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