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