1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 *
5 * Copyright IBM Corporation, 2008
6 *
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
9 * Paul E. McKenney <paulmck@linux.ibm.com>
10 *
11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
13 *
14 * For detailed explanation of Read-Copy Update mechanism see -
15 * Documentation/RCU
16 */
17
18 #define pr_fmt(fmt) "rcu: " fmt
19
20 #include <linux/types.h>
21 #include <linux/kernel.h>
22 #include <linux/init.h>
23 #include <linux/spinlock.h>
24 #include <linux/smp.h>
25 #include <linux/rcupdate_wait.h>
26 #include <linux/interrupt.h>
27 #include <linux/sched.h>
28 #include <linux/sched/debug.h>
29 #include <linux/nmi.h>
30 #include <linux/atomic.h>
31 #include <linux/bitops.h>
32 #include <linux/export.h>
33 #include <linux/completion.h>
34 #include <linux/kmemleak.h>
35 #include <linux/moduleparam.h>
36 #include <linux/panic.h>
37 #include <linux/panic_notifier.h>
38 #include <linux/percpu.h>
39 #include <linux/notifier.h>
40 #include <linux/cpu.h>
41 #include <linux/mutex.h>
42 #include <linux/time.h>
43 #include <linux/kernel_stat.h>
44 #include <linux/wait.h>
45 #include <linux/kthread.h>
46 #include <uapi/linux/sched/types.h>
47 #include <linux/prefetch.h>
48 #include <linux/delay.h>
49 #include <linux/random.h>
50 #include <linux/trace_events.h>
51 #include <linux/suspend.h>
52 #include <linux/ftrace.h>
53 #include <linux/tick.h>
54 #include <linux/sysrq.h>
55 #include <linux/kprobes.h>
56 #include <linux/gfp.h>
57 #include <linux/oom.h>
58 #include <linux/smpboot.h>
59 #include <linux/jiffies.h>
60 #include <linux/slab.h>
61 #include <linux/sched/isolation.h>
62 #include <linux/sched/clock.h>
63 #include <linux/vmalloc.h>
64 #include <linux/mm.h>
65 #include <linux/kasan.h>
66 #include <linux/context_tracking.h>
67 #include "../time/tick-internal.h"
68
69 #include "tree.h"
70 #include "rcu.h"
71
72 #ifdef MODULE_PARAM_PREFIX
73 #undef MODULE_PARAM_PREFIX
74 #endif
75 #define MODULE_PARAM_PREFIX "rcutree."
76
77 /* Data structures. */
78 static void rcu_sr_normal_gp_cleanup_work(struct work_struct *);
79
80 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
81 .gpwrap = true,
82 };
83 static struct rcu_state rcu_state = {
84 .level = { &rcu_state.node[0] },
85 .gp_state = RCU_GP_IDLE,
86 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
87 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
88 .barrier_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.barrier_lock),
89 .name = RCU_NAME,
90 .abbr = RCU_ABBR,
91 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
92 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
93 .ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
94 .srs_cleanup_work = __WORK_INITIALIZER(rcu_state.srs_cleanup_work,
95 rcu_sr_normal_gp_cleanup_work),
96 .srs_cleanups_pending = ATOMIC_INIT(0),
97 #ifdef CONFIG_RCU_NOCB_CPU
98 .nocb_mutex = __MUTEX_INITIALIZER(rcu_state.nocb_mutex),
99 #endif
100 };
101
102 /* Dump rcu_node combining tree at boot to verify correct setup. */
103 static bool dump_tree;
104 module_param(dump_tree, bool, 0444);
105 /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
106 static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
107 #ifndef CONFIG_PREEMPT_RT
108 module_param(use_softirq, bool, 0444);
109 #endif
110 /* Control rcu_node-tree auto-balancing at boot time. */
111 static bool rcu_fanout_exact;
112 module_param(rcu_fanout_exact, bool, 0444);
113 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
114 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
115 module_param(rcu_fanout_leaf, int, 0444);
116 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
117 /* Number of rcu_nodes at specified level. */
118 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
119 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
120
121 /*
122 * The rcu_scheduler_active variable is initialized to the value
123 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
124 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
125 * RCU can assume that there is but one task, allowing RCU to (for example)
126 * optimize synchronize_rcu() to a simple barrier(). When this variable
127 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
128 * to detect real grace periods. This variable is also used to suppress
129 * boot-time false positives from lockdep-RCU error checking. Finally, it
130 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
131 * is fully initialized, including all of its kthreads having been spawned.
132 */
133 int rcu_scheduler_active __read_mostly;
134 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
135
136 /*
137 * The rcu_scheduler_fully_active variable transitions from zero to one
138 * during the early_initcall() processing, which is after the scheduler
139 * is capable of creating new tasks. So RCU processing (for example,
140 * creating tasks for RCU priority boosting) must be delayed until after
141 * rcu_scheduler_fully_active transitions from zero to one. We also
142 * currently delay invocation of any RCU callbacks until after this point.
143 *
144 * It might later prove better for people registering RCU callbacks during
145 * early boot to take responsibility for these callbacks, but one step at
146 * a time.
147 */
148 static int rcu_scheduler_fully_active __read_mostly;
149
150 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
151 unsigned long gps, unsigned long flags);
152 static struct task_struct *rcu_boost_task(struct rcu_node *rnp);
153 static void invoke_rcu_core(void);
154 static void rcu_report_exp_rdp(struct rcu_data *rdp);
155 static void sync_sched_exp_online_cleanup(int cpu);
156 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
157 static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
158 static bool rcu_rdp_cpu_online(struct rcu_data *rdp);
159 static bool rcu_init_invoked(void);
160 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
161 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
162
163 /*
164 * rcuc/rcub/rcuop kthread realtime priority. The "rcuop"
165 * real-time priority(enabling/disabling) is controlled by
166 * the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration.
167 */
168 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
169 module_param(kthread_prio, int, 0444);
170
171 /* Delay in jiffies for grace-period initialization delays, debug only. */
172
173 static int gp_preinit_delay;
174 module_param(gp_preinit_delay, int, 0444);
175 static int gp_init_delay;
176 module_param(gp_init_delay, int, 0444);
177 static int gp_cleanup_delay;
178 module_param(gp_cleanup_delay, int, 0444);
179 static int nohz_full_patience_delay;
180 module_param(nohz_full_patience_delay, int, 0444);
181 static int nohz_full_patience_delay_jiffies;
182
183 // Add delay to rcu_read_unlock() for strict grace periods.
184 static int rcu_unlock_delay;
185 #ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
186 module_param(rcu_unlock_delay, int, 0444);
187 #endif
188
189 /*
190 * This rcu parameter is runtime-read-only. It reflects
191 * a minimum allowed number of objects which can be cached
192 * per-CPU. Object size is equal to one page. This value
193 * can be changed at boot time.
194 */
195 static int rcu_min_cached_objs = 5;
196 module_param(rcu_min_cached_objs, int, 0444);
197
198 // A page shrinker can ask for pages to be freed to make them
199 // available for other parts of the system. This usually happens
200 // under low memory conditions, and in that case we should also
201 // defer page-cache filling for a short time period.
202 //
203 // The default value is 5 seconds, which is long enough to reduce
204 // interference with the shrinker while it asks other systems to
205 // drain their caches.
206 static int rcu_delay_page_cache_fill_msec = 5000;
207 module_param(rcu_delay_page_cache_fill_msec, int, 0444);
208
209 /* Retrieve RCU kthreads priority for rcutorture */
rcu_get_gp_kthreads_prio(void)210 int rcu_get_gp_kthreads_prio(void)
211 {
212 return kthread_prio;
213 }
214 EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
215
216 /*
217 * Number of grace periods between delays, normalized by the duration of
218 * the delay. The longer the delay, the more the grace periods between
219 * each delay. The reason for this normalization is that it means that,
220 * for non-zero delays, the overall slowdown of grace periods is constant
221 * regardless of the duration of the delay. This arrangement balances
222 * the need for long delays to increase some race probabilities with the
223 * need for fast grace periods to increase other race probabilities.
224 */
225 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */
226
227 /*
228 * Return true if an RCU grace period is in progress. The READ_ONCE()s
229 * permit this function to be invoked without holding the root rcu_node
230 * structure's ->lock, but of course results can be subject to change.
231 */
rcu_gp_in_progress(void)232 static int rcu_gp_in_progress(void)
233 {
234 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
235 }
236
237 /*
238 * Return the number of callbacks queued on the specified CPU.
239 * Handles both the nocbs and normal cases.
240 */
rcu_get_n_cbs_cpu(int cpu)241 static long rcu_get_n_cbs_cpu(int cpu)
242 {
243 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
244
245 if (rcu_segcblist_is_enabled(&rdp->cblist))
246 return rcu_segcblist_n_cbs(&rdp->cblist);
247 return 0;
248 }
249
250 /**
251 * rcu_softirq_qs - Provide a set of RCU quiescent states in softirq processing
252 *
253 * Mark a quiescent state for RCU, Tasks RCU, and Tasks Trace RCU.
254 * This is a special-purpose function to be used in the softirq
255 * infrastructure and perhaps the occasional long-running softirq
256 * handler.
257 *
258 * Note that from RCU's viewpoint, a call to rcu_softirq_qs() is
259 * equivalent to momentarily completely enabling preemption. For
260 * example, given this code::
261 *
262 * local_bh_disable();
263 * do_something();
264 * rcu_softirq_qs(); // A
265 * do_something_else();
266 * local_bh_enable(); // B
267 *
268 * A call to synchronize_rcu() that began concurrently with the
269 * call to do_something() would be guaranteed to wait only until
270 * execution reached statement A. Without that rcu_softirq_qs(),
271 * that same synchronize_rcu() would instead be guaranteed to wait
272 * until execution reached statement B.
273 */
rcu_softirq_qs(void)274 void rcu_softirq_qs(void)
275 {
276 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
277 lock_is_held(&rcu_lock_map) ||
278 lock_is_held(&rcu_sched_lock_map),
279 "Illegal rcu_softirq_qs() in RCU read-side critical section");
280 rcu_qs();
281 rcu_preempt_deferred_qs(current);
282 rcu_tasks_qs(current, false);
283 }
284
285 /*
286 * Reset the current CPU's RCU_WATCHING counter to indicate that the
287 * newly onlined CPU is no longer in an extended quiescent state.
288 * This will either leave the counter unchanged, or increment it
289 * to the next non-quiescent value.
290 *
291 * The non-atomic test/increment sequence works because the upper bits
292 * of the ->state variable are manipulated only by the corresponding CPU,
293 * or when the corresponding CPU is offline.
294 */
rcu_watching_online(void)295 static void rcu_watching_online(void)
296 {
297 if (ct_rcu_watching() & CT_RCU_WATCHING)
298 return;
299 ct_state_inc(CT_RCU_WATCHING);
300 }
301
302 /*
303 * Return true if the snapshot returned from ct_rcu_watching()
304 * indicates that RCU is in an extended quiescent state.
305 */
rcu_watching_snap_in_eqs(int snap)306 static bool rcu_watching_snap_in_eqs(int snap)
307 {
308 return !(snap & CT_RCU_WATCHING);
309 }
310
311 /**
312 * rcu_watching_snap_stopped_since() - Has RCU stopped watching a given CPU
313 * since the specified @snap?
314 *
315 * @rdp: The rcu_data corresponding to the CPU for which to check EQS.
316 * @snap: rcu_watching snapshot taken when the CPU wasn't in an EQS.
317 *
318 * Returns true if the CPU corresponding to @rdp has spent some time in an
319 * extended quiescent state since @snap. Note that this doesn't check if it
320 * /still/ is in an EQS, just that it went through one since @snap.
321 *
322 * This is meant to be used in a loop waiting for a CPU to go through an EQS.
323 */
rcu_watching_snap_stopped_since(struct rcu_data * rdp,int snap)324 static bool rcu_watching_snap_stopped_since(struct rcu_data *rdp, int snap)
325 {
326 /*
327 * The first failing snapshot is already ordered against the accesses
328 * performed by the remote CPU after it exits idle.
329 *
330 * The second snapshot therefore only needs to order against accesses
331 * performed by the remote CPU prior to entering idle and therefore can
332 * rely solely on acquire semantics.
333 */
334 if (WARN_ON_ONCE(rcu_watching_snap_in_eqs(snap)))
335 return true;
336
337 return snap != ct_rcu_watching_cpu_acquire(rdp->cpu);
338 }
339
340 /*
341 * Return true if the referenced integer is zero while the specified
342 * CPU remains within a single extended quiescent state.
343 */
rcu_watching_zero_in_eqs(int cpu,int * vp)344 bool rcu_watching_zero_in_eqs(int cpu, int *vp)
345 {
346 int snap;
347
348 // If not quiescent, force back to earlier extended quiescent state.
349 snap = ct_rcu_watching_cpu(cpu) & ~CT_RCU_WATCHING;
350 smp_rmb(); // Order CT state and *vp reads.
351 if (READ_ONCE(*vp))
352 return false; // Non-zero, so report failure;
353 smp_rmb(); // Order *vp read and CT state re-read.
354
355 // If still in the same extended quiescent state, we are good!
356 return snap == ct_rcu_watching_cpu(cpu);
357 }
358
359 /*
360 * Let the RCU core know that this CPU has gone through the scheduler,
361 * which is a quiescent state. This is called when the need for a
362 * quiescent state is urgent, so we burn an atomic operation and full
363 * memory barriers to let the RCU core know about it, regardless of what
364 * this CPU might (or might not) do in the near future.
365 *
366 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
367 *
368 * The caller must have disabled interrupts and must not be idle.
369 */
rcu_momentary_eqs(void)370 notrace void rcu_momentary_eqs(void)
371 {
372 int seq;
373
374 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
375 seq = ct_state_inc(2 * CT_RCU_WATCHING);
376 /* It is illegal to call this from idle state. */
377 WARN_ON_ONCE(!(seq & CT_RCU_WATCHING));
378 rcu_preempt_deferred_qs(current);
379 }
380 EXPORT_SYMBOL_GPL(rcu_momentary_eqs);
381
382 /**
383 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
384 *
385 * If the current CPU is idle and running at a first-level (not nested)
386 * interrupt, or directly, from idle, return true.
387 *
388 * The caller must have at least disabled IRQs.
389 */
rcu_is_cpu_rrupt_from_idle(void)390 static int rcu_is_cpu_rrupt_from_idle(void)
391 {
392 long nesting;
393
394 /*
395 * Usually called from the tick; but also used from smp_function_call()
396 * for expedited grace periods. This latter can result in running from
397 * the idle task, instead of an actual IPI.
398 */
399 lockdep_assert_irqs_disabled();
400
401 /* Check for counter underflows */
402 RCU_LOCKDEP_WARN(ct_nesting() < 0,
403 "RCU nesting counter underflow!");
404 RCU_LOCKDEP_WARN(ct_nmi_nesting() <= 0,
405 "RCU nmi_nesting counter underflow/zero!");
406
407 /* Are we at first interrupt nesting level? */
408 nesting = ct_nmi_nesting();
409 if (nesting > 1)
410 return false;
411
412 /*
413 * If we're not in an interrupt, we must be in the idle task!
414 */
415 WARN_ON_ONCE(!nesting && !is_idle_task(current));
416
417 /* Does CPU appear to be idle from an RCU standpoint? */
418 return ct_nesting() == 0;
419 }
420
421 #define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
422 // Maximum callbacks per rcu_do_batch ...
423 #define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
424 static long blimit = DEFAULT_RCU_BLIMIT;
425 #define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
426 static long qhimark = DEFAULT_RCU_QHIMARK;
427 #define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
428 static long qlowmark = DEFAULT_RCU_QLOMARK;
429 #define DEFAULT_RCU_QOVLD_MULT 2
430 #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
431 static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
432 static long qovld_calc = -1; // No pre-initialization lock acquisitions!
433
434 module_param(blimit, long, 0444);
435 module_param(qhimark, long, 0444);
436 module_param(qlowmark, long, 0444);
437 module_param(qovld, long, 0444);
438
439 static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
440 static ulong jiffies_till_next_fqs = ULONG_MAX;
441 static bool rcu_kick_kthreads;
442 static int rcu_divisor = 7;
443 module_param(rcu_divisor, int, 0644);
444
445 /* Force an exit from rcu_do_batch() after 3 milliseconds. */
446 static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
447 module_param(rcu_resched_ns, long, 0644);
448
449 /*
450 * How long the grace period must be before we start recruiting
451 * quiescent-state help from rcu_note_context_switch().
452 */
453 static ulong jiffies_till_sched_qs = ULONG_MAX;
454 module_param(jiffies_till_sched_qs, ulong, 0444);
455 static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
456 module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
457
458 /*
459 * Make sure that we give the grace-period kthread time to detect any
460 * idle CPUs before taking active measures to force quiescent states.
461 * However, don't go below 100 milliseconds, adjusted upwards for really
462 * large systems.
463 */
adjust_jiffies_till_sched_qs(void)464 static void adjust_jiffies_till_sched_qs(void)
465 {
466 unsigned long j;
467
468 /* If jiffies_till_sched_qs was specified, respect the request. */
469 if (jiffies_till_sched_qs != ULONG_MAX) {
470 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
471 return;
472 }
473 /* Otherwise, set to third fqs scan, but bound below on large system. */
474 j = READ_ONCE(jiffies_till_first_fqs) +
475 2 * READ_ONCE(jiffies_till_next_fqs);
476 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
477 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
478 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
479 WRITE_ONCE(jiffies_to_sched_qs, j);
480 }
481
param_set_first_fqs_jiffies(const char * val,const struct kernel_param * kp)482 static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
483 {
484 ulong j;
485 int ret = kstrtoul(val, 0, &j);
486
487 if (!ret) {
488 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
489 adjust_jiffies_till_sched_qs();
490 }
491 return ret;
492 }
493
param_set_next_fqs_jiffies(const char * val,const struct kernel_param * kp)494 static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
495 {
496 ulong j;
497 int ret = kstrtoul(val, 0, &j);
498
499 if (!ret) {
500 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
501 adjust_jiffies_till_sched_qs();
502 }
503 return ret;
504 }
505
506 static const struct kernel_param_ops first_fqs_jiffies_ops = {
507 .set = param_set_first_fqs_jiffies,
508 .get = param_get_ulong,
509 };
510
511 static const struct kernel_param_ops next_fqs_jiffies_ops = {
512 .set = param_set_next_fqs_jiffies,
513 .get = param_get_ulong,
514 };
515
516 module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
517 module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
518 module_param(rcu_kick_kthreads, bool, 0644);
519
520 static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
521 static int rcu_pending(int user);
522
523 /*
524 * Return the number of RCU GPs completed thus far for debug & stats.
525 */
rcu_get_gp_seq(void)526 unsigned long rcu_get_gp_seq(void)
527 {
528 return READ_ONCE(rcu_state.gp_seq);
529 }
530 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
531
532 /*
533 * Return the number of RCU expedited batches completed thus far for
534 * debug & stats. Odd numbers mean that a batch is in progress, even
535 * numbers mean idle. The value returned will thus be roughly double
536 * the cumulative batches since boot.
537 */
rcu_exp_batches_completed(void)538 unsigned long rcu_exp_batches_completed(void)
539 {
540 return rcu_state.expedited_sequence;
541 }
542 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
543
544 /*
545 * Return the root node of the rcu_state structure.
546 */
rcu_get_root(void)547 static struct rcu_node *rcu_get_root(void)
548 {
549 return &rcu_state.node[0];
550 }
551
552 /*
553 * Send along grace-period-related data for rcutorture diagnostics.
554 */
rcutorture_get_gp_data(int * flags,unsigned long * gp_seq)555 void rcutorture_get_gp_data(int *flags, unsigned long *gp_seq)
556 {
557 *flags = READ_ONCE(rcu_state.gp_flags);
558 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
559 }
560 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
561
562 #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK))
563 /*
564 * An empty function that will trigger a reschedule on
565 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
566 */
late_wakeup_func(struct irq_work * work)567 static void late_wakeup_func(struct irq_work *work)
568 {
569 }
570
571 static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
572 IRQ_WORK_INIT(late_wakeup_func);
573
574 /*
575 * If either:
576 *
577 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
578 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
579 *
580 * In these cases the late RCU wake ups aren't supported in the resched loops and our
581 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
582 * get re-enabled again.
583 */
rcu_irq_work_resched(void)584 noinstr void rcu_irq_work_resched(void)
585 {
586 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
587
588 if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
589 return;
590
591 if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
592 return;
593
594 instrumentation_begin();
595 if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
596 irq_work_queue(this_cpu_ptr(&late_wakeup_work));
597 }
598 instrumentation_end();
599 }
600 #endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */
601
602 #ifdef CONFIG_PROVE_RCU
603 /**
604 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
605 */
rcu_irq_exit_check_preempt(void)606 void rcu_irq_exit_check_preempt(void)
607 {
608 lockdep_assert_irqs_disabled();
609
610 RCU_LOCKDEP_WARN(ct_nesting() <= 0,
611 "RCU nesting counter underflow/zero!");
612 RCU_LOCKDEP_WARN(ct_nmi_nesting() !=
613 CT_NESTING_IRQ_NONIDLE,
614 "Bad RCU nmi_nesting counter\n");
615 RCU_LOCKDEP_WARN(!rcu_is_watching_curr_cpu(),
616 "RCU in extended quiescent state!");
617 }
618 #endif /* #ifdef CONFIG_PROVE_RCU */
619
620 #ifdef CONFIG_NO_HZ_FULL
621 /**
622 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
623 *
624 * The scheduler tick is not normally enabled when CPUs enter the kernel
625 * from nohz_full userspace execution. After all, nohz_full userspace
626 * execution is an RCU quiescent state and the time executing in the kernel
627 * is quite short. Except of course when it isn't. And it is not hard to
628 * cause a large system to spend tens of seconds or even minutes looping
629 * in the kernel, which can cause a number of problems, include RCU CPU
630 * stall warnings.
631 *
632 * Therefore, if a nohz_full CPU fails to report a quiescent state
633 * in a timely manner, the RCU grace-period kthread sets that CPU's
634 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
635 * exception will invoke this function, which will turn on the scheduler
636 * tick, which will enable RCU to detect that CPU's quiescent states,
637 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
638 * The tick will be disabled once a quiescent state is reported for
639 * this CPU.
640 *
641 * Of course, in carefully tuned systems, there might never be an
642 * interrupt or exception. In that case, the RCU grace-period kthread
643 * will eventually cause one to happen. However, in less carefully
644 * controlled environments, this function allows RCU to get what it
645 * needs without creating otherwise useless interruptions.
646 */
__rcu_irq_enter_check_tick(void)647 void __rcu_irq_enter_check_tick(void)
648 {
649 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
650
651 // If we're here from NMI there's nothing to do.
652 if (in_nmi())
653 return;
654
655 RCU_LOCKDEP_WARN(!rcu_is_watching_curr_cpu(),
656 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
657
658 if (!tick_nohz_full_cpu(rdp->cpu) ||
659 !READ_ONCE(rdp->rcu_urgent_qs) ||
660 READ_ONCE(rdp->rcu_forced_tick)) {
661 // RCU doesn't need nohz_full help from this CPU, or it is
662 // already getting that help.
663 return;
664 }
665
666 // We get here only when not in an extended quiescent state and
667 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
668 // already watching and (2) The fact that we are in an interrupt
669 // handler and that the rcu_node lock is an irq-disabled lock
670 // prevents self-deadlock. So we can safely recheck under the lock.
671 // Note that the nohz_full state currently cannot change.
672 raw_spin_lock_rcu_node(rdp->mynode);
673 if (READ_ONCE(rdp->rcu_urgent_qs) && !rdp->rcu_forced_tick) {
674 // A nohz_full CPU is in the kernel and RCU needs a
675 // quiescent state. Turn on the tick!
676 WRITE_ONCE(rdp->rcu_forced_tick, true);
677 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
678 }
679 raw_spin_unlock_rcu_node(rdp->mynode);
680 }
681 NOKPROBE_SYMBOL(__rcu_irq_enter_check_tick);
682 #endif /* CONFIG_NO_HZ_FULL */
683
684 /*
685 * Check to see if any future non-offloaded RCU-related work will need
686 * to be done by the current CPU, even if none need be done immediately,
687 * returning 1 if so. This function is part of the RCU implementation;
688 * it is -not- an exported member of the RCU API. This is used by
689 * the idle-entry code to figure out whether it is safe to disable the
690 * scheduler-clock interrupt.
691 *
692 * Just check whether or not this CPU has non-offloaded RCU callbacks
693 * queued.
694 */
rcu_needs_cpu(void)695 int rcu_needs_cpu(void)
696 {
697 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
698 !rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data));
699 }
700
701 /*
702 * If any sort of urgency was applied to the current CPU (for example,
703 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
704 * to get to a quiescent state, disable it.
705 */
rcu_disable_urgency_upon_qs(struct rcu_data * rdp)706 static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
707 {
708 raw_lockdep_assert_held_rcu_node(rdp->mynode);
709 WRITE_ONCE(rdp->rcu_urgent_qs, false);
710 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
711 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
712 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
713 WRITE_ONCE(rdp->rcu_forced_tick, false);
714 }
715 }
716
717 /**
718 * rcu_is_watching - RCU read-side critical sections permitted on current CPU?
719 *
720 * Return @true if RCU is watching the running CPU and @false otherwise.
721 * An @true return means that this CPU can safely enter RCU read-side
722 * critical sections.
723 *
724 * Although calls to rcu_is_watching() from most parts of the kernel
725 * will return @true, there are important exceptions. For example, if the
726 * current CPU is deep within its idle loop, in kernel entry/exit code,
727 * or offline, rcu_is_watching() will return @false.
728 *
729 * Make notrace because it can be called by the internal functions of
730 * ftrace, and making this notrace removes unnecessary recursion calls.
731 */
rcu_is_watching(void)732 notrace bool rcu_is_watching(void)
733 {
734 bool ret;
735
736 preempt_disable_notrace();
737 ret = rcu_is_watching_curr_cpu();
738 preempt_enable_notrace();
739 return ret;
740 }
741 EXPORT_SYMBOL_GPL(rcu_is_watching);
742
743 /*
744 * If a holdout task is actually running, request an urgent quiescent
745 * state from its CPU. This is unsynchronized, so migrations can cause
746 * the request to go to the wrong CPU. Which is OK, all that will happen
747 * is that the CPU's next context switch will be a bit slower and next
748 * time around this task will generate another request.
749 */
rcu_request_urgent_qs_task(struct task_struct * t)750 void rcu_request_urgent_qs_task(struct task_struct *t)
751 {
752 int cpu;
753
754 barrier();
755 cpu = task_cpu(t);
756 if (!task_curr(t))
757 return; /* This task is not running on that CPU. */
758 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
759 }
760
761 /*
762 * When trying to report a quiescent state on behalf of some other CPU,
763 * it is our responsibility to check for and handle potential overflow
764 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
765 * After all, the CPU might be in deep idle state, and thus executing no
766 * code whatsoever.
767 */
rcu_gpnum_ovf(struct rcu_node * rnp,struct rcu_data * rdp)768 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
769 {
770 raw_lockdep_assert_held_rcu_node(rnp);
771 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
772 rnp->gp_seq))
773 WRITE_ONCE(rdp->gpwrap, true);
774 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
775 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
776 }
777
778 /*
779 * Snapshot the specified CPU's RCU_WATCHING counter so that we can later
780 * credit them with an implicit quiescent state. Return 1 if this CPU
781 * is in dynticks idle mode, which is an extended quiescent state.
782 */
rcu_watching_snap_save(struct rcu_data * rdp)783 static int rcu_watching_snap_save(struct rcu_data *rdp)
784 {
785 /*
786 * Full ordering between remote CPU's post idle accesses and updater's
787 * accesses prior to current GP (and also the started GP sequence number)
788 * is enforced by rcu_seq_start() implicit barrier and even further by
789 * smp_mb__after_unlock_lock() barriers chained all the way throughout the
790 * rnp locking tree since rcu_gp_init() and up to the current leaf rnp
791 * locking.
792 *
793 * Ordering between remote CPU's pre idle accesses and post grace period
794 * updater's accesses is enforced by the below acquire semantic.
795 */
796 rdp->watching_snap = ct_rcu_watching_cpu_acquire(rdp->cpu);
797 if (rcu_watching_snap_in_eqs(rdp->watching_snap)) {
798 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
799 rcu_gpnum_ovf(rdp->mynode, rdp);
800 return 1;
801 }
802 return 0;
803 }
804
805 /*
806 * Returns positive if the specified CPU has passed through a quiescent state
807 * by virtue of being in or having passed through an dynticks idle state since
808 * the last call to rcu_watching_snap_save() for this same CPU, or by
809 * virtue of having been offline.
810 *
811 * Returns negative if the specified CPU needs a force resched.
812 *
813 * Returns zero otherwise.
814 */
rcu_watching_snap_recheck(struct rcu_data * rdp)815 static int rcu_watching_snap_recheck(struct rcu_data *rdp)
816 {
817 unsigned long jtsq;
818 int ret = 0;
819 struct rcu_node *rnp = rdp->mynode;
820
821 /*
822 * If the CPU passed through or entered a dynticks idle phase with
823 * no active irq/NMI handlers, then we can safely pretend that the CPU
824 * already acknowledged the request to pass through a quiescent
825 * state. Either way, that CPU cannot possibly be in an RCU
826 * read-side critical section that started before the beginning
827 * of the current RCU grace period.
828 */
829 if (rcu_watching_snap_stopped_since(rdp, rdp->watching_snap)) {
830 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
831 rcu_gpnum_ovf(rnp, rdp);
832 return 1;
833 }
834
835 /*
836 * Complain if a CPU that is considered to be offline from RCU's
837 * perspective has not yet reported a quiescent state. After all,
838 * the offline CPU should have reported a quiescent state during
839 * the CPU-offline process, or, failing that, by rcu_gp_init()
840 * if it ran concurrently with either the CPU going offline or the
841 * last task on a leaf rcu_node structure exiting its RCU read-side
842 * critical section while all CPUs corresponding to that structure
843 * are offline. This added warning detects bugs in any of these
844 * code paths.
845 *
846 * The rcu_node structure's ->lock is held here, which excludes
847 * the relevant portions the CPU-hotplug code, the grace-period
848 * initialization code, and the rcu_read_unlock() code paths.
849 *
850 * For more detail, please refer to the "Hotplug CPU" section
851 * of RCU's Requirements documentation.
852 */
853 if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp))) {
854 struct rcu_node *rnp1;
855
856 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
857 __func__, rnp->grplo, rnp->grphi, rnp->level,
858 (long)rnp->gp_seq, (long)rnp->completedqs);
859 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
860 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
861 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
862 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
863 __func__, rdp->cpu, ".o"[rcu_rdp_cpu_online(rdp)],
864 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_state,
865 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_state);
866 return 1; /* Break things loose after complaining. */
867 }
868
869 /*
870 * A CPU running for an extended time within the kernel can
871 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
872 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
873 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
874 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
875 * variable are safe because the assignments are repeated if this
876 * CPU failed to pass through a quiescent state. This code
877 * also checks .jiffies_resched in case jiffies_to_sched_qs
878 * is set way high.
879 */
880 jtsq = READ_ONCE(jiffies_to_sched_qs);
881 if (!READ_ONCE(rdp->rcu_need_heavy_qs) &&
882 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
883 time_after(jiffies, rcu_state.jiffies_resched) ||
884 rcu_state.cbovld)) {
885 WRITE_ONCE(rdp->rcu_need_heavy_qs, true);
886 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
887 smp_store_release(&rdp->rcu_urgent_qs, true);
888 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
889 WRITE_ONCE(rdp->rcu_urgent_qs, true);
890 }
891
892 /*
893 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
894 * The above code handles this, but only for straight cond_resched().
895 * And some in-kernel loops check need_resched() before calling
896 * cond_resched(), which defeats the above code for CPUs that are
897 * running in-kernel with scheduling-clock interrupts disabled.
898 * So hit them over the head with the resched_cpu() hammer!
899 */
900 if (tick_nohz_full_cpu(rdp->cpu) &&
901 (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
902 rcu_state.cbovld)) {
903 WRITE_ONCE(rdp->rcu_urgent_qs, true);
904 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
905 ret = -1;
906 }
907
908 /*
909 * If more than halfway to RCU CPU stall-warning time, invoke
910 * resched_cpu() more frequently to try to loosen things up a bit.
911 * Also check to see if the CPU is getting hammered with interrupts,
912 * but only once per grace period, just to keep the IPIs down to
913 * a dull roar.
914 */
915 if (time_after(jiffies, rcu_state.jiffies_resched)) {
916 if (time_after(jiffies,
917 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
918 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
919 ret = -1;
920 }
921 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
922 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
923 (rnp->ffmask & rdp->grpmask)) {
924 rdp->rcu_iw_pending = true;
925 rdp->rcu_iw_gp_seq = rnp->gp_seq;
926 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
927 }
928
929 if (rcu_cpu_stall_cputime && rdp->snap_record.gp_seq != rdp->gp_seq) {
930 int cpu = rdp->cpu;
931 struct rcu_snap_record *rsrp;
932 struct kernel_cpustat *kcsp;
933
934 kcsp = &kcpustat_cpu(cpu);
935
936 rsrp = &rdp->snap_record;
937 rsrp->cputime_irq = kcpustat_field(kcsp, CPUTIME_IRQ, cpu);
938 rsrp->cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu);
939 rsrp->cputime_system = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu);
940 rsrp->nr_hardirqs = kstat_cpu_irqs_sum(rdp->cpu);
941 rsrp->nr_softirqs = kstat_cpu_softirqs_sum(rdp->cpu);
942 rsrp->nr_csw = nr_context_switches_cpu(rdp->cpu);
943 rsrp->jiffies = jiffies;
944 rsrp->gp_seq = rdp->gp_seq;
945 }
946 }
947
948 return ret;
949 }
950
951 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
trace_rcu_this_gp(struct rcu_node * rnp,struct rcu_data * rdp,unsigned long gp_seq_req,const char * s)952 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
953 unsigned long gp_seq_req, const char *s)
954 {
955 trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
956 gp_seq_req, rnp->level,
957 rnp->grplo, rnp->grphi, s);
958 }
959
960 /*
961 * rcu_start_this_gp - Request the start of a particular grace period
962 * @rnp_start: The leaf node of the CPU from which to start.
963 * @rdp: The rcu_data corresponding to the CPU from which to start.
964 * @gp_seq_req: The gp_seq of the grace period to start.
965 *
966 * Start the specified grace period, as needed to handle newly arrived
967 * callbacks. The required future grace periods are recorded in each
968 * rcu_node structure's ->gp_seq_needed field. Returns true if there
969 * is reason to awaken the grace-period kthread.
970 *
971 * The caller must hold the specified rcu_node structure's ->lock, which
972 * is why the caller is responsible for waking the grace-period kthread.
973 *
974 * Returns true if the GP thread needs to be awakened else false.
975 */
rcu_start_this_gp(struct rcu_node * rnp_start,struct rcu_data * rdp,unsigned long gp_seq_req)976 static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
977 unsigned long gp_seq_req)
978 {
979 bool ret = false;
980 struct rcu_node *rnp;
981
982 /*
983 * Use funnel locking to either acquire the root rcu_node
984 * structure's lock or bail out if the need for this grace period
985 * has already been recorded -- or if that grace period has in
986 * fact already started. If there is already a grace period in
987 * progress in a non-leaf node, no recording is needed because the
988 * end of the grace period will scan the leaf rcu_node structures.
989 * Note that rnp_start->lock must not be released.
990 */
991 raw_lockdep_assert_held_rcu_node(rnp_start);
992 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
993 for (rnp = rnp_start; 1; rnp = rnp->parent) {
994 if (rnp != rnp_start)
995 raw_spin_lock_rcu_node(rnp);
996 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
997 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
998 (rnp != rnp_start &&
999 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1000 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1001 TPS("Prestarted"));
1002 goto unlock_out;
1003 }
1004 WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
1005 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1006 /*
1007 * We just marked the leaf or internal node, and a
1008 * grace period is in progress, which means that
1009 * rcu_gp_cleanup() will see the marking. Bail to
1010 * reduce contention.
1011 */
1012 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1013 TPS("Startedleaf"));
1014 goto unlock_out;
1015 }
1016 if (rnp != rnp_start && rnp->parent != NULL)
1017 raw_spin_unlock_rcu_node(rnp);
1018 if (!rnp->parent)
1019 break; /* At root, and perhaps also leaf. */
1020 }
1021
1022 /* If GP already in progress, just leave, otherwise start one. */
1023 if (rcu_gp_in_progress()) {
1024 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1025 goto unlock_out;
1026 }
1027 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1028 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1029 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1030 if (!READ_ONCE(rcu_state.gp_kthread)) {
1031 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1032 goto unlock_out;
1033 }
1034 trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1035 ret = true; /* Caller must wake GP kthread. */
1036 unlock_out:
1037 /* Push furthest requested GP to leaf node and rcu_data structure. */
1038 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1039 WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1040 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1041 }
1042 if (rnp != rnp_start)
1043 raw_spin_unlock_rcu_node(rnp);
1044 return ret;
1045 }
1046
1047 /*
1048 * Clean up any old requests for the just-ended grace period. Also return
1049 * whether any additional grace periods have been requested.
1050 */
rcu_future_gp_cleanup(struct rcu_node * rnp)1051 static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1052 {
1053 bool needmore;
1054 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1055
1056 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1057 if (!needmore)
1058 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1059 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1060 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1061 return needmore;
1062 }
1063
swake_up_one_online_ipi(void * arg)1064 static void swake_up_one_online_ipi(void *arg)
1065 {
1066 struct swait_queue_head *wqh = arg;
1067
1068 swake_up_one(wqh);
1069 }
1070
swake_up_one_online(struct swait_queue_head * wqh)1071 static void swake_up_one_online(struct swait_queue_head *wqh)
1072 {
1073 int cpu = get_cpu();
1074
1075 /*
1076 * If called from rcutree_report_cpu_starting(), wake up
1077 * is dangerous that late in the CPU-down hotplug process. The
1078 * scheduler might queue an ignored hrtimer. Defer the wake up
1079 * to an online CPU instead.
1080 */
1081 if (unlikely(cpu_is_offline(cpu))) {
1082 int target;
1083
1084 target = cpumask_any_and(housekeeping_cpumask(HK_TYPE_RCU),
1085 cpu_online_mask);
1086
1087 smp_call_function_single(target, swake_up_one_online_ipi,
1088 wqh, 0);
1089 put_cpu();
1090 } else {
1091 put_cpu();
1092 swake_up_one(wqh);
1093 }
1094 }
1095
1096 /*
1097 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1098 * interrupt or softirq handler, in which case we just might immediately
1099 * sleep upon return, resulting in a grace-period hang), and don't bother
1100 * awakening when there is nothing for the grace-period kthread to do
1101 * (as in several CPUs raced to awaken, we lost), and finally don't try
1102 * to awaken a kthread that has not yet been created. If all those checks
1103 * are passed, track some debug information and awaken.
1104 *
1105 * So why do the self-wakeup when in an interrupt or softirq handler
1106 * in the grace-period kthread's context? Because the kthread might have
1107 * been interrupted just as it was going to sleep, and just after the final
1108 * pre-sleep check of the awaken condition. In this case, a wakeup really
1109 * is required, and is therefore supplied.
1110 */
rcu_gp_kthread_wake(void)1111 static void rcu_gp_kthread_wake(void)
1112 {
1113 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1114
1115 if ((current == t && !in_hardirq() && !in_serving_softirq()) ||
1116 !READ_ONCE(rcu_state.gp_flags) || !t)
1117 return;
1118 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1119 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1120 swake_up_one_online(&rcu_state.gp_wq);
1121 }
1122
1123 /*
1124 * If there is room, assign a ->gp_seq number to any callbacks on this
1125 * CPU that have not already been assigned. Also accelerate any callbacks
1126 * that were previously assigned a ->gp_seq number that has since proven
1127 * to be too conservative, which can happen if callbacks get assigned a
1128 * ->gp_seq number while RCU is idle, but with reference to a non-root
1129 * rcu_node structure. This function is idempotent, so it does not hurt
1130 * to call it repeatedly. Returns an flag saying that we should awaken
1131 * the RCU grace-period kthread.
1132 *
1133 * The caller must hold rnp->lock with interrupts disabled.
1134 */
rcu_accelerate_cbs(struct rcu_node * rnp,struct rcu_data * rdp)1135 static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1136 {
1137 unsigned long gp_seq_req;
1138 bool ret = false;
1139
1140 rcu_lockdep_assert_cblist_protected(rdp);
1141 raw_lockdep_assert_held_rcu_node(rnp);
1142
1143 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1144 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1145 return false;
1146
1147 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1148
1149 /*
1150 * Callbacks are often registered with incomplete grace-period
1151 * information. Something about the fact that getting exact
1152 * information requires acquiring a global lock... RCU therefore
1153 * makes a conservative estimate of the grace period number at which
1154 * a given callback will become ready to invoke. The following
1155 * code checks this estimate and improves it when possible, thus
1156 * accelerating callback invocation to an earlier grace-period
1157 * number.
1158 */
1159 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1160 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1161 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1162
1163 /* Trace depending on how much we were able to accelerate. */
1164 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1165 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1166 else
1167 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1168
1169 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1170
1171 return ret;
1172 }
1173
1174 /*
1175 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1176 * rcu_node structure's ->lock be held. It consults the cached value
1177 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1178 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1179 * while holding the leaf rcu_node structure's ->lock.
1180 */
rcu_accelerate_cbs_unlocked(struct rcu_node * rnp,struct rcu_data * rdp)1181 static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1182 struct rcu_data *rdp)
1183 {
1184 unsigned long c;
1185 bool needwake;
1186
1187 rcu_lockdep_assert_cblist_protected(rdp);
1188 c = rcu_seq_snap(&rcu_state.gp_seq);
1189 if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1190 /* Old request still live, so mark recent callbacks. */
1191 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1192 return;
1193 }
1194 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1195 needwake = rcu_accelerate_cbs(rnp, rdp);
1196 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1197 if (needwake)
1198 rcu_gp_kthread_wake();
1199 }
1200
1201 /*
1202 * Move any callbacks whose grace period has completed to the
1203 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1204 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1205 * sublist. This function is idempotent, so it does not hurt to
1206 * invoke it repeatedly. As long as it is not invoked -too- often...
1207 * Returns true if the RCU grace-period kthread needs to be awakened.
1208 *
1209 * The caller must hold rnp->lock with interrupts disabled.
1210 */
rcu_advance_cbs(struct rcu_node * rnp,struct rcu_data * rdp)1211 static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1212 {
1213 rcu_lockdep_assert_cblist_protected(rdp);
1214 raw_lockdep_assert_held_rcu_node(rnp);
1215
1216 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1217 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1218 return false;
1219
1220 /*
1221 * Find all callbacks whose ->gp_seq numbers indicate that they
1222 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1223 */
1224 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1225
1226 /* Classify any remaining callbacks. */
1227 return rcu_accelerate_cbs(rnp, rdp);
1228 }
1229
1230 /*
1231 * Move and classify callbacks, but only if doing so won't require
1232 * that the RCU grace-period kthread be awakened.
1233 */
rcu_advance_cbs_nowake(struct rcu_node * rnp,struct rcu_data * rdp)1234 static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1235 struct rcu_data *rdp)
1236 {
1237 rcu_lockdep_assert_cblist_protected(rdp);
1238 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp))
1239 return;
1240 // The grace period cannot end while we hold the rcu_node lock.
1241 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))
1242 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1243 raw_spin_unlock_rcu_node(rnp);
1244 }
1245
1246 /*
1247 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1248 * quiescent state. This is intended to be invoked when the CPU notices
1249 * a new grace period.
1250 */
rcu_strict_gp_check_qs(void)1251 static void rcu_strict_gp_check_qs(void)
1252 {
1253 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1254 rcu_read_lock();
1255 rcu_read_unlock();
1256 }
1257 }
1258
1259 /*
1260 * Update CPU-local rcu_data state to record the beginnings and ends of
1261 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1262 * structure corresponding to the current CPU, and must have irqs disabled.
1263 * Returns true if the grace-period kthread needs to be awakened.
1264 */
__note_gp_changes(struct rcu_node * rnp,struct rcu_data * rdp)1265 static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1266 {
1267 bool ret = false;
1268 bool need_qs;
1269 const bool offloaded = rcu_rdp_is_offloaded(rdp);
1270
1271 raw_lockdep_assert_held_rcu_node(rnp);
1272
1273 if (rdp->gp_seq == rnp->gp_seq)
1274 return false; /* Nothing to do. */
1275
1276 /* Handle the ends of any preceding grace periods first. */
1277 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1278 unlikely(READ_ONCE(rdp->gpwrap))) {
1279 if (!offloaded)
1280 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1281 rdp->core_needs_qs = false;
1282 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1283 } else {
1284 if (!offloaded)
1285 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1286 if (rdp->core_needs_qs)
1287 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1288 }
1289
1290 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1291 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1292 unlikely(READ_ONCE(rdp->gpwrap))) {
1293 /*
1294 * If the current grace period is waiting for this CPU,
1295 * set up to detect a quiescent state, otherwise don't
1296 * go looking for one.
1297 */
1298 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1299 need_qs = !!(rnp->qsmask & rdp->grpmask);
1300 rdp->cpu_no_qs.b.norm = need_qs;
1301 rdp->core_needs_qs = need_qs;
1302 zero_cpu_stall_ticks(rdp);
1303 }
1304 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1305 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1306 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1307 if (IS_ENABLED(CONFIG_PROVE_RCU) && READ_ONCE(rdp->gpwrap))
1308 WRITE_ONCE(rdp->last_sched_clock, jiffies);
1309 WRITE_ONCE(rdp->gpwrap, false);
1310 rcu_gpnum_ovf(rnp, rdp);
1311 return ret;
1312 }
1313
note_gp_changes(struct rcu_data * rdp)1314 static void note_gp_changes(struct rcu_data *rdp)
1315 {
1316 unsigned long flags;
1317 bool needwake;
1318 struct rcu_node *rnp;
1319
1320 local_irq_save(flags);
1321 rnp = rdp->mynode;
1322 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1323 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1324 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1325 local_irq_restore(flags);
1326 return;
1327 }
1328 needwake = __note_gp_changes(rnp, rdp);
1329 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1330 rcu_strict_gp_check_qs();
1331 if (needwake)
1332 rcu_gp_kthread_wake();
1333 }
1334
1335 static atomic_t *rcu_gp_slow_suppress;
1336
1337 /* Register a counter to suppress debugging grace-period delays. */
rcu_gp_slow_register(atomic_t * rgssp)1338 void rcu_gp_slow_register(atomic_t *rgssp)
1339 {
1340 WARN_ON_ONCE(rcu_gp_slow_suppress);
1341
1342 WRITE_ONCE(rcu_gp_slow_suppress, rgssp);
1343 }
1344 EXPORT_SYMBOL_GPL(rcu_gp_slow_register);
1345
1346 /* Unregister a counter, with NULL for not caring which. */
rcu_gp_slow_unregister(atomic_t * rgssp)1347 void rcu_gp_slow_unregister(atomic_t *rgssp)
1348 {
1349 WARN_ON_ONCE(rgssp && rgssp != rcu_gp_slow_suppress && rcu_gp_slow_suppress != NULL);
1350
1351 WRITE_ONCE(rcu_gp_slow_suppress, NULL);
1352 }
1353 EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister);
1354
rcu_gp_slow_is_suppressed(void)1355 static bool rcu_gp_slow_is_suppressed(void)
1356 {
1357 atomic_t *rgssp = READ_ONCE(rcu_gp_slow_suppress);
1358
1359 return rgssp && atomic_read(rgssp);
1360 }
1361
rcu_gp_slow(int delay)1362 static void rcu_gp_slow(int delay)
1363 {
1364 if (!rcu_gp_slow_is_suppressed() && delay > 0 &&
1365 !(rcu_seq_ctr(rcu_state.gp_seq) % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1366 schedule_timeout_idle(delay);
1367 }
1368
1369 static unsigned long sleep_duration;
1370
1371 /* Allow rcutorture to stall the grace-period kthread. */
rcu_gp_set_torture_wait(int duration)1372 void rcu_gp_set_torture_wait(int duration)
1373 {
1374 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1375 WRITE_ONCE(sleep_duration, duration);
1376 }
1377 EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1378
1379 /* Actually implement the aforementioned wait. */
rcu_gp_torture_wait(void)1380 static void rcu_gp_torture_wait(void)
1381 {
1382 unsigned long duration;
1383
1384 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1385 return;
1386 duration = xchg(&sleep_duration, 0UL);
1387 if (duration > 0) {
1388 pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1389 schedule_timeout_idle(duration);
1390 pr_alert("%s: Wait complete\n", __func__);
1391 }
1392 }
1393
1394 /*
1395 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1396 * processing.
1397 */
rcu_strict_gp_boundary(void * unused)1398 static void rcu_strict_gp_boundary(void *unused)
1399 {
1400 invoke_rcu_core();
1401 }
1402
1403 // Make the polled API aware of the beginning of a grace period.
rcu_poll_gp_seq_start(unsigned long * snap)1404 static void rcu_poll_gp_seq_start(unsigned long *snap)
1405 {
1406 struct rcu_node *rnp = rcu_get_root();
1407
1408 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1409 raw_lockdep_assert_held_rcu_node(rnp);
1410
1411 // If RCU was idle, note beginning of GP.
1412 if (!rcu_seq_state(rcu_state.gp_seq_polled))
1413 rcu_seq_start(&rcu_state.gp_seq_polled);
1414
1415 // Either way, record current state.
1416 *snap = rcu_state.gp_seq_polled;
1417 }
1418
1419 // Make the polled API aware of the end of a grace period.
rcu_poll_gp_seq_end(unsigned long * snap)1420 static void rcu_poll_gp_seq_end(unsigned long *snap)
1421 {
1422 struct rcu_node *rnp = rcu_get_root();
1423
1424 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1425 raw_lockdep_assert_held_rcu_node(rnp);
1426
1427 // If the previously noted GP is still in effect, record the
1428 // end of that GP. Either way, zero counter to avoid counter-wrap
1429 // problems.
1430 if (*snap && *snap == rcu_state.gp_seq_polled) {
1431 rcu_seq_end(&rcu_state.gp_seq_polled);
1432 rcu_state.gp_seq_polled_snap = 0;
1433 rcu_state.gp_seq_polled_exp_snap = 0;
1434 } else {
1435 *snap = 0;
1436 }
1437 }
1438
1439 // Make the polled API aware of the beginning of a grace period, but
1440 // where caller does not hold the root rcu_node structure's lock.
rcu_poll_gp_seq_start_unlocked(unsigned long * snap)1441 static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap)
1442 {
1443 unsigned long flags;
1444 struct rcu_node *rnp = rcu_get_root();
1445
1446 if (rcu_init_invoked()) {
1447 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1448 lockdep_assert_irqs_enabled();
1449 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1450 }
1451 rcu_poll_gp_seq_start(snap);
1452 if (rcu_init_invoked())
1453 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1454 }
1455
1456 // Make the polled API aware of the end of a grace period, but where
1457 // caller does not hold the root rcu_node structure's lock.
rcu_poll_gp_seq_end_unlocked(unsigned long * snap)1458 static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
1459 {
1460 unsigned long flags;
1461 struct rcu_node *rnp = rcu_get_root();
1462
1463 if (rcu_init_invoked()) {
1464 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE)
1465 lockdep_assert_irqs_enabled();
1466 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1467 }
1468 rcu_poll_gp_seq_end(snap);
1469 if (rcu_init_invoked())
1470 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1471 }
1472
1473 /*
1474 * There is a single llist, which is used for handling
1475 * synchronize_rcu() users' enqueued rcu_synchronize nodes.
1476 * Within this llist, there are two tail pointers:
1477 *
1478 * wait tail: Tracks the set of nodes, which need to
1479 * wait for the current GP to complete.
1480 * done tail: Tracks the set of nodes, for which grace
1481 * period has elapsed. These nodes processing
1482 * will be done as part of the cleanup work
1483 * execution by a kworker.
1484 *
1485 * At every grace period init, a new wait node is added
1486 * to the llist. This wait node is used as wait tail
1487 * for this new grace period. Given that there are a fixed
1488 * number of wait nodes, if all wait nodes are in use
1489 * (which can happen when kworker callback processing
1490 * is delayed) and additional grace period is requested.
1491 * This means, a system is slow in processing callbacks.
1492 *
1493 * TODO: If a slow processing is detected, a first node
1494 * in the llist should be used as a wait-tail for this
1495 * grace period, therefore users which should wait due
1496 * to a slow process are handled by _this_ grace period
1497 * and not next.
1498 *
1499 * Below is an illustration of how the done and wait
1500 * tail pointers move from one set of rcu_synchronize nodes
1501 * to the other, as grace periods start and finish and
1502 * nodes are processed by kworker.
1503 *
1504 *
1505 * a. Initial llist callbacks list:
1506 *
1507 * +----------+ +--------+ +-------+
1508 * | | | | | |
1509 * | head |---------> | cb2 |--------->| cb1 |
1510 * | | | | | |
1511 * +----------+ +--------+ +-------+
1512 *
1513 *
1514 *
1515 * b. New GP1 Start:
1516 *
1517 * WAIT TAIL
1518 * |
1519 * |
1520 * v
1521 * +----------+ +--------+ +--------+ +-------+
1522 * | | | | | | | |
1523 * | head ------> wait |------> cb2 |------> | cb1 |
1524 * | | | head1 | | | | |
1525 * +----------+ +--------+ +--------+ +-------+
1526 *
1527 *
1528 *
1529 * c. GP completion:
1530 *
1531 * WAIT_TAIL == DONE_TAIL
1532 *
1533 * DONE TAIL
1534 * |
1535 * |
1536 * v
1537 * +----------+ +--------+ +--------+ +-------+
1538 * | | | | | | | |
1539 * | head ------> wait |------> cb2 |------> | cb1 |
1540 * | | | head1 | | | | |
1541 * +----------+ +--------+ +--------+ +-------+
1542 *
1543 *
1544 *
1545 * d. New callbacks and GP2 start:
1546 *
1547 * WAIT TAIL DONE TAIL
1548 * | |
1549 * | |
1550 * v v
1551 * +----------+ +------+ +------+ +------+ +-----+ +-----+ +-----+
1552 * | | | | | | | | | | | | | |
1553 * | head ------> wait |--->| cb4 |--->| cb3 |--->|wait |--->| cb2 |--->| cb1 |
1554 * | | | head2| | | | | |head1| | | | |
1555 * +----------+ +------+ +------+ +------+ +-----+ +-----+ +-----+
1556 *
1557 *
1558 *
1559 * e. GP2 completion:
1560 *
1561 * WAIT_TAIL == DONE_TAIL
1562 * DONE TAIL
1563 * |
1564 * |
1565 * v
1566 * +----------+ +------+ +------+ +------+ +-----+ +-----+ +-----+
1567 * | | | | | | | | | | | | | |
1568 * | head ------> wait |--->| cb4 |--->| cb3 |--->|wait |--->| cb2 |--->| cb1 |
1569 * | | | head2| | | | | |head1| | | | |
1570 * +----------+ +------+ +------+ +------+ +-----+ +-----+ +-----+
1571 *
1572 *
1573 * While the llist state transitions from d to e, a kworker
1574 * can start executing rcu_sr_normal_gp_cleanup_work() and
1575 * can observe either the old done tail (@c) or the new
1576 * done tail (@e). So, done tail updates and reads need
1577 * to use the rel-acq semantics. If the concurrent kworker
1578 * observes the old done tail, the newly queued work
1579 * execution will process the updated done tail. If the
1580 * concurrent kworker observes the new done tail, then
1581 * the newly queued work will skip processing the done
1582 * tail, as workqueue semantics guarantees that the new
1583 * work is executed only after the previous one completes.
1584 *
1585 * f. kworker callbacks processing complete:
1586 *
1587 *
1588 * DONE TAIL
1589 * |
1590 * |
1591 * v
1592 * +----------+ +--------+
1593 * | | | |
1594 * | head ------> wait |
1595 * | | | head2 |
1596 * +----------+ +--------+
1597 *
1598 */
rcu_sr_is_wait_head(struct llist_node * node)1599 static bool rcu_sr_is_wait_head(struct llist_node *node)
1600 {
1601 return &(rcu_state.srs_wait_nodes)[0].node <= node &&
1602 node <= &(rcu_state.srs_wait_nodes)[SR_NORMAL_GP_WAIT_HEAD_MAX - 1].node;
1603 }
1604
rcu_sr_get_wait_head(void)1605 static struct llist_node *rcu_sr_get_wait_head(void)
1606 {
1607 struct sr_wait_node *sr_wn;
1608 int i;
1609
1610 for (i = 0; i < SR_NORMAL_GP_WAIT_HEAD_MAX; i++) {
1611 sr_wn = &(rcu_state.srs_wait_nodes)[i];
1612
1613 if (!atomic_cmpxchg_acquire(&sr_wn->inuse, 0, 1))
1614 return &sr_wn->node;
1615 }
1616
1617 return NULL;
1618 }
1619
rcu_sr_put_wait_head(struct llist_node * node)1620 static void rcu_sr_put_wait_head(struct llist_node *node)
1621 {
1622 struct sr_wait_node *sr_wn = container_of(node, struct sr_wait_node, node);
1623
1624 atomic_set_release(&sr_wn->inuse, 0);
1625 }
1626
1627 /* Disabled by default. */
1628 static int rcu_normal_wake_from_gp;
1629 module_param(rcu_normal_wake_from_gp, int, 0644);
1630 static struct workqueue_struct *sync_wq;
1631
rcu_sr_normal_complete(struct llist_node * node)1632 static void rcu_sr_normal_complete(struct llist_node *node)
1633 {
1634 struct rcu_synchronize *rs = container_of(
1635 (struct rcu_head *) node, struct rcu_synchronize, head);
1636 unsigned long oldstate = (unsigned long) rs->head.func;
1637
1638 WARN_ONCE(IS_ENABLED(CONFIG_PROVE_RCU) &&
1639 !poll_state_synchronize_rcu(oldstate),
1640 "A full grace period is not passed yet: %lu",
1641 rcu_seq_diff(get_state_synchronize_rcu(), oldstate));
1642
1643 /* Finally. */
1644 complete(&rs->completion);
1645 }
1646
rcu_sr_normal_gp_cleanup_work(struct work_struct * work)1647 static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work)
1648 {
1649 struct llist_node *done, *rcu, *next, *head;
1650
1651 /*
1652 * This work execution can potentially execute
1653 * while a new done tail is being updated by
1654 * grace period kthread in rcu_sr_normal_gp_cleanup().
1655 * So, read and updates of done tail need to
1656 * follow acq-rel semantics.
1657 *
1658 * Given that wq semantics guarantees that a single work
1659 * cannot execute concurrently by multiple kworkers,
1660 * the done tail list manipulations are protected here.
1661 */
1662 done = smp_load_acquire(&rcu_state.srs_done_tail);
1663 if (WARN_ON_ONCE(!done))
1664 return;
1665
1666 WARN_ON_ONCE(!rcu_sr_is_wait_head(done));
1667 head = done->next;
1668 done->next = NULL;
1669
1670 /*
1671 * The dummy node, which is pointed to by the
1672 * done tail which is acq-read above is not removed
1673 * here. This allows lockless additions of new
1674 * rcu_synchronize nodes in rcu_sr_normal_add_req(),
1675 * while the cleanup work executes. The dummy
1676 * nodes is removed, in next round of cleanup
1677 * work execution.
1678 */
1679 llist_for_each_safe(rcu, next, head) {
1680 if (!rcu_sr_is_wait_head(rcu)) {
1681 rcu_sr_normal_complete(rcu);
1682 continue;
1683 }
1684
1685 rcu_sr_put_wait_head(rcu);
1686 }
1687
1688 /* Order list manipulations with atomic access. */
1689 atomic_dec_return_release(&rcu_state.srs_cleanups_pending);
1690 }
1691
1692 /*
1693 * Helper function for rcu_gp_cleanup().
1694 */
rcu_sr_normal_gp_cleanup(void)1695 static void rcu_sr_normal_gp_cleanup(void)
1696 {
1697 struct llist_node *wait_tail, *next = NULL, *rcu = NULL;
1698 int done = 0;
1699
1700 wait_tail = rcu_state.srs_wait_tail;
1701 if (wait_tail == NULL)
1702 return;
1703
1704 rcu_state.srs_wait_tail = NULL;
1705 ASSERT_EXCLUSIVE_WRITER(rcu_state.srs_wait_tail);
1706 WARN_ON_ONCE(!rcu_sr_is_wait_head(wait_tail));
1707
1708 /*
1709 * Process (a) and (d) cases. See an illustration.
1710 */
1711 llist_for_each_safe(rcu, next, wait_tail->next) {
1712 if (rcu_sr_is_wait_head(rcu))
1713 break;
1714
1715 rcu_sr_normal_complete(rcu);
1716 // It can be last, update a next on this step.
1717 wait_tail->next = next;
1718
1719 if (++done == SR_MAX_USERS_WAKE_FROM_GP)
1720 break;
1721 }
1722
1723 /*
1724 * Fast path, no more users to process except putting the second last
1725 * wait head if no inflight-workers. If there are in-flight workers,
1726 * they will remove the last wait head.
1727 *
1728 * Note that the ACQUIRE orders atomic access with list manipulation.
1729 */
1730 if (wait_tail->next && wait_tail->next->next == NULL &&
1731 rcu_sr_is_wait_head(wait_tail->next) &&
1732 !atomic_read_acquire(&rcu_state.srs_cleanups_pending)) {
1733 rcu_sr_put_wait_head(wait_tail->next);
1734 wait_tail->next = NULL;
1735 }
1736
1737 /* Concurrent sr_normal_gp_cleanup work might observe this update. */
1738 ASSERT_EXCLUSIVE_WRITER(rcu_state.srs_done_tail);
1739 smp_store_release(&rcu_state.srs_done_tail, wait_tail);
1740
1741 /*
1742 * We schedule a work in order to perform a final processing
1743 * of outstanding users(if still left) and releasing wait-heads
1744 * added by rcu_sr_normal_gp_init() call.
1745 */
1746 if (wait_tail->next) {
1747 atomic_inc(&rcu_state.srs_cleanups_pending);
1748 if (!queue_work(sync_wq, &rcu_state.srs_cleanup_work))
1749 atomic_dec(&rcu_state.srs_cleanups_pending);
1750 }
1751 }
1752
1753 /*
1754 * Helper function for rcu_gp_init().
1755 */
rcu_sr_normal_gp_init(void)1756 static bool rcu_sr_normal_gp_init(void)
1757 {
1758 struct llist_node *first;
1759 struct llist_node *wait_head;
1760 bool start_new_poll = false;
1761
1762 first = READ_ONCE(rcu_state.srs_next.first);
1763 if (!first || rcu_sr_is_wait_head(first))
1764 return start_new_poll;
1765
1766 wait_head = rcu_sr_get_wait_head();
1767 if (!wait_head) {
1768 // Kick another GP to retry.
1769 start_new_poll = true;
1770 return start_new_poll;
1771 }
1772
1773 /* Inject a wait-dummy-node. */
1774 llist_add(wait_head, &rcu_state.srs_next);
1775
1776 /*
1777 * A waiting list of rcu_synchronize nodes should be empty on
1778 * this step, since a GP-kthread, rcu_gp_init() -> gp_cleanup(),
1779 * rolls it over. If not, it is a BUG, warn a user.
1780 */
1781 WARN_ON_ONCE(rcu_state.srs_wait_tail != NULL);
1782 rcu_state.srs_wait_tail = wait_head;
1783 ASSERT_EXCLUSIVE_WRITER(rcu_state.srs_wait_tail);
1784
1785 return start_new_poll;
1786 }
1787
rcu_sr_normal_add_req(struct rcu_synchronize * rs)1788 static void rcu_sr_normal_add_req(struct rcu_synchronize *rs)
1789 {
1790 llist_add((struct llist_node *) &rs->head, &rcu_state.srs_next);
1791 }
1792
1793 /*
1794 * Initialize a new grace period. Return false if no grace period required.
1795 */
rcu_gp_init(void)1796 static noinline_for_stack bool rcu_gp_init(void)
1797 {
1798 unsigned long flags;
1799 unsigned long oldmask;
1800 unsigned long mask;
1801 struct rcu_data *rdp;
1802 struct rcu_node *rnp = rcu_get_root();
1803 bool start_new_poll;
1804
1805 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1806 raw_spin_lock_irq_rcu_node(rnp);
1807 if (!rcu_state.gp_flags) {
1808 /* Spurious wakeup, tell caller to go back to sleep. */
1809 raw_spin_unlock_irq_rcu_node(rnp);
1810 return false;
1811 }
1812 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1813
1814 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1815 /*
1816 * Grace period already in progress, don't start another.
1817 * Not supposed to be able to happen.
1818 */
1819 raw_spin_unlock_irq_rcu_node(rnp);
1820 return false;
1821 }
1822
1823 /* Advance to a new grace period and initialize state. */
1824 record_gp_stall_check_time();
1825 /* Record GP times before starting GP, hence rcu_seq_start(). */
1826 rcu_seq_start(&rcu_state.gp_seq);
1827 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1828 start_new_poll = rcu_sr_normal_gp_init();
1829 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1830 rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap);
1831 raw_spin_unlock_irq_rcu_node(rnp);
1832
1833 /*
1834 * The "start_new_poll" is set to true, only when this GP is not able
1835 * to handle anything and there are outstanding users. It happens when
1836 * the rcu_sr_normal_gp_init() function was not able to insert a dummy
1837 * separator to the llist, because there were no left any dummy-nodes.
1838 *
1839 * Number of dummy-nodes is fixed, it could be that we are run out of
1840 * them, if so we start a new pool request to repeat a try. It is rare
1841 * and it means that a system is doing a slow processing of callbacks.
1842 */
1843 if (start_new_poll)
1844 (void) start_poll_synchronize_rcu();
1845
1846 /*
1847 * Apply per-leaf buffered online and offline operations to
1848 * the rcu_node tree. Note that this new grace period need not
1849 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1850 * offlining path, when combined with checks in this function,
1851 * will handle CPUs that are currently going offline or that will
1852 * go offline later. Please also refer to "Hotplug CPU" section
1853 * of RCU's Requirements documentation.
1854 */
1855 WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
1856 /* Exclude CPU hotplug operations. */
1857 rcu_for_each_leaf_node(rnp) {
1858 local_irq_disable();
1859 arch_spin_lock(&rcu_state.ofl_lock);
1860 raw_spin_lock_rcu_node(rnp);
1861 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1862 !rnp->wait_blkd_tasks) {
1863 /* Nothing to do on this leaf rcu_node structure. */
1864 raw_spin_unlock_rcu_node(rnp);
1865 arch_spin_unlock(&rcu_state.ofl_lock);
1866 local_irq_enable();
1867 continue;
1868 }
1869
1870 /* Record old state, apply changes to ->qsmaskinit field. */
1871 oldmask = rnp->qsmaskinit;
1872 rnp->qsmaskinit = rnp->qsmaskinitnext;
1873
1874 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1875 if (!oldmask != !rnp->qsmaskinit) {
1876 if (!oldmask) { /* First online CPU for rcu_node. */
1877 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1878 rcu_init_new_rnp(rnp);
1879 } else if (rcu_preempt_has_tasks(rnp)) {
1880 rnp->wait_blkd_tasks = true; /* blocked tasks */
1881 } else { /* Last offline CPU and can propagate. */
1882 rcu_cleanup_dead_rnp(rnp);
1883 }
1884 }
1885
1886 /*
1887 * If all waited-on tasks from prior grace period are
1888 * done, and if all this rcu_node structure's CPUs are
1889 * still offline, propagate up the rcu_node tree and
1890 * clear ->wait_blkd_tasks. Otherwise, if one of this
1891 * rcu_node structure's CPUs has since come back online,
1892 * simply clear ->wait_blkd_tasks.
1893 */
1894 if (rnp->wait_blkd_tasks &&
1895 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1896 rnp->wait_blkd_tasks = false;
1897 if (!rnp->qsmaskinit)
1898 rcu_cleanup_dead_rnp(rnp);
1899 }
1900
1901 raw_spin_unlock_rcu_node(rnp);
1902 arch_spin_unlock(&rcu_state.ofl_lock);
1903 local_irq_enable();
1904 }
1905 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1906
1907 /*
1908 * Set the quiescent-state-needed bits in all the rcu_node
1909 * structures for all currently online CPUs in breadth-first
1910 * order, starting from the root rcu_node structure, relying on the
1911 * layout of the tree within the rcu_state.node[] array. Note that
1912 * other CPUs will access only the leaves of the hierarchy, thus
1913 * seeing that no grace period is in progress, at least until the
1914 * corresponding leaf node has been initialized.
1915 *
1916 * The grace period cannot complete until the initialization
1917 * process finishes, because this kthread handles both.
1918 */
1919 WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1920 rcu_for_each_node_breadth_first(rnp) {
1921 rcu_gp_slow(gp_init_delay);
1922 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1923 rdp = this_cpu_ptr(&rcu_data);
1924 rcu_preempt_check_blocked_tasks(rnp);
1925 rnp->qsmask = rnp->qsmaskinit;
1926 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1927 if (rnp == rdp->mynode)
1928 (void)__note_gp_changes(rnp, rdp);
1929 rcu_preempt_boost_start_gp(rnp);
1930 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1931 rnp->level, rnp->grplo,
1932 rnp->grphi, rnp->qsmask);
1933 /* Quiescent states for tasks on any now-offline CPUs. */
1934 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1935 rnp->rcu_gp_init_mask = mask;
1936 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1937 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1938 else
1939 raw_spin_unlock_irq_rcu_node(rnp);
1940 cond_resched_tasks_rcu_qs();
1941 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1942 }
1943
1944 // If strict, make all CPUs aware of new grace period.
1945 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1946 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1947
1948 return true;
1949 }
1950
1951 /*
1952 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1953 * time.
1954 */
rcu_gp_fqs_check_wake(int * gfp)1955 static bool rcu_gp_fqs_check_wake(int *gfp)
1956 {
1957 struct rcu_node *rnp = rcu_get_root();
1958
1959 // If under overload conditions, force an immediate FQS scan.
1960 if (*gfp & RCU_GP_FLAG_OVLD)
1961 return true;
1962
1963 // Someone like call_rcu() requested a force-quiescent-state scan.
1964 *gfp = READ_ONCE(rcu_state.gp_flags);
1965 if (*gfp & RCU_GP_FLAG_FQS)
1966 return true;
1967
1968 // The current grace period has completed.
1969 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1970 return true;
1971
1972 return false;
1973 }
1974
1975 /*
1976 * Do one round of quiescent-state forcing.
1977 */
rcu_gp_fqs(bool first_time)1978 static void rcu_gp_fqs(bool first_time)
1979 {
1980 int nr_fqs = READ_ONCE(rcu_state.nr_fqs_jiffies_stall);
1981 struct rcu_node *rnp = rcu_get_root();
1982
1983 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1984 WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1);
1985
1986 WARN_ON_ONCE(nr_fqs > 3);
1987 /* Only countdown nr_fqs for stall purposes if jiffies moves. */
1988 if (nr_fqs) {
1989 if (nr_fqs == 1) {
1990 WRITE_ONCE(rcu_state.jiffies_stall,
1991 jiffies + rcu_jiffies_till_stall_check());
1992 }
1993 WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, --nr_fqs);
1994 }
1995
1996 if (first_time) {
1997 /* Collect dyntick-idle snapshots. */
1998 force_qs_rnp(rcu_watching_snap_save);
1999 } else {
2000 /* Handle dyntick-idle and offline CPUs. */
2001 force_qs_rnp(rcu_watching_snap_recheck);
2002 }
2003 /* Clear flag to prevent immediate re-entry. */
2004 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2005 raw_spin_lock_irq_rcu_node(rnp);
2006 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & ~RCU_GP_FLAG_FQS);
2007 raw_spin_unlock_irq_rcu_node(rnp);
2008 }
2009 }
2010
2011 /*
2012 * Loop doing repeated quiescent-state forcing until the grace period ends.
2013 */
rcu_gp_fqs_loop(void)2014 static noinline_for_stack void rcu_gp_fqs_loop(void)
2015 {
2016 bool first_gp_fqs = true;
2017 int gf = 0;
2018 unsigned long j;
2019 int ret;
2020 struct rcu_node *rnp = rcu_get_root();
2021
2022 j = READ_ONCE(jiffies_till_first_fqs);
2023 if (rcu_state.cbovld)
2024 gf = RCU_GP_FLAG_OVLD;
2025 ret = 0;
2026 for (;;) {
2027 if (rcu_state.cbovld) {
2028 j = (j + 2) / 3;
2029 if (j <= 0)
2030 j = 1;
2031 }
2032 if (!ret || time_before(jiffies + j, rcu_state.jiffies_force_qs)) {
2033 WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
2034 /*
2035 * jiffies_force_qs before RCU_GP_WAIT_FQS state
2036 * update; required for stall checks.
2037 */
2038 smp_wmb();
2039 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
2040 jiffies + (j ? 3 * j : 2));
2041 }
2042 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2043 TPS("fqswait"));
2044 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
2045 (void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq,
2046 rcu_gp_fqs_check_wake(&gf), j);
2047 rcu_gp_torture_wait();
2048 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
2049 /* Locking provides needed memory barriers. */
2050 /*
2051 * Exit the loop if the root rcu_node structure indicates that the grace period
2052 * has ended, leave the loop. The rcu_preempt_blocked_readers_cgp(rnp) check
2053 * is required only for single-node rcu_node trees because readers blocking
2054 * the current grace period are queued only on leaf rcu_node structures.
2055 * For multi-node trees, checking the root node's ->qsmask suffices, because a
2056 * given root node's ->qsmask bit is cleared only when all CPUs and tasks from
2057 * the corresponding leaf nodes have passed through their quiescent state.
2058 */
2059 if (!READ_ONCE(rnp->qsmask) &&
2060 !rcu_preempt_blocked_readers_cgp(rnp))
2061 break;
2062 /* If time for quiescent-state forcing, do it. */
2063 if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
2064 (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
2065 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2066 TPS("fqsstart"));
2067 rcu_gp_fqs(first_gp_fqs);
2068 gf = 0;
2069 if (first_gp_fqs) {
2070 first_gp_fqs = false;
2071 gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
2072 }
2073 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2074 TPS("fqsend"));
2075 cond_resched_tasks_rcu_qs();
2076 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2077 ret = 0; /* Force full wait till next FQS. */
2078 j = READ_ONCE(jiffies_till_next_fqs);
2079 } else {
2080 /* Deal with stray signal. */
2081 cond_resched_tasks_rcu_qs();
2082 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2083 WARN_ON(signal_pending(current));
2084 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2085 TPS("fqswaitsig"));
2086 ret = 1; /* Keep old FQS timing. */
2087 j = jiffies;
2088 if (time_after(jiffies, rcu_state.jiffies_force_qs))
2089 j = 1;
2090 else
2091 j = rcu_state.jiffies_force_qs - j;
2092 gf = 0;
2093 }
2094 }
2095 }
2096
2097 /*
2098 * Clean up after the old grace period.
2099 */
rcu_gp_cleanup(void)2100 static noinline void rcu_gp_cleanup(void)
2101 {
2102 int cpu;
2103 bool needgp = false;
2104 unsigned long gp_duration;
2105 unsigned long new_gp_seq;
2106 bool offloaded;
2107 struct rcu_data *rdp;
2108 struct rcu_node *rnp = rcu_get_root();
2109 struct swait_queue_head *sq;
2110
2111 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2112 raw_spin_lock_irq_rcu_node(rnp);
2113 rcu_state.gp_end = jiffies;
2114 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
2115 if (gp_duration > rcu_state.gp_max)
2116 rcu_state.gp_max = gp_duration;
2117
2118 /*
2119 * We know the grace period is complete, but to everyone else
2120 * it appears to still be ongoing. But it is also the case
2121 * that to everyone else it looks like there is nothing that
2122 * they can do to advance the grace period. It is therefore
2123 * safe for us to drop the lock in order to mark the grace
2124 * period as completed in all of the rcu_node structures.
2125 */
2126 rcu_poll_gp_seq_end(&rcu_state.gp_seq_polled_snap);
2127 raw_spin_unlock_irq_rcu_node(rnp);
2128
2129 /*
2130 * Propagate new ->gp_seq value to rcu_node structures so that
2131 * other CPUs don't have to wait until the start of the next grace
2132 * period to process their callbacks. This also avoids some nasty
2133 * RCU grace-period initialization races by forcing the end of
2134 * the current grace period to be completely recorded in all of
2135 * the rcu_node structures before the beginning of the next grace
2136 * period is recorded in any of the rcu_node structures.
2137 */
2138 new_gp_seq = rcu_state.gp_seq;
2139 rcu_seq_end(&new_gp_seq);
2140 rcu_for_each_node_breadth_first(rnp) {
2141 raw_spin_lock_irq_rcu_node(rnp);
2142 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2143 dump_blkd_tasks(rnp, 10);
2144 WARN_ON_ONCE(rnp->qsmask);
2145 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2146 if (!rnp->parent)
2147 smp_mb(); // Order against failing poll_state_synchronize_rcu_full().
2148 rdp = this_cpu_ptr(&rcu_data);
2149 if (rnp == rdp->mynode)
2150 needgp = __note_gp_changes(rnp, rdp) || needgp;
2151 /* smp_mb() provided by prior unlock-lock pair. */
2152 needgp = rcu_future_gp_cleanup(rnp) || needgp;
2153 // Reset overload indication for CPUs no longer overloaded
2154 if (rcu_is_leaf_node(rnp))
2155 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
2156 rdp = per_cpu_ptr(&rcu_data, cpu);
2157 check_cb_ovld_locked(rdp, rnp);
2158 }
2159 sq = rcu_nocb_gp_get(rnp);
2160 raw_spin_unlock_irq_rcu_node(rnp);
2161 rcu_nocb_gp_cleanup(sq);
2162 cond_resched_tasks_rcu_qs();
2163 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2164 rcu_gp_slow(gp_cleanup_delay);
2165 }
2166 rnp = rcu_get_root();
2167 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
2168
2169 /* Declare grace period done, trace first to use old GP number. */
2170 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
2171 rcu_seq_end(&rcu_state.gp_seq);
2172 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
2173 WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
2174 /* Check for GP requests since above loop. */
2175 rdp = this_cpu_ptr(&rcu_data);
2176 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2177 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2178 TPS("CleanupMore"));
2179 needgp = true;
2180 }
2181 /* Advance CBs to reduce false positives below. */
2182 offloaded = rcu_rdp_is_offloaded(rdp);
2183 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
2184
2185 // We get here if a grace period was needed (“needgp”)
2186 // and the above call to rcu_accelerate_cbs() did not set
2187 // the RCU_GP_FLAG_INIT bit in ->gp_state (which records
2188 // the need for another grace period). The purpose
2189 // of the “offloaded” check is to avoid invoking
2190 // rcu_accelerate_cbs() on an offloaded CPU because we do not
2191 // hold the ->nocb_lock needed to safely access an offloaded
2192 // ->cblist. We do not want to acquire that lock because
2193 // it can be heavily contended during callback floods.
2194
2195 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
2196 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
2197 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq"));
2198 } else {
2199
2200 // We get here either if there is no need for an
2201 // additional grace period or if rcu_accelerate_cbs() has
2202 // already set the RCU_GP_FLAG_INIT bit in ->gp_flags.
2203 // So all we need to do is to clear all of the other
2204 // ->gp_flags bits.
2205
2206 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT);
2207 }
2208 raw_spin_unlock_irq_rcu_node(rnp);
2209
2210 // Make synchronize_rcu() users aware of the end of old grace period.
2211 rcu_sr_normal_gp_cleanup();
2212
2213 // If strict, make all CPUs aware of the end of the old grace period.
2214 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2215 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
2216 }
2217
2218 /*
2219 * Body of kthread that handles grace periods.
2220 */
rcu_gp_kthread(void * unused)2221 static int __noreturn rcu_gp_kthread(void *unused)
2222 {
2223 rcu_bind_gp_kthread();
2224 for (;;) {
2225
2226 /* Handle grace-period start. */
2227 for (;;) {
2228 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2229 TPS("reqwait"));
2230 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
2231 swait_event_idle_exclusive(rcu_state.gp_wq,
2232 READ_ONCE(rcu_state.gp_flags) &
2233 RCU_GP_FLAG_INIT);
2234 rcu_gp_torture_wait();
2235 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
2236 /* Locking provides needed memory barrier. */
2237 if (rcu_gp_init())
2238 break;
2239 cond_resched_tasks_rcu_qs();
2240 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2241 WARN_ON(signal_pending(current));
2242 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2243 TPS("reqwaitsig"));
2244 }
2245
2246 /* Handle quiescent-state forcing. */
2247 rcu_gp_fqs_loop();
2248
2249 /* Handle grace-period end. */
2250 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
2251 rcu_gp_cleanup();
2252 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
2253 }
2254 }
2255
2256 /*
2257 * Report a full set of quiescent states to the rcu_state data structure.
2258 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
2259 * another grace period is required. Whether we wake the grace-period
2260 * kthread or it awakens itself for the next round of quiescent-state
2261 * forcing, that kthread will clean up after the just-completed grace
2262 * period. Note that the caller must hold rnp->lock, which is released
2263 * before return.
2264 */
rcu_report_qs_rsp(unsigned long flags)2265 static void rcu_report_qs_rsp(unsigned long flags)
2266 __releases(rcu_get_root()->lock)
2267 {
2268 raw_lockdep_assert_held_rcu_node(rcu_get_root());
2269 WARN_ON_ONCE(!rcu_gp_in_progress());
2270 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_FQS);
2271 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
2272 rcu_gp_kthread_wake();
2273 }
2274
2275 /*
2276 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2277 * Allows quiescent states for a group of CPUs to be reported at one go
2278 * to the specified rcu_node structure, though all the CPUs in the group
2279 * must be represented by the same rcu_node structure (which need not be a
2280 * leaf rcu_node structure, though it often will be). The gps parameter
2281 * is the grace-period snapshot, which means that the quiescent states
2282 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2283 * must be held upon entry, and it is released before return.
2284 *
2285 * As a special case, if mask is zero, the bit-already-cleared check is
2286 * disabled. This allows propagating quiescent state due to resumed tasks
2287 * during grace-period initialization.
2288 */
rcu_report_qs_rnp(unsigned long mask,struct rcu_node * rnp,unsigned long gps,unsigned long flags)2289 static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
2290 unsigned long gps, unsigned long flags)
2291 __releases(rnp->lock)
2292 {
2293 unsigned long oldmask = 0;
2294 struct rcu_node *rnp_c;
2295
2296 raw_lockdep_assert_held_rcu_node(rnp);
2297
2298 /* Walk up the rcu_node hierarchy. */
2299 for (;;) {
2300 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2301
2302 /*
2303 * Our bit has already been cleared, or the
2304 * relevant grace period is already over, so done.
2305 */
2306 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2307 return;
2308 }
2309 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2310 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2311 rcu_preempt_blocked_readers_cgp(rnp));
2312 WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
2313 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
2314 mask, rnp->qsmask, rnp->level,
2315 rnp->grplo, rnp->grphi,
2316 !!rnp->gp_tasks);
2317 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2318
2319 /* Other bits still set at this level, so done. */
2320 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2321 return;
2322 }
2323 rnp->completedqs = rnp->gp_seq;
2324 mask = rnp->grpmask;
2325 if (rnp->parent == NULL) {
2326
2327 /* No more levels. Exit loop holding root lock. */
2328
2329 break;
2330 }
2331 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2332 rnp_c = rnp;
2333 rnp = rnp->parent;
2334 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2335 oldmask = READ_ONCE(rnp_c->qsmask);
2336 }
2337
2338 /*
2339 * Get here if we are the last CPU to pass through a quiescent
2340 * state for this grace period. Invoke rcu_report_qs_rsp()
2341 * to clean up and start the next grace period if one is needed.
2342 */
2343 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
2344 }
2345
2346 /*
2347 * Record a quiescent state for all tasks that were previously queued
2348 * on the specified rcu_node structure and that were blocking the current
2349 * RCU grace period. The caller must hold the corresponding rnp->lock with
2350 * irqs disabled, and this lock is released upon return, but irqs remain
2351 * disabled.
2352 */
2353 static void __maybe_unused
rcu_report_unblock_qs_rnp(struct rcu_node * rnp,unsigned long flags)2354 rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
2355 __releases(rnp->lock)
2356 {
2357 unsigned long gps;
2358 unsigned long mask;
2359 struct rcu_node *rnp_p;
2360
2361 raw_lockdep_assert_held_rcu_node(rnp);
2362 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
2363 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2364 rnp->qsmask != 0) {
2365 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2366 return; /* Still need more quiescent states! */
2367 }
2368
2369 rnp->completedqs = rnp->gp_seq;
2370 rnp_p = rnp->parent;
2371 if (rnp_p == NULL) {
2372 /*
2373 * Only one rcu_node structure in the tree, so don't
2374 * try to report up to its nonexistent parent!
2375 */
2376 rcu_report_qs_rsp(flags);
2377 return;
2378 }
2379
2380 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2381 gps = rnp->gp_seq;
2382 mask = rnp->grpmask;
2383 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2384 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2385 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2386 }
2387
2388 /*
2389 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2390 * structure. This must be called from the specified CPU.
2391 */
2392 static void
rcu_report_qs_rdp(struct rcu_data * rdp)2393 rcu_report_qs_rdp(struct rcu_data *rdp)
2394 {
2395 unsigned long flags;
2396 unsigned long mask;
2397 struct rcu_node *rnp;
2398
2399 WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2400 rnp = rdp->mynode;
2401 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2402 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2403 rdp->gpwrap) {
2404
2405 /*
2406 * The grace period in which this quiescent state was
2407 * recorded has ended, so don't report it upwards.
2408 * We will instead need a new quiescent state that lies
2409 * within the current grace period.
2410 */
2411 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2412 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2413 return;
2414 }
2415 mask = rdp->grpmask;
2416 rdp->core_needs_qs = false;
2417 if ((rnp->qsmask & mask) == 0) {
2418 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2419 } else {
2420 /*
2421 * This GP can't end until cpu checks in, so all of our
2422 * callbacks can be processed during the next GP.
2423 *
2424 * NOCB kthreads have their own way to deal with that...
2425 */
2426 if (!rcu_rdp_is_offloaded(rdp)) {
2427 /*
2428 * The current GP has not yet ended, so it
2429 * should not be possible for rcu_accelerate_cbs()
2430 * to return true. So complain, but don't awaken.
2431 */
2432 WARN_ON_ONCE(rcu_accelerate_cbs(rnp, rdp));
2433 }
2434
2435 rcu_disable_urgency_upon_qs(rdp);
2436 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2437 /* ^^^ Released rnp->lock */
2438 }
2439 }
2440
2441 /*
2442 * Check to see if there is a new grace period of which this CPU
2443 * is not yet aware, and if so, set up local rcu_data state for it.
2444 * Otherwise, see if this CPU has just passed through its first
2445 * quiescent state for this grace period, and record that fact if so.
2446 */
2447 static void
rcu_check_quiescent_state(struct rcu_data * rdp)2448 rcu_check_quiescent_state(struct rcu_data *rdp)
2449 {
2450 /* Check for grace-period ends and beginnings. */
2451 note_gp_changes(rdp);
2452
2453 /*
2454 * Does this CPU still need to do its part for current grace period?
2455 * If no, return and let the other CPUs do their part as well.
2456 */
2457 if (!rdp->core_needs_qs)
2458 return;
2459
2460 /*
2461 * Was there a quiescent state since the beginning of the grace
2462 * period? If no, then exit and wait for the next call.
2463 */
2464 if (rdp->cpu_no_qs.b.norm)
2465 return;
2466
2467 /*
2468 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2469 * judge of that).
2470 */
2471 rcu_report_qs_rdp(rdp);
2472 }
2473
2474 /* Return true if callback-invocation time limit exceeded. */
rcu_do_batch_check_time(long count,long tlimit,bool jlimit_check,unsigned long jlimit)2475 static bool rcu_do_batch_check_time(long count, long tlimit,
2476 bool jlimit_check, unsigned long jlimit)
2477 {
2478 // Invoke local_clock() only once per 32 consecutive callbacks.
2479 return unlikely(tlimit) &&
2480 (!likely(count & 31) ||
2481 (IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) &&
2482 jlimit_check && time_after(jiffies, jlimit))) &&
2483 local_clock() >= tlimit;
2484 }
2485
2486 /*
2487 * Invoke any RCU callbacks that have made it to the end of their grace
2488 * period. Throttle as specified by rdp->blimit.
2489 */
rcu_do_batch(struct rcu_data * rdp)2490 static void rcu_do_batch(struct rcu_data *rdp)
2491 {
2492 long bl;
2493 long count = 0;
2494 int div;
2495 bool __maybe_unused empty;
2496 unsigned long flags;
2497 unsigned long jlimit;
2498 bool jlimit_check = false;
2499 long pending;
2500 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2501 struct rcu_head *rhp;
2502 long tlimit = 0;
2503
2504 /* If no callbacks are ready, just return. */
2505 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2506 trace_rcu_batch_start(rcu_state.name,
2507 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2508 trace_rcu_batch_end(rcu_state.name, 0,
2509 !rcu_segcblist_empty(&rdp->cblist),
2510 need_resched(), is_idle_task(current),
2511 rcu_is_callbacks_kthread(rdp));
2512 return;
2513 }
2514
2515 /*
2516 * Extract the list of ready callbacks, disabling IRQs to prevent
2517 * races with call_rcu() from interrupt handlers. Leave the
2518 * callback counts, as rcu_barrier() needs to be conservative.
2519 *
2520 * Callbacks execution is fully ordered against preceding grace period
2521 * completion (materialized by rnp->gp_seq update) thanks to the
2522 * smp_mb__after_unlock_lock() upon node locking required for callbacks
2523 * advancing. In NOCB mode this ordering is then further relayed through
2524 * the nocb locking that protects both callbacks advancing and extraction.
2525 */
2526 rcu_nocb_lock_irqsave(rdp, flags);
2527 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2528 pending = rcu_segcblist_get_seglen(&rdp->cblist, RCU_DONE_TAIL);
2529 div = READ_ONCE(rcu_divisor);
2530 div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2531 bl = max(rdp->blimit, pending >> div);
2532 if ((in_serving_softirq() || rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING) &&
2533 (IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) || unlikely(bl > 100))) {
2534 const long npj = NSEC_PER_SEC / HZ;
2535 long rrn = READ_ONCE(rcu_resched_ns);
2536
2537 rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2538 tlimit = local_clock() + rrn;
2539 jlimit = jiffies + (rrn + npj + 1) / npj;
2540 jlimit_check = true;
2541 }
2542 trace_rcu_batch_start(rcu_state.name,
2543 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2544 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2545 if (rcu_rdp_is_offloaded(rdp))
2546 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2547
2548 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2549 rcu_nocb_unlock_irqrestore(rdp, flags);
2550
2551 /* Invoke callbacks. */
2552 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2553 rhp = rcu_cblist_dequeue(&rcl);
2554
2555 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2556 rcu_callback_t f;
2557
2558 count++;
2559 debug_rcu_head_unqueue(rhp);
2560
2561 rcu_lock_acquire(&rcu_callback_map);
2562 trace_rcu_invoke_callback(rcu_state.name, rhp);
2563
2564 f = rhp->func;
2565 debug_rcu_head_callback(rhp);
2566 WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2567 f(rhp);
2568
2569 rcu_lock_release(&rcu_callback_map);
2570
2571 /*
2572 * Stop only if limit reached and CPU has something to do.
2573 */
2574 if (in_serving_softirq()) {
2575 if (count >= bl && (need_resched() || !is_idle_task(current)))
2576 break;
2577 /*
2578 * Make sure we don't spend too much time here and deprive other
2579 * softirq vectors of CPU cycles.
2580 */
2581 if (rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit))
2582 break;
2583 } else {
2584 // In rcuc/rcuoc context, so no worries about
2585 // depriving other softirq vectors of CPU cycles.
2586 local_bh_enable();
2587 lockdep_assert_irqs_enabled();
2588 cond_resched_tasks_rcu_qs();
2589 lockdep_assert_irqs_enabled();
2590 local_bh_disable();
2591 // But rcuc kthreads can delay quiescent-state
2592 // reporting, so check time limits for them.
2593 if (rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING &&
2594 rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit)) {
2595 rdp->rcu_cpu_has_work = 1;
2596 break;
2597 }
2598 }
2599 }
2600
2601 rcu_nocb_lock_irqsave(rdp, flags);
2602 rdp->n_cbs_invoked += count;
2603 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2604 is_idle_task(current), rcu_is_callbacks_kthread(rdp));
2605
2606 /* Update counts and requeue any remaining callbacks. */
2607 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2608 rcu_segcblist_add_len(&rdp->cblist, -count);
2609
2610 /* Reinstate batch limit if we have worked down the excess. */
2611 count = rcu_segcblist_n_cbs(&rdp->cblist);
2612 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2613 rdp->blimit = blimit;
2614
2615 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2616 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2617 rdp->qlen_last_fqs_check = 0;
2618 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
2619 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2620 rdp->qlen_last_fqs_check = count;
2621
2622 /*
2623 * The following usually indicates a double call_rcu(). To track
2624 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2625 */
2626 empty = rcu_segcblist_empty(&rdp->cblist);
2627 WARN_ON_ONCE(count == 0 && !empty);
2628 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2629 count != 0 && empty);
2630 WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2631 WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2632
2633 rcu_nocb_unlock_irqrestore(rdp, flags);
2634
2635 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2636 }
2637
2638 /*
2639 * This function is invoked from each scheduling-clock interrupt,
2640 * and checks to see if this CPU is in a non-context-switch quiescent
2641 * state, for example, user mode or idle loop. It also schedules RCU
2642 * core processing. If the current grace period has gone on too long,
2643 * it will ask the scheduler to manufacture a context switch for the sole
2644 * purpose of providing the needed quiescent state.
2645 */
rcu_sched_clock_irq(int user)2646 void rcu_sched_clock_irq(int user)
2647 {
2648 unsigned long j;
2649
2650 if (IS_ENABLED(CONFIG_PROVE_RCU)) {
2651 j = jiffies;
2652 WARN_ON_ONCE(time_before(j, __this_cpu_read(rcu_data.last_sched_clock)));
2653 __this_cpu_write(rcu_data.last_sched_clock, j);
2654 }
2655 trace_rcu_utilization(TPS("Start scheduler-tick"));
2656 lockdep_assert_irqs_disabled();
2657 raw_cpu_inc(rcu_data.ticks_this_gp);
2658 /* The load-acquire pairs with the store-release setting to true. */
2659 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2660 /* Idle and userspace execution already are quiescent states. */
2661 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2662 set_tsk_need_resched(current);
2663 set_preempt_need_resched();
2664 }
2665 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2666 }
2667 rcu_flavor_sched_clock_irq(user);
2668 if (rcu_pending(user))
2669 invoke_rcu_core();
2670 if (user || rcu_is_cpu_rrupt_from_idle())
2671 rcu_note_voluntary_context_switch(current);
2672 lockdep_assert_irqs_disabled();
2673
2674 trace_rcu_utilization(TPS("End scheduler-tick"));
2675 }
2676
2677 /*
2678 * Scan the leaf rcu_node structures. For each structure on which all
2679 * CPUs have reported a quiescent state and on which there are tasks
2680 * blocking the current grace period, initiate RCU priority boosting.
2681 * Otherwise, invoke the specified function to check dyntick state for
2682 * each CPU that has not yet reported a quiescent state.
2683 */
force_qs_rnp(int (* f)(struct rcu_data * rdp))2684 static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2685 {
2686 int cpu;
2687 unsigned long flags;
2688 struct rcu_node *rnp;
2689
2690 rcu_state.cbovld = rcu_state.cbovldnext;
2691 rcu_state.cbovldnext = false;
2692 rcu_for_each_leaf_node(rnp) {
2693 unsigned long mask = 0;
2694 unsigned long rsmask = 0;
2695
2696 cond_resched_tasks_rcu_qs();
2697 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2698 rcu_state.cbovldnext |= !!rnp->cbovldmask;
2699 if (rnp->qsmask == 0) {
2700 if (rcu_preempt_blocked_readers_cgp(rnp)) {
2701 /*
2702 * No point in scanning bits because they
2703 * are all zero. But we might need to
2704 * priority-boost blocked readers.
2705 */
2706 rcu_initiate_boost(rnp, flags);
2707 /* rcu_initiate_boost() releases rnp->lock */
2708 continue;
2709 }
2710 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2711 continue;
2712 }
2713 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2714 struct rcu_data *rdp;
2715 int ret;
2716
2717 rdp = per_cpu_ptr(&rcu_data, cpu);
2718 ret = f(rdp);
2719 if (ret > 0) {
2720 mask |= rdp->grpmask;
2721 rcu_disable_urgency_upon_qs(rdp);
2722 }
2723 if (ret < 0)
2724 rsmask |= rdp->grpmask;
2725 }
2726 if (mask != 0) {
2727 /* Idle/offline CPUs, report (releases rnp->lock). */
2728 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2729 } else {
2730 /* Nothing to do here, so just drop the lock. */
2731 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2732 }
2733
2734 for_each_leaf_node_cpu_mask(rnp, cpu, rsmask)
2735 resched_cpu(cpu);
2736 }
2737 }
2738
2739 /*
2740 * Force quiescent states on reluctant CPUs, and also detect which
2741 * CPUs are in dyntick-idle mode.
2742 */
rcu_force_quiescent_state(void)2743 void rcu_force_quiescent_state(void)
2744 {
2745 unsigned long flags;
2746 bool ret;
2747 struct rcu_node *rnp;
2748 struct rcu_node *rnp_old = NULL;
2749
2750 if (!rcu_gp_in_progress())
2751 return;
2752 /* Funnel through hierarchy to reduce memory contention. */
2753 rnp = raw_cpu_read(rcu_data.mynode);
2754 for (; rnp != NULL; rnp = rnp->parent) {
2755 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2756 !raw_spin_trylock(&rnp->fqslock);
2757 if (rnp_old != NULL)
2758 raw_spin_unlock(&rnp_old->fqslock);
2759 if (ret)
2760 return;
2761 rnp_old = rnp;
2762 }
2763 /* rnp_old == rcu_get_root(), rnp == NULL. */
2764
2765 /* Reached the root of the rcu_node tree, acquire lock. */
2766 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2767 raw_spin_unlock(&rnp_old->fqslock);
2768 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2769 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2770 return; /* Someone beat us to it. */
2771 }
2772 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_FQS);
2773 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2774 rcu_gp_kthread_wake();
2775 }
2776 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2777
2778 // Workqueue handler for an RCU reader for kernels enforcing struct RCU
2779 // grace periods.
strict_work_handler(struct work_struct * work)2780 static void strict_work_handler(struct work_struct *work)
2781 {
2782 rcu_read_lock();
2783 rcu_read_unlock();
2784 }
2785
2786 /* Perform RCU core processing work for the current CPU. */
rcu_core(void)2787 static __latent_entropy void rcu_core(void)
2788 {
2789 unsigned long flags;
2790 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2791 struct rcu_node *rnp = rdp->mynode;
2792
2793 if (cpu_is_offline(smp_processor_id()))
2794 return;
2795 trace_rcu_utilization(TPS("Start RCU core"));
2796 WARN_ON_ONCE(!rdp->beenonline);
2797
2798 /* Report any deferred quiescent states if preemption enabled. */
2799 if (IS_ENABLED(CONFIG_PREEMPT_COUNT) && (!(preempt_count() & PREEMPT_MASK))) {
2800 rcu_preempt_deferred_qs(current);
2801 } else if (rcu_preempt_need_deferred_qs(current)) {
2802 set_tsk_need_resched(current);
2803 set_preempt_need_resched();
2804 }
2805
2806 /* Update RCU state based on any recent quiescent states. */
2807 rcu_check_quiescent_state(rdp);
2808
2809 /* No grace period and unregistered callbacks? */
2810 if (!rcu_gp_in_progress() &&
2811 rcu_segcblist_is_enabled(&rdp->cblist) && !rcu_rdp_is_offloaded(rdp)) {
2812 local_irq_save(flags);
2813 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2814 rcu_accelerate_cbs_unlocked(rnp, rdp);
2815 local_irq_restore(flags);
2816 }
2817
2818 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2819
2820 /* If there are callbacks ready, invoke them. */
2821 if (!rcu_rdp_is_offloaded(rdp) && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2822 likely(READ_ONCE(rcu_scheduler_fully_active))) {
2823 rcu_do_batch(rdp);
2824 /* Re-invoke RCU core processing if there are callbacks remaining. */
2825 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2826 invoke_rcu_core();
2827 }
2828
2829 /* Do any needed deferred wakeups of rcuo kthreads. */
2830 do_nocb_deferred_wakeup(rdp);
2831 trace_rcu_utilization(TPS("End RCU core"));
2832
2833 // If strict GPs, schedule an RCU reader in a clean environment.
2834 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2835 queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2836 }
2837
rcu_core_si(void)2838 static void rcu_core_si(void)
2839 {
2840 rcu_core();
2841 }
2842
rcu_wake_cond(struct task_struct * t,int status)2843 static void rcu_wake_cond(struct task_struct *t, int status)
2844 {
2845 /*
2846 * If the thread is yielding, only wake it when this
2847 * is invoked from idle
2848 */
2849 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2850 wake_up_process(t);
2851 }
2852
invoke_rcu_core_kthread(void)2853 static void invoke_rcu_core_kthread(void)
2854 {
2855 struct task_struct *t;
2856 unsigned long flags;
2857
2858 local_irq_save(flags);
2859 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2860 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2861 if (t != NULL && t != current)
2862 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2863 local_irq_restore(flags);
2864 }
2865
2866 /*
2867 * Wake up this CPU's rcuc kthread to do RCU core processing.
2868 */
invoke_rcu_core(void)2869 static void invoke_rcu_core(void)
2870 {
2871 if (!cpu_online(smp_processor_id()))
2872 return;
2873 if (use_softirq)
2874 raise_softirq(RCU_SOFTIRQ);
2875 else
2876 invoke_rcu_core_kthread();
2877 }
2878
rcu_cpu_kthread_park(unsigned int cpu)2879 static void rcu_cpu_kthread_park(unsigned int cpu)
2880 {
2881 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2882 }
2883
rcu_cpu_kthread_should_run(unsigned int cpu)2884 static int rcu_cpu_kthread_should_run(unsigned int cpu)
2885 {
2886 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2887 }
2888
2889 /*
2890 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2891 * the RCU softirq used in configurations of RCU that do not support RCU
2892 * priority boosting.
2893 */
rcu_cpu_kthread(unsigned int cpu)2894 static void rcu_cpu_kthread(unsigned int cpu)
2895 {
2896 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2897 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2898 unsigned long *j = this_cpu_ptr(&rcu_data.rcuc_activity);
2899 int spincnt;
2900
2901 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2902 for (spincnt = 0; spincnt < 10; spincnt++) {
2903 WRITE_ONCE(*j, jiffies);
2904 local_bh_disable();
2905 *statusp = RCU_KTHREAD_RUNNING;
2906 local_irq_disable();
2907 work = *workp;
2908 WRITE_ONCE(*workp, 0);
2909 local_irq_enable();
2910 if (work)
2911 rcu_core();
2912 local_bh_enable();
2913 if (!READ_ONCE(*workp)) {
2914 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2915 *statusp = RCU_KTHREAD_WAITING;
2916 return;
2917 }
2918 }
2919 *statusp = RCU_KTHREAD_YIELDING;
2920 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2921 schedule_timeout_idle(2);
2922 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2923 *statusp = RCU_KTHREAD_WAITING;
2924 WRITE_ONCE(*j, jiffies);
2925 }
2926
2927 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2928 .store = &rcu_data.rcu_cpu_kthread_task,
2929 .thread_should_run = rcu_cpu_kthread_should_run,
2930 .thread_fn = rcu_cpu_kthread,
2931 .thread_comm = "rcuc/%u",
2932 .setup = rcu_cpu_kthread_setup,
2933 .park = rcu_cpu_kthread_park,
2934 };
2935
2936 /*
2937 * Spawn per-CPU RCU core processing kthreads.
2938 */
rcu_spawn_core_kthreads(void)2939 static int __init rcu_spawn_core_kthreads(void)
2940 {
2941 int cpu;
2942
2943 for_each_possible_cpu(cpu)
2944 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2945 if (use_softirq)
2946 return 0;
2947 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2948 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2949 return 0;
2950 }
2951
rcutree_enqueue(struct rcu_data * rdp,struct rcu_head * head,rcu_callback_t func)2952 static void rcutree_enqueue(struct rcu_data *rdp, struct rcu_head *head, rcu_callback_t func)
2953 {
2954 rcu_segcblist_enqueue(&rdp->cblist, head);
2955 if (__is_kvfree_rcu_offset((unsigned long)func))
2956 trace_rcu_kvfree_callback(rcu_state.name, head,
2957 (unsigned long)func,
2958 rcu_segcblist_n_cbs(&rdp->cblist));
2959 else
2960 trace_rcu_callback(rcu_state.name, head,
2961 rcu_segcblist_n_cbs(&rdp->cblist));
2962 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
2963 }
2964
2965 /*
2966 * Handle any core-RCU processing required by a call_rcu() invocation.
2967 */
call_rcu_core(struct rcu_data * rdp,struct rcu_head * head,rcu_callback_t func,unsigned long flags)2968 static void call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2969 rcu_callback_t func, unsigned long flags)
2970 {
2971 rcutree_enqueue(rdp, head, func);
2972 /*
2973 * If called from an extended quiescent state, invoke the RCU
2974 * core in order to force a re-evaluation of RCU's idleness.
2975 */
2976 if (!rcu_is_watching())
2977 invoke_rcu_core();
2978
2979 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2980 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2981 return;
2982
2983 /*
2984 * Force the grace period if too many callbacks or too long waiting.
2985 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2986 * if some other CPU has recently done so. Also, don't bother
2987 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2988 * is the only one waiting for a grace period to complete.
2989 */
2990 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2991 rdp->qlen_last_fqs_check + qhimark)) {
2992
2993 /* Are we ignoring a completed grace period? */
2994 note_gp_changes(rdp);
2995
2996 /* Start a new grace period if one not already started. */
2997 if (!rcu_gp_in_progress()) {
2998 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2999 } else {
3000 /* Give the grace period a kick. */
3001 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
3002 if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap &&
3003 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
3004 rcu_force_quiescent_state();
3005 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
3006 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
3007 }
3008 }
3009 }
3010
3011 /*
3012 * RCU callback function to leak a callback.
3013 */
rcu_leak_callback(struct rcu_head * rhp)3014 static void rcu_leak_callback(struct rcu_head *rhp)
3015 {
3016 }
3017
3018 /*
3019 * Check and if necessary update the leaf rcu_node structure's
3020 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
3021 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
3022 * structure's ->lock.
3023 */
check_cb_ovld_locked(struct rcu_data * rdp,struct rcu_node * rnp)3024 static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
3025 {
3026 raw_lockdep_assert_held_rcu_node(rnp);
3027 if (qovld_calc <= 0)
3028 return; // Early boot and wildcard value set.
3029 if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
3030 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
3031 else
3032 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
3033 }
3034
3035 /*
3036 * Check and if necessary update the leaf rcu_node structure's
3037 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
3038 * number of queued RCU callbacks. No locks need be held, but the
3039 * caller must have disabled interrupts.
3040 *
3041 * Note that this function ignores the possibility that there are a lot
3042 * of callbacks all of which have already seen the end of their respective
3043 * grace periods. This omission is due to the need for no-CBs CPUs to
3044 * be holding ->nocb_lock to do this check, which is too heavy for a
3045 * common-case operation.
3046 */
check_cb_ovld(struct rcu_data * rdp)3047 static void check_cb_ovld(struct rcu_data *rdp)
3048 {
3049 struct rcu_node *const rnp = rdp->mynode;
3050
3051 if (qovld_calc <= 0 ||
3052 ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
3053 !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
3054 return; // Early boot wildcard value or already set correctly.
3055 raw_spin_lock_rcu_node(rnp);
3056 check_cb_ovld_locked(rdp, rnp);
3057 raw_spin_unlock_rcu_node(rnp);
3058 }
3059
3060 static void
__call_rcu_common(struct rcu_head * head,rcu_callback_t func,bool lazy_in)3061 __call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy_in)
3062 {
3063 static atomic_t doublefrees;
3064 unsigned long flags;
3065 bool lazy;
3066 struct rcu_data *rdp;
3067
3068 /* Misaligned rcu_head! */
3069 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3070
3071 if (debug_rcu_head_queue(head)) {
3072 /*
3073 * Probable double call_rcu(), so leak the callback.
3074 * Use rcu:rcu_callback trace event to find the previous
3075 * time callback was passed to call_rcu().
3076 */
3077 if (atomic_inc_return(&doublefrees) < 4) {
3078 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
3079 mem_dump_obj(head);
3080 }
3081 WRITE_ONCE(head->func, rcu_leak_callback);
3082 return;
3083 }
3084 head->func = func;
3085 head->next = NULL;
3086 kasan_record_aux_stack_noalloc(head);
3087 local_irq_save(flags);
3088 rdp = this_cpu_ptr(&rcu_data);
3089 lazy = lazy_in && !rcu_async_should_hurry();
3090
3091 /* Add the callback to our list. */
3092 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
3093 // This can trigger due to call_rcu() from offline CPU:
3094 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
3095 WARN_ON_ONCE(!rcu_is_watching());
3096 // Very early boot, before rcu_init(). Initialize if needed
3097 // and then drop through to queue the callback.
3098 if (rcu_segcblist_empty(&rdp->cblist))
3099 rcu_segcblist_init(&rdp->cblist);
3100 }
3101
3102 check_cb_ovld(rdp);
3103
3104 if (unlikely(rcu_rdp_is_offloaded(rdp)))
3105 call_rcu_nocb(rdp, head, func, flags, lazy);
3106 else
3107 call_rcu_core(rdp, head, func, flags);
3108 local_irq_restore(flags);
3109 }
3110
3111 #ifdef CONFIG_RCU_LAZY
3112 static bool enable_rcu_lazy __read_mostly = !IS_ENABLED(CONFIG_RCU_LAZY_DEFAULT_OFF);
3113 module_param(enable_rcu_lazy, bool, 0444);
3114
3115 /**
3116 * call_rcu_hurry() - Queue RCU callback for invocation after grace period, and
3117 * flush all lazy callbacks (including the new one) to the main ->cblist while
3118 * doing so.
3119 *
3120 * @head: structure to be used for queueing the RCU updates.
3121 * @func: actual callback function to be invoked after the grace period
3122 *
3123 * The callback function will be invoked some time after a full grace
3124 * period elapses, in other words after all pre-existing RCU read-side
3125 * critical sections have completed.
3126 *
3127 * Use this API instead of call_rcu() if you don't want the callback to be
3128 * invoked after very long periods of time, which can happen on systems without
3129 * memory pressure and on systems which are lightly loaded or mostly idle.
3130 * This function will cause callbacks to be invoked sooner than later at the
3131 * expense of extra power. Other than that, this function is identical to, and
3132 * reuses call_rcu()'s logic. Refer to call_rcu() for more details about memory
3133 * ordering and other functionality.
3134 */
call_rcu_hurry(struct rcu_head * head,rcu_callback_t func)3135 void call_rcu_hurry(struct rcu_head *head, rcu_callback_t func)
3136 {
3137 __call_rcu_common(head, func, false);
3138 }
3139 EXPORT_SYMBOL_GPL(call_rcu_hurry);
3140 #else
3141 #define enable_rcu_lazy false
3142 #endif
3143
3144 /**
3145 * call_rcu() - Queue an RCU callback for invocation after a grace period.
3146 * By default the callbacks are 'lazy' and are kept hidden from the main
3147 * ->cblist to prevent starting of grace periods too soon.
3148 * If you desire grace periods to start very soon, use call_rcu_hurry().
3149 *
3150 * @head: structure to be used for queueing the RCU updates.
3151 * @func: actual callback function to be invoked after the grace period
3152 *
3153 * The callback function will be invoked some time after a full grace
3154 * period elapses, in other words after all pre-existing RCU read-side
3155 * critical sections have completed. However, the callback function
3156 * might well execute concurrently with RCU read-side critical sections
3157 * that started after call_rcu() was invoked.
3158 *
3159 * RCU read-side critical sections are delimited by rcu_read_lock()
3160 * and rcu_read_unlock(), and may be nested. In addition, but only in
3161 * v5.0 and later, regions of code across which interrupts, preemption,
3162 * or softirqs have been disabled also serve as RCU read-side critical
3163 * sections. This includes hardware interrupt handlers, softirq handlers,
3164 * and NMI handlers.
3165 *
3166 * Note that all CPUs must agree that the grace period extended beyond
3167 * all pre-existing RCU read-side critical section. On systems with more
3168 * than one CPU, this means that when "func()" is invoked, each CPU is
3169 * guaranteed to have executed a full memory barrier since the end of its
3170 * last RCU read-side critical section whose beginning preceded the call
3171 * to call_rcu(). It also means that each CPU executing an RCU read-side
3172 * critical section that continues beyond the start of "func()" must have
3173 * executed a memory barrier after the call_rcu() but before the beginning
3174 * of that RCU read-side critical section. Note that these guarantees
3175 * include CPUs that are offline, idle, or executing in user mode, as
3176 * well as CPUs that are executing in the kernel.
3177 *
3178 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
3179 * resulting RCU callback function "func()", then both CPU A and CPU B are
3180 * guaranteed to execute a full memory barrier during the time interval
3181 * between the call to call_rcu() and the invocation of "func()" -- even
3182 * if CPU A and CPU B are the same CPU (but again only if the system has
3183 * more than one CPU).
3184 *
3185 * Implementation of these memory-ordering guarantees is described here:
3186 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3187 */
call_rcu(struct rcu_head * head,rcu_callback_t func)3188 void call_rcu(struct rcu_head *head, rcu_callback_t func)
3189 {
3190 __call_rcu_common(head, func, enable_rcu_lazy);
3191 }
3192 EXPORT_SYMBOL_GPL(call_rcu);
3193
3194 /* Maximum number of jiffies to wait before draining a batch. */
3195 #define KFREE_DRAIN_JIFFIES (5 * HZ)
3196 #define KFREE_N_BATCHES 2
3197 #define FREE_N_CHANNELS 2
3198
3199 /**
3200 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
3201 * @list: List node. All blocks are linked between each other
3202 * @gp_snap: Snapshot of RCU state for objects placed to this bulk
3203 * @nr_records: Number of active pointers in the array
3204 * @records: Array of the kvfree_rcu() pointers
3205 */
3206 struct kvfree_rcu_bulk_data {
3207 struct list_head list;
3208 struct rcu_gp_oldstate gp_snap;
3209 unsigned long nr_records;
3210 void *records[] __counted_by(nr_records);
3211 };
3212
3213 /*
3214 * This macro defines how many entries the "records" array
3215 * will contain. It is based on the fact that the size of
3216 * kvfree_rcu_bulk_data structure becomes exactly one page.
3217 */
3218 #define KVFREE_BULK_MAX_ENTR \
3219 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
3220
3221 /**
3222 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
3223 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
3224 * @head_free: List of kfree_rcu() objects waiting for a grace period
3225 * @head_free_gp_snap: Grace-period snapshot to check for attempted premature frees.
3226 * @bulk_head_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
3227 * @krcp: Pointer to @kfree_rcu_cpu structure
3228 */
3229
3230 struct kfree_rcu_cpu_work {
3231 struct rcu_work rcu_work;
3232 struct rcu_head *head_free;
3233 struct rcu_gp_oldstate head_free_gp_snap;
3234 struct list_head bulk_head_free[FREE_N_CHANNELS];
3235 struct kfree_rcu_cpu *krcp;
3236 };
3237
3238 /**
3239 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
3240 * @head: List of kfree_rcu() objects not yet waiting for a grace period
3241 * @head_gp_snap: Snapshot of RCU state for objects placed to "@head"
3242 * @bulk_head: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
3243 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
3244 * @lock: Synchronize access to this structure
3245 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
3246 * @initialized: The @rcu_work fields have been initialized
3247 * @head_count: Number of objects in rcu_head singular list
3248 * @bulk_count: Number of objects in bulk-list
3249 * @bkvcache:
3250 * A simple cache list that contains objects for reuse purpose.
3251 * In order to save some per-cpu space the list is singular.
3252 * Even though it is lockless an access has to be protected by the
3253 * per-cpu lock.
3254 * @page_cache_work: A work to refill the cache when it is empty
3255 * @backoff_page_cache_fill: Delay cache refills
3256 * @work_in_progress: Indicates that page_cache_work is running
3257 * @hrtimer: A hrtimer for scheduling a page_cache_work
3258 * @nr_bkv_objs: number of allocated objects at @bkvcache.
3259 *
3260 * This is a per-CPU structure. The reason that it is not included in
3261 * the rcu_data structure is to permit this code to be extracted from
3262 * the RCU files. Such extraction could allow further optimization of
3263 * the interactions with the slab allocators.
3264 */
3265 struct kfree_rcu_cpu {
3266 // Objects queued on a linked list
3267 // through their rcu_head structures.
3268 struct rcu_head *head;
3269 unsigned long head_gp_snap;
3270 atomic_t head_count;
3271
3272 // Objects queued on a bulk-list.
3273 struct list_head bulk_head[FREE_N_CHANNELS];
3274 atomic_t bulk_count[FREE_N_CHANNELS];
3275
3276 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
3277 raw_spinlock_t lock;
3278 struct delayed_work monitor_work;
3279 bool initialized;
3280
3281 struct delayed_work page_cache_work;
3282 atomic_t backoff_page_cache_fill;
3283 atomic_t work_in_progress;
3284 struct hrtimer hrtimer;
3285
3286 struct llist_head bkvcache;
3287 int nr_bkv_objs;
3288 };
3289
3290 static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
3291 .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
3292 };
3293
3294 static __always_inline void
debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data * bhead)3295 debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
3296 {
3297 #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3298 int i;
3299
3300 for (i = 0; i < bhead->nr_records; i++)
3301 debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
3302 #endif
3303 }
3304
3305 static inline struct kfree_rcu_cpu *
krc_this_cpu_lock(unsigned long * flags)3306 krc_this_cpu_lock(unsigned long *flags)
3307 {
3308 struct kfree_rcu_cpu *krcp;
3309
3310 local_irq_save(*flags); // For safely calling this_cpu_ptr().
3311 krcp = this_cpu_ptr(&krc);
3312 raw_spin_lock(&krcp->lock);
3313
3314 return krcp;
3315 }
3316
3317 static inline void
krc_this_cpu_unlock(struct kfree_rcu_cpu * krcp,unsigned long flags)3318 krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
3319 {
3320 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3321 }
3322
3323 static inline struct kvfree_rcu_bulk_data *
get_cached_bnode(struct kfree_rcu_cpu * krcp)3324 get_cached_bnode(struct kfree_rcu_cpu *krcp)
3325 {
3326 if (!krcp->nr_bkv_objs)
3327 return NULL;
3328
3329 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
3330 return (struct kvfree_rcu_bulk_data *)
3331 llist_del_first(&krcp->bkvcache);
3332 }
3333
3334 static inline bool
put_cached_bnode(struct kfree_rcu_cpu * krcp,struct kvfree_rcu_bulk_data * bnode)3335 put_cached_bnode(struct kfree_rcu_cpu *krcp,
3336 struct kvfree_rcu_bulk_data *bnode)
3337 {
3338 // Check the limit.
3339 if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
3340 return false;
3341
3342 llist_add((struct llist_node *) bnode, &krcp->bkvcache);
3343 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
3344 return true;
3345 }
3346
3347 static int
drain_page_cache(struct kfree_rcu_cpu * krcp)3348 drain_page_cache(struct kfree_rcu_cpu *krcp)
3349 {
3350 unsigned long flags;
3351 struct llist_node *page_list, *pos, *n;
3352 int freed = 0;
3353
3354 if (!rcu_min_cached_objs)
3355 return 0;
3356
3357 raw_spin_lock_irqsave(&krcp->lock, flags);
3358 page_list = llist_del_all(&krcp->bkvcache);
3359 WRITE_ONCE(krcp->nr_bkv_objs, 0);
3360 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3361
3362 llist_for_each_safe(pos, n, page_list) {
3363 free_page((unsigned long)pos);
3364 freed++;
3365 }
3366
3367 return freed;
3368 }
3369
3370 static void
kvfree_rcu_bulk(struct kfree_rcu_cpu * krcp,struct kvfree_rcu_bulk_data * bnode,int idx)3371 kvfree_rcu_bulk(struct kfree_rcu_cpu *krcp,
3372 struct kvfree_rcu_bulk_data *bnode, int idx)
3373 {
3374 unsigned long flags;
3375 int i;
3376
3377 if (!WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&bnode->gp_snap))) {
3378 debug_rcu_bhead_unqueue(bnode);
3379 rcu_lock_acquire(&rcu_callback_map);
3380 if (idx == 0) { // kmalloc() / kfree().
3381 trace_rcu_invoke_kfree_bulk_callback(
3382 rcu_state.name, bnode->nr_records,
3383 bnode->records);
3384
3385 kfree_bulk(bnode->nr_records, bnode->records);
3386 } else { // vmalloc() / vfree().
3387 for (i = 0; i < bnode->nr_records; i++) {
3388 trace_rcu_invoke_kvfree_callback(
3389 rcu_state.name, bnode->records[i], 0);
3390
3391 vfree(bnode->records[i]);
3392 }
3393 }
3394 rcu_lock_release(&rcu_callback_map);
3395 }
3396
3397 raw_spin_lock_irqsave(&krcp->lock, flags);
3398 if (put_cached_bnode(krcp, bnode))
3399 bnode = NULL;
3400 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3401
3402 if (bnode)
3403 free_page((unsigned long) bnode);
3404
3405 cond_resched_tasks_rcu_qs();
3406 }
3407
3408 static void
kvfree_rcu_list(struct rcu_head * head)3409 kvfree_rcu_list(struct rcu_head *head)
3410 {
3411 struct rcu_head *next;
3412
3413 for (; head; head = next) {
3414 void *ptr = (void *) head->func;
3415 unsigned long offset = (void *) head - ptr;
3416
3417 next = head->next;
3418 debug_rcu_head_unqueue((struct rcu_head *)ptr);
3419 rcu_lock_acquire(&rcu_callback_map);
3420 trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3421
3422 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3423 kvfree(ptr);
3424
3425 rcu_lock_release(&rcu_callback_map);
3426 cond_resched_tasks_rcu_qs();
3427 }
3428 }
3429
3430 /*
3431 * This function is invoked in workqueue context after a grace period.
3432 * It frees all the objects queued on ->bulk_head_free or ->head_free.
3433 */
kfree_rcu_work(struct work_struct * work)3434 static void kfree_rcu_work(struct work_struct *work)
3435 {
3436 unsigned long flags;
3437 struct kvfree_rcu_bulk_data *bnode, *n;
3438 struct list_head bulk_head[FREE_N_CHANNELS];
3439 struct rcu_head *head;
3440 struct kfree_rcu_cpu *krcp;
3441 struct kfree_rcu_cpu_work *krwp;
3442 struct rcu_gp_oldstate head_gp_snap;
3443 int i;
3444
3445 krwp = container_of(to_rcu_work(work),
3446 struct kfree_rcu_cpu_work, rcu_work);
3447 krcp = krwp->krcp;
3448
3449 raw_spin_lock_irqsave(&krcp->lock, flags);
3450 // Channels 1 and 2.
3451 for (i = 0; i < FREE_N_CHANNELS; i++)
3452 list_replace_init(&krwp->bulk_head_free[i], &bulk_head[i]);
3453
3454 // Channel 3.
3455 head = krwp->head_free;
3456 krwp->head_free = NULL;
3457 head_gp_snap = krwp->head_free_gp_snap;
3458 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3459
3460 // Handle the first two channels.
3461 for (i = 0; i < FREE_N_CHANNELS; i++) {
3462 // Start from the tail page, so a GP is likely passed for it.
3463 list_for_each_entry_safe(bnode, n, &bulk_head[i], list)
3464 kvfree_rcu_bulk(krcp, bnode, i);
3465 }
3466
3467 /*
3468 * This is used when the "bulk" path can not be used for the
3469 * double-argument of kvfree_rcu(). This happens when the
3470 * page-cache is empty, which means that objects are instead
3471 * queued on a linked list through their rcu_head structures.
3472 * This list is named "Channel 3".
3473 */
3474 if (head && !WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&head_gp_snap)))
3475 kvfree_rcu_list(head);
3476 }
3477
3478 static bool
need_offload_krc(struct kfree_rcu_cpu * krcp)3479 need_offload_krc(struct kfree_rcu_cpu *krcp)
3480 {
3481 int i;
3482
3483 for (i = 0; i < FREE_N_CHANNELS; i++)
3484 if (!list_empty(&krcp->bulk_head[i]))
3485 return true;
3486
3487 return !!READ_ONCE(krcp->head);
3488 }
3489
3490 static bool
need_wait_for_krwp_work(struct kfree_rcu_cpu_work * krwp)3491 need_wait_for_krwp_work(struct kfree_rcu_cpu_work *krwp)
3492 {
3493 int i;
3494
3495 for (i = 0; i < FREE_N_CHANNELS; i++)
3496 if (!list_empty(&krwp->bulk_head_free[i]))
3497 return true;
3498
3499 return !!krwp->head_free;
3500 }
3501
krc_count(struct kfree_rcu_cpu * krcp)3502 static int krc_count(struct kfree_rcu_cpu *krcp)
3503 {
3504 int sum = atomic_read(&krcp->head_count);
3505 int i;
3506
3507 for (i = 0; i < FREE_N_CHANNELS; i++)
3508 sum += atomic_read(&krcp->bulk_count[i]);
3509
3510 return sum;
3511 }
3512
3513 static void
__schedule_delayed_monitor_work(struct kfree_rcu_cpu * krcp)3514 __schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
3515 {
3516 long delay, delay_left;
3517
3518 delay = krc_count(krcp) >= KVFREE_BULK_MAX_ENTR ? 1:KFREE_DRAIN_JIFFIES;
3519 if (delayed_work_pending(&krcp->monitor_work)) {
3520 delay_left = krcp->monitor_work.timer.expires - jiffies;
3521 if (delay < delay_left)
3522 mod_delayed_work(system_unbound_wq, &krcp->monitor_work, delay);
3523 return;
3524 }
3525 queue_delayed_work(system_unbound_wq, &krcp->monitor_work, delay);
3526 }
3527
3528 static void
schedule_delayed_monitor_work(struct kfree_rcu_cpu * krcp)3529 schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
3530 {
3531 unsigned long flags;
3532
3533 raw_spin_lock_irqsave(&krcp->lock, flags);
3534 __schedule_delayed_monitor_work(krcp);
3535 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3536 }
3537
3538 static void
kvfree_rcu_drain_ready(struct kfree_rcu_cpu * krcp)3539 kvfree_rcu_drain_ready(struct kfree_rcu_cpu *krcp)
3540 {
3541 struct list_head bulk_ready[FREE_N_CHANNELS];
3542 struct kvfree_rcu_bulk_data *bnode, *n;
3543 struct rcu_head *head_ready = NULL;
3544 unsigned long flags;
3545 int i;
3546
3547 raw_spin_lock_irqsave(&krcp->lock, flags);
3548 for (i = 0; i < FREE_N_CHANNELS; i++) {
3549 INIT_LIST_HEAD(&bulk_ready[i]);
3550
3551 list_for_each_entry_safe_reverse(bnode, n, &krcp->bulk_head[i], list) {
3552 if (!poll_state_synchronize_rcu_full(&bnode->gp_snap))
3553 break;
3554
3555 atomic_sub(bnode->nr_records, &krcp->bulk_count[i]);
3556 list_move(&bnode->list, &bulk_ready[i]);
3557 }
3558 }
3559
3560 if (krcp->head && poll_state_synchronize_rcu(krcp->head_gp_snap)) {
3561 head_ready = krcp->head;
3562 atomic_set(&krcp->head_count, 0);
3563 WRITE_ONCE(krcp->head, NULL);
3564 }
3565 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3566
3567 for (i = 0; i < FREE_N_CHANNELS; i++) {
3568 list_for_each_entry_safe(bnode, n, &bulk_ready[i], list)
3569 kvfree_rcu_bulk(krcp, bnode, i);
3570 }
3571
3572 if (head_ready)
3573 kvfree_rcu_list(head_ready);
3574 }
3575
3576 /*
3577 * Return: %true if a work is queued, %false otherwise.
3578 */
3579 static bool
kvfree_rcu_queue_batch(struct kfree_rcu_cpu * krcp)3580 kvfree_rcu_queue_batch(struct kfree_rcu_cpu *krcp)
3581 {
3582 unsigned long flags;
3583 bool queued = false;
3584 int i, j;
3585
3586 raw_spin_lock_irqsave(&krcp->lock, flags);
3587
3588 // Attempt to start a new batch.
3589 for (i = 0; i < KFREE_N_BATCHES; i++) {
3590 struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3591
3592 // Try to detach bulk_head or head and attach it, only when
3593 // all channels are free. Any channel is not free means at krwp
3594 // there is on-going rcu work to handle krwp's free business.
3595 if (need_wait_for_krwp_work(krwp))
3596 continue;
3597
3598 // kvfree_rcu_drain_ready() might handle this krcp, if so give up.
3599 if (need_offload_krc(krcp)) {
3600 // Channel 1 corresponds to the SLAB-pointer bulk path.
3601 // Channel 2 corresponds to vmalloc-pointer bulk path.
3602 for (j = 0; j < FREE_N_CHANNELS; j++) {
3603 if (list_empty(&krwp->bulk_head_free[j])) {
3604 atomic_set(&krcp->bulk_count[j], 0);
3605 list_replace_init(&krcp->bulk_head[j],
3606 &krwp->bulk_head_free[j]);
3607 }
3608 }
3609
3610 // Channel 3 corresponds to both SLAB and vmalloc
3611 // objects queued on the linked list.
3612 if (!krwp->head_free) {
3613 krwp->head_free = krcp->head;
3614 get_state_synchronize_rcu_full(&krwp->head_free_gp_snap);
3615 atomic_set(&krcp->head_count, 0);
3616 WRITE_ONCE(krcp->head, NULL);
3617 }
3618
3619 // One work is per one batch, so there are three
3620 // "free channels", the batch can handle. Break
3621 // the loop since it is done with this CPU thus
3622 // queuing an RCU work is _always_ success here.
3623 queued = queue_rcu_work(system_unbound_wq, &krwp->rcu_work);
3624 WARN_ON_ONCE(!queued);
3625 break;
3626 }
3627 }
3628
3629 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3630 return queued;
3631 }
3632
3633 /*
3634 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3635 */
kfree_rcu_monitor(struct work_struct * work)3636 static void kfree_rcu_monitor(struct work_struct *work)
3637 {
3638 struct kfree_rcu_cpu *krcp = container_of(work,
3639 struct kfree_rcu_cpu, monitor_work.work);
3640
3641 // Drain ready for reclaim.
3642 kvfree_rcu_drain_ready(krcp);
3643
3644 // Queue a batch for a rest.
3645 kvfree_rcu_queue_batch(krcp);
3646
3647 // If there is nothing to detach, it means that our job is
3648 // successfully done here. In case of having at least one
3649 // of the channels that is still busy we should rearm the
3650 // work to repeat an attempt. Because previous batches are
3651 // still in progress.
3652 if (need_offload_krc(krcp))
3653 schedule_delayed_monitor_work(krcp);
3654 }
3655
3656 static enum hrtimer_restart
schedule_page_work_fn(struct hrtimer * t)3657 schedule_page_work_fn(struct hrtimer *t)
3658 {
3659 struct kfree_rcu_cpu *krcp =
3660 container_of(t, struct kfree_rcu_cpu, hrtimer);
3661
3662 queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3663 return HRTIMER_NORESTART;
3664 }
3665
fill_page_cache_func(struct work_struct * work)3666 static void fill_page_cache_func(struct work_struct *work)
3667 {
3668 struct kvfree_rcu_bulk_data *bnode;
3669 struct kfree_rcu_cpu *krcp =
3670 container_of(work, struct kfree_rcu_cpu,
3671 page_cache_work.work);
3672 unsigned long flags;
3673 int nr_pages;
3674 bool pushed;
3675 int i;
3676
3677 nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3678 1 : rcu_min_cached_objs;
3679
3680 for (i = READ_ONCE(krcp->nr_bkv_objs); i < nr_pages; i++) {
3681 bnode = (struct kvfree_rcu_bulk_data *)
3682 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3683
3684 if (!bnode)
3685 break;
3686
3687 raw_spin_lock_irqsave(&krcp->lock, flags);
3688 pushed = put_cached_bnode(krcp, bnode);
3689 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3690
3691 if (!pushed) {
3692 free_page((unsigned long) bnode);
3693 break;
3694 }
3695 }
3696
3697 atomic_set(&krcp->work_in_progress, 0);
3698 atomic_set(&krcp->backoff_page_cache_fill, 0);
3699 }
3700
3701 static void
run_page_cache_worker(struct kfree_rcu_cpu * krcp)3702 run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3703 {
3704 // If cache disabled, bail out.
3705 if (!rcu_min_cached_objs)
3706 return;
3707
3708 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3709 !atomic_xchg(&krcp->work_in_progress, 1)) {
3710 if (atomic_read(&krcp->backoff_page_cache_fill)) {
3711 queue_delayed_work(system_unbound_wq,
3712 &krcp->page_cache_work,
3713 msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3714 } else {
3715 hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3716 krcp->hrtimer.function = schedule_page_work_fn;
3717 hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3718 }
3719 }
3720 }
3721
3722 // Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3723 // state specified by flags. If can_alloc is true, the caller must
3724 // be schedulable and not be holding any locks or mutexes that might be
3725 // acquired by the memory allocator or anything that it might invoke.
3726 // Returns true if ptr was successfully recorded, else the caller must
3727 // use a fallback.
3728 static inline bool
add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu ** krcp,unsigned long * flags,void * ptr,bool can_alloc)3729 add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3730 unsigned long *flags, void *ptr, bool can_alloc)
3731 {
3732 struct kvfree_rcu_bulk_data *bnode;
3733 int idx;
3734
3735 *krcp = krc_this_cpu_lock(flags);
3736 if (unlikely(!(*krcp)->initialized))
3737 return false;
3738
3739 idx = !!is_vmalloc_addr(ptr);
3740 bnode = list_first_entry_or_null(&(*krcp)->bulk_head[idx],
3741 struct kvfree_rcu_bulk_data, list);
3742
3743 /* Check if a new block is required. */
3744 if (!bnode || bnode->nr_records == KVFREE_BULK_MAX_ENTR) {
3745 bnode = get_cached_bnode(*krcp);
3746 if (!bnode && can_alloc) {
3747 krc_this_cpu_unlock(*krcp, *flags);
3748
3749 // __GFP_NORETRY - allows a light-weight direct reclaim
3750 // what is OK from minimizing of fallback hitting point of
3751 // view. Apart of that it forbids any OOM invoking what is
3752 // also beneficial since we are about to release memory soon.
3753 //
3754 // __GFP_NOMEMALLOC - prevents from consuming of all the
3755 // memory reserves. Please note we have a fallback path.
3756 //
3757 // __GFP_NOWARN - it is supposed that an allocation can
3758 // be failed under low memory or high memory pressure
3759 // scenarios.
3760 bnode = (struct kvfree_rcu_bulk_data *)
3761 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3762 raw_spin_lock_irqsave(&(*krcp)->lock, *flags);
3763 }
3764
3765 if (!bnode)
3766 return false;
3767
3768 // Initialize the new block and attach it.
3769 bnode->nr_records = 0;
3770 list_add(&bnode->list, &(*krcp)->bulk_head[idx]);
3771 }
3772
3773 // Finally insert and update the GP for this page.
3774 bnode->nr_records++;
3775 bnode->records[bnode->nr_records - 1] = ptr;
3776 get_state_synchronize_rcu_full(&bnode->gp_snap);
3777 atomic_inc(&(*krcp)->bulk_count[idx]);
3778
3779 return true;
3780 }
3781
3782 /*
3783 * Queue a request for lazy invocation of the appropriate free routine
3784 * after a grace period. Please note that three paths are maintained,
3785 * two for the common case using arrays of pointers and a third one that
3786 * is used only when the main paths cannot be used, for example, due to
3787 * memory pressure.
3788 *
3789 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3790 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3791 * be free'd in workqueue context. This allows us to: batch requests together to
3792 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3793 */
kvfree_call_rcu(struct rcu_head * head,void * ptr)3794 void kvfree_call_rcu(struct rcu_head *head, void *ptr)
3795 {
3796 unsigned long flags;
3797 struct kfree_rcu_cpu *krcp;
3798 bool success;
3799
3800 /*
3801 * Please note there is a limitation for the head-less
3802 * variant, that is why there is a clear rule for such
3803 * objects: it can be used from might_sleep() context
3804 * only. For other places please embed an rcu_head to
3805 * your data.
3806 */
3807 if (!head)
3808 might_sleep();
3809
3810 // Queue the object but don't yet schedule the batch.
3811 if (debug_rcu_head_queue(ptr)) {
3812 // Probable double kfree_rcu(), just leak.
3813 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3814 __func__, head);
3815
3816 // Mark as success and leave.
3817 return;
3818 }
3819
3820 kasan_record_aux_stack_noalloc(ptr);
3821 success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3822 if (!success) {
3823 run_page_cache_worker(krcp);
3824
3825 if (head == NULL)
3826 // Inline if kvfree_rcu(one_arg) call.
3827 goto unlock_return;
3828
3829 head->func = ptr;
3830 head->next = krcp->head;
3831 WRITE_ONCE(krcp->head, head);
3832 atomic_inc(&krcp->head_count);
3833
3834 // Take a snapshot for this krcp.
3835 krcp->head_gp_snap = get_state_synchronize_rcu();
3836 success = true;
3837 }
3838
3839 /*
3840 * The kvfree_rcu() caller considers the pointer freed at this point
3841 * and likely removes any references to it. Since the actual slab
3842 * freeing (and kmemleak_free()) is deferred, tell kmemleak to ignore
3843 * this object (no scanning or false positives reporting).
3844 */
3845 kmemleak_ignore(ptr);
3846
3847 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3848 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING)
3849 __schedule_delayed_monitor_work(krcp);
3850
3851 unlock_return:
3852 krc_this_cpu_unlock(krcp, flags);
3853
3854 /*
3855 * Inline kvfree() after synchronize_rcu(). We can do
3856 * it from might_sleep() context only, so the current
3857 * CPU can pass the QS state.
3858 */
3859 if (!success) {
3860 debug_rcu_head_unqueue((struct rcu_head *) ptr);
3861 synchronize_rcu();
3862 kvfree(ptr);
3863 }
3864 }
3865 EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3866
3867 /**
3868 * kvfree_rcu_barrier - Wait until all in-flight kvfree_rcu() complete.
3869 *
3870 * Note that a single argument of kvfree_rcu() call has a slow path that
3871 * triggers synchronize_rcu() following by freeing a pointer. It is done
3872 * before the return from the function. Therefore for any single-argument
3873 * call that will result in a kfree() to a cache that is to be destroyed
3874 * during module exit, it is developer's responsibility to ensure that all
3875 * such calls have returned before the call to kmem_cache_destroy().
3876 */
kvfree_rcu_barrier(void)3877 void kvfree_rcu_barrier(void)
3878 {
3879 struct kfree_rcu_cpu_work *krwp;
3880 struct kfree_rcu_cpu *krcp;
3881 bool queued;
3882 int i, cpu;
3883
3884 /*
3885 * Firstly we detach objects and queue them over an RCU-batch
3886 * for all CPUs. Finally queued works are flushed for each CPU.
3887 *
3888 * Please note. If there are outstanding batches for a particular
3889 * CPU, those have to be finished first following by queuing a new.
3890 */
3891 for_each_possible_cpu(cpu) {
3892 krcp = per_cpu_ptr(&krc, cpu);
3893
3894 /*
3895 * Check if this CPU has any objects which have been queued for a
3896 * new GP completion. If not(means nothing to detach), we are done
3897 * with it. If any batch is pending/running for this "krcp", below
3898 * per-cpu flush_rcu_work() waits its completion(see last step).
3899 */
3900 if (!need_offload_krc(krcp))
3901 continue;
3902
3903 while (1) {
3904 /*
3905 * If we are not able to queue a new RCU work it means:
3906 * - batches for this CPU are still in flight which should
3907 * be flushed first and then repeat;
3908 * - no objects to detach, because of concurrency.
3909 */
3910 queued = kvfree_rcu_queue_batch(krcp);
3911
3912 /*
3913 * Bail out, if there is no need to offload this "krcp"
3914 * anymore. As noted earlier it can run concurrently.
3915 */
3916 if (queued || !need_offload_krc(krcp))
3917 break;
3918
3919 /* There are ongoing batches. */
3920 for (i = 0; i < KFREE_N_BATCHES; i++) {
3921 krwp = &(krcp->krw_arr[i]);
3922 flush_rcu_work(&krwp->rcu_work);
3923 }
3924 }
3925 }
3926
3927 /*
3928 * Now we guarantee that all objects are flushed.
3929 */
3930 for_each_possible_cpu(cpu) {
3931 krcp = per_cpu_ptr(&krc, cpu);
3932
3933 /*
3934 * A monitor work can drain ready to reclaim objects
3935 * directly. Wait its completion if running or pending.
3936 */
3937 cancel_delayed_work_sync(&krcp->monitor_work);
3938
3939 for (i = 0; i < KFREE_N_BATCHES; i++) {
3940 krwp = &(krcp->krw_arr[i]);
3941 flush_rcu_work(&krwp->rcu_work);
3942 }
3943 }
3944 }
3945 EXPORT_SYMBOL_GPL(kvfree_rcu_barrier);
3946
3947 static unsigned long
kfree_rcu_shrink_count(struct shrinker * shrink,struct shrink_control * sc)3948 kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3949 {
3950 int cpu;
3951 unsigned long count = 0;
3952
3953 /* Snapshot count of all CPUs */
3954 for_each_possible_cpu(cpu) {
3955 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3956
3957 count += krc_count(krcp);
3958 count += READ_ONCE(krcp->nr_bkv_objs);
3959 atomic_set(&krcp->backoff_page_cache_fill, 1);
3960 }
3961
3962 return count == 0 ? SHRINK_EMPTY : count;
3963 }
3964
3965 static unsigned long
kfree_rcu_shrink_scan(struct shrinker * shrink,struct shrink_control * sc)3966 kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3967 {
3968 int cpu, freed = 0;
3969
3970 for_each_possible_cpu(cpu) {
3971 int count;
3972 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3973
3974 count = krc_count(krcp);
3975 count += drain_page_cache(krcp);
3976 kfree_rcu_monitor(&krcp->monitor_work.work);
3977
3978 sc->nr_to_scan -= count;
3979 freed += count;
3980
3981 if (sc->nr_to_scan <= 0)
3982 break;
3983 }
3984
3985 return freed == 0 ? SHRINK_STOP : freed;
3986 }
3987
kfree_rcu_scheduler_running(void)3988 void __init kfree_rcu_scheduler_running(void)
3989 {
3990 int cpu;
3991
3992 for_each_possible_cpu(cpu) {
3993 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3994
3995 if (need_offload_krc(krcp))
3996 schedule_delayed_monitor_work(krcp);
3997 }
3998 }
3999
4000 /*
4001 * During early boot, any blocking grace-period wait automatically
4002 * implies a grace period.
4003 *
4004 * Later on, this could in theory be the case for kernels built with
4005 * CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this
4006 * is not a common case. Furthermore, this optimization would cause
4007 * the rcu_gp_oldstate structure to expand by 50%, so this potential
4008 * grace-period optimization is ignored once the scheduler is running.
4009 */
rcu_blocking_is_gp(void)4010 static int rcu_blocking_is_gp(void)
4011 {
4012 if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) {
4013 might_sleep();
4014 return false;
4015 }
4016 return true;
4017 }
4018
4019 /*
4020 * Helper function for the synchronize_rcu() API.
4021 */
synchronize_rcu_normal(void)4022 static void synchronize_rcu_normal(void)
4023 {
4024 struct rcu_synchronize rs;
4025
4026 trace_rcu_sr_normal(rcu_state.name, &rs.head, TPS("request"));
4027
4028 if (!READ_ONCE(rcu_normal_wake_from_gp)) {
4029 wait_rcu_gp(call_rcu_hurry);
4030 goto trace_complete_out;
4031 }
4032
4033 init_rcu_head_on_stack(&rs.head);
4034 init_completion(&rs.completion);
4035
4036 /*
4037 * This code might be preempted, therefore take a GP
4038 * snapshot before adding a request.
4039 */
4040 if (IS_ENABLED(CONFIG_PROVE_RCU))
4041 rs.head.func = (void *) get_state_synchronize_rcu();
4042
4043 rcu_sr_normal_add_req(&rs);
4044
4045 /* Kick a GP and start waiting. */
4046 (void) start_poll_synchronize_rcu();
4047
4048 /* Now we can wait. */
4049 wait_for_completion(&rs.completion);
4050 destroy_rcu_head_on_stack(&rs.head);
4051
4052 trace_complete_out:
4053 trace_rcu_sr_normal(rcu_state.name, &rs.head, TPS("complete"));
4054 }
4055
4056 /**
4057 * synchronize_rcu - wait until a grace period has elapsed.
4058 *
4059 * Control will return to the caller some time after a full grace
4060 * period has elapsed, in other words after all currently executing RCU
4061 * read-side critical sections have completed. Note, however, that
4062 * upon return from synchronize_rcu(), the caller might well be executing
4063 * concurrently with new RCU read-side critical sections that began while
4064 * synchronize_rcu() was waiting.
4065 *
4066 * RCU read-side critical sections are delimited by rcu_read_lock()
4067 * and rcu_read_unlock(), and may be nested. In addition, but only in
4068 * v5.0 and later, regions of code across which interrupts, preemption,
4069 * or softirqs have been disabled also serve as RCU read-side critical
4070 * sections. This includes hardware interrupt handlers, softirq handlers,
4071 * and NMI handlers.
4072 *
4073 * Note that this guarantee implies further memory-ordering guarantees.
4074 * On systems with more than one CPU, when synchronize_rcu() returns,
4075 * each CPU is guaranteed to have executed a full memory barrier since
4076 * the end of its last RCU read-side critical section whose beginning
4077 * preceded the call to synchronize_rcu(). In addition, each CPU having
4078 * an RCU read-side critical section that extends beyond the return from
4079 * synchronize_rcu() is guaranteed to have executed a full memory barrier
4080 * after the beginning of synchronize_rcu() and before the beginning of
4081 * that RCU read-side critical section. Note that these guarantees include
4082 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
4083 * that are executing in the kernel.
4084 *
4085 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
4086 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
4087 * to have executed a full memory barrier during the execution of
4088 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
4089 * again only if the system has more than one CPU).
4090 *
4091 * Implementation of these memory-ordering guarantees is described here:
4092 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
4093 */
synchronize_rcu(void)4094 void synchronize_rcu(void)
4095 {
4096 unsigned long flags;
4097 struct rcu_node *rnp;
4098
4099 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
4100 lock_is_held(&rcu_lock_map) ||
4101 lock_is_held(&rcu_sched_lock_map),
4102 "Illegal synchronize_rcu() in RCU read-side critical section");
4103 if (!rcu_blocking_is_gp()) {
4104 if (rcu_gp_is_expedited())
4105 synchronize_rcu_expedited();
4106 else
4107 synchronize_rcu_normal();
4108 return;
4109 }
4110
4111 // Context allows vacuous grace periods.
4112 // Note well that this code runs with !PREEMPT && !SMP.
4113 // In addition, all code that advances grace periods runs at
4114 // process level. Therefore, this normal GP overlaps with other
4115 // normal GPs only by being fully nested within them, which allows
4116 // reuse of ->gp_seq_polled_snap.
4117 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_snap);
4118 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_snap);
4119
4120 // Update the normal grace-period counters to record
4121 // this grace period, but only those used by the boot CPU.
4122 // The rcu_scheduler_starting() will take care of the rest of
4123 // these counters.
4124 local_irq_save(flags);
4125 WARN_ON_ONCE(num_online_cpus() > 1);
4126 rcu_state.gp_seq += (1 << RCU_SEQ_CTR_SHIFT);
4127 for (rnp = this_cpu_ptr(&rcu_data)->mynode; rnp; rnp = rnp->parent)
4128 rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
4129 local_irq_restore(flags);
4130 }
4131 EXPORT_SYMBOL_GPL(synchronize_rcu);
4132
4133 /**
4134 * get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie
4135 * @rgosp: Place to put state cookie
4136 *
4137 * Stores into @rgosp a value that will always be treated by functions
4138 * like poll_state_synchronize_rcu_full() as a cookie whose grace period
4139 * has already completed.
4140 */
get_completed_synchronize_rcu_full(struct rcu_gp_oldstate * rgosp)4141 void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
4142 {
4143 rgosp->rgos_norm = RCU_GET_STATE_COMPLETED;
4144 rgosp->rgos_exp = RCU_GET_STATE_COMPLETED;
4145 }
4146 EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full);
4147
4148 /**
4149 * get_state_synchronize_rcu - Snapshot current RCU state
4150 *
4151 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
4152 * or poll_state_synchronize_rcu() to determine whether or not a full
4153 * grace period has elapsed in the meantime.
4154 */
get_state_synchronize_rcu(void)4155 unsigned long get_state_synchronize_rcu(void)
4156 {
4157 /*
4158 * Any prior manipulation of RCU-protected data must happen
4159 * before the load from ->gp_seq.
4160 */
4161 smp_mb(); /* ^^^ */
4162 return rcu_seq_snap(&rcu_state.gp_seq_polled);
4163 }
4164 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
4165
4166 /**
4167 * get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited
4168 * @rgosp: location to place combined normal/expedited grace-period state
4169 *
4170 * Places the normal and expedited grace-period states in @rgosp. This
4171 * state value can be passed to a later call to cond_synchronize_rcu_full()
4172 * or poll_state_synchronize_rcu_full() to determine whether or not a
4173 * grace period (whether normal or expedited) has elapsed in the meantime.
4174 * The rcu_gp_oldstate structure takes up twice the memory of an unsigned
4175 * long, but is guaranteed to see all grace periods. In contrast, the
4176 * combined state occupies less memory, but can sometimes fail to take
4177 * grace periods into account.
4178 *
4179 * This does not guarantee that the needed grace period will actually
4180 * start.
4181 */
get_state_synchronize_rcu_full(struct rcu_gp_oldstate * rgosp)4182 void get_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
4183 {
4184 struct rcu_node *rnp = rcu_get_root();
4185
4186 /*
4187 * Any prior manipulation of RCU-protected data must happen
4188 * before the loads from ->gp_seq and ->expedited_sequence.
4189 */
4190 smp_mb(); /* ^^^ */
4191 rgosp->rgos_norm = rcu_seq_snap(&rnp->gp_seq);
4192 rgosp->rgos_exp = rcu_seq_snap(&rcu_state.expedited_sequence);
4193 }
4194 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full);
4195
4196 /*
4197 * Helper function for start_poll_synchronize_rcu() and
4198 * start_poll_synchronize_rcu_full().
4199 */
start_poll_synchronize_rcu_common(void)4200 static void start_poll_synchronize_rcu_common(void)
4201 {
4202 unsigned long flags;
4203 bool needwake;
4204 struct rcu_data *rdp;
4205 struct rcu_node *rnp;
4206
4207 local_irq_save(flags);
4208 rdp = this_cpu_ptr(&rcu_data);
4209 rnp = rdp->mynode;
4210 raw_spin_lock_rcu_node(rnp); // irqs already disabled.
4211 // Note it is possible for a grace period to have elapsed between
4212 // the above call to get_state_synchronize_rcu() and the below call
4213 // to rcu_seq_snap. This is OK, the worst that happens is that we
4214 // get a grace period that no one needed. These accesses are ordered
4215 // by smp_mb(), and we are accessing them in the opposite order
4216 // from which they are updated at grace-period start, as required.
4217 needwake = rcu_start_this_gp(rnp, rdp, rcu_seq_snap(&rcu_state.gp_seq));
4218 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4219 if (needwake)
4220 rcu_gp_kthread_wake();
4221 }
4222
4223 /**
4224 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
4225 *
4226 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
4227 * or poll_state_synchronize_rcu() to determine whether or not a full
4228 * grace period has elapsed in the meantime. If the needed grace period
4229 * is not already slated to start, notifies RCU core of the need for that
4230 * grace period.
4231 */
start_poll_synchronize_rcu(void)4232 unsigned long start_poll_synchronize_rcu(void)
4233 {
4234 unsigned long gp_seq = get_state_synchronize_rcu();
4235
4236 start_poll_synchronize_rcu_common();
4237 return gp_seq;
4238 }
4239 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
4240
4241 /**
4242 * start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period
4243 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
4244 *
4245 * Places the normal and expedited grace-period states in *@rgos. This
4246 * state value can be passed to a later call to cond_synchronize_rcu_full()
4247 * or poll_state_synchronize_rcu_full() to determine whether or not a
4248 * grace period (whether normal or expedited) has elapsed in the meantime.
4249 * If the needed grace period is not already slated to start, notifies
4250 * RCU core of the need for that grace period.
4251 */
start_poll_synchronize_rcu_full(struct rcu_gp_oldstate * rgosp)4252 void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
4253 {
4254 get_state_synchronize_rcu_full(rgosp);
4255
4256 start_poll_synchronize_rcu_common();
4257 }
4258 EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full);
4259
4260 /**
4261 * poll_state_synchronize_rcu - Has the specified RCU grace period completed?
4262 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
4263 *
4264 * If a full RCU grace period has elapsed since the earlier call from
4265 * which @oldstate was obtained, return @true, otherwise return @false.
4266 * If @false is returned, it is the caller's responsibility to invoke this
4267 * function later on until it does return @true. Alternatively, the caller
4268 * can explicitly wait for a grace period, for example, by passing @oldstate
4269 * to either cond_synchronize_rcu() or cond_synchronize_rcu_expedited()
4270 * on the one hand or by directly invoking either synchronize_rcu() or
4271 * synchronize_rcu_expedited() on the other.
4272 *
4273 * Yes, this function does not take counter wrap into account.
4274 * But counter wrap is harmless. If the counter wraps, we have waited for
4275 * more than a billion grace periods (and way more on a 64-bit system!).
4276 * Those needing to keep old state values for very long time periods
4277 * (many hours even on 32-bit systems) should check them occasionally and
4278 * either refresh them or set a flag indicating that the grace period has
4279 * completed. Alternatively, they can use get_completed_synchronize_rcu()
4280 * to get a guaranteed-completed grace-period state.
4281 *
4282 * In addition, because oldstate compresses the grace-period state for
4283 * both normal and expedited grace periods into a single unsigned long,
4284 * it can miss a grace period when synchronize_rcu() runs concurrently
4285 * with synchronize_rcu_expedited(). If this is unacceptable, please
4286 * instead use the _full() variant of these polling APIs.
4287 *
4288 * This function provides the same memory-ordering guarantees that
4289 * would be provided by a synchronize_rcu() that was invoked at the call
4290 * to the function that provided @oldstate, and that returned at the end
4291 * of this function.
4292 */
poll_state_synchronize_rcu(unsigned long oldstate)4293 bool poll_state_synchronize_rcu(unsigned long oldstate)
4294 {
4295 if (oldstate == RCU_GET_STATE_COMPLETED ||
4296 rcu_seq_done_exact(&rcu_state.gp_seq_polled, oldstate)) {
4297 smp_mb(); /* Ensure GP ends before subsequent accesses. */
4298 return true;
4299 }
4300 return false;
4301 }
4302 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
4303
4304 /**
4305 * poll_state_synchronize_rcu_full - Has the specified RCU grace period completed?
4306 * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full()
4307 *
4308 * If a full RCU grace period has elapsed since the earlier call from
4309 * which *rgosp was obtained, return @true, otherwise return @false.
4310 * If @false is returned, it is the caller's responsibility to invoke this
4311 * function later on until it does return @true. Alternatively, the caller
4312 * can explicitly wait for a grace period, for example, by passing @rgosp
4313 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
4314 *
4315 * Yes, this function does not take counter wrap into account.
4316 * But counter wrap is harmless. If the counter wraps, we have waited
4317 * for more than a billion grace periods (and way more on a 64-bit
4318 * system!). Those needing to keep rcu_gp_oldstate values for very
4319 * long time periods (many hours even on 32-bit systems) should check
4320 * them occasionally and either refresh them or set a flag indicating
4321 * that the grace period has completed. Alternatively, they can use
4322 * get_completed_synchronize_rcu_full() to get a guaranteed-completed
4323 * grace-period state.
4324 *
4325 * This function provides the same memory-ordering guarantees that would
4326 * be provided by a synchronize_rcu() that was invoked at the call to
4327 * the function that provided @rgosp, and that returned at the end of this
4328 * function. And this guarantee requires that the root rcu_node structure's
4329 * ->gp_seq field be checked instead of that of the rcu_state structure.
4330 * The problem is that the just-ending grace-period's callbacks can be
4331 * invoked between the time that the root rcu_node structure's ->gp_seq
4332 * field is updated and the time that the rcu_state structure's ->gp_seq
4333 * field is updated. Therefore, if a single synchronize_rcu() is to
4334 * cause a subsequent poll_state_synchronize_rcu_full() to return @true,
4335 * then the root rcu_node structure is the one that needs to be polled.
4336 */
poll_state_synchronize_rcu_full(struct rcu_gp_oldstate * rgosp)4337 bool poll_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
4338 {
4339 struct rcu_node *rnp = rcu_get_root();
4340
4341 smp_mb(); // Order against root rcu_node structure grace-period cleanup.
4342 if (rgosp->rgos_norm == RCU_GET_STATE_COMPLETED ||
4343 rcu_seq_done_exact(&rnp->gp_seq, rgosp->rgos_norm) ||
4344 rgosp->rgos_exp == RCU_GET_STATE_COMPLETED ||
4345 rcu_seq_done_exact(&rcu_state.expedited_sequence, rgosp->rgos_exp)) {
4346 smp_mb(); /* Ensure GP ends before subsequent accesses. */
4347 return true;
4348 }
4349 return false;
4350 }
4351 EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu_full);
4352
4353 /**
4354 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
4355 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
4356 *
4357 * If a full RCU grace period has elapsed since the earlier call to
4358 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
4359 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
4360 *
4361 * Yes, this function does not take counter wrap into account.
4362 * But counter wrap is harmless. If the counter wraps, we have waited for
4363 * more than 2 billion grace periods (and way more on a 64-bit system!),
4364 * so waiting for a couple of additional grace periods should be just fine.
4365 *
4366 * This function provides the same memory-ordering guarantees that
4367 * would be provided by a synchronize_rcu() that was invoked at the call
4368 * to the function that provided @oldstate and that returned at the end
4369 * of this function.
4370 */
cond_synchronize_rcu(unsigned long oldstate)4371 void cond_synchronize_rcu(unsigned long oldstate)
4372 {
4373 if (!poll_state_synchronize_rcu(oldstate))
4374 synchronize_rcu();
4375 }
4376 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
4377
4378 /**
4379 * cond_synchronize_rcu_full - Conditionally wait for an RCU grace period
4380 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
4381 *
4382 * If a full RCU grace period has elapsed since the call to
4383 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
4384 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
4385 * obtained, just return. Otherwise, invoke synchronize_rcu() to wait
4386 * for a full grace period.
4387 *
4388 * Yes, this function does not take counter wrap into account.
4389 * But counter wrap is harmless. If the counter wraps, we have waited for
4390 * more than 2 billion grace periods (and way more on a 64-bit system!),
4391 * so waiting for a couple of additional grace periods should be just fine.
4392 *
4393 * This function provides the same memory-ordering guarantees that
4394 * would be provided by a synchronize_rcu() that was invoked at the call
4395 * to the function that provided @rgosp and that returned at the end of
4396 * this function.
4397 */
cond_synchronize_rcu_full(struct rcu_gp_oldstate * rgosp)4398 void cond_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp)
4399 {
4400 if (!poll_state_synchronize_rcu_full(rgosp))
4401 synchronize_rcu();
4402 }
4403 EXPORT_SYMBOL_GPL(cond_synchronize_rcu_full);
4404
4405 /*
4406 * Check to see if there is any immediate RCU-related work to be done by
4407 * the current CPU, returning 1 if so and zero otherwise. The checks are
4408 * in order of increasing expense: checks that can be carried out against
4409 * CPU-local state are performed first. However, we must check for CPU
4410 * stalls first, else we might not get a chance.
4411 */
rcu_pending(int user)4412 static int rcu_pending(int user)
4413 {
4414 bool gp_in_progress;
4415 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
4416 struct rcu_node *rnp = rdp->mynode;
4417
4418 lockdep_assert_irqs_disabled();
4419
4420 /* Check for CPU stalls, if enabled. */
4421 check_cpu_stall(rdp);
4422
4423 /* Does this CPU need a deferred NOCB wakeup? */
4424 if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
4425 return 1;
4426
4427 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
4428 gp_in_progress = rcu_gp_in_progress();
4429 if ((user || rcu_is_cpu_rrupt_from_idle() ||
4430 (gp_in_progress &&
4431 time_before(jiffies, READ_ONCE(rcu_state.gp_start) +
4432 nohz_full_patience_delay_jiffies))) &&
4433 rcu_nohz_full_cpu())
4434 return 0;
4435
4436 /* Is the RCU core waiting for a quiescent state from this CPU? */
4437 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
4438 return 1;
4439
4440 /* Does this CPU have callbacks ready to invoke? */
4441 if (!rcu_rdp_is_offloaded(rdp) &&
4442 rcu_segcblist_ready_cbs(&rdp->cblist))
4443 return 1;
4444
4445 /* Has RCU gone idle with this CPU needing another grace period? */
4446 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
4447 !rcu_rdp_is_offloaded(rdp) &&
4448 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
4449 return 1;
4450
4451 /* Have RCU grace period completed or started? */
4452 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
4453 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
4454 return 1;
4455
4456 /* nothing to do */
4457 return 0;
4458 }
4459
4460 /*
4461 * Helper function for rcu_barrier() tracing. If tracing is disabled,
4462 * the compiler is expected to optimize this away.
4463 */
rcu_barrier_trace(const char * s,int cpu,unsigned long done)4464 static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
4465 {
4466 trace_rcu_barrier(rcu_state.name, s, cpu,
4467 atomic_read(&rcu_state.barrier_cpu_count), done);
4468 }
4469
4470 /*
4471 * RCU callback function for rcu_barrier(). If we are last, wake
4472 * up the task executing rcu_barrier().
4473 *
4474 * Note that the value of rcu_state.barrier_sequence must be captured
4475 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
4476 * other CPUs might count the value down to zero before this CPU gets
4477 * around to invoking rcu_barrier_trace(), which might result in bogus
4478 * data from the next instance of rcu_barrier().
4479 */
rcu_barrier_callback(struct rcu_head * rhp)4480 static void rcu_barrier_callback(struct rcu_head *rhp)
4481 {
4482 unsigned long __maybe_unused s = rcu_state.barrier_sequence;
4483
4484 rhp->next = rhp; // Mark the callback as having been invoked.
4485 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
4486 rcu_barrier_trace(TPS("LastCB"), -1, s);
4487 complete(&rcu_state.barrier_completion);
4488 } else {
4489 rcu_barrier_trace(TPS("CB"), -1, s);
4490 }
4491 }
4492
4493 /*
4494 * If needed, entrain an rcu_barrier() callback on rdp->cblist.
4495 */
rcu_barrier_entrain(struct rcu_data * rdp)4496 static void rcu_barrier_entrain(struct rcu_data *rdp)
4497 {
4498 unsigned long gseq = READ_ONCE(rcu_state.barrier_sequence);
4499 unsigned long lseq = READ_ONCE(rdp->barrier_seq_snap);
4500 bool wake_nocb = false;
4501 bool was_alldone = false;
4502
4503 lockdep_assert_held(&rcu_state.barrier_lock);
4504 if (rcu_seq_state(lseq) || !rcu_seq_state(gseq) || rcu_seq_ctr(lseq) != rcu_seq_ctr(gseq))
4505 return;
4506 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
4507 rdp->barrier_head.func = rcu_barrier_callback;
4508 debug_rcu_head_queue(&rdp->barrier_head);
4509 rcu_nocb_lock(rdp);
4510 /*
4511 * Flush bypass and wakeup rcuog if we add callbacks to an empty regular
4512 * queue. This way we don't wait for bypass timer that can reach seconds
4513 * if it's fully lazy.
4514 */
4515 was_alldone = rcu_rdp_is_offloaded(rdp) && !rcu_segcblist_pend_cbs(&rdp->cblist);
4516 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
4517 wake_nocb = was_alldone && rcu_segcblist_pend_cbs(&rdp->cblist);
4518 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
4519 atomic_inc(&rcu_state.barrier_cpu_count);
4520 } else {
4521 debug_rcu_head_unqueue(&rdp->barrier_head);
4522 rcu_barrier_trace(TPS("IRQNQ"), -1, rcu_state.barrier_sequence);
4523 }
4524 rcu_nocb_unlock(rdp);
4525 if (wake_nocb)
4526 wake_nocb_gp(rdp, false);
4527 smp_store_release(&rdp->barrier_seq_snap, gseq);
4528 }
4529
4530 /*
4531 * Called with preemption disabled, and from cross-cpu IRQ context.
4532 */
rcu_barrier_handler(void * cpu_in)4533 static void rcu_barrier_handler(void *cpu_in)
4534 {
4535 uintptr_t cpu = (uintptr_t)cpu_in;
4536 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4537
4538 lockdep_assert_irqs_disabled();
4539 WARN_ON_ONCE(cpu != rdp->cpu);
4540 WARN_ON_ONCE(cpu != smp_processor_id());
4541 raw_spin_lock(&rcu_state.barrier_lock);
4542 rcu_barrier_entrain(rdp);
4543 raw_spin_unlock(&rcu_state.barrier_lock);
4544 }
4545
4546 /**
4547 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
4548 *
4549 * Note that this primitive does not necessarily wait for an RCU grace period
4550 * to complete. For example, if there are no RCU callbacks queued anywhere
4551 * in the system, then rcu_barrier() is within its rights to return
4552 * immediately, without waiting for anything, much less an RCU grace period.
4553 */
rcu_barrier(void)4554 void rcu_barrier(void)
4555 {
4556 uintptr_t cpu;
4557 unsigned long flags;
4558 unsigned long gseq;
4559 struct rcu_data *rdp;
4560 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4561
4562 rcu_barrier_trace(TPS("Begin"), -1, s);
4563
4564 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4565 mutex_lock(&rcu_state.barrier_mutex);
4566
4567 /* Did someone else do our work for us? */
4568 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4569 rcu_barrier_trace(TPS("EarlyExit"), -1, rcu_state.barrier_sequence);
4570 smp_mb(); /* caller's subsequent code after above check. */
4571 mutex_unlock(&rcu_state.barrier_mutex);
4572 return;
4573 }
4574
4575 /* Mark the start of the barrier operation. */
4576 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4577 rcu_seq_start(&rcu_state.barrier_sequence);
4578 gseq = rcu_state.barrier_sequence;
4579 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
4580
4581 /*
4582 * Initialize the count to two rather than to zero in order
4583 * to avoid a too-soon return to zero in case of an immediate
4584 * invocation of the just-enqueued callback (or preemption of
4585 * this task). Exclude CPU-hotplug operations to ensure that no
4586 * offline non-offloaded CPU has callbacks queued.
4587 */
4588 init_completion(&rcu_state.barrier_completion);
4589 atomic_set(&rcu_state.barrier_cpu_count, 2);
4590 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4591
4592 /*
4593 * Force each CPU with callbacks to register a new callback.
4594 * When that callback is invoked, we will know that all of the
4595 * corresponding CPU's preceding callbacks have been invoked.
4596 */
4597 for_each_possible_cpu(cpu) {
4598 rdp = per_cpu_ptr(&rcu_data, cpu);
4599 retry:
4600 if (smp_load_acquire(&rdp->barrier_seq_snap) == gseq)
4601 continue;
4602 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
4603 if (!rcu_segcblist_n_cbs(&rdp->cblist)) {
4604 WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4605 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4606 rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence);
4607 continue;
4608 }
4609 if (!rcu_rdp_cpu_online(rdp)) {
4610 rcu_barrier_entrain(rdp);
4611 WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4612 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4613 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, rcu_state.barrier_sequence);
4614 continue;
4615 }
4616 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
4617 if (smp_call_function_single(cpu, rcu_barrier_handler, (void *)cpu, 1)) {
4618 schedule_timeout_uninterruptible(1);
4619 goto retry;
4620 }
4621 WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq);
4622 rcu_barrier_trace(TPS("OnlineQ"), cpu, rcu_state.barrier_sequence);
4623 }
4624
4625 /*
4626 * Now that we have an rcu_barrier_callback() callback on each
4627 * CPU, and thus each counted, remove the initial count.
4628 */
4629 if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
4630 complete(&rcu_state.barrier_completion);
4631
4632 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4633 wait_for_completion(&rcu_state.barrier_completion);
4634
4635 /* Mark the end of the barrier operation. */
4636 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
4637 rcu_seq_end(&rcu_state.barrier_sequence);
4638 gseq = rcu_state.barrier_sequence;
4639 for_each_possible_cpu(cpu) {
4640 rdp = per_cpu_ptr(&rcu_data, cpu);
4641
4642 WRITE_ONCE(rdp->barrier_seq_snap, gseq);
4643 }
4644
4645 /* Other rcu_barrier() invocations can now safely proceed. */
4646 mutex_unlock(&rcu_state.barrier_mutex);
4647 }
4648 EXPORT_SYMBOL_GPL(rcu_barrier);
4649
4650 static unsigned long rcu_barrier_last_throttle;
4651
4652 /**
4653 * rcu_barrier_throttled - Do rcu_barrier(), but limit to one per second
4654 *
4655 * This can be thought of as guard rails around rcu_barrier() that
4656 * permits unrestricted userspace use, at least assuming the hardware's
4657 * try_cmpxchg() is robust. There will be at most one call per second to
4658 * rcu_barrier() system-wide from use of this function, which means that
4659 * callers might needlessly wait a second or three.
4660 *
4661 * This is intended for use by test suites to avoid OOM by flushing RCU
4662 * callbacks from the previous test before starting the next. See the
4663 * rcutree.do_rcu_barrier module parameter for more information.
4664 *
4665 * Why not simply make rcu_barrier() more scalable? That might be
4666 * the eventual endpoint, but let's keep it simple for the time being.
4667 * Note that the module parameter infrastructure serializes calls to a
4668 * given .set() function, but should concurrent .set() invocation ever be
4669 * possible, we are ready!
4670 */
rcu_barrier_throttled(void)4671 static void rcu_barrier_throttled(void)
4672 {
4673 unsigned long j = jiffies;
4674 unsigned long old = READ_ONCE(rcu_barrier_last_throttle);
4675 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4676
4677 while (time_in_range(j, old, old + HZ / 16) ||
4678 !try_cmpxchg(&rcu_barrier_last_throttle, &old, j)) {
4679 schedule_timeout_idle(HZ / 16);
4680 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4681 smp_mb(); /* caller's subsequent code after above check. */
4682 return;
4683 }
4684 j = jiffies;
4685 old = READ_ONCE(rcu_barrier_last_throttle);
4686 }
4687 rcu_barrier();
4688 }
4689
4690 /*
4691 * Invoke rcu_barrier_throttled() when a rcutree.do_rcu_barrier
4692 * request arrives. We insist on a true value to allow for possible
4693 * future expansion.
4694 */
param_set_do_rcu_barrier(const char * val,const struct kernel_param * kp)4695 static int param_set_do_rcu_barrier(const char *val, const struct kernel_param *kp)
4696 {
4697 bool b;
4698 int ret;
4699
4700 if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING)
4701 return -EAGAIN;
4702 ret = kstrtobool(val, &b);
4703 if (!ret && b) {
4704 atomic_inc((atomic_t *)kp->arg);
4705 rcu_barrier_throttled();
4706 atomic_dec((atomic_t *)kp->arg);
4707 }
4708 return ret;
4709 }
4710
4711 /*
4712 * Output the number of outstanding rcutree.do_rcu_barrier requests.
4713 */
param_get_do_rcu_barrier(char * buffer,const struct kernel_param * kp)4714 static int param_get_do_rcu_barrier(char *buffer, const struct kernel_param *kp)
4715 {
4716 return sprintf(buffer, "%d\n", atomic_read((atomic_t *)kp->arg));
4717 }
4718
4719 static const struct kernel_param_ops do_rcu_barrier_ops = {
4720 .set = param_set_do_rcu_barrier,
4721 .get = param_get_do_rcu_barrier,
4722 };
4723 static atomic_t do_rcu_barrier;
4724 module_param_cb(do_rcu_barrier, &do_rcu_barrier_ops, &do_rcu_barrier, 0644);
4725
4726 /*
4727 * Compute the mask of online CPUs for the specified rcu_node structure.
4728 * This will not be stable unless the rcu_node structure's ->lock is
4729 * held, but the bit corresponding to the current CPU will be stable
4730 * in most contexts.
4731 */
rcu_rnp_online_cpus(struct rcu_node * rnp)4732 static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
4733 {
4734 return READ_ONCE(rnp->qsmaskinitnext);
4735 }
4736
4737 /*
4738 * Is the CPU corresponding to the specified rcu_data structure online
4739 * from RCU's perspective? This perspective is given by that structure's
4740 * ->qsmaskinitnext field rather than by the global cpu_online_mask.
4741 */
rcu_rdp_cpu_online(struct rcu_data * rdp)4742 static bool rcu_rdp_cpu_online(struct rcu_data *rdp)
4743 {
4744 return !!(rdp->grpmask & rcu_rnp_online_cpus(rdp->mynode));
4745 }
4746
rcu_cpu_online(int cpu)4747 bool rcu_cpu_online(int cpu)
4748 {
4749 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4750
4751 return rcu_rdp_cpu_online(rdp);
4752 }
4753
4754 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
4755
4756 /*
4757 * Is the current CPU online as far as RCU is concerned?
4758 *
4759 * Disable preemption to avoid false positives that could otherwise
4760 * happen due to the current CPU number being sampled, this task being
4761 * preempted, its old CPU being taken offline, resuming on some other CPU,
4762 * then determining that its old CPU is now offline.
4763 *
4764 * Disable checking if in an NMI handler because we cannot safely
4765 * report errors from NMI handlers anyway. In addition, it is OK to use
4766 * RCU on an offline processor during initial boot, hence the check for
4767 * rcu_scheduler_fully_active.
4768 */
rcu_lockdep_current_cpu_online(void)4769 bool rcu_lockdep_current_cpu_online(void)
4770 {
4771 struct rcu_data *rdp;
4772 bool ret = false;
4773
4774 if (in_nmi() || !rcu_scheduler_fully_active)
4775 return true;
4776 preempt_disable_notrace();
4777 rdp = this_cpu_ptr(&rcu_data);
4778 /*
4779 * Strictly, we care here about the case where the current CPU is
4780 * in rcutree_report_cpu_starting() and thus has an excuse for rdp->grpmask
4781 * not being up to date. So arch_spin_is_locked() might have a
4782 * false positive if it's held by some *other* CPU, but that's
4783 * OK because that just means a false *negative* on the warning.
4784 */
4785 if (rcu_rdp_cpu_online(rdp) || arch_spin_is_locked(&rcu_state.ofl_lock))
4786 ret = true;
4787 preempt_enable_notrace();
4788 return ret;
4789 }
4790 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
4791
4792 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
4793
4794 // Has rcu_init() been invoked? This is used (for example) to determine
4795 // whether spinlocks may be acquired safely.
rcu_init_invoked(void)4796 static bool rcu_init_invoked(void)
4797 {
4798 return !!READ_ONCE(rcu_state.n_online_cpus);
4799 }
4800
4801 /*
4802 * All CPUs for the specified rcu_node structure have gone offline,
4803 * and all tasks that were preempted within an RCU read-side critical
4804 * section while running on one of those CPUs have since exited their RCU
4805 * read-side critical section. Some other CPU is reporting this fact with
4806 * the specified rcu_node structure's ->lock held and interrupts disabled.
4807 * This function therefore goes up the tree of rcu_node structures,
4808 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
4809 * the leaf rcu_node structure's ->qsmaskinit field has already been
4810 * updated.
4811 *
4812 * This function does check that the specified rcu_node structure has
4813 * all CPUs offline and no blocked tasks, so it is OK to invoke it
4814 * prematurely. That said, invoking it after the fact will cost you
4815 * a needless lock acquisition. So once it has done its work, don't
4816 * invoke it again.
4817 */
rcu_cleanup_dead_rnp(struct rcu_node * rnp_leaf)4818 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
4819 {
4820 long mask;
4821 struct rcu_node *rnp = rnp_leaf;
4822
4823 raw_lockdep_assert_held_rcu_node(rnp_leaf);
4824 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
4825 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
4826 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
4827 return;
4828 for (;;) {
4829 mask = rnp->grpmask;
4830 rnp = rnp->parent;
4831 if (!rnp)
4832 break;
4833 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4834 rnp->qsmaskinit &= ~mask;
4835 /* Between grace periods, so better already be zero! */
4836 WARN_ON_ONCE(rnp->qsmask);
4837 if (rnp->qsmaskinit) {
4838 raw_spin_unlock_rcu_node(rnp);
4839 /* irqs remain disabled. */
4840 return;
4841 }
4842 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4843 }
4844 }
4845
4846 /*
4847 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4848 * first CPU in a given leaf rcu_node structure coming online. The caller
4849 * must hold the corresponding leaf rcu_node ->lock with interrupts
4850 * disabled.
4851 */
rcu_init_new_rnp(struct rcu_node * rnp_leaf)4852 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4853 {
4854 long mask;
4855 long oldmask;
4856 struct rcu_node *rnp = rnp_leaf;
4857
4858 raw_lockdep_assert_held_rcu_node(rnp_leaf);
4859 WARN_ON_ONCE(rnp->wait_blkd_tasks);
4860 for (;;) {
4861 mask = rnp->grpmask;
4862 rnp = rnp->parent;
4863 if (rnp == NULL)
4864 return;
4865 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4866 oldmask = rnp->qsmaskinit;
4867 rnp->qsmaskinit |= mask;
4868 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4869 if (oldmask)
4870 return;
4871 }
4872 }
4873
4874 /*
4875 * Do boot-time initialization of a CPU's per-CPU RCU data.
4876 */
4877 static void __init
rcu_boot_init_percpu_data(int cpu)4878 rcu_boot_init_percpu_data(int cpu)
4879 {
4880 struct context_tracking *ct = this_cpu_ptr(&context_tracking);
4881 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4882
4883 /* Set up local state, ensuring consistent view of global state. */
4884 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4885 INIT_WORK(&rdp->strict_work, strict_work_handler);
4886 WARN_ON_ONCE(ct->nesting != 1);
4887 WARN_ON_ONCE(rcu_watching_snap_in_eqs(ct_rcu_watching_cpu(cpu)));
4888 rdp->barrier_seq_snap = rcu_state.barrier_sequence;
4889 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4890 rdp->rcu_ofl_gp_state = RCU_GP_CLEANED;
4891 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4892 rdp->rcu_onl_gp_state = RCU_GP_CLEANED;
4893 rdp->last_sched_clock = jiffies;
4894 rdp->cpu = cpu;
4895 rcu_boot_init_nocb_percpu_data(rdp);
4896 }
4897
4898 struct kthread_worker *rcu_exp_gp_kworker;
4899
rcu_spawn_exp_par_gp_kworker(struct rcu_node * rnp)4900 static void rcu_spawn_exp_par_gp_kworker(struct rcu_node *rnp)
4901 {
4902 struct kthread_worker *kworker;
4903 const char *name = "rcu_exp_par_gp_kthread_worker/%d";
4904 struct sched_param param = { .sched_priority = kthread_prio };
4905 int rnp_index = rnp - rcu_get_root();
4906
4907 if (rnp->exp_kworker)
4908 return;
4909
4910 kworker = kthread_create_worker(0, name, rnp_index);
4911 if (IS_ERR_OR_NULL(kworker)) {
4912 pr_err("Failed to create par gp kworker on %d/%d\n",
4913 rnp->grplo, rnp->grphi);
4914 return;
4915 }
4916 WRITE_ONCE(rnp->exp_kworker, kworker);
4917
4918 if (IS_ENABLED(CONFIG_RCU_EXP_KTHREAD))
4919 sched_setscheduler_nocheck(kworker->task, SCHED_FIFO, ¶m);
4920 }
4921
rcu_exp_par_gp_task(struct rcu_node * rnp)4922 static struct task_struct *rcu_exp_par_gp_task(struct rcu_node *rnp)
4923 {
4924 struct kthread_worker *kworker = READ_ONCE(rnp->exp_kworker);
4925
4926 if (!kworker)
4927 return NULL;
4928
4929 return kworker->task;
4930 }
4931
rcu_start_exp_gp_kworker(void)4932 static void __init rcu_start_exp_gp_kworker(void)
4933 {
4934 const char *name = "rcu_exp_gp_kthread_worker";
4935 struct sched_param param = { .sched_priority = kthread_prio };
4936
4937 rcu_exp_gp_kworker = kthread_create_worker(0, name);
4938 if (IS_ERR_OR_NULL(rcu_exp_gp_kworker)) {
4939 pr_err("Failed to create %s!\n", name);
4940 rcu_exp_gp_kworker = NULL;
4941 return;
4942 }
4943
4944 if (IS_ENABLED(CONFIG_RCU_EXP_KTHREAD))
4945 sched_setscheduler_nocheck(rcu_exp_gp_kworker->task, SCHED_FIFO, ¶m);
4946 }
4947
rcu_spawn_rnp_kthreads(struct rcu_node * rnp)4948 static void rcu_spawn_rnp_kthreads(struct rcu_node *rnp)
4949 {
4950 if (rcu_scheduler_fully_active) {
4951 mutex_lock(&rnp->kthread_mutex);
4952 rcu_spawn_one_boost_kthread(rnp);
4953 rcu_spawn_exp_par_gp_kworker(rnp);
4954 mutex_unlock(&rnp->kthread_mutex);
4955 }
4956 }
4957
4958 /*
4959 * Invoked early in the CPU-online process, when pretty much all services
4960 * are available. The incoming CPU is not present.
4961 *
4962 * Initializes a CPU's per-CPU RCU data. Note that only one online or
4963 * offline event can be happening at a given time. Note also that we can
4964 * accept some slop in the rsp->gp_seq access due to the fact that this
4965 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4966 * And any offloaded callbacks are being numbered elsewhere.
4967 */
rcutree_prepare_cpu(unsigned int cpu)4968 int rcutree_prepare_cpu(unsigned int cpu)
4969 {
4970 unsigned long flags;
4971 struct context_tracking *ct = per_cpu_ptr(&context_tracking, cpu);
4972 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4973 struct rcu_node *rnp = rcu_get_root();
4974
4975 /* Set up local state, ensuring consistent view of global state. */
4976 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4977 rdp->qlen_last_fqs_check = 0;
4978 rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs);
4979 rdp->blimit = blimit;
4980 ct->nesting = 1; /* CPU not up, no tearing. */
4981 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4982
4983 /*
4984 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4985 * (re-)initialized.
4986 */
4987 if (!rcu_segcblist_is_enabled(&rdp->cblist))
4988 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
4989
4990 /*
4991 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4992 * propagation up the rcu_node tree will happen at the beginning
4993 * of the next grace period.
4994 */
4995 rnp = rdp->mynode;
4996 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4997 rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4998 rdp->gp_seq_needed = rdp->gp_seq;
4999 rdp->cpu_no_qs.b.norm = true;
5000 rdp->core_needs_qs = false;
5001 rdp->rcu_iw_pending = false;
5002 rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
5003 rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
5004 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
5005 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
5006 rcu_spawn_rnp_kthreads(rnp);
5007 rcu_spawn_cpu_nocb_kthread(cpu);
5008 ASSERT_EXCLUSIVE_WRITER(rcu_state.n_online_cpus);
5009 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
5010
5011 return 0;
5012 }
5013
5014 /*
5015 * Update kthreads affinity during CPU-hotplug changes.
5016 *
5017 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
5018 * served by the rcu_node in question. The CPU hotplug lock is still
5019 * held, so the value of rnp->qsmaskinit will be stable.
5020 *
5021 * We don't include outgoingcpu in the affinity set, use -1 if there is
5022 * no outgoing CPU. If there are no CPUs left in the affinity set,
5023 * this function allows the kthread to execute on any CPU.
5024 *
5025 * Any future concurrent calls are serialized via ->kthread_mutex.
5026 */
rcutree_affinity_setting(unsigned int cpu,int outgoingcpu)5027 static void rcutree_affinity_setting(unsigned int cpu, int outgoingcpu)
5028 {
5029 cpumask_var_t cm;
5030 unsigned long mask;
5031 struct rcu_data *rdp;
5032 struct rcu_node *rnp;
5033 struct task_struct *task_boost, *task_exp;
5034
5035 rdp = per_cpu_ptr(&rcu_data, cpu);
5036 rnp = rdp->mynode;
5037
5038 task_boost = rcu_boost_task(rnp);
5039 task_exp = rcu_exp_par_gp_task(rnp);
5040
5041 /*
5042 * If CPU is the boot one, those tasks are created later from early
5043 * initcall since kthreadd must be created first.
5044 */
5045 if (!task_boost && !task_exp)
5046 return;
5047
5048 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
5049 return;
5050
5051 mutex_lock(&rnp->kthread_mutex);
5052 mask = rcu_rnp_online_cpus(rnp);
5053 for_each_leaf_node_possible_cpu(rnp, cpu)
5054 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
5055 cpu != outgoingcpu)
5056 cpumask_set_cpu(cpu, cm);
5057 cpumask_and(cm, cm, housekeeping_cpumask(HK_TYPE_RCU));
5058 if (cpumask_empty(cm)) {
5059 cpumask_copy(cm, housekeeping_cpumask(HK_TYPE_RCU));
5060 if (outgoingcpu >= 0)
5061 cpumask_clear_cpu(outgoingcpu, cm);
5062 }
5063
5064 if (task_exp)
5065 set_cpus_allowed_ptr(task_exp, cm);
5066
5067 if (task_boost)
5068 set_cpus_allowed_ptr(task_boost, cm);
5069
5070 mutex_unlock(&rnp->kthread_mutex);
5071
5072 free_cpumask_var(cm);
5073 }
5074
5075 /*
5076 * Has the specified (known valid) CPU ever been fully online?
5077 */
rcu_cpu_beenfullyonline(int cpu)5078 bool rcu_cpu_beenfullyonline(int cpu)
5079 {
5080 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
5081
5082 return smp_load_acquire(&rdp->beenonline);
5083 }
5084
5085 /*
5086 * Near the end of the CPU-online process. Pretty much all services
5087 * enabled, and the CPU is now very much alive.
5088 */
rcutree_online_cpu(unsigned int cpu)5089 int rcutree_online_cpu(unsigned int cpu)
5090 {
5091 unsigned long flags;
5092 struct rcu_data *rdp;
5093 struct rcu_node *rnp;
5094
5095 rdp = per_cpu_ptr(&rcu_data, cpu);
5096 rnp = rdp->mynode;
5097 raw_spin_lock_irqsave_rcu_node(rnp, flags);
5098 rnp->ffmask |= rdp->grpmask;
5099 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
5100 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
5101 return 0; /* Too early in boot for scheduler work. */
5102 sync_sched_exp_online_cleanup(cpu);
5103 rcutree_affinity_setting(cpu, -1);
5104
5105 // Stop-machine done, so allow nohz_full to disable tick.
5106 tick_dep_clear(TICK_DEP_BIT_RCU);
5107 return 0;
5108 }
5109
5110 /*
5111 * Mark the specified CPU as being online so that subsequent grace periods
5112 * (both expedited and normal) will wait on it. Note that this means that
5113 * incoming CPUs are not allowed to use RCU read-side critical sections
5114 * until this function is called. Failing to observe this restriction
5115 * will result in lockdep splats.
5116 *
5117 * Note that this function is special in that it is invoked directly
5118 * from the incoming CPU rather than from the cpuhp_step mechanism.
5119 * This is because this function must be invoked at a precise location.
5120 * This incoming CPU must not have enabled interrupts yet.
5121 *
5122 * This mirrors the effects of rcutree_report_cpu_dead().
5123 */
rcutree_report_cpu_starting(unsigned int cpu)5124 void rcutree_report_cpu_starting(unsigned int cpu)
5125 {
5126 unsigned long mask;
5127 struct rcu_data *rdp;
5128 struct rcu_node *rnp;
5129 bool newcpu;
5130
5131 lockdep_assert_irqs_disabled();
5132 rdp = per_cpu_ptr(&rcu_data, cpu);
5133 if (rdp->cpu_started)
5134 return;
5135 rdp->cpu_started = true;
5136
5137 rnp = rdp->mynode;
5138 mask = rdp->grpmask;
5139 arch_spin_lock(&rcu_state.ofl_lock);
5140 rcu_watching_online();
5141 raw_spin_lock(&rcu_state.barrier_lock);
5142 raw_spin_lock_rcu_node(rnp);
5143 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
5144 raw_spin_unlock(&rcu_state.barrier_lock);
5145 newcpu = !(rnp->expmaskinitnext & mask);
5146 rnp->expmaskinitnext |= mask;
5147 /* Allow lockless access for expedited grace periods. */
5148 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
5149 ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
5150 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
5151 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
5152 rdp->rcu_onl_gp_state = READ_ONCE(rcu_state.gp_state);
5153
5154 /* An incoming CPU should never be blocking a grace period. */
5155 if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
5156 /* rcu_report_qs_rnp() *really* wants some flags to restore */
5157 unsigned long flags;
5158
5159 local_irq_save(flags);
5160 rcu_disable_urgency_upon_qs(rdp);
5161 /* Report QS -after- changing ->qsmaskinitnext! */
5162 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
5163 } else {
5164 raw_spin_unlock_rcu_node(rnp);
5165 }
5166 arch_spin_unlock(&rcu_state.ofl_lock);
5167 smp_store_release(&rdp->beenonline, true);
5168 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
5169 }
5170
5171 /*
5172 * The outgoing function has no further need of RCU, so remove it from
5173 * the rcu_node tree's ->qsmaskinitnext bit masks.
5174 *
5175 * Note that this function is special in that it is invoked directly
5176 * from the outgoing CPU rather than from the cpuhp_step mechanism.
5177 * This is because this function must be invoked at a precise location.
5178 *
5179 * This mirrors the effect of rcutree_report_cpu_starting().
5180 */
rcutree_report_cpu_dead(void)5181 void rcutree_report_cpu_dead(void)
5182 {
5183 unsigned long flags;
5184 unsigned long mask;
5185 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
5186 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
5187
5188 /*
5189 * IRQS must be disabled from now on and until the CPU dies, or an interrupt
5190 * may introduce a new READ-side while it is actually off the QS masks.
5191 */
5192 lockdep_assert_irqs_disabled();
5193 // Do any dangling deferred wakeups.
5194 do_nocb_deferred_wakeup(rdp);
5195
5196 rcu_preempt_deferred_qs(current);
5197
5198 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
5199 mask = rdp->grpmask;
5200 arch_spin_lock(&rcu_state.ofl_lock);
5201 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
5202 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
5203 rdp->rcu_ofl_gp_state = READ_ONCE(rcu_state.gp_state);
5204 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
5205 /* Report quiescent state -before- changing ->qsmaskinitnext! */
5206 rcu_disable_urgency_upon_qs(rdp);
5207 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
5208 raw_spin_lock_irqsave_rcu_node(rnp, flags);
5209 }
5210 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
5211 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
5212 arch_spin_unlock(&rcu_state.ofl_lock);
5213 rdp->cpu_started = false;
5214 }
5215
5216 #ifdef CONFIG_HOTPLUG_CPU
5217 /*
5218 * The outgoing CPU has just passed through the dying-idle state, and we
5219 * are being invoked from the CPU that was IPIed to continue the offline
5220 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
5221 */
rcutree_migrate_callbacks(int cpu)5222 void rcutree_migrate_callbacks(int cpu)
5223 {
5224 unsigned long flags;
5225 struct rcu_data *my_rdp;
5226 struct rcu_node *my_rnp;
5227 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
5228 bool needwake;
5229
5230 if (rcu_rdp_is_offloaded(rdp))
5231 return;
5232
5233 raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags);
5234 if (rcu_segcblist_empty(&rdp->cblist)) {
5235 raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags);
5236 return; /* No callbacks to migrate. */
5237 }
5238
5239 WARN_ON_ONCE(rcu_rdp_cpu_online(rdp));
5240 rcu_barrier_entrain(rdp);
5241 my_rdp = this_cpu_ptr(&rcu_data);
5242 my_rnp = my_rdp->mynode;
5243 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
5244 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies, false));
5245 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
5246 /* Leverage recent GPs and set GP for new callbacks. */
5247 needwake = rcu_advance_cbs(my_rnp, rdp) ||
5248 rcu_advance_cbs(my_rnp, my_rdp);
5249 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
5250 raw_spin_unlock(&rcu_state.barrier_lock); /* irqs remain disabled. */
5251 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
5252 rcu_segcblist_disable(&rdp->cblist);
5253 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != !rcu_segcblist_n_cbs(&my_rdp->cblist));
5254 check_cb_ovld_locked(my_rdp, my_rnp);
5255 if (rcu_rdp_is_offloaded(my_rdp)) {
5256 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
5257 __call_rcu_nocb_wake(my_rdp, true, flags);
5258 } else {
5259 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
5260 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
5261 }
5262 local_irq_restore(flags);
5263 if (needwake)
5264 rcu_gp_kthread_wake();
5265 lockdep_assert_irqs_enabled();
5266 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
5267 !rcu_segcblist_empty(&rdp->cblist),
5268 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
5269 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
5270 rcu_segcblist_first_cb(&rdp->cblist));
5271 }
5272
5273 /*
5274 * The CPU has been completely removed, and some other CPU is reporting
5275 * this fact from process context. Do the remainder of the cleanup.
5276 * There can only be one CPU hotplug operation at a time, so no need for
5277 * explicit locking.
5278 */
rcutree_dead_cpu(unsigned int cpu)5279 int rcutree_dead_cpu(unsigned int cpu)
5280 {
5281 ASSERT_EXCLUSIVE_WRITER(rcu_state.n_online_cpus);
5282 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
5283 // Stop-machine done, so allow nohz_full to disable tick.
5284 tick_dep_clear(TICK_DEP_BIT_RCU);
5285 return 0;
5286 }
5287
5288 /*
5289 * Near the end of the offline process. Trace the fact that this CPU
5290 * is going offline.
5291 */
rcutree_dying_cpu(unsigned int cpu)5292 int rcutree_dying_cpu(unsigned int cpu)
5293 {
5294 bool blkd;
5295 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
5296 struct rcu_node *rnp = rdp->mynode;
5297
5298 blkd = !!(READ_ONCE(rnp->qsmask) & rdp->grpmask);
5299 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
5300 blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
5301 return 0;
5302 }
5303
5304 /*
5305 * Near the beginning of the process. The CPU is still very much alive
5306 * with pretty much all services enabled.
5307 */
rcutree_offline_cpu(unsigned int cpu)5308 int rcutree_offline_cpu(unsigned int cpu)
5309 {
5310 unsigned long flags;
5311 struct rcu_data *rdp;
5312 struct rcu_node *rnp;
5313
5314 rdp = per_cpu_ptr(&rcu_data, cpu);
5315 rnp = rdp->mynode;
5316 raw_spin_lock_irqsave_rcu_node(rnp, flags);
5317 rnp->ffmask &= ~rdp->grpmask;
5318 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
5319
5320 rcutree_affinity_setting(cpu, cpu);
5321
5322 // nohz_full CPUs need the tick for stop-machine to work quickly
5323 tick_dep_set(TICK_DEP_BIT_RCU);
5324 return 0;
5325 }
5326 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
5327
5328 /*
5329 * On non-huge systems, use expedited RCU grace periods to make suspend
5330 * and hibernation run faster.
5331 */
rcu_pm_notify(struct notifier_block * self,unsigned long action,void * hcpu)5332 static int rcu_pm_notify(struct notifier_block *self,
5333 unsigned long action, void *hcpu)
5334 {
5335 switch (action) {
5336 case PM_HIBERNATION_PREPARE:
5337 case PM_SUSPEND_PREPARE:
5338 rcu_async_hurry();
5339 rcu_expedite_gp();
5340 break;
5341 case PM_POST_HIBERNATION:
5342 case PM_POST_SUSPEND:
5343 rcu_unexpedite_gp();
5344 rcu_async_relax();
5345 break;
5346 default:
5347 break;
5348 }
5349 return NOTIFY_OK;
5350 }
5351
5352 /*
5353 * Spawn the kthreads that handle RCU's grace periods.
5354 */
rcu_spawn_gp_kthread(void)5355 static int __init rcu_spawn_gp_kthread(void)
5356 {
5357 unsigned long flags;
5358 struct rcu_node *rnp;
5359 struct sched_param sp;
5360 struct task_struct *t;
5361 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
5362
5363 rcu_scheduler_fully_active = 1;
5364 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
5365 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
5366 return 0;
5367 if (kthread_prio) {
5368 sp.sched_priority = kthread_prio;
5369 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
5370 }
5371 rnp = rcu_get_root();
5372 raw_spin_lock_irqsave_rcu_node(rnp, flags);
5373 WRITE_ONCE(rcu_state.gp_activity, jiffies);
5374 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
5375 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
5376 smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
5377 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
5378 wake_up_process(t);
5379 /* This is a pre-SMP initcall, we expect a single CPU */
5380 WARN_ON(num_online_cpus() > 1);
5381 /*
5382 * Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu()
5383 * due to rcu_scheduler_fully_active.
5384 */
5385 rcu_spawn_cpu_nocb_kthread(smp_processor_id());
5386 rcu_spawn_rnp_kthreads(rdp->mynode);
5387 rcu_spawn_core_kthreads();
5388 /* Create kthread worker for expedited GPs */
5389 rcu_start_exp_gp_kworker();
5390 return 0;
5391 }
5392 early_initcall(rcu_spawn_gp_kthread);
5393
5394 /*
5395 * This function is invoked towards the end of the scheduler's
5396 * initialization process. Before this is called, the idle task might
5397 * contain synchronous grace-period primitives (during which time, this idle
5398 * task is booting the system, and such primitives are no-ops). After this
5399 * function is called, any synchronous grace-period primitives are run as
5400 * expedited, with the requesting task driving the grace period forward.
5401 * A later core_initcall() rcu_set_runtime_mode() will switch to full
5402 * runtime RCU functionality.
5403 */
rcu_scheduler_starting(void)5404 void rcu_scheduler_starting(void)
5405 {
5406 unsigned long flags;
5407 struct rcu_node *rnp;
5408
5409 WARN_ON(num_online_cpus() != 1);
5410 WARN_ON(nr_context_switches() > 0);
5411 rcu_test_sync_prims();
5412
5413 // Fix up the ->gp_seq counters.
5414 local_irq_save(flags);
5415 rcu_for_each_node_breadth_first(rnp)
5416 rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq;
5417 local_irq_restore(flags);
5418
5419 // Switch out of early boot mode.
5420 rcu_scheduler_active = RCU_SCHEDULER_INIT;
5421 rcu_test_sync_prims();
5422 }
5423
5424 /*
5425 * Helper function for rcu_init() that initializes the rcu_state structure.
5426 */
rcu_init_one(void)5427 static void __init rcu_init_one(void)
5428 {
5429 static const char * const buf[] = RCU_NODE_NAME_INIT;
5430 static const char * const fqs[] = RCU_FQS_NAME_INIT;
5431 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
5432 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
5433
5434 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
5435 int cpustride = 1;
5436 int i;
5437 int j;
5438 struct rcu_node *rnp;
5439
5440 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
5441
5442 /* Silence gcc 4.8 false positive about array index out of range. */
5443 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
5444 panic("rcu_init_one: rcu_num_lvls out of range");
5445
5446 /* Initialize the level-tracking arrays. */
5447
5448 for (i = 1; i < rcu_num_lvls; i++)
5449 rcu_state.level[i] =
5450 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
5451 rcu_init_levelspread(levelspread, num_rcu_lvl);
5452
5453 /* Initialize the elements themselves, starting from the leaves. */
5454
5455 for (i = rcu_num_lvls - 1; i >= 0; i--) {
5456 cpustride *= levelspread[i];
5457 rnp = rcu_state.level[i];
5458 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
5459 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
5460 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
5461 &rcu_node_class[i], buf[i]);
5462 raw_spin_lock_init(&rnp->fqslock);
5463 lockdep_set_class_and_name(&rnp->fqslock,
5464 &rcu_fqs_class[i], fqs[i]);
5465 rnp->gp_seq = rcu_state.gp_seq;
5466 rnp->gp_seq_needed = rcu_state.gp_seq;
5467 rnp->completedqs = rcu_state.gp_seq;
5468 rnp->qsmask = 0;
5469 rnp->qsmaskinit = 0;
5470 rnp->grplo = j * cpustride;
5471 rnp->grphi = (j + 1) * cpustride - 1;
5472 if (rnp->grphi >= nr_cpu_ids)
5473 rnp->grphi = nr_cpu_ids - 1;
5474 if (i == 0) {
5475 rnp->grpnum = 0;
5476 rnp->grpmask = 0;
5477 rnp->parent = NULL;
5478 } else {
5479 rnp->grpnum = j % levelspread[i - 1];
5480 rnp->grpmask = BIT(rnp->grpnum);
5481 rnp->parent = rcu_state.level[i - 1] +
5482 j / levelspread[i - 1];
5483 }
5484 rnp->level = i;
5485 INIT_LIST_HEAD(&rnp->blkd_tasks);
5486 rcu_init_one_nocb(rnp);
5487 init_waitqueue_head(&rnp->exp_wq[0]);
5488 init_waitqueue_head(&rnp->exp_wq[1]);
5489 init_waitqueue_head(&rnp->exp_wq[2]);
5490 init_waitqueue_head(&rnp->exp_wq[3]);
5491 spin_lock_init(&rnp->exp_lock);
5492 mutex_init(&rnp->kthread_mutex);
5493 raw_spin_lock_init(&rnp->exp_poll_lock);
5494 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
5495 INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp);
5496 }
5497 }
5498
5499 init_swait_queue_head(&rcu_state.gp_wq);
5500 init_swait_queue_head(&rcu_state.expedited_wq);
5501 rnp = rcu_first_leaf_node();
5502 for_each_possible_cpu(i) {
5503 while (i > rnp->grphi)
5504 rnp++;
5505 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
5506 per_cpu_ptr(&rcu_data, i)->barrier_head.next =
5507 &per_cpu_ptr(&rcu_data, i)->barrier_head;
5508 rcu_boot_init_percpu_data(i);
5509 }
5510 }
5511
5512 /*
5513 * Force priority from the kernel command-line into range.
5514 */
sanitize_kthread_prio(void)5515 static void __init sanitize_kthread_prio(void)
5516 {
5517 int kthread_prio_in = kthread_prio;
5518
5519 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
5520 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
5521 kthread_prio = 2;
5522 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
5523 kthread_prio = 1;
5524 else if (kthread_prio < 0)
5525 kthread_prio = 0;
5526 else if (kthread_prio > 99)
5527 kthread_prio = 99;
5528
5529 if (kthread_prio != kthread_prio_in)
5530 pr_alert("%s: Limited prio to %d from %d\n",
5531 __func__, kthread_prio, kthread_prio_in);
5532 }
5533
5534 /*
5535 * Compute the rcu_node tree geometry from kernel parameters. This cannot
5536 * replace the definitions in tree.h because those are needed to size
5537 * the ->node array in the rcu_state structure.
5538 */
rcu_init_geometry(void)5539 void rcu_init_geometry(void)
5540 {
5541 ulong d;
5542 int i;
5543 static unsigned long old_nr_cpu_ids;
5544 int rcu_capacity[RCU_NUM_LVLS];
5545 static bool initialized;
5546
5547 if (initialized) {
5548 /*
5549 * Warn if setup_nr_cpu_ids() had not yet been invoked,
5550 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
5551 */
5552 WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
5553 return;
5554 }
5555
5556 old_nr_cpu_ids = nr_cpu_ids;
5557 initialized = true;
5558
5559 /*
5560 * Initialize any unspecified boot parameters.
5561 * The default values of jiffies_till_first_fqs and
5562 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
5563 * value, which is a function of HZ, then adding one for each
5564 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
5565 */
5566 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
5567 if (jiffies_till_first_fqs == ULONG_MAX)
5568 jiffies_till_first_fqs = d;
5569 if (jiffies_till_next_fqs == ULONG_MAX)
5570 jiffies_till_next_fqs = d;
5571 adjust_jiffies_till_sched_qs();
5572
5573 /* If the compile-time values are accurate, just leave. */
5574 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
5575 nr_cpu_ids == NR_CPUS)
5576 return;
5577 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
5578 rcu_fanout_leaf, nr_cpu_ids);
5579
5580 /*
5581 * The boot-time rcu_fanout_leaf parameter must be at least two
5582 * and cannot exceed the number of bits in the rcu_node masks.
5583 * Complain and fall back to the compile-time values if this
5584 * limit is exceeded.
5585 */
5586 if (rcu_fanout_leaf < 2 || rcu_fanout_leaf > BITS_PER_LONG) {
5587 rcu_fanout_leaf = RCU_FANOUT_LEAF;
5588 WARN_ON(1);
5589 return;
5590 }
5591
5592 /*
5593 * Compute number of nodes that can be handled an rcu_node tree
5594 * with the given number of levels.
5595 */
5596 rcu_capacity[0] = rcu_fanout_leaf;
5597 for (i = 1; i < RCU_NUM_LVLS; i++)
5598 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
5599
5600 /*
5601 * The tree must be able to accommodate the configured number of CPUs.
5602 * If this limit is exceeded, fall back to the compile-time values.
5603 */
5604 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
5605 rcu_fanout_leaf = RCU_FANOUT_LEAF;
5606 WARN_ON(1);
5607 return;
5608 }
5609
5610 /* Calculate the number of levels in the tree. */
5611 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
5612 }
5613 rcu_num_lvls = i + 1;
5614
5615 /* Calculate the number of rcu_nodes at each level of the tree. */
5616 for (i = 0; i < rcu_num_lvls; i++) {
5617 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
5618 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
5619 }
5620
5621 /* Calculate the total number of rcu_node structures. */
5622 rcu_num_nodes = 0;
5623 for (i = 0; i < rcu_num_lvls; i++)
5624 rcu_num_nodes += num_rcu_lvl[i];
5625 }
5626
5627 /*
5628 * Dump out the structure of the rcu_node combining tree associated
5629 * with the rcu_state structure.
5630 */
rcu_dump_rcu_node_tree(void)5631 static void __init rcu_dump_rcu_node_tree(void)
5632 {
5633 int level = 0;
5634 struct rcu_node *rnp;
5635
5636 pr_info("rcu_node tree layout dump\n");
5637 pr_info(" ");
5638 rcu_for_each_node_breadth_first(rnp) {
5639 if (rnp->level != level) {
5640 pr_cont("\n");
5641 pr_info(" ");
5642 level = rnp->level;
5643 }
5644 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
5645 }
5646 pr_cont("\n");
5647 }
5648
5649 struct workqueue_struct *rcu_gp_wq;
5650
kfree_rcu_batch_init(void)5651 static void __init kfree_rcu_batch_init(void)
5652 {
5653 int cpu;
5654 int i, j;
5655 struct shrinker *kfree_rcu_shrinker;
5656
5657 /* Clamp it to [0:100] seconds interval. */
5658 if (rcu_delay_page_cache_fill_msec < 0 ||
5659 rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
5660
5661 rcu_delay_page_cache_fill_msec =
5662 clamp(rcu_delay_page_cache_fill_msec, 0,
5663 (int) (100 * MSEC_PER_SEC));
5664
5665 pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
5666 rcu_delay_page_cache_fill_msec);
5667 }
5668
5669 for_each_possible_cpu(cpu) {
5670 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
5671
5672 for (i = 0; i < KFREE_N_BATCHES; i++) {
5673 INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
5674 krcp->krw_arr[i].krcp = krcp;
5675
5676 for (j = 0; j < FREE_N_CHANNELS; j++)
5677 INIT_LIST_HEAD(&krcp->krw_arr[i].bulk_head_free[j]);
5678 }
5679
5680 for (i = 0; i < FREE_N_CHANNELS; i++)
5681 INIT_LIST_HEAD(&krcp->bulk_head[i]);
5682
5683 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
5684 INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
5685 krcp->initialized = true;
5686 }
5687
5688 kfree_rcu_shrinker = shrinker_alloc(0, "rcu-kfree");
5689 if (!kfree_rcu_shrinker) {
5690 pr_err("Failed to allocate kfree_rcu() shrinker!\n");
5691 return;
5692 }
5693
5694 kfree_rcu_shrinker->count_objects = kfree_rcu_shrink_count;
5695 kfree_rcu_shrinker->scan_objects = kfree_rcu_shrink_scan;
5696
5697 shrinker_register(kfree_rcu_shrinker);
5698 }
5699
rcu_init(void)5700 void __init rcu_init(void)
5701 {
5702 int cpu = smp_processor_id();
5703
5704 rcu_early_boot_tests();
5705
5706 kfree_rcu_batch_init();
5707 rcu_bootup_announce();
5708 sanitize_kthread_prio();
5709 rcu_init_geometry();
5710 rcu_init_one();
5711 if (dump_tree)
5712 rcu_dump_rcu_node_tree();
5713 if (use_softirq)
5714 open_softirq(RCU_SOFTIRQ, rcu_core_si);
5715
5716 /*
5717 * We don't need protection against CPU-hotplug here because
5718 * this is called early in boot, before either interrupts
5719 * or the scheduler are operational.
5720 */
5721 pm_notifier(rcu_pm_notify, 0);
5722 WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot.
5723 rcutree_prepare_cpu(cpu);
5724 rcutree_report_cpu_starting(cpu);
5725 rcutree_online_cpu(cpu);
5726
5727 /* Create workqueue for Tree SRCU and for expedited GPs. */
5728 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
5729 WARN_ON(!rcu_gp_wq);
5730
5731 sync_wq = alloc_workqueue("sync_wq", WQ_MEM_RECLAIM, 0);
5732 WARN_ON(!sync_wq);
5733
5734 /* Fill in default value for rcutree.qovld boot parameter. */
5735 /* -After- the rcu_node ->lock fields are initialized! */
5736 if (qovld < 0)
5737 qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
5738 else
5739 qovld_calc = qovld;
5740
5741 // Kick-start in case any polled grace periods started early.
5742 (void)start_poll_synchronize_rcu_expedited();
5743
5744 rcu_test_sync_prims();
5745
5746 tasks_cblist_init_generic();
5747 }
5748
5749 #include "tree_stall.h"
5750 #include "tree_exp.h"
5751 #include "tree_nocb.h"
5752 #include "tree_plugin.h"
5753