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