////////////////////////////////// |
126 | // | |
127 | // Generic code. | |
128 | ||
af051ca4 PM |
129 | /* Record grace-period phase and time. */ |
130 | static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate) | |
131 | { | |
132 | rtp->gp_state = newstate; | |
133 | rtp->gp_jiffies = jiffies; | |
134 | } | |
135 | ||
8344496e | 136 | #ifndef CONFIG_TINY_RCU |
af051ca4 PM |
137 | /* Return state name. */ |
138 | static const char *tasks_gp_state_getname(struct rcu_tasks *rtp) | |
139 | { | |
140 | int i = data_race(rtp->gp_state); // Let KCSAN detect update races | |
141 | int j = READ_ONCE(i); // Prevent the compiler from reading twice | |
142 | ||
143 | if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names)) | |
144 | return "???"; | |
145 | return rcu_tasks_gp_state_names[j]; | |
146 | } | |
8344496e | 147 | #endif /* #ifndef CONFIG_TINY_RCU */ |
af051ca4 | 148 | |
5873b8a9 PM |
149 | // Enqueue a callback for the specified flavor of Tasks RCU. |
150 | static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func, | |
151 | struct rcu_tasks *rtp) | |
eacd6f04 PM |
152 | { |
153 | unsigned long flags; | |
154 | bool needwake; | |
155 | ||
156 | rhp->next = NULL; | |
157 | rhp->func = func; | |
07e10515 PM |
158 | raw_spin_lock_irqsave(&rtp->cbs_lock, flags); |
159 | needwake = !rtp->cbs_head; | |
160 | WRITE_ONCE(*rtp->cbs_tail, rhp); | |
161 | rtp->cbs_tail = &rhp->next; | |
162 | raw_spin_unlock_irqrestore(&rtp->cbs_lock, flags); | |
eacd6f04 | 163 | /* We can't create the thread unless interrupts are enabled. */ |
07e10515 PM |
164 | if (needwake && READ_ONCE(rtp->kthread_ptr)) |
165 | wake_up(&rtp->cbs_wq); | |
eacd6f04 | 166 | } |
eacd6f04 | 167 | |
5873b8a9 PM |
168 | // Wait for a grace period for the specified flavor of Tasks RCU. |
169 | static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp) | |
eacd6f04 PM |
170 | { |
171 | /* Complain if the scheduler has not started. */ | |
172 | RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE, | |
173 | "synchronize_rcu_tasks called too soon"); | |
174 | ||
175 | /* Wait for the grace period. */ | |
5873b8a9 | 176 | wait_rcu_gp(rtp->call_func); |
eacd6f04 PM |
177 | } |
178 | ||
179 | /* RCU-tasks kthread that detects grace periods and invokes callbacks. */ | |
180 | static int __noreturn rcu_tasks_kthread(void *arg) | |
181 | { | |
182 | unsigned long flags; | |
eacd6f04 PM |
183 | struct rcu_head *list; |
184 | struct rcu_head *next; | |
07e10515 | 185 | struct rcu_tasks *rtp = arg; |
eacd6f04 PM |
186 | |
187 | /* Run on housekeeping CPUs by default. Sysadm can move if desired. */ | |
188 | housekeeping_affine(current, HK_FLAG_RCU); | |
07e10515 | 189 | WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start! |
eacd6f04 PM |
190 | |
191 | /* | |
192 | * Each pass through the following loop makes one check for | |
193 | * newly arrived callbacks, and, if there are some, waits for | |
194 | * one RCU-tasks grace period and then invokes the callbacks. | |
195 | * This loop is terminated by the system going down. ;-) | |
196 | */ | |
197 | for (;;) { | |
198 | ||
199 | /* Pick up any new callbacks. */ | |
07e10515 | 200 | raw_spin_lock_irqsave(&rtp->cbs_lock, flags); |
43766c3e | 201 | smp_mb__after_spinlock(); // Order updates vs. GP. |
07e10515 PM |
202 | list = rtp->cbs_head; |
203 | rtp->cbs_head = NULL; | |
204 | rtp->cbs_tail = &rtp->cbs_head; | |
205 | raw_spin_unlock_irqrestore(&rtp->cbs_lock, flags); | |
eacd6f04 PM |
206 | |
207 | /* If there were none, wait a bit and start over. */ | |
208 | if (!list) { | |
07e10515 PM |
209 | wait_event_interruptible(rtp->cbs_wq, |
210 | READ_ONCE(rtp->cbs_head)); | |
211 | if (!rtp->cbs_head) { | |
eacd6f04 | 212 | WARN_ON(signal_pending(current)); |
af051ca4 | 213 | set_tasks_gp_state(rtp, RTGS_WAIT_WAIT_CBS); |
ea6eed9f | 214 | schedule_timeout_idle(HZ/10); |
eacd6f04 PM |
215 | } |
216 | continue; | |
217 | } | |
218 | ||
5873b8a9 | 219 | // Wait for one grace period. |
af051ca4 | 220 | set_tasks_gp_state(rtp, RTGS_WAIT_GP); |
88092d0c | 221 | rtp->gp_start = jiffies; |
5873b8a9 | 222 | rtp->gp_func(rtp); |
238dbce3 | 223 | rtp->n_gps++; |
eacd6f04 PM |
224 | |
225 | /* Invoke the callbacks. */ | |
af051ca4 | 226 | set_tasks_gp_state(rtp, RTGS_INVOKE_CBS); |
eacd6f04 PM |
227 | while (list) { |
228 | next = list->next; | |
229 | local_bh_disable(); | |
230 | list->func(list); | |
231 | local_bh_enable(); | |
232 | list = next; | |
233 | cond_resched(); | |
234 | } | |
235 | /* Paranoid sleep to keep this from entering a tight loop */ | |
ea6eed9f | 236 | schedule_timeout_idle(HZ/10); |
af051ca4 PM |
237 | |
238 | set_tasks_gp_state(rtp, RTGS_WAIT_CBS); | |
eacd6f04 PM |
239 | } |
240 | } | |
241 | ||
5873b8a9 PM |
242 | /* Spawn RCU-tasks grace-period kthread, e.g., at core_initcall() time. */ |
243 | static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp) | |
eacd6f04 PM |
244 | { |
245 | struct task_struct *t; | |
246 | ||
c97d12a6 PM |
247 | t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname); |
248 | if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name)) | |
5873b8a9 | 249 | return; |
eacd6f04 | 250 | smp_mb(); /* Ensure others see full kthread. */ |
eacd6f04 | 251 | } |
eacd6f04 | 252 | |
eacd6f04 PM |
253 | #ifndef CONFIG_TINY_RCU |
254 | ||
255 | /* | |
256 | * Print any non-default Tasks RCU settings. | |
257 | */ | |
258 | static void __init rcu_tasks_bootup_oddness(void) | |
259 | { | |
d5f177d3 | 260 | #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) |
eacd6f04 PM |
261 | if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT) |
262 | pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout); | |
d5f177d3 PM |
263 | #endif /* #ifdef CONFIG_TASKS_RCU */ |
264 | #ifdef CONFIG_TASKS_RCU | |
265 | pr_info("\tTrampoline variant of Tasks RCU enabled.\n"); | |
eacd6f04 | 266 | #endif /* #ifdef CONFIG_TASKS_RCU */ |
c84aad76 PM |
267 | #ifdef CONFIG_TASKS_RUDE_RCU |
268 | pr_info("\tRude variant of Tasks RCU enabled.\n"); | |
269 | #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */ | |
d5f177d3 PM |
270 | #ifdef CONFIG_TASKS_TRACE_RCU |
271 | pr_info("\tTracing variant of Tasks RCU enabled.\n"); | |
272 | #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */ | |
eacd6f04 PM |
273 | } |
274 | ||
275 | #endif /* #ifndef CONFIG_TINY_RCU */ | |
5873b8a9 | 276 | |
8344496e | 277 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
278 | /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */ |
279 | static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s) | |
280 | { | |
7e0669c3 | 281 | pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n", |
e21408ce | 282 | rtp->kname, |
7e0669c3 | 283 | tasks_gp_state_getname(rtp), data_race(rtp->gp_state), |
af051ca4 | 284 | jiffies - data_race(rtp->gp_jiffies), |
7e0669c3 PM |
285 | data_race(rtp->n_gps), |
286 | data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis), | |
e21408ce PM |
287 | ".k"[!!data_race(rtp->kthread_ptr)], |
288 | ".C"[!!data_race(rtp->cbs_head)], | |
289 | s); | |
290 | } | |
8344496e | 291 | #endif /* #ifndef CONFIG_TINY_RCU */ |
e21408ce | 292 | |
25246fc8 PM |
293 | static void exit_tasks_rcu_finish_trace(struct task_struct *t); |
294 | ||
295 | #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) | |
5873b8a9 | 296 | |
d01aa263 PM |
297 | //////////////////////////////////////////////////////////////////////// |
298 | // | |
299 | // Shared code between task-list-scanning variants of Tasks RCU. | |
300 | ||
301 | /* Wait for one RCU-tasks grace period. */ | |
302 | static void rcu_tasks_wait_gp(struct rcu_tasks *rtp) | |
303 | { | |
304 | struct task_struct *g, *t; | |
305 | unsigned long lastreport; | |
306 | LIST_HEAD(holdouts); | |
307 | int fract; | |
308 | ||
af051ca4 | 309 | set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP); |
d01aa263 PM |
310 | rtp->pregp_func(); |
311 | ||
312 | /* | |
313 | * There were callbacks, so we need to wait for an RCU-tasks | |
314 | * grace period. Start off by scanning the task list for tasks | |
315 | * that are not already voluntarily blocked. Mark these tasks | |
316 | * and make a list of them in holdouts. | |
317 | */ | |
af051ca4 | 318 | set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST); |
d01aa263 PM |
319 | rcu_read_lock(); |
320 | for_each_process_thread(g, t) | |
321 | rtp->pertask_func(t, &holdouts); | |
322 | rcu_read_unlock(); | |
323 | ||
af051ca4 | 324 | set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST); |
9796e1ae | 325 | rtp->postscan_func(&holdouts); |
d01aa263 PM |
326 | |
327 | /* | |
328 | * Each pass through the following loop scans the list of holdout | |
329 | * tasks, removing any that are no longer holdouts. When the list | |
330 | * is empty, we are done. | |
331 | */ | |
332 | lastreport = jiffies; | |
333 | ||
2393a613 PM |
334 | // Start off with initial wait and slowly back off to 1 HZ wait. |
335 | fract = rtp->init_fract; | |
336 | if (fract > HZ) | |
337 | fract = HZ; | |
d01aa263 PM |
338 | |
339 | for (;;) { | |
340 | bool firstreport; | |
341 | bool needreport; | |
342 | int rtst; | |
343 | ||
344 | if (list_empty(&holdouts)) | |
345 | break; | |
346 | ||
347 | /* Slowly back off waiting for holdouts */ | |
af051ca4 | 348 | set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS); |
ea6eed9f | 349 | schedule_timeout_idle(HZ/fract); |
d01aa263 PM |
350 | |
351 | if (fract > 1) | |
352 | fract--; | |
353 | ||
354 | rtst = READ_ONCE(rcu_task_stall_timeout); | |
355 | needreport = rtst > 0 && time_after(jiffies, lastreport + rtst); | |
356 | if (needreport) | |
357 | lastreport = jiffies; | |
358 | firstreport = true; | |
359 | WARN_ON(signal_pending(current)); | |
af051ca4 | 360 | set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS); |
d01aa263 PM |
361 | rtp->holdouts_func(&holdouts, needreport, &firstreport); |
362 | } | |
363 | ||
af051ca4 PM |
364 | set_tasks_gp_state(rtp, RTGS_POST_GP); |
365 | rtp->postgp_func(rtp); | |
d01aa263 PM |
366 | } |
367 | ||
25246fc8 PM |
368 | #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */ |
369 | ||
370 | #ifdef CONFIG_TASKS_RCU | |
371 | ||
5873b8a9 PM |
372 | //////////////////////////////////////////////////////////////////////// |
373 | // | |
374 | // Simple variant of RCU whose quiescent states are voluntary context | |
375 | // switch, cond_resched_rcu_qs(), user-space execution, and idle. | |
376 | // As such, grace periods can take one good long time. There are no | |
377 | // read-side primitives similar to rcu_read_lock() and rcu_read_unlock() | |
378 | // because this implementation is intended to get the system into a safe | |
379 | // state for some of the manipulations involved in tracing and the like. | |
380 | // Finally, this implementation does not support high call_rcu_tasks() | |
381 | // rates from multiple CPUs. If this is required, per-CPU callback lists | |
382 | // will be needed. | |
383 | ||
e4fe5dd6 PM |
384 | /* Pre-grace-period preparation. */ |
385 | static void rcu_tasks_pregp_step(void) | |
386 | { | |
387 | /* | |
388 | * Wait for all pre-existing t->on_rq and t->nvcsw transitions | |
389 | * to complete. Invoking synchronize_rcu() suffices because all | |
390 | * these transitions occur with interrupts disabled. Without this | |
391 | * synchronize_rcu(), a read-side critical section that started | |
392 | * before the grace period might be incorrectly seen as having | |
393 | * started after the grace period. | |
394 | * | |
395 | * This synchronize_rcu() also dispenses with the need for a | |
396 | * memory barrier on the first store to t->rcu_tasks_holdout, | |
397 | * as it forces the store to happen after the beginning of the | |
398 | * grace period. | |
399 | */ | |
400 | synchronize_rcu(); | |
401 | } | |
402 | ||
403 | /* Per-task initial processing. */ | |
404 | static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop) | |
405 | { | |
406 | if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) { | |
407 | get_task_struct(t); | |
408 | t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw); | |
409 | WRITE_ONCE(t->rcu_tasks_holdout, true); | |
410 | list_add(&t->rcu_tasks_holdout_list, hop); | |
411 | } | |
412 | } | |
413 | ||
414 | /* Processing between scanning taskslist and draining the holdout list. */ | |
04a3c5aa | 415 | static void rcu_tasks_postscan(struct list_head *hop) |
e4fe5dd6 PM |
416 | { |
417 | /* | |
418 | * Wait for tasks that are in the process of exiting. This | |
419 | * does only part of the job, ensuring that all tasks that were | |
420 | * previously exiting reach the point where they have disabled | |
421 | * preemption, allowing the later synchronize_rcu() to finish | |
422 | * the job. | |
423 | */ | |
424 | synchronize_srcu(&tasks_rcu_exit_srcu); | |
425 | } | |
426 | ||
5873b8a9 PM |
427 | /* See if tasks are still holding out, complain if so. */ |
428 | static void check_holdout_task(struct task_struct *t, | |
429 | bool needreport, bool *firstreport) | |
430 | { | |
431 | int cpu; | |
432 | ||
433 | if (!READ_ONCE(t->rcu_tasks_holdout) || | |
434 | t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) || | |
435 | !READ_ONCE(t->on_rq) || | |
436 | (IS_ENABLED(CONFIG_NO_HZ_FULL) && | |
437 | !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) { | |
438 | WRITE_ONCE(t->rcu_tasks_holdout, false); | |
439 | list_del_init(&t->rcu_tasks_holdout_list); | |
440 | put_task_struct(t); | |
441 | return; | |
442 | } | |
443 | rcu_request_urgent_qs_task(t); | |
444 | if (!needreport) | |
445 | return; | |
446 | if (*firstreport) { | |
447 | pr_err("INFO: rcu_tasks detected stalls on tasks:\n"); | |
448 | *firstreport = false; | |
449 | } | |
450 | cpu = task_cpu(t); | |
451 | pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n", | |
452 | t, ".I"[is_idle_task(t)], | |
453 | "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)], | |
454 | t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout, | |
455 | t->rcu_tasks_idle_cpu, cpu); | |
456 | sched_show_task(t); | |
457 | } | |
458 | ||
e4fe5dd6 PM |
459 | /* Scan the holdout lists for tasks no longer holding out. */ |
460 | static void check_all_holdout_tasks(struct list_head *hop, | |
461 | bool needreport, bool *firstreport) | |
462 | { | |
463 | struct task_struct *t, *t1; | |
464 | ||
465 | list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) { | |
466 | check_holdout_task(t, needreport, firstreport); | |
467 | cond_resched(); | |
468 | } | |
469 | } | |
470 | ||
471 | /* Finish off the Tasks-RCU grace period. */ | |
af051ca4 | 472 | static void rcu_tasks_postgp(struct rcu_tasks *rtp) |
e4fe5dd6 PM |
473 | { |
474 | /* | |
475 | * Because ->on_rq and ->nvcsw are not guaranteed to have a full | |
476 | * memory barriers prior to them in the schedule() path, memory | |
477 | * reordering on other CPUs could cause their RCU-tasks read-side | |
478 | * critical sections to extend past the end of the grace period. | |
479 | * However, because these ->nvcsw updates are carried out with | |
480 | * interrupts disabled, we can use synchronize_rcu() to force the | |
481 | * needed ordering on all such CPUs. | |
482 | * | |
483 | * This synchronize_rcu() also confines all ->rcu_tasks_holdout | |
484 | * accesses to be within the grace period, avoiding the need for | |
485 | * memory barriers for ->rcu_tasks_holdout accesses. | |
486 | * | |
487 | * In addition, this synchronize_rcu() waits for exiting tasks | |
488 | * to complete their final preempt_disable() region of execution, | |
489 | * cleaning up after the synchronize_srcu() above. | |
490 | */ | |
491 | synchronize_rcu(); | |
492 | } | |
493 | ||
5873b8a9 | 494 | void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func); |
c97d12a6 | 495 | DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks"); |
5873b8a9 PM |
496 | |
497 | /** | |
498 | * call_rcu_tasks() - Queue an RCU for invocation task-based grace period | |
499 | * @rhp: structure to be used for queueing the RCU updates. | |
500 | * @func: actual callback function to be invoked after the grace period | |
501 | * | |
502 | * The callback function will be invoked some time after a full grace | |
503 | * period elapses, in other words after all currently executing RCU | |
504 | * read-side critical sections have completed. call_rcu_tasks() assumes | |
505 | * that the read-side critical sections end at a voluntary context | |
506 | * switch (not a preemption!), cond_resched_rcu_qs(), entry into idle, | |
507 | * or transition to usermode execution. As such, there are no read-side | |
508 | * primitives analogous to rcu_read_lock() and rcu_read_unlock() because | |
509 | * this primitive is intended to determine that all tasks have passed | |
510 | * through a safe state, not so much for data-strcuture synchronization. | |
511 | * | |
512 | * See the description of call_rcu() for more detailed information on | |
513 | * memory ordering guarantees. | |
514 | */ | |
515 | void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func) | |
516 | { | |
517 | call_rcu_tasks_generic(rhp, func, &rcu_tasks); | |
518 | } | |
519 | EXPORT_SYMBOL_GPL(call_rcu_tasks); | |
520 | ||
521 | /** | |
522 | * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed. | |
523 | * | |
524 | * Control will return to the caller some time after a full rcu-tasks | |
525 | * grace period has elapsed, in other words after all currently | |
526 | * executing rcu-tasks read-side critical sections have elapsed. These | |
527 | * read-side critical sections are delimited by calls to schedule(), | |
528 | * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls | |
529 | * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched(). | |
530 | * | |
531 | * This is a very specialized primitive, intended only for a few uses in | |
532 | * tracing and other situations requiring manipulation of function | |
533 | * preambles and profiling hooks. The synchronize_rcu_tasks() function | |
534 | * is not (yet) intended for heavy use from multiple CPUs. | |
535 | * | |
536 | * See the description of synchronize_rcu() for more detailed information | |
537 | * on memory ordering guarantees. | |
538 | */ | |
539 | void synchronize_rcu_tasks(void) | |
540 | { | |
541 | synchronize_rcu_tasks_generic(&rcu_tasks); | |
542 | } | |
543 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks); | |
544 | ||
545 | /** | |
546 | * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks. | |
547 | * | |
548 | * Although the current implementation is guaranteed to wait, it is not | |
549 | * obligated to, for example, if there are no pending callbacks. | |
550 | */ | |
551 | void rcu_barrier_tasks(void) | |
552 | { | |
553 | /* There is only one callback queue, so this is easy. ;-) */ | |
554 | synchronize_rcu_tasks(); | |
555 | } | |
556 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks); | |
557 | ||
558 | static int __init rcu_spawn_tasks_kthread(void) | |
559 | { | |
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 | } | |
569 | core_initcall(rcu_spawn_tasks_kthread); | |
570 | ||
8344496e | 571 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
572 | static void show_rcu_tasks_classic_gp_kthread(void) |
573 | { | |
574 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks, ""); | |
575 | } | |
8344496e | 576 | #endif /* #ifndef CONFIG_TINY_RCU */ |
e21408ce | 577 | |
25246fc8 PM |
578 | /* Do the srcu_read_lock() for the above synchronize_srcu(). */ |
579 | void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu) | |
580 | { | |
581 | preempt_disable(); | |
582 | current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu); | |
583 | preempt_enable(); | |
584 | } | |
585 | ||
586 | /* Do the srcu_read_unlock() for the above synchronize_srcu(). */ | |
587 | void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu) | |
588 | { | |
589 | struct task_struct *t = current; | |
590 | ||
591 | preempt_disable(); | |
592 | __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx); | |
593 | preempt_enable(); | |
594 | exit_tasks_rcu_finish_trace(t); | |
595 | } | |
596 | ||
e21408ce | 597 | #else /* #ifdef CONFIG_TASKS_RCU */ |
78edc005 | 598 | static inline void show_rcu_tasks_classic_gp_kthread(void) { } |
25246fc8 PM |
599 | void exit_tasks_rcu_start(void) { } |
600 | void 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. | |
615 | static void rcu_tasks_be_rude(struct work_struct *work) | |
616 | { | |
617 | } | |
618 | ||
619 | // Wait for one rude RCU-tasks grace period. | |
620 | static 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 | ||
626 | void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func); | |
c97d12a6 PM |
627 | DEFINE_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 | */ | |
648 | void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func) | |
649 | { | |
650 | call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude); | |
651 | } | |
652 | EXPORT_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 | */ | |
672 | void synchronize_rcu_tasks_rude(void) | |
673 | { | |
674 | synchronize_rcu_tasks_generic(&rcu_tasks_rude); | |
675 | } | |
676 | EXPORT_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 | */ | |
684 | void rcu_barrier_tasks_rude(void) | |
685 | { | |
686 | /* There is only one callback queue, so this is easy. ;-) */ | |
687 | synchronize_rcu_tasks_rude(); | |
688 | } | |
689 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude); | |
690 | ||
691 | static int __init rcu_spawn_tasks_rude_kthread(void) | |
692 | { | |
693 | rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude); | |
694 | return 0; | |
695 | } | |
696 | core_initcall(rcu_spawn_tasks_rude_kthread); | |
697 | ||
8344496e | 698 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
699 | static void show_rcu_tasks_rude_gp_kthread(void) |
700 | { | |
701 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, ""); | |
702 | } | |
8344496e | 703 | #endif /* #ifndef CONFIG_TINY_RCU */ |
e21408ce PM |
704 | |
705 | #else /* #ifdef CONFIG_TASKS_RUDE_RCU */ | |
706 | static void show_rcu_tasks_rude_gp_kthread(void) {} | |
707 | #endif /* #else #ifdef CONFIG_TASKS_RUDE_RCU */ | |
d5f177d3 PM |
708 | |
709 | //////////////////////////////////////////////////////////////////////// | |
710 | // | |
711 | // Tracing variant of Tasks RCU. This variant is designed to be used | |
712 | // to protect tracing hooks, including those of BPF. This variant | |
713 | // therefore: | |
714 | // | |
715 | // 1. Has explicit read-side markers to allow finite grace periods | |
716 | // in the face of in-kernel loops for PREEMPT=n builds. | |
717 | // | |
718 | // 2. Protects code in the idle loop, exception entry/exit, and | |
719 | // CPU-hotplug code paths, similar to the capabilities of SRCU. | |
720 | // | |
721 | // 3. Avoids expensive read-side instruction, having overhead similar | |
722 | // to that of Preemptible RCU. | |
723 | // | |
724 | // There are of course downsides. The grace-period code can send IPIs to | |
725 | // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace. | |
726 | // It is necessary to scan the full tasklist, much as for Tasks RCU. There | |
727 | // is a single callback queue guarded by a single lock, again, much as for | |
728 | // Tasks RCU. If needed, these downsides can be at least partially remedied. | |
729 | // | |
730 | // Perhaps most important, this variant of RCU does not affect the vanilla | |
731 | // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace | |
732 | // readers can operate from idle, offline, and exception entry/exit in no | |
733 | // way allows rcu_preempt and rcu_sched readers to also do so. | |
734 | ||
735 | // The lockdep state must be outside of #ifdef to be useful. | |
736 | #ifdef CONFIG_DEBUG_LOCK_ALLOC | |
737 | static struct lock_class_key rcu_lock_trace_key; | |
738 | struct lockdep_map rcu_trace_lock_map = | |
739 | STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key); | |
740 | EXPORT_SYMBOL_GPL(rcu_trace_lock_map); | |
741 | #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */ | |
742 | ||
743 | #ifdef CONFIG_TASKS_TRACE_RCU | |
744 | ||
30d8aa51 PM |
745 | static atomic_t trc_n_readers_need_end; // Number of waited-for readers. |
746 | static DECLARE_WAIT_QUEUE_HEAD(trc_wait); // List of holdout tasks. | |
d5f177d3 PM |
747 | |
748 | // Record outstanding IPIs to each CPU. No point in sending two... | |
749 | static DEFINE_PER_CPU(bool, trc_ipi_to_cpu); | |
750 | ||
40471509 PM |
751 | // The number of detections of task quiescent state relying on |
752 | // heavyweight readers executing explicit memory barriers. | |
6731da9e PM |
753 | static unsigned long n_heavy_reader_attempts; |
754 | static unsigned long n_heavy_reader_updates; | |
755 | static unsigned long n_heavy_reader_ofl_updates; | |
40471509 | 756 | |
b0afa0f0 PM |
757 | void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func); |
758 | DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace, | |
759 | "RCU Tasks Trace"); | |
760 | ||
b38f57c1 PM |
761 | /* |
762 | * This irq_work handler allows rcu_read_unlock_trace() to be invoked | |
763 | * while the scheduler locks are held. | |
764 | */ | |
765 | static void rcu_read_unlock_iw(struct irq_work *iwp) | |
766 | { | |
767 | wake_up(&trc_wait); | |
768 | } | |
769 | static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw); | |
770 | ||
d5f177d3 | 771 | /* If we are the last reader, wake up the grace-period kthread. */ |
276c4104 | 772 | void rcu_read_unlock_trace_special(struct task_struct *t, int nesting) |
d5f177d3 | 773 | { |
276c4104 PM |
774 | int nq = t->trc_reader_special.b.need_qs; |
775 | ||
9ae58d7b PM |
776 | if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && |
777 | t->trc_reader_special.b.need_mb) | |
276c4104 PM |
778 | smp_mb(); // Pairs with update-side barriers. |
779 | // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers. | |
780 | if (nq) | |
781 | WRITE_ONCE(t->trc_reader_special.b.need_qs, false); | |
782 | WRITE_ONCE(t->trc_reader_nesting, nesting); | |
783 | if (nq && atomic_dec_and_test(&trc_n_readers_need_end)) | |
b38f57c1 | 784 | irq_work_queue(&rcu_tasks_trace_iw); |
d5f177d3 PM |
785 | } |
786 | EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special); | |
787 | ||
788 | /* Add a task to the holdout list, if it is not already on the list. */ | |
789 | static void trc_add_holdout(struct task_struct *t, struct list_head *bhp) | |
790 | { | |
791 | if (list_empty(&t->trc_holdout_list)) { | |
792 | get_task_struct(t); | |
793 | list_add(&t->trc_holdout_list, bhp); | |
794 | } | |
795 | } | |
796 | ||
797 | /* Remove a task from the holdout list, if it is in fact present. */ | |
798 | static void trc_del_holdout(struct task_struct *t) | |
799 | { | |
800 | if (!list_empty(&t->trc_holdout_list)) { | |
801 | list_del_init(&t->trc_holdout_list); | |
802 | put_task_struct(t); | |
803 | } | |
804 | } | |
805 | ||
806 | /* IPI handler to check task state. */ | |
807 | static void trc_read_check_handler(void *t_in) | |
808 | { | |
809 | struct task_struct *t = current; | |
810 | struct task_struct *texp = t_in; | |
811 | ||
812 | // If the task is no longer running on this CPU, leave. | |
813 | if (unlikely(texp != t)) { | |
814 | if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end))) | |
815 | wake_up(&trc_wait); | |
816 | goto reset_ipi; // Already on holdout list, so will check later. | |
817 | } | |
818 | ||
819 | // If the task is not in a read-side critical section, and | |
820 | // if this is the last reader, awaken the grace-period kthread. | |
821 | if (likely(!t->trc_reader_nesting)) { | |
822 | if (WARN_ON_ONCE(atomic_dec_and_test(&trc_n_readers_need_end))) | |
823 | wake_up(&trc_wait); | |
824 | // Mark as checked after decrement to avoid false | |
825 | // positives on the above WARN_ON_ONCE(). | |
826 | WRITE_ONCE(t->trc_reader_checked, true); | |
827 | goto reset_ipi; | |
828 | } | |
829 | WRITE_ONCE(t->trc_reader_checked, true); | |
830 | ||
831 | // Get here if the task is in a read-side critical section. Set | |
832 | // its state so that it will awaken the grace-period kthread upon | |
833 | // exit from that critical section. | |
276c4104 PM |
834 | WARN_ON_ONCE(t->trc_reader_special.b.need_qs); |
835 | WRITE_ONCE(t->trc_reader_special.b.need_qs, true); | |
d5f177d3 PM |
836 | |
837 | reset_ipi: | |
838 | // Allow future IPIs to be sent on CPU and for task. | |
839 | // Also order this IPI handler against any later manipulations of | |
840 | // the intended task. | |
841 | smp_store_release(&per_cpu(trc_ipi_to_cpu, smp_processor_id()), false); // ^^^ | |
842 | smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^ | |
843 | } | |
844 | ||
845 | /* Callback function for scheduler to check locked-down task. */ | |
846 | static bool trc_inspect_reader(struct task_struct *t, void *arg) | |
847 | { | |
7d0c9c50 PM |
848 | int cpu = task_cpu(t); |
849 | bool in_qs = false; | |
7e3b70e0 | 850 | bool ofl = cpu_is_offline(cpu); |
7d0c9c50 PM |
851 | |
852 | if (task_curr(t)) { | |
30d8aa51 | 853 | WARN_ON_ONCE(ofl && !is_idle_task(t)); |
7e3b70e0 | 854 | |
7d0c9c50 | 855 | // If no chance of heavyweight readers, do it the hard way. |
7e3b70e0 | 856 | if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) |
7d0c9c50 PM |
857 | return false; |
858 | ||
859 | // If heavyweight readers are enabled on the remote task, | |
860 | // we can inspect its state despite its currently running. | |
861 | // However, we cannot safely change its state. | |
40471509 | 862 | n_heavy_reader_attempts++; |
7e3b70e0 PM |
863 | if (!ofl && // Check for "running" idle tasks on offline CPUs. |
864 | !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting)) | |
7d0c9c50 | 865 | return false; // No quiescent state, do it the hard way. |
40471509 | 866 | n_heavy_reader_updates++; |
edf3775f PM |
867 | if (ofl) |
868 | n_heavy_reader_ofl_updates++; | |
7d0c9c50 PM |
869 | in_qs = true; |
870 | } else { | |
871 | in_qs = likely(!t->trc_reader_nesting); | |
872 | } | |
d5f177d3 PM |
873 | |
874 | // Mark as checked. Because this is called from the grace-period | |
875 | // kthread, also remove the task from the holdout list. | |
876 | t->trc_reader_checked = true; | |
877 | trc_del_holdout(t); | |
878 | ||
7d0c9c50 PM |
879 | if (in_qs) |
880 | return true; // Already in quiescent state, done!!! | |
881 | ||
882 | // The task is in a read-side critical section, so set up its | |
883 | // state so that it will awaken the grace-period kthread upon exit | |
884 | // from that critical section. | |
885 | atomic_inc(&trc_n_readers_need_end); // One more to wait on. | |
886 | WARN_ON_ONCE(t->trc_reader_special.b.need_qs); | |
887 | WRITE_ONCE(t->trc_reader_special.b.need_qs, true); | |
d5f177d3 PM |
888 | return true; |
889 | } | |
890 | ||
891 | /* Attempt to extract the state for the specified task. */ | |
892 | static void trc_wait_for_one_reader(struct task_struct *t, | |
893 | struct list_head *bhp) | |
894 | { | |
895 | int cpu; | |
896 | ||
897 | // If a previous IPI is still in flight, let it complete. | |
898 | if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI | |
899 | return; | |
900 | ||
901 | // The current task had better be in a quiescent state. | |
902 | if (t == current) { | |
903 | t->trc_reader_checked = true; | |
904 | trc_del_holdout(t); | |
905 | WARN_ON_ONCE(t->trc_reader_nesting); | |
906 | return; | |
907 | } | |
908 | ||
909 | // Attempt to nail down the task for inspection. | |
910 | get_task_struct(t); | |
911 | if (try_invoke_on_locked_down_task(t, trc_inspect_reader, NULL)) { | |
912 | put_task_struct(t); | |
913 | return; | |
914 | } | |
915 | put_task_struct(t); | |
916 | ||
917 | // If currently running, send an IPI, either way, add to list. | |
918 | trc_add_holdout(t, bhp); | |
574de876 PM |
919 | if (task_curr(t) && |
920 | time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) { | |
d5f177d3 PM |
921 | // The task is currently running, so try IPIing it. |
922 | cpu = task_cpu(t); | |
923 | ||
924 | // If there is already an IPI outstanding, let it happen. | |
925 | if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0) | |
926 | return; | |
927 | ||
928 | atomic_inc(&trc_n_readers_need_end); | |
929 | per_cpu(trc_ipi_to_cpu, cpu) = true; | |
930 | t->trc_ipi_to_cpu = cpu; | |
238dbce3 | 931 | rcu_tasks_trace.n_ipis++; |
d5f177d3 PM |
932 | if (smp_call_function_single(cpu, |
933 | trc_read_check_handler, t, 0)) { | |
934 | // Just in case there is some other reason for | |
935 | // failure than the target CPU being offline. | |
7e0669c3 | 936 | rcu_tasks_trace.n_ipis_fails++; |
d5f177d3 PM |
937 | per_cpu(trc_ipi_to_cpu, cpu) = false; |
938 | t->trc_ipi_to_cpu = cpu; | |
939 | if (atomic_dec_and_test(&trc_n_readers_need_end)) { | |
940 | WARN_ON_ONCE(1); | |
941 | wake_up(&trc_wait); | |
942 | } | |
943 | } | |
944 | } | |
945 | } | |
946 | ||
947 | /* Initialize for a new RCU-tasks-trace grace period. */ | |
948 | static void rcu_tasks_trace_pregp_step(void) | |
949 | { | |
950 | int cpu; | |
951 | ||
d5f177d3 PM |
952 | // Allow for fast-acting IPIs. |
953 | atomic_set(&trc_n_readers_need_end, 1); | |
954 | ||
955 | // There shouldn't be any old IPIs, but... | |
956 | for_each_possible_cpu(cpu) | |
957 | WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu)); | |
81b4a7bc PM |
958 | |
959 | // Disable CPU hotplug across the tasklist scan. | |
960 | // This also waits for all readers in CPU-hotplug code paths. | |
961 | cpus_read_lock(); | |
d5f177d3 PM |
962 | } |
963 | ||
964 | /* Do first-round processing for the specified task. */ | |
965 | static void rcu_tasks_trace_pertask(struct task_struct *t, | |
966 | struct list_head *hop) | |
967 | { | |
276c4104 | 968 | WRITE_ONCE(t->trc_reader_special.b.need_qs, false); |
43766c3e | 969 | WRITE_ONCE(t->trc_reader_checked, false); |
d5f177d3 PM |
970 | t->trc_ipi_to_cpu = -1; |
971 | trc_wait_for_one_reader(t, hop); | |
972 | } | |
973 | ||
9796e1ae PM |
974 | /* |
975 | * Do intermediate processing between task and holdout scans and | |
976 | * pick up the idle tasks. | |
977 | */ | |
978 | static void rcu_tasks_trace_postscan(struct list_head *hop) | |
d5f177d3 | 979 | { |
9796e1ae PM |
980 | int cpu; |
981 | ||
982 | for_each_possible_cpu(cpu) | |
983 | rcu_tasks_trace_pertask(idle_task(cpu), hop); | |
984 | ||
81b4a7bc PM |
985 | // Re-enable CPU hotplug now that the tasklist scan has completed. |
986 | cpus_read_unlock(); | |
987 | ||
d5f177d3 PM |
988 | // Wait for late-stage exiting tasks to finish exiting. |
989 | // These might have passed the call to exit_tasks_rcu_finish(). | |
990 | synchronize_rcu(); | |
991 | // Any tasks that exit after this point will set ->trc_reader_checked. | |
992 | } | |
993 | ||
4593e772 PM |
994 | /* Show the state of a task stalling the current RCU tasks trace GP. */ |
995 | static void show_stalled_task_trace(struct task_struct *t, bool *firstreport) | |
996 | { | |
997 | int cpu; | |
998 | ||
999 | if (*firstreport) { | |
1000 | pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n"); | |
1001 | *firstreport = false; | |
1002 | } | |
1003 | // FIXME: This should attempt to use try_invoke_on_nonrunning_task(). | |
1004 | cpu = task_cpu(t); | |
1005 | pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n", | |
1006 | t->pid, | |
1007 | ".I"[READ_ONCE(t->trc_ipi_to_cpu) > 0], | |
1008 | ".i"[is_idle_task(t)], | |
1009 | ".N"[cpu > 0 && tick_nohz_full_cpu(cpu)], | |
1010 | t->trc_reader_nesting, | |
276c4104 | 1011 | " N"[!!t->trc_reader_special.b.need_qs], |
4593e772 PM |
1012 | cpu); |
1013 | sched_show_task(t); | |
1014 | } | |
1015 | ||
1016 | /* List stalled IPIs for RCU tasks trace. */ | |
1017 | static void show_stalled_ipi_trace(void) | |
1018 | { | |
1019 | int cpu; | |
1020 | ||
1021 | for_each_possible_cpu(cpu) | |
1022 | if (per_cpu(trc_ipi_to_cpu, cpu)) | |
1023 | pr_alert("\tIPI outstanding to CPU %d\n", cpu); | |
1024 | } | |
1025 | ||
d5f177d3 PM |
1026 | /* Do one scan of the holdout list. */ |
1027 | static void check_all_holdout_tasks_trace(struct list_head *hop, | |
4593e772 | 1028 | bool needreport, bool *firstreport) |
d5f177d3 PM |
1029 | { |
1030 | struct task_struct *g, *t; | |
1031 | ||
81b4a7bc PM |
1032 | // Disable CPU hotplug across the holdout list scan. |
1033 | cpus_read_lock(); | |
1034 | ||
d5f177d3 PM |
1035 | list_for_each_entry_safe(t, g, hop, trc_holdout_list) { |
1036 | // If safe and needed, try to check the current task. | |
1037 | if (READ_ONCE(t->trc_ipi_to_cpu) == -1 && | |
1038 | !READ_ONCE(t->trc_reader_checked)) | |
1039 | trc_wait_for_one_reader(t, hop); | |
1040 | ||
1041 | // If check succeeded, remove this task from the list. | |
1042 | if (READ_ONCE(t->trc_reader_checked)) | |
1043 | trc_del_holdout(t); | |
4593e772 PM |
1044 | else if (needreport) |
1045 | show_stalled_task_trace(t, firstreport); | |
1046 | } | |
81b4a7bc PM |
1047 | |
1048 | // Re-enable CPU hotplug now that the holdout list scan has completed. | |
1049 | cpus_read_unlock(); | |
1050 | ||
4593e772 PM |
1051 | if (needreport) { |
1052 | if (firstreport) | |
1053 | pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n"); | |
1054 | show_stalled_ipi_trace(); | |
d5f177d3 PM |
1055 | } |
1056 | } | |
1057 | ||
1058 | /* Wait for grace period to complete and provide ordering. */ | |
af051ca4 | 1059 | static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp) |
d5f177d3 | 1060 | { |
4593e772 PM |
1061 | bool firstreport; |
1062 | struct task_struct *g, *t; | |
1063 | LIST_HEAD(holdouts); | |
1064 | long ret; | |
1065 | ||
d5f177d3 PM |
1066 | // Remove the safety count. |
1067 | smp_mb__before_atomic(); // Order vs. earlier atomics | |
1068 | atomic_dec(&trc_n_readers_need_end); | |
1069 | smp_mb__after_atomic(); // Order vs. later atomics | |
1070 | ||
1071 | // Wait for readers. | |
af051ca4 | 1072 | set_tasks_gp_state(rtp, RTGS_WAIT_READERS); |
4593e772 PM |
1073 | for (;;) { |
1074 | ret = wait_event_idle_exclusive_timeout( | |
1075 | trc_wait, | |
1076 | atomic_read(&trc_n_readers_need_end) == 0, | |
1077 | READ_ONCE(rcu_task_stall_timeout)); | |
1078 | if (ret) | |
1079 | break; // Count reached zero. | |
af051ca4 | 1080 | // Stall warning time, so make a list of the offenders. |
4593e772 | 1081 | for_each_process_thread(g, t) |
276c4104 | 1082 | if (READ_ONCE(t->trc_reader_special.b.need_qs)) |
4593e772 PM |
1083 | trc_add_holdout(t, &holdouts); |
1084 | firstreport = true; | |
1085 | list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) | |
276c4104 | 1086 | if (READ_ONCE(t->trc_reader_special.b.need_qs)) { |
4593e772 PM |
1087 | show_stalled_task_trace(t, &firstreport); |
1088 | trc_del_holdout(t); | |
1089 | } | |
1090 | if (firstreport) | |
1091 | pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n"); | |
1092 | show_stalled_ipi_trace(); | |
1093 | pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end)); | |
1094 | } | |
d5f177d3 | 1095 | smp_mb(); // Caller's code must be ordered after wakeup. |
43766c3e | 1096 | // Pairs with pretty much every ordering primitive. |
d5f177d3 PM |
1097 | } |
1098 | ||
1099 | /* Report any needed quiescent state for this exiting task. */ | |
25246fc8 | 1100 | static void exit_tasks_rcu_finish_trace(struct task_struct *t) |
d5f177d3 PM |
1101 | { |
1102 | WRITE_ONCE(t->trc_reader_checked, true); | |
1103 | WARN_ON_ONCE(t->trc_reader_nesting); | |
1104 | WRITE_ONCE(t->trc_reader_nesting, 0); | |
276c4104 PM |
1105 | if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs))) |
1106 | rcu_read_unlock_trace_special(t, 0); | |
d5f177d3 PM |
1107 | } |
1108 | ||
d5f177d3 PM |
1109 | /** |
1110 | * call_rcu_tasks_trace() - Queue a callback trace task-based grace period | |
1111 | * @rhp: structure to be used for queueing the RCU updates. | |
1112 | * @func: actual callback function to be invoked after the grace period | |
1113 | * | |
1114 | * The callback function will be invoked some time after a full grace | |
1115 | * period elapses, in other words after all currently executing RCU | |
1116 | * read-side critical sections have completed. call_rcu_tasks_trace() | |
1117 | * assumes that the read-side critical sections end at context switch, | |
1118 | * cond_resched_rcu_qs(), or transition to usermode execution. As such, | |
1119 | * there are no read-side primitives analogous to rcu_read_lock() and | |
1120 | * rcu_read_unlock() because this primitive is intended to determine | |
1121 | * that all tasks have passed through a safe state, not so much for | |
1122 | * data-strcuture synchronization. | |
1123 | * | |
1124 | * See the description of call_rcu() for more detailed information on | |
1125 | * memory ordering guarantees. | |
1126 | */ | |
1127 | void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func) | |
1128 | { | |
1129 | call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace); | |
1130 | } | |
1131 | EXPORT_SYMBOL_GPL(call_rcu_tasks_trace); | |
1132 | ||
1133 | /** | |
1134 | * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period | |
1135 | * | |
1136 | * Control will return to the caller some time after a trace rcu-tasks | |
c7dcf810 PM |
1137 | * grace period has elapsed, in other words after all currently executing |
1138 | * rcu-tasks read-side critical sections have elapsed. These read-side | |
1139 | * critical sections are delimited by calls to rcu_read_lock_trace() | |
1140 | * and rcu_read_unlock_trace(). | |
d5f177d3 PM |
1141 | * |
1142 | * This is a very specialized primitive, intended only for a few uses in | |
1143 | * tracing and other situations requiring manipulation of function preambles | |
1144 | * and profiling hooks. The synchronize_rcu_tasks_trace() function is not | |
1145 | * (yet) intended for heavy use from multiple CPUs. | |
1146 | * | |
1147 | * See the description of synchronize_rcu() for more detailed information | |
1148 | * on memory ordering guarantees. | |
1149 | */ | |
1150 | void synchronize_rcu_tasks_trace(void) | |
1151 | { | |
1152 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section"); | |
1153 | synchronize_rcu_tasks_generic(&rcu_tasks_trace); | |
1154 | } | |
1155 | EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace); | |
1156 | ||
1157 | /** | |
1158 | * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks. | |
1159 | * | |
1160 | * Although the current implementation is guaranteed to wait, it is not | |
1161 | * obligated to, for example, if there are no pending callbacks. | |
1162 | */ | |
1163 | void rcu_barrier_tasks_trace(void) | |
1164 | { | |
1165 | /* There is only one callback queue, so this is easy. ;-) */ | |
1166 | synchronize_rcu_tasks_trace(); | |
1167 | } | |
1168 | EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace); | |
1169 | ||
1170 | static int __init rcu_spawn_tasks_trace_kthread(void) | |
1171 | { | |
2393a613 PM |
1172 | if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) { |
1173 | rcu_tasks_trace.init_fract = 10; | |
1174 | } else { | |
1175 | rcu_tasks_trace.init_fract = HZ / 5; | |
1176 | if (rcu_tasks_trace.init_fract <= 0) | |
1177 | rcu_tasks_trace.init_fract = 1; | |
1178 | } | |
d5f177d3 PM |
1179 | rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step; |
1180 | rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask; | |
1181 | rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan; | |
1182 | rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace; | |
1183 | rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp; | |
1184 | rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace); | |
1185 | return 0; | |
1186 | } | |
1187 | core_initcall(rcu_spawn_tasks_trace_kthread); | |
1188 | ||
8344496e | 1189 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
1190 | static void show_rcu_tasks_trace_gp_kthread(void) |
1191 | { | |
40471509 | 1192 | char buf[64]; |
e21408ce | 1193 | |
edf3775f PM |
1194 | sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end), |
1195 | data_race(n_heavy_reader_ofl_updates), | |
40471509 PM |
1196 | data_race(n_heavy_reader_updates), |
1197 | data_race(n_heavy_reader_attempts)); | |
e21408ce PM |
1198 | show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf); |
1199 | } | |
8344496e | 1200 | #endif /* #ifndef CONFIG_TINY_RCU */ |
e21408ce | 1201 | |
d5f177d3 | 1202 | #else /* #ifdef CONFIG_TASKS_TRACE_RCU */ |
25246fc8 | 1203 | static void exit_tasks_rcu_finish_trace(struct task_struct *t) { } |
e21408ce | 1204 | static inline void show_rcu_tasks_trace_gp_kthread(void) {} |
d5f177d3 | 1205 | #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */ |
8fd8ca38 | 1206 | |
8344496e | 1207 | #ifndef CONFIG_TINY_RCU |
e21408ce PM |
1208 | void show_rcu_tasks_gp_kthreads(void) |
1209 | { | |
1210 | show_rcu_tasks_classic_gp_kthread(); | |
1211 | show_rcu_tasks_rude_gp_kthread(); | |
1212 | show_rcu_tasks_trace_gp_kthread(); | |
1213 | } | |
8344496e | 1214 | #endif /* #ifndef CONFIG_TINY_RCU */ |
e21408ce | 1215 | |
8fd8ca38 PM |
1216 | #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */ |
1217 | static inline void rcu_tasks_bootup_oddness(void) {} | |
e21408ce | 1218 | void show_rcu_tasks_gp_kthreads(void) {} |
8fd8ca38 | 1219 | #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */ |