]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Read-Copy Update mechanism for mutual exclusion | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public License | |
15 | * along with this program; if not, you can access it online at | |
16 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
17 | * | |
18 | * Copyright IBM Corporation, 2008 | |
19 | * | |
20 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> | |
21 | * Manfred Spraul <manfred@colorfullife.com> | |
22 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version | |
23 | * | |
24 | * Based on the original work by Paul McKenney <paulmck@us.ibm.com> | |
25 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. | |
26 | * | |
27 | * For detailed explanation of Read-Copy Update mechanism see - | |
28 | * Documentation/RCU | |
29 | */ | |
30 | #include <linux/types.h> | |
31 | #include <linux/kernel.h> | |
32 | #include <linux/init.h> | |
33 | #include <linux/spinlock.h> | |
34 | #include <linux/smp.h> | |
35 | #include <linux/rcupdate.h> | |
36 | #include <linux/interrupt.h> | |
37 | #include <linux/sched.h> | |
38 | #include <linux/nmi.h> | |
39 | #include <linux/atomic.h> | |
40 | #include <linux/bitops.h> | |
41 | #include <linux/export.h> | |
42 | #include <linux/completion.h> | |
43 | #include <linux/moduleparam.h> | |
44 | #include <linux/module.h> | |
45 | #include <linux/percpu.h> | |
46 | #include <linux/notifier.h> | |
47 | #include <linux/cpu.h> | |
48 | #include <linux/mutex.h> | |
49 | #include <linux/time.h> | |
50 | #include <linux/kernel_stat.h> | |
51 | #include <linux/wait.h> | |
52 | #include <linux/kthread.h> | |
53 | #include <linux/prefetch.h> | |
54 | #include <linux/delay.h> | |
55 | #include <linux/stop_machine.h> | |
56 | #include <linux/random.h> | |
57 | #include <linux/trace_events.h> | |
58 | #include <linux/suspend.h> | |
59 | ||
60 | #include "tree.h" | |
61 | #include "rcu.h" | |
62 | ||
63 | MODULE_ALIAS("rcutree"); | |
64 | #ifdef MODULE_PARAM_PREFIX | |
65 | #undef MODULE_PARAM_PREFIX | |
66 | #endif | |
67 | #define MODULE_PARAM_PREFIX "rcutree." | |
68 | ||
69 | /* Data structures. */ | |
70 | ||
71 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; | |
72 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; | |
73 | static struct lock_class_key rcu_exp_class[RCU_NUM_LVLS]; | |
74 | static struct lock_class_key rcu_exp_sched_class[RCU_NUM_LVLS]; | |
75 | ||
76 | /* | |
77 | * In order to export the rcu_state name to the tracing tools, it | |
78 | * needs to be added in the __tracepoint_string section. | |
79 | * This requires defining a separate variable tp_<sname>_varname | |
80 | * that points to the string being used, and this will allow | |
81 | * the tracing userspace tools to be able to decipher the string | |
82 | * address to the matching string. | |
83 | */ | |
84 | #ifdef CONFIG_TRACING | |
85 | # define DEFINE_RCU_TPS(sname) \ | |
86 | static char sname##_varname[] = #sname; \ | |
87 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; | |
88 | # define RCU_STATE_NAME(sname) sname##_varname | |
89 | #else | |
90 | # define DEFINE_RCU_TPS(sname) | |
91 | # define RCU_STATE_NAME(sname) __stringify(sname) | |
92 | #endif | |
93 | ||
94 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ | |
95 | DEFINE_RCU_TPS(sname) \ | |
96 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \ | |
97 | struct rcu_state sname##_state = { \ | |
98 | .level = { &sname##_state.node[0] }, \ | |
99 | .rda = &sname##_data, \ | |
100 | .call = cr, \ | |
101 | .fqs_state = RCU_GP_IDLE, \ | |
102 | .gpnum = 0UL - 300UL, \ | |
103 | .completed = 0UL - 300UL, \ | |
104 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ | |
105 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ | |
106 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
107 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ | |
108 | .name = RCU_STATE_NAME(sname), \ | |
109 | .abbr = sabbr, \ | |
110 | } | |
111 | ||
112 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); | |
113 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); | |
114 | ||
115 | static struct rcu_state *const rcu_state_p; | |
116 | static struct rcu_data __percpu *const rcu_data_p; | |
117 | LIST_HEAD(rcu_struct_flavors); | |
118 | ||
119 | /* Dump rcu_node combining tree at boot to verify correct setup. */ | |
120 | static bool dump_tree; | |
121 | module_param(dump_tree, bool, 0444); | |
122 | /* Control rcu_node-tree auto-balancing at boot time. */ | |
123 | static bool rcu_fanout_exact; | |
124 | module_param(rcu_fanout_exact, bool, 0444); | |
125 | /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */ | |
126 | static int rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
127 | module_param(rcu_fanout_leaf, int, 0444); | |
128 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; | |
129 | /* Number of rcu_nodes at specified level. */ | |
130 | static int num_rcu_lvl[] = NUM_RCU_LVL_INIT; | |
131 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ | |
132 | ||
133 | /* | |
134 | * The rcu_scheduler_active variable transitions from zero to one just | |
135 | * before the first task is spawned. So when this variable is zero, RCU | |
136 | * can assume that there is but one task, allowing RCU to (for example) | |
137 | * optimize synchronize_sched() to a simple barrier(). When this variable | |
138 | * is one, RCU must actually do all the hard work required to detect real | |
139 | * grace periods. This variable is also used to suppress boot-time false | |
140 | * positives from lockdep-RCU error checking. | |
141 | */ | |
142 | int rcu_scheduler_active __read_mostly; | |
143 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
144 | ||
145 | /* | |
146 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
147 | * during the early_initcall() processing, which is after the scheduler | |
148 | * is capable of creating new tasks. So RCU processing (for example, | |
149 | * creating tasks for RCU priority boosting) must be delayed until after | |
150 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
151 | * currently delay invocation of any RCU callbacks until after this point. | |
152 | * | |
153 | * It might later prove better for people registering RCU callbacks during | |
154 | * early boot to take responsibility for these callbacks, but one step at | |
155 | * a time. | |
156 | */ | |
157 | static int rcu_scheduler_fully_active __read_mostly; | |
158 | ||
159 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); | |
160 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); | |
161 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); | |
162 | static void invoke_rcu_core(void); | |
163 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
164 | ||
165 | /* rcuc/rcub kthread realtime priority */ | |
166 | #ifdef CONFIG_RCU_KTHREAD_PRIO | |
167 | static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO; | |
168 | #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */ | |
169 | static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0; | |
170 | #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */ | |
171 | module_param(kthread_prio, int, 0644); | |
172 | ||
173 | /* Delay in jiffies for grace-period initialization delays, debug only. */ | |
174 | ||
175 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT | |
176 | static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY; | |
177 | module_param(gp_preinit_delay, int, 0644); | |
178 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ | |
179 | static const int gp_preinit_delay; | |
180 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ | |
181 | ||
182 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT | |
183 | static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY; | |
184 | module_param(gp_init_delay, int, 0644); | |
185 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ | |
186 | static const int gp_init_delay; | |
187 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ | |
188 | ||
189 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP | |
190 | static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY; | |
191 | module_param(gp_cleanup_delay, int, 0644); | |
192 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ | |
193 | static const int gp_cleanup_delay; | |
194 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ | |
195 | ||
196 | /* | |
197 | * Number of grace periods between delays, normalized by the duration of | |
198 | * the delay. The longer the the delay, the more the grace periods between | |
199 | * each delay. The reason for this normalization is that it means that, | |
200 | * for non-zero delays, the overall slowdown of grace periods is constant | |
201 | * regardless of the duration of the delay. This arrangement balances | |
202 | * the need for long delays to increase some race probabilities with the | |
203 | * need for fast grace periods to increase other race probabilities. | |
204 | */ | |
205 | #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */ | |
206 | ||
207 | /* | |
208 | * Track the rcutorture test sequence number and the update version | |
209 | * number within a given test. The rcutorture_testseq is incremented | |
210 | * on every rcutorture module load and unload, so has an odd value | |
211 | * when a test is running. The rcutorture_vernum is set to zero | |
212 | * when rcutorture starts and is incremented on each rcutorture update. | |
213 | * These variables enable correlating rcutorture output with the | |
214 | * RCU tracing information. | |
215 | */ | |
216 | unsigned long rcutorture_testseq; | |
217 | unsigned long rcutorture_vernum; | |
218 | ||
219 | /* | |
220 | * Compute the mask of online CPUs for the specified rcu_node structure. | |
221 | * This will not be stable unless the rcu_node structure's ->lock is | |
222 | * held, but the bit corresponding to the current CPU will be stable | |
223 | * in most contexts. | |
224 | */ | |
225 | unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) | |
226 | { | |
227 | return READ_ONCE(rnp->qsmaskinitnext); | |
228 | } | |
229 | ||
230 | /* | |
231 | * Return true if an RCU grace period is in progress. The READ_ONCE()s | |
232 | * permit this function to be invoked without holding the root rcu_node | |
233 | * structure's ->lock, but of course results can be subject to change. | |
234 | */ | |
235 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
236 | { | |
237 | return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum); | |
238 | } | |
239 | ||
240 | /* | |
241 | * Note a quiescent state. Because we do not need to know | |
242 | * how many quiescent states passed, just if there was at least | |
243 | * one since the start of the grace period, this just sets a flag. | |
244 | * The caller must have disabled preemption. | |
245 | */ | |
246 | void rcu_sched_qs(void) | |
247 | { | |
248 | if (!__this_cpu_read(rcu_sched_data.passed_quiesce)) { | |
249 | trace_rcu_grace_period(TPS("rcu_sched"), | |
250 | __this_cpu_read(rcu_sched_data.gpnum), | |
251 | TPS("cpuqs")); | |
252 | __this_cpu_write(rcu_sched_data.passed_quiesce, 1); | |
253 | } | |
254 | } | |
255 | ||
256 | void rcu_bh_qs(void) | |
257 | { | |
258 | if (!__this_cpu_read(rcu_bh_data.passed_quiesce)) { | |
259 | trace_rcu_grace_period(TPS("rcu_bh"), | |
260 | __this_cpu_read(rcu_bh_data.gpnum), | |
261 | TPS("cpuqs")); | |
262 | __this_cpu_write(rcu_bh_data.passed_quiesce, 1); | |
263 | } | |
264 | } | |
265 | ||
266 | static DEFINE_PER_CPU(int, rcu_sched_qs_mask); | |
267 | ||
268 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { | |
269 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, | |
270 | .dynticks = ATOMIC_INIT(1), | |
271 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
272 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
273 | .dynticks_idle = ATOMIC_INIT(1), | |
274 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
275 | }; | |
276 | ||
277 | DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr); | |
278 | EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr); | |
279 | ||
280 | /* | |
281 | * Let the RCU core know that this CPU has gone through the scheduler, | |
282 | * which is a quiescent state. This is called when the need for a | |
283 | * quiescent state is urgent, so we burn an atomic operation and full | |
284 | * memory barriers to let the RCU core know about it, regardless of what | |
285 | * this CPU might (or might not) do in the near future. | |
286 | * | |
287 | * We inform the RCU core by emulating a zero-duration dyntick-idle | |
288 | * period, which we in turn do by incrementing the ->dynticks counter | |
289 | * by two. | |
290 | */ | |
291 | static void rcu_momentary_dyntick_idle(void) | |
292 | { | |
293 | unsigned long flags; | |
294 | struct rcu_data *rdp; | |
295 | struct rcu_dynticks *rdtp; | |
296 | int resched_mask; | |
297 | struct rcu_state *rsp; | |
298 | ||
299 | local_irq_save(flags); | |
300 | ||
301 | /* | |
302 | * Yes, we can lose flag-setting operations. This is OK, because | |
303 | * the flag will be set again after some delay. | |
304 | */ | |
305 | resched_mask = raw_cpu_read(rcu_sched_qs_mask); | |
306 | raw_cpu_write(rcu_sched_qs_mask, 0); | |
307 | ||
308 | /* Find the flavor that needs a quiescent state. */ | |
309 | for_each_rcu_flavor(rsp) { | |
310 | rdp = raw_cpu_ptr(rsp->rda); | |
311 | if (!(resched_mask & rsp->flavor_mask)) | |
312 | continue; | |
313 | smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */ | |
314 | if (READ_ONCE(rdp->mynode->completed) != | |
315 | READ_ONCE(rdp->cond_resched_completed)) | |
316 | continue; | |
317 | ||
318 | /* | |
319 | * Pretend to be momentarily idle for the quiescent state. | |
320 | * This allows the grace-period kthread to record the | |
321 | * quiescent state, with no need for this CPU to do anything | |
322 | * further. | |
323 | */ | |
324 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
325 | smp_mb__before_atomic(); /* Earlier stuff before QS. */ | |
326 | atomic_add(2, &rdtp->dynticks); /* QS. */ | |
327 | smp_mb__after_atomic(); /* Later stuff after QS. */ | |
328 | break; | |
329 | } | |
330 | local_irq_restore(flags); | |
331 | } | |
332 | ||
333 | /* | |
334 | * Note a context switch. This is a quiescent state for RCU-sched, | |
335 | * and requires special handling for preemptible RCU. | |
336 | * The caller must have disabled preemption. | |
337 | */ | |
338 | void rcu_note_context_switch(void) | |
339 | { | |
340 | barrier(); /* Avoid RCU read-side critical sections leaking down. */ | |
341 | trace_rcu_utilization(TPS("Start context switch")); | |
342 | rcu_sched_qs(); | |
343 | rcu_preempt_note_context_switch(); | |
344 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) | |
345 | rcu_momentary_dyntick_idle(); | |
346 | trace_rcu_utilization(TPS("End context switch")); | |
347 | barrier(); /* Avoid RCU read-side critical sections leaking up. */ | |
348 | } | |
349 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); | |
350 | ||
351 | /* | |
352 | * Register a quiescent state for all RCU flavors. If there is an | |
353 | * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight | |
354 | * dyntick-idle quiescent state visible to other CPUs (but only for those | |
355 | * RCU flavors in desperate need of a quiescent state, which will normally | |
356 | * be none of them). Either way, do a lightweight quiescent state for | |
357 | * all RCU flavors. | |
358 | * | |
359 | * The barrier() calls are redundant in the common case when this is | |
360 | * called externally, but just in case this is called from within this | |
361 | * file. | |
362 | * | |
363 | */ | |
364 | void rcu_all_qs(void) | |
365 | { | |
366 | barrier(); /* Avoid RCU read-side critical sections leaking down. */ | |
367 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) | |
368 | rcu_momentary_dyntick_idle(); | |
369 | this_cpu_inc(rcu_qs_ctr); | |
370 | barrier(); /* Avoid RCU read-side critical sections leaking up. */ | |
371 | } | |
372 | EXPORT_SYMBOL_GPL(rcu_all_qs); | |
373 | ||
374 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ | |
375 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
376 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
377 | ||
378 | module_param(blimit, long, 0444); | |
379 | module_param(qhimark, long, 0444); | |
380 | module_param(qlowmark, long, 0444); | |
381 | ||
382 | static ulong jiffies_till_first_fqs = ULONG_MAX; | |
383 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
384 | ||
385 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
386 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
387 | ||
388 | /* | |
389 | * How long the grace period must be before we start recruiting | |
390 | * quiescent-state help from rcu_note_context_switch(). | |
391 | */ | |
392 | static ulong jiffies_till_sched_qs = HZ / 20; | |
393 | module_param(jiffies_till_sched_qs, ulong, 0644); | |
394 | ||
395 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, | |
396 | struct rcu_data *rdp); | |
397 | static void force_qs_rnp(struct rcu_state *rsp, | |
398 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
399 | unsigned long *maxj), | |
400 | bool *isidle, unsigned long *maxj); | |
401 | static void force_quiescent_state(struct rcu_state *rsp); | |
402 | static int rcu_pending(void); | |
403 | ||
404 | /* | |
405 | * Return the number of RCU batches started thus far for debug & stats. | |
406 | */ | |
407 | unsigned long rcu_batches_started(void) | |
408 | { | |
409 | return rcu_state_p->gpnum; | |
410 | } | |
411 | EXPORT_SYMBOL_GPL(rcu_batches_started); | |
412 | ||
413 | /* | |
414 | * Return the number of RCU-sched batches started thus far for debug & stats. | |
415 | */ | |
416 | unsigned long rcu_batches_started_sched(void) | |
417 | { | |
418 | return rcu_sched_state.gpnum; | |
419 | } | |
420 | EXPORT_SYMBOL_GPL(rcu_batches_started_sched); | |
421 | ||
422 | /* | |
423 | * Return the number of RCU BH batches started thus far for debug & stats. | |
424 | */ | |
425 | unsigned long rcu_batches_started_bh(void) | |
426 | { | |
427 | return rcu_bh_state.gpnum; | |
428 | } | |
429 | EXPORT_SYMBOL_GPL(rcu_batches_started_bh); | |
430 | ||
431 | /* | |
432 | * Return the number of RCU batches completed thus far for debug & stats. | |
433 | */ | |
434 | unsigned long rcu_batches_completed(void) | |
435 | { | |
436 | return rcu_state_p->completed; | |
437 | } | |
438 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
439 | ||
440 | /* | |
441 | * Return the number of RCU-sched batches completed thus far for debug & stats. | |
442 | */ | |
443 | unsigned long rcu_batches_completed_sched(void) | |
444 | { | |
445 | return rcu_sched_state.completed; | |
446 | } | |
447 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); | |
448 | ||
449 | /* | |
450 | * Return the number of RCU BH batches completed thus far for debug & stats. | |
451 | */ | |
452 | unsigned long rcu_batches_completed_bh(void) | |
453 | { | |
454 | return rcu_bh_state.completed; | |
455 | } | |
456 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
457 | ||
458 | /* | |
459 | * Force a quiescent state. | |
460 | */ | |
461 | void rcu_force_quiescent_state(void) | |
462 | { | |
463 | force_quiescent_state(rcu_state_p); | |
464 | } | |
465 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
466 | ||
467 | /* | |
468 | * Force a quiescent state for RCU BH. | |
469 | */ | |
470 | void rcu_bh_force_quiescent_state(void) | |
471 | { | |
472 | force_quiescent_state(&rcu_bh_state); | |
473 | } | |
474 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
475 | ||
476 | /* | |
477 | * Force a quiescent state for RCU-sched. | |
478 | */ | |
479 | void rcu_sched_force_quiescent_state(void) | |
480 | { | |
481 | force_quiescent_state(&rcu_sched_state); | |
482 | } | |
483 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
484 | ||
485 | /* | |
486 | * Show the state of the grace-period kthreads. | |
487 | */ | |
488 | void show_rcu_gp_kthreads(void) | |
489 | { | |
490 | struct rcu_state *rsp; | |
491 | ||
492 | for_each_rcu_flavor(rsp) { | |
493 | pr_info("%s: wait state: %d ->state: %#lx\n", | |
494 | rsp->name, rsp->gp_state, rsp->gp_kthread->state); | |
495 | /* sched_show_task(rsp->gp_kthread); */ | |
496 | } | |
497 | } | |
498 | EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); | |
499 | ||
500 | /* | |
501 | * Record the number of times rcutorture tests have been initiated and | |
502 | * terminated. This information allows the debugfs tracing stats to be | |
503 | * correlated to the rcutorture messages, even when the rcutorture module | |
504 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
505 | * store this state in rcutorture itself. | |
506 | */ | |
507 | void rcutorture_record_test_transition(void) | |
508 | { | |
509 | rcutorture_testseq++; | |
510 | rcutorture_vernum = 0; | |
511 | } | |
512 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
513 | ||
514 | /* | |
515 | * Send along grace-period-related data for rcutorture diagnostics. | |
516 | */ | |
517 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
518 | unsigned long *gpnum, unsigned long *completed) | |
519 | { | |
520 | struct rcu_state *rsp = NULL; | |
521 | ||
522 | switch (test_type) { | |
523 | case RCU_FLAVOR: | |
524 | rsp = rcu_state_p; | |
525 | break; | |
526 | case RCU_BH_FLAVOR: | |
527 | rsp = &rcu_bh_state; | |
528 | break; | |
529 | case RCU_SCHED_FLAVOR: | |
530 | rsp = &rcu_sched_state; | |
531 | break; | |
532 | default: | |
533 | break; | |
534 | } | |
535 | if (rsp != NULL) { | |
536 | *flags = READ_ONCE(rsp->gp_flags); | |
537 | *gpnum = READ_ONCE(rsp->gpnum); | |
538 | *completed = READ_ONCE(rsp->completed); | |
539 | return; | |
540 | } | |
541 | *flags = 0; | |
542 | *gpnum = 0; | |
543 | *completed = 0; | |
544 | } | |
545 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
546 | ||
547 | /* | |
548 | * Record the number of writer passes through the current rcutorture test. | |
549 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
550 | * messages. | |
551 | */ | |
552 | void rcutorture_record_progress(unsigned long vernum) | |
553 | { | |
554 | rcutorture_vernum++; | |
555 | } | |
556 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
557 | ||
558 | /* | |
559 | * Does the CPU have callbacks ready to be invoked? | |
560 | */ | |
561 | static int | |
562 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
563 | { | |
564 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && | |
565 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
566 | } | |
567 | ||
568 | /* | |
569 | * Return the root node of the specified rcu_state structure. | |
570 | */ | |
571 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
572 | { | |
573 | return &rsp->node[0]; | |
574 | } | |
575 | ||
576 | /* | |
577 | * Is there any need for future grace periods? | |
578 | * Interrupts must be disabled. If the caller does not hold the root | |
579 | * rnp_node structure's ->lock, the results are advisory only. | |
580 | */ | |
581 | static int rcu_future_needs_gp(struct rcu_state *rsp) | |
582 | { | |
583 | struct rcu_node *rnp = rcu_get_root(rsp); | |
584 | int idx = (READ_ONCE(rnp->completed) + 1) & 0x1; | |
585 | int *fp = &rnp->need_future_gp[idx]; | |
586 | ||
587 | return READ_ONCE(*fp); | |
588 | } | |
589 | ||
590 | /* | |
591 | * Does the current CPU require a not-yet-started grace period? | |
592 | * The caller must have disabled interrupts to prevent races with | |
593 | * normal callback registry. | |
594 | */ | |
595 | static int | |
596 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
597 | { | |
598 | int i; | |
599 | ||
600 | if (rcu_gp_in_progress(rsp)) | |
601 | return 0; /* No, a grace period is already in progress. */ | |
602 | if (rcu_future_needs_gp(rsp)) | |
603 | return 1; /* Yes, a no-CBs CPU needs one. */ | |
604 | if (!rdp->nxttail[RCU_NEXT_TAIL]) | |
605 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
606 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
607 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
608 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
609 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
610 | ULONG_CMP_LT(READ_ONCE(rsp->completed), | |
611 | rdp->nxtcompleted[i])) | |
612 | return 1; /* Yes, CBs for future grace period. */ | |
613 | return 0; /* No grace period needed. */ | |
614 | } | |
615 | ||
616 | /* | |
617 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state | |
618 | * | |
619 | * If the new value of the ->dynticks_nesting counter now is zero, | |
620 | * we really have entered idle, and must do the appropriate accounting. | |
621 | * The caller must have disabled interrupts. | |
622 | */ | |
623 | static void rcu_eqs_enter_common(long long oldval, bool user) | |
624 | { | |
625 | struct rcu_state *rsp; | |
626 | struct rcu_data *rdp; | |
627 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); | |
628 | ||
629 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); | |
630 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
631 | !user && !is_idle_task(current)) { | |
632 | struct task_struct *idle __maybe_unused = | |
633 | idle_task(smp_processor_id()); | |
634 | ||
635 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); | |
636 | ftrace_dump(DUMP_ORIG); | |
637 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", | |
638 | current->pid, current->comm, | |
639 | idle->pid, idle->comm); /* must be idle task! */ | |
640 | } | |
641 | for_each_rcu_flavor(rsp) { | |
642 | rdp = this_cpu_ptr(rsp->rda); | |
643 | do_nocb_deferred_wakeup(rdp); | |
644 | } | |
645 | rcu_prepare_for_idle(); | |
646 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ | |
647 | smp_mb__before_atomic(); /* See above. */ | |
648 | atomic_inc(&rdtp->dynticks); | |
649 | smp_mb__after_atomic(); /* Force ordering with next sojourn. */ | |
650 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
651 | atomic_read(&rdtp->dynticks) & 0x1); | |
652 | rcu_dynticks_task_enter(); | |
653 | ||
654 | /* | |
655 | * It is illegal to enter an extended quiescent state while | |
656 | * in an RCU read-side critical section. | |
657 | */ | |
658 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map), | |
659 | "Illegal idle entry in RCU read-side critical section."); | |
660 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map), | |
661 | "Illegal idle entry in RCU-bh read-side critical section."); | |
662 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map), | |
663 | "Illegal idle entry in RCU-sched read-side critical section."); | |
664 | } | |
665 | ||
666 | /* | |
667 | * Enter an RCU extended quiescent state, which can be either the | |
668 | * idle loop or adaptive-tickless usermode execution. | |
669 | */ | |
670 | static void rcu_eqs_enter(bool user) | |
671 | { | |
672 | long long oldval; | |
673 | struct rcu_dynticks *rdtp; | |
674 | ||
675 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
676 | oldval = rdtp->dynticks_nesting; | |
677 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
678 | (oldval & DYNTICK_TASK_NEST_MASK) == 0); | |
679 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { | |
680 | rdtp->dynticks_nesting = 0; | |
681 | rcu_eqs_enter_common(oldval, user); | |
682 | } else { | |
683 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; | |
684 | } | |
685 | } | |
686 | ||
687 | /** | |
688 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
689 | * | |
690 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
691 | * read-side critical sections can occur. (Though RCU read-side | |
692 | * critical sections can occur in irq handlers in idle, a possibility | |
693 | * handled by irq_enter() and irq_exit().) | |
694 | * | |
695 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
696 | * the possibility of usermode upcalls having messed up our count | |
697 | * of interrupt nesting level during the prior busy period. | |
698 | */ | |
699 | void rcu_idle_enter(void) | |
700 | { | |
701 | unsigned long flags; | |
702 | ||
703 | local_irq_save(flags); | |
704 | rcu_eqs_enter(false); | |
705 | rcu_sysidle_enter(0); | |
706 | local_irq_restore(flags); | |
707 | } | |
708 | EXPORT_SYMBOL_GPL(rcu_idle_enter); | |
709 | ||
710 | #ifdef CONFIG_NO_HZ_FULL | |
711 | /** | |
712 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
713 | * | |
714 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
715 | * is permitted between this call and rcu_user_exit(). This way the | |
716 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
717 | * when the CPU runs in userspace. | |
718 | */ | |
719 | void rcu_user_enter(void) | |
720 | { | |
721 | rcu_eqs_enter(1); | |
722 | } | |
723 | #endif /* CONFIG_NO_HZ_FULL */ | |
724 | ||
725 | /** | |
726 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
727 | * | |
728 | * Exit from an interrupt handler, which might possibly result in entering | |
729 | * idle mode, in other words, leaving the mode in which read-side critical | |
730 | * sections can occur. | |
731 | * | |
732 | * This code assumes that the idle loop never does anything that might | |
733 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
734 | * architecture violates this assumption, RCU will give you what you | |
735 | * deserve, good and hard. But very infrequently and irreproducibly. | |
736 | * | |
737 | * Use things like work queues to work around this limitation. | |
738 | * | |
739 | * You have been warned. | |
740 | */ | |
741 | void rcu_irq_exit(void) | |
742 | { | |
743 | unsigned long flags; | |
744 | long long oldval; | |
745 | struct rcu_dynticks *rdtp; | |
746 | ||
747 | local_irq_save(flags); | |
748 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
749 | oldval = rdtp->dynticks_nesting; | |
750 | rdtp->dynticks_nesting--; | |
751 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
752 | rdtp->dynticks_nesting < 0); | |
753 | if (rdtp->dynticks_nesting) | |
754 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); | |
755 | else | |
756 | rcu_eqs_enter_common(oldval, true); | |
757 | rcu_sysidle_enter(1); | |
758 | local_irq_restore(flags); | |
759 | } | |
760 | ||
761 | /* | |
762 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state | |
763 | * | |
764 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
765 | * we really have exited idle, and must do the appropriate accounting. | |
766 | * The caller must have disabled interrupts. | |
767 | */ | |
768 | static void rcu_eqs_exit_common(long long oldval, int user) | |
769 | { | |
770 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); | |
771 | ||
772 | rcu_dynticks_task_exit(); | |
773 | smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */ | |
774 | atomic_inc(&rdtp->dynticks); | |
775 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
776 | smp_mb__after_atomic(); /* See above. */ | |
777 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
778 | !(atomic_read(&rdtp->dynticks) & 0x1)); | |
779 | rcu_cleanup_after_idle(); | |
780 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); | |
781 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
782 | !user && !is_idle_task(current)) { | |
783 | struct task_struct *idle __maybe_unused = | |
784 | idle_task(smp_processor_id()); | |
785 | ||
786 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), | |
787 | oldval, rdtp->dynticks_nesting); | |
788 | ftrace_dump(DUMP_ORIG); | |
789 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", | |
790 | current->pid, current->comm, | |
791 | idle->pid, idle->comm); /* must be idle task! */ | |
792 | } | |
793 | } | |
794 | ||
795 | /* | |
796 | * Exit an RCU extended quiescent state, which can be either the | |
797 | * idle loop or adaptive-tickless usermode execution. | |
798 | */ | |
799 | static void rcu_eqs_exit(bool user) | |
800 | { | |
801 | struct rcu_dynticks *rdtp; | |
802 | long long oldval; | |
803 | ||
804 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
805 | oldval = rdtp->dynticks_nesting; | |
806 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0); | |
807 | if (oldval & DYNTICK_TASK_NEST_MASK) { | |
808 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; | |
809 | } else { | |
810 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; | |
811 | rcu_eqs_exit_common(oldval, user); | |
812 | } | |
813 | } | |
814 | ||
815 | /** | |
816 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
817 | * | |
818 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
819 | * read-side critical sections can occur. | |
820 | * | |
821 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
822 | * allow for the possibility of usermode upcalls messing up our count | |
823 | * of interrupt nesting level during the busy period that is just | |
824 | * now starting. | |
825 | */ | |
826 | void rcu_idle_exit(void) | |
827 | { | |
828 | unsigned long flags; | |
829 | ||
830 | local_irq_save(flags); | |
831 | rcu_eqs_exit(false); | |
832 | rcu_sysidle_exit(0); | |
833 | local_irq_restore(flags); | |
834 | } | |
835 | EXPORT_SYMBOL_GPL(rcu_idle_exit); | |
836 | ||
837 | #ifdef CONFIG_NO_HZ_FULL | |
838 | /** | |
839 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
840 | * | |
841 | * Exit RCU idle mode while entering the kernel because it can | |
842 | * run a RCU read side critical section anytime. | |
843 | */ | |
844 | void rcu_user_exit(void) | |
845 | { | |
846 | rcu_eqs_exit(1); | |
847 | } | |
848 | #endif /* CONFIG_NO_HZ_FULL */ | |
849 | ||
850 | /** | |
851 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
852 | * | |
853 | * Enter an interrupt handler, which might possibly result in exiting | |
854 | * idle mode, in other words, entering the mode in which read-side critical | |
855 | * sections can occur. | |
856 | * | |
857 | * Note that the Linux kernel is fully capable of entering an interrupt | |
858 | * handler that it never exits, for example when doing upcalls to | |
859 | * user mode! This code assumes that the idle loop never does upcalls to | |
860 | * user mode. If your architecture does do upcalls from the idle loop (or | |
861 | * does anything else that results in unbalanced calls to the irq_enter() | |
862 | * and irq_exit() functions), RCU will give you what you deserve, good | |
863 | * and hard. But very infrequently and irreproducibly. | |
864 | * | |
865 | * Use things like work queues to work around this limitation. | |
866 | * | |
867 | * You have been warned. | |
868 | */ | |
869 | void rcu_irq_enter(void) | |
870 | { | |
871 | unsigned long flags; | |
872 | struct rcu_dynticks *rdtp; | |
873 | long long oldval; | |
874 | ||
875 | local_irq_save(flags); | |
876 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
877 | oldval = rdtp->dynticks_nesting; | |
878 | rdtp->dynticks_nesting++; | |
879 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
880 | rdtp->dynticks_nesting == 0); | |
881 | if (oldval) | |
882 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); | |
883 | else | |
884 | rcu_eqs_exit_common(oldval, true); | |
885 | rcu_sysidle_exit(1); | |
886 | local_irq_restore(flags); | |
887 | } | |
888 | ||
889 | /** | |
890 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
891 | * | |
892 | * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and | |
893 | * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know | |
894 | * that the CPU is active. This implementation permits nested NMIs, as | |
895 | * long as the nesting level does not overflow an int. (You will probably | |
896 | * run out of stack space first.) | |
897 | */ | |
898 | void rcu_nmi_enter(void) | |
899 | { | |
900 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); | |
901 | int incby = 2; | |
902 | ||
903 | /* Complain about underflow. */ | |
904 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0); | |
905 | ||
906 | /* | |
907 | * If idle from RCU viewpoint, atomically increment ->dynticks | |
908 | * to mark non-idle and increment ->dynticks_nmi_nesting by one. | |
909 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means | |
910 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed | |
911 | * to be in the outermost NMI handler that interrupted an RCU-idle | |
912 | * period (observation due to Andy Lutomirski). | |
913 | */ | |
914 | if (!(atomic_read(&rdtp->dynticks) & 0x1)) { | |
915 | smp_mb__before_atomic(); /* Force delay from prior write. */ | |
916 | atomic_inc(&rdtp->dynticks); | |
917 | /* atomic_inc() before later RCU read-side crit sects */ | |
918 | smp_mb__after_atomic(); /* See above. */ | |
919 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
920 | incby = 1; | |
921 | } | |
922 | rdtp->dynticks_nmi_nesting += incby; | |
923 | barrier(); | |
924 | } | |
925 | ||
926 | /** | |
927 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
928 | * | |
929 | * If we are returning from the outermost NMI handler that interrupted an | |
930 | * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting | |
931 | * to let the RCU grace-period handling know that the CPU is back to | |
932 | * being RCU-idle. | |
933 | */ | |
934 | void rcu_nmi_exit(void) | |
935 | { | |
936 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); | |
937 | ||
938 | /* | |
939 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. | |
940 | * (We are exiting an NMI handler, so RCU better be paying attention | |
941 | * to us!) | |
942 | */ | |
943 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0); | |
944 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
945 | ||
946 | /* | |
947 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so | |
948 | * leave it in non-RCU-idle state. | |
949 | */ | |
950 | if (rdtp->dynticks_nmi_nesting != 1) { | |
951 | rdtp->dynticks_nmi_nesting -= 2; | |
952 | return; | |
953 | } | |
954 | ||
955 | /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ | |
956 | rdtp->dynticks_nmi_nesting = 0; | |
957 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ | |
958 | smp_mb__before_atomic(); /* See above. */ | |
959 | atomic_inc(&rdtp->dynticks); | |
960 | smp_mb__after_atomic(); /* Force delay to next write. */ | |
961 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
962 | } | |
963 | ||
964 | /** | |
965 | * __rcu_is_watching - are RCU read-side critical sections safe? | |
966 | * | |
967 | * Return true if RCU is watching the running CPU, which means that | |
968 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
969 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
970 | * least disabled preemption. | |
971 | */ | |
972 | bool notrace __rcu_is_watching(void) | |
973 | { | |
974 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
975 | } | |
976 | ||
977 | /** | |
978 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
979 | * | |
980 | * If the current CPU is in its idle loop and is neither in an interrupt | |
981 | * or NMI handler, return true. | |
982 | */ | |
983 | bool notrace rcu_is_watching(void) | |
984 | { | |
985 | bool ret; | |
986 | ||
987 | preempt_disable_notrace(); | |
988 | ret = __rcu_is_watching(); | |
989 | preempt_enable_notrace(); | |
990 | return ret; | |
991 | } | |
992 | EXPORT_SYMBOL_GPL(rcu_is_watching); | |
993 | ||
994 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) | |
995 | ||
996 | /* | |
997 | * Is the current CPU online? Disable preemption to avoid false positives | |
998 | * that could otherwise happen due to the current CPU number being sampled, | |
999 | * this task being preempted, its old CPU being taken offline, resuming | |
1000 | * on some other CPU, then determining that its old CPU is now offline. | |
1001 | * It is OK to use RCU on an offline processor during initial boot, hence | |
1002 | * the check for rcu_scheduler_fully_active. Note also that it is OK | |
1003 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
1004 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
1005 | * offline to continue to use RCU for one jiffy after marking itself | |
1006 | * offline in the cpu_online_mask. This leniency is necessary given the | |
1007 | * non-atomic nature of the online and offline processing, for example, | |
1008 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
1009 | * notifiers. | |
1010 | * | |
1011 | * This is also why RCU internally marks CPUs online during the | |
1012 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
1013 | * | |
1014 | * Disable checking if in an NMI handler because we cannot safely report | |
1015 | * errors from NMI handlers anyway. | |
1016 | */ | |
1017 | bool rcu_lockdep_current_cpu_online(void) | |
1018 | { | |
1019 | struct rcu_data *rdp; | |
1020 | struct rcu_node *rnp; | |
1021 | bool ret; | |
1022 | ||
1023 | if (in_nmi()) | |
1024 | return true; | |
1025 | preempt_disable(); | |
1026 | rdp = this_cpu_ptr(&rcu_sched_data); | |
1027 | rnp = rdp->mynode; | |
1028 | ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) || | |
1029 | !rcu_scheduler_fully_active; | |
1030 | preempt_enable(); | |
1031 | return ret; | |
1032 | } | |
1033 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
1034 | ||
1035 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ | |
1036 | ||
1037 | /** | |
1038 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle | |
1039 | * | |
1040 | * If the current CPU is idle or running at a first-level (not nested) | |
1041 | * interrupt from idle, return true. The caller must have at least | |
1042 | * disabled preemption. | |
1043 | */ | |
1044 | static int rcu_is_cpu_rrupt_from_idle(void) | |
1045 | { | |
1046 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; | |
1047 | } | |
1048 | ||
1049 | /* | |
1050 | * Snapshot the specified CPU's dynticks counter so that we can later | |
1051 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1052 | * is in dynticks idle mode, which is an extended quiescent state. | |
1053 | */ | |
1054 | static int dyntick_save_progress_counter(struct rcu_data *rdp, | |
1055 | bool *isidle, unsigned long *maxj) | |
1056 | { | |
1057 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); | |
1058 | rcu_sysidle_check_cpu(rdp, isidle, maxj); | |
1059 | if ((rdp->dynticks_snap & 0x1) == 0) { | |
1060 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); | |
1061 | return 1; | |
1062 | } else { | |
1063 | if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, | |
1064 | rdp->mynode->gpnum)) | |
1065 | WRITE_ONCE(rdp->gpwrap, true); | |
1066 | return 0; | |
1067 | } | |
1068 | } | |
1069 | ||
1070 | /* | |
1071 | * Return true if the specified CPU has passed through a quiescent | |
1072 | * state by virtue of being in or having passed through an dynticks | |
1073 | * idle state since the last call to dyntick_save_progress_counter() | |
1074 | * for this same CPU, or by virtue of having been offline. | |
1075 | */ | |
1076 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, | |
1077 | bool *isidle, unsigned long *maxj) | |
1078 | { | |
1079 | unsigned int curr; | |
1080 | int *rcrmp; | |
1081 | unsigned int snap; | |
1082 | ||
1083 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); | |
1084 | snap = (unsigned int)rdp->dynticks_snap; | |
1085 | ||
1086 | /* | |
1087 | * If the CPU passed through or entered a dynticks idle phase with | |
1088 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
1089 | * already acknowledged the request to pass through a quiescent | |
1090 | * state. Either way, that CPU cannot possibly be in an RCU | |
1091 | * read-side critical section that started before the beginning | |
1092 | * of the current RCU grace period. | |
1093 | */ | |
1094 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { | |
1095 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); | |
1096 | rdp->dynticks_fqs++; | |
1097 | return 1; | |
1098 | } | |
1099 | ||
1100 | /* | |
1101 | * Check for the CPU being offline, but only if the grace period | |
1102 | * is old enough. We don't need to worry about the CPU changing | |
1103 | * state: If we see it offline even once, it has been through a | |
1104 | * quiescent state. | |
1105 | * | |
1106 | * The reason for insisting that the grace period be at least | |
1107 | * one jiffy old is that CPUs that are not quite online and that | |
1108 | * have just gone offline can still execute RCU read-side critical | |
1109 | * sections. | |
1110 | */ | |
1111 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
1112 | return 0; /* Grace period is not old enough. */ | |
1113 | barrier(); | |
1114 | if (cpu_is_offline(rdp->cpu)) { | |
1115 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); | |
1116 | rdp->offline_fqs++; | |
1117 | return 1; | |
1118 | } | |
1119 | ||
1120 | /* | |
1121 | * A CPU running for an extended time within the kernel can | |
1122 | * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode, | |
1123 | * even context-switching back and forth between a pair of | |
1124 | * in-kernel CPU-bound tasks cannot advance grace periods. | |
1125 | * So if the grace period is old enough, make the CPU pay attention. | |
1126 | * Note that the unsynchronized assignments to the per-CPU | |
1127 | * rcu_sched_qs_mask variable are safe. Yes, setting of | |
1128 | * bits can be lost, but they will be set again on the next | |
1129 | * force-quiescent-state pass. So lost bit sets do not result | |
1130 | * in incorrect behavior, merely in a grace period lasting | |
1131 | * a few jiffies longer than it might otherwise. Because | |
1132 | * there are at most four threads involved, and because the | |
1133 | * updates are only once every few jiffies, the probability of | |
1134 | * lossage (and thus of slight grace-period extension) is | |
1135 | * quite low. | |
1136 | * | |
1137 | * Note that if the jiffies_till_sched_qs boot/sysfs parameter | |
1138 | * is set too high, we override with half of the RCU CPU stall | |
1139 | * warning delay. | |
1140 | */ | |
1141 | rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu); | |
1142 | if (ULONG_CMP_GE(jiffies, | |
1143 | rdp->rsp->gp_start + jiffies_till_sched_qs) || | |
1144 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { | |
1145 | if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) { | |
1146 | WRITE_ONCE(rdp->cond_resched_completed, | |
1147 | READ_ONCE(rdp->mynode->completed)); | |
1148 | smp_mb(); /* ->cond_resched_completed before *rcrmp. */ | |
1149 | WRITE_ONCE(*rcrmp, | |
1150 | READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask); | |
1151 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1152 | rdp->rsp->jiffies_resched += 5; /* Enable beating. */ | |
1153 | } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { | |
1154 | /* Time to beat on that CPU again! */ | |
1155 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1156 | rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */ | |
1157 | } | |
1158 | } | |
1159 | ||
1160 | return 0; | |
1161 | } | |
1162 | ||
1163 | static void record_gp_stall_check_time(struct rcu_state *rsp) | |
1164 | { | |
1165 | unsigned long j = jiffies; | |
1166 | unsigned long j1; | |
1167 | ||
1168 | rsp->gp_start = j; | |
1169 | smp_wmb(); /* Record start time before stall time. */ | |
1170 | j1 = rcu_jiffies_till_stall_check(); | |
1171 | WRITE_ONCE(rsp->jiffies_stall, j + j1); | |
1172 | rsp->jiffies_resched = j + j1 / 2; | |
1173 | rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs); | |
1174 | } | |
1175 | ||
1176 | /* | |
1177 | * Complain about starvation of grace-period kthread. | |
1178 | */ | |
1179 | static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp) | |
1180 | { | |
1181 | unsigned long gpa; | |
1182 | unsigned long j; | |
1183 | ||
1184 | j = jiffies; | |
1185 | gpa = READ_ONCE(rsp->gp_activity); | |
1186 | if (j - gpa > 2 * HZ) | |
1187 | pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x s%d ->state=%#lx\n", | |
1188 | rsp->name, j - gpa, | |
1189 | rsp->gpnum, rsp->completed, | |
1190 | rsp->gp_flags, rsp->gp_state, | |
1191 | rsp->gp_kthread ? rsp->gp_kthread->state : 0); | |
1192 | } | |
1193 | ||
1194 | /* | |
1195 | * Dump stacks of all tasks running on stalled CPUs. | |
1196 | */ | |
1197 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
1198 | { | |
1199 | int cpu; | |
1200 | unsigned long flags; | |
1201 | struct rcu_node *rnp; | |
1202 | ||
1203 | rcu_for_each_leaf_node(rsp, rnp) { | |
1204 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1205 | if (rnp->qsmask != 0) { | |
1206 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1207 | if (rnp->qsmask & (1UL << cpu)) | |
1208 | dump_cpu_task(rnp->grplo + cpu); | |
1209 | } | |
1210 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1211 | } | |
1212 | } | |
1213 | ||
1214 | static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum) | |
1215 | { | |
1216 | int cpu; | |
1217 | long delta; | |
1218 | unsigned long flags; | |
1219 | unsigned long gpa; | |
1220 | unsigned long j; | |
1221 | int ndetected = 0; | |
1222 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1223 | long totqlen = 0; | |
1224 | ||
1225 | /* Only let one CPU complain about others per time interval. */ | |
1226 | ||
1227 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1228 | delta = jiffies - READ_ONCE(rsp->jiffies_stall); | |
1229 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { | |
1230 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1231 | return; | |
1232 | } | |
1233 | WRITE_ONCE(rsp->jiffies_stall, | |
1234 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3); | |
1235 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1236 | ||
1237 | /* | |
1238 | * OK, time to rat on our buddy... | |
1239 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1240 | * RCU CPU stall warnings. | |
1241 | */ | |
1242 | pr_err("INFO: %s detected stalls on CPUs/tasks:", | |
1243 | rsp->name); | |
1244 | print_cpu_stall_info_begin(); | |
1245 | rcu_for_each_leaf_node(rsp, rnp) { | |
1246 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1247 | ndetected += rcu_print_task_stall(rnp); | |
1248 | if (rnp->qsmask != 0) { | |
1249 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1250 | if (rnp->qsmask & (1UL << cpu)) { | |
1251 | print_cpu_stall_info(rsp, | |
1252 | rnp->grplo + cpu); | |
1253 | ndetected++; | |
1254 | } | |
1255 | } | |
1256 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1257 | } | |
1258 | ||
1259 | print_cpu_stall_info_end(); | |
1260 | for_each_possible_cpu(cpu) | |
1261 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
1262 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", | |
1263 | smp_processor_id(), (long)(jiffies - rsp->gp_start), | |
1264 | (long)rsp->gpnum, (long)rsp->completed, totqlen); | |
1265 | if (ndetected) { | |
1266 | rcu_dump_cpu_stacks(rsp); | |
1267 | } else { | |
1268 | if (READ_ONCE(rsp->gpnum) != gpnum || | |
1269 | READ_ONCE(rsp->completed) == gpnum) { | |
1270 | pr_err("INFO: Stall ended before state dump start\n"); | |
1271 | } else { | |
1272 | j = jiffies; | |
1273 | gpa = READ_ONCE(rsp->gp_activity); | |
1274 | pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n", | |
1275 | rsp->name, j - gpa, j, gpa, | |
1276 | jiffies_till_next_fqs, | |
1277 | rcu_get_root(rsp)->qsmask); | |
1278 | /* In this case, the current CPU might be at fault. */ | |
1279 | sched_show_task(current); | |
1280 | } | |
1281 | } | |
1282 | ||
1283 | /* Complain about tasks blocking the grace period. */ | |
1284 | rcu_print_detail_task_stall(rsp); | |
1285 | ||
1286 | rcu_check_gp_kthread_starvation(rsp); | |
1287 | ||
1288 | force_quiescent_state(rsp); /* Kick them all. */ | |
1289 | } | |
1290 | ||
1291 | static void print_cpu_stall(struct rcu_state *rsp) | |
1292 | { | |
1293 | int cpu; | |
1294 | unsigned long flags; | |
1295 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1296 | long totqlen = 0; | |
1297 | ||
1298 | /* | |
1299 | * OK, time to rat on ourselves... | |
1300 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1301 | * RCU CPU stall warnings. | |
1302 | */ | |
1303 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); | |
1304 | print_cpu_stall_info_begin(); | |
1305 | print_cpu_stall_info(rsp, smp_processor_id()); | |
1306 | print_cpu_stall_info_end(); | |
1307 | for_each_possible_cpu(cpu) | |
1308 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
1309 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", | |
1310 | jiffies - rsp->gp_start, | |
1311 | (long)rsp->gpnum, (long)rsp->completed, totqlen); | |
1312 | ||
1313 | rcu_check_gp_kthread_starvation(rsp); | |
1314 | ||
1315 | rcu_dump_cpu_stacks(rsp); | |
1316 | ||
1317 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1318 | if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall))) | |
1319 | WRITE_ONCE(rsp->jiffies_stall, | |
1320 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3); | |
1321 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1322 | ||
1323 | /* | |
1324 | * Attempt to revive the RCU machinery by forcing a context switch. | |
1325 | * | |
1326 | * A context switch would normally allow the RCU state machine to make | |
1327 | * progress and it could be we're stuck in kernel space without context | |
1328 | * switches for an entirely unreasonable amount of time. | |
1329 | */ | |
1330 | resched_cpu(smp_processor_id()); | |
1331 | } | |
1332 | ||
1333 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
1334 | { | |
1335 | unsigned long completed; | |
1336 | unsigned long gpnum; | |
1337 | unsigned long gps; | |
1338 | unsigned long j; | |
1339 | unsigned long js; | |
1340 | struct rcu_node *rnp; | |
1341 | ||
1342 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) | |
1343 | return; | |
1344 | j = jiffies; | |
1345 | ||
1346 | /* | |
1347 | * Lots of memory barriers to reject false positives. | |
1348 | * | |
1349 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
1350 | * then rsp->gp_start, and finally rsp->completed. These values | |
1351 | * are updated in the opposite order with memory barriers (or | |
1352 | * equivalent) during grace-period initialization and cleanup. | |
1353 | * Now, a false positive can occur if we get an new value of | |
1354 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
1355 | * the memory barriers, the only way that this can happen is if one | |
1356 | * grace period ends and another starts between these two fetches. | |
1357 | * Detect this by comparing rsp->completed with the previous fetch | |
1358 | * from rsp->gpnum. | |
1359 | * | |
1360 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
1361 | * and rsp->gp_start suffice to forestall false positives. | |
1362 | */ | |
1363 | gpnum = READ_ONCE(rsp->gpnum); | |
1364 | smp_rmb(); /* Pick up ->gpnum first... */ | |
1365 | js = READ_ONCE(rsp->jiffies_stall); | |
1366 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ | |
1367 | gps = READ_ONCE(rsp->gp_start); | |
1368 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ | |
1369 | completed = READ_ONCE(rsp->completed); | |
1370 | if (ULONG_CMP_GE(completed, gpnum) || | |
1371 | ULONG_CMP_LT(j, js) || | |
1372 | ULONG_CMP_GE(gps, js)) | |
1373 | return; /* No stall or GP completed since entering function. */ | |
1374 | rnp = rdp->mynode; | |
1375 | if (rcu_gp_in_progress(rsp) && | |
1376 | (READ_ONCE(rnp->qsmask) & rdp->grpmask)) { | |
1377 | ||
1378 | /* We haven't checked in, so go dump stack. */ | |
1379 | print_cpu_stall(rsp); | |
1380 | ||
1381 | } else if (rcu_gp_in_progress(rsp) && | |
1382 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
1383 | ||
1384 | /* They had a few time units to dump stack, so complain. */ | |
1385 | print_other_cpu_stall(rsp, gpnum); | |
1386 | } | |
1387 | } | |
1388 | ||
1389 | /** | |
1390 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1391 | * | |
1392 | * Set the stall-warning timeout way off into the future, thus preventing | |
1393 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1394 | * RCU grace periods. | |
1395 | * | |
1396 | * The caller must disable hard irqs. | |
1397 | */ | |
1398 | void rcu_cpu_stall_reset(void) | |
1399 | { | |
1400 | struct rcu_state *rsp; | |
1401 | ||
1402 | for_each_rcu_flavor(rsp) | |
1403 | WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2); | |
1404 | } | |
1405 | ||
1406 | /* | |
1407 | * Initialize the specified rcu_data structure's default callback list | |
1408 | * to empty. The default callback list is the one that is not used by | |
1409 | * no-callbacks CPUs. | |
1410 | */ | |
1411 | static void init_default_callback_list(struct rcu_data *rdp) | |
1412 | { | |
1413 | int i; | |
1414 | ||
1415 | rdp->nxtlist = NULL; | |
1416 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1417 | rdp->nxttail[i] = &rdp->nxtlist; | |
1418 | } | |
1419 | ||
1420 | /* | |
1421 | * Initialize the specified rcu_data structure's callback list to empty. | |
1422 | */ | |
1423 | static void init_callback_list(struct rcu_data *rdp) | |
1424 | { | |
1425 | if (init_nocb_callback_list(rdp)) | |
1426 | return; | |
1427 | init_default_callback_list(rdp); | |
1428 | } | |
1429 | ||
1430 | /* | |
1431 | * Determine the value that ->completed will have at the end of the | |
1432 | * next subsequent grace period. This is used to tag callbacks so that | |
1433 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1434 | * been dyntick-idle for an extended period with callbacks under the | |
1435 | * influence of RCU_FAST_NO_HZ. | |
1436 | * | |
1437 | * The caller must hold rnp->lock with interrupts disabled. | |
1438 | */ | |
1439 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1440 | struct rcu_node *rnp) | |
1441 | { | |
1442 | /* | |
1443 | * If RCU is idle, we just wait for the next grace period. | |
1444 | * But we can only be sure that RCU is idle if we are looking | |
1445 | * at the root rcu_node structure -- otherwise, a new grace | |
1446 | * period might have started, but just not yet gotten around | |
1447 | * to initializing the current non-root rcu_node structure. | |
1448 | */ | |
1449 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1450 | return rnp->completed + 1; | |
1451 | ||
1452 | /* | |
1453 | * Otherwise, wait for a possible partial grace period and | |
1454 | * then the subsequent full grace period. | |
1455 | */ | |
1456 | return rnp->completed + 2; | |
1457 | } | |
1458 | ||
1459 | /* | |
1460 | * Trace-event helper function for rcu_start_future_gp() and | |
1461 | * rcu_nocb_wait_gp(). | |
1462 | */ | |
1463 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
1464 | unsigned long c, const char *s) | |
1465 | { | |
1466 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1467 | rnp->completed, c, rnp->level, | |
1468 | rnp->grplo, rnp->grphi, s); | |
1469 | } | |
1470 | ||
1471 | /* | |
1472 | * Start some future grace period, as needed to handle newly arrived | |
1473 | * callbacks. The required future grace periods are recorded in each | |
1474 | * rcu_node structure's ->need_future_gp field. Returns true if there | |
1475 | * is reason to awaken the grace-period kthread. | |
1476 | * | |
1477 | * The caller must hold the specified rcu_node structure's ->lock. | |
1478 | */ | |
1479 | static bool __maybe_unused | |
1480 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
1481 | unsigned long *c_out) | |
1482 | { | |
1483 | unsigned long c; | |
1484 | int i; | |
1485 | bool ret = false; | |
1486 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); | |
1487 | ||
1488 | /* | |
1489 | * Pick up grace-period number for new callbacks. If this | |
1490 | * grace period is already marked as needed, return to the caller. | |
1491 | */ | |
1492 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
1493 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); | |
1494 | if (rnp->need_future_gp[c & 0x1]) { | |
1495 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); | |
1496 | goto out; | |
1497 | } | |
1498 | ||
1499 | /* | |
1500 | * If either this rcu_node structure or the root rcu_node structure | |
1501 | * believe that a grace period is in progress, then we must wait | |
1502 | * for the one following, which is in "c". Because our request | |
1503 | * will be noticed at the end of the current grace period, we don't | |
1504 | * need to explicitly start one. We only do the lockless check | |
1505 | * of rnp_root's fields if the current rcu_node structure thinks | |
1506 | * there is no grace period in flight, and because we hold rnp->lock, | |
1507 | * the only possible change is when rnp_root's two fields are | |
1508 | * equal, in which case rnp_root->gpnum might be concurrently | |
1509 | * incremented. But that is OK, as it will just result in our | |
1510 | * doing some extra useless work. | |
1511 | */ | |
1512 | if (rnp->gpnum != rnp->completed || | |
1513 | READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) { | |
1514 | rnp->need_future_gp[c & 0x1]++; | |
1515 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); | |
1516 | goto out; | |
1517 | } | |
1518 | ||
1519 | /* | |
1520 | * There might be no grace period in progress. If we don't already | |
1521 | * hold it, acquire the root rcu_node structure's lock in order to | |
1522 | * start one (if needed). | |
1523 | */ | |
1524 | if (rnp != rnp_root) { | |
1525 | raw_spin_lock(&rnp_root->lock); | |
1526 | smp_mb__after_unlock_lock(); | |
1527 | } | |
1528 | ||
1529 | /* | |
1530 | * Get a new grace-period number. If there really is no grace | |
1531 | * period in progress, it will be smaller than the one we obtained | |
1532 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1533 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1534 | */ | |
1535 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1536 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1537 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1538 | rdp->nxtcompleted[i] = c; | |
1539 | ||
1540 | /* | |
1541 | * If the needed for the required grace period is already | |
1542 | * recorded, trace and leave. | |
1543 | */ | |
1544 | if (rnp_root->need_future_gp[c & 0x1]) { | |
1545 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); | |
1546 | goto unlock_out; | |
1547 | } | |
1548 | ||
1549 | /* Record the need for the future grace period. */ | |
1550 | rnp_root->need_future_gp[c & 0x1]++; | |
1551 | ||
1552 | /* If a grace period is not already in progress, start one. */ | |
1553 | if (rnp_root->gpnum != rnp_root->completed) { | |
1554 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); | |
1555 | } else { | |
1556 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); | |
1557 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); | |
1558 | } | |
1559 | unlock_out: | |
1560 | if (rnp != rnp_root) | |
1561 | raw_spin_unlock(&rnp_root->lock); | |
1562 | out: | |
1563 | if (c_out != NULL) | |
1564 | *c_out = c; | |
1565 | return ret; | |
1566 | } | |
1567 | ||
1568 | /* | |
1569 | * Clean up any old requests for the just-ended grace period. Also return | |
1570 | * whether any additional grace periods have been requested. Also invoke | |
1571 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1572 | * waiting for this grace period to complete. | |
1573 | */ | |
1574 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1575 | { | |
1576 | int c = rnp->completed; | |
1577 | int needmore; | |
1578 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1579 | ||
1580 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1581 | rnp->need_future_gp[c & 0x1] = 0; | |
1582 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
1583 | trace_rcu_future_gp(rnp, rdp, c, | |
1584 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
1585 | return needmore; | |
1586 | } | |
1587 | ||
1588 | /* | |
1589 | * Awaken the grace-period kthread for the specified flavor of RCU. | |
1590 | * Don't do a self-awaken, and don't bother awakening when there is | |
1591 | * nothing for the grace-period kthread to do (as in several CPUs | |
1592 | * raced to awaken, and we lost), and finally don't try to awaken | |
1593 | * a kthread that has not yet been created. | |
1594 | */ | |
1595 | static void rcu_gp_kthread_wake(struct rcu_state *rsp) | |
1596 | { | |
1597 | if (current == rsp->gp_kthread || | |
1598 | !READ_ONCE(rsp->gp_flags) || | |
1599 | !rsp->gp_kthread) | |
1600 | return; | |
1601 | wake_up(&rsp->gp_wq); | |
1602 | } | |
1603 | ||
1604 | /* | |
1605 | * If there is room, assign a ->completed number to any callbacks on | |
1606 | * this CPU that have not already been assigned. Also accelerate any | |
1607 | * callbacks that were previously assigned a ->completed number that has | |
1608 | * since proven to be too conservative, which can happen if callbacks get | |
1609 | * assigned a ->completed number while RCU is idle, but with reference to | |
1610 | * a non-root rcu_node structure. This function is idempotent, so it does | |
1611 | * not hurt to call it repeatedly. Returns an flag saying that we should | |
1612 | * awaken the RCU grace-period kthread. | |
1613 | * | |
1614 | * The caller must hold rnp->lock with interrupts disabled. | |
1615 | */ | |
1616 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1617 | struct rcu_data *rdp) | |
1618 | { | |
1619 | unsigned long c; | |
1620 | int i; | |
1621 | bool ret; | |
1622 | ||
1623 | /* If the CPU has no callbacks, nothing to do. */ | |
1624 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1625 | return false; | |
1626 | ||
1627 | /* | |
1628 | * Starting from the sublist containing the callbacks most | |
1629 | * recently assigned a ->completed number and working down, find the | |
1630 | * first sublist that is not assignable to an upcoming grace period. | |
1631 | * Such a sublist has something in it (first two tests) and has | |
1632 | * a ->completed number assigned that will complete sooner than | |
1633 | * the ->completed number for newly arrived callbacks (last test). | |
1634 | * | |
1635 | * The key point is that any later sublist can be assigned the | |
1636 | * same ->completed number as the newly arrived callbacks, which | |
1637 | * means that the callbacks in any of these later sublist can be | |
1638 | * grouped into a single sublist, whether or not they have already | |
1639 | * been assigned a ->completed number. | |
1640 | */ | |
1641 | c = rcu_cbs_completed(rsp, rnp); | |
1642 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1643 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1644 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1645 | break; | |
1646 | ||
1647 | /* | |
1648 | * If there are no sublist for unassigned callbacks, leave. | |
1649 | * At the same time, advance "i" one sublist, so that "i" will | |
1650 | * index into the sublist where all the remaining callbacks should | |
1651 | * be grouped into. | |
1652 | */ | |
1653 | if (++i >= RCU_NEXT_TAIL) | |
1654 | return false; | |
1655 | ||
1656 | /* | |
1657 | * Assign all subsequent callbacks' ->completed number to the next | |
1658 | * full grace period and group them all in the sublist initially | |
1659 | * indexed by "i". | |
1660 | */ | |
1661 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1662 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1663 | rdp->nxtcompleted[i] = c; | |
1664 | } | |
1665 | /* Record any needed additional grace periods. */ | |
1666 | ret = rcu_start_future_gp(rnp, rdp, NULL); | |
1667 | ||
1668 | /* Trace depending on how much we were able to accelerate. */ | |
1669 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
1670 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); | |
1671 | else | |
1672 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); | |
1673 | return ret; | |
1674 | } | |
1675 | ||
1676 | /* | |
1677 | * Move any callbacks whose grace period has completed to the | |
1678 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1679 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1680 | * sublist. This function is idempotent, so it does not hurt to | |
1681 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
1682 | * Returns true if the RCU grace-period kthread needs to be awakened. | |
1683 | * | |
1684 | * The caller must hold rnp->lock with interrupts disabled. | |
1685 | */ | |
1686 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1687 | struct rcu_data *rdp) | |
1688 | { | |
1689 | int i, j; | |
1690 | ||
1691 | /* If the CPU has no callbacks, nothing to do. */ | |
1692 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1693 | return false; | |
1694 | ||
1695 | /* | |
1696 | * Find all callbacks whose ->completed numbers indicate that they | |
1697 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1698 | */ | |
1699 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1700 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1701 | break; | |
1702 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1703 | } | |
1704 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1705 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1706 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1707 | ||
1708 | /* Copy down callbacks to fill in empty sublists. */ | |
1709 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1710 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1711 | break; | |
1712 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1713 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1714 | } | |
1715 | ||
1716 | /* Classify any remaining callbacks. */ | |
1717 | return rcu_accelerate_cbs(rsp, rnp, rdp); | |
1718 | } | |
1719 | ||
1720 | /* | |
1721 | * Update CPU-local rcu_data state to record the beginnings and ends of | |
1722 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1723 | * structure corresponding to the current CPU, and must have irqs disabled. | |
1724 | * Returns true if the grace-period kthread needs to be awakened. | |
1725 | */ | |
1726 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, | |
1727 | struct rcu_data *rdp) | |
1728 | { | |
1729 | bool ret; | |
1730 | ||
1731 | /* Handle the ends of any preceding grace periods first. */ | |
1732 | if (rdp->completed == rnp->completed && | |
1733 | !unlikely(READ_ONCE(rdp->gpwrap))) { | |
1734 | ||
1735 | /* No grace period end, so just accelerate recent callbacks. */ | |
1736 | ret = rcu_accelerate_cbs(rsp, rnp, rdp); | |
1737 | ||
1738 | } else { | |
1739 | ||
1740 | /* Advance callbacks. */ | |
1741 | ret = rcu_advance_cbs(rsp, rnp, rdp); | |
1742 | ||
1743 | /* Remember that we saw this grace-period completion. */ | |
1744 | rdp->completed = rnp->completed; | |
1745 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); | |
1746 | } | |
1747 | ||
1748 | if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) { | |
1749 | /* | |
1750 | * If the current grace period is waiting for this CPU, | |
1751 | * set up to detect a quiescent state, otherwise don't | |
1752 | * go looking for one. | |
1753 | */ | |
1754 | rdp->gpnum = rnp->gpnum; | |
1755 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); | |
1756 | rdp->passed_quiesce = 0; | |
1757 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); | |
1758 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); | |
1759 | zero_cpu_stall_ticks(rdp); | |
1760 | WRITE_ONCE(rdp->gpwrap, false); | |
1761 | } | |
1762 | return ret; | |
1763 | } | |
1764 | ||
1765 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) | |
1766 | { | |
1767 | unsigned long flags; | |
1768 | bool needwake; | |
1769 | struct rcu_node *rnp; | |
1770 | ||
1771 | local_irq_save(flags); | |
1772 | rnp = rdp->mynode; | |
1773 | if ((rdp->gpnum == READ_ONCE(rnp->gpnum) && | |
1774 | rdp->completed == READ_ONCE(rnp->completed) && | |
1775 | !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */ | |
1776 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ | |
1777 | local_irq_restore(flags); | |
1778 | return; | |
1779 | } | |
1780 | smp_mb__after_unlock_lock(); | |
1781 | needwake = __note_gp_changes(rsp, rnp, rdp); | |
1782 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1783 | if (needwake) | |
1784 | rcu_gp_kthread_wake(rsp); | |
1785 | } | |
1786 | ||
1787 | static void rcu_gp_slow(struct rcu_state *rsp, int delay) | |
1788 | { | |
1789 | if (delay > 0 && | |
1790 | !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay))) | |
1791 | schedule_timeout_uninterruptible(delay); | |
1792 | } | |
1793 | ||
1794 | /* | |
1795 | * Initialize a new grace period. Return 0 if no grace period required. | |
1796 | */ | |
1797 | static int rcu_gp_init(struct rcu_state *rsp) | |
1798 | { | |
1799 | unsigned long oldmask; | |
1800 | struct rcu_data *rdp; | |
1801 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1802 | ||
1803 | WRITE_ONCE(rsp->gp_activity, jiffies); | |
1804 | raw_spin_lock_irq(&rnp->lock); | |
1805 | smp_mb__after_unlock_lock(); | |
1806 | if (!READ_ONCE(rsp->gp_flags)) { | |
1807 | /* Spurious wakeup, tell caller to go back to sleep. */ | |
1808 | raw_spin_unlock_irq(&rnp->lock); | |
1809 | return 0; | |
1810 | } | |
1811 | WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */ | |
1812 | ||
1813 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { | |
1814 | /* | |
1815 | * Grace period already in progress, don't start another. | |
1816 | * Not supposed to be able to happen. | |
1817 | */ | |
1818 | raw_spin_unlock_irq(&rnp->lock); | |
1819 | return 0; | |
1820 | } | |
1821 | ||
1822 | /* Advance to a new grace period and initialize state. */ | |
1823 | record_gp_stall_check_time(rsp); | |
1824 | /* Record GP times before starting GP, hence smp_store_release(). */ | |
1825 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); | |
1826 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); | |
1827 | raw_spin_unlock_irq(&rnp->lock); | |
1828 | ||
1829 | /* | |
1830 | * Apply per-leaf buffered online and offline operations to the | |
1831 | * rcu_node tree. Note that this new grace period need not wait | |
1832 | * for subsequent online CPUs, and that quiescent-state forcing | |
1833 | * will handle subsequent offline CPUs. | |
1834 | */ | |
1835 | rcu_for_each_leaf_node(rsp, rnp) { | |
1836 | rcu_gp_slow(rsp, gp_preinit_delay); | |
1837 | raw_spin_lock_irq(&rnp->lock); | |
1838 | smp_mb__after_unlock_lock(); | |
1839 | if (rnp->qsmaskinit == rnp->qsmaskinitnext && | |
1840 | !rnp->wait_blkd_tasks) { | |
1841 | /* Nothing to do on this leaf rcu_node structure. */ | |
1842 | raw_spin_unlock_irq(&rnp->lock); | |
1843 | continue; | |
1844 | } | |
1845 | ||
1846 | /* Record old state, apply changes to ->qsmaskinit field. */ | |
1847 | oldmask = rnp->qsmaskinit; | |
1848 | rnp->qsmaskinit = rnp->qsmaskinitnext; | |
1849 | ||
1850 | /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ | |
1851 | if (!oldmask != !rnp->qsmaskinit) { | |
1852 | if (!oldmask) /* First online CPU for this rcu_node. */ | |
1853 | rcu_init_new_rnp(rnp); | |
1854 | else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */ | |
1855 | rnp->wait_blkd_tasks = true; | |
1856 | else /* Last offline CPU and can propagate. */ | |
1857 | rcu_cleanup_dead_rnp(rnp); | |
1858 | } | |
1859 | ||
1860 | /* | |
1861 | * If all waited-on tasks from prior grace period are | |
1862 | * done, and if all this rcu_node structure's CPUs are | |
1863 | * still offline, propagate up the rcu_node tree and | |
1864 | * clear ->wait_blkd_tasks. Otherwise, if one of this | |
1865 | * rcu_node structure's CPUs has since come back online, | |
1866 | * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp() | |
1867 | * checks for this, so just call it unconditionally). | |
1868 | */ | |
1869 | if (rnp->wait_blkd_tasks && | |
1870 | (!rcu_preempt_has_tasks(rnp) || | |
1871 | rnp->qsmaskinit)) { | |
1872 | rnp->wait_blkd_tasks = false; | |
1873 | rcu_cleanup_dead_rnp(rnp); | |
1874 | } | |
1875 | ||
1876 | raw_spin_unlock_irq(&rnp->lock); | |
1877 | } | |
1878 | ||
1879 | /* | |
1880 | * Set the quiescent-state-needed bits in all the rcu_node | |
1881 | * structures for all currently online CPUs in breadth-first order, | |
1882 | * starting from the root rcu_node structure, relying on the layout | |
1883 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1884 | * will access only the leaves of the hierarchy, thus seeing that no | |
1885 | * grace period is in progress, at least until the corresponding | |
1886 | * leaf node has been initialized. In addition, we have excluded | |
1887 | * CPU-hotplug operations. | |
1888 | * | |
1889 | * The grace period cannot complete until the initialization | |
1890 | * process finishes, because this kthread handles both. | |
1891 | */ | |
1892 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
1893 | rcu_gp_slow(rsp, gp_init_delay); | |
1894 | raw_spin_lock_irq(&rnp->lock); | |
1895 | smp_mb__after_unlock_lock(); | |
1896 | rdp = this_cpu_ptr(rsp->rda); | |
1897 | rcu_preempt_check_blocked_tasks(rnp); | |
1898 | rnp->qsmask = rnp->qsmaskinit; | |
1899 | WRITE_ONCE(rnp->gpnum, rsp->gpnum); | |
1900 | if (WARN_ON_ONCE(rnp->completed != rsp->completed)) | |
1901 | WRITE_ONCE(rnp->completed, rsp->completed); | |
1902 | if (rnp == rdp->mynode) | |
1903 | (void)__note_gp_changes(rsp, rnp, rdp); | |
1904 | rcu_preempt_boost_start_gp(rnp); | |
1905 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1906 | rnp->level, rnp->grplo, | |
1907 | rnp->grphi, rnp->qsmask); | |
1908 | raw_spin_unlock_irq(&rnp->lock); | |
1909 | cond_resched_rcu_qs(); | |
1910 | WRITE_ONCE(rsp->gp_activity, jiffies); | |
1911 | } | |
1912 | ||
1913 | return 1; | |
1914 | } | |
1915 | ||
1916 | /* | |
1917 | * Helper function for wait_event_interruptible_timeout() wakeup | |
1918 | * at force-quiescent-state time. | |
1919 | */ | |
1920 | static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp) | |
1921 | { | |
1922 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1923 | ||
1924 | /* Someone like call_rcu() requested a force-quiescent-state scan. */ | |
1925 | *gfp = READ_ONCE(rsp->gp_flags); | |
1926 | if (*gfp & RCU_GP_FLAG_FQS) | |
1927 | return true; | |
1928 | ||
1929 | /* The current grace period has completed. */ | |
1930 | if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp)) | |
1931 | return true; | |
1932 | ||
1933 | return false; | |
1934 | } | |
1935 | ||
1936 | /* | |
1937 | * Do one round of quiescent-state forcing. | |
1938 | */ | |
1939 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) | |
1940 | { | |
1941 | int fqs_state = fqs_state_in; | |
1942 | bool isidle = false; | |
1943 | unsigned long maxj; | |
1944 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1945 | ||
1946 | WRITE_ONCE(rsp->gp_activity, jiffies); | |
1947 | rsp->n_force_qs++; | |
1948 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1949 | /* Collect dyntick-idle snapshots. */ | |
1950 | if (is_sysidle_rcu_state(rsp)) { | |
1951 | isidle = true; | |
1952 | maxj = jiffies - ULONG_MAX / 4; | |
1953 | } | |
1954 | force_qs_rnp(rsp, dyntick_save_progress_counter, | |
1955 | &isidle, &maxj); | |
1956 | rcu_sysidle_report_gp(rsp, isidle, maxj); | |
1957 | fqs_state = RCU_FORCE_QS; | |
1958 | } else { | |
1959 | /* Handle dyntick-idle and offline CPUs. */ | |
1960 | isidle = true; | |
1961 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); | |
1962 | } | |
1963 | /* Clear flag to prevent immediate re-entry. */ | |
1964 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1965 | raw_spin_lock_irq(&rnp->lock); | |
1966 | smp_mb__after_unlock_lock(); | |
1967 | WRITE_ONCE(rsp->gp_flags, | |
1968 | READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS); | |
1969 | raw_spin_unlock_irq(&rnp->lock); | |
1970 | } | |
1971 | return fqs_state; | |
1972 | } | |
1973 | ||
1974 | /* | |
1975 | * Clean up after the old grace period. | |
1976 | */ | |
1977 | static void rcu_gp_cleanup(struct rcu_state *rsp) | |
1978 | { | |
1979 | unsigned long gp_duration; | |
1980 | bool needgp = false; | |
1981 | int nocb = 0; | |
1982 | struct rcu_data *rdp; | |
1983 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1984 | ||
1985 | WRITE_ONCE(rsp->gp_activity, jiffies); | |
1986 | raw_spin_lock_irq(&rnp->lock); | |
1987 | smp_mb__after_unlock_lock(); | |
1988 | gp_duration = jiffies - rsp->gp_start; | |
1989 | if (gp_duration > rsp->gp_max) | |
1990 | rsp->gp_max = gp_duration; | |
1991 | ||
1992 | /* | |
1993 | * We know the grace period is complete, but to everyone else | |
1994 | * it appears to still be ongoing. But it is also the case | |
1995 | * that to everyone else it looks like there is nothing that | |
1996 | * they can do to advance the grace period. It is therefore | |
1997 | * safe for us to drop the lock in order to mark the grace | |
1998 | * period as completed in all of the rcu_node structures. | |
1999 | */ | |
2000 | raw_spin_unlock_irq(&rnp->lock); | |
2001 | ||
2002 | /* | |
2003 | * Propagate new ->completed value to rcu_node structures so | |
2004 | * that other CPUs don't have to wait until the start of the next | |
2005 | * grace period to process their callbacks. This also avoids | |
2006 | * some nasty RCU grace-period initialization races by forcing | |
2007 | * the end of the current grace period to be completely recorded in | |
2008 | * all of the rcu_node structures before the beginning of the next | |
2009 | * grace period is recorded in any of the rcu_node structures. | |
2010 | */ | |
2011 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
2012 | raw_spin_lock_irq(&rnp->lock); | |
2013 | smp_mb__after_unlock_lock(); | |
2014 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); | |
2015 | WARN_ON_ONCE(rnp->qsmask); | |
2016 | WRITE_ONCE(rnp->completed, rsp->gpnum); | |
2017 | rdp = this_cpu_ptr(rsp->rda); | |
2018 | if (rnp == rdp->mynode) | |
2019 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; | |
2020 | /* smp_mb() provided by prior unlock-lock pair. */ | |
2021 | nocb += rcu_future_gp_cleanup(rsp, rnp); | |
2022 | raw_spin_unlock_irq(&rnp->lock); | |
2023 | cond_resched_rcu_qs(); | |
2024 | WRITE_ONCE(rsp->gp_activity, jiffies); | |
2025 | rcu_gp_slow(rsp, gp_cleanup_delay); | |
2026 | } | |
2027 | rnp = rcu_get_root(rsp); | |
2028 | raw_spin_lock_irq(&rnp->lock); | |
2029 | smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */ | |
2030 | rcu_nocb_gp_set(rnp, nocb); | |
2031 | ||
2032 | /* Declare grace period done. */ | |
2033 | WRITE_ONCE(rsp->completed, rsp->gpnum); | |
2034 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); | |
2035 | rsp->fqs_state = RCU_GP_IDLE; | |
2036 | rdp = this_cpu_ptr(rsp->rda); | |
2037 | /* Advance CBs to reduce false positives below. */ | |
2038 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; | |
2039 | if (needgp || cpu_needs_another_gp(rsp, rdp)) { | |
2040 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); | |
2041 | trace_rcu_grace_period(rsp->name, | |
2042 | READ_ONCE(rsp->gpnum), | |
2043 | TPS("newreq")); | |
2044 | } | |
2045 | raw_spin_unlock_irq(&rnp->lock); | |
2046 | } | |
2047 | ||
2048 | /* | |
2049 | * Body of kthread that handles grace periods. | |
2050 | */ | |
2051 | static int __noreturn rcu_gp_kthread(void *arg) | |
2052 | { | |
2053 | int fqs_state; | |
2054 | int gf; | |
2055 | unsigned long j; | |
2056 | int ret; | |
2057 | struct rcu_state *rsp = arg; | |
2058 | struct rcu_node *rnp = rcu_get_root(rsp); | |
2059 | ||
2060 | rcu_bind_gp_kthread(); | |
2061 | for (;;) { | |
2062 | ||
2063 | /* Handle grace-period start. */ | |
2064 | for (;;) { | |
2065 | trace_rcu_grace_period(rsp->name, | |
2066 | READ_ONCE(rsp->gpnum), | |
2067 | TPS("reqwait")); | |
2068 | rsp->gp_state = RCU_GP_WAIT_GPS; | |
2069 | wait_event_interruptible(rsp->gp_wq, | |
2070 | READ_ONCE(rsp->gp_flags) & | |
2071 | RCU_GP_FLAG_INIT); | |
2072 | rsp->gp_state = RCU_GP_DONE_GPS; | |
2073 | /* Locking provides needed memory barrier. */ | |
2074 | if (rcu_gp_init(rsp)) | |
2075 | break; | |
2076 | cond_resched_rcu_qs(); | |
2077 | WRITE_ONCE(rsp->gp_activity, jiffies); | |
2078 | WARN_ON(signal_pending(current)); | |
2079 | trace_rcu_grace_period(rsp->name, | |
2080 | READ_ONCE(rsp->gpnum), | |
2081 | TPS("reqwaitsig")); | |
2082 | } | |
2083 | ||
2084 | /* Handle quiescent-state forcing. */ | |
2085 | fqs_state = RCU_SAVE_DYNTICK; | |
2086 | j = jiffies_till_first_fqs; | |
2087 | if (j > HZ) { | |
2088 | j = HZ; | |
2089 | jiffies_till_first_fqs = HZ; | |
2090 | } | |
2091 | ret = 0; | |
2092 | for (;;) { | |
2093 | if (!ret) | |
2094 | rsp->jiffies_force_qs = jiffies + j; | |
2095 | trace_rcu_grace_period(rsp->name, | |
2096 | READ_ONCE(rsp->gpnum), | |
2097 | TPS("fqswait")); | |
2098 | rsp->gp_state = RCU_GP_WAIT_FQS; | |
2099 | ret = wait_event_interruptible_timeout(rsp->gp_wq, | |
2100 | rcu_gp_fqs_check_wake(rsp, &gf), j); | |
2101 | rsp->gp_state = RCU_GP_DOING_FQS; | |
2102 | /* Locking provides needed memory barriers. */ | |
2103 | /* If grace period done, leave loop. */ | |
2104 | if (!READ_ONCE(rnp->qsmask) && | |
2105 | !rcu_preempt_blocked_readers_cgp(rnp)) | |
2106 | break; | |
2107 | /* If time for quiescent-state forcing, do it. */ | |
2108 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || | |
2109 | (gf & RCU_GP_FLAG_FQS)) { | |
2110 | trace_rcu_grace_period(rsp->name, | |
2111 | READ_ONCE(rsp->gpnum), | |
2112 | TPS("fqsstart")); | |
2113 | fqs_state = rcu_gp_fqs(rsp, fqs_state); | |
2114 | trace_rcu_grace_period(rsp->name, | |
2115 | READ_ONCE(rsp->gpnum), | |
2116 | TPS("fqsend")); | |
2117 | cond_resched_rcu_qs(); | |
2118 | WRITE_ONCE(rsp->gp_activity, jiffies); | |
2119 | } else { | |
2120 | /* Deal with stray signal. */ | |
2121 | cond_resched_rcu_qs(); | |
2122 | WRITE_ONCE(rsp->gp_activity, jiffies); | |
2123 | WARN_ON(signal_pending(current)); | |
2124 | trace_rcu_grace_period(rsp->name, | |
2125 | READ_ONCE(rsp->gpnum), | |
2126 | TPS("fqswaitsig")); | |
2127 | } | |
2128 | j = jiffies_till_next_fqs; | |
2129 | if (j > HZ) { | |
2130 | j = HZ; | |
2131 | jiffies_till_next_fqs = HZ; | |
2132 | } else if (j < 1) { | |
2133 | j = 1; | |
2134 | jiffies_till_next_fqs = 1; | |
2135 | } | |
2136 | } | |
2137 | ||
2138 | /* Handle grace-period end. */ | |
2139 | rsp->gp_state = RCU_GP_CLEANUP; | |
2140 | rcu_gp_cleanup(rsp); | |
2141 | rsp->gp_state = RCU_GP_CLEANED; | |
2142 | } | |
2143 | } | |
2144 | ||
2145 | /* | |
2146 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
2147 | * in preparation for detecting the next grace period. The caller must hold | |
2148 | * the root node's ->lock and hard irqs must be disabled. | |
2149 | * | |
2150 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
2151 | * invoke this function. This can happen when the dying CPU reports its | |
2152 | * quiescent state. | |
2153 | * | |
2154 | * Returns true if the grace-period kthread must be awakened. | |
2155 | */ | |
2156 | static bool | |
2157 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, | |
2158 | struct rcu_data *rdp) | |
2159 | { | |
2160 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { | |
2161 | /* | |
2162 | * Either we have not yet spawned the grace-period | |
2163 | * task, this CPU does not need another grace period, | |
2164 | * or a grace period is already in progress. | |
2165 | * Either way, don't start a new grace period. | |
2166 | */ | |
2167 | return false; | |
2168 | } | |
2169 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); | |
2170 | trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum), | |
2171 | TPS("newreq")); | |
2172 | ||
2173 | /* | |
2174 | * We can't do wakeups while holding the rnp->lock, as that | |
2175 | * could cause possible deadlocks with the rq->lock. Defer | |
2176 | * the wakeup to our caller. | |
2177 | */ | |
2178 | return true; | |
2179 | } | |
2180 | ||
2181 | /* | |
2182 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
2183 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
2184 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
2185 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
2186 | * that is encountered beforehand. | |
2187 | * | |
2188 | * Returns true if the grace-period kthread needs to be awakened. | |
2189 | */ | |
2190 | static bool rcu_start_gp(struct rcu_state *rsp) | |
2191 | { | |
2192 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
2193 | struct rcu_node *rnp = rcu_get_root(rsp); | |
2194 | bool ret = false; | |
2195 | ||
2196 | /* | |
2197 | * If there is no grace period in progress right now, any | |
2198 | * callbacks we have up to this point will be satisfied by the | |
2199 | * next grace period. Also, advancing the callbacks reduces the | |
2200 | * probability of false positives from cpu_needs_another_gp() | |
2201 | * resulting in pointless grace periods. So, advance callbacks | |
2202 | * then start the grace period! | |
2203 | */ | |
2204 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; | |
2205 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; | |
2206 | return ret; | |
2207 | } | |
2208 | ||
2209 | /* | |
2210 | * Report a full set of quiescent states to the specified rcu_state | |
2211 | * data structure. This involves cleaning up after the prior grace | |
2212 | * period and letting rcu_start_gp() start up the next grace period | |
2213 | * if one is needed. Note that the caller must hold rnp->lock, which | |
2214 | * is released before return. | |
2215 | */ | |
2216 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) | |
2217 | __releases(rcu_get_root(rsp)->lock) | |
2218 | { | |
2219 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); | |
2220 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); | |
2221 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); | |
2222 | rcu_gp_kthread_wake(rsp); | |
2223 | } | |
2224 | ||
2225 | /* | |
2226 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. | |
2227 | * Allows quiescent states for a group of CPUs to be reported at one go | |
2228 | * to the specified rcu_node structure, though all the CPUs in the group | |
2229 | * must be represented by the same rcu_node structure (which need not be a | |
2230 | * leaf rcu_node structure, though it often will be). The gps parameter | |
2231 | * is the grace-period snapshot, which means that the quiescent states | |
2232 | * are valid only if rnp->gpnum is equal to gps. That structure's lock | |
2233 | * must be held upon entry, and it is released before return. | |
2234 | */ | |
2235 | static void | |
2236 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, | |
2237 | struct rcu_node *rnp, unsigned long gps, unsigned long flags) | |
2238 | __releases(rnp->lock) | |
2239 | { | |
2240 | unsigned long oldmask = 0; | |
2241 | struct rcu_node *rnp_c; | |
2242 | ||
2243 | /* Walk up the rcu_node hierarchy. */ | |
2244 | for (;;) { | |
2245 | if (!(rnp->qsmask & mask) || rnp->gpnum != gps) { | |
2246 | ||
2247 | /* | |
2248 | * Our bit has already been cleared, or the | |
2249 | * relevant grace period is already over, so done. | |
2250 | */ | |
2251 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2252 | return; | |
2253 | } | |
2254 | WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ | |
2255 | rnp->qsmask &= ~mask; | |
2256 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, | |
2257 | mask, rnp->qsmask, rnp->level, | |
2258 | rnp->grplo, rnp->grphi, | |
2259 | !!rnp->gp_tasks); | |
2260 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { | |
2261 | ||
2262 | /* Other bits still set at this level, so done. */ | |
2263 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2264 | return; | |
2265 | } | |
2266 | mask = rnp->grpmask; | |
2267 | if (rnp->parent == NULL) { | |
2268 | ||
2269 | /* No more levels. Exit loop holding root lock. */ | |
2270 | ||
2271 | break; | |
2272 | } | |
2273 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2274 | rnp_c = rnp; | |
2275 | rnp = rnp->parent; | |
2276 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2277 | smp_mb__after_unlock_lock(); | |
2278 | oldmask = rnp_c->qsmask; | |
2279 | } | |
2280 | ||
2281 | /* | |
2282 | * Get here if we are the last CPU to pass through a quiescent | |
2283 | * state for this grace period. Invoke rcu_report_qs_rsp() | |
2284 | * to clean up and start the next grace period if one is needed. | |
2285 | */ | |
2286 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ | |
2287 | } | |
2288 | ||
2289 | /* | |
2290 | * Record a quiescent state for all tasks that were previously queued | |
2291 | * on the specified rcu_node structure and that were blocking the current | |
2292 | * RCU grace period. The caller must hold the specified rnp->lock with | |
2293 | * irqs disabled, and this lock is released upon return, but irqs remain | |
2294 | * disabled. | |
2295 | */ | |
2296 | static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp, | |
2297 | struct rcu_node *rnp, unsigned long flags) | |
2298 | __releases(rnp->lock) | |
2299 | { | |
2300 | unsigned long gps; | |
2301 | unsigned long mask; | |
2302 | struct rcu_node *rnp_p; | |
2303 | ||
2304 | if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p || | |
2305 | rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { | |
2306 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2307 | return; /* Still need more quiescent states! */ | |
2308 | } | |
2309 | ||
2310 | rnp_p = rnp->parent; | |
2311 | if (rnp_p == NULL) { | |
2312 | /* | |
2313 | * Only one rcu_node structure in the tree, so don't | |
2314 | * try to report up to its nonexistent parent! | |
2315 | */ | |
2316 | rcu_report_qs_rsp(rsp, flags); | |
2317 | return; | |
2318 | } | |
2319 | ||
2320 | /* Report up the rest of the hierarchy, tracking current ->gpnum. */ | |
2321 | gps = rnp->gpnum; | |
2322 | mask = rnp->grpmask; | |
2323 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2324 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
2325 | smp_mb__after_unlock_lock(); | |
2326 | rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags); | |
2327 | } | |
2328 | ||
2329 | /* | |
2330 | * Record a quiescent state for the specified CPU to that CPU's rcu_data | |
2331 | * structure. This must be either called from the specified CPU, or | |
2332 | * called when the specified CPU is known to be offline (and when it is | |
2333 | * also known that no other CPU is concurrently trying to help the offline | |
2334 | * CPU). The lastcomp argument is used to make sure we are still in the | |
2335 | * grace period of interest. We don't want to end the current grace period | |
2336 | * based on quiescent states detected in an earlier grace period! | |
2337 | */ | |
2338 | static void | |
2339 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) | |
2340 | { | |
2341 | unsigned long flags; | |
2342 | unsigned long mask; | |
2343 | bool needwake; | |
2344 | struct rcu_node *rnp; | |
2345 | ||
2346 | rnp = rdp->mynode; | |
2347 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2348 | smp_mb__after_unlock_lock(); | |
2349 | if ((rdp->passed_quiesce == 0 && | |
2350 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) || | |
2351 | rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum || | |
2352 | rdp->gpwrap) { | |
2353 | ||
2354 | /* | |
2355 | * The grace period in which this quiescent state was | |
2356 | * recorded has ended, so don't report it upwards. | |
2357 | * We will instead need a new quiescent state that lies | |
2358 | * within the current grace period. | |
2359 | */ | |
2360 | rdp->passed_quiesce = 0; /* need qs for new gp. */ | |
2361 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); | |
2362 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2363 | return; | |
2364 | } | |
2365 | mask = rdp->grpmask; | |
2366 | if ((rnp->qsmask & mask) == 0) { | |
2367 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2368 | } else { | |
2369 | rdp->qs_pending = 0; | |
2370 | ||
2371 | /* | |
2372 | * This GP can't end until cpu checks in, so all of our | |
2373 | * callbacks can be processed during the next GP. | |
2374 | */ | |
2375 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); | |
2376 | ||
2377 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); | |
2378 | /* ^^^ Released rnp->lock */ | |
2379 | if (needwake) | |
2380 | rcu_gp_kthread_wake(rsp); | |
2381 | } | |
2382 | } | |
2383 | ||
2384 | /* | |
2385 | * Check to see if there is a new grace period of which this CPU | |
2386 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2387 | * Otherwise, see if this CPU has just passed through its first | |
2388 | * quiescent state for this grace period, and record that fact if so. | |
2389 | */ | |
2390 | static void | |
2391 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
2392 | { | |
2393 | /* Check for grace-period ends and beginnings. */ | |
2394 | note_gp_changes(rsp, rdp); | |
2395 | ||
2396 | /* | |
2397 | * Does this CPU still need to do its part for current grace period? | |
2398 | * If no, return and let the other CPUs do their part as well. | |
2399 | */ | |
2400 | if (!rdp->qs_pending) | |
2401 | return; | |
2402 | ||
2403 | /* | |
2404 | * Was there a quiescent state since the beginning of the grace | |
2405 | * period? If no, then exit and wait for the next call. | |
2406 | */ | |
2407 | if (!rdp->passed_quiesce && | |
2408 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) | |
2409 | return; | |
2410 | ||
2411 | /* | |
2412 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2413 | * judge of that). | |
2414 | */ | |
2415 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); | |
2416 | } | |
2417 | ||
2418 | /* | |
2419 | * Send the specified CPU's RCU callbacks to the orphanage. The | |
2420 | * specified CPU must be offline, and the caller must hold the | |
2421 | * ->orphan_lock. | |
2422 | */ | |
2423 | static void | |
2424 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
2425 | struct rcu_node *rnp, struct rcu_data *rdp) | |
2426 | { | |
2427 | /* No-CBs CPUs do not have orphanable callbacks. */ | |
2428 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu)) | |
2429 | return; | |
2430 | ||
2431 | /* | |
2432 | * Orphan the callbacks. First adjust the counts. This is safe | |
2433 | * because _rcu_barrier() excludes CPU-hotplug operations, so it | |
2434 | * cannot be running now. Thus no memory barrier is required. | |
2435 | */ | |
2436 | if (rdp->nxtlist != NULL) { | |
2437 | rsp->qlen_lazy += rdp->qlen_lazy; | |
2438 | rsp->qlen += rdp->qlen; | |
2439 | rdp->n_cbs_orphaned += rdp->qlen; | |
2440 | rdp->qlen_lazy = 0; | |
2441 | WRITE_ONCE(rdp->qlen, 0); | |
2442 | } | |
2443 | ||
2444 | /* | |
2445 | * Next, move those callbacks still needing a grace period to | |
2446 | * the orphanage, where some other CPU will pick them up. | |
2447 | * Some of the callbacks might have gone partway through a grace | |
2448 | * period, but that is too bad. They get to start over because we | |
2449 | * cannot assume that grace periods are synchronized across CPUs. | |
2450 | * We don't bother updating the ->nxttail[] array yet, instead | |
2451 | * we just reset the whole thing later on. | |
2452 | */ | |
2453 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { | |
2454 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2455 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2456 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2457 | } | |
2458 | ||
2459 | /* | |
2460 | * Then move the ready-to-invoke callbacks to the orphanage, | |
2461 | * where some other CPU will pick them up. These will not be | |
2462 | * required to pass though another grace period: They are done. | |
2463 | */ | |
2464 | if (rdp->nxtlist != NULL) { | |
2465 | *rsp->orphan_donetail = rdp->nxtlist; | |
2466 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
2467 | } | |
2468 | ||
2469 | /* | |
2470 | * Finally, initialize the rcu_data structure's list to empty and | |
2471 | * disallow further callbacks on this CPU. | |
2472 | */ | |
2473 | init_callback_list(rdp); | |
2474 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; | |
2475 | } | |
2476 | ||
2477 | /* | |
2478 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
2479 | * orphanage. The caller must hold the ->orphan_lock. | |
2480 | */ | |
2481 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) | |
2482 | { | |
2483 | int i; | |
2484 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); | |
2485 | ||
2486 | /* No-CBs CPUs are handled specially. */ | |
2487 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || | |
2488 | rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) | |
2489 | return; | |
2490 | ||
2491 | /* Do the accounting first. */ | |
2492 | rdp->qlen_lazy += rsp->qlen_lazy; | |
2493 | rdp->qlen += rsp->qlen; | |
2494 | rdp->n_cbs_adopted += rsp->qlen; | |
2495 | if (rsp->qlen_lazy != rsp->qlen) | |
2496 | rcu_idle_count_callbacks_posted(); | |
2497 | rsp->qlen_lazy = 0; | |
2498 | rsp->qlen = 0; | |
2499 | ||
2500 | /* | |
2501 | * We do not need a memory barrier here because the only way we | |
2502 | * can get here if there is an rcu_barrier() in flight is if | |
2503 | * we are the task doing the rcu_barrier(). | |
2504 | */ | |
2505 | ||
2506 | /* First adopt the ready-to-invoke callbacks. */ | |
2507 | if (rsp->orphan_donelist != NULL) { | |
2508 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2509 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
2510 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
2511 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2512 | rdp->nxttail[i] = rsp->orphan_donetail; | |
2513 | rsp->orphan_donelist = NULL; | |
2514 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2515 | } | |
2516 | ||
2517 | /* And then adopt the callbacks that still need a grace period. */ | |
2518 | if (rsp->orphan_nxtlist != NULL) { | |
2519 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
2520 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
2521 | rsp->orphan_nxtlist = NULL; | |
2522 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2523 | } | |
2524 | } | |
2525 | ||
2526 | /* | |
2527 | * Trace the fact that this CPU is going offline. | |
2528 | */ | |
2529 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
2530 | { | |
2531 | RCU_TRACE(unsigned long mask); | |
2532 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
2533 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
2534 | ||
2535 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) | |
2536 | return; | |
2537 | ||
2538 | RCU_TRACE(mask = rdp->grpmask); | |
2539 | trace_rcu_grace_period(rsp->name, | |
2540 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
2541 | TPS("cpuofl")); | |
2542 | } | |
2543 | ||
2544 | /* | |
2545 | * All CPUs for the specified rcu_node structure have gone offline, | |
2546 | * and all tasks that were preempted within an RCU read-side critical | |
2547 | * section while running on one of those CPUs have since exited their RCU | |
2548 | * read-side critical section. Some other CPU is reporting this fact with | |
2549 | * the specified rcu_node structure's ->lock held and interrupts disabled. | |
2550 | * This function therefore goes up the tree of rcu_node structures, | |
2551 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that | |
2552 | * the leaf rcu_node structure's ->qsmaskinit field has already been | |
2553 | * updated | |
2554 | * | |
2555 | * This function does check that the specified rcu_node structure has | |
2556 | * all CPUs offline and no blocked tasks, so it is OK to invoke it | |
2557 | * prematurely. That said, invoking it after the fact will cost you | |
2558 | * a needless lock acquisition. So once it has done its work, don't | |
2559 | * invoke it again. | |
2560 | */ | |
2561 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) | |
2562 | { | |
2563 | long mask; | |
2564 | struct rcu_node *rnp = rnp_leaf; | |
2565 | ||
2566 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || | |
2567 | rnp->qsmaskinit || rcu_preempt_has_tasks(rnp)) | |
2568 | return; | |
2569 | for (;;) { | |
2570 | mask = rnp->grpmask; | |
2571 | rnp = rnp->parent; | |
2572 | if (!rnp) | |
2573 | break; | |
2574 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
2575 | smp_mb__after_unlock_lock(); /* GP memory ordering. */ | |
2576 | rnp->qsmaskinit &= ~mask; | |
2577 | rnp->qsmask &= ~mask; | |
2578 | if (rnp->qsmaskinit) { | |
2579 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2580 | return; | |
2581 | } | |
2582 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2583 | } | |
2584 | } | |
2585 | ||
2586 | /* | |
2587 | * The CPU is exiting the idle loop into the arch_cpu_idle_dead() | |
2588 | * function. We now remove it from the rcu_node tree's ->qsmaskinit | |
2589 | * bit masks. | |
2590 | */ | |
2591 | static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp) | |
2592 | { | |
2593 | unsigned long flags; | |
2594 | unsigned long mask; | |
2595 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
2596 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ | |
2597 | ||
2598 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) | |
2599 | return; | |
2600 | ||
2601 | /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ | |
2602 | mask = rdp->grpmask; | |
2603 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2604 | smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */ | |
2605 | rnp->qsmaskinitnext &= ~mask; | |
2606 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2607 | } | |
2608 | ||
2609 | /* | |
2610 | * The CPU has been completely removed, and some other CPU is reporting | |
2611 | * this fact from process context. Do the remainder of the cleanup, | |
2612 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
2613 | * adopting them. There can only be one CPU hotplug operation at a time, | |
2614 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
2615 | */ | |
2616 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) | |
2617 | { | |
2618 | unsigned long flags; | |
2619 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
2620 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ | |
2621 | ||
2622 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) | |
2623 | return; | |
2624 | ||
2625 | /* Adjust any no-longer-needed kthreads. */ | |
2626 | rcu_boost_kthread_setaffinity(rnp, -1); | |
2627 | ||
2628 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ | |
2629 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); | |
2630 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
2631 | rcu_adopt_orphan_cbs(rsp, flags); | |
2632 | raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags); | |
2633 | ||
2634 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, | |
2635 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2636 | cpu, rdp->qlen, rdp->nxtlist); | |
2637 | } | |
2638 | ||
2639 | /* | |
2640 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2641 | * period. Thottle as specified by rdp->blimit. | |
2642 | */ | |
2643 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) | |
2644 | { | |
2645 | unsigned long flags; | |
2646 | struct rcu_head *next, *list, **tail; | |
2647 | long bl, count, count_lazy; | |
2648 | int i; | |
2649 | ||
2650 | /* If no callbacks are ready, just return. */ | |
2651 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { | |
2652 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); | |
2653 | trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist), | |
2654 | need_resched(), is_idle_task(current), | |
2655 | rcu_is_callbacks_kthread()); | |
2656 | return; | |
2657 | } | |
2658 | ||
2659 | /* | |
2660 | * Extract the list of ready callbacks, disabling to prevent | |
2661 | * races with call_rcu() from interrupt handlers. | |
2662 | */ | |
2663 | local_irq_save(flags); | |
2664 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); | |
2665 | bl = rdp->blimit; | |
2666 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); | |
2667 | list = rdp->nxtlist; | |
2668 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2669 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2670 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
2671 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) | |
2672 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2673 | rdp->nxttail[i] = &rdp->nxtlist; | |
2674 | local_irq_restore(flags); | |
2675 | ||
2676 | /* Invoke callbacks. */ | |
2677 | count = count_lazy = 0; | |
2678 | while (list) { | |
2679 | next = list->next; | |
2680 | prefetch(next); | |
2681 | debug_rcu_head_unqueue(list); | |
2682 | if (__rcu_reclaim(rsp->name, list)) | |
2683 | count_lazy++; | |
2684 | list = next; | |
2685 | /* Stop only if limit reached and CPU has something to do. */ | |
2686 | if (++count >= bl && | |
2687 | (need_resched() || | |
2688 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
2689 | break; | |
2690 | } | |
2691 | ||
2692 | local_irq_save(flags); | |
2693 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), | |
2694 | is_idle_task(current), | |
2695 | rcu_is_callbacks_kthread()); | |
2696 | ||
2697 | /* Update count, and requeue any remaining callbacks. */ | |
2698 | if (list != NULL) { | |
2699 | *tail = rdp->nxtlist; | |
2700 | rdp->nxtlist = list; | |
2701 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
2702 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2703 | rdp->nxttail[i] = tail; | |
2704 | else | |
2705 | break; | |
2706 | } | |
2707 | smp_mb(); /* List handling before counting for rcu_barrier(). */ | |
2708 | rdp->qlen_lazy -= count_lazy; | |
2709 | WRITE_ONCE(rdp->qlen, rdp->qlen - count); | |
2710 | rdp->n_cbs_invoked += count; | |
2711 | ||
2712 | /* Reinstate batch limit if we have worked down the excess. */ | |
2713 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2714 | rdp->blimit = blimit; | |
2715 | ||
2716 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ | |
2717 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2718 | rdp->qlen_last_fqs_check = 0; | |
2719 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2720 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2721 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2722 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); | |
2723 | ||
2724 | local_irq_restore(flags); | |
2725 | ||
2726 | /* Re-invoke RCU core processing if there are callbacks remaining. */ | |
2727 | if (cpu_has_callbacks_ready_to_invoke(rdp)) | |
2728 | invoke_rcu_core(); | |
2729 | } | |
2730 | ||
2731 | /* | |
2732 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2733 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
2734 | * Also schedule RCU core processing. | |
2735 | * | |
2736 | * This function must be called from hardirq context. It is normally | |
2737 | * invoked from the scheduling-clock interrupt. If rcu_pending returns | |
2738 | * false, there is no point in invoking rcu_check_callbacks(). | |
2739 | */ | |
2740 | void rcu_check_callbacks(int user) | |
2741 | { | |
2742 | trace_rcu_utilization(TPS("Start scheduler-tick")); | |
2743 | increment_cpu_stall_ticks(); | |
2744 | if (user || rcu_is_cpu_rrupt_from_idle()) { | |
2745 | ||
2746 | /* | |
2747 | * Get here if this CPU took its interrupt from user | |
2748 | * mode or from the idle loop, and if this is not a | |
2749 | * nested interrupt. In this case, the CPU is in | |
2750 | * a quiescent state, so note it. | |
2751 | * | |
2752 | * No memory barrier is required here because both | |
2753 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local | |
2754 | * variables that other CPUs neither access nor modify, | |
2755 | * at least not while the corresponding CPU is online. | |
2756 | */ | |
2757 | ||
2758 | rcu_sched_qs(); | |
2759 | rcu_bh_qs(); | |
2760 | ||
2761 | } else if (!in_softirq()) { | |
2762 | ||
2763 | /* | |
2764 | * Get here if this CPU did not take its interrupt from | |
2765 | * softirq, in other words, if it is not interrupting | |
2766 | * a rcu_bh read-side critical section. This is an _bh | |
2767 | * critical section, so note it. | |
2768 | */ | |
2769 | ||
2770 | rcu_bh_qs(); | |
2771 | } | |
2772 | rcu_preempt_check_callbacks(); | |
2773 | if (rcu_pending()) | |
2774 | invoke_rcu_core(); | |
2775 | if (user) | |
2776 | rcu_note_voluntary_context_switch(current); | |
2777 | trace_rcu_utilization(TPS("End scheduler-tick")); | |
2778 | } | |
2779 | ||
2780 | /* | |
2781 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2782 | * have not yet encountered a quiescent state, using the function specified. | |
2783 | * Also initiate boosting for any threads blocked on the root rcu_node. | |
2784 | * | |
2785 | * The caller must have suppressed start of new grace periods. | |
2786 | */ | |
2787 | static void force_qs_rnp(struct rcu_state *rsp, | |
2788 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2789 | unsigned long *maxj), | |
2790 | bool *isidle, unsigned long *maxj) | |
2791 | { | |
2792 | unsigned long bit; | |
2793 | int cpu; | |
2794 | unsigned long flags; | |
2795 | unsigned long mask; | |
2796 | struct rcu_node *rnp; | |
2797 | ||
2798 | rcu_for_each_leaf_node(rsp, rnp) { | |
2799 | cond_resched_rcu_qs(); | |
2800 | mask = 0; | |
2801 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2802 | smp_mb__after_unlock_lock(); | |
2803 | if (rnp->qsmask == 0) { | |
2804 | if (rcu_state_p == &rcu_sched_state || | |
2805 | rsp != rcu_state_p || | |
2806 | rcu_preempt_blocked_readers_cgp(rnp)) { | |
2807 | /* | |
2808 | * No point in scanning bits because they | |
2809 | * are all zero. But we might need to | |
2810 | * priority-boost blocked readers. | |
2811 | */ | |
2812 | rcu_initiate_boost(rnp, flags); | |
2813 | /* rcu_initiate_boost() releases rnp->lock */ | |
2814 | continue; | |
2815 | } | |
2816 | if (rnp->parent && | |
2817 | (rnp->parent->qsmask & rnp->grpmask)) { | |
2818 | /* | |
2819 | * Race between grace-period | |
2820 | * initialization and task exiting RCU | |
2821 | * read-side critical section: Report. | |
2822 | */ | |
2823 | rcu_report_unblock_qs_rnp(rsp, rnp, flags); | |
2824 | /* rcu_report_unblock_qs_rnp() rlses ->lock */ | |
2825 | continue; | |
2826 | } | |
2827 | } | |
2828 | cpu = rnp->grplo; | |
2829 | bit = 1; | |
2830 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { | |
2831 | if ((rnp->qsmask & bit) != 0) { | |
2832 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) | |
2833 | mask |= bit; | |
2834 | } | |
2835 | } | |
2836 | if (mask != 0) { | |
2837 | /* Idle/offline CPUs, report (releases rnp->lock. */ | |
2838 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); | |
2839 | } else { | |
2840 | /* Nothing to do here, so just drop the lock. */ | |
2841 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2842 | } | |
2843 | } | |
2844 | } | |
2845 | ||
2846 | /* | |
2847 | * Force quiescent states on reluctant CPUs, and also detect which | |
2848 | * CPUs are in dyntick-idle mode. | |
2849 | */ | |
2850 | static void force_quiescent_state(struct rcu_state *rsp) | |
2851 | { | |
2852 | unsigned long flags; | |
2853 | bool ret; | |
2854 | struct rcu_node *rnp; | |
2855 | struct rcu_node *rnp_old = NULL; | |
2856 | ||
2857 | /* Funnel through hierarchy to reduce memory contention. */ | |
2858 | rnp = __this_cpu_read(rsp->rda->mynode); | |
2859 | for (; rnp != NULL; rnp = rnp->parent) { | |
2860 | ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2861 | !raw_spin_trylock(&rnp->fqslock); | |
2862 | if (rnp_old != NULL) | |
2863 | raw_spin_unlock(&rnp_old->fqslock); | |
2864 | if (ret) { | |
2865 | rsp->n_force_qs_lh++; | |
2866 | return; | |
2867 | } | |
2868 | rnp_old = rnp; | |
2869 | } | |
2870 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
2871 | ||
2872 | /* Reached the root of the rcu_node tree, acquire lock. */ | |
2873 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
2874 | smp_mb__after_unlock_lock(); | |
2875 | raw_spin_unlock(&rnp_old->fqslock); | |
2876 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
2877 | rsp->n_force_qs_lh++; | |
2878 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); | |
2879 | return; /* Someone beat us to it. */ | |
2880 | } | |
2881 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); | |
2882 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); | |
2883 | rcu_gp_kthread_wake(rsp); | |
2884 | } | |
2885 | ||
2886 | /* | |
2887 | * This does the RCU core processing work for the specified rcu_state | |
2888 | * and rcu_data structures. This may be called only from the CPU to | |
2889 | * whom the rdp belongs. | |
2890 | */ | |
2891 | static void | |
2892 | __rcu_process_callbacks(struct rcu_state *rsp) | |
2893 | { | |
2894 | unsigned long flags; | |
2895 | bool needwake; | |
2896 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); | |
2897 | ||
2898 | WARN_ON_ONCE(rdp->beenonline == 0); | |
2899 | ||
2900 | /* Update RCU state based on any recent quiescent states. */ | |
2901 | rcu_check_quiescent_state(rsp, rdp); | |
2902 | ||
2903 | /* Does this CPU require a not-yet-started grace period? */ | |
2904 | local_irq_save(flags); | |
2905 | if (cpu_needs_another_gp(rsp, rdp)) { | |
2906 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ | |
2907 | needwake = rcu_start_gp(rsp); | |
2908 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); | |
2909 | if (needwake) | |
2910 | rcu_gp_kthread_wake(rsp); | |
2911 | } else { | |
2912 | local_irq_restore(flags); | |
2913 | } | |
2914 | ||
2915 | /* If there are callbacks ready, invoke them. */ | |
2916 | if (cpu_has_callbacks_ready_to_invoke(rdp)) | |
2917 | invoke_rcu_callbacks(rsp, rdp); | |
2918 | ||
2919 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2920 | do_nocb_deferred_wakeup(rdp); | |
2921 | } | |
2922 | ||
2923 | /* | |
2924 | * Do RCU core processing for the current CPU. | |
2925 | */ | |
2926 | static void rcu_process_callbacks(struct softirq_action *unused) | |
2927 | { | |
2928 | struct rcu_state *rsp; | |
2929 | ||
2930 | if (cpu_is_offline(smp_processor_id())) | |
2931 | return; | |
2932 | trace_rcu_utilization(TPS("Start RCU core")); | |
2933 | for_each_rcu_flavor(rsp) | |
2934 | __rcu_process_callbacks(rsp); | |
2935 | trace_rcu_utilization(TPS("End RCU core")); | |
2936 | } | |
2937 | ||
2938 | /* | |
2939 | * Schedule RCU callback invocation. If the specified type of RCU | |
2940 | * does not support RCU priority boosting, just do a direct call, | |
2941 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
2942 | * are running on the current CPU with softirqs disabled, the | |
2943 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
2944 | */ | |
2945 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) | |
2946 | { | |
2947 | if (unlikely(!READ_ONCE(rcu_scheduler_fully_active))) | |
2948 | return; | |
2949 | if (likely(!rsp->boost)) { | |
2950 | rcu_do_batch(rsp, rdp); | |
2951 | return; | |
2952 | } | |
2953 | invoke_rcu_callbacks_kthread(); | |
2954 | } | |
2955 | ||
2956 | static void invoke_rcu_core(void) | |
2957 | { | |
2958 | if (cpu_online(smp_processor_id())) | |
2959 | raise_softirq(RCU_SOFTIRQ); | |
2960 | } | |
2961 | ||
2962 | /* | |
2963 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2964 | */ | |
2965 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2966 | struct rcu_head *head, unsigned long flags) | |
2967 | { | |
2968 | bool needwake; | |
2969 | ||
2970 | /* | |
2971 | * If called from an extended quiescent state, invoke the RCU | |
2972 | * core in order to force a re-evaluation of RCU's idleness. | |
2973 | */ | |
2974 | if (!rcu_is_watching()) | |
2975 | invoke_rcu_core(); | |
2976 | ||
2977 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ | |
2978 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) | |
2979 | return; | |
2980 | ||
2981 | /* | |
2982 | * Force the grace period if too many callbacks or too long waiting. | |
2983 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2984 | * if some other CPU has recently done so. Also, don't bother | |
2985 | * invoking force_quiescent_state() if the newly enqueued callback | |
2986 | * is the only one waiting for a grace period to complete. | |
2987 | */ | |
2988 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { | |
2989 | ||
2990 | /* Are we ignoring a completed grace period? */ | |
2991 | note_gp_changes(rsp, rdp); | |
2992 | ||
2993 | /* Start a new grace period if one not already started. */ | |
2994 | if (!rcu_gp_in_progress(rsp)) { | |
2995 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
2996 | ||
2997 | raw_spin_lock(&rnp_root->lock); | |
2998 | smp_mb__after_unlock_lock(); | |
2999 | needwake = rcu_start_gp(rsp); | |
3000 | raw_spin_unlock(&rnp_root->lock); | |
3001 | if (needwake) | |
3002 | rcu_gp_kthread_wake(rsp); | |
3003 | } else { | |
3004 | /* Give the grace period a kick. */ | |
3005 | rdp->blimit = LONG_MAX; | |
3006 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
3007 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
3008 | force_quiescent_state(rsp); | |
3009 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
3010 | rdp->qlen_last_fqs_check = rdp->qlen; | |
3011 | } | |
3012 | } | |
3013 | } | |
3014 | ||
3015 | /* | |
3016 | * RCU callback function to leak a callback. | |
3017 | */ | |
3018 | static void rcu_leak_callback(struct rcu_head *rhp) | |
3019 | { | |
3020 | } | |
3021 | ||
3022 | /* | |
3023 | * Helper function for call_rcu() and friends. The cpu argument will | |
3024 | * normally be -1, indicating "currently running CPU". It may specify | |
3025 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
3026 | * is expected to specify a CPU. | |
3027 | */ | |
3028 | static void | |
3029 | __call_rcu(struct rcu_head *head, rcu_callback_t func, | |
3030 | struct rcu_state *rsp, int cpu, bool lazy) | |
3031 | { | |
3032 | unsigned long flags; | |
3033 | struct rcu_data *rdp; | |
3034 | ||
3035 | WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */ | |
3036 | if (debug_rcu_head_queue(head)) { | |
3037 | /* Probable double call_rcu(), so leak the callback. */ | |
3038 | WRITE_ONCE(head->func, rcu_leak_callback); | |
3039 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); | |
3040 | return; | |
3041 | } | |
3042 | head->func = func; | |
3043 | head->next = NULL; | |
3044 | ||
3045 | /* | |
3046 | * Opportunistically note grace-period endings and beginnings. | |
3047 | * Note that we might see a beginning right after we see an | |
3048 | * end, but never vice versa, since this CPU has to pass through | |
3049 | * a quiescent state betweentimes. | |
3050 | */ | |
3051 | local_irq_save(flags); | |
3052 | rdp = this_cpu_ptr(rsp->rda); | |
3053 | ||
3054 | /* Add the callback to our list. */ | |
3055 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { | |
3056 | int offline; | |
3057 | ||
3058 | if (cpu != -1) | |
3059 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
3060 | if (likely(rdp->mynode)) { | |
3061 | /* Post-boot, so this should be for a no-CBs CPU. */ | |
3062 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); | |
3063 | WARN_ON_ONCE(offline); | |
3064 | /* Offline CPU, _call_rcu() illegal, leak callback. */ | |
3065 | local_irq_restore(flags); | |
3066 | return; | |
3067 | } | |
3068 | /* | |
3069 | * Very early boot, before rcu_init(). Initialize if needed | |
3070 | * and then drop through to queue the callback. | |
3071 | */ | |
3072 | BUG_ON(cpu != -1); | |
3073 | WARN_ON_ONCE(!rcu_is_watching()); | |
3074 | if (!likely(rdp->nxtlist)) | |
3075 | init_default_callback_list(rdp); | |
3076 | } | |
3077 | WRITE_ONCE(rdp->qlen, rdp->qlen + 1); | |
3078 | if (lazy) | |
3079 | rdp->qlen_lazy++; | |
3080 | else | |
3081 | rcu_idle_count_callbacks_posted(); | |
3082 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ | |
3083 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
3084 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
3085 | ||
3086 | if (__is_kfree_rcu_offset((unsigned long)func)) | |
3087 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
3088 | rdp->qlen_lazy, rdp->qlen); | |
3089 | else | |
3090 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); | |
3091 | ||
3092 | /* Go handle any RCU core processing required. */ | |
3093 | __call_rcu_core(rsp, rdp, head, flags); | |
3094 | local_irq_restore(flags); | |
3095 | } | |
3096 | ||
3097 | /* | |
3098 | * Queue an RCU-sched callback for invocation after a grace period. | |
3099 | */ | |
3100 | void call_rcu_sched(struct rcu_head *head, rcu_callback_t func) | |
3101 | { | |
3102 | __call_rcu(head, func, &rcu_sched_state, -1, 0); | |
3103 | } | |
3104 | EXPORT_SYMBOL_GPL(call_rcu_sched); | |
3105 | ||
3106 | /* | |
3107 | * Queue an RCU callback for invocation after a quicker grace period. | |
3108 | */ | |
3109 | void call_rcu_bh(struct rcu_head *head, rcu_callback_t func) | |
3110 | { | |
3111 | __call_rcu(head, func, &rcu_bh_state, -1, 0); | |
3112 | } | |
3113 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
3114 | ||
3115 | /* | |
3116 | * Queue an RCU callback for lazy invocation after a grace period. | |
3117 | * This will likely be later named something like "call_rcu_lazy()", | |
3118 | * but this change will require some way of tagging the lazy RCU | |
3119 | * callbacks in the list of pending callbacks. Until then, this | |
3120 | * function may only be called from __kfree_rcu(). | |
3121 | */ | |
3122 | void kfree_call_rcu(struct rcu_head *head, | |
3123 | rcu_callback_t func) | |
3124 | { | |
3125 | __call_rcu(head, func, rcu_state_p, -1, 1); | |
3126 | } | |
3127 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
3128 | ||
3129 | /* | |
3130 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
3131 | * any blocking grace-period wait automatically implies a grace period | |
3132 | * if there is only one CPU online at any point time during execution | |
3133 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
3134 | * occasionally incorrectly indicate that there are multiple CPUs online | |
3135 | * when there was in fact only one the whole time, as this just adds | |
3136 | * some overhead: RCU still operates correctly. | |
3137 | */ | |
3138 | static inline int rcu_blocking_is_gp(void) | |
3139 | { | |
3140 | int ret; | |
3141 | ||
3142 | might_sleep(); /* Check for RCU read-side critical section. */ | |
3143 | preempt_disable(); | |
3144 | ret = num_online_cpus() <= 1; | |
3145 | preempt_enable(); | |
3146 | return ret; | |
3147 | } | |
3148 | ||
3149 | /** | |
3150 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
3151 | * | |
3152 | * Control will return to the caller some time after a full rcu-sched | |
3153 | * grace period has elapsed, in other words after all currently executing | |
3154 | * rcu-sched read-side critical sections have completed. These read-side | |
3155 | * critical sections are delimited by rcu_read_lock_sched() and | |
3156 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
3157 | * local_irq_disable(), and so on may be used in place of | |
3158 | * rcu_read_lock_sched(). | |
3159 | * | |
3160 | * This means that all preempt_disable code sequences, including NMI and | |
3161 | * non-threaded hardware-interrupt handlers, in progress on entry will | |
3162 | * have completed before this primitive returns. However, this does not | |
3163 | * guarantee that softirq handlers will have completed, since in some | |
3164 | * kernels, these handlers can run in process context, and can block. | |
3165 | * | |
3166 | * Note that this guarantee implies further memory-ordering guarantees. | |
3167 | * On systems with more than one CPU, when synchronize_sched() returns, | |
3168 | * each CPU is guaranteed to have executed a full memory barrier since the | |
3169 | * end of its last RCU-sched read-side critical section whose beginning | |
3170 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
3171 | * an RCU read-side critical section that extends beyond the return from | |
3172 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
3173 | * after the beginning of synchronize_sched() and before the beginning of | |
3174 | * that RCU read-side critical section. Note that these guarantees include | |
3175 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
3176 | * that are executing in the kernel. | |
3177 | * | |
3178 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
3179 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
3180 | * to have executed a full memory barrier during the execution of | |
3181 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
3182 | * again only if the system has more than one CPU). | |
3183 | * | |
3184 | * This primitive provides the guarantees made by the (now removed) | |
3185 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
3186 | * guarantees that rcu_read_lock() sections will have completed. | |
3187 | * In "classic RCU", these two guarantees happen to be one and | |
3188 | * the same, but can differ in realtime RCU implementations. | |
3189 | */ | |
3190 | void synchronize_sched(void) | |
3191 | { | |
3192 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || | |
3193 | lock_is_held(&rcu_lock_map) || | |
3194 | lock_is_held(&rcu_sched_lock_map), | |
3195 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
3196 | if (rcu_blocking_is_gp()) | |
3197 | return; | |
3198 | if (rcu_gp_is_expedited()) | |
3199 | synchronize_sched_expedited(); | |
3200 | else | |
3201 | wait_rcu_gp(call_rcu_sched); | |
3202 | } | |
3203 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
3204 | ||
3205 | /** | |
3206 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
3207 | * | |
3208 | * Control will return to the caller some time after a full rcu_bh grace | |
3209 | * period has elapsed, in other words after all currently executing rcu_bh | |
3210 | * read-side critical sections have completed. RCU read-side critical | |
3211 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
3212 | * and may be nested. | |
3213 | * | |
3214 | * See the description of synchronize_sched() for more detailed information | |
3215 | * on memory ordering guarantees. | |
3216 | */ | |
3217 | void synchronize_rcu_bh(void) | |
3218 | { | |
3219 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || | |
3220 | lock_is_held(&rcu_lock_map) || | |
3221 | lock_is_held(&rcu_sched_lock_map), | |
3222 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
3223 | if (rcu_blocking_is_gp()) | |
3224 | return; | |
3225 | if (rcu_gp_is_expedited()) | |
3226 | synchronize_rcu_bh_expedited(); | |
3227 | else | |
3228 | wait_rcu_gp(call_rcu_bh); | |
3229 | } | |
3230 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
3231 | ||
3232 | /** | |
3233 | * get_state_synchronize_rcu - Snapshot current RCU state | |
3234 | * | |
3235 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
3236 | * to determine whether or not a full grace period has elapsed in the | |
3237 | * meantime. | |
3238 | */ | |
3239 | unsigned long get_state_synchronize_rcu(void) | |
3240 | { | |
3241 | /* | |
3242 | * Any prior manipulation of RCU-protected data must happen | |
3243 | * before the load from ->gpnum. | |
3244 | */ | |
3245 | smp_mb(); /* ^^^ */ | |
3246 | ||
3247 | /* | |
3248 | * Make sure this load happens before the purportedly | |
3249 | * time-consuming work between get_state_synchronize_rcu() | |
3250 | * and cond_synchronize_rcu(). | |
3251 | */ | |
3252 | return smp_load_acquire(&rcu_state_p->gpnum); | |
3253 | } | |
3254 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
3255 | ||
3256 | /** | |
3257 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
3258 | * | |
3259 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
3260 | * | |
3261 | * If a full RCU grace period has elapsed since the earlier call to | |
3262 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
3263 | * synchronize_rcu() to wait for a full grace period. | |
3264 | * | |
3265 | * Yes, this function does not take counter wrap into account. But | |
3266 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3267 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3268 | * so waiting for one additional grace period should be just fine. | |
3269 | */ | |
3270 | void cond_synchronize_rcu(unsigned long oldstate) | |
3271 | { | |
3272 | unsigned long newstate; | |
3273 | ||
3274 | /* | |
3275 | * Ensure that this load happens before any RCU-destructive | |
3276 | * actions the caller might carry out after we return. | |
3277 | */ | |
3278 | newstate = smp_load_acquire(&rcu_state_p->completed); | |
3279 | if (ULONG_CMP_GE(oldstate, newstate)) | |
3280 | synchronize_rcu(); | |
3281 | } | |
3282 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
3283 | ||
3284 | /** | |
3285 | * get_state_synchronize_sched - Snapshot current RCU-sched state | |
3286 | * | |
3287 | * Returns a cookie that is used by a later call to cond_synchronize_sched() | |
3288 | * to determine whether or not a full grace period has elapsed in the | |
3289 | * meantime. | |
3290 | */ | |
3291 | unsigned long get_state_synchronize_sched(void) | |
3292 | { | |
3293 | /* | |
3294 | * Any prior manipulation of RCU-protected data must happen | |
3295 | * before the load from ->gpnum. | |
3296 | */ | |
3297 | smp_mb(); /* ^^^ */ | |
3298 | ||
3299 | /* | |
3300 | * Make sure this load happens before the purportedly | |
3301 | * time-consuming work between get_state_synchronize_sched() | |
3302 | * and cond_synchronize_sched(). | |
3303 | */ | |
3304 | return smp_load_acquire(&rcu_sched_state.gpnum); | |
3305 | } | |
3306 | EXPORT_SYMBOL_GPL(get_state_synchronize_sched); | |
3307 | ||
3308 | /** | |
3309 | * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period | |
3310 | * | |
3311 | * @oldstate: return value from earlier call to get_state_synchronize_sched() | |
3312 | * | |
3313 | * If a full RCU-sched grace period has elapsed since the earlier call to | |
3314 | * get_state_synchronize_sched(), just return. Otherwise, invoke | |
3315 | * synchronize_sched() to wait for a full grace period. | |
3316 | * | |
3317 | * Yes, this function does not take counter wrap into account. But | |
3318 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3319 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3320 | * so waiting for one additional grace period should be just fine. | |
3321 | */ | |
3322 | void cond_synchronize_sched(unsigned long oldstate) | |
3323 | { | |
3324 | unsigned long newstate; | |
3325 | ||
3326 | /* | |
3327 | * Ensure that this load happens before any RCU-destructive | |
3328 | * actions the caller might carry out after we return. | |
3329 | */ | |
3330 | newstate = smp_load_acquire(&rcu_sched_state.completed); | |
3331 | if (ULONG_CMP_GE(oldstate, newstate)) | |
3332 | synchronize_sched(); | |
3333 | } | |
3334 | EXPORT_SYMBOL_GPL(cond_synchronize_sched); | |
3335 | ||
3336 | /* Adjust sequence number for start of update-side operation. */ | |
3337 | static void rcu_seq_start(unsigned long *sp) | |
3338 | { | |
3339 | WRITE_ONCE(*sp, *sp + 1); | |
3340 | smp_mb(); /* Ensure update-side operation after counter increment. */ | |
3341 | WARN_ON_ONCE(!(*sp & 0x1)); | |
3342 | } | |
3343 | ||
3344 | /* Adjust sequence number for end of update-side operation. */ | |
3345 | static void rcu_seq_end(unsigned long *sp) | |
3346 | { | |
3347 | smp_mb(); /* Ensure update-side operation before counter increment. */ | |
3348 | WRITE_ONCE(*sp, *sp + 1); | |
3349 | WARN_ON_ONCE(*sp & 0x1); | |
3350 | } | |
3351 | ||
3352 | /* Take a snapshot of the update side's sequence number. */ | |
3353 | static unsigned long rcu_seq_snap(unsigned long *sp) | |
3354 | { | |
3355 | unsigned long s; | |
3356 | ||
3357 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
3358 | s = (READ_ONCE(*sp) + 3) & ~0x1; | |
3359 | smp_mb(); /* Above access must not bleed into critical section. */ | |
3360 | return s; | |
3361 | } | |
3362 | ||
3363 | /* | |
3364 | * Given a snapshot from rcu_seq_snap(), determine whether or not a | |
3365 | * full update-side operation has occurred. | |
3366 | */ | |
3367 | static bool rcu_seq_done(unsigned long *sp, unsigned long s) | |
3368 | { | |
3369 | return ULONG_CMP_GE(READ_ONCE(*sp), s); | |
3370 | } | |
3371 | ||
3372 | /* Wrapper functions for expedited grace periods. */ | |
3373 | static void rcu_exp_gp_seq_start(struct rcu_state *rsp) | |
3374 | { | |
3375 | rcu_seq_start(&rsp->expedited_sequence); | |
3376 | } | |
3377 | static void rcu_exp_gp_seq_end(struct rcu_state *rsp) | |
3378 | { | |
3379 | rcu_seq_end(&rsp->expedited_sequence); | |
3380 | smp_mb(); /* Ensure that consecutive grace periods serialize. */ | |
3381 | } | |
3382 | static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp) | |
3383 | { | |
3384 | return rcu_seq_snap(&rsp->expedited_sequence); | |
3385 | } | |
3386 | static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s) | |
3387 | { | |
3388 | return rcu_seq_done(&rsp->expedited_sequence, s); | |
3389 | } | |
3390 | ||
3391 | /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */ | |
3392 | static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp, | |
3393 | struct rcu_data *rdp, | |
3394 | atomic_long_t *stat, unsigned long s) | |
3395 | { | |
3396 | if (rcu_exp_gp_seq_done(rsp, s)) { | |
3397 | if (rnp) | |
3398 | mutex_unlock(&rnp->exp_funnel_mutex); | |
3399 | else if (rdp) | |
3400 | mutex_unlock(&rdp->exp_funnel_mutex); | |
3401 | /* Ensure test happens before caller kfree(). */ | |
3402 | smp_mb__before_atomic(); /* ^^^ */ | |
3403 | atomic_long_inc(stat); | |
3404 | return true; | |
3405 | } | |
3406 | return false; | |
3407 | } | |
3408 | ||
3409 | /* | |
3410 | * Funnel-lock acquisition for expedited grace periods. Returns a | |
3411 | * pointer to the root rcu_node structure, or NULL if some other | |
3412 | * task did the expedited grace period for us. | |
3413 | */ | |
3414 | static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s) | |
3415 | { | |
3416 | struct rcu_data *rdp; | |
3417 | struct rcu_node *rnp0; | |
3418 | struct rcu_node *rnp1 = NULL; | |
3419 | ||
3420 | /* | |
3421 | * First try directly acquiring the root lock in order to reduce | |
3422 | * latency in the common case where expedited grace periods are | |
3423 | * rare. We check mutex_is_locked() to avoid pathological levels of | |
3424 | * memory contention on ->exp_funnel_mutex in the heavy-load case. | |
3425 | */ | |
3426 | rnp0 = rcu_get_root(rsp); | |
3427 | if (!mutex_is_locked(&rnp0->exp_funnel_mutex)) { | |
3428 | if (mutex_trylock(&rnp0->exp_funnel_mutex)) { | |
3429 | if (sync_exp_work_done(rsp, rnp0, NULL, | |
3430 | &rsp->expedited_workdone0, s)) | |
3431 | return NULL; | |
3432 | return rnp0; | |
3433 | } | |
3434 | } | |
3435 | ||
3436 | /* | |
3437 | * Each pass through the following loop works its way | |
3438 | * up the rcu_node tree, returning if others have done the | |
3439 | * work or otherwise falls through holding the root rnp's | |
3440 | * ->exp_funnel_mutex. The mapping from CPU to rcu_node structure | |
3441 | * can be inexact, as it is just promoting locality and is not | |
3442 | * strictly needed for correctness. | |
3443 | */ | |
3444 | rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id()); | |
3445 | if (sync_exp_work_done(rsp, NULL, NULL, &rsp->expedited_workdone1, s)) | |
3446 | return NULL; | |
3447 | mutex_lock(&rdp->exp_funnel_mutex); | |
3448 | rnp0 = rdp->mynode; | |
3449 | for (; rnp0 != NULL; rnp0 = rnp0->parent) { | |
3450 | if (sync_exp_work_done(rsp, rnp1, rdp, | |
3451 | &rsp->expedited_workdone2, s)) | |
3452 | return NULL; | |
3453 | mutex_lock(&rnp0->exp_funnel_mutex); | |
3454 | if (rnp1) | |
3455 | mutex_unlock(&rnp1->exp_funnel_mutex); | |
3456 | else | |
3457 | mutex_unlock(&rdp->exp_funnel_mutex); | |
3458 | rnp1 = rnp0; | |
3459 | } | |
3460 | if (sync_exp_work_done(rsp, rnp1, rdp, | |
3461 | &rsp->expedited_workdone3, s)) | |
3462 | return NULL; | |
3463 | return rnp1; | |
3464 | } | |
3465 | ||
3466 | /* Invoked on each online non-idle CPU for expedited quiescent state. */ | |
3467 | static int synchronize_sched_expedited_cpu_stop(void *data) | |
3468 | { | |
3469 | struct rcu_data *rdp = data; | |
3470 | struct rcu_state *rsp = rdp->rsp; | |
3471 | ||
3472 | /* We are here: If we are last, do the wakeup. */ | |
3473 | rdp->exp_done = true; | |
3474 | if (atomic_dec_and_test(&rsp->expedited_need_qs)) | |
3475 | wake_up(&rsp->expedited_wq); | |
3476 | return 0; | |
3477 | } | |
3478 | ||
3479 | static void synchronize_sched_expedited_wait(struct rcu_state *rsp) | |
3480 | { | |
3481 | int cpu; | |
3482 | unsigned long jiffies_stall; | |
3483 | unsigned long jiffies_start; | |
3484 | struct rcu_data *rdp; | |
3485 | int ret; | |
3486 | ||
3487 | jiffies_stall = rcu_jiffies_till_stall_check(); | |
3488 | jiffies_start = jiffies; | |
3489 | ||
3490 | for (;;) { | |
3491 | ret = wait_event_interruptible_timeout( | |
3492 | rsp->expedited_wq, | |
3493 | !atomic_read(&rsp->expedited_need_qs), | |
3494 | jiffies_stall); | |
3495 | if (ret > 0) | |
3496 | return; | |
3497 | if (ret < 0) { | |
3498 | /* Hit a signal, disable CPU stall warnings. */ | |
3499 | wait_event(rsp->expedited_wq, | |
3500 | !atomic_read(&rsp->expedited_need_qs)); | |
3501 | return; | |
3502 | } | |
3503 | pr_err("INFO: %s detected expedited stalls on CPUs: {", | |
3504 | rsp->name); | |
3505 | for_each_online_cpu(cpu) { | |
3506 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
3507 | ||
3508 | if (rdp->exp_done) | |
3509 | continue; | |
3510 | pr_cont(" %d", cpu); | |
3511 | } | |
3512 | pr_cont(" } %lu jiffies s: %lu\n", | |
3513 | jiffies - jiffies_start, rsp->expedited_sequence); | |
3514 | for_each_online_cpu(cpu) { | |
3515 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
3516 | ||
3517 | if (rdp->exp_done) | |
3518 | continue; | |
3519 | dump_cpu_task(cpu); | |
3520 | } | |
3521 | jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3; | |
3522 | } | |
3523 | } | |
3524 | ||
3525 | /** | |
3526 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
3527 | * | |
3528 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
3529 | * approach to force the grace period to end quickly. This consumes | |
3530 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
3531 | * so is thus not recommended for any sort of common-case code. In fact, | |
3532 | * if you are using synchronize_sched_expedited() in a loop, please | |
3533 | * restructure your code to batch your updates, and then use a single | |
3534 | * synchronize_sched() instead. | |
3535 | * | |
3536 | * This implementation can be thought of as an application of sequence | |
3537 | * locking to expedited grace periods, but using the sequence counter to | |
3538 | * determine when someone else has already done the work instead of for | |
3539 | * retrying readers. | |
3540 | */ | |
3541 | void synchronize_sched_expedited(void) | |
3542 | { | |
3543 | int cpu; | |
3544 | unsigned long s; | |
3545 | struct rcu_node *rnp; | |
3546 | struct rcu_state *rsp = &rcu_sched_state; | |
3547 | ||
3548 | /* Take a snapshot of the sequence number. */ | |
3549 | s = rcu_exp_gp_seq_snap(rsp); | |
3550 | ||
3551 | if (!try_get_online_cpus()) { | |
3552 | /* CPU hotplug operation in flight, fall back to normal GP. */ | |
3553 | wait_rcu_gp(call_rcu_sched); | |
3554 | atomic_long_inc(&rsp->expedited_normal); | |
3555 | return; | |
3556 | } | |
3557 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); | |
3558 | ||
3559 | rnp = exp_funnel_lock(rsp, s); | |
3560 | if (rnp == NULL) { | |
3561 | put_online_cpus(); | |
3562 | return; /* Someone else did our work for us. */ | |
3563 | } | |
3564 | ||
3565 | rcu_exp_gp_seq_start(rsp); | |
3566 | ||
3567 | /* Stop each CPU that is online, non-idle, and not us. */ | |
3568 | init_waitqueue_head(&rsp->expedited_wq); | |
3569 | atomic_set(&rsp->expedited_need_qs, 1); /* Extra count avoids race. */ | |
3570 | for_each_online_cpu(cpu) { | |
3571 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
3572 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); | |
3573 | ||
3574 | rdp->exp_done = false; | |
3575 | ||
3576 | /* Skip our CPU and any idle CPUs. */ | |
3577 | if (raw_smp_processor_id() == cpu || | |
3578 | !(atomic_add_return(0, &rdtp->dynticks) & 0x1)) | |
3579 | continue; | |
3580 | atomic_inc(&rsp->expedited_need_qs); | |
3581 | stop_one_cpu_nowait(cpu, synchronize_sched_expedited_cpu_stop, | |
3582 | rdp, &rdp->exp_stop_work); | |
3583 | } | |
3584 | ||
3585 | /* Remove extra count and, if necessary, wait for CPUs to stop. */ | |
3586 | if (!atomic_dec_and_test(&rsp->expedited_need_qs)) | |
3587 | synchronize_sched_expedited_wait(rsp); | |
3588 | ||
3589 | rcu_exp_gp_seq_end(rsp); | |
3590 | mutex_unlock(&rnp->exp_funnel_mutex); | |
3591 | ||
3592 | put_online_cpus(); | |
3593 | } | |
3594 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
3595 | ||
3596 | /* | |
3597 | * Check to see if there is any immediate RCU-related work to be done | |
3598 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
3599 | * The checks are in order of increasing expense: checks that can be | |
3600 | * carried out against CPU-local state are performed first. However, | |
3601 | * we must check for CPU stalls first, else we might not get a chance. | |
3602 | */ | |
3603 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
3604 | { | |
3605 | struct rcu_node *rnp = rdp->mynode; | |
3606 | ||
3607 | rdp->n_rcu_pending++; | |
3608 | ||
3609 | /* Check for CPU stalls, if enabled. */ | |
3610 | check_cpu_stall(rsp, rdp); | |
3611 | ||
3612 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ | |
3613 | if (rcu_nohz_full_cpu(rsp)) | |
3614 | return 0; | |
3615 | ||
3616 | /* Is the RCU core waiting for a quiescent state from this CPU? */ | |
3617 | if (rcu_scheduler_fully_active && | |
3618 | rdp->qs_pending && !rdp->passed_quiesce && | |
3619 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) { | |
3620 | rdp->n_rp_qs_pending++; | |
3621 | } else if (rdp->qs_pending && | |
3622 | (rdp->passed_quiesce || | |
3623 | rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) { | |
3624 | rdp->n_rp_report_qs++; | |
3625 | return 1; | |
3626 | } | |
3627 | ||
3628 | /* Does this CPU have callbacks ready to invoke? */ | |
3629 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { | |
3630 | rdp->n_rp_cb_ready++; | |
3631 | return 1; | |
3632 | } | |
3633 | ||
3634 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
3635 | if (cpu_needs_another_gp(rsp, rdp)) { | |
3636 | rdp->n_rp_cpu_needs_gp++; | |
3637 | return 1; | |
3638 | } | |
3639 | ||
3640 | /* Has another RCU grace period completed? */ | |
3641 | if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ | |
3642 | rdp->n_rp_gp_completed++; | |
3643 | return 1; | |
3644 | } | |
3645 | ||
3646 | /* Has a new RCU grace period started? */ | |
3647 | if (READ_ONCE(rnp->gpnum) != rdp->gpnum || | |
3648 | unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */ | |
3649 | rdp->n_rp_gp_started++; | |
3650 | return 1; | |
3651 | } | |
3652 | ||
3653 | /* Does this CPU need a deferred NOCB wakeup? */ | |
3654 | if (rcu_nocb_need_deferred_wakeup(rdp)) { | |
3655 | rdp->n_rp_nocb_defer_wakeup++; | |
3656 | return 1; | |
3657 | } | |
3658 | ||
3659 | /* nothing to do */ | |
3660 | rdp->n_rp_need_nothing++; | |
3661 | return 0; | |
3662 | } | |
3663 | ||
3664 | /* | |
3665 | * Check to see if there is any immediate RCU-related work to be done | |
3666 | * by the current CPU, returning 1 if so. This function is part of the | |
3667 | * RCU implementation; it is -not- an exported member of the RCU API. | |
3668 | */ | |
3669 | static int rcu_pending(void) | |
3670 | { | |
3671 | struct rcu_state *rsp; | |
3672 | ||
3673 | for_each_rcu_flavor(rsp) | |
3674 | if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda))) | |
3675 | return 1; | |
3676 | return 0; | |
3677 | } | |
3678 | ||
3679 | /* | |
3680 | * Return true if the specified CPU has any callback. If all_lazy is | |
3681 | * non-NULL, store an indication of whether all callbacks are lazy. | |
3682 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
3683 | */ | |
3684 | static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy) | |
3685 | { | |
3686 | bool al = true; | |
3687 | bool hc = false; | |
3688 | struct rcu_data *rdp; | |
3689 | struct rcu_state *rsp; | |
3690 | ||
3691 | for_each_rcu_flavor(rsp) { | |
3692 | rdp = this_cpu_ptr(rsp->rda); | |
3693 | if (!rdp->nxtlist) | |
3694 | continue; | |
3695 | hc = true; | |
3696 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
3697 | al = false; | |
3698 | break; | |
3699 | } | |
3700 | } | |
3701 | if (all_lazy) | |
3702 | *all_lazy = al; | |
3703 | return hc; | |
3704 | } | |
3705 | ||
3706 | /* | |
3707 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
3708 | * the compiler is expected to optimize this away. | |
3709 | */ | |
3710 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, | |
3711 | int cpu, unsigned long done) | |
3712 | { | |
3713 | trace_rcu_barrier(rsp->name, s, cpu, | |
3714 | atomic_read(&rsp->barrier_cpu_count), done); | |
3715 | } | |
3716 | ||
3717 | /* | |
3718 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
3719 | * up the task executing _rcu_barrier(). | |
3720 | */ | |
3721 | static void rcu_barrier_callback(struct rcu_head *rhp) | |
3722 | { | |
3723 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); | |
3724 | struct rcu_state *rsp = rdp->rsp; | |
3725 | ||
3726 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { | |
3727 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence); | |
3728 | complete(&rsp->barrier_completion); | |
3729 | } else { | |
3730 | _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence); | |
3731 | } | |
3732 | } | |
3733 | ||
3734 | /* | |
3735 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3736 | */ | |
3737 | static void rcu_barrier_func(void *type) | |
3738 | { | |
3739 | struct rcu_state *rsp = type; | |
3740 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); | |
3741 | ||
3742 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence); | |
3743 | atomic_inc(&rsp->barrier_cpu_count); | |
3744 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); | |
3745 | } | |
3746 | ||
3747 | /* | |
3748 | * Orchestrate the specified type of RCU barrier, waiting for all | |
3749 | * RCU callbacks of the specified type to complete. | |
3750 | */ | |
3751 | static void _rcu_barrier(struct rcu_state *rsp) | |
3752 | { | |
3753 | int cpu; | |
3754 | struct rcu_data *rdp; | |
3755 | unsigned long s = rcu_seq_snap(&rsp->barrier_sequence); | |
3756 | ||
3757 | _rcu_barrier_trace(rsp, "Begin", -1, s); | |
3758 | ||
3759 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ | |
3760 | mutex_lock(&rsp->barrier_mutex); | |
3761 | ||
3762 | /* Did someone else do our work for us? */ | |
3763 | if (rcu_seq_done(&rsp->barrier_sequence, s)) { | |
3764 | _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence); | |
3765 | smp_mb(); /* caller's subsequent code after above check. */ | |
3766 | mutex_unlock(&rsp->barrier_mutex); | |
3767 | return; | |
3768 | } | |
3769 | ||
3770 | /* Mark the start of the barrier operation. */ | |
3771 | rcu_seq_start(&rsp->barrier_sequence); | |
3772 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence); | |
3773 | ||
3774 | /* | |
3775 | * Initialize the count to one rather than to zero in order to | |
3776 | * avoid a too-soon return to zero in case of a short grace period | |
3777 | * (or preemption of this task). Exclude CPU-hotplug operations | |
3778 | * to ensure that no offline CPU has callbacks queued. | |
3779 | */ | |
3780 | init_completion(&rsp->barrier_completion); | |
3781 | atomic_set(&rsp->barrier_cpu_count, 1); | |
3782 | get_online_cpus(); | |
3783 | ||
3784 | /* | |
3785 | * Force each CPU with callbacks to register a new callback. | |
3786 | * When that callback is invoked, we will know that all of the | |
3787 | * corresponding CPU's preceding callbacks have been invoked. | |
3788 | */ | |
3789 | for_each_possible_cpu(cpu) { | |
3790 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) | |
3791 | continue; | |
3792 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
3793 | if (rcu_is_nocb_cpu(cpu)) { | |
3794 | if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) { | |
3795 | _rcu_barrier_trace(rsp, "OfflineNoCB", cpu, | |
3796 | rsp->barrier_sequence); | |
3797 | } else { | |
3798 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, | |
3799 | rsp->barrier_sequence); | |
3800 | smp_mb__before_atomic(); | |
3801 | atomic_inc(&rsp->barrier_cpu_count); | |
3802 | __call_rcu(&rdp->barrier_head, | |
3803 | rcu_barrier_callback, rsp, cpu, 0); | |
3804 | } | |
3805 | } else if (READ_ONCE(rdp->qlen)) { | |
3806 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, | |
3807 | rsp->barrier_sequence); | |
3808 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); | |
3809 | } else { | |
3810 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, | |
3811 | rsp->barrier_sequence); | |
3812 | } | |
3813 | } | |
3814 | put_online_cpus(); | |
3815 | ||
3816 | /* | |
3817 | * Now that we have an rcu_barrier_callback() callback on each | |
3818 | * CPU, and thus each counted, remove the initial count. | |
3819 | */ | |
3820 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) | |
3821 | complete(&rsp->barrier_completion); | |
3822 | ||
3823 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ | |
3824 | wait_for_completion(&rsp->barrier_completion); | |
3825 | ||
3826 | /* Mark the end of the barrier operation. */ | |
3827 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence); | |
3828 | rcu_seq_end(&rsp->barrier_sequence); | |
3829 | ||
3830 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
3831 | mutex_unlock(&rsp->barrier_mutex); | |
3832 | } | |
3833 | ||
3834 | /** | |
3835 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
3836 | */ | |
3837 | void rcu_barrier_bh(void) | |
3838 | { | |
3839 | _rcu_barrier(&rcu_bh_state); | |
3840 | } | |
3841 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
3842 | ||
3843 | /** | |
3844 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
3845 | */ | |
3846 | void rcu_barrier_sched(void) | |
3847 | { | |
3848 | _rcu_barrier(&rcu_sched_state); | |
3849 | } | |
3850 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
3851 | ||
3852 | /* | |
3853 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the | |
3854 | * first CPU in a given leaf rcu_node structure coming online. The caller | |
3855 | * must hold the corresponding leaf rcu_node ->lock with interrrupts | |
3856 | * disabled. | |
3857 | */ | |
3858 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) | |
3859 | { | |
3860 | long mask; | |
3861 | struct rcu_node *rnp = rnp_leaf; | |
3862 | ||
3863 | for (;;) { | |
3864 | mask = rnp->grpmask; | |
3865 | rnp = rnp->parent; | |
3866 | if (rnp == NULL) | |
3867 | return; | |
3868 | raw_spin_lock(&rnp->lock); /* Interrupts already disabled. */ | |
3869 | rnp->qsmaskinit |= mask; | |
3870 | raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */ | |
3871 | } | |
3872 | } | |
3873 | ||
3874 | /* | |
3875 | * Do boot-time initialization of a CPU's per-CPU RCU data. | |
3876 | */ | |
3877 | static void __init | |
3878 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
3879 | { | |
3880 | static struct lock_class_key rcu_exp_sched_rdp_class; | |
3881 | unsigned long flags; | |
3882 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
3883 | struct rcu_node *rnp = rcu_get_root(rsp); | |
3884 | ||
3885 | /* Set up local state, ensuring consistent view of global state. */ | |
3886 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3887 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); | |
3888 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); | |
3889 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); | |
3890 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); | |
3891 | rdp->cpu = cpu; | |
3892 | rdp->rsp = rsp; | |
3893 | mutex_init(&rdp->exp_funnel_mutex); | |
3894 | rcu_boot_init_nocb_percpu_data(rdp); | |
3895 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3896 | if (rsp == &rcu_sched_state) | |
3897 | lockdep_set_class_and_name(&rdp->exp_funnel_mutex, | |
3898 | &rcu_exp_sched_rdp_class, | |
3899 | "rcu_data_exp_sched"); | |
3900 | } | |
3901 | ||
3902 | /* | |
3903 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
3904 | * offline event can be happening at a given time. Note also that we | |
3905 | * can accept some slop in the rsp->completed access due to the fact | |
3906 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
3907 | */ | |
3908 | static void | |
3909 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp) | |
3910 | { | |
3911 | unsigned long flags; | |
3912 | unsigned long mask; | |
3913 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
3914 | struct rcu_node *rnp = rcu_get_root(rsp); | |
3915 | ||
3916 | /* Set up local state, ensuring consistent view of global state. */ | |
3917 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3918 | rdp->beenonline = 1; /* We have now been online. */ | |
3919 | rdp->qlen_last_fqs_check = 0; | |
3920 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
3921 | rdp->blimit = blimit; | |
3922 | if (!rdp->nxtlist) | |
3923 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ | |
3924 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; | |
3925 | rcu_sysidle_init_percpu_data(rdp->dynticks); | |
3926 | atomic_set(&rdp->dynticks->dynticks, | |
3927 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
3928 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
3929 | ||
3930 | /* | |
3931 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed | |
3932 | * propagation up the rcu_node tree will happen at the beginning | |
3933 | * of the next grace period. | |
3934 | */ | |
3935 | rnp = rdp->mynode; | |
3936 | mask = rdp->grpmask; | |
3937 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
3938 | smp_mb__after_unlock_lock(); | |
3939 | rnp->qsmaskinitnext |= mask; | |
3940 | rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */ | |
3941 | rdp->completed = rnp->completed; | |
3942 | rdp->passed_quiesce = false; | |
3943 | rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu); | |
3944 | rdp->qs_pending = false; | |
3945 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); | |
3946 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3947 | } | |
3948 | ||
3949 | static void rcu_prepare_cpu(int cpu) | |
3950 | { | |
3951 | struct rcu_state *rsp; | |
3952 | ||
3953 | for_each_rcu_flavor(rsp) | |
3954 | rcu_init_percpu_data(cpu, rsp); | |
3955 | } | |
3956 | ||
3957 | /* | |
3958 | * Handle CPU online/offline notification events. | |
3959 | */ | |
3960 | int rcu_cpu_notify(struct notifier_block *self, | |
3961 | unsigned long action, void *hcpu) | |
3962 | { | |
3963 | long cpu = (long)hcpu; | |
3964 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); | |
3965 | struct rcu_node *rnp = rdp->mynode; | |
3966 | struct rcu_state *rsp; | |
3967 | ||
3968 | switch (action) { | |
3969 | case CPU_UP_PREPARE: | |
3970 | case CPU_UP_PREPARE_FROZEN: | |
3971 | rcu_prepare_cpu(cpu); | |
3972 | rcu_prepare_kthreads(cpu); | |
3973 | rcu_spawn_all_nocb_kthreads(cpu); | |
3974 | break; | |
3975 | case CPU_ONLINE: | |
3976 | case CPU_DOWN_FAILED: | |
3977 | rcu_boost_kthread_setaffinity(rnp, -1); | |
3978 | break; | |
3979 | case CPU_DOWN_PREPARE: | |
3980 | rcu_boost_kthread_setaffinity(rnp, cpu); | |
3981 | break; | |
3982 | case CPU_DYING: | |
3983 | case CPU_DYING_FROZEN: | |
3984 | for_each_rcu_flavor(rsp) | |
3985 | rcu_cleanup_dying_cpu(rsp); | |
3986 | break; | |
3987 | case CPU_DYING_IDLE: | |
3988 | for_each_rcu_flavor(rsp) { | |
3989 | rcu_cleanup_dying_idle_cpu(cpu, rsp); | |
3990 | } | |
3991 | break; | |
3992 | case CPU_DEAD: | |
3993 | case CPU_DEAD_FROZEN: | |
3994 | case CPU_UP_CANCELED: | |
3995 | case CPU_UP_CANCELED_FROZEN: | |
3996 | for_each_rcu_flavor(rsp) { | |
3997 | rcu_cleanup_dead_cpu(cpu, rsp); | |
3998 | do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu)); | |
3999 | } | |
4000 | break; | |
4001 | default: | |
4002 | break; | |
4003 | } | |
4004 | return NOTIFY_OK; | |
4005 | } | |
4006 | ||
4007 | static int rcu_pm_notify(struct notifier_block *self, | |
4008 | unsigned long action, void *hcpu) | |
4009 | { | |
4010 | switch (action) { | |
4011 | case PM_HIBERNATION_PREPARE: | |
4012 | case PM_SUSPEND_PREPARE: | |
4013 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
4014 | rcu_expedite_gp(); | |
4015 | break; | |
4016 | case PM_POST_HIBERNATION: | |
4017 | case PM_POST_SUSPEND: | |
4018 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
4019 | rcu_unexpedite_gp(); | |
4020 | break; | |
4021 | default: | |
4022 | break; | |
4023 | } | |
4024 | return NOTIFY_OK; | |
4025 | } | |
4026 | ||
4027 | /* | |
4028 | * Spawn the kthreads that handle each RCU flavor's grace periods. | |
4029 | */ | |
4030 | static int __init rcu_spawn_gp_kthread(void) | |
4031 | { | |
4032 | unsigned long flags; | |
4033 | int kthread_prio_in = kthread_prio; | |
4034 | struct rcu_node *rnp; | |
4035 | struct rcu_state *rsp; | |
4036 | struct sched_param sp; | |
4037 | struct task_struct *t; | |
4038 | ||
4039 | /* Force priority into range. */ | |
4040 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) | |
4041 | kthread_prio = 1; | |
4042 | else if (kthread_prio < 0) | |
4043 | kthread_prio = 0; | |
4044 | else if (kthread_prio > 99) | |
4045 | kthread_prio = 99; | |
4046 | if (kthread_prio != kthread_prio_in) | |
4047 | pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n", | |
4048 | kthread_prio, kthread_prio_in); | |
4049 | ||
4050 | rcu_scheduler_fully_active = 1; | |
4051 | for_each_rcu_flavor(rsp) { | |
4052 | t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name); | |
4053 | BUG_ON(IS_ERR(t)); | |
4054 | rnp = rcu_get_root(rsp); | |
4055 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
4056 | rsp->gp_kthread = t; | |
4057 | if (kthread_prio) { | |
4058 | sp.sched_priority = kthread_prio; | |
4059 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
4060 | } | |
4061 | wake_up_process(t); | |
4062 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
4063 | } | |
4064 | rcu_spawn_nocb_kthreads(); | |
4065 | rcu_spawn_boost_kthreads(); | |
4066 | return 0; | |
4067 | } | |
4068 | early_initcall(rcu_spawn_gp_kthread); | |
4069 | ||
4070 | /* | |
4071 | * This function is invoked towards the end of the scheduler's initialization | |
4072 | * process. Before this is called, the idle task might contain | |
4073 | * RCU read-side critical sections (during which time, this idle | |
4074 | * task is booting the system). After this function is called, the | |
4075 | * idle tasks are prohibited from containing RCU read-side critical | |
4076 | * sections. This function also enables RCU lockdep checking. | |
4077 | */ | |
4078 | void rcu_scheduler_starting(void) | |
4079 | { | |
4080 | WARN_ON(num_online_cpus() != 1); | |
4081 | WARN_ON(nr_context_switches() > 0); | |
4082 | rcu_scheduler_active = 1; | |
4083 | } | |
4084 | ||
4085 | /* | |
4086 | * Compute the per-level fanout, either using the exact fanout specified | |
4087 | * or balancing the tree, depending on the rcu_fanout_exact boot parameter. | |
4088 | */ | |
4089 | static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt) | |
4090 | { | |
4091 | int i; | |
4092 | ||
4093 | if (rcu_fanout_exact) { | |
4094 | levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; | |
4095 | for (i = rcu_num_lvls - 2; i >= 0; i--) | |
4096 | levelspread[i] = RCU_FANOUT; | |
4097 | } else { | |
4098 | int ccur; | |
4099 | int cprv; | |
4100 | ||
4101 | cprv = nr_cpu_ids; | |
4102 | for (i = rcu_num_lvls - 1; i >= 0; i--) { | |
4103 | ccur = levelcnt[i]; | |
4104 | levelspread[i] = (cprv + ccur - 1) / ccur; | |
4105 | cprv = ccur; | |
4106 | } | |
4107 | } | |
4108 | } | |
4109 | ||
4110 | /* | |
4111 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
4112 | */ | |
4113 | static void __init rcu_init_one(struct rcu_state *rsp, | |
4114 | struct rcu_data __percpu *rda) | |
4115 | { | |
4116 | static const char * const buf[] = RCU_NODE_NAME_INIT; | |
4117 | static const char * const fqs[] = RCU_FQS_NAME_INIT; | |
4118 | static const char * const exp[] = RCU_EXP_NAME_INIT; | |
4119 | static const char * const exp_sched[] = RCU_EXP_SCHED_NAME_INIT; | |
4120 | static u8 fl_mask = 0x1; | |
4121 | ||
4122 | int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */ | |
4123 | int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */ | |
4124 | int cpustride = 1; | |
4125 | int i; | |
4126 | int j; | |
4127 | struct rcu_node *rnp; | |
4128 | ||
4129 | BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ | |
4130 | ||
4131 | /* Silence gcc 4.8 false positive about array index out of range. */ | |
4132 | if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS) | |
4133 | panic("rcu_init_one: rcu_num_lvls out of range"); | |
4134 | ||
4135 | /* Initialize the level-tracking arrays. */ | |
4136 | ||
4137 | for (i = 0; i < rcu_num_lvls; i++) | |
4138 | levelcnt[i] = num_rcu_lvl[i]; | |
4139 | for (i = 1; i < rcu_num_lvls; i++) | |
4140 | rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1]; | |
4141 | rcu_init_levelspread(levelspread, levelcnt); | |
4142 | rsp->flavor_mask = fl_mask; | |
4143 | fl_mask <<= 1; | |
4144 | ||
4145 | /* Initialize the elements themselves, starting from the leaves. */ | |
4146 | ||
4147 | for (i = rcu_num_lvls - 1; i >= 0; i--) { | |
4148 | cpustride *= levelspread[i]; | |
4149 | rnp = rsp->level[i]; | |
4150 | for (j = 0; j < levelcnt[i]; j++, rnp++) { | |
4151 | raw_spin_lock_init(&rnp->lock); | |
4152 | lockdep_set_class_and_name(&rnp->lock, | |
4153 | &rcu_node_class[i], buf[i]); | |
4154 | raw_spin_lock_init(&rnp->fqslock); | |
4155 | lockdep_set_class_and_name(&rnp->fqslock, | |
4156 | &rcu_fqs_class[i], fqs[i]); | |
4157 | rnp->gpnum = rsp->gpnum; | |
4158 | rnp->completed = rsp->completed; | |
4159 | rnp->qsmask = 0; | |
4160 | rnp->qsmaskinit = 0; | |
4161 | rnp->grplo = j * cpustride; | |
4162 | rnp->grphi = (j + 1) * cpustride - 1; | |
4163 | if (rnp->grphi >= nr_cpu_ids) | |
4164 | rnp->grphi = nr_cpu_ids - 1; | |
4165 | if (i == 0) { | |
4166 | rnp->grpnum = 0; | |
4167 | rnp->grpmask = 0; | |
4168 | rnp->parent = NULL; | |
4169 | } else { | |
4170 | rnp->grpnum = j % levelspread[i - 1]; | |
4171 | rnp->grpmask = 1UL << rnp->grpnum; | |
4172 | rnp->parent = rsp->level[i - 1] + | |
4173 | j / levelspread[i - 1]; | |
4174 | } | |
4175 | rnp->level = i; | |
4176 | INIT_LIST_HEAD(&rnp->blkd_tasks); | |
4177 | rcu_init_one_nocb(rnp); | |
4178 | mutex_init(&rnp->exp_funnel_mutex); | |
4179 | if (rsp == &rcu_sched_state) | |
4180 | lockdep_set_class_and_name( | |
4181 | &rnp->exp_funnel_mutex, | |
4182 | &rcu_exp_sched_class[i], exp_sched[i]); | |
4183 | else | |
4184 | lockdep_set_class_and_name( | |
4185 | &rnp->exp_funnel_mutex, | |
4186 | &rcu_exp_class[i], exp[i]); | |
4187 | } | |
4188 | } | |
4189 | ||
4190 | init_waitqueue_head(&rsp->gp_wq); | |
4191 | rnp = rsp->level[rcu_num_lvls - 1]; | |
4192 | for_each_possible_cpu(i) { | |
4193 | while (i > rnp->grphi) | |
4194 | rnp++; | |
4195 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; | |
4196 | rcu_boot_init_percpu_data(i, rsp); | |
4197 | } | |
4198 | list_add(&rsp->flavors, &rcu_struct_flavors); | |
4199 | } | |
4200 | ||
4201 | /* | |
4202 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4203 | * replace the definitions in tree.h because those are needed to size | |
4204 | * the ->node array in the rcu_state structure. | |
4205 | */ | |
4206 | static void __init rcu_init_geometry(void) | |
4207 | { | |
4208 | ulong d; | |
4209 | int i; | |
4210 | int rcu_capacity[RCU_NUM_LVLS]; | |
4211 | ||
4212 | /* | |
4213 | * Initialize any unspecified boot parameters. | |
4214 | * The default values of jiffies_till_first_fqs and | |
4215 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
4216 | * value, which is a function of HZ, then adding one for each | |
4217 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
4218 | */ | |
4219 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
4220 | if (jiffies_till_first_fqs == ULONG_MAX) | |
4221 | jiffies_till_first_fqs = d; | |
4222 | if (jiffies_till_next_fqs == ULONG_MAX) | |
4223 | jiffies_till_next_fqs = d; | |
4224 | ||
4225 | /* If the compile-time values are accurate, just leave. */ | |
4226 | if (rcu_fanout_leaf == RCU_FANOUT_LEAF && | |
4227 | nr_cpu_ids == NR_CPUS) | |
4228 | return; | |
4229 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", | |
4230 | rcu_fanout_leaf, nr_cpu_ids); | |
4231 | ||
4232 | /* | |
4233 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
4234 | * to increase the leaf-level fanout, not decrease it. Of course, | |
4235 | * the leaf-level fanout cannot exceed the number of bits in | |
4236 | * the rcu_node masks. Complain and fall back to the compile- | |
4237 | * time values if these limits are exceeded. | |
4238 | */ | |
4239 | if (rcu_fanout_leaf < RCU_FANOUT_LEAF || | |
4240 | rcu_fanout_leaf > sizeof(unsigned long) * 8) { | |
4241 | rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
4242 | WARN_ON(1); | |
4243 | return; | |
4244 | } | |
4245 | ||
4246 | /* | |
4247 | * Compute number of nodes that can be handled an rcu_node tree | |
4248 | * with the given number of levels. | |
4249 | */ | |
4250 | rcu_capacity[0] = rcu_fanout_leaf; | |
4251 | for (i = 1; i < RCU_NUM_LVLS; i++) | |
4252 | rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT; | |
4253 | ||
4254 | /* | |
4255 | * The tree must be able to accommodate the configured number of CPUs. | |
4256 | * If this limit is exceeded than we have a serious problem elsewhere. | |
4257 | */ | |
4258 | if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) | |
4259 | panic("rcu_init_geometry: rcu_capacity[] is too small"); | |
4260 | ||
4261 | /* Calculate the number of levels in the tree. */ | |
4262 | for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) { | |
4263 | } | |
4264 | rcu_num_lvls = i + 1; | |
4265 | ||
4266 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
4267 | for (i = 0; i < rcu_num_lvls; i++) { | |
4268 | int cap = rcu_capacity[(rcu_num_lvls - 1) - i]; | |
4269 | num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap); | |
4270 | } | |
4271 | ||
4272 | /* Calculate the total number of rcu_node structures. */ | |
4273 | rcu_num_nodes = 0; | |
4274 | for (i = 0; i < rcu_num_lvls; i++) | |
4275 | rcu_num_nodes += num_rcu_lvl[i]; | |
4276 | } | |
4277 | ||
4278 | /* | |
4279 | * Dump out the structure of the rcu_node combining tree associated | |
4280 | * with the rcu_state structure referenced by rsp. | |
4281 | */ | |
4282 | static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp) | |
4283 | { | |
4284 | int level = 0; | |
4285 | struct rcu_node *rnp; | |
4286 | ||
4287 | pr_info("rcu_node tree layout dump\n"); | |
4288 | pr_info(" "); | |
4289 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
4290 | if (rnp->level != level) { | |
4291 | pr_cont("\n"); | |
4292 | pr_info(" "); | |
4293 | level = rnp->level; | |
4294 | } | |
4295 | pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum); | |
4296 | } | |
4297 | pr_cont("\n"); | |
4298 | } | |
4299 | ||
4300 | void __init rcu_init(void) | |
4301 | { | |
4302 | int cpu; | |
4303 | ||
4304 | rcu_early_boot_tests(); | |
4305 | ||
4306 | rcu_bootup_announce(); | |
4307 | rcu_init_geometry(); | |
4308 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); | |
4309 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); | |
4310 | if (dump_tree) | |
4311 | rcu_dump_rcu_node_tree(&rcu_sched_state); | |
4312 | __rcu_init_preempt(); | |
4313 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); | |
4314 | ||
4315 | /* | |
4316 | * We don't need protection against CPU-hotplug here because | |
4317 | * this is called early in boot, before either interrupts | |
4318 | * or the scheduler are operational. | |
4319 | */ | |
4320 | cpu_notifier(rcu_cpu_notify, 0); | |
4321 | pm_notifier(rcu_pm_notify, 0); | |
4322 | for_each_online_cpu(cpu) | |
4323 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
4324 | } | |
4325 | ||
4326 | #include "tree_plugin.h" |