]>
Commit | Line | Data |
---|---|---|
64db4cff PM |
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 | |
87de1cfd PM |
15 | * along with this program; if not, you can access it online at |
16 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
64db4cff PM |
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 - | |
a71fca58 | 28 | * Documentation/RCU |
64db4cff PM |
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> | |
c1dc0b9c | 38 | #include <linux/nmi.h> |
8826f3b0 | 39 | #include <linux/atomic.h> |
64db4cff | 40 | #include <linux/bitops.h> |
9984de1a | 41 | #include <linux/export.h> |
64db4cff PM |
42 | #include <linux/completion.h> |
43 | #include <linux/moduleparam.h> | |
4102adab | 44 | #include <linux/module.h> |
64db4cff PM |
45 | #include <linux/percpu.h> |
46 | #include <linux/notifier.h> | |
47 | #include <linux/cpu.h> | |
48 | #include <linux/mutex.h> | |
49 | #include <linux/time.h> | |
bbad9379 | 50 | #include <linux/kernel_stat.h> |
a26ac245 PM |
51 | #include <linux/wait.h> |
52 | #include <linux/kthread.h> | |
268bb0ce | 53 | #include <linux/prefetch.h> |
3d3b7db0 PM |
54 | #include <linux/delay.h> |
55 | #include <linux/stop_machine.h> | |
661a85dc | 56 | #include <linux/random.h> |
af658dca | 57 | #include <linux/trace_events.h> |
d1d74d14 | 58 | #include <linux/suspend.h> |
64db4cff | 59 | |
4102adab | 60 | #include "tree.h" |
29c00b4a | 61 | #include "rcu.h" |
9f77da9f | 62 | |
4102adab PM |
63 | MODULE_ALIAS("rcutree"); |
64 | #ifdef MODULE_PARAM_PREFIX | |
65 | #undef MODULE_PARAM_PREFIX | |
66 | #endif | |
67 | #define MODULE_PARAM_PREFIX "rcutree." | |
68 | ||
64db4cff PM |
69 | /* Data structures. */ |
70 | ||
f885b7f2 | 71 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
394f2769 | 72 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; |
88b91c7c | 73 | |
f7f7bac9 SRRH |
74 | /* |
75 | * In order to export the rcu_state name to the tracing tools, it | |
76 | * needs to be added in the __tracepoint_string section. | |
77 | * This requires defining a separate variable tp_<sname>_varname | |
78 | * that points to the string being used, and this will allow | |
79 | * the tracing userspace tools to be able to decipher the string | |
80 | * address to the matching string. | |
81 | */ | |
a8a29b3b AB |
82 | #ifdef CONFIG_TRACING |
83 | # define DEFINE_RCU_TPS(sname) \ | |
f7f7bac9 | 84 | static char sname##_varname[] = #sname; \ |
a8a29b3b AB |
85 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; |
86 | # define RCU_STATE_NAME(sname) sname##_varname | |
87 | #else | |
88 | # define DEFINE_RCU_TPS(sname) | |
89 | # define RCU_STATE_NAME(sname) __stringify(sname) | |
90 | #endif | |
91 | ||
92 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ | |
93 | DEFINE_RCU_TPS(sname) \ | |
c92fb057 | 94 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \ |
a41bfeb2 | 95 | struct rcu_state sname##_state = { \ |
6c90cc7b | 96 | .level = { &sname##_state.node[0] }, \ |
2723249a | 97 | .rda = &sname##_data, \ |
037b64ed | 98 | .call = cr, \ |
af446b70 | 99 | .fqs_state = RCU_GP_IDLE, \ |
42c3533e PM |
100 | .gpnum = 0UL - 300UL, \ |
101 | .completed = 0UL - 300UL, \ | |
7b2e6011 | 102 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
6c90cc7b PM |
103 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
104 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 105 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
a8a29b3b | 106 | .name = RCU_STATE_NAME(sname), \ |
a4889858 | 107 | .abbr = sabbr, \ |
2723249a | 108 | } |
64db4cff | 109 | |
a41bfeb2 SRRH |
110 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); |
111 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); | |
b1f77b05 | 112 | |
b28a7c01 | 113 | static struct rcu_state *const rcu_state_p; |
2927a689 | 114 | static struct rcu_data __percpu *const rcu_data_p; |
6ce75a23 | 115 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 116 | |
a3dc2948 PM |
117 | /* Dump rcu_node combining tree at boot to verify correct setup. */ |
118 | static bool dump_tree; | |
119 | module_param(dump_tree, bool, 0444); | |
7fa27001 PM |
120 | /* Control rcu_node-tree auto-balancing at boot time. */ |
121 | static bool rcu_fanout_exact; | |
122 | module_param(rcu_fanout_exact, bool, 0444); | |
47d631af PM |
123 | /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */ |
124 | static int rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
7e5c2dfb | 125 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 | 126 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
cb007102 AG |
127 | /* Number of rcu_nodes at specified level. */ |
128 | static int num_rcu_lvl[] = NUM_RCU_LVL_INIT; | |
f885b7f2 PM |
129 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ |
130 | ||
b0d30417 PM |
131 | /* |
132 | * The rcu_scheduler_active variable transitions from zero to one just | |
133 | * before the first task is spawned. So when this variable is zero, RCU | |
134 | * can assume that there is but one task, allowing RCU to (for example) | |
b44f6656 | 135 | * optimize synchronize_sched() to a simple barrier(). When this variable |
b0d30417 PM |
136 | * is one, RCU must actually do all the hard work required to detect real |
137 | * grace periods. This variable is also used to suppress boot-time false | |
138 | * positives from lockdep-RCU error checking. | |
139 | */ | |
bbad9379 PM |
140 | int rcu_scheduler_active __read_mostly; |
141 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
142 | ||
b0d30417 PM |
143 | /* |
144 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
145 | * during the early_initcall() processing, which is after the scheduler | |
146 | * is capable of creating new tasks. So RCU processing (for example, | |
147 | * creating tasks for RCU priority boosting) must be delayed until after | |
148 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
149 | * currently delay invocation of any RCU callbacks until after this point. | |
150 | * | |
151 | * It might later prove better for people registering RCU callbacks during | |
152 | * early boot to take responsibility for these callbacks, but one step at | |
153 | * a time. | |
154 | */ | |
155 | static int rcu_scheduler_fully_active __read_mostly; | |
156 | ||
0aa04b05 PM |
157 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); |
158 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); | |
5d01bbd1 | 159 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
160 | static void invoke_rcu_core(void); |
161 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 162 | |
a94844b2 | 163 | /* rcuc/rcub kthread realtime priority */ |
26730f55 | 164 | #ifdef CONFIG_RCU_KTHREAD_PRIO |
a94844b2 | 165 | static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO; |
26730f55 PM |
166 | #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */ |
167 | static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0; | |
168 | #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */ | |
a94844b2 PM |
169 | module_param(kthread_prio, int, 0644); |
170 | ||
8d7dc928 | 171 | /* Delay in jiffies for grace-period initialization delays, debug only. */ |
0f41c0dd PM |
172 | |
173 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT | |
174 | static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY; | |
175 | module_param(gp_preinit_delay, int, 0644); | |
176 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ | |
177 | static const int gp_preinit_delay; | |
178 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ | |
179 | ||
8d7dc928 PM |
180 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT |
181 | static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY; | |
37745d28 | 182 | module_param(gp_init_delay, int, 0644); |
8d7dc928 PM |
183 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ |
184 | static const int gp_init_delay; | |
185 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ | |
eab128e8 | 186 | |
0f41c0dd PM |
187 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP |
188 | static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY; | |
189 | module_param(gp_cleanup_delay, int, 0644); | |
190 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ | |
191 | static const int gp_cleanup_delay; | |
192 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ | |
193 | ||
eab128e8 PM |
194 | /* |
195 | * Number of grace periods between delays, normalized by the duration of | |
196 | * the delay. The longer the the delay, the more the grace periods between | |
197 | * each delay. The reason for this normalization is that it means that, | |
198 | * for non-zero delays, the overall slowdown of grace periods is constant | |
199 | * regardless of the duration of the delay. This arrangement balances | |
200 | * the need for long delays to increase some race probabilities with the | |
201 | * need for fast grace periods to increase other race probabilities. | |
202 | */ | |
203 | #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */ | |
37745d28 | 204 | |
4a298656 PM |
205 | /* |
206 | * Track the rcutorture test sequence number and the update version | |
207 | * number within a given test. The rcutorture_testseq is incremented | |
208 | * on every rcutorture module load and unload, so has an odd value | |
209 | * when a test is running. The rcutorture_vernum is set to zero | |
210 | * when rcutorture starts and is incremented on each rcutorture update. | |
211 | * These variables enable correlating rcutorture output with the | |
212 | * RCU tracing information. | |
213 | */ | |
214 | unsigned long rcutorture_testseq; | |
215 | unsigned long rcutorture_vernum; | |
216 | ||
0aa04b05 PM |
217 | /* |
218 | * Compute the mask of online CPUs for the specified rcu_node structure. | |
219 | * This will not be stable unless the rcu_node structure's ->lock is | |
220 | * held, but the bit corresponding to the current CPU will be stable | |
221 | * in most contexts. | |
222 | */ | |
223 | unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) | |
224 | { | |
7d0ae808 | 225 | return READ_ONCE(rnp->qsmaskinitnext); |
0aa04b05 PM |
226 | } |
227 | ||
fc2219d4 | 228 | /* |
7d0ae808 | 229 | * Return true if an RCU grace period is in progress. The READ_ONCE()s |
fc2219d4 PM |
230 | * permit this function to be invoked without holding the root rcu_node |
231 | * structure's ->lock, but of course results can be subject to change. | |
232 | */ | |
233 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
234 | { | |
7d0ae808 | 235 | return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum); |
fc2219d4 PM |
236 | } |
237 | ||
b1f77b05 | 238 | /* |
d6714c22 | 239 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 240 | * how many quiescent states passed, just if there was at least |
d6714c22 | 241 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 242 | * The caller must have disabled preemption. |
b1f77b05 | 243 | */ |
284a8c93 | 244 | void rcu_sched_qs(void) |
b1f77b05 | 245 | { |
284a8c93 PM |
246 | if (!__this_cpu_read(rcu_sched_data.passed_quiesce)) { |
247 | trace_rcu_grace_period(TPS("rcu_sched"), | |
248 | __this_cpu_read(rcu_sched_data.gpnum), | |
249 | TPS("cpuqs")); | |
250 | __this_cpu_write(rcu_sched_data.passed_quiesce, 1); | |
251 | } | |
b1f77b05 IM |
252 | } |
253 | ||
284a8c93 | 254 | void rcu_bh_qs(void) |
b1f77b05 | 255 | { |
284a8c93 PM |
256 | if (!__this_cpu_read(rcu_bh_data.passed_quiesce)) { |
257 | trace_rcu_grace_period(TPS("rcu_bh"), | |
258 | __this_cpu_read(rcu_bh_data.gpnum), | |
259 | TPS("cpuqs")); | |
260 | __this_cpu_write(rcu_bh_data.passed_quiesce, 1); | |
261 | } | |
b1f77b05 | 262 | } |
64db4cff | 263 | |
4a81e832 PM |
264 | static DEFINE_PER_CPU(int, rcu_sched_qs_mask); |
265 | ||
266 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { | |
267 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, | |
268 | .dynticks = ATOMIC_INIT(1), | |
269 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
270 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
271 | .dynticks_idle = ATOMIC_INIT(1), | |
272 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
273 | }; | |
274 | ||
5cd37193 PM |
275 | DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr); |
276 | EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr); | |
277 | ||
4a81e832 PM |
278 | /* |
279 | * Let the RCU core know that this CPU has gone through the scheduler, | |
280 | * which is a quiescent state. This is called when the need for a | |
281 | * quiescent state is urgent, so we burn an atomic operation and full | |
282 | * memory barriers to let the RCU core know about it, regardless of what | |
283 | * this CPU might (or might not) do in the near future. | |
284 | * | |
285 | * We inform the RCU core by emulating a zero-duration dyntick-idle | |
286 | * period, which we in turn do by incrementing the ->dynticks counter | |
287 | * by two. | |
288 | */ | |
289 | static void rcu_momentary_dyntick_idle(void) | |
290 | { | |
291 | unsigned long flags; | |
292 | struct rcu_data *rdp; | |
293 | struct rcu_dynticks *rdtp; | |
294 | int resched_mask; | |
295 | struct rcu_state *rsp; | |
296 | ||
297 | local_irq_save(flags); | |
298 | ||
299 | /* | |
300 | * Yes, we can lose flag-setting operations. This is OK, because | |
301 | * the flag will be set again after some delay. | |
302 | */ | |
303 | resched_mask = raw_cpu_read(rcu_sched_qs_mask); | |
304 | raw_cpu_write(rcu_sched_qs_mask, 0); | |
305 | ||
306 | /* Find the flavor that needs a quiescent state. */ | |
307 | for_each_rcu_flavor(rsp) { | |
308 | rdp = raw_cpu_ptr(rsp->rda); | |
309 | if (!(resched_mask & rsp->flavor_mask)) | |
310 | continue; | |
311 | smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */ | |
7d0ae808 PM |
312 | if (READ_ONCE(rdp->mynode->completed) != |
313 | READ_ONCE(rdp->cond_resched_completed)) | |
4a81e832 PM |
314 | continue; |
315 | ||
316 | /* | |
317 | * Pretend to be momentarily idle for the quiescent state. | |
318 | * This allows the grace-period kthread to record the | |
319 | * quiescent state, with no need for this CPU to do anything | |
320 | * further. | |
321 | */ | |
322 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
323 | smp_mb__before_atomic(); /* Earlier stuff before QS. */ | |
324 | atomic_add(2, &rdtp->dynticks); /* QS. */ | |
325 | smp_mb__after_atomic(); /* Later stuff after QS. */ | |
326 | break; | |
327 | } | |
328 | local_irq_restore(flags); | |
329 | } | |
330 | ||
25502a6c PM |
331 | /* |
332 | * Note a context switch. This is a quiescent state for RCU-sched, | |
333 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 334 | * The caller must have disabled preemption. |
25502a6c | 335 | */ |
38200cf2 | 336 | void rcu_note_context_switch(void) |
25502a6c | 337 | { |
f7f7bac9 | 338 | trace_rcu_utilization(TPS("Start context switch")); |
284a8c93 | 339 | rcu_sched_qs(); |
38200cf2 | 340 | rcu_preempt_note_context_switch(); |
4a81e832 PM |
341 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) |
342 | rcu_momentary_dyntick_idle(); | |
f7f7bac9 | 343 | trace_rcu_utilization(TPS("End context switch")); |
25502a6c | 344 | } |
29ce8310 | 345 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 346 | |
5cd37193 | 347 | /* |
1925d196 | 348 | * Register a quiescent state for all RCU flavors. If there is an |
5cd37193 PM |
349 | * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight |
350 | * dyntick-idle quiescent state visible to other CPUs (but only for those | |
1925d196 | 351 | * RCU flavors in desperate need of a quiescent state, which will normally |
5cd37193 PM |
352 | * be none of them). Either way, do a lightweight quiescent state for |
353 | * all RCU flavors. | |
354 | */ | |
355 | void rcu_all_qs(void) | |
356 | { | |
357 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) | |
358 | rcu_momentary_dyntick_idle(); | |
359 | this_cpu_inc(rcu_qs_ctr); | |
360 | } | |
361 | EXPORT_SYMBOL_GPL(rcu_all_qs); | |
362 | ||
878d7439 ED |
363 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
364 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
365 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 366 | |
878d7439 ED |
367 | module_param(blimit, long, 0444); |
368 | module_param(qhimark, long, 0444); | |
369 | module_param(qlowmark, long, 0444); | |
3d76c082 | 370 | |
026ad283 PM |
371 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
372 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
373 | |
374 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
375 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
376 | ||
4a81e832 PM |
377 | /* |
378 | * How long the grace period must be before we start recruiting | |
379 | * quiescent-state help from rcu_note_context_switch(). | |
380 | */ | |
381 | static ulong jiffies_till_sched_qs = HZ / 20; | |
382 | module_param(jiffies_till_sched_qs, ulong, 0644); | |
383 | ||
48a7639c | 384 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
910ee45d | 385 | struct rcu_data *rdp); |
217af2a2 PM |
386 | static void force_qs_rnp(struct rcu_state *rsp, |
387 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
388 | unsigned long *maxj), | |
389 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 390 | static void force_quiescent_state(struct rcu_state *rsp); |
e3950ecd | 391 | static int rcu_pending(void); |
64db4cff PM |
392 | |
393 | /* | |
917963d0 | 394 | * Return the number of RCU batches started thus far for debug & stats. |
64db4cff | 395 | */ |
917963d0 PM |
396 | unsigned long rcu_batches_started(void) |
397 | { | |
398 | return rcu_state_p->gpnum; | |
399 | } | |
400 | EXPORT_SYMBOL_GPL(rcu_batches_started); | |
401 | ||
402 | /* | |
403 | * Return the number of RCU-sched batches started thus far for debug & stats. | |
64db4cff | 404 | */ |
917963d0 PM |
405 | unsigned long rcu_batches_started_sched(void) |
406 | { | |
407 | return rcu_sched_state.gpnum; | |
408 | } | |
409 | EXPORT_SYMBOL_GPL(rcu_batches_started_sched); | |
410 | ||
411 | /* | |
412 | * Return the number of RCU BH batches started thus far for debug & stats. | |
413 | */ | |
414 | unsigned long rcu_batches_started_bh(void) | |
415 | { | |
416 | return rcu_bh_state.gpnum; | |
417 | } | |
418 | EXPORT_SYMBOL_GPL(rcu_batches_started_bh); | |
419 | ||
420 | /* | |
421 | * Return the number of RCU batches completed thus far for debug & stats. | |
422 | */ | |
423 | unsigned long rcu_batches_completed(void) | |
424 | { | |
425 | return rcu_state_p->completed; | |
426 | } | |
427 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
428 | ||
429 | /* | |
430 | * Return the number of RCU-sched batches completed thus far for debug & stats. | |
64db4cff | 431 | */ |
9733e4f0 | 432 | unsigned long rcu_batches_completed_sched(void) |
64db4cff | 433 | { |
d6714c22 | 434 | return rcu_sched_state.completed; |
64db4cff | 435 | } |
d6714c22 | 436 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
437 | |
438 | /* | |
917963d0 | 439 | * Return the number of RCU BH batches completed thus far for debug & stats. |
64db4cff | 440 | */ |
9733e4f0 | 441 | unsigned long rcu_batches_completed_bh(void) |
64db4cff PM |
442 | { |
443 | return rcu_bh_state.completed; | |
444 | } | |
445 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
446 | ||
a381d757 ACB |
447 | /* |
448 | * Force a quiescent state. | |
449 | */ | |
450 | void rcu_force_quiescent_state(void) | |
451 | { | |
e534165b | 452 | force_quiescent_state(rcu_state_p); |
a381d757 ACB |
453 | } |
454 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
455 | ||
bf66f18e PM |
456 | /* |
457 | * Force a quiescent state for RCU BH. | |
458 | */ | |
459 | void rcu_bh_force_quiescent_state(void) | |
460 | { | |
4cdfc175 | 461 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
462 | } |
463 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
464 | ||
e7580f33 PM |
465 | /* |
466 | * Force a quiescent state for RCU-sched. | |
467 | */ | |
468 | void rcu_sched_force_quiescent_state(void) | |
469 | { | |
470 | force_quiescent_state(&rcu_sched_state); | |
471 | } | |
472 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
473 | ||
afea227f PM |
474 | /* |
475 | * Show the state of the grace-period kthreads. | |
476 | */ | |
477 | void show_rcu_gp_kthreads(void) | |
478 | { | |
479 | struct rcu_state *rsp; | |
480 | ||
481 | for_each_rcu_flavor(rsp) { | |
482 | pr_info("%s: wait state: %d ->state: %#lx\n", | |
483 | rsp->name, rsp->gp_state, rsp->gp_kthread->state); | |
484 | /* sched_show_task(rsp->gp_kthread); */ | |
485 | } | |
486 | } | |
487 | EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); | |
488 | ||
4a298656 PM |
489 | /* |
490 | * Record the number of times rcutorture tests have been initiated and | |
491 | * terminated. This information allows the debugfs tracing stats to be | |
492 | * correlated to the rcutorture messages, even when the rcutorture module | |
493 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
494 | * store this state in rcutorture itself. | |
495 | */ | |
496 | void rcutorture_record_test_transition(void) | |
497 | { | |
498 | rcutorture_testseq++; | |
499 | rcutorture_vernum = 0; | |
500 | } | |
501 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
502 | ||
ad0dc7f9 PM |
503 | /* |
504 | * Send along grace-period-related data for rcutorture diagnostics. | |
505 | */ | |
506 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
507 | unsigned long *gpnum, unsigned long *completed) | |
508 | { | |
509 | struct rcu_state *rsp = NULL; | |
510 | ||
511 | switch (test_type) { | |
512 | case RCU_FLAVOR: | |
e534165b | 513 | rsp = rcu_state_p; |
ad0dc7f9 PM |
514 | break; |
515 | case RCU_BH_FLAVOR: | |
516 | rsp = &rcu_bh_state; | |
517 | break; | |
518 | case RCU_SCHED_FLAVOR: | |
519 | rsp = &rcu_sched_state; | |
520 | break; | |
521 | default: | |
522 | break; | |
523 | } | |
524 | if (rsp != NULL) { | |
7d0ae808 PM |
525 | *flags = READ_ONCE(rsp->gp_flags); |
526 | *gpnum = READ_ONCE(rsp->gpnum); | |
527 | *completed = READ_ONCE(rsp->completed); | |
ad0dc7f9 PM |
528 | return; |
529 | } | |
530 | *flags = 0; | |
531 | *gpnum = 0; | |
532 | *completed = 0; | |
533 | } | |
534 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
535 | ||
4a298656 PM |
536 | /* |
537 | * Record the number of writer passes through the current rcutorture test. | |
538 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
539 | * messages. | |
540 | */ | |
541 | void rcutorture_record_progress(unsigned long vernum) | |
542 | { | |
543 | rcutorture_vernum++; | |
544 | } | |
545 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
546 | ||
64db4cff PM |
547 | /* |
548 | * Does the CPU have callbacks ready to be invoked? | |
549 | */ | |
550 | static int | |
551 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
552 | { | |
3fbfbf7a PM |
553 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
554 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
555 | } |
556 | ||
365187fb PM |
557 | /* |
558 | * Return the root node of the specified rcu_state structure. | |
559 | */ | |
560 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
561 | { | |
562 | return &rsp->node[0]; | |
563 | } | |
564 | ||
565 | /* | |
566 | * Is there any need for future grace periods? | |
567 | * Interrupts must be disabled. If the caller does not hold the root | |
568 | * rnp_node structure's ->lock, the results are advisory only. | |
569 | */ | |
570 | static int rcu_future_needs_gp(struct rcu_state *rsp) | |
571 | { | |
572 | struct rcu_node *rnp = rcu_get_root(rsp); | |
7d0ae808 | 573 | int idx = (READ_ONCE(rnp->completed) + 1) & 0x1; |
365187fb PM |
574 | int *fp = &rnp->need_future_gp[idx]; |
575 | ||
7d0ae808 | 576 | return READ_ONCE(*fp); |
365187fb PM |
577 | } |
578 | ||
64db4cff | 579 | /* |
dc35c893 PM |
580 | * Does the current CPU require a not-yet-started grace period? |
581 | * The caller must have disabled interrupts to prevent races with | |
582 | * normal callback registry. | |
64db4cff PM |
583 | */ |
584 | static int | |
585 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
586 | { | |
dc35c893 | 587 | int i; |
3fbfbf7a | 588 | |
dc35c893 PM |
589 | if (rcu_gp_in_progress(rsp)) |
590 | return 0; /* No, a grace period is already in progress. */ | |
365187fb | 591 | if (rcu_future_needs_gp(rsp)) |
34ed6246 | 592 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
593 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
594 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
595 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
596 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
597 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
598 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
7d0ae808 | 599 | ULONG_CMP_LT(READ_ONCE(rsp->completed), |
dc35c893 PM |
600 | rdp->nxtcompleted[i])) |
601 | return 1; /* Yes, CBs for future grace period. */ | |
602 | return 0; /* No grace period needed. */ | |
64db4cff PM |
603 | } |
604 | ||
9b2e4f18 | 605 | /* |
adf5091e | 606 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
607 | * |
608 | * If the new value of the ->dynticks_nesting counter now is zero, | |
609 | * we really have entered idle, and must do the appropriate accounting. | |
610 | * The caller must have disabled interrupts. | |
611 | */ | |
28ced795 | 612 | static void rcu_eqs_enter_common(long long oldval, bool user) |
9b2e4f18 | 613 | { |
96d3fd0d PM |
614 | struct rcu_state *rsp; |
615 | struct rcu_data *rdp; | |
28ced795 | 616 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
96d3fd0d | 617 | |
f7f7bac9 | 618 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
1ce46ee5 PM |
619 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
620 | !user && !is_idle_task(current)) { | |
289828e6 PM |
621 | struct task_struct *idle __maybe_unused = |
622 | idle_task(smp_processor_id()); | |
0989cb46 | 623 | |
f7f7bac9 | 624 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 625 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
626 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
627 | current->pid, current->comm, | |
628 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 629 | } |
96d3fd0d PM |
630 | for_each_rcu_flavor(rsp) { |
631 | rdp = this_cpu_ptr(rsp->rda); | |
632 | do_nocb_deferred_wakeup(rdp); | |
633 | } | |
198bbf81 | 634 | rcu_prepare_for_idle(); |
9b2e4f18 | 635 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 636 | smp_mb__before_atomic(); /* See above. */ |
9b2e4f18 | 637 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 638 | smp_mb__after_atomic(); /* Force ordering with next sojourn. */ |
1ce46ee5 PM |
639 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
640 | atomic_read(&rdtp->dynticks) & 0x1); | |
176f8f7a | 641 | rcu_dynticks_task_enter(); |
c44e2cdd PM |
642 | |
643 | /* | |
adf5091e | 644 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
645 | * in an RCU read-side critical section. |
646 | */ | |
647 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
648 | "Illegal idle entry in RCU read-side critical section."); | |
649 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
650 | "Illegal idle entry in RCU-bh read-side critical section."); | |
651 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
652 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 653 | } |
64db4cff | 654 | |
adf5091e FW |
655 | /* |
656 | * Enter an RCU extended quiescent state, which can be either the | |
657 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 658 | */ |
adf5091e | 659 | static void rcu_eqs_enter(bool user) |
64db4cff | 660 | { |
4145fa7f | 661 | long long oldval; |
64db4cff PM |
662 | struct rcu_dynticks *rdtp; |
663 | ||
c9d4b0af | 664 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 665 | oldval = rdtp->dynticks_nesting; |
1ce46ee5 PM |
666 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
667 | (oldval & DYNTICK_TASK_NEST_MASK) == 0); | |
3a592405 | 668 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { |
29e37d81 | 669 | rdtp->dynticks_nesting = 0; |
28ced795 | 670 | rcu_eqs_enter_common(oldval, user); |
3a592405 | 671 | } else { |
29e37d81 | 672 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; |
3a592405 | 673 | } |
64db4cff | 674 | } |
adf5091e FW |
675 | |
676 | /** | |
677 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
678 | * | |
679 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
680 | * read-side critical sections can occur. (Though RCU read-side | |
681 | * critical sections can occur in irq handlers in idle, a possibility | |
682 | * handled by irq_enter() and irq_exit().) | |
683 | * | |
684 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
685 | * the possibility of usermode upcalls having messed up our count | |
686 | * of interrupt nesting level during the prior busy period. | |
687 | */ | |
688 | void rcu_idle_enter(void) | |
689 | { | |
c5d900bf FW |
690 | unsigned long flags; |
691 | ||
692 | local_irq_save(flags); | |
cb349ca9 | 693 | rcu_eqs_enter(false); |
28ced795 | 694 | rcu_sysidle_enter(0); |
c5d900bf | 695 | local_irq_restore(flags); |
adf5091e | 696 | } |
8a2ecf47 | 697 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 698 | |
2b1d5024 | 699 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
700 | /** |
701 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
702 | * | |
703 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
704 | * is permitted between this call and rcu_user_exit(). This way the | |
705 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
706 | * when the CPU runs in userspace. | |
707 | */ | |
708 | void rcu_user_enter(void) | |
709 | { | |
91d1aa43 | 710 | rcu_eqs_enter(1); |
adf5091e | 711 | } |
2b1d5024 | 712 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 713 | |
9b2e4f18 PM |
714 | /** |
715 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
716 | * | |
717 | * Exit from an interrupt handler, which might possibly result in entering | |
718 | * idle mode, in other words, leaving the mode in which read-side critical | |
719 | * sections can occur. | |
64db4cff | 720 | * |
9b2e4f18 PM |
721 | * This code assumes that the idle loop never does anything that might |
722 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
723 | * architecture violates this assumption, RCU will give you what you | |
724 | * deserve, good and hard. But very infrequently and irreproducibly. | |
725 | * | |
726 | * Use things like work queues to work around this limitation. | |
727 | * | |
728 | * You have been warned. | |
64db4cff | 729 | */ |
9b2e4f18 | 730 | void rcu_irq_exit(void) |
64db4cff PM |
731 | { |
732 | unsigned long flags; | |
4145fa7f | 733 | long long oldval; |
64db4cff PM |
734 | struct rcu_dynticks *rdtp; |
735 | ||
736 | local_irq_save(flags); | |
c9d4b0af | 737 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 738 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 | 739 | rdtp->dynticks_nesting--; |
1ce46ee5 PM |
740 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
741 | rdtp->dynticks_nesting < 0); | |
b6fc6020 | 742 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 743 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 744 | else |
28ced795 CL |
745 | rcu_eqs_enter_common(oldval, true); |
746 | rcu_sysidle_enter(1); | |
9b2e4f18 PM |
747 | local_irq_restore(flags); |
748 | } | |
749 | ||
750 | /* | |
adf5091e | 751 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
752 | * |
753 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
754 | * we really have exited idle, and must do the appropriate accounting. | |
755 | * The caller must have disabled interrupts. | |
756 | */ | |
28ced795 | 757 | static void rcu_eqs_exit_common(long long oldval, int user) |
9b2e4f18 | 758 | { |
28ced795 CL |
759 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
760 | ||
176f8f7a | 761 | rcu_dynticks_task_exit(); |
4e857c58 | 762 | smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */ |
23b5c8fa PM |
763 | atomic_inc(&rdtp->dynticks); |
764 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
4e857c58 | 765 | smp_mb__after_atomic(); /* See above. */ |
1ce46ee5 PM |
766 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
767 | !(atomic_read(&rdtp->dynticks) & 0x1)); | |
8fa7845d | 768 | rcu_cleanup_after_idle(); |
f7f7bac9 | 769 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
1ce46ee5 PM |
770 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
771 | !user && !is_idle_task(current)) { | |
289828e6 PM |
772 | struct task_struct *idle __maybe_unused = |
773 | idle_task(smp_processor_id()); | |
0989cb46 | 774 | |
f7f7bac9 | 775 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 776 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 777 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
778 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
779 | current->pid, current->comm, | |
780 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
781 | } |
782 | } | |
783 | ||
adf5091e FW |
784 | /* |
785 | * Exit an RCU extended quiescent state, which can be either the | |
786 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 787 | */ |
adf5091e | 788 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 789 | { |
9b2e4f18 PM |
790 | struct rcu_dynticks *rdtp; |
791 | long long oldval; | |
792 | ||
c9d4b0af | 793 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 | 794 | oldval = rdtp->dynticks_nesting; |
1ce46ee5 | 795 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0); |
3a592405 | 796 | if (oldval & DYNTICK_TASK_NEST_MASK) { |
29e37d81 | 797 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; |
3a592405 | 798 | } else { |
29e37d81 | 799 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
28ced795 | 800 | rcu_eqs_exit_common(oldval, user); |
3a592405 | 801 | } |
9b2e4f18 | 802 | } |
adf5091e FW |
803 | |
804 | /** | |
805 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
806 | * | |
807 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
808 | * read-side critical sections can occur. | |
809 | * | |
810 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
811 | * allow for the possibility of usermode upcalls messing up our count | |
812 | * of interrupt nesting level during the busy period that is just | |
813 | * now starting. | |
814 | */ | |
815 | void rcu_idle_exit(void) | |
816 | { | |
c5d900bf FW |
817 | unsigned long flags; |
818 | ||
819 | local_irq_save(flags); | |
cb349ca9 | 820 | rcu_eqs_exit(false); |
28ced795 | 821 | rcu_sysidle_exit(0); |
c5d900bf | 822 | local_irq_restore(flags); |
adf5091e | 823 | } |
8a2ecf47 | 824 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 825 | |
2b1d5024 | 826 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
827 | /** |
828 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
829 | * | |
830 | * Exit RCU idle mode while entering the kernel because it can | |
831 | * run a RCU read side critical section anytime. | |
832 | */ | |
833 | void rcu_user_exit(void) | |
834 | { | |
91d1aa43 | 835 | rcu_eqs_exit(1); |
adf5091e | 836 | } |
2b1d5024 | 837 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 838 | |
9b2e4f18 PM |
839 | /** |
840 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
841 | * | |
842 | * Enter an interrupt handler, which might possibly result in exiting | |
843 | * idle mode, in other words, entering the mode in which read-side critical | |
844 | * sections can occur. | |
845 | * | |
846 | * Note that the Linux kernel is fully capable of entering an interrupt | |
847 | * handler that it never exits, for example when doing upcalls to | |
848 | * user mode! This code assumes that the idle loop never does upcalls to | |
849 | * user mode. If your architecture does do upcalls from the idle loop (or | |
850 | * does anything else that results in unbalanced calls to the irq_enter() | |
851 | * and irq_exit() functions), RCU will give you what you deserve, good | |
852 | * and hard. But very infrequently and irreproducibly. | |
853 | * | |
854 | * Use things like work queues to work around this limitation. | |
855 | * | |
856 | * You have been warned. | |
857 | */ | |
858 | void rcu_irq_enter(void) | |
859 | { | |
860 | unsigned long flags; | |
861 | struct rcu_dynticks *rdtp; | |
862 | long long oldval; | |
863 | ||
864 | local_irq_save(flags); | |
c9d4b0af | 865 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 PM |
866 | oldval = rdtp->dynticks_nesting; |
867 | rdtp->dynticks_nesting++; | |
1ce46ee5 PM |
868 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
869 | rdtp->dynticks_nesting == 0); | |
b6fc6020 | 870 | if (oldval) |
f7f7bac9 | 871 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 872 | else |
28ced795 CL |
873 | rcu_eqs_exit_common(oldval, true); |
874 | rcu_sysidle_exit(1); | |
64db4cff | 875 | local_irq_restore(flags); |
64db4cff PM |
876 | } |
877 | ||
878 | /** | |
879 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
880 | * | |
734d1680 PM |
881 | * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and |
882 | * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know | |
883 | * that the CPU is active. This implementation permits nested NMIs, as | |
884 | * long as the nesting level does not overflow an int. (You will probably | |
885 | * run out of stack space first.) | |
64db4cff PM |
886 | */ |
887 | void rcu_nmi_enter(void) | |
888 | { | |
c9d4b0af | 889 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
734d1680 | 890 | int incby = 2; |
64db4cff | 891 | |
734d1680 PM |
892 | /* Complain about underflow. */ |
893 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0); | |
894 | ||
895 | /* | |
896 | * If idle from RCU viewpoint, atomically increment ->dynticks | |
897 | * to mark non-idle and increment ->dynticks_nmi_nesting by one. | |
898 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means | |
899 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed | |
900 | * to be in the outermost NMI handler that interrupted an RCU-idle | |
901 | * period (observation due to Andy Lutomirski). | |
902 | */ | |
903 | if (!(atomic_read(&rdtp->dynticks) & 0x1)) { | |
904 | smp_mb__before_atomic(); /* Force delay from prior write. */ | |
905 | atomic_inc(&rdtp->dynticks); | |
906 | /* atomic_inc() before later RCU read-side crit sects */ | |
907 | smp_mb__after_atomic(); /* See above. */ | |
908 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
909 | incby = 1; | |
910 | } | |
911 | rdtp->dynticks_nmi_nesting += incby; | |
912 | barrier(); | |
64db4cff PM |
913 | } |
914 | ||
915 | /** | |
916 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
917 | * | |
734d1680 PM |
918 | * If we are returning from the outermost NMI handler that interrupted an |
919 | * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting | |
920 | * to let the RCU grace-period handling know that the CPU is back to | |
921 | * being RCU-idle. | |
64db4cff PM |
922 | */ |
923 | void rcu_nmi_exit(void) | |
924 | { | |
c9d4b0af | 925 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 926 | |
734d1680 PM |
927 | /* |
928 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. | |
929 | * (We are exiting an NMI handler, so RCU better be paying attention | |
930 | * to us!) | |
931 | */ | |
932 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0); | |
933 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
934 | ||
935 | /* | |
936 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so | |
937 | * leave it in non-RCU-idle state. | |
938 | */ | |
939 | if (rdtp->dynticks_nmi_nesting != 1) { | |
940 | rdtp->dynticks_nmi_nesting -= 2; | |
64db4cff | 941 | return; |
734d1680 PM |
942 | } |
943 | ||
944 | /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ | |
945 | rdtp->dynticks_nmi_nesting = 0; | |
23b5c8fa | 946 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 947 | smp_mb__before_atomic(); /* See above. */ |
23b5c8fa | 948 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 949 | smp_mb__after_atomic(); /* Force delay to next write. */ |
23b5c8fa | 950 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
64db4cff PM |
951 | } |
952 | ||
953 | /** | |
5c173eb8 PM |
954 | * __rcu_is_watching - are RCU read-side critical sections safe? |
955 | * | |
956 | * Return true if RCU is watching the running CPU, which means that | |
957 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
958 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
959 | * least disabled preemption. | |
960 | */ | |
9418fb20 | 961 | bool notrace __rcu_is_watching(void) |
5c173eb8 PM |
962 | { |
963 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
964 | } | |
965 | ||
966 | /** | |
967 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
64db4cff | 968 | * |
9b2e4f18 | 969 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 970 | * or NMI handler, return true. |
64db4cff | 971 | */ |
9418fb20 | 972 | bool notrace rcu_is_watching(void) |
64db4cff | 973 | { |
f534ed1f | 974 | bool ret; |
34240697 PM |
975 | |
976 | preempt_disable(); | |
5c173eb8 | 977 | ret = __rcu_is_watching(); |
34240697 PM |
978 | preempt_enable(); |
979 | return ret; | |
64db4cff | 980 | } |
5c173eb8 | 981 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 982 | |
62fde6ed | 983 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
984 | |
985 | /* | |
986 | * Is the current CPU online? Disable preemption to avoid false positives | |
987 | * that could otherwise happen due to the current CPU number being sampled, | |
988 | * this task being preempted, its old CPU being taken offline, resuming | |
989 | * on some other CPU, then determining that its old CPU is now offline. | |
990 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
991 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
992 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
993 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
994 | * offline to continue to use RCU for one jiffy after marking itself | |
995 | * offline in the cpu_online_mask. This leniency is necessary given the | |
996 | * non-atomic nature of the online and offline processing, for example, | |
997 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
998 | * notifiers. | |
999 | * | |
1000 | * This is also why RCU internally marks CPUs online during the | |
1001 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
1002 | * |
1003 | * Disable checking if in an NMI handler because we cannot safely report | |
1004 | * errors from NMI handlers anyway. | |
1005 | */ | |
1006 | bool rcu_lockdep_current_cpu_online(void) | |
1007 | { | |
2036d94a PM |
1008 | struct rcu_data *rdp; |
1009 | struct rcu_node *rnp; | |
c0d6d01b PM |
1010 | bool ret; |
1011 | ||
1012 | if (in_nmi()) | |
f6f7ee9a | 1013 | return true; |
c0d6d01b | 1014 | preempt_disable(); |
c9d4b0af | 1015 | rdp = this_cpu_ptr(&rcu_sched_data); |
2036d94a | 1016 | rnp = rdp->mynode; |
0aa04b05 | 1017 | ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) || |
c0d6d01b PM |
1018 | !rcu_scheduler_fully_active; |
1019 | preempt_enable(); | |
1020 | return ret; | |
1021 | } | |
1022 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
1023 | ||
62fde6ed | 1024 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 1025 | |
64db4cff | 1026 | /** |
9b2e4f18 | 1027 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 1028 | * |
9b2e4f18 PM |
1029 | * If the current CPU is idle or running at a first-level (not nested) |
1030 | * interrupt from idle, return true. The caller must have at least | |
1031 | * disabled preemption. | |
64db4cff | 1032 | */ |
62e3cb14 | 1033 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 1034 | { |
c9d4b0af | 1035 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
64db4cff PM |
1036 | } |
1037 | ||
64db4cff PM |
1038 | /* |
1039 | * Snapshot the specified CPU's dynticks counter so that we can later | |
1040 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 1041 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 1042 | */ |
217af2a2 PM |
1043 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
1044 | bool *isidle, unsigned long *maxj) | |
64db4cff | 1045 | { |
23b5c8fa | 1046 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 1047 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
7941dbde ACB |
1048 | if ((rdp->dynticks_snap & 0x1) == 0) { |
1049 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); | |
1050 | return 1; | |
1051 | } else { | |
7d0ae808 | 1052 | if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, |
e3663b10 | 1053 | rdp->mynode->gpnum)) |
7d0ae808 | 1054 | WRITE_ONCE(rdp->gpwrap, true); |
7941dbde ACB |
1055 | return 0; |
1056 | } | |
64db4cff PM |
1057 | } |
1058 | ||
1059 | /* | |
1060 | * Return true if the specified CPU has passed through a quiescent | |
1061 | * state by virtue of being in or having passed through an dynticks | |
1062 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 1063 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 1064 | */ |
217af2a2 PM |
1065 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
1066 | bool *isidle, unsigned long *maxj) | |
64db4cff | 1067 | { |
7eb4f455 | 1068 | unsigned int curr; |
4a81e832 | 1069 | int *rcrmp; |
7eb4f455 | 1070 | unsigned int snap; |
64db4cff | 1071 | |
7eb4f455 PM |
1072 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
1073 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
1074 | |
1075 | /* | |
1076 | * If the CPU passed through or entered a dynticks idle phase with | |
1077 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
1078 | * already acknowledged the request to pass through a quiescent | |
1079 | * state. Either way, that CPU cannot possibly be in an RCU | |
1080 | * read-side critical section that started before the beginning | |
1081 | * of the current RCU grace period. | |
1082 | */ | |
7eb4f455 | 1083 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 1084 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
1085 | rdp->dynticks_fqs++; |
1086 | return 1; | |
1087 | } | |
1088 | ||
a82dcc76 PM |
1089 | /* |
1090 | * Check for the CPU being offline, but only if the grace period | |
1091 | * is old enough. We don't need to worry about the CPU changing | |
1092 | * state: If we see it offline even once, it has been through a | |
1093 | * quiescent state. | |
1094 | * | |
1095 | * The reason for insisting that the grace period be at least | |
1096 | * one jiffy old is that CPUs that are not quite online and that | |
1097 | * have just gone offline can still execute RCU read-side critical | |
1098 | * sections. | |
1099 | */ | |
1100 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
1101 | return 0; /* Grace period is not old enough. */ | |
1102 | barrier(); | |
1103 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 1104 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
1105 | rdp->offline_fqs++; |
1106 | return 1; | |
1107 | } | |
65d798f0 PM |
1108 | |
1109 | /* | |
4a81e832 PM |
1110 | * A CPU running for an extended time within the kernel can |
1111 | * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode, | |
1112 | * even context-switching back and forth between a pair of | |
1113 | * in-kernel CPU-bound tasks cannot advance grace periods. | |
1114 | * So if the grace period is old enough, make the CPU pay attention. | |
1115 | * Note that the unsynchronized assignments to the per-CPU | |
1116 | * rcu_sched_qs_mask variable are safe. Yes, setting of | |
1117 | * bits can be lost, but they will be set again on the next | |
1118 | * force-quiescent-state pass. So lost bit sets do not result | |
1119 | * in incorrect behavior, merely in a grace period lasting | |
1120 | * a few jiffies longer than it might otherwise. Because | |
1121 | * there are at most four threads involved, and because the | |
1122 | * updates are only once every few jiffies, the probability of | |
1123 | * lossage (and thus of slight grace-period extension) is | |
1124 | * quite low. | |
1125 | * | |
1126 | * Note that if the jiffies_till_sched_qs boot/sysfs parameter | |
1127 | * is set too high, we override with half of the RCU CPU stall | |
1128 | * warning delay. | |
6193c76a | 1129 | */ |
4a81e832 PM |
1130 | rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu); |
1131 | if (ULONG_CMP_GE(jiffies, | |
1132 | rdp->rsp->gp_start + jiffies_till_sched_qs) || | |
cb1e78cf | 1133 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { |
7d0ae808 PM |
1134 | if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) { |
1135 | WRITE_ONCE(rdp->cond_resched_completed, | |
1136 | READ_ONCE(rdp->mynode->completed)); | |
4a81e832 | 1137 | smp_mb(); /* ->cond_resched_completed before *rcrmp. */ |
7d0ae808 PM |
1138 | WRITE_ONCE(*rcrmp, |
1139 | READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask); | |
4a81e832 PM |
1140 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ |
1141 | rdp->rsp->jiffies_resched += 5; /* Enable beating. */ | |
1142 | } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { | |
1143 | /* Time to beat on that CPU again! */ | |
1144 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1145 | rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */ | |
1146 | } | |
6193c76a PM |
1147 | } |
1148 | ||
a82dcc76 | 1149 | return 0; |
64db4cff PM |
1150 | } |
1151 | ||
64db4cff PM |
1152 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
1153 | { | |
cb1e78cf | 1154 | unsigned long j = jiffies; |
6193c76a | 1155 | unsigned long j1; |
26cdfedf PM |
1156 | |
1157 | rsp->gp_start = j; | |
1158 | smp_wmb(); /* Record start time before stall time. */ | |
6193c76a | 1159 | j1 = rcu_jiffies_till_stall_check(); |
7d0ae808 | 1160 | WRITE_ONCE(rsp->jiffies_stall, j + j1); |
6193c76a | 1161 | rsp->jiffies_resched = j + j1 / 2; |
7d0ae808 | 1162 | rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs); |
64db4cff PM |
1163 | } |
1164 | ||
fb81a44b PM |
1165 | /* |
1166 | * Complain about starvation of grace-period kthread. | |
1167 | */ | |
1168 | static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp) | |
1169 | { | |
1170 | unsigned long gpa; | |
1171 | unsigned long j; | |
1172 | ||
1173 | j = jiffies; | |
7d0ae808 | 1174 | gpa = READ_ONCE(rsp->gp_activity); |
fb81a44b | 1175 | if (j - gpa > 2 * HZ) |
319362c9 | 1176 | pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x s%d ->state=%#lx\n", |
81e701e4 | 1177 | rsp->name, j - gpa, |
319362c9 PM |
1178 | rsp->gpnum, rsp->completed, |
1179 | rsp->gp_flags, rsp->gp_state, | |
1180 | rsp->gp_kthread ? rsp->gp_kthread->state : 0); | |
64db4cff PM |
1181 | } |
1182 | ||
b637a328 | 1183 | /* |
bc1dce51 | 1184 | * Dump stacks of all tasks running on stalled CPUs. |
b637a328 PM |
1185 | */ |
1186 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
1187 | { | |
1188 | int cpu; | |
1189 | unsigned long flags; | |
1190 | struct rcu_node *rnp; | |
1191 | ||
1192 | rcu_for_each_leaf_node(rsp, rnp) { | |
1193 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1194 | if (rnp->qsmask != 0) { | |
1195 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1196 | if (rnp->qsmask & (1UL << cpu)) | |
1197 | dump_cpu_task(rnp->grplo + cpu); | |
1198 | } | |
1199 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1200 | } | |
1201 | } | |
1202 | ||
6ccd2ecd | 1203 | static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum) |
64db4cff PM |
1204 | { |
1205 | int cpu; | |
1206 | long delta; | |
1207 | unsigned long flags; | |
6ccd2ecd PM |
1208 | unsigned long gpa; |
1209 | unsigned long j; | |
285fe294 | 1210 | int ndetected = 0; |
64db4cff | 1211 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 1212 | long totqlen = 0; |
64db4cff PM |
1213 | |
1214 | /* Only let one CPU complain about others per time interval. */ | |
1215 | ||
1304afb2 | 1216 | raw_spin_lock_irqsave(&rnp->lock, flags); |
7d0ae808 | 1217 | delta = jiffies - READ_ONCE(rsp->jiffies_stall); |
fc2219d4 | 1218 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 1219 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1220 | return; |
1221 | } | |
7d0ae808 PM |
1222 | WRITE_ONCE(rsp->jiffies_stall, |
1223 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3); | |
1304afb2 | 1224 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1225 | |
8cdd32a9 PM |
1226 | /* |
1227 | * OK, time to rat on our buddy... | |
1228 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1229 | * RCU CPU stall warnings. | |
1230 | */ | |
d7f3e207 | 1231 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 1232 | rsp->name); |
a858af28 | 1233 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 1234 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 1235 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 1236 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
1237 | if (rnp->qsmask != 0) { |
1238 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1239 | if (rnp->qsmask & (1UL << cpu)) { | |
1240 | print_cpu_stall_info(rsp, | |
1241 | rnp->grplo + cpu); | |
1242 | ndetected++; | |
1243 | } | |
1244 | } | |
3acd9eb3 | 1245 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1246 | } |
a858af28 | 1247 | |
a858af28 | 1248 | print_cpu_stall_info_end(); |
53bb857c PM |
1249 | for_each_possible_cpu(cpu) |
1250 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e | 1251 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", |
eee05882 | 1252 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
83ebe63e | 1253 | (long)rsp->gpnum, (long)rsp->completed, totqlen); |
6ccd2ecd | 1254 | if (ndetected) { |
b637a328 | 1255 | rcu_dump_cpu_stacks(rsp); |
6ccd2ecd | 1256 | } else { |
7d0ae808 PM |
1257 | if (READ_ONCE(rsp->gpnum) != gpnum || |
1258 | READ_ONCE(rsp->completed) == gpnum) { | |
6ccd2ecd PM |
1259 | pr_err("INFO: Stall ended before state dump start\n"); |
1260 | } else { | |
1261 | j = jiffies; | |
7d0ae808 | 1262 | gpa = READ_ONCE(rsp->gp_activity); |
237a0f21 | 1263 | pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n", |
6ccd2ecd | 1264 | rsp->name, j - gpa, j, gpa, |
237a0f21 PM |
1265 | jiffies_till_next_fqs, |
1266 | rcu_get_root(rsp)->qsmask); | |
6ccd2ecd PM |
1267 | /* In this case, the current CPU might be at fault. */ |
1268 | sched_show_task(current); | |
1269 | } | |
1270 | } | |
c1dc0b9c | 1271 | |
4cdfc175 | 1272 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
1273 | rcu_print_detail_task_stall(rsp); |
1274 | ||
fb81a44b PM |
1275 | rcu_check_gp_kthread_starvation(rsp); |
1276 | ||
4cdfc175 | 1277 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
1278 | } |
1279 | ||
1280 | static void print_cpu_stall(struct rcu_state *rsp) | |
1281 | { | |
53bb857c | 1282 | int cpu; |
64db4cff PM |
1283 | unsigned long flags; |
1284 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 1285 | long totqlen = 0; |
64db4cff | 1286 | |
8cdd32a9 PM |
1287 | /* |
1288 | * OK, time to rat on ourselves... | |
1289 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1290 | * RCU CPU stall warnings. | |
1291 | */ | |
d7f3e207 | 1292 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
1293 | print_cpu_stall_info_begin(); |
1294 | print_cpu_stall_info(rsp, smp_processor_id()); | |
1295 | print_cpu_stall_info_end(); | |
53bb857c PM |
1296 | for_each_possible_cpu(cpu) |
1297 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e PM |
1298 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", |
1299 | jiffies - rsp->gp_start, | |
1300 | (long)rsp->gpnum, (long)rsp->completed, totqlen); | |
fb81a44b PM |
1301 | |
1302 | rcu_check_gp_kthread_starvation(rsp); | |
1303 | ||
bc1dce51 | 1304 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 1305 | |
1304afb2 | 1306 | raw_spin_lock_irqsave(&rnp->lock, flags); |
7d0ae808 PM |
1307 | if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall))) |
1308 | WRITE_ONCE(rsp->jiffies_stall, | |
1309 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3); | |
1304afb2 | 1310 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 1311 | |
b021fe3e PZ |
1312 | /* |
1313 | * Attempt to revive the RCU machinery by forcing a context switch. | |
1314 | * | |
1315 | * A context switch would normally allow the RCU state machine to make | |
1316 | * progress and it could be we're stuck in kernel space without context | |
1317 | * switches for an entirely unreasonable amount of time. | |
1318 | */ | |
1319 | resched_cpu(smp_processor_id()); | |
64db4cff PM |
1320 | } |
1321 | ||
1322 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
1323 | { | |
26cdfedf PM |
1324 | unsigned long completed; |
1325 | unsigned long gpnum; | |
1326 | unsigned long gps; | |
bad6e139 PM |
1327 | unsigned long j; |
1328 | unsigned long js; | |
64db4cff PM |
1329 | struct rcu_node *rnp; |
1330 | ||
26cdfedf | 1331 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
c68de209 | 1332 | return; |
cb1e78cf | 1333 | j = jiffies; |
26cdfedf PM |
1334 | |
1335 | /* | |
1336 | * Lots of memory barriers to reject false positives. | |
1337 | * | |
1338 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
1339 | * then rsp->gp_start, and finally rsp->completed. These values | |
1340 | * are updated in the opposite order with memory barriers (or | |
1341 | * equivalent) during grace-period initialization and cleanup. | |
1342 | * Now, a false positive can occur if we get an new value of | |
1343 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
1344 | * the memory barriers, the only way that this can happen is if one | |
1345 | * grace period ends and another starts between these two fetches. | |
1346 | * Detect this by comparing rsp->completed with the previous fetch | |
1347 | * from rsp->gpnum. | |
1348 | * | |
1349 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
1350 | * and rsp->gp_start suffice to forestall false positives. | |
1351 | */ | |
7d0ae808 | 1352 | gpnum = READ_ONCE(rsp->gpnum); |
26cdfedf | 1353 | smp_rmb(); /* Pick up ->gpnum first... */ |
7d0ae808 | 1354 | js = READ_ONCE(rsp->jiffies_stall); |
26cdfedf | 1355 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
7d0ae808 | 1356 | gps = READ_ONCE(rsp->gp_start); |
26cdfedf | 1357 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ |
7d0ae808 | 1358 | completed = READ_ONCE(rsp->completed); |
26cdfedf PM |
1359 | if (ULONG_CMP_GE(completed, gpnum) || |
1360 | ULONG_CMP_LT(j, js) || | |
1361 | ULONG_CMP_GE(gps, js)) | |
1362 | return; /* No stall or GP completed since entering function. */ | |
64db4cff | 1363 | rnp = rdp->mynode; |
c96ea7cf | 1364 | if (rcu_gp_in_progress(rsp) && |
7d0ae808 | 1365 | (READ_ONCE(rnp->qsmask) & rdp->grpmask)) { |
64db4cff PM |
1366 | |
1367 | /* We haven't checked in, so go dump stack. */ | |
1368 | print_cpu_stall(rsp); | |
1369 | ||
bad6e139 PM |
1370 | } else if (rcu_gp_in_progress(rsp) && |
1371 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 1372 | |
bad6e139 | 1373 | /* They had a few time units to dump stack, so complain. */ |
6ccd2ecd | 1374 | print_other_cpu_stall(rsp, gpnum); |
64db4cff PM |
1375 | } |
1376 | } | |
1377 | ||
53d84e00 PM |
1378 | /** |
1379 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1380 | * | |
1381 | * Set the stall-warning timeout way off into the future, thus preventing | |
1382 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1383 | * RCU grace periods. | |
1384 | * | |
1385 | * The caller must disable hard irqs. | |
1386 | */ | |
1387 | void rcu_cpu_stall_reset(void) | |
1388 | { | |
6ce75a23 PM |
1389 | struct rcu_state *rsp; |
1390 | ||
1391 | for_each_rcu_flavor(rsp) | |
7d0ae808 | 1392 | WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2); |
53d84e00 PM |
1393 | } |
1394 | ||
3f5d3ea6 | 1395 | /* |
d3f3f3f2 PM |
1396 | * Initialize the specified rcu_data structure's default callback list |
1397 | * to empty. The default callback list is the one that is not used by | |
1398 | * no-callbacks CPUs. | |
3f5d3ea6 | 1399 | */ |
d3f3f3f2 | 1400 | static void init_default_callback_list(struct rcu_data *rdp) |
3f5d3ea6 PM |
1401 | { |
1402 | int i; | |
1403 | ||
1404 | rdp->nxtlist = NULL; | |
1405 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1406 | rdp->nxttail[i] = &rdp->nxtlist; | |
1407 | } | |
1408 | ||
d3f3f3f2 PM |
1409 | /* |
1410 | * Initialize the specified rcu_data structure's callback list to empty. | |
1411 | */ | |
1412 | static void init_callback_list(struct rcu_data *rdp) | |
1413 | { | |
1414 | if (init_nocb_callback_list(rdp)) | |
1415 | return; | |
1416 | init_default_callback_list(rdp); | |
1417 | } | |
1418 | ||
dc35c893 PM |
1419 | /* |
1420 | * Determine the value that ->completed will have at the end of the | |
1421 | * next subsequent grace period. This is used to tag callbacks so that | |
1422 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1423 | * been dyntick-idle for an extended period with callbacks under the | |
1424 | * influence of RCU_FAST_NO_HZ. | |
1425 | * | |
1426 | * The caller must hold rnp->lock with interrupts disabled. | |
1427 | */ | |
1428 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1429 | struct rcu_node *rnp) | |
1430 | { | |
1431 | /* | |
1432 | * If RCU is idle, we just wait for the next grace period. | |
1433 | * But we can only be sure that RCU is idle if we are looking | |
1434 | * at the root rcu_node structure -- otherwise, a new grace | |
1435 | * period might have started, but just not yet gotten around | |
1436 | * to initializing the current non-root rcu_node structure. | |
1437 | */ | |
1438 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1439 | return rnp->completed + 1; | |
1440 | ||
1441 | /* | |
1442 | * Otherwise, wait for a possible partial grace period and | |
1443 | * then the subsequent full grace period. | |
1444 | */ | |
1445 | return rnp->completed + 2; | |
1446 | } | |
1447 | ||
0446be48 PM |
1448 | /* |
1449 | * Trace-event helper function for rcu_start_future_gp() and | |
1450 | * rcu_nocb_wait_gp(). | |
1451 | */ | |
1452 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1453 | unsigned long c, const char *s) |
0446be48 PM |
1454 | { |
1455 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1456 | rnp->completed, c, rnp->level, | |
1457 | rnp->grplo, rnp->grphi, s); | |
1458 | } | |
1459 | ||
1460 | /* | |
1461 | * Start some future grace period, as needed to handle newly arrived | |
1462 | * callbacks. The required future grace periods are recorded in each | |
48a7639c PM |
1463 | * rcu_node structure's ->need_future_gp field. Returns true if there |
1464 | * is reason to awaken the grace-period kthread. | |
0446be48 PM |
1465 | * |
1466 | * The caller must hold the specified rcu_node structure's ->lock. | |
1467 | */ | |
48a7639c PM |
1468 | static bool __maybe_unused |
1469 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
1470 | unsigned long *c_out) | |
0446be48 PM |
1471 | { |
1472 | unsigned long c; | |
1473 | int i; | |
48a7639c | 1474 | bool ret = false; |
0446be48 PM |
1475 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); |
1476 | ||
1477 | /* | |
1478 | * Pick up grace-period number for new callbacks. If this | |
1479 | * grace period is already marked as needed, return to the caller. | |
1480 | */ | |
1481 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1482 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1483 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1484 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
48a7639c | 1485 | goto out; |
0446be48 PM |
1486 | } |
1487 | ||
1488 | /* | |
1489 | * If either this rcu_node structure or the root rcu_node structure | |
1490 | * believe that a grace period is in progress, then we must wait | |
1491 | * for the one following, which is in "c". Because our request | |
1492 | * will be noticed at the end of the current grace period, we don't | |
48bd8e9b PK |
1493 | * need to explicitly start one. We only do the lockless check |
1494 | * of rnp_root's fields if the current rcu_node structure thinks | |
1495 | * there is no grace period in flight, and because we hold rnp->lock, | |
1496 | * the only possible change is when rnp_root's two fields are | |
1497 | * equal, in which case rnp_root->gpnum might be concurrently | |
1498 | * incremented. But that is OK, as it will just result in our | |
1499 | * doing some extra useless work. | |
0446be48 PM |
1500 | */ |
1501 | if (rnp->gpnum != rnp->completed || | |
7d0ae808 | 1502 | READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) { |
0446be48 | 1503 | rnp->need_future_gp[c & 0x1]++; |
f7f7bac9 | 1504 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
48a7639c | 1505 | goto out; |
0446be48 PM |
1506 | } |
1507 | ||
1508 | /* | |
1509 | * There might be no grace period in progress. If we don't already | |
1510 | * hold it, acquire the root rcu_node structure's lock in order to | |
1511 | * start one (if needed). | |
1512 | */ | |
6303b9c8 | 1513 | if (rnp != rnp_root) { |
0446be48 | 1514 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 PM |
1515 | smp_mb__after_unlock_lock(); |
1516 | } | |
0446be48 PM |
1517 | |
1518 | /* | |
1519 | * Get a new grace-period number. If there really is no grace | |
1520 | * period in progress, it will be smaller than the one we obtained | |
1521 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1522 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1523 | */ | |
1524 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1525 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1526 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1527 | rdp->nxtcompleted[i] = c; | |
1528 | ||
1529 | /* | |
1530 | * If the needed for the required grace period is already | |
1531 | * recorded, trace and leave. | |
1532 | */ | |
1533 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1534 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1535 | goto unlock_out; |
1536 | } | |
1537 | ||
1538 | /* Record the need for the future grace period. */ | |
1539 | rnp_root->need_future_gp[c & 0x1]++; | |
1540 | ||
1541 | /* If a grace period is not already in progress, start one. */ | |
1542 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1543 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1544 | } else { |
f7f7bac9 | 1545 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
48a7639c | 1546 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1547 | } |
1548 | unlock_out: | |
1549 | if (rnp != rnp_root) | |
1550 | raw_spin_unlock(&rnp_root->lock); | |
48a7639c PM |
1551 | out: |
1552 | if (c_out != NULL) | |
1553 | *c_out = c; | |
1554 | return ret; | |
0446be48 PM |
1555 | } |
1556 | ||
1557 | /* | |
1558 | * Clean up any old requests for the just-ended grace period. Also return | |
1559 | * whether any additional grace periods have been requested. Also invoke | |
1560 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1561 | * waiting for this grace period to complete. | |
1562 | */ | |
1563 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1564 | { | |
1565 | int c = rnp->completed; | |
1566 | int needmore; | |
1567 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1568 | ||
1569 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1570 | rnp->need_future_gp[c & 0x1] = 0; | |
1571 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1572 | trace_rcu_future_gp(rnp, rdp, c, |
1573 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1574 | return needmore; |
1575 | } | |
1576 | ||
48a7639c PM |
1577 | /* |
1578 | * Awaken the grace-period kthread for the specified flavor of RCU. | |
1579 | * Don't do a self-awaken, and don't bother awakening when there is | |
1580 | * nothing for the grace-period kthread to do (as in several CPUs | |
1581 | * raced to awaken, and we lost), and finally don't try to awaken | |
1582 | * a kthread that has not yet been created. | |
1583 | */ | |
1584 | static void rcu_gp_kthread_wake(struct rcu_state *rsp) | |
1585 | { | |
1586 | if (current == rsp->gp_kthread || | |
7d0ae808 | 1587 | !READ_ONCE(rsp->gp_flags) || |
48a7639c PM |
1588 | !rsp->gp_kthread) |
1589 | return; | |
1590 | wake_up(&rsp->gp_wq); | |
1591 | } | |
1592 | ||
dc35c893 PM |
1593 | /* |
1594 | * If there is room, assign a ->completed number to any callbacks on | |
1595 | * this CPU that have not already been assigned. Also accelerate any | |
1596 | * callbacks that were previously assigned a ->completed number that has | |
1597 | * since proven to be too conservative, which can happen if callbacks get | |
1598 | * assigned a ->completed number while RCU is idle, but with reference to | |
1599 | * a non-root rcu_node structure. This function is idempotent, so it does | |
48a7639c PM |
1600 | * not hurt to call it repeatedly. Returns an flag saying that we should |
1601 | * awaken the RCU grace-period kthread. | |
dc35c893 PM |
1602 | * |
1603 | * The caller must hold rnp->lock with interrupts disabled. | |
1604 | */ | |
48a7639c | 1605 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1606 | struct rcu_data *rdp) |
1607 | { | |
1608 | unsigned long c; | |
1609 | int i; | |
48a7639c | 1610 | bool ret; |
dc35c893 PM |
1611 | |
1612 | /* If the CPU has no callbacks, nothing to do. */ | |
1613 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1614 | return false; |
dc35c893 PM |
1615 | |
1616 | /* | |
1617 | * Starting from the sublist containing the callbacks most | |
1618 | * recently assigned a ->completed number and working down, find the | |
1619 | * first sublist that is not assignable to an upcoming grace period. | |
1620 | * Such a sublist has something in it (first two tests) and has | |
1621 | * a ->completed number assigned that will complete sooner than | |
1622 | * the ->completed number for newly arrived callbacks (last test). | |
1623 | * | |
1624 | * The key point is that any later sublist can be assigned the | |
1625 | * same ->completed number as the newly arrived callbacks, which | |
1626 | * means that the callbacks in any of these later sublist can be | |
1627 | * grouped into a single sublist, whether or not they have already | |
1628 | * been assigned a ->completed number. | |
1629 | */ | |
1630 | c = rcu_cbs_completed(rsp, rnp); | |
1631 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1632 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1633 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1634 | break; | |
1635 | ||
1636 | /* | |
1637 | * If there are no sublist for unassigned callbacks, leave. | |
1638 | * At the same time, advance "i" one sublist, so that "i" will | |
1639 | * index into the sublist where all the remaining callbacks should | |
1640 | * be grouped into. | |
1641 | */ | |
1642 | if (++i >= RCU_NEXT_TAIL) | |
48a7639c | 1643 | return false; |
dc35c893 PM |
1644 | |
1645 | /* | |
1646 | * Assign all subsequent callbacks' ->completed number to the next | |
1647 | * full grace period and group them all in the sublist initially | |
1648 | * indexed by "i". | |
1649 | */ | |
1650 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1651 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1652 | rdp->nxtcompleted[i] = c; | |
1653 | } | |
910ee45d | 1654 | /* Record any needed additional grace periods. */ |
48a7639c | 1655 | ret = rcu_start_future_gp(rnp, rdp, NULL); |
6d4b418c PM |
1656 | |
1657 | /* Trace depending on how much we were able to accelerate. */ | |
1658 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1659 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1660 | else |
f7f7bac9 | 1661 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
48a7639c | 1662 | return ret; |
dc35c893 PM |
1663 | } |
1664 | ||
1665 | /* | |
1666 | * Move any callbacks whose grace period has completed to the | |
1667 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1668 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1669 | * sublist. This function is idempotent, so it does not hurt to | |
1670 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1671 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1672 | * |
1673 | * The caller must hold rnp->lock with interrupts disabled. | |
1674 | */ | |
48a7639c | 1675 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1676 | struct rcu_data *rdp) |
1677 | { | |
1678 | int i, j; | |
1679 | ||
1680 | /* If the CPU has no callbacks, nothing to do. */ | |
1681 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1682 | return false; |
dc35c893 PM |
1683 | |
1684 | /* | |
1685 | * Find all callbacks whose ->completed numbers indicate that they | |
1686 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1687 | */ | |
1688 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1689 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1690 | break; | |
1691 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1692 | } | |
1693 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1694 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1695 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1696 | ||
1697 | /* Copy down callbacks to fill in empty sublists. */ | |
1698 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1699 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1700 | break; | |
1701 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1702 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1703 | } | |
1704 | ||
1705 | /* Classify any remaining callbacks. */ | |
48a7639c | 1706 | return rcu_accelerate_cbs(rsp, rnp, rdp); |
dc35c893 PM |
1707 | } |
1708 | ||
d09b62df | 1709 | /* |
ba9fbe95 PM |
1710 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1711 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1712 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1713 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1714 | */ |
48a7639c PM |
1715 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, |
1716 | struct rcu_data *rdp) | |
d09b62df | 1717 | { |
48a7639c PM |
1718 | bool ret; |
1719 | ||
ba9fbe95 | 1720 | /* Handle the ends of any preceding grace periods first. */ |
e3663b10 | 1721 | if (rdp->completed == rnp->completed && |
7d0ae808 | 1722 | !unlikely(READ_ONCE(rdp->gpwrap))) { |
d09b62df | 1723 | |
ba9fbe95 | 1724 | /* No grace period end, so just accelerate recent callbacks. */ |
48a7639c | 1725 | ret = rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1726 | |
dc35c893 PM |
1727 | } else { |
1728 | ||
1729 | /* Advance callbacks. */ | |
48a7639c | 1730 | ret = rcu_advance_cbs(rsp, rnp, rdp); |
d09b62df PM |
1731 | |
1732 | /* Remember that we saw this grace-period completion. */ | |
1733 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1734 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1735 | } |
398ebe60 | 1736 | |
7d0ae808 | 1737 | if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) { |
6eaef633 PM |
1738 | /* |
1739 | * If the current grace period is waiting for this CPU, | |
1740 | * set up to detect a quiescent state, otherwise don't | |
1741 | * go looking for one. | |
1742 | */ | |
1743 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1744 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
6eaef633 | 1745 | rdp->passed_quiesce = 0; |
5cd37193 | 1746 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); |
6eaef633 PM |
1747 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); |
1748 | zero_cpu_stall_ticks(rdp); | |
7d0ae808 | 1749 | WRITE_ONCE(rdp->gpwrap, false); |
6eaef633 | 1750 | } |
48a7639c | 1751 | return ret; |
6eaef633 PM |
1752 | } |
1753 | ||
d34ea322 | 1754 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1755 | { |
1756 | unsigned long flags; | |
48a7639c | 1757 | bool needwake; |
6eaef633 PM |
1758 | struct rcu_node *rnp; |
1759 | ||
1760 | local_irq_save(flags); | |
1761 | rnp = rdp->mynode; | |
7d0ae808 PM |
1762 | if ((rdp->gpnum == READ_ONCE(rnp->gpnum) && |
1763 | rdp->completed == READ_ONCE(rnp->completed) && | |
1764 | !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */ | |
6eaef633 PM |
1765 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
1766 | local_irq_restore(flags); | |
1767 | return; | |
1768 | } | |
6303b9c8 | 1769 | smp_mb__after_unlock_lock(); |
48a7639c | 1770 | needwake = __note_gp_changes(rsp, rnp, rdp); |
6eaef633 | 1771 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
48a7639c PM |
1772 | if (needwake) |
1773 | rcu_gp_kthread_wake(rsp); | |
6eaef633 PM |
1774 | } |
1775 | ||
0f41c0dd PM |
1776 | static void rcu_gp_slow(struct rcu_state *rsp, int delay) |
1777 | { | |
1778 | if (delay > 0 && | |
1779 | !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay))) | |
1780 | schedule_timeout_uninterruptible(delay); | |
1781 | } | |
1782 | ||
b3dbec76 | 1783 | /* |
f7be8209 | 1784 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1785 | */ |
7fdefc10 | 1786 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 | 1787 | { |
0aa04b05 | 1788 | unsigned long oldmask; |
b3dbec76 | 1789 | struct rcu_data *rdp; |
7fdefc10 | 1790 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1791 | |
7d0ae808 | 1792 | WRITE_ONCE(rsp->gp_activity, jiffies); |
7fdefc10 | 1793 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1794 | smp_mb__after_unlock_lock(); |
7d0ae808 | 1795 | if (!READ_ONCE(rsp->gp_flags)) { |
f7be8209 PM |
1796 | /* Spurious wakeup, tell caller to go back to sleep. */ |
1797 | raw_spin_unlock_irq(&rnp->lock); | |
1798 | return 0; | |
1799 | } | |
7d0ae808 | 1800 | WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */ |
b3dbec76 | 1801 | |
f7be8209 PM |
1802 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1803 | /* | |
1804 | * Grace period already in progress, don't start another. | |
1805 | * Not supposed to be able to happen. | |
1806 | */ | |
7fdefc10 PM |
1807 | raw_spin_unlock_irq(&rnp->lock); |
1808 | return 0; | |
1809 | } | |
1810 | ||
7fdefc10 | 1811 | /* Advance to a new grace period and initialize state. */ |
26cdfedf | 1812 | record_gp_stall_check_time(rsp); |
765a3f4f PM |
1813 | /* Record GP times before starting GP, hence smp_store_release(). */ |
1814 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); | |
f7f7bac9 | 1815 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1816 | raw_spin_unlock_irq(&rnp->lock); |
1817 | ||
0aa04b05 PM |
1818 | /* |
1819 | * Apply per-leaf buffered online and offline operations to the | |
1820 | * rcu_node tree. Note that this new grace period need not wait | |
1821 | * for subsequent online CPUs, and that quiescent-state forcing | |
1822 | * will handle subsequent offline CPUs. | |
1823 | */ | |
1824 | rcu_for_each_leaf_node(rsp, rnp) { | |
0f41c0dd | 1825 | rcu_gp_slow(rsp, gp_preinit_delay); |
0aa04b05 PM |
1826 | raw_spin_lock_irq(&rnp->lock); |
1827 | smp_mb__after_unlock_lock(); | |
1828 | if (rnp->qsmaskinit == rnp->qsmaskinitnext && | |
1829 | !rnp->wait_blkd_tasks) { | |
1830 | /* Nothing to do on this leaf rcu_node structure. */ | |
1831 | raw_spin_unlock_irq(&rnp->lock); | |
1832 | continue; | |
1833 | } | |
1834 | ||
1835 | /* Record old state, apply changes to ->qsmaskinit field. */ | |
1836 | oldmask = rnp->qsmaskinit; | |
1837 | rnp->qsmaskinit = rnp->qsmaskinitnext; | |
1838 | ||
1839 | /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ | |
1840 | if (!oldmask != !rnp->qsmaskinit) { | |
1841 | if (!oldmask) /* First online CPU for this rcu_node. */ | |
1842 | rcu_init_new_rnp(rnp); | |
1843 | else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */ | |
1844 | rnp->wait_blkd_tasks = true; | |
1845 | else /* Last offline CPU and can propagate. */ | |
1846 | rcu_cleanup_dead_rnp(rnp); | |
1847 | } | |
1848 | ||
1849 | /* | |
1850 | * If all waited-on tasks from prior grace period are | |
1851 | * done, and if all this rcu_node structure's CPUs are | |
1852 | * still offline, propagate up the rcu_node tree and | |
1853 | * clear ->wait_blkd_tasks. Otherwise, if one of this | |
1854 | * rcu_node structure's CPUs has since come back online, | |
1855 | * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp() | |
1856 | * checks for this, so just call it unconditionally). | |
1857 | */ | |
1858 | if (rnp->wait_blkd_tasks && | |
1859 | (!rcu_preempt_has_tasks(rnp) || | |
1860 | rnp->qsmaskinit)) { | |
1861 | rnp->wait_blkd_tasks = false; | |
1862 | rcu_cleanup_dead_rnp(rnp); | |
1863 | } | |
1864 | ||
1865 | raw_spin_unlock_irq(&rnp->lock); | |
1866 | } | |
7fdefc10 PM |
1867 | |
1868 | /* | |
1869 | * Set the quiescent-state-needed bits in all the rcu_node | |
1870 | * structures for all currently online CPUs in breadth-first order, | |
1871 | * starting from the root rcu_node structure, relying on the layout | |
1872 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1873 | * will access only the leaves of the hierarchy, thus seeing that no | |
1874 | * grace period is in progress, at least until the corresponding | |
1875 | * leaf node has been initialized. In addition, we have excluded | |
1876 | * CPU-hotplug operations. | |
1877 | * | |
1878 | * The grace period cannot complete until the initialization | |
1879 | * process finishes, because this kthread handles both. | |
1880 | */ | |
1881 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
0f41c0dd | 1882 | rcu_gp_slow(rsp, gp_init_delay); |
b3dbec76 | 1883 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1884 | smp_mb__after_unlock_lock(); |
b3dbec76 | 1885 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1886 | rcu_preempt_check_blocked_tasks(rnp); |
1887 | rnp->qsmask = rnp->qsmaskinit; | |
7d0ae808 | 1888 | WRITE_ONCE(rnp->gpnum, rsp->gpnum); |
3f47da0f | 1889 | if (WARN_ON_ONCE(rnp->completed != rsp->completed)) |
7d0ae808 | 1890 | WRITE_ONCE(rnp->completed, rsp->completed); |
7fdefc10 | 1891 | if (rnp == rdp->mynode) |
48a7639c | 1892 | (void)__note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1893 | rcu_preempt_boost_start_gp(rnp); |
1894 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1895 | rnp->level, rnp->grplo, | |
1896 | rnp->grphi, rnp->qsmask); | |
1897 | raw_spin_unlock_irq(&rnp->lock); | |
bde6c3aa | 1898 | cond_resched_rcu_qs(); |
7d0ae808 | 1899 | WRITE_ONCE(rsp->gp_activity, jiffies); |
7fdefc10 | 1900 | } |
b3dbec76 | 1901 | |
7fdefc10 PM |
1902 | return 1; |
1903 | } | |
b3dbec76 | 1904 | |
4cdfc175 PM |
1905 | /* |
1906 | * Do one round of quiescent-state forcing. | |
1907 | */ | |
01896f7e | 1908 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) |
4cdfc175 PM |
1909 | { |
1910 | int fqs_state = fqs_state_in; | |
217af2a2 PM |
1911 | bool isidle = false; |
1912 | unsigned long maxj; | |
4cdfc175 PM |
1913 | struct rcu_node *rnp = rcu_get_root(rsp); |
1914 | ||
7d0ae808 | 1915 | WRITE_ONCE(rsp->gp_activity, jiffies); |
4cdfc175 PM |
1916 | rsp->n_force_qs++; |
1917 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1918 | /* Collect dyntick-idle snapshots. */ | |
0edd1b17 | 1919 | if (is_sysidle_rcu_state(rsp)) { |
e02b2edf | 1920 | isidle = true; |
0edd1b17 PM |
1921 | maxj = jiffies - ULONG_MAX / 4; |
1922 | } | |
217af2a2 PM |
1923 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1924 | &isidle, &maxj); | |
0edd1b17 | 1925 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1926 | fqs_state = RCU_FORCE_QS; |
1927 | } else { | |
1928 | /* Handle dyntick-idle and offline CPUs. */ | |
675da67f | 1929 | isidle = true; |
217af2a2 | 1930 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1931 | } |
1932 | /* Clear flag to prevent immediate re-entry. */ | |
7d0ae808 | 1933 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { |
4cdfc175 | 1934 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1935 | smp_mb__after_unlock_lock(); |
7d0ae808 PM |
1936 | WRITE_ONCE(rsp->gp_flags, |
1937 | READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS); | |
4cdfc175 PM |
1938 | raw_spin_unlock_irq(&rnp->lock); |
1939 | } | |
1940 | return fqs_state; | |
1941 | } | |
1942 | ||
7fdefc10 PM |
1943 | /* |
1944 | * Clean up after the old grace period. | |
1945 | */ | |
4cdfc175 | 1946 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1947 | { |
1948 | unsigned long gp_duration; | |
48a7639c | 1949 | bool needgp = false; |
dae6e64d | 1950 | int nocb = 0; |
7fdefc10 PM |
1951 | struct rcu_data *rdp; |
1952 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1953 | |
7d0ae808 | 1954 | WRITE_ONCE(rsp->gp_activity, jiffies); |
7fdefc10 | 1955 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1956 | smp_mb__after_unlock_lock(); |
7fdefc10 PM |
1957 | gp_duration = jiffies - rsp->gp_start; |
1958 | if (gp_duration > rsp->gp_max) | |
1959 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1960 | |
7fdefc10 PM |
1961 | /* |
1962 | * We know the grace period is complete, but to everyone else | |
1963 | * it appears to still be ongoing. But it is also the case | |
1964 | * that to everyone else it looks like there is nothing that | |
1965 | * they can do to advance the grace period. It is therefore | |
1966 | * safe for us to drop the lock in order to mark the grace | |
1967 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1968 | */ |
5d4b8659 | 1969 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1970 | |
5d4b8659 PM |
1971 | /* |
1972 | * Propagate new ->completed value to rcu_node structures so | |
1973 | * that other CPUs don't have to wait until the start of the next | |
1974 | * grace period to process their callbacks. This also avoids | |
1975 | * some nasty RCU grace-period initialization races by forcing | |
1976 | * the end of the current grace period to be completely recorded in | |
1977 | * all of the rcu_node structures before the beginning of the next | |
1978 | * grace period is recorded in any of the rcu_node structures. | |
1979 | */ | |
1980 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1981 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1982 | smp_mb__after_unlock_lock(); |
5c60d25f PM |
1983 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
1984 | WARN_ON_ONCE(rnp->qsmask); | |
7d0ae808 | 1985 | WRITE_ONCE(rnp->completed, rsp->gpnum); |
b11cc576 PM |
1986 | rdp = this_cpu_ptr(rsp->rda); |
1987 | if (rnp == rdp->mynode) | |
48a7639c | 1988 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; |
78e4bc34 | 1989 | /* smp_mb() provided by prior unlock-lock pair. */ |
0446be48 | 1990 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 | 1991 | raw_spin_unlock_irq(&rnp->lock); |
bde6c3aa | 1992 | cond_resched_rcu_qs(); |
7d0ae808 | 1993 | WRITE_ONCE(rsp->gp_activity, jiffies); |
0f41c0dd | 1994 | rcu_gp_slow(rsp, gp_cleanup_delay); |
7fdefc10 | 1995 | } |
5d4b8659 PM |
1996 | rnp = rcu_get_root(rsp); |
1997 | raw_spin_lock_irq(&rnp->lock); | |
765a3f4f | 1998 | smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */ |
dae6e64d | 1999 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 | 2000 | |
765a3f4f | 2001 | /* Declare grace period done. */ |
7d0ae808 | 2002 | WRITE_ONCE(rsp->completed, rsp->gpnum); |
f7f7bac9 | 2003 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
7fdefc10 | 2004 | rsp->fqs_state = RCU_GP_IDLE; |
5d4b8659 | 2005 | rdp = this_cpu_ptr(rsp->rda); |
48a7639c PM |
2006 | /* Advance CBs to reduce false positives below. */ |
2007 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; | |
2008 | if (needgp || cpu_needs_another_gp(rsp, rdp)) { | |
7d0ae808 | 2009 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); |
bb311ecc | 2010 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2011 | READ_ONCE(rsp->gpnum), |
bb311ecc PM |
2012 | TPS("newreq")); |
2013 | } | |
7fdefc10 | 2014 | raw_spin_unlock_irq(&rnp->lock); |
7fdefc10 PM |
2015 | } |
2016 | ||
2017 | /* | |
2018 | * Body of kthread that handles grace periods. | |
2019 | */ | |
2020 | static int __noreturn rcu_gp_kthread(void *arg) | |
2021 | { | |
4cdfc175 | 2022 | int fqs_state; |
88d6df61 | 2023 | int gf; |
d40011f6 | 2024 | unsigned long j; |
4cdfc175 | 2025 | int ret; |
7fdefc10 PM |
2026 | struct rcu_state *rsp = arg; |
2027 | struct rcu_node *rnp = rcu_get_root(rsp); | |
2028 | ||
5871968d | 2029 | rcu_bind_gp_kthread(); |
7fdefc10 PM |
2030 | for (;;) { |
2031 | ||
2032 | /* Handle grace-period start. */ | |
2033 | for (;;) { | |
63c4db78 | 2034 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2035 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2036 | TPS("reqwait")); |
afea227f | 2037 | rsp->gp_state = RCU_GP_WAIT_GPS; |
4cdfc175 | 2038 | wait_event_interruptible(rsp->gp_wq, |
7d0ae808 | 2039 | READ_ONCE(rsp->gp_flags) & |
4cdfc175 | 2040 | RCU_GP_FLAG_INIT); |
319362c9 | 2041 | rsp->gp_state = RCU_GP_DONE_GPS; |
78e4bc34 | 2042 | /* Locking provides needed memory barrier. */ |
f7be8209 | 2043 | if (rcu_gp_init(rsp)) |
7fdefc10 | 2044 | break; |
bde6c3aa | 2045 | cond_resched_rcu_qs(); |
7d0ae808 | 2046 | WRITE_ONCE(rsp->gp_activity, jiffies); |
73a860cd | 2047 | WARN_ON(signal_pending(current)); |
63c4db78 | 2048 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2049 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2050 | TPS("reqwaitsig")); |
7fdefc10 | 2051 | } |
cabc49c1 | 2052 | |
4cdfc175 PM |
2053 | /* Handle quiescent-state forcing. */ |
2054 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
2055 | j = jiffies_till_first_fqs; |
2056 | if (j > HZ) { | |
2057 | j = HZ; | |
2058 | jiffies_till_first_fqs = HZ; | |
2059 | } | |
88d6df61 | 2060 | ret = 0; |
cabc49c1 | 2061 | for (;;) { |
88d6df61 PM |
2062 | if (!ret) |
2063 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 | 2064 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2065 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2066 | TPS("fqswait")); |
afea227f | 2067 | rsp->gp_state = RCU_GP_WAIT_FQS; |
4cdfc175 | 2068 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
7d0ae808 | 2069 | ((gf = READ_ONCE(rsp->gp_flags)) & |
88d6df61 | 2070 | RCU_GP_FLAG_FQS) || |
7d0ae808 | 2071 | (!READ_ONCE(rnp->qsmask) && |
4cdfc175 | 2072 | !rcu_preempt_blocked_readers_cgp(rnp)), |
d40011f6 | 2073 | j); |
319362c9 | 2074 | rsp->gp_state = RCU_GP_DONE_FQS; |
78e4bc34 | 2075 | /* Locking provides needed memory barriers. */ |
4cdfc175 | 2076 | /* If grace period done, leave loop. */ |
7d0ae808 | 2077 | if (!READ_ONCE(rnp->qsmask) && |
4cdfc175 | 2078 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 2079 | break; |
4cdfc175 | 2080 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
2081 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
2082 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 | 2083 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2084 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2085 | TPS("fqsstart")); |
4cdfc175 | 2086 | fqs_state = rcu_gp_fqs(rsp, fqs_state); |
63c4db78 | 2087 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2088 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2089 | TPS("fqsend")); |
bde6c3aa | 2090 | cond_resched_rcu_qs(); |
7d0ae808 | 2091 | WRITE_ONCE(rsp->gp_activity, jiffies); |
4cdfc175 PM |
2092 | } else { |
2093 | /* Deal with stray signal. */ | |
bde6c3aa | 2094 | cond_resched_rcu_qs(); |
7d0ae808 | 2095 | WRITE_ONCE(rsp->gp_activity, jiffies); |
73a860cd | 2096 | WARN_ON(signal_pending(current)); |
63c4db78 | 2097 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2098 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2099 | TPS("fqswaitsig")); |
4cdfc175 | 2100 | } |
d40011f6 PM |
2101 | j = jiffies_till_next_fqs; |
2102 | if (j > HZ) { | |
2103 | j = HZ; | |
2104 | jiffies_till_next_fqs = HZ; | |
2105 | } else if (j < 1) { | |
2106 | j = 1; | |
2107 | jiffies_till_next_fqs = 1; | |
2108 | } | |
cabc49c1 | 2109 | } |
4cdfc175 PM |
2110 | |
2111 | /* Handle grace-period end. */ | |
319362c9 | 2112 | rsp->gp_state = RCU_GP_CLEANUP; |
4cdfc175 | 2113 | rcu_gp_cleanup(rsp); |
319362c9 | 2114 | rsp->gp_state = RCU_GP_CLEANED; |
b3dbec76 | 2115 | } |
b3dbec76 PM |
2116 | } |
2117 | ||
64db4cff PM |
2118 | /* |
2119 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
2120 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 2121 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
2122 | * |
2123 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
2124 | * invoke this function. This can happen when the dying CPU reports its | |
2125 | * quiescent state. | |
48a7639c PM |
2126 | * |
2127 | * Returns true if the grace-period kthread must be awakened. | |
64db4cff | 2128 | */ |
48a7639c | 2129 | static bool |
910ee45d PM |
2130 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
2131 | struct rcu_data *rdp) | |
64db4cff | 2132 | { |
b8462084 | 2133 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 2134 | /* |
b3dbec76 | 2135 | * Either we have not yet spawned the grace-period |
62da1921 PM |
2136 | * task, this CPU does not need another grace period, |
2137 | * or a grace period is already in progress. | |
b3dbec76 | 2138 | * Either way, don't start a new grace period. |
afe24b12 | 2139 | */ |
48a7639c | 2140 | return false; |
afe24b12 | 2141 | } |
7d0ae808 PM |
2142 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); |
2143 | trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum), | |
bb311ecc | 2144 | TPS("newreq")); |
62da1921 | 2145 | |
016a8d5b SR |
2146 | /* |
2147 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c | 2148 | * could cause possible deadlocks with the rq->lock. Defer |
48a7639c | 2149 | * the wakeup to our caller. |
016a8d5b | 2150 | */ |
48a7639c | 2151 | return true; |
64db4cff PM |
2152 | } |
2153 | ||
910ee45d PM |
2154 | /* |
2155 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
2156 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
2157 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
2158 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
2159 | * that is encountered beforehand. | |
48a7639c PM |
2160 | * |
2161 | * Returns true if the grace-period kthread needs to be awakened. | |
910ee45d | 2162 | */ |
48a7639c | 2163 | static bool rcu_start_gp(struct rcu_state *rsp) |
910ee45d PM |
2164 | { |
2165 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
2166 | struct rcu_node *rnp = rcu_get_root(rsp); | |
48a7639c | 2167 | bool ret = false; |
910ee45d PM |
2168 | |
2169 | /* | |
2170 | * If there is no grace period in progress right now, any | |
2171 | * callbacks we have up to this point will be satisfied by the | |
2172 | * next grace period. Also, advancing the callbacks reduces the | |
2173 | * probability of false positives from cpu_needs_another_gp() | |
2174 | * resulting in pointless grace periods. So, advance callbacks | |
2175 | * then start the grace period! | |
2176 | */ | |
48a7639c PM |
2177 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; |
2178 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; | |
2179 | return ret; | |
910ee45d PM |
2180 | } |
2181 | ||
f41d911f | 2182 | /* |
d3f6bad3 PM |
2183 | * Report a full set of quiescent states to the specified rcu_state |
2184 | * data structure. This involves cleaning up after the prior grace | |
2185 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
2186 | * if one is needed. Note that the caller must hold rnp->lock, which |
2187 | * is released before return. | |
f41d911f | 2188 | */ |
d3f6bad3 | 2189 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 2190 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 2191 | { |
fc2219d4 | 2192 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cd73ca21 | 2193 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); |
cabc49c1 | 2194 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
2aa792e6 | 2195 | rcu_gp_kthread_wake(rsp); |
f41d911f PM |
2196 | } |
2197 | ||
64db4cff | 2198 | /* |
d3f6bad3 PM |
2199 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
2200 | * Allows quiescent states for a group of CPUs to be reported at one go | |
2201 | * to the specified rcu_node structure, though all the CPUs in the group | |
654e9533 PM |
2202 | * must be represented by the same rcu_node structure (which need not be a |
2203 | * leaf rcu_node structure, though it often will be). The gps parameter | |
2204 | * is the grace-period snapshot, which means that the quiescent states | |
2205 | * are valid only if rnp->gpnum is equal to gps. That structure's lock | |
2206 | * must be held upon entry, and it is released before return. | |
64db4cff PM |
2207 | */ |
2208 | static void | |
d3f6bad3 | 2209 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
654e9533 | 2210 | struct rcu_node *rnp, unsigned long gps, unsigned long flags) |
64db4cff PM |
2211 | __releases(rnp->lock) |
2212 | { | |
654e9533 | 2213 | unsigned long oldmask = 0; |
28ecd580 PM |
2214 | struct rcu_node *rnp_c; |
2215 | ||
64db4cff PM |
2216 | /* Walk up the rcu_node hierarchy. */ |
2217 | for (;;) { | |
654e9533 | 2218 | if (!(rnp->qsmask & mask) || rnp->gpnum != gps) { |
64db4cff | 2219 | |
654e9533 PM |
2220 | /* |
2221 | * Our bit has already been cleared, or the | |
2222 | * relevant grace period is already over, so done. | |
2223 | */ | |
1304afb2 | 2224 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2225 | return; |
2226 | } | |
654e9533 | 2227 | WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ |
64db4cff | 2228 | rnp->qsmask &= ~mask; |
d4c08f2a PM |
2229 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
2230 | mask, rnp->qsmask, rnp->level, | |
2231 | rnp->grplo, rnp->grphi, | |
2232 | !!rnp->gp_tasks); | |
27f4d280 | 2233 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
2234 | |
2235 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 2236 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2237 | return; |
2238 | } | |
2239 | mask = rnp->grpmask; | |
2240 | if (rnp->parent == NULL) { | |
2241 | ||
2242 | /* No more levels. Exit loop holding root lock. */ | |
2243 | ||
2244 | break; | |
2245 | } | |
1304afb2 | 2246 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 2247 | rnp_c = rnp; |
64db4cff | 2248 | rnp = rnp->parent; |
1304afb2 | 2249 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2250 | smp_mb__after_unlock_lock(); |
654e9533 | 2251 | oldmask = rnp_c->qsmask; |
64db4cff PM |
2252 | } |
2253 | ||
2254 | /* | |
2255 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 2256 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 2257 | * to clean up and start the next grace period if one is needed. |
64db4cff | 2258 | */ |
d3f6bad3 | 2259 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
2260 | } |
2261 | ||
cc99a310 PM |
2262 | /* |
2263 | * Record a quiescent state for all tasks that were previously queued | |
2264 | * on the specified rcu_node structure and that were blocking the current | |
2265 | * RCU grace period. The caller must hold the specified rnp->lock with | |
2266 | * irqs disabled, and this lock is released upon return, but irqs remain | |
2267 | * disabled. | |
2268 | */ | |
0aa04b05 | 2269 | static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp, |
cc99a310 PM |
2270 | struct rcu_node *rnp, unsigned long flags) |
2271 | __releases(rnp->lock) | |
2272 | { | |
654e9533 | 2273 | unsigned long gps; |
cc99a310 PM |
2274 | unsigned long mask; |
2275 | struct rcu_node *rnp_p; | |
2276 | ||
a77da14c PM |
2277 | if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p || |
2278 | rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { | |
cc99a310 PM |
2279 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
2280 | return; /* Still need more quiescent states! */ | |
2281 | } | |
2282 | ||
2283 | rnp_p = rnp->parent; | |
2284 | if (rnp_p == NULL) { | |
2285 | /* | |
a77da14c PM |
2286 | * Only one rcu_node structure in the tree, so don't |
2287 | * try to report up to its nonexistent parent! | |
cc99a310 PM |
2288 | */ |
2289 | rcu_report_qs_rsp(rsp, flags); | |
2290 | return; | |
2291 | } | |
2292 | ||
654e9533 PM |
2293 | /* Report up the rest of the hierarchy, tracking current ->gpnum. */ |
2294 | gps = rnp->gpnum; | |
cc99a310 PM |
2295 | mask = rnp->grpmask; |
2296 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2297 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
2298 | smp_mb__after_unlock_lock(); | |
654e9533 | 2299 | rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags); |
cc99a310 PM |
2300 | } |
2301 | ||
64db4cff | 2302 | /* |
d3f6bad3 PM |
2303 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
2304 | * structure. This must be either called from the specified CPU, or | |
2305 | * called when the specified CPU is known to be offline (and when it is | |
2306 | * also known that no other CPU is concurrently trying to help the offline | |
2307 | * CPU). The lastcomp argument is used to make sure we are still in the | |
2308 | * grace period of interest. We don't want to end the current grace period | |
2309 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
2310 | */ |
2311 | static void | |
d7d6a11e | 2312 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2313 | { |
2314 | unsigned long flags; | |
2315 | unsigned long mask; | |
48a7639c | 2316 | bool needwake; |
64db4cff PM |
2317 | struct rcu_node *rnp; |
2318 | ||
2319 | rnp = rdp->mynode; | |
1304afb2 | 2320 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2321 | smp_mb__after_unlock_lock(); |
5cd37193 PM |
2322 | if ((rdp->passed_quiesce == 0 && |
2323 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) || | |
2324 | rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum || | |
2325 | rdp->gpwrap) { | |
64db4cff PM |
2326 | |
2327 | /* | |
e4cc1f22 PM |
2328 | * The grace period in which this quiescent state was |
2329 | * recorded has ended, so don't report it upwards. | |
2330 | * We will instead need a new quiescent state that lies | |
2331 | * within the current grace period. | |
64db4cff | 2332 | */ |
e4cc1f22 | 2333 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
5cd37193 | 2334 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); |
1304afb2 | 2335 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2336 | return; |
2337 | } | |
2338 | mask = rdp->grpmask; | |
2339 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 2340 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2341 | } else { |
2342 | rdp->qs_pending = 0; | |
2343 | ||
2344 | /* | |
2345 | * This GP can't end until cpu checks in, so all of our | |
2346 | * callbacks can be processed during the next GP. | |
2347 | */ | |
48a7639c | 2348 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 2349 | |
654e9533 PM |
2350 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); |
2351 | /* ^^^ Released rnp->lock */ | |
48a7639c PM |
2352 | if (needwake) |
2353 | rcu_gp_kthread_wake(rsp); | |
64db4cff PM |
2354 | } |
2355 | } | |
2356 | ||
2357 | /* | |
2358 | * Check to see if there is a new grace period of which this CPU | |
2359 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2360 | * Otherwise, see if this CPU has just passed through its first | |
2361 | * quiescent state for this grace period, and record that fact if so. | |
2362 | */ | |
2363 | static void | |
2364 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
2365 | { | |
05eb552b PM |
2366 | /* Check for grace-period ends and beginnings. */ |
2367 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
2368 | |
2369 | /* | |
2370 | * Does this CPU still need to do its part for current grace period? | |
2371 | * If no, return and let the other CPUs do their part as well. | |
2372 | */ | |
2373 | if (!rdp->qs_pending) | |
2374 | return; | |
2375 | ||
2376 | /* | |
2377 | * Was there a quiescent state since the beginning of the grace | |
2378 | * period? If no, then exit and wait for the next call. | |
2379 | */ | |
5cd37193 PM |
2380 | if (!rdp->passed_quiesce && |
2381 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) | |
64db4cff PM |
2382 | return; |
2383 | ||
d3f6bad3 PM |
2384 | /* |
2385 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2386 | * judge of that). | |
2387 | */ | |
d7d6a11e | 2388 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
2389 | } |
2390 | ||
e74f4c45 | 2391 | /* |
b1420f1c PM |
2392 | * Send the specified CPU's RCU callbacks to the orphanage. The |
2393 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 2394 | * ->orphan_lock. |
e74f4c45 | 2395 | */ |
b1420f1c PM |
2396 | static void |
2397 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
2398 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 2399 | { |
3fbfbf7a | 2400 | /* No-CBs CPUs do not have orphanable callbacks. */ |
ea46351c | 2401 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
2402 | return; |
2403 | ||
b1420f1c PM |
2404 | /* |
2405 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
2406 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
2407 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 2408 | */ |
a50c3af9 | 2409 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2410 | rsp->qlen_lazy += rdp->qlen_lazy; |
2411 | rsp->qlen += rdp->qlen; | |
2412 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 2413 | rdp->qlen_lazy = 0; |
7d0ae808 | 2414 | WRITE_ONCE(rdp->qlen, 0); |
a50c3af9 PM |
2415 | } |
2416 | ||
2417 | /* | |
b1420f1c PM |
2418 | * Next, move those callbacks still needing a grace period to |
2419 | * the orphanage, where some other CPU will pick them up. | |
2420 | * Some of the callbacks might have gone partway through a grace | |
2421 | * period, but that is too bad. They get to start over because we | |
2422 | * cannot assume that grace periods are synchronized across CPUs. | |
2423 | * We don't bother updating the ->nxttail[] array yet, instead | |
2424 | * we just reset the whole thing later on. | |
a50c3af9 | 2425 | */ |
b1420f1c PM |
2426 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
2427 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2428 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2429 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
2430 | } |
2431 | ||
2432 | /* | |
b1420f1c PM |
2433 | * Then move the ready-to-invoke callbacks to the orphanage, |
2434 | * where some other CPU will pick them up. These will not be | |
2435 | * required to pass though another grace period: They are done. | |
a50c3af9 | 2436 | */ |
e5601400 | 2437 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2438 | *rsp->orphan_donetail = rdp->nxtlist; |
2439 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 2440 | } |
e74f4c45 | 2441 | |
b33078b6 PM |
2442 | /* |
2443 | * Finally, initialize the rcu_data structure's list to empty and | |
2444 | * disallow further callbacks on this CPU. | |
2445 | */ | |
3f5d3ea6 | 2446 | init_callback_list(rdp); |
b33078b6 | 2447 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
b1420f1c PM |
2448 | } |
2449 | ||
2450 | /* | |
2451 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 2452 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c | 2453 | */ |
96d3fd0d | 2454 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) |
b1420f1c PM |
2455 | { |
2456 | int i; | |
fa07a58f | 2457 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
b1420f1c | 2458 | |
3fbfbf7a | 2459 | /* No-CBs CPUs are handled specially. */ |
ea46351c PM |
2460 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || |
2461 | rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) | |
3fbfbf7a PM |
2462 | return; |
2463 | ||
b1420f1c PM |
2464 | /* Do the accounting first. */ |
2465 | rdp->qlen_lazy += rsp->qlen_lazy; | |
2466 | rdp->qlen += rsp->qlen; | |
2467 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
2468 | if (rsp->qlen_lazy != rsp->qlen) |
2469 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2470 | rsp->qlen_lazy = 0; |
2471 | rsp->qlen = 0; | |
2472 | ||
2473 | /* | |
2474 | * We do not need a memory barrier here because the only way we | |
2475 | * can get here if there is an rcu_barrier() in flight is if | |
2476 | * we are the task doing the rcu_barrier(). | |
2477 | */ | |
2478 | ||
2479 | /* First adopt the ready-to-invoke callbacks. */ | |
2480 | if (rsp->orphan_donelist != NULL) { | |
2481 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2482 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
2483 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
2484 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2485 | rdp->nxttail[i] = rsp->orphan_donetail; | |
2486 | rsp->orphan_donelist = NULL; | |
2487 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2488 | } | |
2489 | ||
2490 | /* And then adopt the callbacks that still need a grace period. */ | |
2491 | if (rsp->orphan_nxtlist != NULL) { | |
2492 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
2493 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
2494 | rsp->orphan_nxtlist = NULL; | |
2495 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2496 | } | |
2497 | } | |
2498 | ||
2499 | /* | |
2500 | * Trace the fact that this CPU is going offline. | |
2501 | */ | |
2502 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
2503 | { | |
2504 | RCU_TRACE(unsigned long mask); | |
2505 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
2506 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
2507 | ||
ea46351c PM |
2508 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2509 | return; | |
2510 | ||
b1420f1c | 2511 | RCU_TRACE(mask = rdp->grpmask); |
e5601400 PM |
2512 | trace_rcu_grace_period(rsp->name, |
2513 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 2514 | TPS("cpuofl")); |
64db4cff PM |
2515 | } |
2516 | ||
8af3a5e7 PM |
2517 | /* |
2518 | * All CPUs for the specified rcu_node structure have gone offline, | |
2519 | * and all tasks that were preempted within an RCU read-side critical | |
2520 | * section while running on one of those CPUs have since exited their RCU | |
2521 | * read-side critical section. Some other CPU is reporting this fact with | |
2522 | * the specified rcu_node structure's ->lock held and interrupts disabled. | |
2523 | * This function therefore goes up the tree of rcu_node structures, | |
2524 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that | |
2525 | * the leaf rcu_node structure's ->qsmaskinit field has already been | |
2526 | * updated | |
2527 | * | |
2528 | * This function does check that the specified rcu_node structure has | |
2529 | * all CPUs offline and no blocked tasks, so it is OK to invoke it | |
2530 | * prematurely. That said, invoking it after the fact will cost you | |
2531 | * a needless lock acquisition. So once it has done its work, don't | |
2532 | * invoke it again. | |
2533 | */ | |
2534 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) | |
2535 | { | |
2536 | long mask; | |
2537 | struct rcu_node *rnp = rnp_leaf; | |
2538 | ||
ea46351c PM |
2539 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || |
2540 | rnp->qsmaskinit || rcu_preempt_has_tasks(rnp)) | |
8af3a5e7 PM |
2541 | return; |
2542 | for (;;) { | |
2543 | mask = rnp->grpmask; | |
2544 | rnp = rnp->parent; | |
2545 | if (!rnp) | |
2546 | break; | |
2547 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
2548 | smp_mb__after_unlock_lock(); /* GP memory ordering. */ | |
2549 | rnp->qsmaskinit &= ~mask; | |
0aa04b05 | 2550 | rnp->qsmask &= ~mask; |
8af3a5e7 PM |
2551 | if (rnp->qsmaskinit) { |
2552 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2553 | return; | |
2554 | } | |
2555 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2556 | } | |
2557 | } | |
2558 | ||
88428cc5 PM |
2559 | /* |
2560 | * The CPU is exiting the idle loop into the arch_cpu_idle_dead() | |
2561 | * function. We now remove it from the rcu_node tree's ->qsmaskinit | |
2562 | * bit masks. | |
2563 | */ | |
2564 | static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp) | |
2565 | { | |
2566 | unsigned long flags; | |
2567 | unsigned long mask; | |
2568 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
2569 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ | |
2570 | ||
ea46351c PM |
2571 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2572 | return; | |
2573 | ||
88428cc5 PM |
2574 | /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ |
2575 | mask = rdp->grpmask; | |
2576 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2577 | smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */ | |
2578 | rnp->qsmaskinitnext &= ~mask; | |
2579 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2580 | } | |
2581 | ||
64db4cff | 2582 | /* |
e5601400 | 2583 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
2584 | * this fact from process context. Do the remainder of the cleanup, |
2585 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
2586 | * adopting them. There can only be one CPU hotplug operation at a time, |
2587 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 2588 | */ |
e5601400 | 2589 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 2590 | { |
2036d94a | 2591 | unsigned long flags; |
e5601400 | 2592 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 2593 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2594 | |
ea46351c PM |
2595 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2596 | return; | |
2597 | ||
2036d94a | 2598 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2599 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 2600 | |
b1420f1c | 2601 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
78043c46 | 2602 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
b1420f1c | 2603 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); |
96d3fd0d | 2604 | rcu_adopt_orphan_cbs(rsp, flags); |
a8f4cbad | 2605 | raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags); |
b1420f1c | 2606 | |
cf01537e PM |
2607 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
2608 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2609 | cpu, rdp->qlen, rdp->nxtlist); | |
64db4cff PM |
2610 | } |
2611 | ||
64db4cff PM |
2612 | /* |
2613 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2614 | * period. Thottle as specified by rdp->blimit. | |
2615 | */ | |
37c72e56 | 2616 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2617 | { |
2618 | unsigned long flags; | |
2619 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
2620 | long bl, count, count_lazy; |
2621 | int i; | |
64db4cff | 2622 | |
dc35c893 | 2623 | /* If no callbacks are ready, just return. */ |
29c00b4a | 2624 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 2625 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
7d0ae808 | 2626 | trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist), |
4968c300 PM |
2627 | need_resched(), is_idle_task(current), |
2628 | rcu_is_callbacks_kthread()); | |
64db4cff | 2629 | return; |
29c00b4a | 2630 | } |
64db4cff PM |
2631 | |
2632 | /* | |
2633 | * Extract the list of ready callbacks, disabling to prevent | |
2634 | * races with call_rcu() from interrupt handlers. | |
2635 | */ | |
2636 | local_irq_save(flags); | |
8146c4e2 | 2637 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2638 | bl = rdp->blimit; |
486e2593 | 2639 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2640 | list = rdp->nxtlist; |
2641 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2642 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2643 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2644 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2645 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2646 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2647 | local_irq_restore(flags); |
2648 | ||
2649 | /* Invoke callbacks. */ | |
486e2593 | 2650 | count = count_lazy = 0; |
64db4cff PM |
2651 | while (list) { |
2652 | next = list->next; | |
2653 | prefetch(next); | |
551d55a9 | 2654 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2655 | if (__rcu_reclaim(rsp->name, list)) |
2656 | count_lazy++; | |
64db4cff | 2657 | list = next; |
dff1672d PM |
2658 | /* Stop only if limit reached and CPU has something to do. */ |
2659 | if (++count >= bl && | |
2660 | (need_resched() || | |
2661 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2662 | break; |
2663 | } | |
2664 | ||
2665 | local_irq_save(flags); | |
4968c300 PM |
2666 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2667 | is_idle_task(current), | |
2668 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2669 | |
2670 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2671 | if (list != NULL) { |
2672 | *tail = rdp->nxtlist; | |
2673 | rdp->nxtlist = list; | |
b41772ab PM |
2674 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2675 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2676 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2677 | else |
2678 | break; | |
2679 | } | |
b1420f1c PM |
2680 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2681 | rdp->qlen_lazy -= count_lazy; | |
7d0ae808 | 2682 | WRITE_ONCE(rdp->qlen, rdp->qlen - count); |
b1420f1c | 2683 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2684 | |
2685 | /* Reinstate batch limit if we have worked down the excess. */ | |
2686 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2687 | rdp->blimit = blimit; | |
2688 | ||
37c72e56 PM |
2689 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2690 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2691 | rdp->qlen_last_fqs_check = 0; | |
2692 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2693 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2694 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2695 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2696 | |
64db4cff PM |
2697 | local_irq_restore(flags); |
2698 | ||
e0f23060 | 2699 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2700 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2701 | invoke_rcu_core(); |
64db4cff PM |
2702 | } |
2703 | ||
2704 | /* | |
2705 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2706 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2707 | * Also schedule RCU core processing. |
64db4cff | 2708 | * |
9b2e4f18 | 2709 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2710 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2711 | * false, there is no point in invoking rcu_check_callbacks(). | |
2712 | */ | |
c3377c2d | 2713 | void rcu_check_callbacks(int user) |
64db4cff | 2714 | { |
f7f7bac9 | 2715 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2716 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2717 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2718 | |
2719 | /* | |
2720 | * Get here if this CPU took its interrupt from user | |
2721 | * mode or from the idle loop, and if this is not a | |
2722 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2723 | * a quiescent state, so note it. |
64db4cff PM |
2724 | * |
2725 | * No memory barrier is required here because both | |
d6714c22 PM |
2726 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2727 | * variables that other CPUs neither access nor modify, | |
2728 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2729 | */ |
2730 | ||
284a8c93 PM |
2731 | rcu_sched_qs(); |
2732 | rcu_bh_qs(); | |
64db4cff PM |
2733 | |
2734 | } else if (!in_softirq()) { | |
2735 | ||
2736 | /* | |
2737 | * Get here if this CPU did not take its interrupt from | |
2738 | * softirq, in other words, if it is not interrupting | |
2739 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2740 | * critical section, so note it. |
64db4cff PM |
2741 | */ |
2742 | ||
284a8c93 | 2743 | rcu_bh_qs(); |
64db4cff | 2744 | } |
86aea0e6 | 2745 | rcu_preempt_check_callbacks(); |
e3950ecd | 2746 | if (rcu_pending()) |
a46e0899 | 2747 | invoke_rcu_core(); |
8315f422 PM |
2748 | if (user) |
2749 | rcu_note_voluntary_context_switch(current); | |
f7f7bac9 | 2750 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2751 | } |
2752 | ||
64db4cff PM |
2753 | /* |
2754 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2755 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2756 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2757 | * | |
ee47eb9f | 2758 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2759 | */ |
217af2a2 PM |
2760 | static void force_qs_rnp(struct rcu_state *rsp, |
2761 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2762 | unsigned long *maxj), | |
2763 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2764 | { |
2765 | unsigned long bit; | |
2766 | int cpu; | |
2767 | unsigned long flags; | |
2768 | unsigned long mask; | |
a0b6c9a7 | 2769 | struct rcu_node *rnp; |
64db4cff | 2770 | |
a0b6c9a7 | 2771 | rcu_for_each_leaf_node(rsp, rnp) { |
bde6c3aa | 2772 | cond_resched_rcu_qs(); |
64db4cff | 2773 | mask = 0; |
1304afb2 | 2774 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2775 | smp_mb__after_unlock_lock(); |
a0b6c9a7 | 2776 | if (rnp->qsmask == 0) { |
a77da14c PM |
2777 | if (rcu_state_p == &rcu_sched_state || |
2778 | rsp != rcu_state_p || | |
2779 | rcu_preempt_blocked_readers_cgp(rnp)) { | |
2780 | /* | |
2781 | * No point in scanning bits because they | |
2782 | * are all zero. But we might need to | |
2783 | * priority-boost blocked readers. | |
2784 | */ | |
2785 | rcu_initiate_boost(rnp, flags); | |
2786 | /* rcu_initiate_boost() releases rnp->lock */ | |
2787 | continue; | |
2788 | } | |
2789 | if (rnp->parent && | |
2790 | (rnp->parent->qsmask & rnp->grpmask)) { | |
2791 | /* | |
2792 | * Race between grace-period | |
2793 | * initialization and task exiting RCU | |
2794 | * read-side critical section: Report. | |
2795 | */ | |
2796 | rcu_report_unblock_qs_rnp(rsp, rnp, flags); | |
2797 | /* rcu_report_unblock_qs_rnp() rlses ->lock */ | |
2798 | continue; | |
2799 | } | |
64db4cff | 2800 | } |
a0b6c9a7 | 2801 | cpu = rnp->grplo; |
64db4cff | 2802 | bit = 1; |
a0b6c9a7 | 2803 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 | 2804 | if ((rnp->qsmask & bit) != 0) { |
0edd1b17 PM |
2805 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) |
2806 | mask |= bit; | |
2807 | } | |
64db4cff | 2808 | } |
45f014c5 | 2809 | if (mask != 0) { |
654e9533 PM |
2810 | /* Idle/offline CPUs, report (releases rnp->lock. */ |
2811 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); | |
0aa04b05 PM |
2812 | } else { |
2813 | /* Nothing to do here, so just drop the lock. */ | |
2814 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff | 2815 | } |
64db4cff | 2816 | } |
64db4cff PM |
2817 | } |
2818 | ||
2819 | /* | |
2820 | * Force quiescent states on reluctant CPUs, and also detect which | |
2821 | * CPUs are in dyntick-idle mode. | |
2822 | */ | |
4cdfc175 | 2823 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2824 | { |
2825 | unsigned long flags; | |
394f2769 PM |
2826 | bool ret; |
2827 | struct rcu_node *rnp; | |
2828 | struct rcu_node *rnp_old = NULL; | |
2829 | ||
2830 | /* Funnel through hierarchy to reduce memory contention. */ | |
d860d403 | 2831 | rnp = __this_cpu_read(rsp->rda->mynode); |
394f2769 | 2832 | for (; rnp != NULL; rnp = rnp->parent) { |
7d0ae808 | 2833 | ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || |
394f2769 PM |
2834 | !raw_spin_trylock(&rnp->fqslock); |
2835 | if (rnp_old != NULL) | |
2836 | raw_spin_unlock(&rnp_old->fqslock); | |
2837 | if (ret) { | |
a792563b | 2838 | rsp->n_force_qs_lh++; |
394f2769 PM |
2839 | return; |
2840 | } | |
2841 | rnp_old = rnp; | |
2842 | } | |
2843 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2844 | |
394f2769 PM |
2845 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2846 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
6303b9c8 | 2847 | smp_mb__after_unlock_lock(); |
394f2769 | 2848 | raw_spin_unlock(&rnp_old->fqslock); |
7d0ae808 | 2849 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { |
a792563b | 2850 | rsp->n_force_qs_lh++; |
394f2769 | 2851 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2852 | return; /* Someone beat us to it. */ |
46a1e34e | 2853 | } |
7d0ae808 | 2854 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); |
394f2769 | 2855 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
2aa792e6 | 2856 | rcu_gp_kthread_wake(rsp); |
64db4cff PM |
2857 | } |
2858 | ||
64db4cff | 2859 | /* |
e0f23060 PM |
2860 | * This does the RCU core processing work for the specified rcu_state |
2861 | * and rcu_data structures. This may be called only from the CPU to | |
2862 | * whom the rdp belongs. | |
64db4cff PM |
2863 | */ |
2864 | static void | |
1bca8cf1 | 2865 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2866 | { |
2867 | unsigned long flags; | |
48a7639c | 2868 | bool needwake; |
fa07a58f | 2869 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
64db4cff | 2870 | |
2e597558 PM |
2871 | WARN_ON_ONCE(rdp->beenonline == 0); |
2872 | ||
64db4cff PM |
2873 | /* Update RCU state based on any recent quiescent states. */ |
2874 | rcu_check_quiescent_state(rsp, rdp); | |
2875 | ||
2876 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2877 | local_irq_save(flags); |
64db4cff | 2878 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2879 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
48a7639c | 2880 | needwake = rcu_start_gp(rsp); |
b8462084 | 2881 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
48a7639c PM |
2882 | if (needwake) |
2883 | rcu_gp_kthread_wake(rsp); | |
dc35c893 PM |
2884 | } else { |
2885 | local_irq_restore(flags); | |
64db4cff PM |
2886 | } |
2887 | ||
2888 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2889 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2890 | invoke_rcu_callbacks(rsp, rdp); |
96d3fd0d PM |
2891 | |
2892 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2893 | do_nocb_deferred_wakeup(rdp); | |
09223371 SL |
2894 | } |
2895 | ||
64db4cff | 2896 | /* |
e0f23060 | 2897 | * Do RCU core processing for the current CPU. |
64db4cff | 2898 | */ |
09223371 | 2899 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2900 | { |
6ce75a23 PM |
2901 | struct rcu_state *rsp; |
2902 | ||
bfa00b4c PM |
2903 | if (cpu_is_offline(smp_processor_id())) |
2904 | return; | |
f7f7bac9 | 2905 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2906 | for_each_rcu_flavor(rsp) |
2907 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2908 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2909 | } |
2910 | ||
a26ac245 | 2911 | /* |
e0f23060 PM |
2912 | * Schedule RCU callback invocation. If the specified type of RCU |
2913 | * does not support RCU priority boosting, just do a direct call, | |
2914 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
924df8a0 | 2915 | * are running on the current CPU with softirqs disabled, the |
e0f23060 | 2916 | * rcu_cpu_kthread_task cannot disappear out from under us. |
a26ac245 | 2917 | */ |
a46e0899 | 2918 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2919 | { |
7d0ae808 | 2920 | if (unlikely(!READ_ONCE(rcu_scheduler_fully_active))) |
b0d30417 | 2921 | return; |
a46e0899 PM |
2922 | if (likely(!rsp->boost)) { |
2923 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2924 | return; |
2925 | } | |
a46e0899 | 2926 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2927 | } |
2928 | ||
a46e0899 | 2929 | static void invoke_rcu_core(void) |
09223371 | 2930 | { |
b0f74036 PM |
2931 | if (cpu_online(smp_processor_id())) |
2932 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2933 | } |
2934 | ||
29154c57 PM |
2935 | /* |
2936 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2937 | */ | |
2938 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2939 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2940 | { |
48a7639c PM |
2941 | bool needwake; |
2942 | ||
62fde6ed PM |
2943 | /* |
2944 | * If called from an extended quiescent state, invoke the RCU | |
2945 | * core in order to force a re-evaluation of RCU's idleness. | |
2946 | */ | |
9910affa | 2947 | if (!rcu_is_watching()) |
62fde6ed PM |
2948 | invoke_rcu_core(); |
2949 | ||
a16b7a69 | 2950 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2951 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2952 | return; |
64db4cff | 2953 | |
37c72e56 PM |
2954 | /* |
2955 | * Force the grace period if too many callbacks or too long waiting. | |
2956 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2957 | * if some other CPU has recently done so. Also, don't bother | |
2958 | * invoking force_quiescent_state() if the newly enqueued callback | |
2959 | * is the only one waiting for a grace period to complete. | |
2960 | */ | |
2655d57e | 2961 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2962 | |
2963 | /* Are we ignoring a completed grace period? */ | |
470716fc | 2964 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
2965 | |
2966 | /* Start a new grace period if one not already started. */ | |
2967 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
2968 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
2969 | ||
b8462084 | 2970 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 | 2971 | smp_mb__after_unlock_lock(); |
48a7639c | 2972 | needwake = rcu_start_gp(rsp); |
b8462084 | 2973 | raw_spin_unlock(&rnp_root->lock); |
48a7639c PM |
2974 | if (needwake) |
2975 | rcu_gp_kthread_wake(rsp); | |
b52573d2 PM |
2976 | } else { |
2977 | /* Give the grace period a kick. */ | |
2978 | rdp->blimit = LONG_MAX; | |
2979 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2980 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2981 | force_quiescent_state(rsp); |
b52573d2 PM |
2982 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2983 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2984 | } | |
4cdfc175 | 2985 | } |
29154c57 PM |
2986 | } |
2987 | ||
ae150184 PM |
2988 | /* |
2989 | * RCU callback function to leak a callback. | |
2990 | */ | |
2991 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2992 | { | |
2993 | } | |
2994 | ||
3fbfbf7a PM |
2995 | /* |
2996 | * Helper function for call_rcu() and friends. The cpu argument will | |
2997 | * normally be -1, indicating "currently running CPU". It may specify | |
2998 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2999 | * is expected to specify a CPU. | |
3000 | */ | |
64db4cff PM |
3001 | static void |
3002 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 3003 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
3004 | { |
3005 | unsigned long flags; | |
3006 | struct rcu_data *rdp; | |
3007 | ||
1146edcb | 3008 | WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */ |
ae150184 PM |
3009 | if (debug_rcu_head_queue(head)) { |
3010 | /* Probable double call_rcu(), so leak the callback. */ | |
7d0ae808 | 3011 | WRITE_ONCE(head->func, rcu_leak_callback); |
ae150184 PM |
3012 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); |
3013 | return; | |
3014 | } | |
64db4cff PM |
3015 | head->func = func; |
3016 | head->next = NULL; | |
3017 | ||
64db4cff PM |
3018 | /* |
3019 | * Opportunistically note grace-period endings and beginnings. | |
3020 | * Note that we might see a beginning right after we see an | |
3021 | * end, but never vice versa, since this CPU has to pass through | |
3022 | * a quiescent state betweentimes. | |
3023 | */ | |
3024 | local_irq_save(flags); | |
394f99a9 | 3025 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
3026 | |
3027 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
3028 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
3029 | int offline; | |
3030 | ||
3031 | if (cpu != -1) | |
3032 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
143da9c2 PM |
3033 | if (likely(rdp->mynode)) { |
3034 | /* Post-boot, so this should be for a no-CBs CPU. */ | |
3035 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); | |
3036 | WARN_ON_ONCE(offline); | |
3037 | /* Offline CPU, _call_rcu() illegal, leak callback. */ | |
3038 | local_irq_restore(flags); | |
3039 | return; | |
3040 | } | |
3041 | /* | |
3042 | * Very early boot, before rcu_init(). Initialize if needed | |
3043 | * and then drop through to queue the callback. | |
3044 | */ | |
3045 | BUG_ON(cpu != -1); | |
34404ca8 | 3046 | WARN_ON_ONCE(!rcu_is_watching()); |
143da9c2 PM |
3047 | if (!likely(rdp->nxtlist)) |
3048 | init_default_callback_list(rdp); | |
0d8ee37e | 3049 | } |
7d0ae808 | 3050 | WRITE_ONCE(rdp->qlen, rdp->qlen + 1); |
486e2593 PM |
3051 | if (lazy) |
3052 | rdp->qlen_lazy++; | |
c57afe80 PM |
3053 | else |
3054 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
3055 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
3056 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
3057 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 3058 | |
d4c08f2a PM |
3059 | if (__is_kfree_rcu_offset((unsigned long)func)) |
3060 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 3061 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 3062 | else |
486e2593 | 3063 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 3064 | |
29154c57 PM |
3065 | /* Go handle any RCU core processing required. */ |
3066 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
3067 | local_irq_restore(flags); |
3068 | } | |
3069 | ||
3070 | /* | |
d6714c22 | 3071 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 3072 | */ |
d6714c22 | 3073 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 3074 | { |
3fbfbf7a | 3075 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 3076 | } |
d6714c22 | 3077 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
3078 | |
3079 | /* | |
486e2593 | 3080 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
3081 | */ |
3082 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
3083 | { | |
3fbfbf7a | 3084 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
3085 | } |
3086 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
3087 | ||
495aa969 ACB |
3088 | /* |
3089 | * Queue an RCU callback for lazy invocation after a grace period. | |
3090 | * This will likely be later named something like "call_rcu_lazy()", | |
3091 | * but this change will require some way of tagging the lazy RCU | |
3092 | * callbacks in the list of pending callbacks. Until then, this | |
3093 | * function may only be called from __kfree_rcu(). | |
3094 | */ | |
3095 | void kfree_call_rcu(struct rcu_head *head, | |
3096 | void (*func)(struct rcu_head *rcu)) | |
3097 | { | |
e534165b | 3098 | __call_rcu(head, func, rcu_state_p, -1, 1); |
495aa969 ACB |
3099 | } |
3100 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
3101 | ||
6d813391 PM |
3102 | /* |
3103 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
3104 | * any blocking grace-period wait automatically implies a grace period | |
3105 | * if there is only one CPU online at any point time during execution | |
3106 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
3107 | * occasionally incorrectly indicate that there are multiple CPUs online | |
3108 | * when there was in fact only one the whole time, as this just adds | |
3109 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
3110 | */ |
3111 | static inline int rcu_blocking_is_gp(void) | |
3112 | { | |
95f0c1de PM |
3113 | int ret; |
3114 | ||
6d813391 | 3115 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
3116 | preempt_disable(); |
3117 | ret = num_online_cpus() <= 1; | |
3118 | preempt_enable(); | |
3119 | return ret; | |
6d813391 PM |
3120 | } |
3121 | ||
6ebb237b PM |
3122 | /** |
3123 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
3124 | * | |
3125 | * Control will return to the caller some time after a full rcu-sched | |
3126 | * grace period has elapsed, in other words after all currently executing | |
3127 | * rcu-sched read-side critical sections have completed. These read-side | |
3128 | * critical sections are delimited by rcu_read_lock_sched() and | |
3129 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
3130 | * local_irq_disable(), and so on may be used in place of | |
3131 | * rcu_read_lock_sched(). | |
3132 | * | |
3133 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
3134 | * non-threaded hardware-interrupt handlers, in progress on entry will |
3135 | * have completed before this primitive returns. However, this does not | |
3136 | * guarantee that softirq handlers will have completed, since in some | |
3137 | * kernels, these handlers can run in process context, and can block. | |
3138 | * | |
3139 | * Note that this guarantee implies further memory-ordering guarantees. | |
3140 | * On systems with more than one CPU, when synchronize_sched() returns, | |
3141 | * each CPU is guaranteed to have executed a full memory barrier since the | |
3142 | * end of its last RCU-sched read-side critical section whose beginning | |
3143 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
3144 | * an RCU read-side critical section that extends beyond the return from | |
3145 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
3146 | * after the beginning of synchronize_sched() and before the beginning of | |
3147 | * that RCU read-side critical section. Note that these guarantees include | |
3148 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
3149 | * that are executing in the kernel. | |
3150 | * | |
3151 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
3152 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
3153 | * to have executed a full memory barrier during the execution of | |
3154 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
3155 | * again only if the system has more than one CPU). | |
6ebb237b PM |
3156 | * |
3157 | * This primitive provides the guarantees made by the (now removed) | |
3158 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
3159 | * guarantees that rcu_read_lock() sections will have completed. | |
3160 | * In "classic RCU", these two guarantees happen to be one and | |
3161 | * the same, but can differ in realtime RCU implementations. | |
3162 | */ | |
3163 | void synchronize_sched(void) | |
3164 | { | |
fe15d706 PM |
3165 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
3166 | !lock_is_held(&rcu_lock_map) && | |
3167 | !lock_is_held(&rcu_sched_lock_map), | |
3168 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
3169 | if (rcu_blocking_is_gp()) |
3170 | return; | |
5afff48b | 3171 | if (rcu_gp_is_expedited()) |
3705b88d AM |
3172 | synchronize_sched_expedited(); |
3173 | else | |
3174 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
3175 | } |
3176 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
3177 | ||
3178 | /** | |
3179 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
3180 | * | |
3181 | * Control will return to the caller some time after a full rcu_bh grace | |
3182 | * period has elapsed, in other words after all currently executing rcu_bh | |
3183 | * read-side critical sections have completed. RCU read-side critical | |
3184 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
3185 | * and may be nested. | |
f0a0e6f2 PM |
3186 | * |
3187 | * See the description of synchronize_sched() for more detailed information | |
3188 | * on memory ordering guarantees. | |
6ebb237b PM |
3189 | */ |
3190 | void synchronize_rcu_bh(void) | |
3191 | { | |
fe15d706 PM |
3192 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
3193 | !lock_is_held(&rcu_lock_map) && | |
3194 | !lock_is_held(&rcu_sched_lock_map), | |
3195 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
3196 | if (rcu_blocking_is_gp()) |
3197 | return; | |
5afff48b | 3198 | if (rcu_gp_is_expedited()) |
3705b88d AM |
3199 | synchronize_rcu_bh_expedited(); |
3200 | else | |
3201 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
3202 | } |
3203 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
3204 | ||
765a3f4f PM |
3205 | /** |
3206 | * get_state_synchronize_rcu - Snapshot current RCU state | |
3207 | * | |
3208 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
3209 | * to determine whether or not a full grace period has elapsed in the | |
3210 | * meantime. | |
3211 | */ | |
3212 | unsigned long get_state_synchronize_rcu(void) | |
3213 | { | |
3214 | /* | |
3215 | * Any prior manipulation of RCU-protected data must happen | |
3216 | * before the load from ->gpnum. | |
3217 | */ | |
3218 | smp_mb(); /* ^^^ */ | |
3219 | ||
3220 | /* | |
3221 | * Make sure this load happens before the purportedly | |
3222 | * time-consuming work between get_state_synchronize_rcu() | |
3223 | * and cond_synchronize_rcu(). | |
3224 | */ | |
e534165b | 3225 | return smp_load_acquire(&rcu_state_p->gpnum); |
765a3f4f PM |
3226 | } |
3227 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
3228 | ||
3229 | /** | |
3230 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
3231 | * | |
3232 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
3233 | * | |
3234 | * If a full RCU grace period has elapsed since the earlier call to | |
3235 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
3236 | * synchronize_rcu() to wait for a full grace period. | |
3237 | * | |
3238 | * Yes, this function does not take counter wrap into account. But | |
3239 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3240 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3241 | * so waiting for one additional grace period should be just fine. | |
3242 | */ | |
3243 | void cond_synchronize_rcu(unsigned long oldstate) | |
3244 | { | |
3245 | unsigned long newstate; | |
3246 | ||
3247 | /* | |
3248 | * Ensure that this load happens before any RCU-destructive | |
3249 | * actions the caller might carry out after we return. | |
3250 | */ | |
e534165b | 3251 | newstate = smp_load_acquire(&rcu_state_p->completed); |
765a3f4f PM |
3252 | if (ULONG_CMP_GE(oldstate, newstate)) |
3253 | synchronize_rcu(); | |
3254 | } | |
3255 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
3256 | ||
3d3b7db0 PM |
3257 | static int synchronize_sched_expedited_cpu_stop(void *data) |
3258 | { | |
3259 | /* | |
3260 | * There must be a full memory barrier on each affected CPU | |
3261 | * between the time that try_stop_cpus() is called and the | |
3262 | * time that it returns. | |
3263 | * | |
3264 | * In the current initial implementation of cpu_stop, the | |
3265 | * above condition is already met when the control reaches | |
3266 | * this point and the following smp_mb() is not strictly | |
3267 | * necessary. Do smp_mb() anyway for documentation and | |
3268 | * robustness against future implementation changes. | |
3269 | */ | |
3270 | smp_mb(); /* See above comment block. */ | |
3271 | return 0; | |
3272 | } | |
3273 | ||
236fefaf PM |
3274 | /** |
3275 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
3276 | * | |
3277 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
3278 | * approach to force the grace period to end quickly. This consumes | |
3279 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
3280 | * so is thus not recommended for any sort of common-case code. In fact, | |
3281 | * if you are using synchronize_sched_expedited() in a loop, please | |
3282 | * restructure your code to batch your updates, and then use a single | |
3283 | * synchronize_sched() instead. | |
3d3b7db0 | 3284 | * |
3d3b7db0 PM |
3285 | * This implementation can be thought of as an application of ticket |
3286 | * locking to RCU, with sync_sched_expedited_started and | |
3287 | * sync_sched_expedited_done taking on the roles of the halves | |
3288 | * of the ticket-lock word. Each task atomically increments | |
3289 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
3290 | * then attempts to stop all the CPUs. If this succeeds, then each | |
3291 | * CPU will have executed a context switch, resulting in an RCU-sched | |
3292 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
3293 | * update sync_sched_expedited_done to match our snapshot -- but | |
3294 | * only if someone else has not already advanced past our snapshot. | |
3295 | * | |
3296 | * On the other hand, if try_stop_cpus() fails, we check the value | |
3297 | * of sync_sched_expedited_done. If it has advanced past our | |
3298 | * initial snapshot, then someone else must have forced a grace period | |
3299 | * some time after we took our snapshot. In this case, our work is | |
3300 | * done for us, and we can simply return. Otherwise, we try again, | |
3301 | * but keep our initial snapshot for purposes of checking for someone | |
3302 | * doing our work for us. | |
3303 | * | |
3304 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
3305 | */ | |
3306 | void synchronize_sched_expedited(void) | |
3307 | { | |
e0775cef PM |
3308 | cpumask_var_t cm; |
3309 | bool cma = false; | |
3310 | int cpu; | |
1924bcb0 PM |
3311 | long firstsnap, s, snap; |
3312 | int trycount = 0; | |
40694d66 | 3313 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 3314 | |
1924bcb0 PM |
3315 | /* |
3316 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
3317 | * By allowing sync_sched_expedited_started to advance no more than | |
3318 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
3319 | * that more than 3.5 billion CPUs would be required to force a | |
3320 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
3321 | * course be required on a 64-bit system. | |
3322 | */ | |
40694d66 PM |
3323 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
3324 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 | 3325 | ULONG_MAX / 8)) { |
30ff1533 | 3326 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 3327 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
3328 | return; |
3329 | } | |
3d3b7db0 | 3330 | |
1924bcb0 PM |
3331 | /* |
3332 | * Take a ticket. Note that atomic_inc_return() implies a | |
3333 | * full memory barrier. | |
3334 | */ | |
40694d66 | 3335 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 3336 | firstsnap = snap; |
dd56af42 PM |
3337 | if (!try_get_online_cpus()) { |
3338 | /* CPU hotplug operation in flight, fall back to normal GP. */ | |
3339 | wait_rcu_gp(call_rcu_sched); | |
3340 | atomic_long_inc(&rsp->expedited_normal); | |
3341 | return; | |
3342 | } | |
1cc85961 | 3343 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 | 3344 | |
e0775cef PM |
3345 | /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */ |
3346 | cma = zalloc_cpumask_var(&cm, GFP_KERNEL); | |
3347 | if (cma) { | |
3348 | cpumask_copy(cm, cpu_online_mask); | |
3349 | cpumask_clear_cpu(raw_smp_processor_id(), cm); | |
3350 | for_each_cpu(cpu, cm) { | |
3351 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); | |
3352 | ||
3353 | if (!(atomic_add_return(0, &rdtp->dynticks) & 0x1)) | |
3354 | cpumask_clear_cpu(cpu, cm); | |
3355 | } | |
3356 | if (cpumask_weight(cm) == 0) | |
3357 | goto all_cpus_idle; | |
3358 | } | |
3359 | ||
3d3b7db0 PM |
3360 | /* |
3361 | * Each pass through the following loop attempts to force a | |
3362 | * context switch on each CPU. | |
3363 | */ | |
e0775cef | 3364 | while (try_stop_cpus(cma ? cm : cpu_online_mask, |
3d3b7db0 PM |
3365 | synchronize_sched_expedited_cpu_stop, |
3366 | NULL) == -EAGAIN) { | |
3367 | put_online_cpus(); | |
a30489c5 | 3368 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 3369 | |
1924bcb0 | 3370 | /* Check to see if someone else did our work for us. */ |
40694d66 | 3371 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3372 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 | 3373 | /* ensure test happens before caller kfree */ |
4e857c58 | 3374 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3375 | atomic_long_inc(&rsp->expedited_workdone1); |
e0775cef | 3376 | free_cpumask_var(cm); |
1924bcb0 PM |
3377 | return; |
3378 | } | |
3d3b7db0 PM |
3379 | |
3380 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 3381 | if (trycount++ < 10) { |
3d3b7db0 | 3382 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 3383 | } else { |
3705b88d | 3384 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 3385 | atomic_long_inc(&rsp->expedited_normal); |
e0775cef | 3386 | free_cpumask_var(cm); |
3d3b7db0 PM |
3387 | return; |
3388 | } | |
3389 | ||
1924bcb0 | 3390 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 3391 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3392 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 | 3393 | /* ensure test happens before caller kfree */ |
4e857c58 | 3394 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3395 | atomic_long_inc(&rsp->expedited_workdone2); |
e0775cef | 3396 | free_cpumask_var(cm); |
3d3b7db0 PM |
3397 | return; |
3398 | } | |
3399 | ||
3400 | /* | |
3401 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
3402 | * callers to piggyback on our grace period. We retry |
3403 | * after they started, so our grace period works for them, | |
3404 | * and they started after our first try, so their grace | |
3405 | * period works for us. | |
3d3b7db0 | 3406 | */ |
dd56af42 PM |
3407 | if (!try_get_online_cpus()) { |
3408 | /* CPU hotplug operation in flight, use normal GP. */ | |
3409 | wait_rcu_gp(call_rcu_sched); | |
3410 | atomic_long_inc(&rsp->expedited_normal); | |
e0775cef | 3411 | free_cpumask_var(cm); |
dd56af42 PM |
3412 | return; |
3413 | } | |
40694d66 | 3414 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
3415 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
3416 | } | |
a30489c5 | 3417 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 | 3418 | |
e0775cef PM |
3419 | all_cpus_idle: |
3420 | free_cpumask_var(cm); | |
3421 | ||
3d3b7db0 PM |
3422 | /* |
3423 | * Everyone up to our most recent fetch is covered by our grace | |
3424 | * period. Update the counter, but only if our work is still | |
3425 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 3426 | * than we did already did their update. |
3d3b7db0 PM |
3427 | */ |
3428 | do { | |
a30489c5 | 3429 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 3430 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3431 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 | 3432 | /* ensure test happens before caller kfree */ |
4e857c58 | 3433 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3434 | atomic_long_inc(&rsp->expedited_done_lost); |
3d3b7db0 PM |
3435 | break; |
3436 | } | |
40694d66 | 3437 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 3438 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
3439 | |
3440 | put_online_cpus(); | |
3441 | } | |
3442 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
3443 | ||
64db4cff PM |
3444 | /* |
3445 | * Check to see if there is any immediate RCU-related work to be done | |
3446 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
3447 | * The checks are in order of increasing expense: checks that can be | |
3448 | * carried out against CPU-local state are performed first. However, | |
3449 | * we must check for CPU stalls first, else we might not get a chance. | |
3450 | */ | |
3451 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
3452 | { | |
2f51f988 PM |
3453 | struct rcu_node *rnp = rdp->mynode; |
3454 | ||
64db4cff PM |
3455 | rdp->n_rcu_pending++; |
3456 | ||
3457 | /* Check for CPU stalls, if enabled. */ | |
3458 | check_cpu_stall(rsp, rdp); | |
3459 | ||
a096932f PM |
3460 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ |
3461 | if (rcu_nohz_full_cpu(rsp)) | |
3462 | return 0; | |
3463 | ||
64db4cff | 3464 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
5c51dd73 | 3465 | if (rcu_scheduler_fully_active && |
5cd37193 PM |
3466 | rdp->qs_pending && !rdp->passed_quiesce && |
3467 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) { | |
d21670ac | 3468 | rdp->n_rp_qs_pending++; |
5cd37193 PM |
3469 | } else if (rdp->qs_pending && |
3470 | (rdp->passed_quiesce || | |
3471 | rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) { | |
d21670ac | 3472 | rdp->n_rp_report_qs++; |
64db4cff | 3473 | return 1; |
7ba5c840 | 3474 | } |
64db4cff PM |
3475 | |
3476 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
3477 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
3478 | rdp->n_rp_cb_ready++; | |
64db4cff | 3479 | return 1; |
7ba5c840 | 3480 | } |
64db4cff PM |
3481 | |
3482 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
3483 | if (cpu_needs_another_gp(rsp, rdp)) { |
3484 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 3485 | return 1; |
7ba5c840 | 3486 | } |
64db4cff PM |
3487 | |
3488 | /* Has another RCU grace period completed? */ | |
7d0ae808 | 3489 | if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 3490 | rdp->n_rp_gp_completed++; |
64db4cff | 3491 | return 1; |
7ba5c840 | 3492 | } |
64db4cff PM |
3493 | |
3494 | /* Has a new RCU grace period started? */ | |
7d0ae808 PM |
3495 | if (READ_ONCE(rnp->gpnum) != rdp->gpnum || |
3496 | unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */ | |
7ba5c840 | 3497 | rdp->n_rp_gp_started++; |
64db4cff | 3498 | return 1; |
7ba5c840 | 3499 | } |
64db4cff | 3500 | |
96d3fd0d PM |
3501 | /* Does this CPU need a deferred NOCB wakeup? */ |
3502 | if (rcu_nocb_need_deferred_wakeup(rdp)) { | |
3503 | rdp->n_rp_nocb_defer_wakeup++; | |
3504 | return 1; | |
3505 | } | |
3506 | ||
64db4cff | 3507 | /* nothing to do */ |
7ba5c840 | 3508 | rdp->n_rp_need_nothing++; |
64db4cff PM |
3509 | return 0; |
3510 | } | |
3511 | ||
3512 | /* | |
3513 | * Check to see if there is any immediate RCU-related work to be done | |
3514 | * by the current CPU, returning 1 if so. This function is part of the | |
3515 | * RCU implementation; it is -not- an exported member of the RCU API. | |
3516 | */ | |
e3950ecd | 3517 | static int rcu_pending(void) |
64db4cff | 3518 | { |
6ce75a23 PM |
3519 | struct rcu_state *rsp; |
3520 | ||
3521 | for_each_rcu_flavor(rsp) | |
e3950ecd | 3522 | if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda))) |
6ce75a23 PM |
3523 | return 1; |
3524 | return 0; | |
64db4cff PM |
3525 | } |
3526 | ||
3527 | /* | |
c0f4dfd4 PM |
3528 | * Return true if the specified CPU has any callback. If all_lazy is |
3529 | * non-NULL, store an indication of whether all callbacks are lazy. | |
3530 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 3531 | */ |
82072c4f | 3532 | static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy) |
64db4cff | 3533 | { |
c0f4dfd4 PM |
3534 | bool al = true; |
3535 | bool hc = false; | |
3536 | struct rcu_data *rdp; | |
6ce75a23 PM |
3537 | struct rcu_state *rsp; |
3538 | ||
c0f4dfd4 | 3539 | for_each_rcu_flavor(rsp) { |
aa6da514 | 3540 | rdp = this_cpu_ptr(rsp->rda); |
69c8d28c PM |
3541 | if (!rdp->nxtlist) |
3542 | continue; | |
3543 | hc = true; | |
3544 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
c0f4dfd4 | 3545 | al = false; |
69c8d28c PM |
3546 | break; |
3547 | } | |
c0f4dfd4 PM |
3548 | } |
3549 | if (all_lazy) | |
3550 | *all_lazy = al; | |
3551 | return hc; | |
64db4cff PM |
3552 | } |
3553 | ||
a83eff0a PM |
3554 | /* |
3555 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
3556 | * the compiler is expected to optimize this away. | |
3557 | */ | |
e66c33d5 | 3558 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
3559 | int cpu, unsigned long done) |
3560 | { | |
3561 | trace_rcu_barrier(rsp->name, s, cpu, | |
3562 | atomic_read(&rsp->barrier_cpu_count), done); | |
3563 | } | |
3564 | ||
b1420f1c PM |
3565 | /* |
3566 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
3567 | * up the task executing _rcu_barrier(). | |
3568 | */ | |
24ebbca8 | 3569 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3570 | { |
24ebbca8 PM |
3571 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
3572 | struct rcu_state *rsp = rdp->rsp; | |
3573 | ||
a83eff0a PM |
3574 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
3575 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 3576 | complete(&rsp->barrier_completion); |
a83eff0a PM |
3577 | } else { |
3578 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
3579 | } | |
d0ec774c PM |
3580 | } |
3581 | ||
3582 | /* | |
3583 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3584 | */ | |
3585 | static void rcu_barrier_func(void *type) | |
3586 | { | |
037b64ed | 3587 | struct rcu_state *rsp = type; |
fa07a58f | 3588 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
d0ec774c | 3589 | |
a83eff0a | 3590 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 3591 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 3592 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
3593 | } |
3594 | ||
d0ec774c PM |
3595 | /* |
3596 | * Orchestrate the specified type of RCU barrier, waiting for all | |
3597 | * RCU callbacks of the specified type to complete. | |
3598 | */ | |
037b64ed | 3599 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 3600 | { |
b1420f1c | 3601 | int cpu; |
b1420f1c | 3602 | struct rcu_data *rdp; |
7d0ae808 | 3603 | unsigned long snap = READ_ONCE(rsp->n_barrier_done); |
cf3a9c48 | 3604 | unsigned long snap_done; |
b1420f1c | 3605 | |
a83eff0a | 3606 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 3607 | |
e74f4c45 | 3608 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 3609 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 3610 | |
cf3a9c48 PM |
3611 | /* |
3612 | * Ensure that all prior references, including to ->n_barrier_done, | |
3613 | * are ordered before the _rcu_barrier() machinery. | |
3614 | */ | |
3615 | smp_mb(); /* See above block comment. */ | |
3616 | ||
3617 | /* | |
3618 | * Recheck ->n_barrier_done to see if others did our work for us. | |
3619 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
3620 | * transition. The "if" expression below therefore rounds the old | |
3621 | * value up to the next even number and adds two before comparing. | |
3622 | */ | |
458fb381 | 3623 | snap_done = rsp->n_barrier_done; |
a83eff0a | 3624 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
458fb381 PM |
3625 | |
3626 | /* | |
3627 | * If the value in snap is odd, we needed to wait for the current | |
3628 | * rcu_barrier() to complete, then wait for the next one, in other | |
3629 | * words, we need the value of snap_done to be three larger than | |
3630 | * the value of snap. On the other hand, if the value in snap is | |
3631 | * even, we only had to wait for the next rcu_barrier() to complete, | |
3632 | * in other words, we need the value of snap_done to be only two | |
3633 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes | |
3634 | * this for us (thank you, Linus!). | |
3635 | */ | |
3636 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) { | |
a83eff0a | 3637 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
3638 | smp_mb(); /* caller's subsequent code after above check. */ |
3639 | mutex_unlock(&rsp->barrier_mutex); | |
3640 | return; | |
3641 | } | |
3642 | ||
3643 | /* | |
3644 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
7d0ae808 | 3645 | * WRITE_ONCE() to prevent the compiler from speculating |
cf3a9c48 PM |
3646 | * the increment to precede the early-exit check. |
3647 | */ | |
7d0ae808 | 3648 | WRITE_ONCE(rsp->n_barrier_done, rsp->n_barrier_done + 1); |
cf3a9c48 | 3649 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); |
a83eff0a | 3650 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 3651 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 3652 | |
d0ec774c | 3653 | /* |
b1420f1c PM |
3654 | * Initialize the count to one rather than to zero in order to |
3655 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
3656 | * (or preemption of this task). Exclude CPU-hotplug operations |
3657 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 3658 | */ |
7db74df8 | 3659 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 3660 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 3661 | get_online_cpus(); |
b1420f1c PM |
3662 | |
3663 | /* | |
1331e7a1 PM |
3664 | * Force each CPU with callbacks to register a new callback. |
3665 | * When that callback is invoked, we will know that all of the | |
3666 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 3667 | */ |
3fbfbf7a | 3668 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 3669 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 3670 | continue; |
b1420f1c | 3671 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 3672 | if (rcu_is_nocb_cpu(cpu)) { |
d7e29933 PM |
3673 | if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) { |
3674 | _rcu_barrier_trace(rsp, "OfflineNoCB", cpu, | |
3675 | rsp->n_barrier_done); | |
3676 | } else { | |
3677 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, | |
3678 | rsp->n_barrier_done); | |
41050a00 | 3679 | smp_mb__before_atomic(); |
d7e29933 PM |
3680 | atomic_inc(&rsp->barrier_cpu_count); |
3681 | __call_rcu(&rdp->barrier_head, | |
3682 | rcu_barrier_callback, rsp, cpu, 0); | |
3683 | } | |
7d0ae808 | 3684 | } else if (READ_ONCE(rdp->qlen)) { |
a83eff0a PM |
3685 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
3686 | rsp->n_barrier_done); | |
037b64ed | 3687 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 3688 | } else { |
a83eff0a PM |
3689 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
3690 | rsp->n_barrier_done); | |
b1420f1c PM |
3691 | } |
3692 | } | |
1331e7a1 | 3693 | put_online_cpus(); |
b1420f1c PM |
3694 | |
3695 | /* | |
3696 | * Now that we have an rcu_barrier_callback() callback on each | |
3697 | * CPU, and thus each counted, remove the initial count. | |
3698 | */ | |
24ebbca8 | 3699 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 3700 | complete(&rsp->barrier_completion); |
b1420f1c | 3701 | |
cf3a9c48 PM |
3702 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
3703 | smp_mb(); /* Keep increment after above mechanism. */ | |
7d0ae808 | 3704 | WRITE_ONCE(rsp->n_barrier_done, rsp->n_barrier_done + 1); |
cf3a9c48 | 3705 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); |
a83eff0a | 3706 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
3707 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
3708 | ||
b1420f1c | 3709 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 3710 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
3711 | |
3712 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 3713 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 3714 | } |
d0ec774c PM |
3715 | |
3716 | /** | |
3717 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
3718 | */ | |
3719 | void rcu_barrier_bh(void) | |
3720 | { | |
037b64ed | 3721 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
3722 | } |
3723 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
3724 | ||
3725 | /** | |
3726 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
3727 | */ | |
3728 | void rcu_barrier_sched(void) | |
3729 | { | |
037b64ed | 3730 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
3731 | } |
3732 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
3733 | ||
0aa04b05 PM |
3734 | /* |
3735 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the | |
3736 | * first CPU in a given leaf rcu_node structure coming online. The caller | |
3737 | * must hold the corresponding leaf rcu_node ->lock with interrrupts | |
3738 | * disabled. | |
3739 | */ | |
3740 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) | |
3741 | { | |
3742 | long mask; | |
3743 | struct rcu_node *rnp = rnp_leaf; | |
3744 | ||
3745 | for (;;) { | |
3746 | mask = rnp->grpmask; | |
3747 | rnp = rnp->parent; | |
3748 | if (rnp == NULL) | |
3749 | return; | |
3750 | raw_spin_lock(&rnp->lock); /* Interrupts already disabled. */ | |
3751 | rnp->qsmaskinit |= mask; | |
3752 | raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */ | |
3753 | } | |
3754 | } | |
3755 | ||
64db4cff | 3756 | /* |
27569620 | 3757 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 3758 | */ |
27569620 PM |
3759 | static void __init |
3760 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
3761 | { |
3762 | unsigned long flags; | |
394f99a9 | 3763 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
3764 | struct rcu_node *rnp = rcu_get_root(rsp); |
3765 | ||
3766 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 3767 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 3768 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
27569620 | 3769 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 3770 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 3771 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 3772 | rdp->cpu = cpu; |
d4c08f2a | 3773 | rdp->rsp = rsp; |
3fbfbf7a | 3774 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 3775 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
3776 | } |
3777 | ||
3778 | /* | |
3779 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
3780 | * offline event can be happening at a given time. Note also that we | |
3781 | * can accept some slop in the rsp->completed access due to the fact | |
3782 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 3783 | */ |
49fb4c62 | 3784 | static void |
9b67122a | 3785 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp) |
64db4cff PM |
3786 | { |
3787 | unsigned long flags; | |
64db4cff | 3788 | unsigned long mask; |
394f99a9 | 3789 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
3790 | struct rcu_node *rnp = rcu_get_root(rsp); |
3791 | ||
3792 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 3793 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 3794 | rdp->beenonline = 1; /* We have now been online. */ |
37c72e56 PM |
3795 | rdp->qlen_last_fqs_check = 0; |
3796 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 3797 | rdp->blimit = blimit; |
39c8d313 PM |
3798 | if (!rdp->nxtlist) |
3799 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ | |
29e37d81 | 3800 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 3801 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
3802 | atomic_set(&rdp->dynticks->dynticks, |
3803 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 3804 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 3805 | |
0aa04b05 PM |
3806 | /* |
3807 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed | |
3808 | * propagation up the rcu_node tree will happen at the beginning | |
3809 | * of the next grace period. | |
3810 | */ | |
64db4cff PM |
3811 | rnp = rdp->mynode; |
3812 | mask = rdp->grpmask; | |
0aa04b05 PM |
3813 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
3814 | smp_mb__after_unlock_lock(); | |
3815 | rnp->qsmaskinitnext |= mask; | |
3816 | rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */ | |
3817 | rdp->completed = rnp->completed; | |
3818 | rdp->passed_quiesce = false; | |
a738eec6 | 3819 | rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu); |
0aa04b05 PM |
3820 | rdp->qs_pending = false; |
3821 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); | |
3822 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff PM |
3823 | } |
3824 | ||
49fb4c62 | 3825 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 3826 | { |
6ce75a23 PM |
3827 | struct rcu_state *rsp; |
3828 | ||
3829 | for_each_rcu_flavor(rsp) | |
9b67122a | 3830 | rcu_init_percpu_data(cpu, rsp); |
64db4cff PM |
3831 | } |
3832 | ||
3833 | /* | |
f41d911f | 3834 | * Handle CPU online/offline notification events. |
64db4cff | 3835 | */ |
88428cc5 PM |
3836 | int rcu_cpu_notify(struct notifier_block *self, |
3837 | unsigned long action, void *hcpu) | |
64db4cff PM |
3838 | { |
3839 | long cpu = (long)hcpu; | |
e534165b | 3840 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
a26ac245 | 3841 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 3842 | struct rcu_state *rsp; |
64db4cff PM |
3843 | |
3844 | switch (action) { | |
3845 | case CPU_UP_PREPARE: | |
3846 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
3847 | rcu_prepare_cpu(cpu); |
3848 | rcu_prepare_kthreads(cpu); | |
35ce7f29 | 3849 | rcu_spawn_all_nocb_kthreads(cpu); |
a26ac245 PM |
3850 | break; |
3851 | case CPU_ONLINE: | |
0f962a5e | 3852 | case CPU_DOWN_FAILED: |
5d01bbd1 | 3853 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
3854 | break; |
3855 | case CPU_DOWN_PREPARE: | |
34ed6246 | 3856 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 3857 | break; |
d0ec774c PM |
3858 | case CPU_DYING: |
3859 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
3860 | for_each_rcu_flavor(rsp) |
3861 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 3862 | break; |
88428cc5 PM |
3863 | case CPU_DYING_IDLE: |
3864 | for_each_rcu_flavor(rsp) { | |
3865 | rcu_cleanup_dying_idle_cpu(cpu, rsp); | |
3866 | } | |
3867 | break; | |
64db4cff PM |
3868 | case CPU_DEAD: |
3869 | case CPU_DEAD_FROZEN: | |
3870 | case CPU_UP_CANCELED: | |
3871 | case CPU_UP_CANCELED_FROZEN: | |
776d6807 | 3872 | for_each_rcu_flavor(rsp) { |
6ce75a23 | 3873 | rcu_cleanup_dead_cpu(cpu, rsp); |
776d6807 PM |
3874 | do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu)); |
3875 | } | |
64db4cff PM |
3876 | break; |
3877 | default: | |
3878 | break; | |
3879 | } | |
34ed6246 | 3880 | return NOTIFY_OK; |
64db4cff PM |
3881 | } |
3882 | ||
d1d74d14 BP |
3883 | static int rcu_pm_notify(struct notifier_block *self, |
3884 | unsigned long action, void *hcpu) | |
3885 | { | |
3886 | switch (action) { | |
3887 | case PM_HIBERNATION_PREPARE: | |
3888 | case PM_SUSPEND_PREPARE: | |
3889 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
5afff48b | 3890 | rcu_expedite_gp(); |
d1d74d14 BP |
3891 | break; |
3892 | case PM_POST_HIBERNATION: | |
3893 | case PM_POST_SUSPEND: | |
5afff48b PM |
3894 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ |
3895 | rcu_unexpedite_gp(); | |
d1d74d14 BP |
3896 | break; |
3897 | default: | |
3898 | break; | |
3899 | } | |
3900 | return NOTIFY_OK; | |
3901 | } | |
3902 | ||
b3dbec76 | 3903 | /* |
9386c0b7 | 3904 | * Spawn the kthreads that handle each RCU flavor's grace periods. |
b3dbec76 PM |
3905 | */ |
3906 | static int __init rcu_spawn_gp_kthread(void) | |
3907 | { | |
3908 | unsigned long flags; | |
a94844b2 | 3909 | int kthread_prio_in = kthread_prio; |
b3dbec76 PM |
3910 | struct rcu_node *rnp; |
3911 | struct rcu_state *rsp; | |
a94844b2 | 3912 | struct sched_param sp; |
b3dbec76 PM |
3913 | struct task_struct *t; |
3914 | ||
a94844b2 PM |
3915 | /* Force priority into range. */ |
3916 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) | |
3917 | kthread_prio = 1; | |
3918 | else if (kthread_prio < 0) | |
3919 | kthread_prio = 0; | |
3920 | else if (kthread_prio > 99) | |
3921 | kthread_prio = 99; | |
3922 | if (kthread_prio != kthread_prio_in) | |
3923 | pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n", | |
3924 | kthread_prio, kthread_prio_in); | |
3925 | ||
9386c0b7 | 3926 | rcu_scheduler_fully_active = 1; |
b3dbec76 | 3927 | for_each_rcu_flavor(rsp) { |
a94844b2 | 3928 | t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
3929 | BUG_ON(IS_ERR(t)); |
3930 | rnp = rcu_get_root(rsp); | |
3931 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3932 | rsp->gp_kthread = t; | |
a94844b2 PM |
3933 | if (kthread_prio) { |
3934 | sp.sched_priority = kthread_prio; | |
3935 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
3936 | } | |
3937 | wake_up_process(t); | |
b3dbec76 PM |
3938 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
3939 | } | |
35ce7f29 | 3940 | rcu_spawn_nocb_kthreads(); |
9386c0b7 | 3941 | rcu_spawn_boost_kthreads(); |
b3dbec76 PM |
3942 | return 0; |
3943 | } | |
3944 | early_initcall(rcu_spawn_gp_kthread); | |
3945 | ||
bbad9379 PM |
3946 | /* |
3947 | * This function is invoked towards the end of the scheduler's initialization | |
3948 | * process. Before this is called, the idle task might contain | |
3949 | * RCU read-side critical sections (during which time, this idle | |
3950 | * task is booting the system). After this function is called, the | |
3951 | * idle tasks are prohibited from containing RCU read-side critical | |
3952 | * sections. This function also enables RCU lockdep checking. | |
3953 | */ | |
3954 | void rcu_scheduler_starting(void) | |
3955 | { | |
3956 | WARN_ON(num_online_cpus() != 1); | |
3957 | WARN_ON(nr_context_switches() > 0); | |
3958 | rcu_scheduler_active = 1; | |
3959 | } | |
3960 | ||
64db4cff PM |
3961 | /* |
3962 | * Compute the per-level fanout, either using the exact fanout specified | |
7fa27001 | 3963 | * or balancing the tree, depending on the rcu_fanout_exact boot parameter. |
64db4cff | 3964 | */ |
199977bf | 3965 | static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt) |
64db4cff | 3966 | { |
64db4cff PM |
3967 | int i; |
3968 | ||
7fa27001 | 3969 | if (rcu_fanout_exact) { |
199977bf | 3970 | levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; |
66292405 | 3971 | for (i = rcu_num_lvls - 2; i >= 0; i--) |
199977bf | 3972 | levelspread[i] = RCU_FANOUT; |
66292405 PM |
3973 | } else { |
3974 | int ccur; | |
3975 | int cprv; | |
3976 | ||
3977 | cprv = nr_cpu_ids; | |
3978 | for (i = rcu_num_lvls - 1; i >= 0; i--) { | |
199977bf AG |
3979 | ccur = levelcnt[i]; |
3980 | levelspread[i] = (cprv + ccur - 1) / ccur; | |
66292405 PM |
3981 | cprv = ccur; |
3982 | } | |
64db4cff PM |
3983 | } |
3984 | } | |
64db4cff PM |
3985 | |
3986 | /* | |
3987 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3988 | */ | |
394f99a9 LJ |
3989 | static void __init rcu_init_one(struct rcu_state *rsp, |
3990 | struct rcu_data __percpu *rda) | |
64db4cff | 3991 | { |
cb007102 AG |
3992 | static const char * const buf[] = RCU_NODE_NAME_INIT; |
3993 | static const char * const fqs[] = RCU_FQS_NAME_INIT; | |
4a81e832 | 3994 | static u8 fl_mask = 0x1; |
199977bf AG |
3995 | |
3996 | int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */ | |
3997 | int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */ | |
64db4cff PM |
3998 | int cpustride = 1; |
3999 | int i; | |
4000 | int j; | |
4001 | struct rcu_node *rnp; | |
4002 | ||
05b84aec | 4003 | BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
b6407e86 | 4004 | |
3eaaaf6c PM |
4005 | /* Silence gcc 4.8 false positive about array index out of range. */ |
4006 | if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS) | |
4007 | panic("rcu_init_one: rcu_num_lvls out of range"); | |
4930521a | 4008 | |
64db4cff PM |
4009 | /* Initialize the level-tracking arrays. */ |
4010 | ||
f885b7f2 | 4011 | for (i = 0; i < rcu_num_lvls; i++) |
199977bf | 4012 | levelcnt[i] = num_rcu_lvl[i]; |
f885b7f2 | 4013 | for (i = 1; i < rcu_num_lvls; i++) |
199977bf AG |
4014 | rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1]; |
4015 | rcu_init_levelspread(levelspread, levelcnt); | |
4a81e832 PM |
4016 | rsp->flavor_mask = fl_mask; |
4017 | fl_mask <<= 1; | |
64db4cff PM |
4018 | |
4019 | /* Initialize the elements themselves, starting from the leaves. */ | |
4020 | ||
f885b7f2 | 4021 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
199977bf | 4022 | cpustride *= levelspread[i]; |
64db4cff | 4023 | rnp = rsp->level[i]; |
199977bf | 4024 | for (j = 0; j < levelcnt[i]; j++, rnp++) { |
1304afb2 | 4025 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
4026 | lockdep_set_class_and_name(&rnp->lock, |
4027 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
4028 | raw_spin_lock_init(&rnp->fqslock); |
4029 | lockdep_set_class_and_name(&rnp->fqslock, | |
4030 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
4031 | rnp->gpnum = rsp->gpnum; |
4032 | rnp->completed = rsp->completed; | |
64db4cff PM |
4033 | rnp->qsmask = 0; |
4034 | rnp->qsmaskinit = 0; | |
4035 | rnp->grplo = j * cpustride; | |
4036 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
4037 | if (rnp->grphi >= nr_cpu_ids) |
4038 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
4039 | if (i == 0) { |
4040 | rnp->grpnum = 0; | |
4041 | rnp->grpmask = 0; | |
4042 | rnp->parent = NULL; | |
4043 | } else { | |
199977bf | 4044 | rnp->grpnum = j % levelspread[i - 1]; |
64db4cff PM |
4045 | rnp->grpmask = 1UL << rnp->grpnum; |
4046 | rnp->parent = rsp->level[i - 1] + | |
199977bf | 4047 | j / levelspread[i - 1]; |
64db4cff PM |
4048 | } |
4049 | rnp->level = i; | |
12f5f524 | 4050 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 4051 | rcu_init_one_nocb(rnp); |
64db4cff PM |
4052 | } |
4053 | } | |
0c34029a | 4054 | |
b3dbec76 | 4055 | init_waitqueue_head(&rsp->gp_wq); |
f885b7f2 | 4056 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 4057 | for_each_possible_cpu(i) { |
4a90a068 | 4058 | while (i > rnp->grphi) |
0c34029a | 4059 | rnp++; |
394f99a9 | 4060 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
4061 | rcu_boot_init_percpu_data(i, rsp); |
4062 | } | |
6ce75a23 | 4063 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
4064 | } |
4065 | ||
f885b7f2 PM |
4066 | /* |
4067 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 4068 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
4069 | * the ->node array in the rcu_state structure. |
4070 | */ | |
4071 | static void __init rcu_init_geometry(void) | |
4072 | { | |
026ad283 | 4073 | ulong d; |
f885b7f2 | 4074 | int i; |
05b84aec | 4075 | int rcu_capacity[RCU_NUM_LVLS]; |
f885b7f2 | 4076 | |
026ad283 PM |
4077 | /* |
4078 | * Initialize any unspecified boot parameters. | |
4079 | * The default values of jiffies_till_first_fqs and | |
4080 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
4081 | * value, which is a function of HZ, then adding one for each | |
4082 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
4083 | */ | |
4084 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
4085 | if (jiffies_till_first_fqs == ULONG_MAX) | |
4086 | jiffies_till_first_fqs = d; | |
4087 | if (jiffies_till_next_fqs == ULONG_MAX) | |
4088 | jiffies_till_next_fqs = d; | |
4089 | ||
f885b7f2 | 4090 | /* If the compile-time values are accurate, just leave. */ |
47d631af | 4091 | if (rcu_fanout_leaf == RCU_FANOUT_LEAF && |
b17c7035 | 4092 | nr_cpu_ids == NR_CPUS) |
f885b7f2 | 4093 | return; |
39479098 PM |
4094 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", |
4095 | rcu_fanout_leaf, nr_cpu_ids); | |
f885b7f2 | 4096 | |
75cf15a4 AG |
4097 | /* |
4098 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
4099 | * to increase the leaf-level fanout, not decrease it. Of course, | |
4100 | * the leaf-level fanout cannot exceed the number of bits in | |
4101 | * the rcu_node masks. Complain and fall back to the compile- | |
4102 | * time values if these limits are exceeded. | |
4103 | */ | |
4104 | if (rcu_fanout_leaf < RCU_FANOUT_LEAF || | |
4105 | rcu_fanout_leaf > sizeof(unsigned long) * 8) { | |
4106 | WARN_ON(1); | |
4107 | return; | |
4108 | } | |
4109 | ||
f885b7f2 PM |
4110 | /* |
4111 | * Compute number of nodes that can be handled an rcu_node tree | |
9618138b | 4112 | * with the given number of levels. |
f885b7f2 | 4113 | */ |
9618138b | 4114 | rcu_capacity[0] = rcu_fanout_leaf; |
05b84aec | 4115 | for (i = 1; i < RCU_NUM_LVLS; i++) |
05c5df31 | 4116 | rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT; |
f885b7f2 PM |
4117 | |
4118 | /* | |
75cf15a4 AG |
4119 | * The tree must be able to accommodate the configured number of CPUs. |
4120 | * If this limit is exceeded than we have a serious problem elsewhere. | |
f885b7f2 | 4121 | */ |
05b84aec | 4122 | if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) |
75cf15a4 | 4123 | panic("rcu_init_geometry: rcu_capacity[] is too small"); |
f885b7f2 | 4124 | |
679f9858 | 4125 | /* Calculate the number of levels in the tree. */ |
9618138b | 4126 | for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) { |
679f9858 | 4127 | } |
9618138b | 4128 | rcu_num_lvls = i + 1; |
679f9858 | 4129 | |
f885b7f2 | 4130 | /* Calculate the number of rcu_nodes at each level of the tree. */ |
679f9858 | 4131 | for (i = 0; i < rcu_num_lvls; i++) { |
9618138b | 4132 | int cap = rcu_capacity[(rcu_num_lvls - 1) - i]; |
679f9858 AG |
4133 | num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap); |
4134 | } | |
f885b7f2 PM |
4135 | |
4136 | /* Calculate the total number of rcu_node structures. */ | |
4137 | rcu_num_nodes = 0; | |
679f9858 | 4138 | for (i = 0; i < rcu_num_lvls; i++) |
f885b7f2 | 4139 | rcu_num_nodes += num_rcu_lvl[i]; |
f885b7f2 PM |
4140 | } |
4141 | ||
a3dc2948 PM |
4142 | /* |
4143 | * Dump out the structure of the rcu_node combining tree associated | |
4144 | * with the rcu_state structure referenced by rsp. | |
4145 | */ | |
4146 | static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp) | |
4147 | { | |
4148 | int level = 0; | |
4149 | struct rcu_node *rnp; | |
4150 | ||
4151 | pr_info("rcu_node tree layout dump\n"); | |
4152 | pr_info(" "); | |
4153 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
4154 | if (rnp->level != level) { | |
4155 | pr_cont("\n"); | |
4156 | pr_info(" "); | |
4157 | level = rnp->level; | |
4158 | } | |
4159 | pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum); | |
4160 | } | |
4161 | pr_cont("\n"); | |
4162 | } | |
4163 | ||
9f680ab4 | 4164 | void __init rcu_init(void) |
64db4cff | 4165 | { |
017c4261 | 4166 | int cpu; |
9f680ab4 | 4167 | |
47627678 PM |
4168 | rcu_early_boot_tests(); |
4169 | ||
f41d911f | 4170 | rcu_bootup_announce(); |
f885b7f2 | 4171 | rcu_init_geometry(); |
394f99a9 | 4172 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); |
69c8d28c | 4173 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
a3dc2948 PM |
4174 | if (dump_tree) |
4175 | rcu_dump_rcu_node_tree(&rcu_sched_state); | |
f41d911f | 4176 | __rcu_init_preempt(); |
b5b39360 | 4177 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
4178 | |
4179 | /* | |
4180 | * We don't need protection against CPU-hotplug here because | |
4181 | * this is called early in boot, before either interrupts | |
4182 | * or the scheduler are operational. | |
4183 | */ | |
4184 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 4185 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
4186 | for_each_online_cpu(cpu) |
4187 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
4188 | } |
4189 | ||
4102adab | 4190 | #include "tree_plugin.h" |