]>
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> |
f7f7bac9 | 57 | #include <linux/ftrace_event.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) \ | |
2723249a | 94 | 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 | |
e534165b | 113 | static struct rcu_state *rcu_state_p; |
6ce75a23 | 114 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 115 | |
f885b7f2 PM |
116 | /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */ |
117 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF; | |
7e5c2dfb | 118 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 PM |
119 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
120 | static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */ | |
121 | NUM_RCU_LVL_0, | |
122 | NUM_RCU_LVL_1, | |
123 | NUM_RCU_LVL_2, | |
124 | NUM_RCU_LVL_3, | |
125 | NUM_RCU_LVL_4, | |
126 | }; | |
127 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ | |
128 | ||
b0d30417 PM |
129 | /* |
130 | * The rcu_scheduler_active variable transitions from zero to one just | |
131 | * before the first task is spawned. So when this variable is zero, RCU | |
132 | * can assume that there is but one task, allowing RCU to (for example) | |
b44f6656 | 133 | * optimize synchronize_sched() to a simple barrier(). When this variable |
b0d30417 PM |
134 | * is one, RCU must actually do all the hard work required to detect real |
135 | * grace periods. This variable is also used to suppress boot-time false | |
136 | * positives from lockdep-RCU error checking. | |
137 | */ | |
bbad9379 PM |
138 | int rcu_scheduler_active __read_mostly; |
139 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
140 | ||
b0d30417 PM |
141 | /* |
142 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
143 | * during the early_initcall() processing, which is after the scheduler | |
144 | * is capable of creating new tasks. So RCU processing (for example, | |
145 | * creating tasks for RCU priority boosting) must be delayed until after | |
146 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
147 | * currently delay invocation of any RCU callbacks until after this point. | |
148 | * | |
149 | * It might later prove better for people registering RCU callbacks during | |
150 | * early boot to take responsibility for these callbacks, but one step at | |
151 | * a time. | |
152 | */ | |
153 | static int rcu_scheduler_fully_active __read_mostly; | |
154 | ||
0aa04b05 PM |
155 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); |
156 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); | |
5d01bbd1 | 157 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
158 | static void invoke_rcu_core(void); |
159 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 160 | |
a94844b2 PM |
161 | /* rcuc/rcub kthread realtime priority */ |
162 | static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO; | |
163 | module_param(kthread_prio, int, 0644); | |
164 | ||
8d7dc928 PM |
165 | /* Delay in jiffies for grace-period initialization delays, debug only. */ |
166 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT | |
167 | static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY; | |
37745d28 | 168 | module_param(gp_init_delay, int, 0644); |
8d7dc928 PM |
169 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ |
170 | static const int gp_init_delay; | |
171 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ | |
172 | #define PER_RCU_NODE_PERIOD 10 /* Number of grace periods between delays. */ | |
37745d28 | 173 | |
4a298656 PM |
174 | /* |
175 | * Track the rcutorture test sequence number and the update version | |
176 | * number within a given test. The rcutorture_testseq is incremented | |
177 | * on every rcutorture module load and unload, so has an odd value | |
178 | * when a test is running. The rcutorture_vernum is set to zero | |
179 | * when rcutorture starts and is incremented on each rcutorture update. | |
180 | * These variables enable correlating rcutorture output with the | |
181 | * RCU tracing information. | |
182 | */ | |
183 | unsigned long rcutorture_testseq; | |
184 | unsigned long rcutorture_vernum; | |
185 | ||
0aa04b05 PM |
186 | /* |
187 | * Compute the mask of online CPUs for the specified rcu_node structure. | |
188 | * This will not be stable unless the rcu_node structure's ->lock is | |
189 | * held, but the bit corresponding to the current CPU will be stable | |
190 | * in most contexts. | |
191 | */ | |
192 | unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) | |
193 | { | |
194 | return ACCESS_ONCE(rnp->qsmaskinitnext); | |
195 | } | |
196 | ||
fc2219d4 PM |
197 | /* |
198 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
199 | * permit this function to be invoked without holding the root rcu_node | |
200 | * structure's ->lock, but of course results can be subject to change. | |
201 | */ | |
202 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
203 | { | |
204 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
205 | } | |
206 | ||
b1f77b05 | 207 | /* |
d6714c22 | 208 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 209 | * how many quiescent states passed, just if there was at least |
d6714c22 | 210 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 211 | * The caller must have disabled preemption. |
b1f77b05 | 212 | */ |
284a8c93 | 213 | void rcu_sched_qs(void) |
b1f77b05 | 214 | { |
284a8c93 PM |
215 | if (!__this_cpu_read(rcu_sched_data.passed_quiesce)) { |
216 | trace_rcu_grace_period(TPS("rcu_sched"), | |
217 | __this_cpu_read(rcu_sched_data.gpnum), | |
218 | TPS("cpuqs")); | |
219 | __this_cpu_write(rcu_sched_data.passed_quiesce, 1); | |
220 | } | |
b1f77b05 IM |
221 | } |
222 | ||
284a8c93 | 223 | void rcu_bh_qs(void) |
b1f77b05 | 224 | { |
284a8c93 PM |
225 | if (!__this_cpu_read(rcu_bh_data.passed_quiesce)) { |
226 | trace_rcu_grace_period(TPS("rcu_bh"), | |
227 | __this_cpu_read(rcu_bh_data.gpnum), | |
228 | TPS("cpuqs")); | |
229 | __this_cpu_write(rcu_bh_data.passed_quiesce, 1); | |
230 | } | |
b1f77b05 | 231 | } |
64db4cff | 232 | |
4a81e832 PM |
233 | static DEFINE_PER_CPU(int, rcu_sched_qs_mask); |
234 | ||
235 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { | |
236 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, | |
237 | .dynticks = ATOMIC_INIT(1), | |
238 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
239 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
240 | .dynticks_idle = ATOMIC_INIT(1), | |
241 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
242 | }; | |
243 | ||
5cd37193 PM |
244 | DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr); |
245 | EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr); | |
246 | ||
4a81e832 PM |
247 | /* |
248 | * Let the RCU core know that this CPU has gone through the scheduler, | |
249 | * which is a quiescent state. This is called when the need for a | |
250 | * quiescent state is urgent, so we burn an atomic operation and full | |
251 | * memory barriers to let the RCU core know about it, regardless of what | |
252 | * this CPU might (or might not) do in the near future. | |
253 | * | |
254 | * We inform the RCU core by emulating a zero-duration dyntick-idle | |
255 | * period, which we in turn do by incrementing the ->dynticks counter | |
256 | * by two. | |
257 | */ | |
258 | static void rcu_momentary_dyntick_idle(void) | |
259 | { | |
260 | unsigned long flags; | |
261 | struct rcu_data *rdp; | |
262 | struct rcu_dynticks *rdtp; | |
263 | int resched_mask; | |
264 | struct rcu_state *rsp; | |
265 | ||
266 | local_irq_save(flags); | |
267 | ||
268 | /* | |
269 | * Yes, we can lose flag-setting operations. This is OK, because | |
270 | * the flag will be set again after some delay. | |
271 | */ | |
272 | resched_mask = raw_cpu_read(rcu_sched_qs_mask); | |
273 | raw_cpu_write(rcu_sched_qs_mask, 0); | |
274 | ||
275 | /* Find the flavor that needs a quiescent state. */ | |
276 | for_each_rcu_flavor(rsp) { | |
277 | rdp = raw_cpu_ptr(rsp->rda); | |
278 | if (!(resched_mask & rsp->flavor_mask)) | |
279 | continue; | |
280 | smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */ | |
281 | if (ACCESS_ONCE(rdp->mynode->completed) != | |
282 | ACCESS_ONCE(rdp->cond_resched_completed)) | |
283 | continue; | |
284 | ||
285 | /* | |
286 | * Pretend to be momentarily idle for the quiescent state. | |
287 | * This allows the grace-period kthread to record the | |
288 | * quiescent state, with no need for this CPU to do anything | |
289 | * further. | |
290 | */ | |
291 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
292 | smp_mb__before_atomic(); /* Earlier stuff before QS. */ | |
293 | atomic_add(2, &rdtp->dynticks); /* QS. */ | |
294 | smp_mb__after_atomic(); /* Later stuff after QS. */ | |
295 | break; | |
296 | } | |
297 | local_irq_restore(flags); | |
298 | } | |
299 | ||
25502a6c PM |
300 | /* |
301 | * Note a context switch. This is a quiescent state for RCU-sched, | |
302 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 303 | * The caller must have disabled preemption. |
25502a6c | 304 | */ |
38200cf2 | 305 | void rcu_note_context_switch(void) |
25502a6c | 306 | { |
f7f7bac9 | 307 | trace_rcu_utilization(TPS("Start context switch")); |
284a8c93 | 308 | rcu_sched_qs(); |
38200cf2 | 309 | rcu_preempt_note_context_switch(); |
4a81e832 PM |
310 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) |
311 | rcu_momentary_dyntick_idle(); | |
f7f7bac9 | 312 | trace_rcu_utilization(TPS("End context switch")); |
25502a6c | 313 | } |
29ce8310 | 314 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 315 | |
5cd37193 | 316 | /* |
1925d196 | 317 | * Register a quiescent state for all RCU flavors. If there is an |
5cd37193 PM |
318 | * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight |
319 | * dyntick-idle quiescent state visible to other CPUs (but only for those | |
1925d196 | 320 | * RCU flavors in desperate need of a quiescent state, which will normally |
5cd37193 PM |
321 | * be none of them). Either way, do a lightweight quiescent state for |
322 | * all RCU flavors. | |
323 | */ | |
324 | void rcu_all_qs(void) | |
325 | { | |
326 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) | |
327 | rcu_momentary_dyntick_idle(); | |
328 | this_cpu_inc(rcu_qs_ctr); | |
329 | } | |
330 | EXPORT_SYMBOL_GPL(rcu_all_qs); | |
331 | ||
878d7439 ED |
332 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
333 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
334 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 335 | |
878d7439 ED |
336 | module_param(blimit, long, 0444); |
337 | module_param(qhimark, long, 0444); | |
338 | module_param(qlowmark, long, 0444); | |
3d76c082 | 339 | |
026ad283 PM |
340 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
341 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
342 | |
343 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
344 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
345 | ||
4a81e832 PM |
346 | /* |
347 | * How long the grace period must be before we start recruiting | |
348 | * quiescent-state help from rcu_note_context_switch(). | |
349 | */ | |
350 | static ulong jiffies_till_sched_qs = HZ / 20; | |
351 | module_param(jiffies_till_sched_qs, ulong, 0644); | |
352 | ||
48a7639c | 353 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
910ee45d | 354 | struct rcu_data *rdp); |
217af2a2 PM |
355 | static void force_qs_rnp(struct rcu_state *rsp, |
356 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
357 | unsigned long *maxj), | |
358 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 359 | static void force_quiescent_state(struct rcu_state *rsp); |
e3950ecd | 360 | static int rcu_pending(void); |
64db4cff PM |
361 | |
362 | /* | |
917963d0 | 363 | * Return the number of RCU batches started thus far for debug & stats. |
64db4cff | 364 | */ |
917963d0 PM |
365 | unsigned long rcu_batches_started(void) |
366 | { | |
367 | return rcu_state_p->gpnum; | |
368 | } | |
369 | EXPORT_SYMBOL_GPL(rcu_batches_started); | |
370 | ||
371 | /* | |
372 | * Return the number of RCU-sched batches started thus far for debug & stats. | |
64db4cff | 373 | */ |
917963d0 PM |
374 | unsigned long rcu_batches_started_sched(void) |
375 | { | |
376 | return rcu_sched_state.gpnum; | |
377 | } | |
378 | EXPORT_SYMBOL_GPL(rcu_batches_started_sched); | |
379 | ||
380 | /* | |
381 | * Return the number of RCU BH batches started thus far for debug & stats. | |
382 | */ | |
383 | unsigned long rcu_batches_started_bh(void) | |
384 | { | |
385 | return rcu_bh_state.gpnum; | |
386 | } | |
387 | EXPORT_SYMBOL_GPL(rcu_batches_started_bh); | |
388 | ||
389 | /* | |
390 | * Return the number of RCU batches completed thus far for debug & stats. | |
391 | */ | |
392 | unsigned long rcu_batches_completed(void) | |
393 | { | |
394 | return rcu_state_p->completed; | |
395 | } | |
396 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
397 | ||
398 | /* | |
399 | * Return the number of RCU-sched batches completed thus far for debug & stats. | |
64db4cff | 400 | */ |
9733e4f0 | 401 | unsigned long rcu_batches_completed_sched(void) |
64db4cff | 402 | { |
d6714c22 | 403 | return rcu_sched_state.completed; |
64db4cff | 404 | } |
d6714c22 | 405 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
406 | |
407 | /* | |
917963d0 | 408 | * Return the number of RCU BH batches completed thus far for debug & stats. |
64db4cff | 409 | */ |
9733e4f0 | 410 | unsigned long rcu_batches_completed_bh(void) |
64db4cff PM |
411 | { |
412 | return rcu_bh_state.completed; | |
413 | } | |
414 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
415 | ||
a381d757 ACB |
416 | /* |
417 | * Force a quiescent state. | |
418 | */ | |
419 | void rcu_force_quiescent_state(void) | |
420 | { | |
e534165b | 421 | force_quiescent_state(rcu_state_p); |
a381d757 ACB |
422 | } |
423 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
424 | ||
bf66f18e PM |
425 | /* |
426 | * Force a quiescent state for RCU BH. | |
427 | */ | |
428 | void rcu_bh_force_quiescent_state(void) | |
429 | { | |
4cdfc175 | 430 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
431 | } |
432 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
433 | ||
e7580f33 PM |
434 | /* |
435 | * Force a quiescent state for RCU-sched. | |
436 | */ | |
437 | void rcu_sched_force_quiescent_state(void) | |
438 | { | |
439 | force_quiescent_state(&rcu_sched_state); | |
440 | } | |
441 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
442 | ||
afea227f PM |
443 | /* |
444 | * Show the state of the grace-period kthreads. | |
445 | */ | |
446 | void show_rcu_gp_kthreads(void) | |
447 | { | |
448 | struct rcu_state *rsp; | |
449 | ||
450 | for_each_rcu_flavor(rsp) { | |
451 | pr_info("%s: wait state: %d ->state: %#lx\n", | |
452 | rsp->name, rsp->gp_state, rsp->gp_kthread->state); | |
453 | /* sched_show_task(rsp->gp_kthread); */ | |
454 | } | |
455 | } | |
456 | EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); | |
457 | ||
4a298656 PM |
458 | /* |
459 | * Record the number of times rcutorture tests have been initiated and | |
460 | * terminated. This information allows the debugfs tracing stats to be | |
461 | * correlated to the rcutorture messages, even when the rcutorture module | |
462 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
463 | * store this state in rcutorture itself. | |
464 | */ | |
465 | void rcutorture_record_test_transition(void) | |
466 | { | |
467 | rcutorture_testseq++; | |
468 | rcutorture_vernum = 0; | |
469 | } | |
470 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
471 | ||
ad0dc7f9 PM |
472 | /* |
473 | * Send along grace-period-related data for rcutorture diagnostics. | |
474 | */ | |
475 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
476 | unsigned long *gpnum, unsigned long *completed) | |
477 | { | |
478 | struct rcu_state *rsp = NULL; | |
479 | ||
480 | switch (test_type) { | |
481 | case RCU_FLAVOR: | |
e534165b | 482 | rsp = rcu_state_p; |
ad0dc7f9 PM |
483 | break; |
484 | case RCU_BH_FLAVOR: | |
485 | rsp = &rcu_bh_state; | |
486 | break; | |
487 | case RCU_SCHED_FLAVOR: | |
488 | rsp = &rcu_sched_state; | |
489 | break; | |
490 | default: | |
491 | break; | |
492 | } | |
493 | if (rsp != NULL) { | |
494 | *flags = ACCESS_ONCE(rsp->gp_flags); | |
495 | *gpnum = ACCESS_ONCE(rsp->gpnum); | |
496 | *completed = ACCESS_ONCE(rsp->completed); | |
497 | return; | |
498 | } | |
499 | *flags = 0; | |
500 | *gpnum = 0; | |
501 | *completed = 0; | |
502 | } | |
503 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
504 | ||
4a298656 PM |
505 | /* |
506 | * Record the number of writer passes through the current rcutorture test. | |
507 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
508 | * messages. | |
509 | */ | |
510 | void rcutorture_record_progress(unsigned long vernum) | |
511 | { | |
512 | rcutorture_vernum++; | |
513 | } | |
514 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
515 | ||
64db4cff PM |
516 | /* |
517 | * Does the CPU have callbacks ready to be invoked? | |
518 | */ | |
519 | static int | |
520 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
521 | { | |
3fbfbf7a PM |
522 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
523 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
524 | } |
525 | ||
365187fb PM |
526 | /* |
527 | * Return the root node of the specified rcu_state structure. | |
528 | */ | |
529 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
530 | { | |
531 | return &rsp->node[0]; | |
532 | } | |
533 | ||
534 | /* | |
535 | * Is there any need for future grace periods? | |
536 | * Interrupts must be disabled. If the caller does not hold the root | |
537 | * rnp_node structure's ->lock, the results are advisory only. | |
538 | */ | |
539 | static int rcu_future_needs_gp(struct rcu_state *rsp) | |
540 | { | |
541 | struct rcu_node *rnp = rcu_get_root(rsp); | |
542 | int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1; | |
543 | int *fp = &rnp->need_future_gp[idx]; | |
544 | ||
545 | return ACCESS_ONCE(*fp); | |
546 | } | |
547 | ||
64db4cff | 548 | /* |
dc35c893 PM |
549 | * Does the current CPU require a not-yet-started grace period? |
550 | * The caller must have disabled interrupts to prevent races with | |
551 | * normal callback registry. | |
64db4cff PM |
552 | */ |
553 | static int | |
554 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
555 | { | |
dc35c893 | 556 | int i; |
3fbfbf7a | 557 | |
dc35c893 PM |
558 | if (rcu_gp_in_progress(rsp)) |
559 | return 0; /* No, a grace period is already in progress. */ | |
365187fb | 560 | if (rcu_future_needs_gp(rsp)) |
34ed6246 | 561 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
562 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
563 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
564 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
565 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
566 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
567 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
568 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), | |
569 | rdp->nxtcompleted[i])) | |
570 | return 1; /* Yes, CBs for future grace period. */ | |
571 | return 0; /* No grace period needed. */ | |
64db4cff PM |
572 | } |
573 | ||
9b2e4f18 | 574 | /* |
adf5091e | 575 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
576 | * |
577 | * If the new value of the ->dynticks_nesting counter now is zero, | |
578 | * we really have entered idle, and must do the appropriate accounting. | |
579 | * The caller must have disabled interrupts. | |
580 | */ | |
28ced795 | 581 | static void rcu_eqs_enter_common(long long oldval, bool user) |
9b2e4f18 | 582 | { |
96d3fd0d PM |
583 | struct rcu_state *rsp; |
584 | struct rcu_data *rdp; | |
28ced795 | 585 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
96d3fd0d | 586 | |
f7f7bac9 | 587 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 588 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
589 | struct task_struct *idle __maybe_unused = |
590 | idle_task(smp_processor_id()); | |
0989cb46 | 591 | |
f7f7bac9 | 592 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 593 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
594 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
595 | current->pid, current->comm, | |
596 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 597 | } |
96d3fd0d PM |
598 | for_each_rcu_flavor(rsp) { |
599 | rdp = this_cpu_ptr(rsp->rda); | |
600 | do_nocb_deferred_wakeup(rdp); | |
601 | } | |
198bbf81 | 602 | rcu_prepare_for_idle(); |
9b2e4f18 | 603 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 604 | smp_mb__before_atomic(); /* See above. */ |
9b2e4f18 | 605 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 606 | smp_mb__after_atomic(); /* Force ordering with next sojourn. */ |
9b2e4f18 | 607 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
176f8f7a | 608 | rcu_dynticks_task_enter(); |
c44e2cdd PM |
609 | |
610 | /* | |
adf5091e | 611 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
612 | * in an RCU read-side critical section. |
613 | */ | |
614 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
615 | "Illegal idle entry in RCU read-side critical section."); | |
616 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
617 | "Illegal idle entry in RCU-bh read-side critical section."); | |
618 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
619 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 620 | } |
64db4cff | 621 | |
adf5091e FW |
622 | /* |
623 | * Enter an RCU extended quiescent state, which can be either the | |
624 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 625 | */ |
adf5091e | 626 | static void rcu_eqs_enter(bool user) |
64db4cff | 627 | { |
4145fa7f | 628 | long long oldval; |
64db4cff PM |
629 | struct rcu_dynticks *rdtp; |
630 | ||
c9d4b0af | 631 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 632 | oldval = rdtp->dynticks_nesting; |
29e37d81 | 633 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
3a592405 | 634 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { |
29e37d81 | 635 | rdtp->dynticks_nesting = 0; |
28ced795 | 636 | rcu_eqs_enter_common(oldval, user); |
3a592405 | 637 | } else { |
29e37d81 | 638 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; |
3a592405 | 639 | } |
64db4cff | 640 | } |
adf5091e FW |
641 | |
642 | /** | |
643 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
644 | * | |
645 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
646 | * read-side critical sections can occur. (Though RCU read-side | |
647 | * critical sections can occur in irq handlers in idle, a possibility | |
648 | * handled by irq_enter() and irq_exit().) | |
649 | * | |
650 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
651 | * the possibility of usermode upcalls having messed up our count | |
652 | * of interrupt nesting level during the prior busy period. | |
653 | */ | |
654 | void rcu_idle_enter(void) | |
655 | { | |
c5d900bf FW |
656 | unsigned long flags; |
657 | ||
658 | local_irq_save(flags); | |
cb349ca9 | 659 | rcu_eqs_enter(false); |
28ced795 | 660 | rcu_sysidle_enter(0); |
c5d900bf | 661 | local_irq_restore(flags); |
adf5091e | 662 | } |
8a2ecf47 | 663 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 664 | |
2b1d5024 | 665 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
666 | /** |
667 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
668 | * | |
669 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
670 | * is permitted between this call and rcu_user_exit(). This way the | |
671 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
672 | * when the CPU runs in userspace. | |
673 | */ | |
674 | void rcu_user_enter(void) | |
675 | { | |
91d1aa43 | 676 | rcu_eqs_enter(1); |
adf5091e | 677 | } |
2b1d5024 | 678 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 679 | |
9b2e4f18 PM |
680 | /** |
681 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
682 | * | |
683 | * Exit from an interrupt handler, which might possibly result in entering | |
684 | * idle mode, in other words, leaving the mode in which read-side critical | |
685 | * sections can occur. | |
64db4cff | 686 | * |
9b2e4f18 PM |
687 | * This code assumes that the idle loop never does anything that might |
688 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
689 | * architecture violates this assumption, RCU will give you what you | |
690 | * deserve, good and hard. But very infrequently and irreproducibly. | |
691 | * | |
692 | * Use things like work queues to work around this limitation. | |
693 | * | |
694 | * You have been warned. | |
64db4cff | 695 | */ |
9b2e4f18 | 696 | void rcu_irq_exit(void) |
64db4cff PM |
697 | { |
698 | unsigned long flags; | |
4145fa7f | 699 | long long oldval; |
64db4cff PM |
700 | struct rcu_dynticks *rdtp; |
701 | ||
702 | local_irq_save(flags); | |
c9d4b0af | 703 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 704 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
705 | rdtp->dynticks_nesting--; |
706 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 | 707 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 708 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 709 | else |
28ced795 CL |
710 | rcu_eqs_enter_common(oldval, true); |
711 | rcu_sysidle_enter(1); | |
9b2e4f18 PM |
712 | local_irq_restore(flags); |
713 | } | |
714 | ||
715 | /* | |
adf5091e | 716 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
717 | * |
718 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
719 | * we really have exited idle, and must do the appropriate accounting. | |
720 | * The caller must have disabled interrupts. | |
721 | */ | |
28ced795 | 722 | static void rcu_eqs_exit_common(long long oldval, int user) |
9b2e4f18 | 723 | { |
28ced795 CL |
724 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
725 | ||
176f8f7a | 726 | rcu_dynticks_task_exit(); |
4e857c58 | 727 | smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */ |
23b5c8fa PM |
728 | atomic_inc(&rdtp->dynticks); |
729 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
4e857c58 | 730 | smp_mb__after_atomic(); /* See above. */ |
23b5c8fa | 731 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); |
8fa7845d | 732 | rcu_cleanup_after_idle(); |
f7f7bac9 | 733 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 734 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
735 | struct task_struct *idle __maybe_unused = |
736 | idle_task(smp_processor_id()); | |
0989cb46 | 737 | |
f7f7bac9 | 738 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 739 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 740 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
741 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
742 | current->pid, current->comm, | |
743 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
744 | } |
745 | } | |
746 | ||
adf5091e FW |
747 | /* |
748 | * Exit an RCU extended quiescent state, which can be either the | |
749 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 750 | */ |
adf5091e | 751 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 752 | { |
9b2e4f18 PM |
753 | struct rcu_dynticks *rdtp; |
754 | long long oldval; | |
755 | ||
c9d4b0af | 756 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 | 757 | oldval = rdtp->dynticks_nesting; |
29e37d81 | 758 | WARN_ON_ONCE(oldval < 0); |
3a592405 | 759 | if (oldval & DYNTICK_TASK_NEST_MASK) { |
29e37d81 | 760 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; |
3a592405 | 761 | } else { |
29e37d81 | 762 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
28ced795 | 763 | rcu_eqs_exit_common(oldval, user); |
3a592405 | 764 | } |
9b2e4f18 | 765 | } |
adf5091e FW |
766 | |
767 | /** | |
768 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
769 | * | |
770 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
771 | * read-side critical sections can occur. | |
772 | * | |
773 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
774 | * allow for the possibility of usermode upcalls messing up our count | |
775 | * of interrupt nesting level during the busy period that is just | |
776 | * now starting. | |
777 | */ | |
778 | void rcu_idle_exit(void) | |
779 | { | |
c5d900bf FW |
780 | unsigned long flags; |
781 | ||
782 | local_irq_save(flags); | |
cb349ca9 | 783 | rcu_eqs_exit(false); |
28ced795 | 784 | rcu_sysidle_exit(0); |
c5d900bf | 785 | local_irq_restore(flags); |
adf5091e | 786 | } |
8a2ecf47 | 787 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 788 | |
2b1d5024 | 789 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
790 | /** |
791 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
792 | * | |
793 | * Exit RCU idle mode while entering the kernel because it can | |
794 | * run a RCU read side critical section anytime. | |
795 | */ | |
796 | void rcu_user_exit(void) | |
797 | { | |
91d1aa43 | 798 | rcu_eqs_exit(1); |
adf5091e | 799 | } |
2b1d5024 | 800 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 801 | |
9b2e4f18 PM |
802 | /** |
803 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
804 | * | |
805 | * Enter an interrupt handler, which might possibly result in exiting | |
806 | * idle mode, in other words, entering the mode in which read-side critical | |
807 | * sections can occur. | |
808 | * | |
809 | * Note that the Linux kernel is fully capable of entering an interrupt | |
810 | * handler that it never exits, for example when doing upcalls to | |
811 | * user mode! This code assumes that the idle loop never does upcalls to | |
812 | * user mode. If your architecture does do upcalls from the idle loop (or | |
813 | * does anything else that results in unbalanced calls to the irq_enter() | |
814 | * and irq_exit() functions), RCU will give you what you deserve, good | |
815 | * and hard. But very infrequently and irreproducibly. | |
816 | * | |
817 | * Use things like work queues to work around this limitation. | |
818 | * | |
819 | * You have been warned. | |
820 | */ | |
821 | void rcu_irq_enter(void) | |
822 | { | |
823 | unsigned long flags; | |
824 | struct rcu_dynticks *rdtp; | |
825 | long long oldval; | |
826 | ||
827 | local_irq_save(flags); | |
c9d4b0af | 828 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 PM |
829 | oldval = rdtp->dynticks_nesting; |
830 | rdtp->dynticks_nesting++; | |
831 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 | 832 | if (oldval) |
f7f7bac9 | 833 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 834 | else |
28ced795 CL |
835 | rcu_eqs_exit_common(oldval, true); |
836 | rcu_sysidle_exit(1); | |
64db4cff | 837 | local_irq_restore(flags); |
64db4cff PM |
838 | } |
839 | ||
840 | /** | |
841 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
842 | * | |
734d1680 PM |
843 | * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and |
844 | * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know | |
845 | * that the CPU is active. This implementation permits nested NMIs, as | |
846 | * long as the nesting level does not overflow an int. (You will probably | |
847 | * run out of stack space first.) | |
64db4cff PM |
848 | */ |
849 | void rcu_nmi_enter(void) | |
850 | { | |
c9d4b0af | 851 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
734d1680 | 852 | int incby = 2; |
64db4cff | 853 | |
734d1680 PM |
854 | /* Complain about underflow. */ |
855 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0); | |
856 | ||
857 | /* | |
858 | * If idle from RCU viewpoint, atomically increment ->dynticks | |
859 | * to mark non-idle and increment ->dynticks_nmi_nesting by one. | |
860 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means | |
861 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed | |
862 | * to be in the outermost NMI handler that interrupted an RCU-idle | |
863 | * period (observation due to Andy Lutomirski). | |
864 | */ | |
865 | if (!(atomic_read(&rdtp->dynticks) & 0x1)) { | |
866 | smp_mb__before_atomic(); /* Force delay from prior write. */ | |
867 | atomic_inc(&rdtp->dynticks); | |
868 | /* atomic_inc() before later RCU read-side crit sects */ | |
869 | smp_mb__after_atomic(); /* See above. */ | |
870 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
871 | incby = 1; | |
872 | } | |
873 | rdtp->dynticks_nmi_nesting += incby; | |
874 | barrier(); | |
64db4cff PM |
875 | } |
876 | ||
877 | /** | |
878 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
879 | * | |
734d1680 PM |
880 | * If we are returning from the outermost NMI handler that interrupted an |
881 | * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting | |
882 | * to let the RCU grace-period handling know that the CPU is back to | |
883 | * being RCU-idle. | |
64db4cff PM |
884 | */ |
885 | void rcu_nmi_exit(void) | |
886 | { | |
c9d4b0af | 887 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 888 | |
734d1680 PM |
889 | /* |
890 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. | |
891 | * (We are exiting an NMI handler, so RCU better be paying attention | |
892 | * to us!) | |
893 | */ | |
894 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0); | |
895 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
896 | ||
897 | /* | |
898 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so | |
899 | * leave it in non-RCU-idle state. | |
900 | */ | |
901 | if (rdtp->dynticks_nmi_nesting != 1) { | |
902 | rdtp->dynticks_nmi_nesting -= 2; | |
64db4cff | 903 | return; |
734d1680 PM |
904 | } |
905 | ||
906 | /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ | |
907 | rdtp->dynticks_nmi_nesting = 0; | |
23b5c8fa | 908 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 909 | smp_mb__before_atomic(); /* See above. */ |
23b5c8fa | 910 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 911 | smp_mb__after_atomic(); /* Force delay to next write. */ |
23b5c8fa | 912 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
64db4cff PM |
913 | } |
914 | ||
915 | /** | |
5c173eb8 PM |
916 | * __rcu_is_watching - are RCU read-side critical sections safe? |
917 | * | |
918 | * Return true if RCU is watching the running CPU, which means that | |
919 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
920 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
921 | * least disabled preemption. | |
922 | */ | |
9418fb20 | 923 | bool notrace __rcu_is_watching(void) |
5c173eb8 PM |
924 | { |
925 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
926 | } | |
927 | ||
928 | /** | |
929 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
64db4cff | 930 | * |
9b2e4f18 | 931 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 932 | * or NMI handler, return true. |
64db4cff | 933 | */ |
9418fb20 | 934 | bool notrace rcu_is_watching(void) |
64db4cff | 935 | { |
f534ed1f | 936 | bool ret; |
34240697 PM |
937 | |
938 | preempt_disable(); | |
5c173eb8 | 939 | ret = __rcu_is_watching(); |
34240697 PM |
940 | preempt_enable(); |
941 | return ret; | |
64db4cff | 942 | } |
5c173eb8 | 943 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 944 | |
62fde6ed | 945 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
946 | |
947 | /* | |
948 | * Is the current CPU online? Disable preemption to avoid false positives | |
949 | * that could otherwise happen due to the current CPU number being sampled, | |
950 | * this task being preempted, its old CPU being taken offline, resuming | |
951 | * on some other CPU, then determining that its old CPU is now offline. | |
952 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
953 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
954 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
955 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
956 | * offline to continue to use RCU for one jiffy after marking itself | |
957 | * offline in the cpu_online_mask. This leniency is necessary given the | |
958 | * non-atomic nature of the online and offline processing, for example, | |
959 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
960 | * notifiers. | |
961 | * | |
962 | * This is also why RCU internally marks CPUs online during the | |
963 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
964 | * |
965 | * Disable checking if in an NMI handler because we cannot safely report | |
966 | * errors from NMI handlers anyway. | |
967 | */ | |
968 | bool rcu_lockdep_current_cpu_online(void) | |
969 | { | |
2036d94a PM |
970 | struct rcu_data *rdp; |
971 | struct rcu_node *rnp; | |
c0d6d01b PM |
972 | bool ret; |
973 | ||
974 | if (in_nmi()) | |
f6f7ee9a | 975 | return true; |
c0d6d01b | 976 | preempt_disable(); |
c9d4b0af | 977 | rdp = this_cpu_ptr(&rcu_sched_data); |
2036d94a | 978 | rnp = rdp->mynode; |
0aa04b05 | 979 | ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) || |
c0d6d01b PM |
980 | !rcu_scheduler_fully_active; |
981 | preempt_enable(); | |
982 | return ret; | |
983 | } | |
984 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
985 | ||
62fde6ed | 986 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 987 | |
64db4cff | 988 | /** |
9b2e4f18 | 989 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 990 | * |
9b2e4f18 PM |
991 | * If the current CPU is idle or running at a first-level (not nested) |
992 | * interrupt from idle, return true. The caller must have at least | |
993 | * disabled preemption. | |
64db4cff | 994 | */ |
62e3cb14 | 995 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 996 | { |
c9d4b0af | 997 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
64db4cff PM |
998 | } |
999 | ||
64db4cff PM |
1000 | /* |
1001 | * Snapshot the specified CPU's dynticks counter so that we can later | |
1002 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 1003 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 1004 | */ |
217af2a2 PM |
1005 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
1006 | bool *isidle, unsigned long *maxj) | |
64db4cff | 1007 | { |
23b5c8fa | 1008 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 1009 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
7941dbde ACB |
1010 | if ((rdp->dynticks_snap & 0x1) == 0) { |
1011 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); | |
1012 | return 1; | |
1013 | } else { | |
e3663b10 PM |
1014 | if (ULONG_CMP_LT(ACCESS_ONCE(rdp->gpnum) + ULONG_MAX / 4, |
1015 | rdp->mynode->gpnum)) | |
1016 | ACCESS_ONCE(rdp->gpwrap) = true; | |
7941dbde ACB |
1017 | return 0; |
1018 | } | |
64db4cff PM |
1019 | } |
1020 | ||
1021 | /* | |
1022 | * Return true if the specified CPU has passed through a quiescent | |
1023 | * state by virtue of being in or having passed through an dynticks | |
1024 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 1025 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 1026 | */ |
217af2a2 PM |
1027 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
1028 | bool *isidle, unsigned long *maxj) | |
64db4cff | 1029 | { |
7eb4f455 | 1030 | unsigned int curr; |
4a81e832 | 1031 | int *rcrmp; |
7eb4f455 | 1032 | unsigned int snap; |
64db4cff | 1033 | |
7eb4f455 PM |
1034 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
1035 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
1036 | |
1037 | /* | |
1038 | * If the CPU passed through or entered a dynticks idle phase with | |
1039 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
1040 | * already acknowledged the request to pass through a quiescent | |
1041 | * state. Either way, that CPU cannot possibly be in an RCU | |
1042 | * read-side critical section that started before the beginning | |
1043 | * of the current RCU grace period. | |
1044 | */ | |
7eb4f455 | 1045 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 1046 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
1047 | rdp->dynticks_fqs++; |
1048 | return 1; | |
1049 | } | |
1050 | ||
a82dcc76 PM |
1051 | /* |
1052 | * Check for the CPU being offline, but only if the grace period | |
1053 | * is old enough. We don't need to worry about the CPU changing | |
1054 | * state: If we see it offline even once, it has been through a | |
1055 | * quiescent state. | |
1056 | * | |
1057 | * The reason for insisting that the grace period be at least | |
1058 | * one jiffy old is that CPUs that are not quite online and that | |
1059 | * have just gone offline can still execute RCU read-side critical | |
1060 | * sections. | |
1061 | */ | |
1062 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
1063 | return 0; /* Grace period is not old enough. */ | |
1064 | barrier(); | |
1065 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 1066 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
1067 | rdp->offline_fqs++; |
1068 | return 1; | |
1069 | } | |
65d798f0 PM |
1070 | |
1071 | /* | |
4a81e832 PM |
1072 | * A CPU running for an extended time within the kernel can |
1073 | * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode, | |
1074 | * even context-switching back and forth between a pair of | |
1075 | * in-kernel CPU-bound tasks cannot advance grace periods. | |
1076 | * So if the grace period is old enough, make the CPU pay attention. | |
1077 | * Note that the unsynchronized assignments to the per-CPU | |
1078 | * rcu_sched_qs_mask variable are safe. Yes, setting of | |
1079 | * bits can be lost, but they will be set again on the next | |
1080 | * force-quiescent-state pass. So lost bit sets do not result | |
1081 | * in incorrect behavior, merely in a grace period lasting | |
1082 | * a few jiffies longer than it might otherwise. Because | |
1083 | * there are at most four threads involved, and because the | |
1084 | * updates are only once every few jiffies, the probability of | |
1085 | * lossage (and thus of slight grace-period extension) is | |
1086 | * quite low. | |
1087 | * | |
1088 | * Note that if the jiffies_till_sched_qs boot/sysfs parameter | |
1089 | * is set too high, we override with half of the RCU CPU stall | |
1090 | * warning delay. | |
6193c76a | 1091 | */ |
4a81e832 PM |
1092 | rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu); |
1093 | if (ULONG_CMP_GE(jiffies, | |
1094 | rdp->rsp->gp_start + jiffies_till_sched_qs) || | |
cb1e78cf | 1095 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { |
4a81e832 PM |
1096 | if (!(ACCESS_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) { |
1097 | ACCESS_ONCE(rdp->cond_resched_completed) = | |
1098 | ACCESS_ONCE(rdp->mynode->completed); | |
1099 | smp_mb(); /* ->cond_resched_completed before *rcrmp. */ | |
1100 | ACCESS_ONCE(*rcrmp) = | |
1101 | ACCESS_ONCE(*rcrmp) + rdp->rsp->flavor_mask; | |
1102 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1103 | rdp->rsp->jiffies_resched += 5; /* Enable beating. */ | |
1104 | } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { | |
1105 | /* Time to beat on that CPU again! */ | |
1106 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1107 | rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */ | |
1108 | } | |
6193c76a PM |
1109 | } |
1110 | ||
a82dcc76 | 1111 | return 0; |
64db4cff PM |
1112 | } |
1113 | ||
64db4cff PM |
1114 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
1115 | { | |
cb1e78cf | 1116 | unsigned long j = jiffies; |
6193c76a | 1117 | unsigned long j1; |
26cdfedf PM |
1118 | |
1119 | rsp->gp_start = j; | |
1120 | smp_wmb(); /* Record start time before stall time. */ | |
6193c76a | 1121 | j1 = rcu_jiffies_till_stall_check(); |
4fc5b755 | 1122 | ACCESS_ONCE(rsp->jiffies_stall) = j + j1; |
6193c76a | 1123 | rsp->jiffies_resched = j + j1 / 2; |
fc908ed3 | 1124 | rsp->n_force_qs_gpstart = ACCESS_ONCE(rsp->n_force_qs); |
64db4cff PM |
1125 | } |
1126 | ||
fb81a44b PM |
1127 | /* |
1128 | * Complain about starvation of grace-period kthread. | |
1129 | */ | |
1130 | static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp) | |
1131 | { | |
1132 | unsigned long gpa; | |
1133 | unsigned long j; | |
1134 | ||
1135 | j = jiffies; | |
1136 | gpa = ACCESS_ONCE(rsp->gp_activity); | |
1137 | if (j - gpa > 2 * HZ) | |
1138 | pr_err("%s kthread starved for %ld jiffies!\n", | |
1139 | rsp->name, j - gpa); | |
64db4cff PM |
1140 | } |
1141 | ||
b637a328 | 1142 | /* |
bc1dce51 | 1143 | * Dump stacks of all tasks running on stalled CPUs. |
b637a328 PM |
1144 | */ |
1145 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
1146 | { | |
1147 | int cpu; | |
1148 | unsigned long flags; | |
1149 | struct rcu_node *rnp; | |
1150 | ||
1151 | rcu_for_each_leaf_node(rsp, rnp) { | |
1152 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1153 | if (rnp->qsmask != 0) { | |
1154 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1155 | if (rnp->qsmask & (1UL << cpu)) | |
1156 | dump_cpu_task(rnp->grplo + cpu); | |
1157 | } | |
1158 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1159 | } | |
1160 | } | |
1161 | ||
6ccd2ecd | 1162 | static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum) |
64db4cff PM |
1163 | { |
1164 | int cpu; | |
1165 | long delta; | |
1166 | unsigned long flags; | |
6ccd2ecd PM |
1167 | unsigned long gpa; |
1168 | unsigned long j; | |
285fe294 | 1169 | int ndetected = 0; |
64db4cff | 1170 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 1171 | long totqlen = 0; |
64db4cff PM |
1172 | |
1173 | /* Only let one CPU complain about others per time interval. */ | |
1174 | ||
1304afb2 | 1175 | raw_spin_lock_irqsave(&rnp->lock, flags); |
4fc5b755 | 1176 | delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall); |
fc2219d4 | 1177 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 1178 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1179 | return; |
1180 | } | |
4fc5b755 | 1181 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 1182 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1183 | |
8cdd32a9 PM |
1184 | /* |
1185 | * OK, time to rat on our buddy... | |
1186 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1187 | * RCU CPU stall warnings. | |
1188 | */ | |
d7f3e207 | 1189 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 1190 | rsp->name); |
a858af28 | 1191 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 1192 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 1193 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 1194 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
1195 | if (rnp->qsmask != 0) { |
1196 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1197 | if (rnp->qsmask & (1UL << cpu)) { | |
1198 | print_cpu_stall_info(rsp, | |
1199 | rnp->grplo + cpu); | |
1200 | ndetected++; | |
1201 | } | |
1202 | } | |
3acd9eb3 | 1203 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1204 | } |
a858af28 | 1205 | |
a858af28 | 1206 | print_cpu_stall_info_end(); |
53bb857c PM |
1207 | for_each_possible_cpu(cpu) |
1208 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e | 1209 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", |
eee05882 | 1210 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
83ebe63e | 1211 | (long)rsp->gpnum, (long)rsp->completed, totqlen); |
6ccd2ecd | 1212 | if (ndetected) { |
b637a328 | 1213 | rcu_dump_cpu_stacks(rsp); |
6ccd2ecd PM |
1214 | } else { |
1215 | if (ACCESS_ONCE(rsp->gpnum) != gpnum || | |
1216 | ACCESS_ONCE(rsp->completed) == gpnum) { | |
1217 | pr_err("INFO: Stall ended before state dump start\n"); | |
1218 | } else { | |
1219 | j = jiffies; | |
1220 | gpa = ACCESS_ONCE(rsp->gp_activity); | |
237a0f21 | 1221 | pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n", |
6ccd2ecd | 1222 | rsp->name, j - gpa, j, gpa, |
237a0f21 PM |
1223 | jiffies_till_next_fqs, |
1224 | rcu_get_root(rsp)->qsmask); | |
6ccd2ecd PM |
1225 | /* In this case, the current CPU might be at fault. */ |
1226 | sched_show_task(current); | |
1227 | } | |
1228 | } | |
c1dc0b9c | 1229 | |
4cdfc175 | 1230 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
1231 | rcu_print_detail_task_stall(rsp); |
1232 | ||
fb81a44b PM |
1233 | rcu_check_gp_kthread_starvation(rsp); |
1234 | ||
4cdfc175 | 1235 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
1236 | } |
1237 | ||
1238 | static void print_cpu_stall(struct rcu_state *rsp) | |
1239 | { | |
53bb857c | 1240 | int cpu; |
64db4cff PM |
1241 | unsigned long flags; |
1242 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 1243 | long totqlen = 0; |
64db4cff | 1244 | |
8cdd32a9 PM |
1245 | /* |
1246 | * OK, time to rat on ourselves... | |
1247 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1248 | * RCU CPU stall warnings. | |
1249 | */ | |
d7f3e207 | 1250 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
1251 | print_cpu_stall_info_begin(); |
1252 | print_cpu_stall_info(rsp, smp_processor_id()); | |
1253 | print_cpu_stall_info_end(); | |
53bb857c PM |
1254 | for_each_possible_cpu(cpu) |
1255 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e PM |
1256 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", |
1257 | jiffies - rsp->gp_start, | |
1258 | (long)rsp->gpnum, (long)rsp->completed, totqlen); | |
fb81a44b PM |
1259 | |
1260 | rcu_check_gp_kthread_starvation(rsp); | |
1261 | ||
bc1dce51 | 1262 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 1263 | |
1304afb2 | 1264 | raw_spin_lock_irqsave(&rnp->lock, flags); |
4fc5b755 IM |
1265 | if (ULONG_CMP_GE(jiffies, ACCESS_ONCE(rsp->jiffies_stall))) |
1266 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + | |
6bfc09e2 | 1267 | 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 1268 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 1269 | |
b021fe3e PZ |
1270 | /* |
1271 | * Attempt to revive the RCU machinery by forcing a context switch. | |
1272 | * | |
1273 | * A context switch would normally allow the RCU state machine to make | |
1274 | * progress and it could be we're stuck in kernel space without context | |
1275 | * switches for an entirely unreasonable amount of time. | |
1276 | */ | |
1277 | resched_cpu(smp_processor_id()); | |
64db4cff PM |
1278 | } |
1279 | ||
1280 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
1281 | { | |
26cdfedf PM |
1282 | unsigned long completed; |
1283 | unsigned long gpnum; | |
1284 | unsigned long gps; | |
bad6e139 PM |
1285 | unsigned long j; |
1286 | unsigned long js; | |
64db4cff PM |
1287 | struct rcu_node *rnp; |
1288 | ||
26cdfedf | 1289 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
c68de209 | 1290 | return; |
cb1e78cf | 1291 | j = jiffies; |
26cdfedf PM |
1292 | |
1293 | /* | |
1294 | * Lots of memory barriers to reject false positives. | |
1295 | * | |
1296 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
1297 | * then rsp->gp_start, and finally rsp->completed. These values | |
1298 | * are updated in the opposite order with memory barriers (or | |
1299 | * equivalent) during grace-period initialization and cleanup. | |
1300 | * Now, a false positive can occur if we get an new value of | |
1301 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
1302 | * the memory barriers, the only way that this can happen is if one | |
1303 | * grace period ends and another starts between these two fetches. | |
1304 | * Detect this by comparing rsp->completed with the previous fetch | |
1305 | * from rsp->gpnum. | |
1306 | * | |
1307 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
1308 | * and rsp->gp_start suffice to forestall false positives. | |
1309 | */ | |
1310 | gpnum = ACCESS_ONCE(rsp->gpnum); | |
1311 | smp_rmb(); /* Pick up ->gpnum first... */ | |
bad6e139 | 1312 | js = ACCESS_ONCE(rsp->jiffies_stall); |
26cdfedf PM |
1313 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
1314 | gps = ACCESS_ONCE(rsp->gp_start); | |
1315 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ | |
1316 | completed = ACCESS_ONCE(rsp->completed); | |
1317 | if (ULONG_CMP_GE(completed, gpnum) || | |
1318 | ULONG_CMP_LT(j, js) || | |
1319 | ULONG_CMP_GE(gps, js)) | |
1320 | return; /* No stall or GP completed since entering function. */ | |
64db4cff | 1321 | rnp = rdp->mynode; |
c96ea7cf | 1322 | if (rcu_gp_in_progress(rsp) && |
26cdfedf | 1323 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) { |
64db4cff PM |
1324 | |
1325 | /* We haven't checked in, so go dump stack. */ | |
1326 | print_cpu_stall(rsp); | |
1327 | ||
bad6e139 PM |
1328 | } else if (rcu_gp_in_progress(rsp) && |
1329 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 1330 | |
bad6e139 | 1331 | /* They had a few time units to dump stack, so complain. */ |
6ccd2ecd | 1332 | print_other_cpu_stall(rsp, gpnum); |
64db4cff PM |
1333 | } |
1334 | } | |
1335 | ||
53d84e00 PM |
1336 | /** |
1337 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1338 | * | |
1339 | * Set the stall-warning timeout way off into the future, thus preventing | |
1340 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1341 | * RCU grace periods. | |
1342 | * | |
1343 | * The caller must disable hard irqs. | |
1344 | */ | |
1345 | void rcu_cpu_stall_reset(void) | |
1346 | { | |
6ce75a23 PM |
1347 | struct rcu_state *rsp; |
1348 | ||
1349 | for_each_rcu_flavor(rsp) | |
4fc5b755 | 1350 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + ULONG_MAX / 2; |
53d84e00 PM |
1351 | } |
1352 | ||
3f5d3ea6 | 1353 | /* |
d3f3f3f2 PM |
1354 | * Initialize the specified rcu_data structure's default callback list |
1355 | * to empty. The default callback list is the one that is not used by | |
1356 | * no-callbacks CPUs. | |
3f5d3ea6 | 1357 | */ |
d3f3f3f2 | 1358 | static void init_default_callback_list(struct rcu_data *rdp) |
3f5d3ea6 PM |
1359 | { |
1360 | int i; | |
1361 | ||
1362 | rdp->nxtlist = NULL; | |
1363 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1364 | rdp->nxttail[i] = &rdp->nxtlist; | |
1365 | } | |
1366 | ||
d3f3f3f2 PM |
1367 | /* |
1368 | * Initialize the specified rcu_data structure's callback list to empty. | |
1369 | */ | |
1370 | static void init_callback_list(struct rcu_data *rdp) | |
1371 | { | |
1372 | if (init_nocb_callback_list(rdp)) | |
1373 | return; | |
1374 | init_default_callback_list(rdp); | |
1375 | } | |
1376 | ||
dc35c893 PM |
1377 | /* |
1378 | * Determine the value that ->completed will have at the end of the | |
1379 | * next subsequent grace period. This is used to tag callbacks so that | |
1380 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1381 | * been dyntick-idle for an extended period with callbacks under the | |
1382 | * influence of RCU_FAST_NO_HZ. | |
1383 | * | |
1384 | * The caller must hold rnp->lock with interrupts disabled. | |
1385 | */ | |
1386 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1387 | struct rcu_node *rnp) | |
1388 | { | |
1389 | /* | |
1390 | * If RCU is idle, we just wait for the next grace period. | |
1391 | * But we can only be sure that RCU is idle if we are looking | |
1392 | * at the root rcu_node structure -- otherwise, a new grace | |
1393 | * period might have started, but just not yet gotten around | |
1394 | * to initializing the current non-root rcu_node structure. | |
1395 | */ | |
1396 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1397 | return rnp->completed + 1; | |
1398 | ||
1399 | /* | |
1400 | * Otherwise, wait for a possible partial grace period and | |
1401 | * then the subsequent full grace period. | |
1402 | */ | |
1403 | return rnp->completed + 2; | |
1404 | } | |
1405 | ||
0446be48 PM |
1406 | /* |
1407 | * Trace-event helper function for rcu_start_future_gp() and | |
1408 | * rcu_nocb_wait_gp(). | |
1409 | */ | |
1410 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1411 | unsigned long c, const char *s) |
0446be48 PM |
1412 | { |
1413 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1414 | rnp->completed, c, rnp->level, | |
1415 | rnp->grplo, rnp->grphi, s); | |
1416 | } | |
1417 | ||
1418 | /* | |
1419 | * Start some future grace period, as needed to handle newly arrived | |
1420 | * callbacks. The required future grace periods are recorded in each | |
48a7639c PM |
1421 | * rcu_node structure's ->need_future_gp field. Returns true if there |
1422 | * is reason to awaken the grace-period kthread. | |
0446be48 PM |
1423 | * |
1424 | * The caller must hold the specified rcu_node structure's ->lock. | |
1425 | */ | |
48a7639c PM |
1426 | static bool __maybe_unused |
1427 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
1428 | unsigned long *c_out) | |
0446be48 PM |
1429 | { |
1430 | unsigned long c; | |
1431 | int i; | |
48a7639c | 1432 | bool ret = false; |
0446be48 PM |
1433 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); |
1434 | ||
1435 | /* | |
1436 | * Pick up grace-period number for new callbacks. If this | |
1437 | * grace period is already marked as needed, return to the caller. | |
1438 | */ | |
1439 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1440 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1441 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1442 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
48a7639c | 1443 | goto out; |
0446be48 PM |
1444 | } |
1445 | ||
1446 | /* | |
1447 | * If either this rcu_node structure or the root rcu_node structure | |
1448 | * believe that a grace period is in progress, then we must wait | |
1449 | * for the one following, which is in "c". Because our request | |
1450 | * will be noticed at the end of the current grace period, we don't | |
48bd8e9b PK |
1451 | * need to explicitly start one. We only do the lockless check |
1452 | * of rnp_root's fields if the current rcu_node structure thinks | |
1453 | * there is no grace period in flight, and because we hold rnp->lock, | |
1454 | * the only possible change is when rnp_root's two fields are | |
1455 | * equal, in which case rnp_root->gpnum might be concurrently | |
1456 | * incremented. But that is OK, as it will just result in our | |
1457 | * doing some extra useless work. | |
0446be48 PM |
1458 | */ |
1459 | if (rnp->gpnum != rnp->completed || | |
48bd8e9b | 1460 | ACCESS_ONCE(rnp_root->gpnum) != ACCESS_ONCE(rnp_root->completed)) { |
0446be48 | 1461 | rnp->need_future_gp[c & 0x1]++; |
f7f7bac9 | 1462 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
48a7639c | 1463 | goto out; |
0446be48 PM |
1464 | } |
1465 | ||
1466 | /* | |
1467 | * There might be no grace period in progress. If we don't already | |
1468 | * hold it, acquire the root rcu_node structure's lock in order to | |
1469 | * start one (if needed). | |
1470 | */ | |
6303b9c8 | 1471 | if (rnp != rnp_root) { |
0446be48 | 1472 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 PM |
1473 | smp_mb__after_unlock_lock(); |
1474 | } | |
0446be48 PM |
1475 | |
1476 | /* | |
1477 | * Get a new grace-period number. If there really is no grace | |
1478 | * period in progress, it will be smaller than the one we obtained | |
1479 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1480 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1481 | */ | |
1482 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1483 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1484 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1485 | rdp->nxtcompleted[i] = c; | |
1486 | ||
1487 | /* | |
1488 | * If the needed for the required grace period is already | |
1489 | * recorded, trace and leave. | |
1490 | */ | |
1491 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1492 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1493 | goto unlock_out; |
1494 | } | |
1495 | ||
1496 | /* Record the need for the future grace period. */ | |
1497 | rnp_root->need_future_gp[c & 0x1]++; | |
1498 | ||
1499 | /* If a grace period is not already in progress, start one. */ | |
1500 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1501 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1502 | } else { |
f7f7bac9 | 1503 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
48a7639c | 1504 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1505 | } |
1506 | unlock_out: | |
1507 | if (rnp != rnp_root) | |
1508 | raw_spin_unlock(&rnp_root->lock); | |
48a7639c PM |
1509 | out: |
1510 | if (c_out != NULL) | |
1511 | *c_out = c; | |
1512 | return ret; | |
0446be48 PM |
1513 | } |
1514 | ||
1515 | /* | |
1516 | * Clean up any old requests for the just-ended grace period. Also return | |
1517 | * whether any additional grace periods have been requested. Also invoke | |
1518 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1519 | * waiting for this grace period to complete. | |
1520 | */ | |
1521 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1522 | { | |
1523 | int c = rnp->completed; | |
1524 | int needmore; | |
1525 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1526 | ||
1527 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1528 | rnp->need_future_gp[c & 0x1] = 0; | |
1529 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1530 | trace_rcu_future_gp(rnp, rdp, c, |
1531 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1532 | return needmore; |
1533 | } | |
1534 | ||
48a7639c PM |
1535 | /* |
1536 | * Awaken the grace-period kthread for the specified flavor of RCU. | |
1537 | * Don't do a self-awaken, and don't bother awakening when there is | |
1538 | * nothing for the grace-period kthread to do (as in several CPUs | |
1539 | * raced to awaken, and we lost), and finally don't try to awaken | |
1540 | * a kthread that has not yet been created. | |
1541 | */ | |
1542 | static void rcu_gp_kthread_wake(struct rcu_state *rsp) | |
1543 | { | |
1544 | if (current == rsp->gp_kthread || | |
1545 | !ACCESS_ONCE(rsp->gp_flags) || | |
1546 | !rsp->gp_kthread) | |
1547 | return; | |
1548 | wake_up(&rsp->gp_wq); | |
1549 | } | |
1550 | ||
dc35c893 PM |
1551 | /* |
1552 | * If there is room, assign a ->completed number to any callbacks on | |
1553 | * this CPU that have not already been assigned. Also accelerate any | |
1554 | * callbacks that were previously assigned a ->completed number that has | |
1555 | * since proven to be too conservative, which can happen if callbacks get | |
1556 | * assigned a ->completed number while RCU is idle, but with reference to | |
1557 | * a non-root rcu_node structure. This function is idempotent, so it does | |
48a7639c PM |
1558 | * not hurt to call it repeatedly. Returns an flag saying that we should |
1559 | * awaken the RCU grace-period kthread. | |
dc35c893 PM |
1560 | * |
1561 | * The caller must hold rnp->lock with interrupts disabled. | |
1562 | */ | |
48a7639c | 1563 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1564 | struct rcu_data *rdp) |
1565 | { | |
1566 | unsigned long c; | |
1567 | int i; | |
48a7639c | 1568 | bool ret; |
dc35c893 PM |
1569 | |
1570 | /* If the CPU has no callbacks, nothing to do. */ | |
1571 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1572 | return false; |
dc35c893 PM |
1573 | |
1574 | /* | |
1575 | * Starting from the sublist containing the callbacks most | |
1576 | * recently assigned a ->completed number and working down, find the | |
1577 | * first sublist that is not assignable to an upcoming grace period. | |
1578 | * Such a sublist has something in it (first two tests) and has | |
1579 | * a ->completed number assigned that will complete sooner than | |
1580 | * the ->completed number for newly arrived callbacks (last test). | |
1581 | * | |
1582 | * The key point is that any later sublist can be assigned the | |
1583 | * same ->completed number as the newly arrived callbacks, which | |
1584 | * means that the callbacks in any of these later sublist can be | |
1585 | * grouped into a single sublist, whether or not they have already | |
1586 | * been assigned a ->completed number. | |
1587 | */ | |
1588 | c = rcu_cbs_completed(rsp, rnp); | |
1589 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1590 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1591 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1592 | break; | |
1593 | ||
1594 | /* | |
1595 | * If there are no sublist for unassigned callbacks, leave. | |
1596 | * At the same time, advance "i" one sublist, so that "i" will | |
1597 | * index into the sublist where all the remaining callbacks should | |
1598 | * be grouped into. | |
1599 | */ | |
1600 | if (++i >= RCU_NEXT_TAIL) | |
48a7639c | 1601 | return false; |
dc35c893 PM |
1602 | |
1603 | /* | |
1604 | * Assign all subsequent callbacks' ->completed number to the next | |
1605 | * full grace period and group them all in the sublist initially | |
1606 | * indexed by "i". | |
1607 | */ | |
1608 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1609 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1610 | rdp->nxtcompleted[i] = c; | |
1611 | } | |
910ee45d | 1612 | /* Record any needed additional grace periods. */ |
48a7639c | 1613 | ret = rcu_start_future_gp(rnp, rdp, NULL); |
6d4b418c PM |
1614 | |
1615 | /* Trace depending on how much we were able to accelerate. */ | |
1616 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1617 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1618 | else |
f7f7bac9 | 1619 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
48a7639c | 1620 | return ret; |
dc35c893 PM |
1621 | } |
1622 | ||
1623 | /* | |
1624 | * Move any callbacks whose grace period has completed to the | |
1625 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1626 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1627 | * sublist. This function is idempotent, so it does not hurt to | |
1628 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1629 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1630 | * |
1631 | * The caller must hold rnp->lock with interrupts disabled. | |
1632 | */ | |
48a7639c | 1633 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1634 | struct rcu_data *rdp) |
1635 | { | |
1636 | int i, j; | |
1637 | ||
1638 | /* If the CPU has no callbacks, nothing to do. */ | |
1639 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1640 | return false; |
dc35c893 PM |
1641 | |
1642 | /* | |
1643 | * Find all callbacks whose ->completed numbers indicate that they | |
1644 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1645 | */ | |
1646 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1647 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1648 | break; | |
1649 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1650 | } | |
1651 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1652 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1653 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1654 | ||
1655 | /* Copy down callbacks to fill in empty sublists. */ | |
1656 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1657 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1658 | break; | |
1659 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1660 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1661 | } | |
1662 | ||
1663 | /* Classify any remaining callbacks. */ | |
48a7639c | 1664 | return rcu_accelerate_cbs(rsp, rnp, rdp); |
dc35c893 PM |
1665 | } |
1666 | ||
d09b62df | 1667 | /* |
ba9fbe95 PM |
1668 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1669 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1670 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1671 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1672 | */ |
48a7639c PM |
1673 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, |
1674 | struct rcu_data *rdp) | |
d09b62df | 1675 | { |
48a7639c PM |
1676 | bool ret; |
1677 | ||
ba9fbe95 | 1678 | /* Handle the ends of any preceding grace periods first. */ |
e3663b10 PM |
1679 | if (rdp->completed == rnp->completed && |
1680 | !unlikely(ACCESS_ONCE(rdp->gpwrap))) { | |
d09b62df | 1681 | |
ba9fbe95 | 1682 | /* No grace period end, so just accelerate recent callbacks. */ |
48a7639c | 1683 | ret = rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1684 | |
dc35c893 PM |
1685 | } else { |
1686 | ||
1687 | /* Advance callbacks. */ | |
48a7639c | 1688 | ret = rcu_advance_cbs(rsp, rnp, rdp); |
d09b62df PM |
1689 | |
1690 | /* Remember that we saw this grace-period completion. */ | |
1691 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1692 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1693 | } |
398ebe60 | 1694 | |
e3663b10 | 1695 | if (rdp->gpnum != rnp->gpnum || unlikely(ACCESS_ONCE(rdp->gpwrap))) { |
6eaef633 PM |
1696 | /* |
1697 | * If the current grace period is waiting for this CPU, | |
1698 | * set up to detect a quiescent state, otherwise don't | |
1699 | * go looking for one. | |
1700 | */ | |
1701 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1702 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
6eaef633 | 1703 | rdp->passed_quiesce = 0; |
5cd37193 | 1704 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); |
6eaef633 PM |
1705 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); |
1706 | zero_cpu_stall_ticks(rdp); | |
e3663b10 | 1707 | ACCESS_ONCE(rdp->gpwrap) = false; |
6eaef633 | 1708 | } |
48a7639c | 1709 | return ret; |
6eaef633 PM |
1710 | } |
1711 | ||
d34ea322 | 1712 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1713 | { |
1714 | unsigned long flags; | |
48a7639c | 1715 | bool needwake; |
6eaef633 PM |
1716 | struct rcu_node *rnp; |
1717 | ||
1718 | local_irq_save(flags); | |
1719 | rnp = rdp->mynode; | |
d34ea322 | 1720 | if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) && |
e3663b10 PM |
1721 | rdp->completed == ACCESS_ONCE(rnp->completed) && |
1722 | !unlikely(ACCESS_ONCE(rdp->gpwrap))) || /* w/out lock. */ | |
6eaef633 PM |
1723 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
1724 | local_irq_restore(flags); | |
1725 | return; | |
1726 | } | |
6303b9c8 | 1727 | smp_mb__after_unlock_lock(); |
48a7639c | 1728 | needwake = __note_gp_changes(rsp, rnp, rdp); |
6eaef633 | 1729 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
48a7639c PM |
1730 | if (needwake) |
1731 | rcu_gp_kthread_wake(rsp); | |
6eaef633 PM |
1732 | } |
1733 | ||
b3dbec76 | 1734 | /* |
f7be8209 | 1735 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1736 | */ |
7fdefc10 | 1737 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 | 1738 | { |
0aa04b05 | 1739 | unsigned long oldmask; |
b3dbec76 | 1740 | struct rcu_data *rdp; |
7fdefc10 | 1741 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1742 | |
6ccd2ecd | 1743 | ACCESS_ONCE(rsp->gp_activity) = jiffies; |
7fdefc10 | 1744 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1745 | smp_mb__after_unlock_lock(); |
91dc9542 | 1746 | if (!ACCESS_ONCE(rsp->gp_flags)) { |
f7be8209 PM |
1747 | /* Spurious wakeup, tell caller to go back to sleep. */ |
1748 | raw_spin_unlock_irq(&rnp->lock); | |
1749 | return 0; | |
1750 | } | |
91dc9542 | 1751 | ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */ |
b3dbec76 | 1752 | |
f7be8209 PM |
1753 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1754 | /* | |
1755 | * Grace period already in progress, don't start another. | |
1756 | * Not supposed to be able to happen. | |
1757 | */ | |
7fdefc10 PM |
1758 | raw_spin_unlock_irq(&rnp->lock); |
1759 | return 0; | |
1760 | } | |
1761 | ||
7fdefc10 | 1762 | /* Advance to a new grace period and initialize state. */ |
26cdfedf | 1763 | record_gp_stall_check_time(rsp); |
765a3f4f PM |
1764 | /* Record GP times before starting GP, hence smp_store_release(). */ |
1765 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); | |
f7f7bac9 | 1766 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1767 | raw_spin_unlock_irq(&rnp->lock); |
1768 | ||
0aa04b05 PM |
1769 | /* |
1770 | * Apply per-leaf buffered online and offline operations to the | |
1771 | * rcu_node tree. Note that this new grace period need not wait | |
1772 | * for subsequent online CPUs, and that quiescent-state forcing | |
1773 | * will handle subsequent offline CPUs. | |
1774 | */ | |
1775 | rcu_for_each_leaf_node(rsp, rnp) { | |
1776 | raw_spin_lock_irq(&rnp->lock); | |
1777 | smp_mb__after_unlock_lock(); | |
1778 | if (rnp->qsmaskinit == rnp->qsmaskinitnext && | |
1779 | !rnp->wait_blkd_tasks) { | |
1780 | /* Nothing to do on this leaf rcu_node structure. */ | |
1781 | raw_spin_unlock_irq(&rnp->lock); | |
1782 | continue; | |
1783 | } | |
1784 | ||
1785 | /* Record old state, apply changes to ->qsmaskinit field. */ | |
1786 | oldmask = rnp->qsmaskinit; | |
1787 | rnp->qsmaskinit = rnp->qsmaskinitnext; | |
1788 | ||
1789 | /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ | |
1790 | if (!oldmask != !rnp->qsmaskinit) { | |
1791 | if (!oldmask) /* First online CPU for this rcu_node. */ | |
1792 | rcu_init_new_rnp(rnp); | |
1793 | else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */ | |
1794 | rnp->wait_blkd_tasks = true; | |
1795 | else /* Last offline CPU and can propagate. */ | |
1796 | rcu_cleanup_dead_rnp(rnp); | |
1797 | } | |
1798 | ||
1799 | /* | |
1800 | * If all waited-on tasks from prior grace period are | |
1801 | * done, and if all this rcu_node structure's CPUs are | |
1802 | * still offline, propagate up the rcu_node tree and | |
1803 | * clear ->wait_blkd_tasks. Otherwise, if one of this | |
1804 | * rcu_node structure's CPUs has since come back online, | |
1805 | * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp() | |
1806 | * checks for this, so just call it unconditionally). | |
1807 | */ | |
1808 | if (rnp->wait_blkd_tasks && | |
1809 | (!rcu_preempt_has_tasks(rnp) || | |
1810 | rnp->qsmaskinit)) { | |
1811 | rnp->wait_blkd_tasks = false; | |
1812 | rcu_cleanup_dead_rnp(rnp); | |
1813 | } | |
1814 | ||
1815 | raw_spin_unlock_irq(&rnp->lock); | |
1816 | } | |
7fdefc10 PM |
1817 | |
1818 | /* | |
1819 | * Set the quiescent-state-needed bits in all the rcu_node | |
1820 | * structures for all currently online CPUs in breadth-first order, | |
1821 | * starting from the root rcu_node structure, relying on the layout | |
1822 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1823 | * will access only the leaves of the hierarchy, thus seeing that no | |
1824 | * grace period is in progress, at least until the corresponding | |
1825 | * leaf node has been initialized. In addition, we have excluded | |
1826 | * CPU-hotplug operations. | |
1827 | * | |
1828 | * The grace period cannot complete until the initialization | |
1829 | * process finishes, because this kthread handles both. | |
1830 | */ | |
1831 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
b3dbec76 | 1832 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1833 | smp_mb__after_unlock_lock(); |
b3dbec76 | 1834 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1835 | rcu_preempt_check_blocked_tasks(rnp); |
1836 | rnp->qsmask = rnp->qsmaskinit; | |
0446be48 | 1837 | ACCESS_ONCE(rnp->gpnum) = rsp->gpnum; |
3f47da0f LJ |
1838 | if (WARN_ON_ONCE(rnp->completed != rsp->completed)) |
1839 | ACCESS_ONCE(rnp->completed) = rsp->completed; | |
7fdefc10 | 1840 | if (rnp == rdp->mynode) |
48a7639c | 1841 | (void)__note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1842 | rcu_preempt_boost_start_gp(rnp); |
1843 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1844 | rnp->level, rnp->grplo, | |
1845 | rnp->grphi, rnp->qsmask); | |
1846 | raw_spin_unlock_irq(&rnp->lock); | |
bde6c3aa | 1847 | cond_resched_rcu_qs(); |
6ccd2ecd | 1848 | ACCESS_ONCE(rsp->gp_activity) = jiffies; |
8d7dc928 PM |
1849 | if (gp_init_delay > 0 && |
1850 | !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD))) | |
37745d28 | 1851 | schedule_timeout_uninterruptible(gp_init_delay); |
7fdefc10 | 1852 | } |
b3dbec76 | 1853 | |
7fdefc10 PM |
1854 | return 1; |
1855 | } | |
b3dbec76 | 1856 | |
4cdfc175 PM |
1857 | /* |
1858 | * Do one round of quiescent-state forcing. | |
1859 | */ | |
01896f7e | 1860 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) |
4cdfc175 PM |
1861 | { |
1862 | int fqs_state = fqs_state_in; | |
217af2a2 PM |
1863 | bool isidle = false; |
1864 | unsigned long maxj; | |
4cdfc175 PM |
1865 | struct rcu_node *rnp = rcu_get_root(rsp); |
1866 | ||
6ccd2ecd | 1867 | ACCESS_ONCE(rsp->gp_activity) = jiffies; |
4cdfc175 PM |
1868 | rsp->n_force_qs++; |
1869 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1870 | /* Collect dyntick-idle snapshots. */ | |
0edd1b17 | 1871 | if (is_sysidle_rcu_state(rsp)) { |
e02b2edf | 1872 | isidle = true; |
0edd1b17 PM |
1873 | maxj = jiffies - ULONG_MAX / 4; |
1874 | } | |
217af2a2 PM |
1875 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1876 | &isidle, &maxj); | |
0edd1b17 | 1877 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1878 | fqs_state = RCU_FORCE_QS; |
1879 | } else { | |
1880 | /* Handle dyntick-idle and offline CPUs. */ | |
675da67f | 1881 | isidle = true; |
217af2a2 | 1882 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1883 | } |
1884 | /* Clear flag to prevent immediate re-entry. */ | |
1885 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1886 | raw_spin_lock_irq(&rnp->lock); | |
6303b9c8 | 1887 | smp_mb__after_unlock_lock(); |
4de376a1 PK |
1888 | ACCESS_ONCE(rsp->gp_flags) = |
1889 | ACCESS_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS; | |
4cdfc175 PM |
1890 | raw_spin_unlock_irq(&rnp->lock); |
1891 | } | |
1892 | return fqs_state; | |
1893 | } | |
1894 | ||
7fdefc10 PM |
1895 | /* |
1896 | * Clean up after the old grace period. | |
1897 | */ | |
4cdfc175 | 1898 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1899 | { |
1900 | unsigned long gp_duration; | |
48a7639c | 1901 | bool needgp = false; |
dae6e64d | 1902 | int nocb = 0; |
7fdefc10 PM |
1903 | struct rcu_data *rdp; |
1904 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1905 | |
6ccd2ecd | 1906 | ACCESS_ONCE(rsp->gp_activity) = jiffies; |
7fdefc10 | 1907 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1908 | smp_mb__after_unlock_lock(); |
7fdefc10 PM |
1909 | gp_duration = jiffies - rsp->gp_start; |
1910 | if (gp_duration > rsp->gp_max) | |
1911 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1912 | |
7fdefc10 PM |
1913 | /* |
1914 | * We know the grace period is complete, but to everyone else | |
1915 | * it appears to still be ongoing. But it is also the case | |
1916 | * that to everyone else it looks like there is nothing that | |
1917 | * they can do to advance the grace period. It is therefore | |
1918 | * safe for us to drop the lock in order to mark the grace | |
1919 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1920 | */ |
5d4b8659 | 1921 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1922 | |
5d4b8659 PM |
1923 | /* |
1924 | * Propagate new ->completed value to rcu_node structures so | |
1925 | * that other CPUs don't have to wait until the start of the next | |
1926 | * grace period to process their callbacks. This also avoids | |
1927 | * some nasty RCU grace-period initialization races by forcing | |
1928 | * the end of the current grace period to be completely recorded in | |
1929 | * all of the rcu_node structures before the beginning of the next | |
1930 | * grace period is recorded in any of the rcu_node structures. | |
1931 | */ | |
1932 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1933 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1934 | smp_mb__after_unlock_lock(); |
5c60d25f PM |
1935 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
1936 | WARN_ON_ONCE(rnp->qsmask); | |
0446be48 | 1937 | ACCESS_ONCE(rnp->completed) = rsp->gpnum; |
b11cc576 PM |
1938 | rdp = this_cpu_ptr(rsp->rda); |
1939 | if (rnp == rdp->mynode) | |
48a7639c | 1940 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; |
78e4bc34 | 1941 | /* smp_mb() provided by prior unlock-lock pair. */ |
0446be48 | 1942 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 | 1943 | raw_spin_unlock_irq(&rnp->lock); |
bde6c3aa | 1944 | cond_resched_rcu_qs(); |
6ccd2ecd | 1945 | ACCESS_ONCE(rsp->gp_activity) = jiffies; |
7fdefc10 | 1946 | } |
5d4b8659 PM |
1947 | rnp = rcu_get_root(rsp); |
1948 | raw_spin_lock_irq(&rnp->lock); | |
765a3f4f | 1949 | smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */ |
dae6e64d | 1950 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 | 1951 | |
765a3f4f PM |
1952 | /* Declare grace period done. */ |
1953 | ACCESS_ONCE(rsp->completed) = rsp->gpnum; | |
f7f7bac9 | 1954 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
7fdefc10 | 1955 | rsp->fqs_state = RCU_GP_IDLE; |
5d4b8659 | 1956 | rdp = this_cpu_ptr(rsp->rda); |
48a7639c PM |
1957 | /* Advance CBs to reduce false positives below. */ |
1958 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; | |
1959 | if (needgp || cpu_needs_another_gp(rsp, rdp)) { | |
91dc9542 | 1960 | ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1961 | trace_rcu_grace_period(rsp->name, |
1962 | ACCESS_ONCE(rsp->gpnum), | |
1963 | TPS("newreq")); | |
1964 | } | |
7fdefc10 | 1965 | raw_spin_unlock_irq(&rnp->lock); |
7fdefc10 PM |
1966 | } |
1967 | ||
1968 | /* | |
1969 | * Body of kthread that handles grace periods. | |
1970 | */ | |
1971 | static int __noreturn rcu_gp_kthread(void *arg) | |
1972 | { | |
4cdfc175 | 1973 | int fqs_state; |
88d6df61 | 1974 | int gf; |
d40011f6 | 1975 | unsigned long j; |
4cdfc175 | 1976 | int ret; |
7fdefc10 PM |
1977 | struct rcu_state *rsp = arg; |
1978 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1979 | ||
5871968d | 1980 | rcu_bind_gp_kthread(); |
7fdefc10 PM |
1981 | for (;;) { |
1982 | ||
1983 | /* Handle grace-period start. */ | |
1984 | for (;;) { | |
63c4db78 PM |
1985 | trace_rcu_grace_period(rsp->name, |
1986 | ACCESS_ONCE(rsp->gpnum), | |
1987 | TPS("reqwait")); | |
afea227f | 1988 | rsp->gp_state = RCU_GP_WAIT_GPS; |
4cdfc175 | 1989 | wait_event_interruptible(rsp->gp_wq, |
591c6d17 | 1990 | ACCESS_ONCE(rsp->gp_flags) & |
4cdfc175 | 1991 | RCU_GP_FLAG_INIT); |
78e4bc34 | 1992 | /* Locking provides needed memory barrier. */ |
f7be8209 | 1993 | if (rcu_gp_init(rsp)) |
7fdefc10 | 1994 | break; |
bde6c3aa | 1995 | cond_resched_rcu_qs(); |
6ccd2ecd | 1996 | ACCESS_ONCE(rsp->gp_activity) = jiffies; |
73a860cd | 1997 | WARN_ON(signal_pending(current)); |
63c4db78 PM |
1998 | trace_rcu_grace_period(rsp->name, |
1999 | ACCESS_ONCE(rsp->gpnum), | |
2000 | TPS("reqwaitsig")); | |
7fdefc10 | 2001 | } |
cabc49c1 | 2002 | |
4cdfc175 PM |
2003 | /* Handle quiescent-state forcing. */ |
2004 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
2005 | j = jiffies_till_first_fqs; |
2006 | if (j > HZ) { | |
2007 | j = HZ; | |
2008 | jiffies_till_first_fqs = HZ; | |
2009 | } | |
88d6df61 | 2010 | ret = 0; |
cabc49c1 | 2011 | for (;;) { |
88d6df61 PM |
2012 | if (!ret) |
2013 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 PM |
2014 | trace_rcu_grace_period(rsp->name, |
2015 | ACCESS_ONCE(rsp->gpnum), | |
2016 | TPS("fqswait")); | |
afea227f | 2017 | rsp->gp_state = RCU_GP_WAIT_FQS; |
4cdfc175 | 2018 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
88d6df61 PM |
2019 | ((gf = ACCESS_ONCE(rsp->gp_flags)) & |
2020 | RCU_GP_FLAG_FQS) || | |
4cdfc175 PM |
2021 | (!ACCESS_ONCE(rnp->qsmask) && |
2022 | !rcu_preempt_blocked_readers_cgp(rnp)), | |
d40011f6 | 2023 | j); |
78e4bc34 | 2024 | /* Locking provides needed memory barriers. */ |
4cdfc175 | 2025 | /* If grace period done, leave loop. */ |
cabc49c1 | 2026 | if (!ACCESS_ONCE(rnp->qsmask) && |
4cdfc175 | 2027 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 2028 | break; |
4cdfc175 | 2029 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
2030 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
2031 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 PM |
2032 | trace_rcu_grace_period(rsp->name, |
2033 | ACCESS_ONCE(rsp->gpnum), | |
2034 | TPS("fqsstart")); | |
4cdfc175 | 2035 | fqs_state = rcu_gp_fqs(rsp, fqs_state); |
63c4db78 PM |
2036 | trace_rcu_grace_period(rsp->name, |
2037 | ACCESS_ONCE(rsp->gpnum), | |
2038 | TPS("fqsend")); | |
bde6c3aa | 2039 | cond_resched_rcu_qs(); |
6ccd2ecd | 2040 | ACCESS_ONCE(rsp->gp_activity) = jiffies; |
4cdfc175 PM |
2041 | } else { |
2042 | /* Deal with stray signal. */ | |
bde6c3aa | 2043 | cond_resched_rcu_qs(); |
6ccd2ecd | 2044 | ACCESS_ONCE(rsp->gp_activity) = jiffies; |
73a860cd | 2045 | WARN_ON(signal_pending(current)); |
63c4db78 PM |
2046 | trace_rcu_grace_period(rsp->name, |
2047 | ACCESS_ONCE(rsp->gpnum), | |
2048 | TPS("fqswaitsig")); | |
4cdfc175 | 2049 | } |
d40011f6 PM |
2050 | j = jiffies_till_next_fqs; |
2051 | if (j > HZ) { | |
2052 | j = HZ; | |
2053 | jiffies_till_next_fqs = HZ; | |
2054 | } else if (j < 1) { | |
2055 | j = 1; | |
2056 | jiffies_till_next_fqs = 1; | |
2057 | } | |
cabc49c1 | 2058 | } |
4cdfc175 PM |
2059 | |
2060 | /* Handle grace-period end. */ | |
2061 | rcu_gp_cleanup(rsp); | |
b3dbec76 | 2062 | } |
b3dbec76 PM |
2063 | } |
2064 | ||
64db4cff PM |
2065 | /* |
2066 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
2067 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 2068 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
2069 | * |
2070 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
2071 | * invoke this function. This can happen when the dying CPU reports its | |
2072 | * quiescent state. | |
48a7639c PM |
2073 | * |
2074 | * Returns true if the grace-period kthread must be awakened. | |
64db4cff | 2075 | */ |
48a7639c | 2076 | static bool |
910ee45d PM |
2077 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
2078 | struct rcu_data *rdp) | |
64db4cff | 2079 | { |
b8462084 | 2080 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 2081 | /* |
b3dbec76 | 2082 | * Either we have not yet spawned the grace-period |
62da1921 PM |
2083 | * task, this CPU does not need another grace period, |
2084 | * or a grace period is already in progress. | |
b3dbec76 | 2085 | * Either way, don't start a new grace period. |
afe24b12 | 2086 | */ |
48a7639c | 2087 | return false; |
afe24b12 | 2088 | } |
91dc9542 | 2089 | ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT; |
bb311ecc PM |
2090 | trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum), |
2091 | TPS("newreq")); | |
62da1921 | 2092 | |
016a8d5b SR |
2093 | /* |
2094 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c | 2095 | * could cause possible deadlocks with the rq->lock. Defer |
48a7639c | 2096 | * the wakeup to our caller. |
016a8d5b | 2097 | */ |
48a7639c | 2098 | return true; |
64db4cff PM |
2099 | } |
2100 | ||
910ee45d PM |
2101 | /* |
2102 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
2103 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
2104 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
2105 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
2106 | * that is encountered beforehand. | |
48a7639c PM |
2107 | * |
2108 | * Returns true if the grace-period kthread needs to be awakened. | |
910ee45d | 2109 | */ |
48a7639c | 2110 | static bool rcu_start_gp(struct rcu_state *rsp) |
910ee45d PM |
2111 | { |
2112 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
2113 | struct rcu_node *rnp = rcu_get_root(rsp); | |
48a7639c | 2114 | bool ret = false; |
910ee45d PM |
2115 | |
2116 | /* | |
2117 | * If there is no grace period in progress right now, any | |
2118 | * callbacks we have up to this point will be satisfied by the | |
2119 | * next grace period. Also, advancing the callbacks reduces the | |
2120 | * probability of false positives from cpu_needs_another_gp() | |
2121 | * resulting in pointless grace periods. So, advance callbacks | |
2122 | * then start the grace period! | |
2123 | */ | |
48a7639c PM |
2124 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; |
2125 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; | |
2126 | return ret; | |
910ee45d PM |
2127 | } |
2128 | ||
f41d911f | 2129 | /* |
d3f6bad3 PM |
2130 | * Report a full set of quiescent states to the specified rcu_state |
2131 | * data structure. This involves cleaning up after the prior grace | |
2132 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
2133 | * if one is needed. Note that the caller must hold rnp->lock, which |
2134 | * is released before return. | |
f41d911f | 2135 | */ |
d3f6bad3 | 2136 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 2137 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 2138 | { |
fc2219d4 | 2139 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cabc49c1 | 2140 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
2aa792e6 | 2141 | rcu_gp_kthread_wake(rsp); |
f41d911f PM |
2142 | } |
2143 | ||
64db4cff | 2144 | /* |
d3f6bad3 PM |
2145 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
2146 | * Allows quiescent states for a group of CPUs to be reported at one go | |
2147 | * to the specified rcu_node structure, though all the CPUs in the group | |
654e9533 PM |
2148 | * must be represented by the same rcu_node structure (which need not be a |
2149 | * leaf rcu_node structure, though it often will be). The gps parameter | |
2150 | * is the grace-period snapshot, which means that the quiescent states | |
2151 | * are valid only if rnp->gpnum is equal to gps. That structure's lock | |
2152 | * must be held upon entry, and it is released before return. | |
64db4cff PM |
2153 | */ |
2154 | static void | |
d3f6bad3 | 2155 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
654e9533 | 2156 | struct rcu_node *rnp, unsigned long gps, unsigned long flags) |
64db4cff PM |
2157 | __releases(rnp->lock) |
2158 | { | |
654e9533 | 2159 | unsigned long oldmask = 0; |
28ecd580 PM |
2160 | struct rcu_node *rnp_c; |
2161 | ||
64db4cff PM |
2162 | /* Walk up the rcu_node hierarchy. */ |
2163 | for (;;) { | |
654e9533 | 2164 | if (!(rnp->qsmask & mask) || rnp->gpnum != gps) { |
64db4cff | 2165 | |
654e9533 PM |
2166 | /* |
2167 | * Our bit has already been cleared, or the | |
2168 | * relevant grace period is already over, so done. | |
2169 | */ | |
1304afb2 | 2170 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2171 | return; |
2172 | } | |
654e9533 | 2173 | WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ |
64db4cff | 2174 | rnp->qsmask &= ~mask; |
d4c08f2a PM |
2175 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
2176 | mask, rnp->qsmask, rnp->level, | |
2177 | rnp->grplo, rnp->grphi, | |
2178 | !!rnp->gp_tasks); | |
27f4d280 | 2179 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
2180 | |
2181 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 2182 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2183 | return; |
2184 | } | |
2185 | mask = rnp->grpmask; | |
2186 | if (rnp->parent == NULL) { | |
2187 | ||
2188 | /* No more levels. Exit loop holding root lock. */ | |
2189 | ||
2190 | break; | |
2191 | } | |
1304afb2 | 2192 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 2193 | rnp_c = rnp; |
64db4cff | 2194 | rnp = rnp->parent; |
1304afb2 | 2195 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2196 | smp_mb__after_unlock_lock(); |
654e9533 | 2197 | oldmask = rnp_c->qsmask; |
64db4cff PM |
2198 | } |
2199 | ||
2200 | /* | |
2201 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 2202 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 2203 | * to clean up and start the next grace period if one is needed. |
64db4cff | 2204 | */ |
d3f6bad3 | 2205 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
2206 | } |
2207 | ||
cc99a310 PM |
2208 | /* |
2209 | * Record a quiescent state for all tasks that were previously queued | |
2210 | * on the specified rcu_node structure and that were blocking the current | |
2211 | * RCU grace period. The caller must hold the specified rnp->lock with | |
2212 | * irqs disabled, and this lock is released upon return, but irqs remain | |
2213 | * disabled. | |
2214 | */ | |
0aa04b05 | 2215 | static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp, |
cc99a310 PM |
2216 | struct rcu_node *rnp, unsigned long flags) |
2217 | __releases(rnp->lock) | |
2218 | { | |
654e9533 | 2219 | unsigned long gps; |
cc99a310 PM |
2220 | unsigned long mask; |
2221 | struct rcu_node *rnp_p; | |
2222 | ||
a77da14c PM |
2223 | if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p || |
2224 | rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { | |
cc99a310 PM |
2225 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
2226 | return; /* Still need more quiescent states! */ | |
2227 | } | |
2228 | ||
2229 | rnp_p = rnp->parent; | |
2230 | if (rnp_p == NULL) { | |
2231 | /* | |
a77da14c PM |
2232 | * Only one rcu_node structure in the tree, so don't |
2233 | * try to report up to its nonexistent parent! | |
cc99a310 PM |
2234 | */ |
2235 | rcu_report_qs_rsp(rsp, flags); | |
2236 | return; | |
2237 | } | |
2238 | ||
654e9533 PM |
2239 | /* Report up the rest of the hierarchy, tracking current ->gpnum. */ |
2240 | gps = rnp->gpnum; | |
cc99a310 PM |
2241 | mask = rnp->grpmask; |
2242 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2243 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
2244 | smp_mb__after_unlock_lock(); | |
654e9533 | 2245 | rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags); |
cc99a310 PM |
2246 | } |
2247 | ||
64db4cff | 2248 | /* |
d3f6bad3 PM |
2249 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
2250 | * structure. This must be either called from the specified CPU, or | |
2251 | * called when the specified CPU is known to be offline (and when it is | |
2252 | * also known that no other CPU is concurrently trying to help the offline | |
2253 | * CPU). The lastcomp argument is used to make sure we are still in the | |
2254 | * grace period of interest. We don't want to end the current grace period | |
2255 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
2256 | */ |
2257 | static void | |
d7d6a11e | 2258 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2259 | { |
2260 | unsigned long flags; | |
2261 | unsigned long mask; | |
48a7639c | 2262 | bool needwake; |
64db4cff PM |
2263 | struct rcu_node *rnp; |
2264 | ||
2265 | rnp = rdp->mynode; | |
1304afb2 | 2266 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2267 | smp_mb__after_unlock_lock(); |
5cd37193 PM |
2268 | if ((rdp->passed_quiesce == 0 && |
2269 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) || | |
2270 | rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum || | |
2271 | rdp->gpwrap) { | |
64db4cff PM |
2272 | |
2273 | /* | |
e4cc1f22 PM |
2274 | * The grace period in which this quiescent state was |
2275 | * recorded has ended, so don't report it upwards. | |
2276 | * We will instead need a new quiescent state that lies | |
2277 | * within the current grace period. | |
64db4cff | 2278 | */ |
e4cc1f22 | 2279 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
5cd37193 | 2280 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); |
1304afb2 | 2281 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2282 | return; |
2283 | } | |
2284 | mask = rdp->grpmask; | |
2285 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 2286 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2287 | } else { |
2288 | rdp->qs_pending = 0; | |
2289 | ||
2290 | /* | |
2291 | * This GP can't end until cpu checks in, so all of our | |
2292 | * callbacks can be processed during the next GP. | |
2293 | */ | |
48a7639c | 2294 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 2295 | |
654e9533 PM |
2296 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); |
2297 | /* ^^^ Released rnp->lock */ | |
48a7639c PM |
2298 | if (needwake) |
2299 | rcu_gp_kthread_wake(rsp); | |
64db4cff PM |
2300 | } |
2301 | } | |
2302 | ||
2303 | /* | |
2304 | * Check to see if there is a new grace period of which this CPU | |
2305 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2306 | * Otherwise, see if this CPU has just passed through its first | |
2307 | * quiescent state for this grace period, and record that fact if so. | |
2308 | */ | |
2309 | static void | |
2310 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
2311 | { | |
05eb552b PM |
2312 | /* Check for grace-period ends and beginnings. */ |
2313 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
2314 | |
2315 | /* | |
2316 | * Does this CPU still need to do its part for current grace period? | |
2317 | * If no, return and let the other CPUs do their part as well. | |
2318 | */ | |
2319 | if (!rdp->qs_pending) | |
2320 | return; | |
2321 | ||
2322 | /* | |
2323 | * Was there a quiescent state since the beginning of the grace | |
2324 | * period? If no, then exit and wait for the next call. | |
2325 | */ | |
5cd37193 PM |
2326 | if (!rdp->passed_quiesce && |
2327 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) | |
64db4cff PM |
2328 | return; |
2329 | ||
d3f6bad3 PM |
2330 | /* |
2331 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2332 | * judge of that). | |
2333 | */ | |
d7d6a11e | 2334 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
2335 | } |
2336 | ||
2337 | #ifdef CONFIG_HOTPLUG_CPU | |
2338 | ||
e74f4c45 | 2339 | /* |
b1420f1c PM |
2340 | * Send the specified CPU's RCU callbacks to the orphanage. The |
2341 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 2342 | * ->orphan_lock. |
e74f4c45 | 2343 | */ |
b1420f1c PM |
2344 | static void |
2345 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
2346 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 2347 | { |
3fbfbf7a | 2348 | /* No-CBs CPUs do not have orphanable callbacks. */ |
d1e43fa5 | 2349 | if (rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
2350 | return; |
2351 | ||
b1420f1c PM |
2352 | /* |
2353 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
2354 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
2355 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 2356 | */ |
a50c3af9 | 2357 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2358 | rsp->qlen_lazy += rdp->qlen_lazy; |
2359 | rsp->qlen += rdp->qlen; | |
2360 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 2361 | rdp->qlen_lazy = 0; |
1d1fb395 | 2362 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
2363 | } |
2364 | ||
2365 | /* | |
b1420f1c PM |
2366 | * Next, move those callbacks still needing a grace period to |
2367 | * the orphanage, where some other CPU will pick them up. | |
2368 | * Some of the callbacks might have gone partway through a grace | |
2369 | * period, but that is too bad. They get to start over because we | |
2370 | * cannot assume that grace periods are synchronized across CPUs. | |
2371 | * We don't bother updating the ->nxttail[] array yet, instead | |
2372 | * we just reset the whole thing later on. | |
a50c3af9 | 2373 | */ |
b1420f1c PM |
2374 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
2375 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2376 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2377 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
2378 | } |
2379 | ||
2380 | /* | |
b1420f1c PM |
2381 | * Then move the ready-to-invoke callbacks to the orphanage, |
2382 | * where some other CPU will pick them up. These will not be | |
2383 | * required to pass though another grace period: They are done. | |
a50c3af9 | 2384 | */ |
e5601400 | 2385 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2386 | *rsp->orphan_donetail = rdp->nxtlist; |
2387 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 2388 | } |
e74f4c45 | 2389 | |
b33078b6 PM |
2390 | /* |
2391 | * Finally, initialize the rcu_data structure's list to empty and | |
2392 | * disallow further callbacks on this CPU. | |
2393 | */ | |
3f5d3ea6 | 2394 | init_callback_list(rdp); |
b33078b6 | 2395 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
b1420f1c PM |
2396 | } |
2397 | ||
2398 | /* | |
2399 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 2400 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c | 2401 | */ |
96d3fd0d | 2402 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) |
b1420f1c PM |
2403 | { |
2404 | int i; | |
fa07a58f | 2405 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
b1420f1c | 2406 | |
3fbfbf7a | 2407 | /* No-CBs CPUs are handled specially. */ |
96d3fd0d | 2408 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) |
3fbfbf7a PM |
2409 | return; |
2410 | ||
b1420f1c PM |
2411 | /* Do the accounting first. */ |
2412 | rdp->qlen_lazy += rsp->qlen_lazy; | |
2413 | rdp->qlen += rsp->qlen; | |
2414 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
2415 | if (rsp->qlen_lazy != rsp->qlen) |
2416 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2417 | rsp->qlen_lazy = 0; |
2418 | rsp->qlen = 0; | |
2419 | ||
2420 | /* | |
2421 | * We do not need a memory barrier here because the only way we | |
2422 | * can get here if there is an rcu_barrier() in flight is if | |
2423 | * we are the task doing the rcu_barrier(). | |
2424 | */ | |
2425 | ||
2426 | /* First adopt the ready-to-invoke callbacks. */ | |
2427 | if (rsp->orphan_donelist != NULL) { | |
2428 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2429 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
2430 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
2431 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2432 | rdp->nxttail[i] = rsp->orphan_donetail; | |
2433 | rsp->orphan_donelist = NULL; | |
2434 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2435 | } | |
2436 | ||
2437 | /* And then adopt the callbacks that still need a grace period. */ | |
2438 | if (rsp->orphan_nxtlist != NULL) { | |
2439 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
2440 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
2441 | rsp->orphan_nxtlist = NULL; | |
2442 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2443 | } | |
2444 | } | |
2445 | ||
2446 | /* | |
2447 | * Trace the fact that this CPU is going offline. | |
2448 | */ | |
2449 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
2450 | { | |
2451 | RCU_TRACE(unsigned long mask); | |
2452 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
2453 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
2454 | ||
2455 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
2456 | trace_rcu_grace_period(rsp->name, |
2457 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 2458 | TPS("cpuofl")); |
64db4cff PM |
2459 | } |
2460 | ||
8af3a5e7 PM |
2461 | /* |
2462 | * All CPUs for the specified rcu_node structure have gone offline, | |
2463 | * and all tasks that were preempted within an RCU read-side critical | |
2464 | * section while running on one of those CPUs have since exited their RCU | |
2465 | * read-side critical section. Some other CPU is reporting this fact with | |
2466 | * the specified rcu_node structure's ->lock held and interrupts disabled. | |
2467 | * This function therefore goes up the tree of rcu_node structures, | |
2468 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that | |
2469 | * the leaf rcu_node structure's ->qsmaskinit field has already been | |
2470 | * updated | |
2471 | * | |
2472 | * This function does check that the specified rcu_node structure has | |
2473 | * all CPUs offline and no blocked tasks, so it is OK to invoke it | |
2474 | * prematurely. That said, invoking it after the fact will cost you | |
2475 | * a needless lock acquisition. So once it has done its work, don't | |
2476 | * invoke it again. | |
2477 | */ | |
2478 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) | |
2479 | { | |
2480 | long mask; | |
2481 | struct rcu_node *rnp = rnp_leaf; | |
2482 | ||
2483 | if (rnp->qsmaskinit || rcu_preempt_has_tasks(rnp)) | |
2484 | return; | |
2485 | for (;;) { | |
2486 | mask = rnp->grpmask; | |
2487 | rnp = rnp->parent; | |
2488 | if (!rnp) | |
2489 | break; | |
2490 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
2491 | smp_mb__after_unlock_lock(); /* GP memory ordering. */ | |
2492 | rnp->qsmaskinit &= ~mask; | |
0aa04b05 | 2493 | rnp->qsmask &= ~mask; |
8af3a5e7 PM |
2494 | if (rnp->qsmaskinit) { |
2495 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2496 | return; | |
2497 | } | |
2498 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2499 | } | |
2500 | } | |
2501 | ||
88428cc5 PM |
2502 | /* |
2503 | * The CPU is exiting the idle loop into the arch_cpu_idle_dead() | |
2504 | * function. We now remove it from the rcu_node tree's ->qsmaskinit | |
2505 | * bit masks. | |
2506 | */ | |
2507 | static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp) | |
2508 | { | |
2509 | unsigned long flags; | |
2510 | unsigned long mask; | |
2511 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
2512 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ | |
2513 | ||
2514 | /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ | |
2515 | mask = rdp->grpmask; | |
2516 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2517 | smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */ | |
2518 | rnp->qsmaskinitnext &= ~mask; | |
2519 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2520 | } | |
2521 | ||
64db4cff | 2522 | /* |
e5601400 | 2523 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
2524 | * this fact from process context. Do the remainder of the cleanup, |
2525 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
2526 | * adopting them. There can only be one CPU hotplug operation at a time, |
2527 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 2528 | */ |
e5601400 | 2529 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 2530 | { |
2036d94a | 2531 | unsigned long flags; |
e5601400 | 2532 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 2533 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2534 | |
2036d94a | 2535 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2536 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 2537 | |
b1420f1c | 2538 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
78043c46 | 2539 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
b1420f1c | 2540 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); |
96d3fd0d | 2541 | rcu_adopt_orphan_cbs(rsp, flags); |
a8f4cbad | 2542 | raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags); |
b1420f1c | 2543 | |
cf01537e PM |
2544 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
2545 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2546 | cpu, rdp->qlen, rdp->nxtlist); | |
64db4cff PM |
2547 | } |
2548 | ||
2549 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
2550 | ||
e5601400 | 2551 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
2552 | { |
2553 | } | |
2554 | ||
b6a932d1 PM |
2555 | static void __maybe_unused rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) |
2556 | { | |
2557 | } | |
2558 | ||
88428cc5 PM |
2559 | static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp) |
2560 | { | |
2561 | } | |
2562 | ||
e5601400 | 2563 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
2564 | { |
2565 | } | |
2566 | ||
2567 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
2568 | ||
2569 | /* | |
2570 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2571 | * period. Thottle as specified by rdp->blimit. | |
2572 | */ | |
37c72e56 | 2573 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2574 | { |
2575 | unsigned long flags; | |
2576 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
2577 | long bl, count, count_lazy; |
2578 | int i; | |
64db4cff | 2579 | |
dc35c893 | 2580 | /* If no callbacks are ready, just return. */ |
29c00b4a | 2581 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 2582 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
2583 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
2584 | need_resched(), is_idle_task(current), | |
2585 | rcu_is_callbacks_kthread()); | |
64db4cff | 2586 | return; |
29c00b4a | 2587 | } |
64db4cff PM |
2588 | |
2589 | /* | |
2590 | * Extract the list of ready callbacks, disabling to prevent | |
2591 | * races with call_rcu() from interrupt handlers. | |
2592 | */ | |
2593 | local_irq_save(flags); | |
8146c4e2 | 2594 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2595 | bl = rdp->blimit; |
486e2593 | 2596 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2597 | list = rdp->nxtlist; |
2598 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2599 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2600 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2601 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2602 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2603 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2604 | local_irq_restore(flags); |
2605 | ||
2606 | /* Invoke callbacks. */ | |
486e2593 | 2607 | count = count_lazy = 0; |
64db4cff PM |
2608 | while (list) { |
2609 | next = list->next; | |
2610 | prefetch(next); | |
551d55a9 | 2611 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2612 | if (__rcu_reclaim(rsp->name, list)) |
2613 | count_lazy++; | |
64db4cff | 2614 | list = next; |
dff1672d PM |
2615 | /* Stop only if limit reached and CPU has something to do. */ |
2616 | if (++count >= bl && | |
2617 | (need_resched() || | |
2618 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2619 | break; |
2620 | } | |
2621 | ||
2622 | local_irq_save(flags); | |
4968c300 PM |
2623 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2624 | is_idle_task(current), | |
2625 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2626 | |
2627 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2628 | if (list != NULL) { |
2629 | *tail = rdp->nxtlist; | |
2630 | rdp->nxtlist = list; | |
b41772ab PM |
2631 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2632 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2633 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2634 | else |
2635 | break; | |
2636 | } | |
b1420f1c PM |
2637 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2638 | rdp->qlen_lazy -= count_lazy; | |
a792563b | 2639 | ACCESS_ONCE(rdp->qlen) = rdp->qlen - count; |
b1420f1c | 2640 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2641 | |
2642 | /* Reinstate batch limit if we have worked down the excess. */ | |
2643 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2644 | rdp->blimit = blimit; | |
2645 | ||
37c72e56 PM |
2646 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2647 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2648 | rdp->qlen_last_fqs_check = 0; | |
2649 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2650 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2651 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2652 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2653 | |
64db4cff PM |
2654 | local_irq_restore(flags); |
2655 | ||
e0f23060 | 2656 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2657 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2658 | invoke_rcu_core(); |
64db4cff PM |
2659 | } |
2660 | ||
2661 | /* | |
2662 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2663 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2664 | * Also schedule RCU core processing. |
64db4cff | 2665 | * |
9b2e4f18 | 2666 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2667 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2668 | * false, there is no point in invoking rcu_check_callbacks(). | |
2669 | */ | |
c3377c2d | 2670 | void rcu_check_callbacks(int user) |
64db4cff | 2671 | { |
f7f7bac9 | 2672 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2673 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2674 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2675 | |
2676 | /* | |
2677 | * Get here if this CPU took its interrupt from user | |
2678 | * mode or from the idle loop, and if this is not a | |
2679 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2680 | * a quiescent state, so note it. |
64db4cff PM |
2681 | * |
2682 | * No memory barrier is required here because both | |
d6714c22 PM |
2683 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2684 | * variables that other CPUs neither access nor modify, | |
2685 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2686 | */ |
2687 | ||
284a8c93 PM |
2688 | rcu_sched_qs(); |
2689 | rcu_bh_qs(); | |
64db4cff PM |
2690 | |
2691 | } else if (!in_softirq()) { | |
2692 | ||
2693 | /* | |
2694 | * Get here if this CPU did not take its interrupt from | |
2695 | * softirq, in other words, if it is not interrupting | |
2696 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2697 | * critical section, so note it. |
64db4cff PM |
2698 | */ |
2699 | ||
284a8c93 | 2700 | rcu_bh_qs(); |
64db4cff | 2701 | } |
86aea0e6 | 2702 | rcu_preempt_check_callbacks(); |
e3950ecd | 2703 | if (rcu_pending()) |
a46e0899 | 2704 | invoke_rcu_core(); |
8315f422 PM |
2705 | if (user) |
2706 | rcu_note_voluntary_context_switch(current); | |
f7f7bac9 | 2707 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2708 | } |
2709 | ||
64db4cff PM |
2710 | /* |
2711 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2712 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2713 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2714 | * | |
ee47eb9f | 2715 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2716 | */ |
217af2a2 PM |
2717 | static void force_qs_rnp(struct rcu_state *rsp, |
2718 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2719 | unsigned long *maxj), | |
2720 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2721 | { |
2722 | unsigned long bit; | |
2723 | int cpu; | |
2724 | unsigned long flags; | |
2725 | unsigned long mask; | |
a0b6c9a7 | 2726 | struct rcu_node *rnp; |
64db4cff | 2727 | |
a0b6c9a7 | 2728 | rcu_for_each_leaf_node(rsp, rnp) { |
bde6c3aa | 2729 | cond_resched_rcu_qs(); |
64db4cff | 2730 | mask = 0; |
1304afb2 | 2731 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2732 | smp_mb__after_unlock_lock(); |
ee47eb9f | 2733 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 2734 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 2735 | return; |
64db4cff | 2736 | } |
a0b6c9a7 | 2737 | if (rnp->qsmask == 0) { |
a77da14c PM |
2738 | if (rcu_state_p == &rcu_sched_state || |
2739 | rsp != rcu_state_p || | |
2740 | rcu_preempt_blocked_readers_cgp(rnp)) { | |
2741 | /* | |
2742 | * No point in scanning bits because they | |
2743 | * are all zero. But we might need to | |
2744 | * priority-boost blocked readers. | |
2745 | */ | |
2746 | rcu_initiate_boost(rnp, flags); | |
2747 | /* rcu_initiate_boost() releases rnp->lock */ | |
2748 | continue; | |
2749 | } | |
2750 | if (rnp->parent && | |
2751 | (rnp->parent->qsmask & rnp->grpmask)) { | |
2752 | /* | |
2753 | * Race between grace-period | |
2754 | * initialization and task exiting RCU | |
2755 | * read-side critical section: Report. | |
2756 | */ | |
2757 | rcu_report_unblock_qs_rnp(rsp, rnp, flags); | |
2758 | /* rcu_report_unblock_qs_rnp() rlses ->lock */ | |
2759 | continue; | |
2760 | } | |
64db4cff | 2761 | } |
a0b6c9a7 | 2762 | cpu = rnp->grplo; |
64db4cff | 2763 | bit = 1; |
a0b6c9a7 | 2764 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 | 2765 | if ((rnp->qsmask & bit) != 0) { |
675da67f PM |
2766 | if ((rnp->qsmaskinit & bit) == 0) |
2767 | *isidle = false; /* Pending hotplug. */ | |
0edd1b17 PM |
2768 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) |
2769 | mask |= bit; | |
2770 | } | |
64db4cff | 2771 | } |
45f014c5 | 2772 | if (mask != 0) { |
654e9533 PM |
2773 | /* Idle/offline CPUs, report (releases rnp->lock. */ |
2774 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); | |
0aa04b05 PM |
2775 | } else { |
2776 | /* Nothing to do here, so just drop the lock. */ | |
2777 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff | 2778 | } |
64db4cff | 2779 | } |
64db4cff PM |
2780 | } |
2781 | ||
2782 | /* | |
2783 | * Force quiescent states on reluctant CPUs, and also detect which | |
2784 | * CPUs are in dyntick-idle mode. | |
2785 | */ | |
4cdfc175 | 2786 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2787 | { |
2788 | unsigned long flags; | |
394f2769 PM |
2789 | bool ret; |
2790 | struct rcu_node *rnp; | |
2791 | struct rcu_node *rnp_old = NULL; | |
2792 | ||
2793 | /* Funnel through hierarchy to reduce memory contention. */ | |
d860d403 | 2794 | rnp = __this_cpu_read(rsp->rda->mynode); |
394f2769 PM |
2795 | for (; rnp != NULL; rnp = rnp->parent) { |
2796 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2797 | !raw_spin_trylock(&rnp->fqslock); | |
2798 | if (rnp_old != NULL) | |
2799 | raw_spin_unlock(&rnp_old->fqslock); | |
2800 | if (ret) { | |
a792563b | 2801 | rsp->n_force_qs_lh++; |
394f2769 PM |
2802 | return; |
2803 | } | |
2804 | rnp_old = rnp; | |
2805 | } | |
2806 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2807 | |
394f2769 PM |
2808 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2809 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
6303b9c8 | 2810 | smp_mb__after_unlock_lock(); |
394f2769 PM |
2811 | raw_spin_unlock(&rnp_old->fqslock); |
2812 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
a792563b | 2813 | rsp->n_force_qs_lh++; |
394f2769 | 2814 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2815 | return; /* Someone beat us to it. */ |
46a1e34e | 2816 | } |
4de376a1 PK |
2817 | ACCESS_ONCE(rsp->gp_flags) = |
2818 | ACCESS_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS; | |
394f2769 | 2819 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
2aa792e6 | 2820 | rcu_gp_kthread_wake(rsp); |
64db4cff PM |
2821 | } |
2822 | ||
64db4cff | 2823 | /* |
e0f23060 PM |
2824 | * This does the RCU core processing work for the specified rcu_state |
2825 | * and rcu_data structures. This may be called only from the CPU to | |
2826 | * whom the rdp belongs. | |
64db4cff PM |
2827 | */ |
2828 | static void | |
1bca8cf1 | 2829 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2830 | { |
2831 | unsigned long flags; | |
48a7639c | 2832 | bool needwake; |
fa07a58f | 2833 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
64db4cff | 2834 | |
2e597558 PM |
2835 | WARN_ON_ONCE(rdp->beenonline == 0); |
2836 | ||
64db4cff PM |
2837 | /* Update RCU state based on any recent quiescent states. */ |
2838 | rcu_check_quiescent_state(rsp, rdp); | |
2839 | ||
2840 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2841 | local_irq_save(flags); |
64db4cff | 2842 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2843 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
48a7639c | 2844 | needwake = rcu_start_gp(rsp); |
b8462084 | 2845 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
48a7639c PM |
2846 | if (needwake) |
2847 | rcu_gp_kthread_wake(rsp); | |
dc35c893 PM |
2848 | } else { |
2849 | local_irq_restore(flags); | |
64db4cff PM |
2850 | } |
2851 | ||
2852 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2853 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2854 | invoke_rcu_callbacks(rsp, rdp); |
96d3fd0d PM |
2855 | |
2856 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2857 | do_nocb_deferred_wakeup(rdp); | |
09223371 SL |
2858 | } |
2859 | ||
64db4cff | 2860 | /* |
e0f23060 | 2861 | * Do RCU core processing for the current CPU. |
64db4cff | 2862 | */ |
09223371 | 2863 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2864 | { |
6ce75a23 PM |
2865 | struct rcu_state *rsp; |
2866 | ||
bfa00b4c PM |
2867 | if (cpu_is_offline(smp_processor_id())) |
2868 | return; | |
f7f7bac9 | 2869 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2870 | for_each_rcu_flavor(rsp) |
2871 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2872 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2873 | } |
2874 | ||
a26ac245 | 2875 | /* |
e0f23060 PM |
2876 | * Schedule RCU callback invocation. If the specified type of RCU |
2877 | * does not support RCU priority boosting, just do a direct call, | |
2878 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
924df8a0 | 2879 | * are running on the current CPU with softirqs disabled, the |
e0f23060 | 2880 | * rcu_cpu_kthread_task cannot disappear out from under us. |
a26ac245 | 2881 | */ |
a46e0899 | 2882 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2883 | { |
b0d30417 PM |
2884 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
2885 | return; | |
a46e0899 PM |
2886 | if (likely(!rsp->boost)) { |
2887 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2888 | return; |
2889 | } | |
a46e0899 | 2890 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2891 | } |
2892 | ||
a46e0899 | 2893 | static void invoke_rcu_core(void) |
09223371 | 2894 | { |
b0f74036 PM |
2895 | if (cpu_online(smp_processor_id())) |
2896 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2897 | } |
2898 | ||
29154c57 PM |
2899 | /* |
2900 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2901 | */ | |
2902 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2903 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2904 | { |
48a7639c PM |
2905 | bool needwake; |
2906 | ||
62fde6ed PM |
2907 | /* |
2908 | * If called from an extended quiescent state, invoke the RCU | |
2909 | * core in order to force a re-evaluation of RCU's idleness. | |
2910 | */ | |
9910affa | 2911 | if (!rcu_is_watching()) |
62fde6ed PM |
2912 | invoke_rcu_core(); |
2913 | ||
a16b7a69 | 2914 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2915 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2916 | return; |
64db4cff | 2917 | |
37c72e56 PM |
2918 | /* |
2919 | * Force the grace period if too many callbacks or too long waiting. | |
2920 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2921 | * if some other CPU has recently done so. Also, don't bother | |
2922 | * invoking force_quiescent_state() if the newly enqueued callback | |
2923 | * is the only one waiting for a grace period to complete. | |
2924 | */ | |
2655d57e | 2925 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2926 | |
2927 | /* Are we ignoring a completed grace period? */ | |
470716fc | 2928 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
2929 | |
2930 | /* Start a new grace period if one not already started. */ | |
2931 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
2932 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
2933 | ||
b8462084 | 2934 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 | 2935 | smp_mb__after_unlock_lock(); |
48a7639c | 2936 | needwake = rcu_start_gp(rsp); |
b8462084 | 2937 | raw_spin_unlock(&rnp_root->lock); |
48a7639c PM |
2938 | if (needwake) |
2939 | rcu_gp_kthread_wake(rsp); | |
b52573d2 PM |
2940 | } else { |
2941 | /* Give the grace period a kick. */ | |
2942 | rdp->blimit = LONG_MAX; | |
2943 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2944 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2945 | force_quiescent_state(rsp); |
b52573d2 PM |
2946 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2947 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2948 | } | |
4cdfc175 | 2949 | } |
29154c57 PM |
2950 | } |
2951 | ||
ae150184 PM |
2952 | /* |
2953 | * RCU callback function to leak a callback. | |
2954 | */ | |
2955 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2956 | { | |
2957 | } | |
2958 | ||
3fbfbf7a PM |
2959 | /* |
2960 | * Helper function for call_rcu() and friends. The cpu argument will | |
2961 | * normally be -1, indicating "currently running CPU". It may specify | |
2962 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2963 | * is expected to specify a CPU. | |
2964 | */ | |
64db4cff PM |
2965 | static void |
2966 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 2967 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
2968 | { |
2969 | unsigned long flags; | |
2970 | struct rcu_data *rdp; | |
2971 | ||
1146edcb | 2972 | WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */ |
ae150184 PM |
2973 | if (debug_rcu_head_queue(head)) { |
2974 | /* Probable double call_rcu(), so leak the callback. */ | |
2975 | ACCESS_ONCE(head->func) = rcu_leak_callback; | |
2976 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); | |
2977 | return; | |
2978 | } | |
64db4cff PM |
2979 | head->func = func; |
2980 | head->next = NULL; | |
2981 | ||
64db4cff PM |
2982 | /* |
2983 | * Opportunistically note grace-period endings and beginnings. | |
2984 | * Note that we might see a beginning right after we see an | |
2985 | * end, but never vice versa, since this CPU has to pass through | |
2986 | * a quiescent state betweentimes. | |
2987 | */ | |
2988 | local_irq_save(flags); | |
394f99a9 | 2989 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
2990 | |
2991 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
2992 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
2993 | int offline; | |
2994 | ||
2995 | if (cpu != -1) | |
2996 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
143da9c2 PM |
2997 | if (likely(rdp->mynode)) { |
2998 | /* Post-boot, so this should be for a no-CBs CPU. */ | |
2999 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); | |
3000 | WARN_ON_ONCE(offline); | |
3001 | /* Offline CPU, _call_rcu() illegal, leak callback. */ | |
3002 | local_irq_restore(flags); | |
3003 | return; | |
3004 | } | |
3005 | /* | |
3006 | * Very early boot, before rcu_init(). Initialize if needed | |
3007 | * and then drop through to queue the callback. | |
3008 | */ | |
3009 | BUG_ON(cpu != -1); | |
34404ca8 | 3010 | WARN_ON_ONCE(!rcu_is_watching()); |
143da9c2 PM |
3011 | if (!likely(rdp->nxtlist)) |
3012 | init_default_callback_list(rdp); | |
0d8ee37e | 3013 | } |
a792563b | 3014 | ACCESS_ONCE(rdp->qlen) = rdp->qlen + 1; |
486e2593 PM |
3015 | if (lazy) |
3016 | rdp->qlen_lazy++; | |
c57afe80 PM |
3017 | else |
3018 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
3019 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
3020 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
3021 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 3022 | |
d4c08f2a PM |
3023 | if (__is_kfree_rcu_offset((unsigned long)func)) |
3024 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 3025 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 3026 | else |
486e2593 | 3027 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 3028 | |
29154c57 PM |
3029 | /* Go handle any RCU core processing required. */ |
3030 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
3031 | local_irq_restore(flags); |
3032 | } | |
3033 | ||
3034 | /* | |
d6714c22 | 3035 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 3036 | */ |
d6714c22 | 3037 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 3038 | { |
3fbfbf7a | 3039 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 3040 | } |
d6714c22 | 3041 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
3042 | |
3043 | /* | |
486e2593 | 3044 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
3045 | */ |
3046 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
3047 | { | |
3fbfbf7a | 3048 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
3049 | } |
3050 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
3051 | ||
495aa969 ACB |
3052 | /* |
3053 | * Queue an RCU callback for lazy invocation after a grace period. | |
3054 | * This will likely be later named something like "call_rcu_lazy()", | |
3055 | * but this change will require some way of tagging the lazy RCU | |
3056 | * callbacks in the list of pending callbacks. Until then, this | |
3057 | * function may only be called from __kfree_rcu(). | |
3058 | */ | |
3059 | void kfree_call_rcu(struct rcu_head *head, | |
3060 | void (*func)(struct rcu_head *rcu)) | |
3061 | { | |
e534165b | 3062 | __call_rcu(head, func, rcu_state_p, -1, 1); |
495aa969 ACB |
3063 | } |
3064 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
3065 | ||
6d813391 PM |
3066 | /* |
3067 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
3068 | * any blocking grace-period wait automatically implies a grace period | |
3069 | * if there is only one CPU online at any point time during execution | |
3070 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
3071 | * occasionally incorrectly indicate that there are multiple CPUs online | |
3072 | * when there was in fact only one the whole time, as this just adds | |
3073 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
3074 | */ |
3075 | static inline int rcu_blocking_is_gp(void) | |
3076 | { | |
95f0c1de PM |
3077 | int ret; |
3078 | ||
6d813391 | 3079 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
3080 | preempt_disable(); |
3081 | ret = num_online_cpus() <= 1; | |
3082 | preempt_enable(); | |
3083 | return ret; | |
6d813391 PM |
3084 | } |
3085 | ||
6ebb237b PM |
3086 | /** |
3087 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
3088 | * | |
3089 | * Control will return to the caller some time after a full rcu-sched | |
3090 | * grace period has elapsed, in other words after all currently executing | |
3091 | * rcu-sched read-side critical sections have completed. These read-side | |
3092 | * critical sections are delimited by rcu_read_lock_sched() and | |
3093 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
3094 | * local_irq_disable(), and so on may be used in place of | |
3095 | * rcu_read_lock_sched(). | |
3096 | * | |
3097 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
3098 | * non-threaded hardware-interrupt handlers, in progress on entry will |
3099 | * have completed before this primitive returns. However, this does not | |
3100 | * guarantee that softirq handlers will have completed, since in some | |
3101 | * kernels, these handlers can run in process context, and can block. | |
3102 | * | |
3103 | * Note that this guarantee implies further memory-ordering guarantees. | |
3104 | * On systems with more than one CPU, when synchronize_sched() returns, | |
3105 | * each CPU is guaranteed to have executed a full memory barrier since the | |
3106 | * end of its last RCU-sched read-side critical section whose beginning | |
3107 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
3108 | * an RCU read-side critical section that extends beyond the return from | |
3109 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
3110 | * after the beginning of synchronize_sched() and before the beginning of | |
3111 | * that RCU read-side critical section. Note that these guarantees include | |
3112 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
3113 | * that are executing in the kernel. | |
3114 | * | |
3115 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
3116 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
3117 | * to have executed a full memory barrier during the execution of | |
3118 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
3119 | * again only if the system has more than one CPU). | |
6ebb237b PM |
3120 | * |
3121 | * This primitive provides the guarantees made by the (now removed) | |
3122 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
3123 | * guarantees that rcu_read_lock() sections will have completed. | |
3124 | * In "classic RCU", these two guarantees happen to be one and | |
3125 | * the same, but can differ in realtime RCU implementations. | |
3126 | */ | |
3127 | void synchronize_sched(void) | |
3128 | { | |
fe15d706 PM |
3129 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
3130 | !lock_is_held(&rcu_lock_map) && | |
3131 | !lock_is_held(&rcu_sched_lock_map), | |
3132 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
3133 | if (rcu_blocking_is_gp()) |
3134 | return; | |
5afff48b | 3135 | if (rcu_gp_is_expedited()) |
3705b88d AM |
3136 | synchronize_sched_expedited(); |
3137 | else | |
3138 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
3139 | } |
3140 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
3141 | ||
3142 | /** | |
3143 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
3144 | * | |
3145 | * Control will return to the caller some time after a full rcu_bh grace | |
3146 | * period has elapsed, in other words after all currently executing rcu_bh | |
3147 | * read-side critical sections have completed. RCU read-side critical | |
3148 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
3149 | * and may be nested. | |
f0a0e6f2 PM |
3150 | * |
3151 | * See the description of synchronize_sched() for more detailed information | |
3152 | * on memory ordering guarantees. | |
6ebb237b PM |
3153 | */ |
3154 | void synchronize_rcu_bh(void) | |
3155 | { | |
fe15d706 PM |
3156 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
3157 | !lock_is_held(&rcu_lock_map) && | |
3158 | !lock_is_held(&rcu_sched_lock_map), | |
3159 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
3160 | if (rcu_blocking_is_gp()) |
3161 | return; | |
5afff48b | 3162 | if (rcu_gp_is_expedited()) |
3705b88d AM |
3163 | synchronize_rcu_bh_expedited(); |
3164 | else | |
3165 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
3166 | } |
3167 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
3168 | ||
765a3f4f PM |
3169 | /** |
3170 | * get_state_synchronize_rcu - Snapshot current RCU state | |
3171 | * | |
3172 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
3173 | * to determine whether or not a full grace period has elapsed in the | |
3174 | * meantime. | |
3175 | */ | |
3176 | unsigned long get_state_synchronize_rcu(void) | |
3177 | { | |
3178 | /* | |
3179 | * Any prior manipulation of RCU-protected data must happen | |
3180 | * before the load from ->gpnum. | |
3181 | */ | |
3182 | smp_mb(); /* ^^^ */ | |
3183 | ||
3184 | /* | |
3185 | * Make sure this load happens before the purportedly | |
3186 | * time-consuming work between get_state_synchronize_rcu() | |
3187 | * and cond_synchronize_rcu(). | |
3188 | */ | |
e534165b | 3189 | return smp_load_acquire(&rcu_state_p->gpnum); |
765a3f4f PM |
3190 | } |
3191 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
3192 | ||
3193 | /** | |
3194 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
3195 | * | |
3196 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
3197 | * | |
3198 | * If a full RCU grace period has elapsed since the earlier call to | |
3199 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
3200 | * synchronize_rcu() to wait for a full grace period. | |
3201 | * | |
3202 | * Yes, this function does not take counter wrap into account. But | |
3203 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3204 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3205 | * so waiting for one additional grace period should be just fine. | |
3206 | */ | |
3207 | void cond_synchronize_rcu(unsigned long oldstate) | |
3208 | { | |
3209 | unsigned long newstate; | |
3210 | ||
3211 | /* | |
3212 | * Ensure that this load happens before any RCU-destructive | |
3213 | * actions the caller might carry out after we return. | |
3214 | */ | |
e534165b | 3215 | newstate = smp_load_acquire(&rcu_state_p->completed); |
765a3f4f PM |
3216 | if (ULONG_CMP_GE(oldstate, newstate)) |
3217 | synchronize_rcu(); | |
3218 | } | |
3219 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
3220 | ||
3d3b7db0 PM |
3221 | static int synchronize_sched_expedited_cpu_stop(void *data) |
3222 | { | |
3223 | /* | |
3224 | * There must be a full memory barrier on each affected CPU | |
3225 | * between the time that try_stop_cpus() is called and the | |
3226 | * time that it returns. | |
3227 | * | |
3228 | * In the current initial implementation of cpu_stop, the | |
3229 | * above condition is already met when the control reaches | |
3230 | * this point and the following smp_mb() is not strictly | |
3231 | * necessary. Do smp_mb() anyway for documentation and | |
3232 | * robustness against future implementation changes. | |
3233 | */ | |
3234 | smp_mb(); /* See above comment block. */ | |
3235 | return 0; | |
3236 | } | |
3237 | ||
236fefaf PM |
3238 | /** |
3239 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
3240 | * | |
3241 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
3242 | * approach to force the grace period to end quickly. This consumes | |
3243 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
3244 | * so is thus not recommended for any sort of common-case code. In fact, | |
3245 | * if you are using synchronize_sched_expedited() in a loop, please | |
3246 | * restructure your code to batch your updates, and then use a single | |
3247 | * synchronize_sched() instead. | |
3d3b7db0 | 3248 | * |
3d3b7db0 PM |
3249 | * This implementation can be thought of as an application of ticket |
3250 | * locking to RCU, with sync_sched_expedited_started and | |
3251 | * sync_sched_expedited_done taking on the roles of the halves | |
3252 | * of the ticket-lock word. Each task atomically increments | |
3253 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
3254 | * then attempts to stop all the CPUs. If this succeeds, then each | |
3255 | * CPU will have executed a context switch, resulting in an RCU-sched | |
3256 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
3257 | * update sync_sched_expedited_done to match our snapshot -- but | |
3258 | * only if someone else has not already advanced past our snapshot. | |
3259 | * | |
3260 | * On the other hand, if try_stop_cpus() fails, we check the value | |
3261 | * of sync_sched_expedited_done. If it has advanced past our | |
3262 | * initial snapshot, then someone else must have forced a grace period | |
3263 | * some time after we took our snapshot. In this case, our work is | |
3264 | * done for us, and we can simply return. Otherwise, we try again, | |
3265 | * but keep our initial snapshot for purposes of checking for someone | |
3266 | * doing our work for us. | |
3267 | * | |
3268 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
3269 | */ | |
3270 | void synchronize_sched_expedited(void) | |
3271 | { | |
e0775cef PM |
3272 | cpumask_var_t cm; |
3273 | bool cma = false; | |
3274 | int cpu; | |
1924bcb0 PM |
3275 | long firstsnap, s, snap; |
3276 | int trycount = 0; | |
40694d66 | 3277 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 3278 | |
1924bcb0 PM |
3279 | /* |
3280 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
3281 | * By allowing sync_sched_expedited_started to advance no more than | |
3282 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
3283 | * that more than 3.5 billion CPUs would be required to force a | |
3284 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
3285 | * course be required on a 64-bit system. | |
3286 | */ | |
40694d66 PM |
3287 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
3288 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 PM |
3289 | ULONG_MAX / 8)) { |
3290 | synchronize_sched(); | |
a30489c5 | 3291 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
3292 | return; |
3293 | } | |
3d3b7db0 | 3294 | |
1924bcb0 PM |
3295 | /* |
3296 | * Take a ticket. Note that atomic_inc_return() implies a | |
3297 | * full memory barrier. | |
3298 | */ | |
40694d66 | 3299 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 3300 | firstsnap = snap; |
dd56af42 PM |
3301 | if (!try_get_online_cpus()) { |
3302 | /* CPU hotplug operation in flight, fall back to normal GP. */ | |
3303 | wait_rcu_gp(call_rcu_sched); | |
3304 | atomic_long_inc(&rsp->expedited_normal); | |
3305 | return; | |
3306 | } | |
1cc85961 | 3307 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 | 3308 | |
e0775cef PM |
3309 | /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */ |
3310 | cma = zalloc_cpumask_var(&cm, GFP_KERNEL); | |
3311 | if (cma) { | |
3312 | cpumask_copy(cm, cpu_online_mask); | |
3313 | cpumask_clear_cpu(raw_smp_processor_id(), cm); | |
3314 | for_each_cpu(cpu, cm) { | |
3315 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); | |
3316 | ||
3317 | if (!(atomic_add_return(0, &rdtp->dynticks) & 0x1)) | |
3318 | cpumask_clear_cpu(cpu, cm); | |
3319 | } | |
3320 | if (cpumask_weight(cm) == 0) | |
3321 | goto all_cpus_idle; | |
3322 | } | |
3323 | ||
3d3b7db0 PM |
3324 | /* |
3325 | * Each pass through the following loop attempts to force a | |
3326 | * context switch on each CPU. | |
3327 | */ | |
e0775cef | 3328 | while (try_stop_cpus(cma ? cm : cpu_online_mask, |
3d3b7db0 PM |
3329 | synchronize_sched_expedited_cpu_stop, |
3330 | NULL) == -EAGAIN) { | |
3331 | put_online_cpus(); | |
a30489c5 | 3332 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 3333 | |
1924bcb0 | 3334 | /* Check to see if someone else did our work for us. */ |
40694d66 | 3335 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3336 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 | 3337 | /* ensure test happens before caller kfree */ |
4e857c58 | 3338 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3339 | atomic_long_inc(&rsp->expedited_workdone1); |
e0775cef | 3340 | free_cpumask_var(cm); |
1924bcb0 PM |
3341 | return; |
3342 | } | |
3d3b7db0 PM |
3343 | |
3344 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 3345 | if (trycount++ < 10) { |
3d3b7db0 | 3346 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 3347 | } else { |
3705b88d | 3348 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 3349 | atomic_long_inc(&rsp->expedited_normal); |
e0775cef | 3350 | free_cpumask_var(cm); |
3d3b7db0 PM |
3351 | return; |
3352 | } | |
3353 | ||
1924bcb0 | 3354 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 3355 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3356 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 | 3357 | /* ensure test happens before caller kfree */ |
4e857c58 | 3358 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3359 | atomic_long_inc(&rsp->expedited_workdone2); |
e0775cef | 3360 | free_cpumask_var(cm); |
3d3b7db0 PM |
3361 | return; |
3362 | } | |
3363 | ||
3364 | /* | |
3365 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
3366 | * callers to piggyback on our grace period. We retry |
3367 | * after they started, so our grace period works for them, | |
3368 | * and they started after our first try, so their grace | |
3369 | * period works for us. | |
3d3b7db0 | 3370 | */ |
dd56af42 PM |
3371 | if (!try_get_online_cpus()) { |
3372 | /* CPU hotplug operation in flight, use normal GP. */ | |
3373 | wait_rcu_gp(call_rcu_sched); | |
3374 | atomic_long_inc(&rsp->expedited_normal); | |
e0775cef | 3375 | free_cpumask_var(cm); |
dd56af42 PM |
3376 | return; |
3377 | } | |
40694d66 | 3378 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
3379 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
3380 | } | |
a30489c5 | 3381 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 | 3382 | |
e0775cef PM |
3383 | all_cpus_idle: |
3384 | free_cpumask_var(cm); | |
3385 | ||
3d3b7db0 PM |
3386 | /* |
3387 | * Everyone up to our most recent fetch is covered by our grace | |
3388 | * period. Update the counter, but only if our work is still | |
3389 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 3390 | * than we did already did their update. |
3d3b7db0 PM |
3391 | */ |
3392 | do { | |
a30489c5 | 3393 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 3394 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3395 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 | 3396 | /* ensure test happens before caller kfree */ |
4e857c58 | 3397 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3398 | atomic_long_inc(&rsp->expedited_done_lost); |
3d3b7db0 PM |
3399 | break; |
3400 | } | |
40694d66 | 3401 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 3402 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
3403 | |
3404 | put_online_cpus(); | |
3405 | } | |
3406 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
3407 | ||
64db4cff PM |
3408 | /* |
3409 | * Check to see if there is any immediate RCU-related work to be done | |
3410 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
3411 | * The checks are in order of increasing expense: checks that can be | |
3412 | * carried out against CPU-local state are performed first. However, | |
3413 | * we must check for CPU stalls first, else we might not get a chance. | |
3414 | */ | |
3415 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
3416 | { | |
2f51f988 PM |
3417 | struct rcu_node *rnp = rdp->mynode; |
3418 | ||
64db4cff PM |
3419 | rdp->n_rcu_pending++; |
3420 | ||
3421 | /* Check for CPU stalls, if enabled. */ | |
3422 | check_cpu_stall(rsp, rdp); | |
3423 | ||
a096932f PM |
3424 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ |
3425 | if (rcu_nohz_full_cpu(rsp)) | |
3426 | return 0; | |
3427 | ||
64db4cff | 3428 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
5c51dd73 | 3429 | if (rcu_scheduler_fully_active && |
5cd37193 PM |
3430 | rdp->qs_pending && !rdp->passed_quiesce && |
3431 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) { | |
d21670ac | 3432 | rdp->n_rp_qs_pending++; |
5cd37193 PM |
3433 | } else if (rdp->qs_pending && |
3434 | (rdp->passed_quiesce || | |
3435 | rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) { | |
d21670ac | 3436 | rdp->n_rp_report_qs++; |
64db4cff | 3437 | return 1; |
7ba5c840 | 3438 | } |
64db4cff PM |
3439 | |
3440 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
3441 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
3442 | rdp->n_rp_cb_ready++; | |
64db4cff | 3443 | return 1; |
7ba5c840 | 3444 | } |
64db4cff PM |
3445 | |
3446 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
3447 | if (cpu_needs_another_gp(rsp, rdp)) { |
3448 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 3449 | return 1; |
7ba5c840 | 3450 | } |
64db4cff PM |
3451 | |
3452 | /* Has another RCU grace period completed? */ | |
2f51f988 | 3453 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 3454 | rdp->n_rp_gp_completed++; |
64db4cff | 3455 | return 1; |
7ba5c840 | 3456 | } |
64db4cff PM |
3457 | |
3458 | /* Has a new RCU grace period started? */ | |
e3663b10 PM |
3459 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum || |
3460 | unlikely(ACCESS_ONCE(rdp->gpwrap))) { /* outside lock */ | |
7ba5c840 | 3461 | rdp->n_rp_gp_started++; |
64db4cff | 3462 | return 1; |
7ba5c840 | 3463 | } |
64db4cff | 3464 | |
96d3fd0d PM |
3465 | /* Does this CPU need a deferred NOCB wakeup? */ |
3466 | if (rcu_nocb_need_deferred_wakeup(rdp)) { | |
3467 | rdp->n_rp_nocb_defer_wakeup++; | |
3468 | return 1; | |
3469 | } | |
3470 | ||
64db4cff | 3471 | /* nothing to do */ |
7ba5c840 | 3472 | rdp->n_rp_need_nothing++; |
64db4cff PM |
3473 | return 0; |
3474 | } | |
3475 | ||
3476 | /* | |
3477 | * Check to see if there is any immediate RCU-related work to be done | |
3478 | * by the current CPU, returning 1 if so. This function is part of the | |
3479 | * RCU implementation; it is -not- an exported member of the RCU API. | |
3480 | */ | |
e3950ecd | 3481 | static int rcu_pending(void) |
64db4cff | 3482 | { |
6ce75a23 PM |
3483 | struct rcu_state *rsp; |
3484 | ||
3485 | for_each_rcu_flavor(rsp) | |
e3950ecd | 3486 | if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda))) |
6ce75a23 PM |
3487 | return 1; |
3488 | return 0; | |
64db4cff PM |
3489 | } |
3490 | ||
3491 | /* | |
c0f4dfd4 PM |
3492 | * Return true if the specified CPU has any callback. If all_lazy is |
3493 | * non-NULL, store an indication of whether all callbacks are lazy. | |
3494 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 3495 | */ |
aa6da514 | 3496 | static int __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy) |
64db4cff | 3497 | { |
c0f4dfd4 PM |
3498 | bool al = true; |
3499 | bool hc = false; | |
3500 | struct rcu_data *rdp; | |
6ce75a23 PM |
3501 | struct rcu_state *rsp; |
3502 | ||
c0f4dfd4 | 3503 | for_each_rcu_flavor(rsp) { |
aa6da514 | 3504 | rdp = this_cpu_ptr(rsp->rda); |
69c8d28c PM |
3505 | if (!rdp->nxtlist) |
3506 | continue; | |
3507 | hc = true; | |
3508 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
c0f4dfd4 | 3509 | al = false; |
69c8d28c PM |
3510 | break; |
3511 | } | |
c0f4dfd4 PM |
3512 | } |
3513 | if (all_lazy) | |
3514 | *all_lazy = al; | |
3515 | return hc; | |
64db4cff PM |
3516 | } |
3517 | ||
a83eff0a PM |
3518 | /* |
3519 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
3520 | * the compiler is expected to optimize this away. | |
3521 | */ | |
e66c33d5 | 3522 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
3523 | int cpu, unsigned long done) |
3524 | { | |
3525 | trace_rcu_barrier(rsp->name, s, cpu, | |
3526 | atomic_read(&rsp->barrier_cpu_count), done); | |
3527 | } | |
3528 | ||
b1420f1c PM |
3529 | /* |
3530 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
3531 | * up the task executing _rcu_barrier(). | |
3532 | */ | |
24ebbca8 | 3533 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3534 | { |
24ebbca8 PM |
3535 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
3536 | struct rcu_state *rsp = rdp->rsp; | |
3537 | ||
a83eff0a PM |
3538 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
3539 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 3540 | complete(&rsp->barrier_completion); |
a83eff0a PM |
3541 | } else { |
3542 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
3543 | } | |
d0ec774c PM |
3544 | } |
3545 | ||
3546 | /* | |
3547 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3548 | */ | |
3549 | static void rcu_barrier_func(void *type) | |
3550 | { | |
037b64ed | 3551 | struct rcu_state *rsp = type; |
fa07a58f | 3552 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
d0ec774c | 3553 | |
a83eff0a | 3554 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 3555 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 3556 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
3557 | } |
3558 | ||
d0ec774c PM |
3559 | /* |
3560 | * Orchestrate the specified type of RCU barrier, waiting for all | |
3561 | * RCU callbacks of the specified type to complete. | |
3562 | */ | |
037b64ed | 3563 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 3564 | { |
b1420f1c | 3565 | int cpu; |
b1420f1c | 3566 | struct rcu_data *rdp; |
cf3a9c48 PM |
3567 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
3568 | unsigned long snap_done; | |
b1420f1c | 3569 | |
a83eff0a | 3570 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 3571 | |
e74f4c45 | 3572 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 3573 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 3574 | |
cf3a9c48 PM |
3575 | /* |
3576 | * Ensure that all prior references, including to ->n_barrier_done, | |
3577 | * are ordered before the _rcu_barrier() machinery. | |
3578 | */ | |
3579 | smp_mb(); /* See above block comment. */ | |
3580 | ||
3581 | /* | |
3582 | * Recheck ->n_barrier_done to see if others did our work for us. | |
3583 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
3584 | * transition. The "if" expression below therefore rounds the old | |
3585 | * value up to the next even number and adds two before comparing. | |
3586 | */ | |
458fb381 | 3587 | snap_done = rsp->n_barrier_done; |
a83eff0a | 3588 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
458fb381 PM |
3589 | |
3590 | /* | |
3591 | * If the value in snap is odd, we needed to wait for the current | |
3592 | * rcu_barrier() to complete, then wait for the next one, in other | |
3593 | * words, we need the value of snap_done to be three larger than | |
3594 | * the value of snap. On the other hand, if the value in snap is | |
3595 | * even, we only had to wait for the next rcu_barrier() to complete, | |
3596 | * in other words, we need the value of snap_done to be only two | |
3597 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes | |
3598 | * this for us (thank you, Linus!). | |
3599 | */ | |
3600 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) { | |
a83eff0a | 3601 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
3602 | smp_mb(); /* caller's subsequent code after above check. */ |
3603 | mutex_unlock(&rsp->barrier_mutex); | |
3604 | return; | |
3605 | } | |
3606 | ||
3607 | /* | |
3608 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
3609 | * ACCESS_ONCE() to prevent the compiler from speculating | |
3610 | * the increment to precede the early-exit check. | |
3611 | */ | |
a792563b | 3612 | ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1; |
cf3a9c48 | 3613 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); |
a83eff0a | 3614 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 3615 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 3616 | |
d0ec774c | 3617 | /* |
b1420f1c PM |
3618 | * Initialize the count to one rather than to zero in order to |
3619 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
3620 | * (or preemption of this task). Exclude CPU-hotplug operations |
3621 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 3622 | */ |
7db74df8 | 3623 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 3624 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 3625 | get_online_cpus(); |
b1420f1c PM |
3626 | |
3627 | /* | |
1331e7a1 PM |
3628 | * Force each CPU with callbacks to register a new callback. |
3629 | * When that callback is invoked, we will know that all of the | |
3630 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 3631 | */ |
3fbfbf7a | 3632 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 3633 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 3634 | continue; |
b1420f1c | 3635 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 3636 | if (rcu_is_nocb_cpu(cpu)) { |
d7e29933 PM |
3637 | if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) { |
3638 | _rcu_barrier_trace(rsp, "OfflineNoCB", cpu, | |
3639 | rsp->n_barrier_done); | |
3640 | } else { | |
3641 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, | |
3642 | rsp->n_barrier_done); | |
41050a00 | 3643 | smp_mb__before_atomic(); |
d7e29933 PM |
3644 | atomic_inc(&rsp->barrier_cpu_count); |
3645 | __call_rcu(&rdp->barrier_head, | |
3646 | rcu_barrier_callback, rsp, cpu, 0); | |
3647 | } | |
3fbfbf7a | 3648 | } else if (ACCESS_ONCE(rdp->qlen)) { |
a83eff0a PM |
3649 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
3650 | rsp->n_barrier_done); | |
037b64ed | 3651 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 3652 | } else { |
a83eff0a PM |
3653 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
3654 | rsp->n_barrier_done); | |
b1420f1c PM |
3655 | } |
3656 | } | |
1331e7a1 | 3657 | put_online_cpus(); |
b1420f1c PM |
3658 | |
3659 | /* | |
3660 | * Now that we have an rcu_barrier_callback() callback on each | |
3661 | * CPU, and thus each counted, remove the initial count. | |
3662 | */ | |
24ebbca8 | 3663 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 3664 | complete(&rsp->barrier_completion); |
b1420f1c | 3665 | |
cf3a9c48 PM |
3666 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
3667 | smp_mb(); /* Keep increment after above mechanism. */ | |
a792563b | 3668 | ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1; |
cf3a9c48 | 3669 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); |
a83eff0a | 3670 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
3671 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
3672 | ||
b1420f1c | 3673 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 3674 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
3675 | |
3676 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 3677 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 3678 | } |
d0ec774c PM |
3679 | |
3680 | /** | |
3681 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
3682 | */ | |
3683 | void rcu_barrier_bh(void) | |
3684 | { | |
037b64ed | 3685 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
3686 | } |
3687 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
3688 | ||
3689 | /** | |
3690 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
3691 | */ | |
3692 | void rcu_barrier_sched(void) | |
3693 | { | |
037b64ed | 3694 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
3695 | } |
3696 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
3697 | ||
0aa04b05 PM |
3698 | /* |
3699 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the | |
3700 | * first CPU in a given leaf rcu_node structure coming online. The caller | |
3701 | * must hold the corresponding leaf rcu_node ->lock with interrrupts | |
3702 | * disabled. | |
3703 | */ | |
3704 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) | |
3705 | { | |
3706 | long mask; | |
3707 | struct rcu_node *rnp = rnp_leaf; | |
3708 | ||
3709 | for (;;) { | |
3710 | mask = rnp->grpmask; | |
3711 | rnp = rnp->parent; | |
3712 | if (rnp == NULL) | |
3713 | return; | |
3714 | raw_spin_lock(&rnp->lock); /* Interrupts already disabled. */ | |
3715 | rnp->qsmaskinit |= mask; | |
3716 | raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */ | |
3717 | } | |
3718 | } | |
3719 | ||
64db4cff | 3720 | /* |
27569620 | 3721 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 3722 | */ |
27569620 PM |
3723 | static void __init |
3724 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
3725 | { |
3726 | unsigned long flags; | |
394f99a9 | 3727 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
3728 | struct rcu_node *rnp = rcu_get_root(rsp); |
3729 | ||
3730 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 3731 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 3732 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
27569620 | 3733 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 3734 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 3735 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 3736 | rdp->cpu = cpu; |
d4c08f2a | 3737 | rdp->rsp = rsp; |
3fbfbf7a | 3738 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 3739 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
3740 | } |
3741 | ||
3742 | /* | |
3743 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
3744 | * offline event can be happening at a given time. Note also that we | |
3745 | * can accept some slop in the rsp->completed access due to the fact | |
3746 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 3747 | */ |
49fb4c62 | 3748 | static void |
9b67122a | 3749 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp) |
64db4cff PM |
3750 | { |
3751 | unsigned long flags; | |
64db4cff | 3752 | unsigned long mask; |
394f99a9 | 3753 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
3754 | struct rcu_node *rnp = rcu_get_root(rsp); |
3755 | ||
3756 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 3757 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 3758 | rdp->beenonline = 1; /* We have now been online. */ |
37c72e56 PM |
3759 | rdp->qlen_last_fqs_check = 0; |
3760 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 3761 | rdp->blimit = blimit; |
39c8d313 PM |
3762 | if (!rdp->nxtlist) |
3763 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ | |
29e37d81 | 3764 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 3765 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
3766 | atomic_set(&rdp->dynticks->dynticks, |
3767 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 3768 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 3769 | |
0aa04b05 PM |
3770 | /* |
3771 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed | |
3772 | * propagation up the rcu_node tree will happen at the beginning | |
3773 | * of the next grace period. | |
3774 | */ | |
64db4cff PM |
3775 | rnp = rdp->mynode; |
3776 | mask = rdp->grpmask; | |
0aa04b05 PM |
3777 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
3778 | smp_mb__after_unlock_lock(); | |
3779 | rnp->qsmaskinitnext |= mask; | |
3780 | rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */ | |
3781 | rdp->completed = rnp->completed; | |
3782 | rdp->passed_quiesce = false; | |
3783 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); | |
3784 | rdp->qs_pending = false; | |
3785 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); | |
3786 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff PM |
3787 | } |
3788 | ||
49fb4c62 | 3789 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 3790 | { |
6ce75a23 PM |
3791 | struct rcu_state *rsp; |
3792 | ||
3793 | for_each_rcu_flavor(rsp) | |
9b67122a | 3794 | rcu_init_percpu_data(cpu, rsp); |
64db4cff PM |
3795 | } |
3796 | ||
3797 | /* | |
f41d911f | 3798 | * Handle CPU online/offline notification events. |
64db4cff | 3799 | */ |
88428cc5 PM |
3800 | int rcu_cpu_notify(struct notifier_block *self, |
3801 | unsigned long action, void *hcpu) | |
64db4cff PM |
3802 | { |
3803 | long cpu = (long)hcpu; | |
e534165b | 3804 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
a26ac245 | 3805 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 3806 | struct rcu_state *rsp; |
64db4cff PM |
3807 | |
3808 | switch (action) { | |
3809 | case CPU_UP_PREPARE: | |
3810 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
3811 | rcu_prepare_cpu(cpu); |
3812 | rcu_prepare_kthreads(cpu); | |
35ce7f29 | 3813 | rcu_spawn_all_nocb_kthreads(cpu); |
a26ac245 PM |
3814 | break; |
3815 | case CPU_ONLINE: | |
0f962a5e | 3816 | case CPU_DOWN_FAILED: |
5d01bbd1 | 3817 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
3818 | break; |
3819 | case CPU_DOWN_PREPARE: | |
34ed6246 | 3820 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 3821 | break; |
d0ec774c PM |
3822 | case CPU_DYING: |
3823 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
3824 | for_each_rcu_flavor(rsp) |
3825 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 3826 | break; |
88428cc5 PM |
3827 | case CPU_DYING_IDLE: |
3828 | for_each_rcu_flavor(rsp) { | |
3829 | rcu_cleanup_dying_idle_cpu(cpu, rsp); | |
3830 | } | |
3831 | break; | |
64db4cff PM |
3832 | case CPU_DEAD: |
3833 | case CPU_DEAD_FROZEN: | |
3834 | case CPU_UP_CANCELED: | |
3835 | case CPU_UP_CANCELED_FROZEN: | |
776d6807 | 3836 | for_each_rcu_flavor(rsp) { |
6ce75a23 | 3837 | rcu_cleanup_dead_cpu(cpu, rsp); |
776d6807 PM |
3838 | do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu)); |
3839 | } | |
64db4cff PM |
3840 | break; |
3841 | default: | |
3842 | break; | |
3843 | } | |
34ed6246 | 3844 | return NOTIFY_OK; |
64db4cff PM |
3845 | } |
3846 | ||
d1d74d14 BP |
3847 | static int rcu_pm_notify(struct notifier_block *self, |
3848 | unsigned long action, void *hcpu) | |
3849 | { | |
3850 | switch (action) { | |
3851 | case PM_HIBERNATION_PREPARE: | |
3852 | case PM_SUSPEND_PREPARE: | |
3853 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
5afff48b | 3854 | rcu_expedite_gp(); |
d1d74d14 BP |
3855 | break; |
3856 | case PM_POST_HIBERNATION: | |
3857 | case PM_POST_SUSPEND: | |
5afff48b PM |
3858 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ |
3859 | rcu_unexpedite_gp(); | |
d1d74d14 BP |
3860 | break; |
3861 | default: | |
3862 | break; | |
3863 | } | |
3864 | return NOTIFY_OK; | |
3865 | } | |
3866 | ||
b3dbec76 | 3867 | /* |
9386c0b7 | 3868 | * Spawn the kthreads that handle each RCU flavor's grace periods. |
b3dbec76 PM |
3869 | */ |
3870 | static int __init rcu_spawn_gp_kthread(void) | |
3871 | { | |
3872 | unsigned long flags; | |
a94844b2 | 3873 | int kthread_prio_in = kthread_prio; |
b3dbec76 PM |
3874 | struct rcu_node *rnp; |
3875 | struct rcu_state *rsp; | |
a94844b2 | 3876 | struct sched_param sp; |
b3dbec76 PM |
3877 | struct task_struct *t; |
3878 | ||
a94844b2 PM |
3879 | /* Force priority into range. */ |
3880 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) | |
3881 | kthread_prio = 1; | |
3882 | else if (kthread_prio < 0) | |
3883 | kthread_prio = 0; | |
3884 | else if (kthread_prio > 99) | |
3885 | kthread_prio = 99; | |
3886 | if (kthread_prio != kthread_prio_in) | |
3887 | pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n", | |
3888 | kthread_prio, kthread_prio_in); | |
3889 | ||
9386c0b7 | 3890 | rcu_scheduler_fully_active = 1; |
b3dbec76 | 3891 | for_each_rcu_flavor(rsp) { |
a94844b2 | 3892 | t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
3893 | BUG_ON(IS_ERR(t)); |
3894 | rnp = rcu_get_root(rsp); | |
3895 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3896 | rsp->gp_kthread = t; | |
a94844b2 PM |
3897 | if (kthread_prio) { |
3898 | sp.sched_priority = kthread_prio; | |
3899 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
3900 | } | |
3901 | wake_up_process(t); | |
b3dbec76 PM |
3902 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
3903 | } | |
35ce7f29 | 3904 | rcu_spawn_nocb_kthreads(); |
9386c0b7 | 3905 | rcu_spawn_boost_kthreads(); |
b3dbec76 PM |
3906 | return 0; |
3907 | } | |
3908 | early_initcall(rcu_spawn_gp_kthread); | |
3909 | ||
bbad9379 PM |
3910 | /* |
3911 | * This function is invoked towards the end of the scheduler's initialization | |
3912 | * process. Before this is called, the idle task might contain | |
3913 | * RCU read-side critical sections (during which time, this idle | |
3914 | * task is booting the system). After this function is called, the | |
3915 | * idle tasks are prohibited from containing RCU read-side critical | |
3916 | * sections. This function also enables RCU lockdep checking. | |
3917 | */ | |
3918 | void rcu_scheduler_starting(void) | |
3919 | { | |
3920 | WARN_ON(num_online_cpus() != 1); | |
3921 | WARN_ON(nr_context_switches() > 0); | |
3922 | rcu_scheduler_active = 1; | |
3923 | } | |
3924 | ||
64db4cff PM |
3925 | /* |
3926 | * Compute the per-level fanout, either using the exact fanout specified | |
3927 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
3928 | */ | |
64db4cff PM |
3929 | static void __init rcu_init_levelspread(struct rcu_state *rsp) |
3930 | { | |
64db4cff PM |
3931 | int i; |
3932 | ||
66292405 PM |
3933 | if (IS_ENABLED(CONFIG_RCU_FANOUT_EXACT)) { |
3934 | rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; | |
3935 | for (i = rcu_num_lvls - 2; i >= 0; i--) | |
3936 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; | |
3937 | } else { | |
3938 | int ccur; | |
3939 | int cprv; | |
3940 | ||
3941 | cprv = nr_cpu_ids; | |
3942 | for (i = rcu_num_lvls - 1; i >= 0; i--) { | |
3943 | ccur = rsp->levelcnt[i]; | |
3944 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
3945 | cprv = ccur; | |
3946 | } | |
64db4cff PM |
3947 | } |
3948 | } | |
64db4cff PM |
3949 | |
3950 | /* | |
3951 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3952 | */ | |
394f99a9 LJ |
3953 | static void __init rcu_init_one(struct rcu_state *rsp, |
3954 | struct rcu_data __percpu *rda) | |
64db4cff | 3955 | { |
b4426b49 FF |
3956 | static const char * const buf[] = { |
3957 | "rcu_node_0", | |
3958 | "rcu_node_1", | |
3959 | "rcu_node_2", | |
3960 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ | |
3961 | static const char * const fqs[] = { | |
3962 | "rcu_node_fqs_0", | |
3963 | "rcu_node_fqs_1", | |
3964 | "rcu_node_fqs_2", | |
3965 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ | |
4a81e832 | 3966 | static u8 fl_mask = 0x1; |
64db4cff PM |
3967 | int cpustride = 1; |
3968 | int i; | |
3969 | int j; | |
3970 | struct rcu_node *rnp; | |
3971 | ||
b6407e86 PM |
3972 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
3973 | ||
4930521a PM |
3974 | /* Silence gcc 4.8 warning about array index out of range. */ |
3975 | if (rcu_num_lvls > RCU_NUM_LVLS) | |
3976 | panic("rcu_init_one: rcu_num_lvls overflow"); | |
3977 | ||
64db4cff PM |
3978 | /* Initialize the level-tracking arrays. */ |
3979 | ||
f885b7f2 PM |
3980 | for (i = 0; i < rcu_num_lvls; i++) |
3981 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
3982 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
3983 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
3984 | rcu_init_levelspread(rsp); | |
4a81e832 PM |
3985 | rsp->flavor_mask = fl_mask; |
3986 | fl_mask <<= 1; | |
64db4cff PM |
3987 | |
3988 | /* Initialize the elements themselves, starting from the leaves. */ | |
3989 | ||
f885b7f2 | 3990 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3991 | cpustride *= rsp->levelspread[i]; |
3992 | rnp = rsp->level[i]; | |
3993 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 3994 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
3995 | lockdep_set_class_and_name(&rnp->lock, |
3996 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
3997 | raw_spin_lock_init(&rnp->fqslock); |
3998 | lockdep_set_class_and_name(&rnp->fqslock, | |
3999 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
4000 | rnp->gpnum = rsp->gpnum; |
4001 | rnp->completed = rsp->completed; | |
64db4cff PM |
4002 | rnp->qsmask = 0; |
4003 | rnp->qsmaskinit = 0; | |
4004 | rnp->grplo = j * cpustride; | |
4005 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
4006 | if (rnp->grphi >= nr_cpu_ids) |
4007 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
4008 | if (i == 0) { |
4009 | rnp->grpnum = 0; | |
4010 | rnp->grpmask = 0; | |
4011 | rnp->parent = NULL; | |
4012 | } else { | |
4013 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
4014 | rnp->grpmask = 1UL << rnp->grpnum; | |
4015 | rnp->parent = rsp->level[i - 1] + | |
4016 | j / rsp->levelspread[i - 1]; | |
4017 | } | |
4018 | rnp->level = i; | |
12f5f524 | 4019 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 4020 | rcu_init_one_nocb(rnp); |
64db4cff PM |
4021 | } |
4022 | } | |
0c34029a | 4023 | |
b3dbec76 | 4024 | init_waitqueue_head(&rsp->gp_wq); |
f885b7f2 | 4025 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 4026 | for_each_possible_cpu(i) { |
4a90a068 | 4027 | while (i > rnp->grphi) |
0c34029a | 4028 | rnp++; |
394f99a9 | 4029 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
4030 | rcu_boot_init_percpu_data(i, rsp); |
4031 | } | |
6ce75a23 | 4032 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
4033 | } |
4034 | ||
f885b7f2 PM |
4035 | /* |
4036 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 4037 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
4038 | * the ->node array in the rcu_state structure. |
4039 | */ | |
4040 | static void __init rcu_init_geometry(void) | |
4041 | { | |
026ad283 | 4042 | ulong d; |
f885b7f2 PM |
4043 | int i; |
4044 | int j; | |
cca6f393 | 4045 | int n = nr_cpu_ids; |
f885b7f2 PM |
4046 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
4047 | ||
026ad283 PM |
4048 | /* |
4049 | * Initialize any unspecified boot parameters. | |
4050 | * The default values of jiffies_till_first_fqs and | |
4051 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
4052 | * value, which is a function of HZ, then adding one for each | |
4053 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
4054 | */ | |
4055 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
4056 | if (jiffies_till_first_fqs == ULONG_MAX) | |
4057 | jiffies_till_first_fqs = d; | |
4058 | if (jiffies_till_next_fqs == ULONG_MAX) | |
4059 | jiffies_till_next_fqs = d; | |
4060 | ||
f885b7f2 | 4061 | /* If the compile-time values are accurate, just leave. */ |
b17c7035 PM |
4062 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
4063 | nr_cpu_ids == NR_CPUS) | |
f885b7f2 | 4064 | return; |
39479098 PM |
4065 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", |
4066 | rcu_fanout_leaf, nr_cpu_ids); | |
f885b7f2 PM |
4067 | |
4068 | /* | |
4069 | * Compute number of nodes that can be handled an rcu_node tree | |
4070 | * with the given number of levels. Setting rcu_capacity[0] makes | |
4071 | * some of the arithmetic easier. | |
4072 | */ | |
4073 | rcu_capacity[0] = 1; | |
4074 | rcu_capacity[1] = rcu_fanout_leaf; | |
4075 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
4076 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
4077 | ||
4078 | /* | |
4079 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
4080 | * to increase the leaf-level fanout, not decrease it. Of course, | |
4081 | * the leaf-level fanout cannot exceed the number of bits in | |
4082 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
4083 | * the configured number of CPUs. Complain and fall back to the | |
4084 | * compile-time values if these limits are exceeded. | |
4085 | */ | |
4086 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
4087 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
4088 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
4089 | WARN_ON(1); | |
4090 | return; | |
4091 | } | |
4092 | ||
4093 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
4094 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
4095 | if (n <= rcu_capacity[i]) { | |
4096 | for (j = 0; j <= i; j++) | |
4097 | num_rcu_lvl[j] = | |
4098 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
4099 | rcu_num_lvls = i; | |
4100 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
4101 | num_rcu_lvl[j] = 0; | |
4102 | break; | |
4103 | } | |
4104 | ||
4105 | /* Calculate the total number of rcu_node structures. */ | |
4106 | rcu_num_nodes = 0; | |
4107 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
4108 | rcu_num_nodes += num_rcu_lvl[i]; | |
4109 | rcu_num_nodes -= n; | |
4110 | } | |
4111 | ||
9f680ab4 | 4112 | void __init rcu_init(void) |
64db4cff | 4113 | { |
017c4261 | 4114 | int cpu; |
9f680ab4 | 4115 | |
47627678 PM |
4116 | rcu_early_boot_tests(); |
4117 | ||
f41d911f | 4118 | rcu_bootup_announce(); |
f885b7f2 | 4119 | rcu_init_geometry(); |
394f99a9 | 4120 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); |
69c8d28c | 4121 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
f41d911f | 4122 | __rcu_init_preempt(); |
b5b39360 | 4123 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
4124 | |
4125 | /* | |
4126 | * We don't need protection against CPU-hotplug here because | |
4127 | * this is called early in boot, before either interrupts | |
4128 | * or the scheduler are operational. | |
4129 | */ | |
4130 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 4131 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
4132 | for_each_online_cpu(cpu) |
4133 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
4134 | } |
4135 | ||
4102adab | 4136 | #include "tree_plugin.h" |