1 /* SPDX-License-Identifier: GPL-2.0+ */
2 /*
3 * Task-based RCU implementations.
4 *
5 * Copyright (C) 2020 Paul E. McKenney
6 */
7
8 #ifdef CONFIG_TASKS_RCU_GENERIC
9 #include "rcu_segcblist.h"
10
11 ////////////////////////////////////////////////////////////////////////
12 //
13 // Generic data structures.
14
15 struct rcu_tasks;
16 typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp);
17 typedef void (*pregp_func_t)(struct list_head *hop);
18 typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop);
19 typedef void (*postscan_func_t)(struct list_head *hop);
20 typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp);
21 typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
22
23 /**
24 * struct rcu_tasks_percpu - Per-CPU component of definition for a Tasks-RCU-like mechanism.
25 * @cblist: Callback list.
26 * @lock: Lock protecting per-CPU callback list.
27 * @rtp_jiffies: Jiffies counter value for statistics.
28 * @rtp_n_lock_retries: Rough lock-contention statistic.
29 * @rtp_work: Work queue for invoking callbacks.
30 * @rtp_irq_work: IRQ work queue for deferred wakeups.
31 * @barrier_q_head: RCU callback for barrier operation.
32 * @rtp_blkd_tasks: List of tasks blocked as readers.
33 * @cpu: CPU number corresponding to this entry.
34 * @rtpp: Pointer to the rcu_tasks structure.
35 */
36 struct rcu_tasks_percpu {
37 struct rcu_segcblist cblist;
38 raw_spinlock_t __private lock;
39 unsigned long rtp_jiffies;
40 unsigned long rtp_n_lock_retries;
41 struct work_struct rtp_work;
42 struct irq_work rtp_irq_work;
43 struct rcu_head barrier_q_head;
44 struct list_head rtp_blkd_tasks;
45 int cpu;
46 struct rcu_tasks *rtpp;
47 };
48
49 /**
50 * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
51 * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
52 * @cbs_gbl_lock: Lock protecting callback list.
53 * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone.
54 * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
55 * @gp_func: This flavor's grace-period-wait function.
56 * @gp_state: Grace period's most recent state transition (debugging).
57 * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
58 * @init_fract: Initial backoff sleep interval.
59 * @gp_jiffies: Time of last @gp_state transition.
60 * @gp_start: Most recent grace-period start in jiffies.
61 * @tasks_gp_seq: Number of grace periods completed since boot.
62 * @n_ipis: Number of IPIs sent to encourage grace periods to end.
63 * @n_ipis_fails: Number of IPI-send failures.
64 * @pregp_func: This flavor's pre-grace-period function (optional).
65 * @pertask_func: This flavor's per-task scan function (optional).
66 * @postscan_func: This flavor's post-task scan function (optional).
67 * @holdouts_func: This flavor's holdout-list scan function (optional).
68 * @postgp_func: This flavor's post-grace-period function (optional).
69 * @call_func: This flavor's call_rcu()-equivalent function.
70 * @rtpcpu: This flavor's rcu_tasks_percpu structure.
71 * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
72 * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
73 * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
74 * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
75 * @barrier_q_mutex: Serialize barrier operations.
76 * @barrier_q_count: Number of queues being waited on.
77 * @barrier_q_completion: Barrier wait/wakeup mechanism.
78 * @barrier_q_seq: Sequence number for barrier operations.
79 * @name: This flavor's textual name.
80 * @kname: This flavor's kthread name.
81 */
82 struct rcu_tasks {
83 struct rcuwait cbs_wait;
84 raw_spinlock_t cbs_gbl_lock;
85 struct mutex tasks_gp_mutex;
86 int gp_state;
87 int gp_sleep;
88 int init_fract;
89 unsigned long gp_jiffies;
90 unsigned long gp_start;
91 unsigned long tasks_gp_seq;
92 unsigned long n_ipis;
93 unsigned long n_ipis_fails;
94 struct task_struct *kthread_ptr;
95 rcu_tasks_gp_func_t gp_func;
96 pregp_func_t pregp_func;
97 pertask_func_t pertask_func;
98 postscan_func_t postscan_func;
99 holdouts_func_t holdouts_func;
100 postgp_func_t postgp_func;
101 call_rcu_func_t call_func;
102 struct rcu_tasks_percpu __percpu *rtpcpu;
103 int percpu_enqueue_shift;
104 int percpu_enqueue_lim;
105 int percpu_dequeue_lim;
106 unsigned long percpu_dequeue_gpseq;
107 struct mutex barrier_q_mutex;
108 atomic_t barrier_q_count;
109 struct completion barrier_q_completion;
110 unsigned long barrier_q_seq;
111 char *name;
112 char *kname;
113 };
114
115 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
116
117 #define DEFINE_RCU_TASKS(rt_name, gp, call, n) \
118 static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = { \
119 .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock), \
120 .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup), \
121 }; \
122 static struct rcu_tasks rt_name = \
123 { \
124 .cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait), \
125 .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock), \
126 .tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex), \
127 .gp_func = gp, \
128 .call_func = call, \
129 .rtpcpu = &rt_name ## __percpu, \
130 .name = n, \
131 .percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS), \
132 .percpu_enqueue_lim = 1, \
133 .percpu_dequeue_lim = 1, \
134 .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex), \
135 .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT, \
136 .kname = #rt_name, \
137 }
138
139 /* Track exiting tasks in order to allow them to be waited for. */
140 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
141
142 /* Avoid IPIing CPUs early in the grace period. */
143 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
144 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
145 module_param(rcu_task_ipi_delay, int, 0644);
146
147 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
148 #define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
149 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
150 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
151 module_param(rcu_task_stall_timeout, int, 0644);
152 #define RCU_TASK_STALL_INFO (HZ * 10)
153 static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
154 module_param(rcu_task_stall_info, int, 0644);
155 static int rcu_task_stall_info_mult __read_mostly = 3;
156 module_param(rcu_task_stall_info_mult, int, 0444);
157
158 static int rcu_task_enqueue_lim __read_mostly = -1;
159 module_param(rcu_task_enqueue_lim, int, 0444);
160
161 static bool rcu_task_cb_adjust;
162 static int rcu_task_contend_lim __read_mostly = 100;
163 module_param(rcu_task_contend_lim, int, 0444);
164 static int rcu_task_collapse_lim __read_mostly = 10;
165 module_param(rcu_task_collapse_lim, int, 0444);
166
167 /* RCU tasks grace-period state for debugging. */
168 #define RTGS_INIT 0
169 #define RTGS_WAIT_WAIT_CBS 1
170 #define RTGS_WAIT_GP 2
171 #define RTGS_PRE_WAIT_GP 3
172 #define RTGS_SCAN_TASKLIST 4
173 #define RTGS_POST_SCAN_TASKLIST 5
174 #define RTGS_WAIT_SCAN_HOLDOUTS 6
175 #define RTGS_SCAN_HOLDOUTS 7
176 #define RTGS_POST_GP 8
177 #define RTGS_WAIT_READERS 9
178 #define RTGS_INVOKE_CBS 10
179 #define RTGS_WAIT_CBS 11
180 #ifndef CONFIG_TINY_RCU
181 static const char * const rcu_tasks_gp_state_names[] = {
182 "RTGS_INIT",
183 "RTGS_WAIT_WAIT_CBS",
184 "RTGS_WAIT_GP",
185 "RTGS_PRE_WAIT_GP",
186 "RTGS_SCAN_TASKLIST",
187 "RTGS_POST_SCAN_TASKLIST",
188 "RTGS_WAIT_SCAN_HOLDOUTS",
189 "RTGS_SCAN_HOLDOUTS",
190 "RTGS_POST_GP",
191 "RTGS_WAIT_READERS",
192 "RTGS_INVOKE_CBS",
193 "RTGS_WAIT_CBS",
194 };
195 #endif /* #ifndef CONFIG_TINY_RCU */
196
197 ////////////////////////////////////////////////////////////////////////
198 //
199 // Generic code.
200
201 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
202
203 /* Record grace-period phase and time. */
set_tasks_gp_state(struct rcu_tasks * rtp,int newstate)204 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
205 {
206 rtp->gp_state = newstate;
207 rtp->gp_jiffies = jiffies;
208 }
209
210 #ifndef CONFIG_TINY_RCU
211 /* Return state name. */
tasks_gp_state_getname(struct rcu_tasks * rtp)212 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
213 {
214 int i = data_race(rtp->gp_state); // Let KCSAN detect update races
215 int j = READ_ONCE(i); // Prevent the compiler from reading twice
216
217 if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
218 return "???";
219 return rcu_tasks_gp_state_names[j];
220 }
221 #endif /* #ifndef CONFIG_TINY_RCU */
222
223 // Initialize per-CPU callback lists for the specified flavor of
224 // Tasks RCU.
cblist_init_generic(struct rcu_tasks * rtp)225 static void cblist_init_generic(struct rcu_tasks *rtp)
226 {
227 int cpu;
228 unsigned long flags;
229 int lim;
230 int shift;
231
232 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
233 if (rcu_task_enqueue_lim < 0) {
234 rcu_task_enqueue_lim = 1;
235 rcu_task_cb_adjust = true;
236 } else if (rcu_task_enqueue_lim == 0) {
237 rcu_task_enqueue_lim = 1;
238 }
239 lim = rcu_task_enqueue_lim;
240
241 if (lim > nr_cpu_ids)
242 lim = nr_cpu_ids;
243 shift = ilog2(nr_cpu_ids / lim);
244 if (((nr_cpu_ids - 1) >> shift) >= lim)
245 shift++;
246 WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
247 WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
248 smp_store_release(&rtp->percpu_enqueue_lim, lim);
249 for_each_possible_cpu(cpu) {
250 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
251
252 WARN_ON_ONCE(!rtpcp);
253 if (cpu)
254 raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
255 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
256 if (rcu_segcblist_empty(&rtpcp->cblist))
257 rcu_segcblist_init(&rtpcp->cblist);
258 INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
259 rtpcp->cpu = cpu;
260 rtpcp->rtpp = rtp;
261 if (!rtpcp->rtp_blkd_tasks.next)
262 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
263 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
264 }
265 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
266
267 if (rcu_task_cb_adjust)
268 pr_info("%s: Setting adjustable number of callback queues.\n", __func__);
269
270 pr_info("%s: Setting shift to %d and lim to %d.\n", __func__, data_race(rtp->percpu_enqueue_shift), data_race(rtp->percpu_enqueue_lim));
271 }
272
273 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
call_rcu_tasks_iw_wakeup(struct irq_work * iwp)274 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
275 {
276 struct rcu_tasks *rtp;
277 struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
278
279 rtp = rtpcp->rtpp;
280 rcuwait_wake_up(&rtp->cbs_wait);
281 }
282
283 // Enqueue a callback for the specified flavor of Tasks RCU.
call_rcu_tasks_generic(struct rcu_head * rhp,rcu_callback_t func,struct rcu_tasks * rtp)284 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
285 struct rcu_tasks *rtp)
286 {
287 int chosen_cpu;
288 unsigned long flags;
289 int ideal_cpu;
290 unsigned long j;
291 bool needadjust = false;
292 bool needwake;
293 struct rcu_tasks_percpu *rtpcp;
294
295 rhp->next = NULL;
296 rhp->func = func;
297 local_irq_save(flags);
298 rcu_read_lock();
299 ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
300 chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
301 rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
302 if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
303 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
304 j = jiffies;
305 if (rtpcp->rtp_jiffies != j) {
306 rtpcp->rtp_jiffies = j;
307 rtpcp->rtp_n_lock_retries = 0;
308 }
309 if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
310 READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
311 needadjust = true; // Defer adjustment to avoid deadlock.
312 }
313 if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) {
314 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
315 cblist_init_generic(rtp);
316 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
317 }
318 needwake = rcu_segcblist_empty(&rtpcp->cblist);
319 rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
320 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
321 if (unlikely(needadjust)) {
322 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
323 if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
324 WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
325 WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
326 smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
327 pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
328 }
329 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
330 }
331 rcu_read_unlock();
332 /* We can't create the thread unless interrupts are enabled. */
333 if (needwake && READ_ONCE(rtp->kthread_ptr))
334 irq_work_queue(&rtpcp->rtp_irq_work);
335 }
336
337 // RCU callback function for rcu_barrier_tasks_generic().
rcu_barrier_tasks_generic_cb(struct rcu_head * rhp)338 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
339 {
340 struct rcu_tasks *rtp;
341 struct rcu_tasks_percpu *rtpcp;
342
343 rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
344 rtp = rtpcp->rtpp;
345 if (atomic_dec_and_test(&rtp->barrier_q_count))
346 complete(&rtp->barrier_q_completion);
347 }
348
349 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
350 // Operates in a manner similar to rcu_barrier().
rcu_barrier_tasks_generic(struct rcu_tasks * rtp)351 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
352 {
353 int cpu;
354 unsigned long flags;
355 struct rcu_tasks_percpu *rtpcp;
356 unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
357
358 mutex_lock(&rtp->barrier_q_mutex);
359 if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
360 smp_mb();
361 mutex_unlock(&rtp->barrier_q_mutex);
362 return;
363 }
364 rcu_seq_start(&rtp->barrier_q_seq);
365 init_completion(&rtp->barrier_q_completion);
366 atomic_set(&rtp->barrier_q_count, 2);
367 for_each_possible_cpu(cpu) {
368 if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
369 break;
370 rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
371 rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
372 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
373 if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
374 atomic_inc(&rtp->barrier_q_count);
375 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
376 }
377 if (atomic_sub_and_test(2, &rtp->barrier_q_count))
378 complete(&rtp->barrier_q_completion);
379 wait_for_completion(&rtp->barrier_q_completion);
380 rcu_seq_end(&rtp->barrier_q_seq);
381 mutex_unlock(&rtp->barrier_q_mutex);
382 }
383
384 // Advance callbacks and indicate whether either a grace period or
385 // callback invocation is needed.
rcu_tasks_need_gpcb(struct rcu_tasks * rtp)386 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
387 {
388 int cpu;
389 unsigned long flags;
390 bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
391 long n;
392 long ncbs = 0;
393 long ncbsnz = 0;
394 int needgpcb = 0;
395
396 for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
397 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
398
399 /* Advance and accelerate any new callbacks. */
400 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
401 continue;
402 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
403 // Should we shrink down to a single callback queue?
404 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
405 if (n) {
406 ncbs += n;
407 if (cpu > 0)
408 ncbsnz += n;
409 }
410 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
411 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
412 if (rcu_segcblist_pend_cbs(&rtpcp->cblist))
413 needgpcb |= 0x3;
414 if (!rcu_segcblist_empty(&rtpcp->cblist))
415 needgpcb |= 0x1;
416 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
417 }
418
419 // Shrink down to a single callback queue if appropriate.
420 // This is done in two stages: (1) If there are no more than
421 // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
422 // CPU, limit enqueueing to CPU 0. (2) After an RCU grace period,
423 // if there has not been an increase in callbacks, limit dequeuing
424 // to CPU 0. Note the matching RCU read-side critical section in
425 // call_rcu_tasks_generic().
426 if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
427 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
428 if (rtp->percpu_enqueue_lim > 1) {
429 WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
430 smp_store_release(&rtp->percpu_enqueue_lim, 1);
431 rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
432 gpdone = false;
433 pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
434 }
435 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
436 }
437 if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
438 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
439 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
440 WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
441 pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
442 }
443 if (rtp->percpu_dequeue_lim == 1) {
444 for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) {
445 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
446
447 WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
448 }
449 }
450 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
451 }
452
453 return needgpcb;
454 }
455
456 // Advance callbacks and invoke any that are ready.
rcu_tasks_invoke_cbs(struct rcu_tasks * rtp,struct rcu_tasks_percpu * rtpcp)457 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
458 {
459 int cpu;
460 int cpunext;
461 int cpuwq;
462 unsigned long flags;
463 int len;
464 struct rcu_head *rhp;
465 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
466 struct rcu_tasks_percpu *rtpcp_next;
467
468 cpu = rtpcp->cpu;
469 cpunext = cpu * 2 + 1;
470 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
471 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
472 cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
473 queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
474 cpunext++;
475 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
476 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
477 cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
478 queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
479 }
480 }
481
482 if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
483 return;
484 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
485 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
486 rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
487 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
488 len = rcl.len;
489 for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
490 local_bh_disable();
491 rhp->func(rhp);
492 local_bh_enable();
493 cond_resched();
494 }
495 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
496 rcu_segcblist_add_len(&rtpcp->cblist, -len);
497 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
498 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
499 }
500
501 // Workqueue flood to advance callbacks and invoke any that are ready.
rcu_tasks_invoke_cbs_wq(struct work_struct * wp)502 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
503 {
504 struct rcu_tasks *rtp;
505 struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
506
507 rtp = rtpcp->rtpp;
508 rcu_tasks_invoke_cbs(rtp, rtpcp);
509 }
510
511 // Wait for one grace period.
rcu_tasks_one_gp(struct rcu_tasks * rtp,bool midboot)512 static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
513 {
514 int needgpcb;
515
516 mutex_lock(&rtp->tasks_gp_mutex);
517
518 // If there were none, wait a bit and start over.
519 if (unlikely(midboot)) {
520 needgpcb = 0x2;
521 } else {
522 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
523 rcuwait_wait_event(&rtp->cbs_wait,
524 (needgpcb = rcu_tasks_need_gpcb(rtp)),
525 TASK_IDLE);
526 }
527
528 if (needgpcb & 0x2) {
529 // Wait for one grace period.
530 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
531 rtp->gp_start = jiffies;
532 rcu_seq_start(&rtp->tasks_gp_seq);
533 rtp->gp_func(rtp);
534 rcu_seq_end(&rtp->tasks_gp_seq);
535 }
536
537 // Invoke callbacks.
538 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
539 rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
540 mutex_unlock(&rtp->tasks_gp_mutex);
541 }
542
543 // RCU-tasks kthread that detects grace periods and invokes callbacks.
rcu_tasks_kthread(void * arg)544 static int __noreturn rcu_tasks_kthread(void *arg)
545 {
546 struct rcu_tasks *rtp = arg;
547
548 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
549 housekeeping_affine(current, HK_TYPE_RCU);
550 WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
551
552 /*
553 * Each pass through the following loop makes one check for
554 * newly arrived callbacks, and, if there are some, waits for
555 * one RCU-tasks grace period and then invokes the callbacks.
556 * This loop is terminated by the system going down. ;-)
557 */
558 for (;;) {
559 // Wait for one grace period and invoke any callbacks
560 // that are ready.
561 rcu_tasks_one_gp(rtp, false);
562
563 // Paranoid sleep to keep this from entering a tight loop.
564 schedule_timeout_idle(rtp->gp_sleep);
565 }
566 }
567
568 // Wait for a grace period for the specified flavor of Tasks RCU.
synchronize_rcu_tasks_generic(struct rcu_tasks * rtp)569 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
570 {
571 /* Complain if the scheduler has not started. */
572 if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
573 "synchronize_%s() called too soon", rtp->name))
574 return;
575
576 // If the grace-period kthread is running, use it.
577 if (READ_ONCE(rtp->kthread_ptr)) {
578 wait_rcu_gp(rtp->call_func);
579 return;
580 }
581 rcu_tasks_one_gp(rtp, true);
582 }
583
584 /* Spawn RCU-tasks grace-period kthread. */
rcu_spawn_tasks_kthread_generic(struct rcu_tasks * rtp)585 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
586 {
587 struct task_struct *t;
588
589 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
590 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
591 return;
592 smp_mb(); /* Ensure others see full kthread. */
593 }
594
595 #ifndef CONFIG_TINY_RCU
596
597 /*
598 * Print any non-default Tasks RCU settings.
599 */
rcu_tasks_bootup_oddness(void)600 static void __init rcu_tasks_bootup_oddness(void)
601 {
602 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
603 int rtsimc;
604
605 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
606 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
607 rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
608 if (rtsimc != rcu_task_stall_info_mult) {
609 pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
610 rcu_task_stall_info_mult = rtsimc;
611 }
612 #endif /* #ifdef CONFIG_TASKS_RCU */
613 #ifdef CONFIG_TASKS_RCU
614 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
615 #endif /* #ifdef CONFIG_TASKS_RCU */
616 #ifdef CONFIG_TASKS_RUDE_RCU
617 pr_info("\tRude variant of Tasks RCU enabled.\n");
618 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
619 #ifdef CONFIG_TASKS_TRACE_RCU
620 pr_info("\tTracing variant of Tasks RCU enabled.\n");
621 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
622 }
623
624 #endif /* #ifndef CONFIG_TINY_RCU */
625
626 #ifndef CONFIG_TINY_RCU
627 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
show_rcu_tasks_generic_gp_kthread(struct rcu_tasks * rtp,char * s)628 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
629 {
630 int cpu;
631 bool havecbs = false;
632
633 for_each_possible_cpu(cpu) {
634 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
635
636 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist))) {
637 havecbs = true;
638 break;
639 }
640 }
641 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
642 rtp->kname,
643 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
644 jiffies - data_race(rtp->gp_jiffies),
645 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
646 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
647 ".k"[!!data_race(rtp->kthread_ptr)],
648 ".C"[havecbs],
649 s);
650 }
651 #endif // #ifndef CONFIG_TINY_RCU
652
653 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
654
655 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
656
657 ////////////////////////////////////////////////////////////////////////
658 //
659 // Shared code between task-list-scanning variants of Tasks RCU.
660
661 /* Wait for one RCU-tasks grace period. */
rcu_tasks_wait_gp(struct rcu_tasks * rtp)662 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
663 {
664 struct task_struct *g;
665 int fract;
666 LIST_HEAD(holdouts);
667 unsigned long j;
668 unsigned long lastinfo;
669 unsigned long lastreport;
670 bool reported = false;
671 int rtsi;
672 struct task_struct *t;
673
674 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
675 rtp->pregp_func(&holdouts);
676
677 /*
678 * There were callbacks, so we need to wait for an RCU-tasks
679 * grace period. Start off by scanning the task list for tasks
680 * that are not already voluntarily blocked. Mark these tasks
681 * and make a list of them in holdouts.
682 */
683 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
684 if (rtp->pertask_func) {
685 rcu_read_lock();
686 for_each_process_thread(g, t)
687 rtp->pertask_func(t, &holdouts);
688 rcu_read_unlock();
689 }
690
691 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
692 rtp->postscan_func(&holdouts);
693
694 /*
695 * Each pass through the following loop scans the list of holdout
696 * tasks, removing any that are no longer holdouts. When the list
697 * is empty, we are done.
698 */
699 lastreport = jiffies;
700 lastinfo = lastreport;
701 rtsi = READ_ONCE(rcu_task_stall_info);
702
703 // Start off with initial wait and slowly back off to 1 HZ wait.
704 fract = rtp->init_fract;
705
706 while (!list_empty(&holdouts)) {
707 ktime_t exp;
708 bool firstreport;
709 bool needreport;
710 int rtst;
711
712 // Slowly back off waiting for holdouts
713 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
714 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
715 schedule_timeout_idle(fract);
716 } else {
717 exp = jiffies_to_nsecs(fract);
718 __set_current_state(TASK_IDLE);
719 schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
720 }
721
722 if (fract < HZ)
723 fract++;
724
725 rtst = READ_ONCE(rcu_task_stall_timeout);
726 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
727 if (needreport) {
728 lastreport = jiffies;
729 reported = true;
730 }
731 firstreport = true;
732 WARN_ON(signal_pending(current));
733 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
734 rtp->holdouts_func(&holdouts, needreport, &firstreport);
735
736 // Print pre-stall informational messages if needed.
737 j = jiffies;
738 if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
739 lastinfo = j;
740 rtsi = rtsi * rcu_task_stall_info_mult;
741 pr_info("%s: %s grace period %lu is %lu jiffies old.\n",
742 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
743 }
744 }
745
746 set_tasks_gp_state(rtp, RTGS_POST_GP);
747 rtp->postgp_func(rtp);
748 }
749
750 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
751
752 #ifdef CONFIG_TASKS_RCU
753
754 ////////////////////////////////////////////////////////////////////////
755 //
756 // Simple variant of RCU whose quiescent states are voluntary context
757 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
758 // As such, grace periods can take one good long time. There are no
759 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
760 // because this implementation is intended to get the system into a safe
761 // state for some of the manipulations involved in tracing and the like.
762 // Finally, this implementation does not support high call_rcu_tasks()
763 // rates from multiple CPUs. If this is required, per-CPU callback lists
764 // will be needed.
765 //
766 // The implementation uses rcu_tasks_wait_gp(), which relies on function
767 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
768 // function sets these function pointers up so that rcu_tasks_wait_gp()
769 // invokes these functions in this order:
770 //
771 // rcu_tasks_pregp_step():
772 // Invokes synchronize_rcu() in order to wait for all in-flight
773 // t->on_rq and t->nvcsw transitions to complete. This works because
774 // all such transitions are carried out with interrupts disabled.
775 // rcu_tasks_pertask(), invoked on every non-idle task:
776 // For every runnable non-idle task other than the current one, use
777 // get_task_struct() to pin down that task, snapshot that task's
778 // number of voluntary context switches, and add that task to the
779 // holdout list.
780 // rcu_tasks_postscan():
781 // Invoke synchronize_srcu() to ensure that all tasks that were
782 // in the process of exiting (and which thus might not know to
783 // synchronize with this RCU Tasks grace period) have completed
784 // exiting.
785 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
786 // Scans the holdout list, attempting to identify a quiescent state
787 // for each task on the list. If there is a quiescent state, the
788 // corresponding task is removed from the holdout list.
789 // rcu_tasks_postgp():
790 // Invokes synchronize_rcu() in order to ensure that all prior
791 // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
792 // to have happened before the end of this RCU Tasks grace period.
793 // Again, this works because all such transitions are carried out
794 // with interrupts disabled.
795 //
796 // For each exiting task, the exit_tasks_rcu_start() and
797 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
798 // read-side critical sections waited for by rcu_tasks_postscan().
799 //
800 // Pre-grace-period update-side code is ordered before the grace
801 // via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code
802 // is ordered before the grace period via synchronize_rcu() call in
803 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
804 // disabling.
805
806 /* Pre-grace-period preparation. */
rcu_tasks_pregp_step(struct list_head * hop)807 static void rcu_tasks_pregp_step(struct list_head *hop)
808 {
809 /*
810 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
811 * to complete. Invoking synchronize_rcu() suffices because all
812 * these transitions occur with interrupts disabled. Without this
813 * synchronize_rcu(), a read-side critical section that started
814 * before the grace period might be incorrectly seen as having
815 * started after the grace period.
816 *
817 * This synchronize_rcu() also dispenses with the need for a
818 * memory barrier on the first store to t->rcu_tasks_holdout,
819 * as it forces the store to happen after the beginning of the
820 * grace period.
821 */
822 synchronize_rcu();
823 }
824
825 /* Per-task initial processing. */
rcu_tasks_pertask(struct task_struct * t,struct list_head * hop)826 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
827 {
828 if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
829 get_task_struct(t);
830 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
831 WRITE_ONCE(t->rcu_tasks_holdout, true);
832 list_add(&t->rcu_tasks_holdout_list, hop);
833 }
834 }
835
836 /* Processing between scanning taskslist and draining the holdout list. */
rcu_tasks_postscan(struct list_head * hop)837 static void rcu_tasks_postscan(struct list_head *hop)
838 {
839 /*
840 * Exiting tasks may escape the tasklist scan. Those are vulnerable
841 * until their final schedule() with TASK_DEAD state. To cope with
842 * this, divide the fragile exit path part in two intersecting
843 * read side critical sections:
844 *
845 * 1) An _SRCU_ read side starting before calling exit_notify(),
846 * which may remove the task from the tasklist, and ending after
847 * the final preempt_disable() call in do_exit().
848 *
849 * 2) An _RCU_ read side starting with the final preempt_disable()
850 * call in do_exit() and ending with the final call to schedule()
851 * with TASK_DEAD state.
852 *
853 * This handles the part 1). And postgp will handle part 2) with a
854 * call to synchronize_rcu().
855 */
856 synchronize_srcu(&tasks_rcu_exit_srcu);
857 }
858
859 /* See if tasks are still holding out, complain if so. */
check_holdout_task(struct task_struct * t,bool needreport,bool * firstreport)860 static void check_holdout_task(struct task_struct *t,
861 bool needreport, bool *firstreport)
862 {
863 int cpu;
864
865 if (!READ_ONCE(t->rcu_tasks_holdout) ||
866 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
867 !READ_ONCE(t->on_rq) ||
868 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
869 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
870 WRITE_ONCE(t->rcu_tasks_holdout, false);
871 list_del_init(&t->rcu_tasks_holdout_list);
872 put_task_struct(t);
873 return;
874 }
875 rcu_request_urgent_qs_task(t);
876 if (!needreport)
877 return;
878 if (*firstreport) {
879 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
880 *firstreport = false;
881 }
882 cpu = task_cpu(t);
883 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
884 t, ".I"[is_idle_task(t)],
885 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
886 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
887 t->rcu_tasks_idle_cpu, cpu);
888 sched_show_task(t);
889 }
890
891 /* Scan the holdout lists for tasks no longer holding out. */
check_all_holdout_tasks(struct list_head * hop,bool needreport,bool * firstreport)892 static void check_all_holdout_tasks(struct list_head *hop,
893 bool needreport, bool *firstreport)
894 {
895 struct task_struct *t, *t1;
896
897 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
898 check_holdout_task(t, needreport, firstreport);
899 cond_resched();
900 }
901 }
902
903 /* Finish off the Tasks-RCU grace period. */
rcu_tasks_postgp(struct rcu_tasks * rtp)904 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
905 {
906 /*
907 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
908 * memory barriers prior to them in the schedule() path, memory
909 * reordering on other CPUs could cause their RCU-tasks read-side
910 * critical sections to extend past the end of the grace period.
911 * However, because these ->nvcsw updates are carried out with
912 * interrupts disabled, we can use synchronize_rcu() to force the
913 * needed ordering on all such CPUs.
914 *
915 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
916 * accesses to be within the grace period, avoiding the need for
917 * memory barriers for ->rcu_tasks_holdout accesses.
918 *
919 * In addition, this synchronize_rcu() waits for exiting tasks
920 * to complete their final preempt_disable() region of execution,
921 * cleaning up after synchronize_srcu(&tasks_rcu_exit_srcu),
922 * enforcing the whole region before tasklist removal until
923 * the final schedule() with TASK_DEAD state to be an RCU TASKS
924 * read side critical section.
925 */
926 synchronize_rcu();
927 }
928
929 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
930 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
931
932 /**
933 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
934 * @rhp: structure to be used for queueing the RCU updates.
935 * @func: actual callback function to be invoked after the grace period
936 *
937 * The callback function will be invoked some time after a full grace
938 * period elapses, in other words after all currently executing RCU
939 * read-side critical sections have completed. call_rcu_tasks() assumes
940 * that the read-side critical sections end at a voluntary context
941 * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
942 * or transition to usermode execution. As such, there are no read-side
943 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
944 * this primitive is intended to determine that all tasks have passed
945 * through a safe state, not so much for data-structure synchronization.
946 *
947 * See the description of call_rcu() for more detailed information on
948 * memory ordering guarantees.
949 */
call_rcu_tasks(struct rcu_head * rhp,rcu_callback_t func)950 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
951 {
952 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
953 }
954 EXPORT_SYMBOL_GPL(call_rcu_tasks);
955
956 /**
957 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
958 *
959 * Control will return to the caller some time after a full rcu-tasks
960 * grace period has elapsed, in other words after all currently
961 * executing rcu-tasks read-side critical sections have elapsed. These
962 * read-side critical sections are delimited by calls to schedule(),
963 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
964 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
965 *
966 * This is a very specialized primitive, intended only for a few uses in
967 * tracing and other situations requiring manipulation of function
968 * preambles and profiling hooks. The synchronize_rcu_tasks() function
969 * is not (yet) intended for heavy use from multiple CPUs.
970 *
971 * See the description of synchronize_rcu() for more detailed information
972 * on memory ordering guarantees.
973 */
synchronize_rcu_tasks(void)974 void synchronize_rcu_tasks(void)
975 {
976 synchronize_rcu_tasks_generic(&rcu_tasks);
977 }
978 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
979
980 /**
981 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
982 *
983 * Although the current implementation is guaranteed to wait, it is not
984 * obligated to, for example, if there are no pending callbacks.
985 */
rcu_barrier_tasks(void)986 void rcu_barrier_tasks(void)
987 {
988 rcu_barrier_tasks_generic(&rcu_tasks);
989 }
990 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
991
rcu_spawn_tasks_kthread(void)992 static int __init rcu_spawn_tasks_kthread(void)
993 {
994 cblist_init_generic(&rcu_tasks);
995 rcu_tasks.gp_sleep = HZ / 10;
996 rcu_tasks.init_fract = HZ / 10;
997 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
998 rcu_tasks.pertask_func = rcu_tasks_pertask;
999 rcu_tasks.postscan_func = rcu_tasks_postscan;
1000 rcu_tasks.holdouts_func = check_all_holdout_tasks;
1001 rcu_tasks.postgp_func = rcu_tasks_postgp;
1002 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1003 return 0;
1004 }
1005
1006 #if !defined(CONFIG_TINY_RCU)
show_rcu_tasks_classic_gp_kthread(void)1007 void show_rcu_tasks_classic_gp_kthread(void)
1008 {
1009 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1010 }
1011 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1012 #endif // !defined(CONFIG_TINY_RCU)
1013
1014 /*
1015 * Contribute to protect against tasklist scan blind spot while the
1016 * task is exiting and may be removed from the tasklist. See
1017 * corresponding synchronize_srcu() for further details.
1018 */
exit_tasks_rcu_start(void)1019 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
1020 {
1021 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
1022 }
1023
1024 /*
1025 * Contribute to protect against tasklist scan blind spot while the
1026 * task is exiting and may be removed from the tasklist. See
1027 * corresponding synchronize_srcu() for further details.
1028 */
exit_tasks_rcu_stop(void)1029 void exit_tasks_rcu_stop(void) __releases(&tasks_rcu_exit_srcu)
1030 {
1031 struct task_struct *t = current;
1032
1033 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
1034 }
1035
1036 /*
1037 * Contribute to protect against tasklist scan blind spot while the
1038 * task is exiting and may be removed from the tasklist. See
1039 * corresponding synchronize_srcu() for further details.
1040 */
exit_tasks_rcu_finish(void)1041 void exit_tasks_rcu_finish(void)
1042 {
1043 exit_tasks_rcu_stop();
1044 exit_tasks_rcu_finish_trace(current);
1045 }
1046
1047 #else /* #ifdef CONFIG_TASKS_RCU */
exit_tasks_rcu_start(void)1048 void exit_tasks_rcu_start(void) { }
exit_tasks_rcu_stop(void)1049 void exit_tasks_rcu_stop(void) { }
exit_tasks_rcu_finish(void)1050 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1051 #endif /* #else #ifdef CONFIG_TASKS_RCU */
1052
1053 #ifdef CONFIG_TASKS_RUDE_RCU
1054
1055 ////////////////////////////////////////////////////////////////////////
1056 //
1057 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
1058 // passing an empty function to schedule_on_each_cpu(). This approach
1059 // provides an asynchronous call_rcu_tasks_rude() API and batching of
1060 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
1061 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
1062 // and induces otherwise unnecessary context switches on all online CPUs,
1063 // whether idle or not.
1064 //
1065 // Callback handling is provided by the rcu_tasks_kthread() function.
1066 //
1067 // Ordering is provided by the scheduler's context-switch code.
1068
1069 // Empty function to allow workqueues to force a context switch.
rcu_tasks_be_rude(struct work_struct * work)1070 static void rcu_tasks_be_rude(struct work_struct *work)
1071 {
1072 }
1073
1074 // Wait for one rude RCU-tasks grace period.
rcu_tasks_rude_wait_gp(struct rcu_tasks * rtp)1075 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1076 {
1077 rtp->n_ipis += cpumask_weight(cpu_online_mask);
1078 schedule_on_each_cpu(rcu_tasks_be_rude);
1079 }
1080
1081 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1082 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1083 "RCU Tasks Rude");
1084
1085 /**
1086 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1087 * @rhp: structure to be used for queueing the RCU updates.
1088 * @func: actual callback function to be invoked after the grace period
1089 *
1090 * The callback function will be invoked some time after a full grace
1091 * period elapses, in other words after all currently executing RCU
1092 * read-side critical sections have completed. call_rcu_tasks_rude()
1093 * assumes that the read-side critical sections end at context switch,
1094 * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1095 * usermode execution is schedulable). As such, there are no read-side
1096 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1097 * this primitive is intended to determine that all tasks have passed
1098 * through a safe state, not so much for data-structure synchronization.
1099 *
1100 * See the description of call_rcu() for more detailed information on
1101 * memory ordering guarantees.
1102 */
call_rcu_tasks_rude(struct rcu_head * rhp,rcu_callback_t func)1103 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1104 {
1105 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1106 }
1107 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1108
1109 /**
1110 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1111 *
1112 * Control will return to the caller some time after a rude rcu-tasks
1113 * grace period has elapsed, in other words after all currently
1114 * executing rcu-tasks read-side critical sections have elapsed. These
1115 * read-side critical sections are delimited by calls to schedule(),
1116 * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1117 * context), and (in theory, anyway) cond_resched().
1118 *
1119 * This is a very specialized primitive, intended only for a few uses in
1120 * tracing and other situations requiring manipulation of function preambles
1121 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
1122 * (yet) intended for heavy use from multiple CPUs.
1123 *
1124 * See the description of synchronize_rcu() for more detailed information
1125 * on memory ordering guarantees.
1126 */
synchronize_rcu_tasks_rude(void)1127 void synchronize_rcu_tasks_rude(void)
1128 {
1129 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1130 }
1131 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1132
1133 /**
1134 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1135 *
1136 * Although the current implementation is guaranteed to wait, it is not
1137 * obligated to, for example, if there are no pending callbacks.
1138 */
rcu_barrier_tasks_rude(void)1139 void rcu_barrier_tasks_rude(void)
1140 {
1141 rcu_barrier_tasks_generic(&rcu_tasks_rude);
1142 }
1143 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1144
rcu_spawn_tasks_rude_kthread(void)1145 static int __init rcu_spawn_tasks_rude_kthread(void)
1146 {
1147 cblist_init_generic(&rcu_tasks_rude);
1148 rcu_tasks_rude.gp_sleep = HZ / 10;
1149 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1150 return 0;
1151 }
1152
1153 #if !defined(CONFIG_TINY_RCU)
show_rcu_tasks_rude_gp_kthread(void)1154 void show_rcu_tasks_rude_gp_kthread(void)
1155 {
1156 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1157 }
1158 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1159 #endif // !defined(CONFIG_TINY_RCU)
1160 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1161
1162 ////////////////////////////////////////////////////////////////////////
1163 //
1164 // Tracing variant of Tasks RCU. This variant is designed to be used
1165 // to protect tracing hooks, including those of BPF. This variant
1166 // therefore:
1167 //
1168 // 1. Has explicit read-side markers to allow finite grace periods
1169 // in the face of in-kernel loops for PREEMPT=n builds.
1170 //
1171 // 2. Protects code in the idle loop, exception entry/exit, and
1172 // CPU-hotplug code paths, similar to the capabilities of SRCU.
1173 //
1174 // 3. Avoids expensive read-side instructions, having overhead similar
1175 // to that of Preemptible RCU.
1176 //
1177 // There are of course downsides. For example, the grace-period code
1178 // can send IPIs to CPUs, even when those CPUs are in the idle loop or
1179 // in nohz_full userspace. If needed, these downsides can be at least
1180 // partially remedied.
1181 //
1182 // Perhaps most important, this variant of RCU does not affect the vanilla
1183 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
1184 // readers can operate from idle, offline, and exception entry/exit in no
1185 // way allows rcu_preempt and rcu_sched readers to also do so.
1186 //
1187 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1188 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
1189 // function sets these function pointers up so that rcu_tasks_wait_gp()
1190 // invokes these functions in this order:
1191 //
1192 // rcu_tasks_trace_pregp_step():
1193 // Disables CPU hotplug, adds all currently executing tasks to the
1194 // holdout list, then checks the state of all tasks that blocked
1195 // or were preempted within their current RCU Tasks Trace read-side
1196 // critical section, adding them to the holdout list if appropriate.
1197 // Finally, this function re-enables CPU hotplug.
1198 // The ->pertask_func() pointer is NULL, so there is no per-task processing.
1199 // rcu_tasks_trace_postscan():
1200 // Invokes synchronize_rcu() to wait for late-stage exiting tasks
1201 // to finish exiting.
1202 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1203 // Scans the holdout list, attempting to identify a quiescent state
1204 // for each task on the list. If there is a quiescent state, the
1205 // corresponding task is removed from the holdout list. Once this
1206 // list is empty, the grace period has completed.
1207 // rcu_tasks_trace_postgp():
1208 // Provides the needed full memory barrier and does debug checks.
1209 //
1210 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1211 //
1212 // Pre-grace-period update-side code is ordered before the grace period
1213 // via the ->cbs_lock and barriers in rcu_tasks_kthread(). Pre-grace-period
1214 // read-side code is ordered before the grace period by atomic operations
1215 // on .b.need_qs flag of each task involved in this process, or by scheduler
1216 // context-switch ordering (for locked-down non-running readers).
1217
1218 // The lockdep state must be outside of #ifdef to be useful.
1219 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1220 static struct lock_class_key rcu_lock_trace_key;
1221 struct lockdep_map rcu_trace_lock_map =
1222 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1223 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1224 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1225
1226 #ifdef CONFIG_TASKS_TRACE_RCU
1227
1228 // Record outstanding IPIs to each CPU. No point in sending two...
1229 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1230
1231 // The number of detections of task quiescent state relying on
1232 // heavyweight readers executing explicit memory barriers.
1233 static unsigned long n_heavy_reader_attempts;
1234 static unsigned long n_heavy_reader_updates;
1235 static unsigned long n_heavy_reader_ofl_updates;
1236 static unsigned long n_trc_holdouts;
1237
1238 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1239 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1240 "RCU Tasks Trace");
1241
1242 /* Load from ->trc_reader_special.b.need_qs with proper ordering. */
rcu_ld_need_qs(struct task_struct * t)1243 static u8 rcu_ld_need_qs(struct task_struct *t)
1244 {
1245 smp_mb(); // Enforce full grace-period ordering.
1246 return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1247 }
1248
1249 /* Store to ->trc_reader_special.b.need_qs with proper ordering. */
rcu_st_need_qs(struct task_struct * t,u8 v)1250 static void rcu_st_need_qs(struct task_struct *t, u8 v)
1251 {
1252 smp_store_release(&t->trc_reader_special.b.need_qs, v);
1253 smp_mb(); // Enforce full grace-period ordering.
1254 }
1255
1256 /*
1257 * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1258 * the four-byte operand-size restriction of some platforms.
1259 * Returns the old value, which is often ignored.
1260 */
rcu_trc_cmpxchg_need_qs(struct task_struct * t,u8 old,u8 new)1261 u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1262 {
1263 union rcu_special ret;
1264 union rcu_special trs_old = READ_ONCE(t->trc_reader_special);
1265 union rcu_special trs_new = trs_old;
1266
1267 if (trs_old.b.need_qs != old)
1268 return trs_old.b.need_qs;
1269 trs_new.b.need_qs = new;
1270 ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s);
1271 return ret.b.need_qs;
1272 }
1273 EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1274
1275 /*
1276 * If we are the last reader, signal the grace-period kthread.
1277 * Also remove from the per-CPU list of blocked tasks.
1278 */
rcu_read_unlock_trace_special(struct task_struct * t)1279 void rcu_read_unlock_trace_special(struct task_struct *t)
1280 {
1281 unsigned long flags;
1282 struct rcu_tasks_percpu *rtpcp;
1283 union rcu_special trs;
1284
1285 // Open-coded full-word version of rcu_ld_need_qs().
1286 smp_mb(); // Enforce full grace-period ordering.
1287 trs = smp_load_acquire(&t->trc_reader_special);
1288
1289 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1290 smp_mb(); // Pairs with update-side barriers.
1291 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1292 if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1293 u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1294 TRC_NEED_QS_CHECKED);
1295
1296 WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1297 }
1298 if (trs.b.blocked) {
1299 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1300 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1301 list_del_init(&t->trc_blkd_node);
1302 WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1303 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1304 }
1305 WRITE_ONCE(t->trc_reader_nesting, 0);
1306 }
1307 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1308
1309 /* Add a newly blocked reader task to its CPU's list. */
rcu_tasks_trace_qs_blkd(struct task_struct * t)1310 void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1311 {
1312 unsigned long flags;
1313 struct rcu_tasks_percpu *rtpcp;
1314
1315 local_irq_save(flags);
1316 rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1317 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1318 t->trc_blkd_cpu = smp_processor_id();
1319 if (!rtpcp->rtp_blkd_tasks.next)
1320 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1321 list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1322 WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1323 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1324 }
1325 EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1326
1327 /* Add a task to the holdout list, if it is not already on the list. */
trc_add_holdout(struct task_struct * t,struct list_head * bhp)1328 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1329 {
1330 if (list_empty(&t->trc_holdout_list)) {
1331 get_task_struct(t);
1332 list_add(&t->trc_holdout_list, bhp);
1333 n_trc_holdouts++;
1334 }
1335 }
1336
1337 /* Remove a task from the holdout list, if it is in fact present. */
trc_del_holdout(struct task_struct * t)1338 static void trc_del_holdout(struct task_struct *t)
1339 {
1340 if (!list_empty(&t->trc_holdout_list)) {
1341 list_del_init(&t->trc_holdout_list);
1342 put_task_struct(t);
1343 n_trc_holdouts--;
1344 }
1345 }
1346
1347 /* IPI handler to check task state. */
trc_read_check_handler(void * t_in)1348 static void trc_read_check_handler(void *t_in)
1349 {
1350 int nesting;
1351 struct task_struct *t = current;
1352 struct task_struct *texp = t_in;
1353
1354 // If the task is no longer running on this CPU, leave.
1355 if (unlikely(texp != t))
1356 goto reset_ipi; // Already on holdout list, so will check later.
1357
1358 // If the task is not in a read-side critical section, and
1359 // if this is the last reader, awaken the grace-period kthread.
1360 nesting = READ_ONCE(t->trc_reader_nesting);
1361 if (likely(!nesting)) {
1362 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1363 goto reset_ipi;
1364 }
1365 // If we are racing with an rcu_read_unlock_trace(), try again later.
1366 if (unlikely(nesting < 0))
1367 goto reset_ipi;
1368
1369 // Get here if the task is in a read-side critical section.
1370 // Set its state so that it will update state for the grace-period
1371 // kthread upon exit from that critical section.
1372 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1373
1374 reset_ipi:
1375 // Allow future IPIs to be sent on CPU and for task.
1376 // Also order this IPI handler against any later manipulations of
1377 // the intended task.
1378 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1379 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1380 }
1381
1382 /* Callback function for scheduler to check locked-down task. */
trc_inspect_reader(struct task_struct * t,void * bhp_in)1383 static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1384 {
1385 struct list_head *bhp = bhp_in;
1386 int cpu = task_cpu(t);
1387 int nesting;
1388 bool ofl = cpu_is_offline(cpu);
1389
1390 if (task_curr(t) && !ofl) {
1391 // If no chance of heavyweight readers, do it the hard way.
1392 if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1393 return -EINVAL;
1394
1395 // If heavyweight readers are enabled on the remote task,
1396 // we can inspect its state despite its currently running.
1397 // However, we cannot safely change its state.
1398 n_heavy_reader_attempts++;
1399 // Check for "running" idle tasks on offline CPUs.
1400 if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1401 return -EINVAL; // No quiescent state, do it the hard way.
1402 n_heavy_reader_updates++;
1403 nesting = 0;
1404 } else {
1405 // The task is not running, so C-language access is safe.
1406 nesting = t->trc_reader_nesting;
1407 WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
1408 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1409 n_heavy_reader_ofl_updates++;
1410 }
1411
1412 // If not exiting a read-side critical section, mark as checked
1413 // so that the grace-period kthread will remove it from the
1414 // holdout list.
1415 if (!nesting) {
1416 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1417 return 0; // In QS, so done.
1418 }
1419 if (nesting < 0)
1420 return -EINVAL; // Reader transitioning, try again later.
1421
1422 // The task is in a read-side critical section, so set up its
1423 // state so that it will update state upon exit from that critical
1424 // section.
1425 if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1426 trc_add_holdout(t, bhp);
1427 return 0;
1428 }
1429
1430 /* Attempt to extract the state for the specified task. */
trc_wait_for_one_reader(struct task_struct * t,struct list_head * bhp)1431 static void trc_wait_for_one_reader(struct task_struct *t,
1432 struct list_head *bhp)
1433 {
1434 int cpu;
1435
1436 // If a previous IPI is still in flight, let it complete.
1437 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1438 return;
1439
1440 // The current task had better be in a quiescent state.
1441 if (t == current) {
1442 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1443 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1444 return;
1445 }
1446
1447 // Attempt to nail down the task for inspection.
1448 get_task_struct(t);
1449 if (!task_call_func(t, trc_inspect_reader, bhp)) {
1450 put_task_struct(t);
1451 return;
1452 }
1453 put_task_struct(t);
1454
1455 // If this task is not yet on the holdout list, then we are in
1456 // an RCU read-side critical section. Otherwise, the invocation of
1457 // trc_add_holdout() that added it to the list did the necessary
1458 // get_task_struct(). Either way, the task cannot be freed out
1459 // from under this code.
1460
1461 // If currently running, send an IPI, either way, add to list.
1462 trc_add_holdout(t, bhp);
1463 if (task_curr(t) &&
1464 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1465 // The task is currently running, so try IPIing it.
1466 cpu = task_cpu(t);
1467
1468 // If there is already an IPI outstanding, let it happen.
1469 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1470 return;
1471
1472 per_cpu(trc_ipi_to_cpu, cpu) = true;
1473 t->trc_ipi_to_cpu = cpu;
1474 rcu_tasks_trace.n_ipis++;
1475 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1476 // Just in case there is some other reason for
1477 // failure than the target CPU being offline.
1478 WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n",
1479 __func__, cpu);
1480 rcu_tasks_trace.n_ipis_fails++;
1481 per_cpu(trc_ipi_to_cpu, cpu) = false;
1482 t->trc_ipi_to_cpu = -1;
1483 }
1484 }
1485 }
1486
1487 /*
1488 * Initialize for first-round processing for the specified task.
1489 * Return false if task is NULL or already taken care of, true otherwise.
1490 */
rcu_tasks_trace_pertask_prep(struct task_struct * t,bool notself)1491 static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1492 {
1493 // During early boot when there is only the one boot CPU, there
1494 // is no idle task for the other CPUs. Also, the grace-period
1495 // kthread is always in a quiescent state. In addition, just return
1496 // if this task is already on the list.
1497 if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1498 return false;
1499
1500 rcu_st_need_qs(t, 0);
1501 t->trc_ipi_to_cpu = -1;
1502 return true;
1503 }
1504
1505 /* Do first-round processing for the specified task. */
rcu_tasks_trace_pertask(struct task_struct * t,struct list_head * hop)1506 static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1507 {
1508 if (rcu_tasks_trace_pertask_prep(t, true))
1509 trc_wait_for_one_reader(t, hop);
1510 }
1511
1512 /* Initialize for a new RCU-tasks-trace grace period. */
rcu_tasks_trace_pregp_step(struct list_head * hop)1513 static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1514 {
1515 LIST_HEAD(blkd_tasks);
1516 int cpu;
1517 unsigned long flags;
1518 struct rcu_tasks_percpu *rtpcp;
1519 struct task_struct *t;
1520
1521 // There shouldn't be any old IPIs, but...
1522 for_each_possible_cpu(cpu)
1523 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1524
1525 // Disable CPU hotplug across the CPU scan for the benefit of
1526 // any IPIs that might be needed. This also waits for all readers
1527 // in CPU-hotplug code paths.
1528 cpus_read_lock();
1529
1530 // These rcu_tasks_trace_pertask_prep() calls are serialized to
1531 // allow safe access to the hop list.
1532 for_each_online_cpu(cpu) {
1533 rcu_read_lock();
1534 t = cpu_curr_snapshot(cpu);
1535 if (rcu_tasks_trace_pertask_prep(t, true))
1536 trc_add_holdout(t, hop);
1537 rcu_read_unlock();
1538 cond_resched_tasks_rcu_qs();
1539 }
1540
1541 // Only after all running tasks have been accounted for is it
1542 // safe to take care of the tasks that have blocked within their
1543 // current RCU tasks trace read-side critical section.
1544 for_each_possible_cpu(cpu) {
1545 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1546 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1547 list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1548 while (!list_empty(&blkd_tasks)) {
1549 rcu_read_lock();
1550 t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1551 list_del_init(&t->trc_blkd_node);
1552 list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1553 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1554 rcu_tasks_trace_pertask(t, hop);
1555 rcu_read_unlock();
1556 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1557 }
1558 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1559 cond_resched_tasks_rcu_qs();
1560 }
1561
1562 // Re-enable CPU hotplug now that the holdout list is populated.
1563 cpus_read_unlock();
1564 }
1565
1566 /*
1567 * Do intermediate processing between task and holdout scans.
1568 */
rcu_tasks_trace_postscan(struct list_head * hop)1569 static void rcu_tasks_trace_postscan(struct list_head *hop)
1570 {
1571 // Wait for late-stage exiting tasks to finish exiting.
1572 // These might have passed the call to exit_tasks_rcu_finish().
1573
1574 // If you remove the following line, update rcu_trace_implies_rcu_gp()!!!
1575 synchronize_rcu();
1576 // Any tasks that exit after this point will set
1577 // TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1578 }
1579
1580 /* Communicate task state back to the RCU tasks trace stall warning request. */
1581 struct trc_stall_chk_rdr {
1582 int nesting;
1583 int ipi_to_cpu;
1584 u8 needqs;
1585 };
1586
trc_check_slow_task(struct task_struct * t,void * arg)1587 static int trc_check_slow_task(struct task_struct *t, void *arg)
1588 {
1589 struct trc_stall_chk_rdr *trc_rdrp = arg;
1590
1591 if (task_curr(t) && cpu_online(task_cpu(t)))
1592 return false; // It is running, so decline to inspect it.
1593 trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1594 trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1595 trc_rdrp->needqs = rcu_ld_need_qs(t);
1596 return true;
1597 }
1598
1599 /* Show the state of a task stalling the current RCU tasks trace GP. */
show_stalled_task_trace(struct task_struct * t,bool * firstreport)1600 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1601 {
1602 int cpu;
1603 struct trc_stall_chk_rdr trc_rdr;
1604 bool is_idle_tsk = is_idle_task(t);
1605
1606 if (*firstreport) {
1607 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1608 *firstreport = false;
1609 }
1610 cpu = task_cpu(t);
1611 if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1612 pr_alert("P%d: %c%c\n",
1613 t->pid,
1614 ".I"[t->trc_ipi_to_cpu >= 0],
1615 ".i"[is_idle_tsk]);
1616 else
1617 pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1618 t->pid,
1619 ".I"[trc_rdr.ipi_to_cpu >= 0],
1620 ".i"[is_idle_tsk],
1621 ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1622 ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1623 trc_rdr.nesting,
1624 " !CN"[trc_rdr.needqs & 0x3],
1625 " ?"[trc_rdr.needqs > 0x3],
1626 cpu, cpu_online(cpu) ? "" : "(offline)");
1627 sched_show_task(t);
1628 }
1629
1630 /* List stalled IPIs for RCU tasks trace. */
show_stalled_ipi_trace(void)1631 static void show_stalled_ipi_trace(void)
1632 {
1633 int cpu;
1634
1635 for_each_possible_cpu(cpu)
1636 if (per_cpu(trc_ipi_to_cpu, cpu))
1637 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1638 }
1639
1640 /* Do one scan of the holdout list. */
check_all_holdout_tasks_trace(struct list_head * hop,bool needreport,bool * firstreport)1641 static void check_all_holdout_tasks_trace(struct list_head *hop,
1642 bool needreport, bool *firstreport)
1643 {
1644 struct task_struct *g, *t;
1645
1646 // Disable CPU hotplug across the holdout list scan for IPIs.
1647 cpus_read_lock();
1648
1649 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1650 // If safe and needed, try to check the current task.
1651 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1652 !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1653 trc_wait_for_one_reader(t, hop);
1654
1655 // If check succeeded, remove this task from the list.
1656 if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1657 rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1658 trc_del_holdout(t);
1659 else if (needreport)
1660 show_stalled_task_trace(t, firstreport);
1661 cond_resched_tasks_rcu_qs();
1662 }
1663
1664 // Re-enable CPU hotplug now that the holdout list scan has completed.
1665 cpus_read_unlock();
1666
1667 if (needreport) {
1668 if (*firstreport)
1669 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1670 show_stalled_ipi_trace();
1671 }
1672 }
1673
rcu_tasks_trace_empty_fn(void * unused)1674 static void rcu_tasks_trace_empty_fn(void *unused)
1675 {
1676 }
1677
1678 /* Wait for grace period to complete and provide ordering. */
rcu_tasks_trace_postgp(struct rcu_tasks * rtp)1679 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1680 {
1681 int cpu;
1682
1683 // Wait for any lingering IPI handlers to complete. Note that
1684 // if a CPU has gone offline or transitioned to userspace in the
1685 // meantime, all IPI handlers should have been drained beforehand.
1686 // Yes, this assumes that CPUs process IPIs in order. If that ever
1687 // changes, there will need to be a recheck and/or timed wait.
1688 for_each_online_cpu(cpu)
1689 if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1690 smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1691
1692 smp_mb(); // Caller's code must be ordered after wakeup.
1693 // Pairs with pretty much every ordering primitive.
1694 }
1695
1696 /* Report any needed quiescent state for this exiting task. */
exit_tasks_rcu_finish_trace(struct task_struct * t)1697 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1698 {
1699 union rcu_special trs = READ_ONCE(t->trc_reader_special);
1700
1701 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1702 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1703 if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1704 rcu_read_unlock_trace_special(t);
1705 else
1706 WRITE_ONCE(t->trc_reader_nesting, 0);
1707 }
1708
1709 /**
1710 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1711 * @rhp: structure to be used for queueing the RCU updates.
1712 * @func: actual callback function to be invoked after the grace period
1713 *
1714 * The callback function will be invoked some time after a trace rcu-tasks
1715 * grace period elapses, in other words after all currently executing
1716 * trace rcu-tasks read-side critical sections have completed. These
1717 * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1718 * and rcu_read_unlock_trace().
1719 *
1720 * See the description of call_rcu() for more detailed information on
1721 * memory ordering guarantees.
1722 */
call_rcu_tasks_trace(struct rcu_head * rhp,rcu_callback_t func)1723 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1724 {
1725 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1726 }
1727 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1728
1729 /**
1730 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1731 *
1732 * Control will return to the caller some time after a trace rcu-tasks
1733 * grace period has elapsed, in other words after all currently executing
1734 * trace rcu-tasks read-side critical sections have elapsed. These read-side
1735 * critical sections are delimited by calls to rcu_read_lock_trace()
1736 * and rcu_read_unlock_trace().
1737 *
1738 * This is a very specialized primitive, intended only for a few uses in
1739 * tracing and other situations requiring manipulation of function preambles
1740 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1741 * (yet) intended for heavy use from multiple CPUs.
1742 *
1743 * See the description of synchronize_rcu() for more detailed information
1744 * on memory ordering guarantees.
1745 */
synchronize_rcu_tasks_trace(void)1746 void synchronize_rcu_tasks_trace(void)
1747 {
1748 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1749 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1750 }
1751 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1752
1753 /**
1754 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1755 *
1756 * Although the current implementation is guaranteed to wait, it is not
1757 * obligated to, for example, if there are no pending callbacks.
1758 */
rcu_barrier_tasks_trace(void)1759 void rcu_barrier_tasks_trace(void)
1760 {
1761 rcu_barrier_tasks_generic(&rcu_tasks_trace);
1762 }
1763 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1764
rcu_spawn_tasks_trace_kthread(void)1765 static int __init rcu_spawn_tasks_trace_kthread(void)
1766 {
1767 cblist_init_generic(&rcu_tasks_trace);
1768 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1769 rcu_tasks_trace.gp_sleep = HZ / 10;
1770 rcu_tasks_trace.init_fract = HZ / 10;
1771 } else {
1772 rcu_tasks_trace.gp_sleep = HZ / 200;
1773 if (rcu_tasks_trace.gp_sleep <= 0)
1774 rcu_tasks_trace.gp_sleep = 1;
1775 rcu_tasks_trace.init_fract = HZ / 200;
1776 if (rcu_tasks_trace.init_fract <= 0)
1777 rcu_tasks_trace.init_fract = 1;
1778 }
1779 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1780 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1781 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1782 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1783 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1784 return 0;
1785 }
1786
1787 #if !defined(CONFIG_TINY_RCU)
show_rcu_tasks_trace_gp_kthread(void)1788 void show_rcu_tasks_trace_gp_kthread(void)
1789 {
1790 char buf[64];
1791
1792 sprintf(buf, "N%lu h:%lu/%lu/%lu",
1793 data_race(n_trc_holdouts),
1794 data_race(n_heavy_reader_ofl_updates),
1795 data_race(n_heavy_reader_updates),
1796 data_race(n_heavy_reader_attempts));
1797 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1798 }
1799 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1800 #endif // !defined(CONFIG_TINY_RCU)
1801
1802 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
exit_tasks_rcu_finish_trace(struct task_struct * t)1803 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1804 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1805
1806 #ifndef CONFIG_TINY_RCU
show_rcu_tasks_gp_kthreads(void)1807 void show_rcu_tasks_gp_kthreads(void)
1808 {
1809 show_rcu_tasks_classic_gp_kthread();
1810 show_rcu_tasks_rude_gp_kthread();
1811 show_rcu_tasks_trace_gp_kthread();
1812 }
1813 #endif /* #ifndef CONFIG_TINY_RCU */
1814
1815 #ifdef CONFIG_PROVE_RCU
1816 struct rcu_tasks_test_desc {
1817 struct rcu_head rh;
1818 const char *name;
1819 bool notrun;
1820 unsigned long runstart;
1821 };
1822
1823 static struct rcu_tasks_test_desc tests[] = {
1824 {
1825 .name = "call_rcu_tasks()",
1826 /* If not defined, the test is skipped. */
1827 .notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1828 },
1829 {
1830 .name = "call_rcu_tasks_rude()",
1831 /* If not defined, the test is skipped. */
1832 .notrun = IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1833 },
1834 {
1835 .name = "call_rcu_tasks_trace()",
1836 /* If not defined, the test is skipped. */
1837 .notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1838 }
1839 };
1840
test_rcu_tasks_callback(struct rcu_head * rhp)1841 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1842 {
1843 struct rcu_tasks_test_desc *rttd =
1844 container_of(rhp, struct rcu_tasks_test_desc, rh);
1845
1846 pr_info("Callback from %s invoked.\n", rttd->name);
1847
1848 rttd->notrun = false;
1849 }
1850
rcu_tasks_initiate_self_tests(void)1851 static void rcu_tasks_initiate_self_tests(void)
1852 {
1853 unsigned long j = jiffies;
1854
1855 pr_info("Running RCU-tasks wait API self tests\n");
1856 #ifdef CONFIG_TASKS_RCU
1857 tests[0].runstart = j;
1858 synchronize_rcu_tasks();
1859 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1860 #endif
1861
1862 #ifdef CONFIG_TASKS_RUDE_RCU
1863 tests[1].runstart = j;
1864 synchronize_rcu_tasks_rude();
1865 call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1866 #endif
1867
1868 #ifdef CONFIG_TASKS_TRACE_RCU
1869 tests[2].runstart = j;
1870 synchronize_rcu_tasks_trace();
1871 call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1872 #endif
1873 }
1874
1875 /*
1876 * Return: 0 - test passed
1877 * 1 - test failed, but have not timed out yet
1878 * -1 - test failed and timed out
1879 */
rcu_tasks_verify_self_tests(void)1880 static int rcu_tasks_verify_self_tests(void)
1881 {
1882 int ret = 0;
1883 int i;
1884 unsigned long bst = rcu_task_stall_timeout;
1885
1886 if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
1887 bst = RCU_TASK_BOOT_STALL_TIMEOUT;
1888 for (i = 0; i < ARRAY_SIZE(tests); i++) {
1889 while (tests[i].notrun) { // still hanging.
1890 if (time_after(jiffies, tests[i].runstart + bst)) {
1891 pr_err("%s has failed boot-time tests.\n", tests[i].name);
1892 ret = -1;
1893 break;
1894 }
1895 ret = 1;
1896 break;
1897 }
1898 }
1899 WARN_ON(ret < 0);
1900
1901 return ret;
1902 }
1903
1904 /*
1905 * Repeat the rcu_tasks_verify_self_tests() call once every second until the
1906 * test passes or has timed out.
1907 */
1908 static struct delayed_work rcu_tasks_verify_work;
rcu_tasks_verify_work_fn(struct work_struct * work __maybe_unused)1909 static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
1910 {
1911 int ret = rcu_tasks_verify_self_tests();
1912
1913 if (ret <= 0)
1914 return;
1915
1916 /* Test fails but not timed out yet, reschedule another check */
1917 schedule_delayed_work(&rcu_tasks_verify_work, HZ);
1918 }
1919
rcu_tasks_verify_schedule_work(void)1920 static int rcu_tasks_verify_schedule_work(void)
1921 {
1922 INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
1923 rcu_tasks_verify_work_fn(NULL);
1924 return 0;
1925 }
1926 late_initcall(rcu_tasks_verify_schedule_work);
1927 #else /* #ifdef CONFIG_PROVE_RCU */
rcu_tasks_initiate_self_tests(void)1928 static void rcu_tasks_initiate_self_tests(void) { }
1929 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1930
rcu_init_tasks_generic(void)1931 void __init rcu_init_tasks_generic(void)
1932 {
1933 #ifdef CONFIG_TASKS_RCU
1934 rcu_spawn_tasks_kthread();
1935 #endif
1936
1937 #ifdef CONFIG_TASKS_RUDE_RCU
1938 rcu_spawn_tasks_rude_kthread();
1939 #endif
1940
1941 #ifdef CONFIG_TASKS_TRACE_RCU
1942 rcu_spawn_tasks_trace_kthread();
1943 #endif
1944
1945 // Run the self-tests.
1946 rcu_tasks_initiate_self_tests();
1947 }
1948
1949 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
rcu_tasks_bootup_oddness(void)1950 static inline void rcu_tasks_bootup_oddness(void) {}
1951 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */
1952