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) \ | |
a41bfeb2 | 94 | struct rcu_state sname##_state = { \ |
6c90cc7b | 95 | .level = { &sname##_state.node[0] }, \ |
037b64ed | 96 | .call = cr, \ |
af446b70 | 97 | .fqs_state = RCU_GP_IDLE, \ |
42c3533e PM |
98 | .gpnum = 0UL - 300UL, \ |
99 | .completed = 0UL - 300UL, \ | |
7b2e6011 | 100 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
6c90cc7b PM |
101 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
102 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 103 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
a4fbe35a | 104 | .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \ |
a8a29b3b | 105 | .name = RCU_STATE_NAME(sname), \ |
a4889858 | 106 | .abbr = sabbr, \ |
a41bfeb2 | 107 | }; \ |
11bbb235 | 108 | DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data) |
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 | ||
5d01bbd1 | 155 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
156 | static void invoke_rcu_core(void); |
157 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 158 | |
4a298656 PM |
159 | /* |
160 | * Track the rcutorture test sequence number and the update version | |
161 | * number within a given test. The rcutorture_testseq is incremented | |
162 | * on every rcutorture module load and unload, so has an odd value | |
163 | * when a test is running. The rcutorture_vernum is set to zero | |
164 | * when rcutorture starts and is incremented on each rcutorture update. | |
165 | * These variables enable correlating rcutorture output with the | |
166 | * RCU tracing information. | |
167 | */ | |
168 | unsigned long rcutorture_testseq; | |
169 | unsigned long rcutorture_vernum; | |
170 | ||
fc2219d4 PM |
171 | /* |
172 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
173 | * permit this function to be invoked without holding the root rcu_node | |
174 | * structure's ->lock, but of course results can be subject to change. | |
175 | */ | |
176 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
177 | { | |
178 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
179 | } | |
180 | ||
b1f77b05 | 181 | /* |
d6714c22 | 182 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 183 | * how many quiescent states passed, just if there was at least |
d6714c22 | 184 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 185 | * The caller must have disabled preemption. |
b1f77b05 | 186 | */ |
284a8c93 | 187 | void rcu_sched_qs(void) |
b1f77b05 | 188 | { |
284a8c93 PM |
189 | if (!__this_cpu_read(rcu_sched_data.passed_quiesce)) { |
190 | trace_rcu_grace_period(TPS("rcu_sched"), | |
191 | __this_cpu_read(rcu_sched_data.gpnum), | |
192 | TPS("cpuqs")); | |
193 | __this_cpu_write(rcu_sched_data.passed_quiesce, 1); | |
194 | } | |
b1f77b05 IM |
195 | } |
196 | ||
284a8c93 | 197 | void rcu_bh_qs(void) |
b1f77b05 | 198 | { |
284a8c93 PM |
199 | if (!__this_cpu_read(rcu_bh_data.passed_quiesce)) { |
200 | trace_rcu_grace_period(TPS("rcu_bh"), | |
201 | __this_cpu_read(rcu_bh_data.gpnum), | |
202 | TPS("cpuqs")); | |
203 | __this_cpu_write(rcu_bh_data.passed_quiesce, 1); | |
204 | } | |
b1f77b05 | 205 | } |
64db4cff | 206 | |
4a81e832 PM |
207 | static DEFINE_PER_CPU(int, rcu_sched_qs_mask); |
208 | ||
209 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { | |
210 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, | |
211 | .dynticks = ATOMIC_INIT(1), | |
212 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
213 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
214 | .dynticks_idle = ATOMIC_INIT(1), | |
215 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
216 | }; | |
217 | ||
218 | /* | |
219 | * Let the RCU core know that this CPU has gone through the scheduler, | |
220 | * which is a quiescent state. This is called when the need for a | |
221 | * quiescent state is urgent, so we burn an atomic operation and full | |
222 | * memory barriers to let the RCU core know about it, regardless of what | |
223 | * this CPU might (or might not) do in the near future. | |
224 | * | |
225 | * We inform the RCU core by emulating a zero-duration dyntick-idle | |
226 | * period, which we in turn do by incrementing the ->dynticks counter | |
227 | * by two. | |
228 | */ | |
229 | static void rcu_momentary_dyntick_idle(void) | |
230 | { | |
231 | unsigned long flags; | |
232 | struct rcu_data *rdp; | |
233 | struct rcu_dynticks *rdtp; | |
234 | int resched_mask; | |
235 | struct rcu_state *rsp; | |
236 | ||
237 | local_irq_save(flags); | |
238 | ||
239 | /* | |
240 | * Yes, we can lose flag-setting operations. This is OK, because | |
241 | * the flag will be set again after some delay. | |
242 | */ | |
243 | resched_mask = raw_cpu_read(rcu_sched_qs_mask); | |
244 | raw_cpu_write(rcu_sched_qs_mask, 0); | |
245 | ||
246 | /* Find the flavor that needs a quiescent state. */ | |
247 | for_each_rcu_flavor(rsp) { | |
248 | rdp = raw_cpu_ptr(rsp->rda); | |
249 | if (!(resched_mask & rsp->flavor_mask)) | |
250 | continue; | |
251 | smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */ | |
252 | if (ACCESS_ONCE(rdp->mynode->completed) != | |
253 | ACCESS_ONCE(rdp->cond_resched_completed)) | |
254 | continue; | |
255 | ||
256 | /* | |
257 | * Pretend to be momentarily idle for the quiescent state. | |
258 | * This allows the grace-period kthread to record the | |
259 | * quiescent state, with no need for this CPU to do anything | |
260 | * further. | |
261 | */ | |
262 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
263 | smp_mb__before_atomic(); /* Earlier stuff before QS. */ | |
264 | atomic_add(2, &rdtp->dynticks); /* QS. */ | |
265 | smp_mb__after_atomic(); /* Later stuff after QS. */ | |
266 | break; | |
267 | } | |
268 | local_irq_restore(flags); | |
269 | } | |
270 | ||
25502a6c PM |
271 | /* |
272 | * Note a context switch. This is a quiescent state for RCU-sched, | |
273 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 274 | * The caller must have disabled preemption. |
25502a6c | 275 | */ |
38200cf2 | 276 | void rcu_note_context_switch(void) |
25502a6c | 277 | { |
f7f7bac9 | 278 | trace_rcu_utilization(TPS("Start context switch")); |
284a8c93 | 279 | rcu_sched_qs(); |
38200cf2 | 280 | rcu_preempt_note_context_switch(); |
4a81e832 PM |
281 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) |
282 | rcu_momentary_dyntick_idle(); | |
f7f7bac9 | 283 | trace_rcu_utilization(TPS("End context switch")); |
25502a6c | 284 | } |
29ce8310 | 285 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 286 | |
878d7439 ED |
287 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
288 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
289 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 290 | |
878d7439 ED |
291 | module_param(blimit, long, 0444); |
292 | module_param(qhimark, long, 0444); | |
293 | module_param(qlowmark, long, 0444); | |
3d76c082 | 294 | |
026ad283 PM |
295 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
296 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
297 | |
298 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
299 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
300 | ||
4a81e832 PM |
301 | /* |
302 | * How long the grace period must be before we start recruiting | |
303 | * quiescent-state help from rcu_note_context_switch(). | |
304 | */ | |
305 | static ulong jiffies_till_sched_qs = HZ / 20; | |
306 | module_param(jiffies_till_sched_qs, ulong, 0644); | |
307 | ||
48a7639c | 308 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
910ee45d | 309 | struct rcu_data *rdp); |
217af2a2 PM |
310 | static void force_qs_rnp(struct rcu_state *rsp, |
311 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
312 | unsigned long *maxj), | |
313 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 314 | static void force_quiescent_state(struct rcu_state *rsp); |
e3950ecd | 315 | static int rcu_pending(void); |
64db4cff PM |
316 | |
317 | /* | |
d6714c22 | 318 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
64db4cff | 319 | */ |
d6714c22 | 320 | long rcu_batches_completed_sched(void) |
64db4cff | 321 | { |
d6714c22 | 322 | return rcu_sched_state.completed; |
64db4cff | 323 | } |
d6714c22 | 324 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
325 | |
326 | /* | |
327 | * Return the number of RCU BH batches processed thus far for debug & stats. | |
328 | */ | |
329 | long rcu_batches_completed_bh(void) | |
330 | { | |
331 | return rcu_bh_state.completed; | |
332 | } | |
333 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
334 | ||
a381d757 ACB |
335 | /* |
336 | * Force a quiescent state. | |
337 | */ | |
338 | void rcu_force_quiescent_state(void) | |
339 | { | |
e534165b | 340 | force_quiescent_state(rcu_state_p); |
a381d757 ACB |
341 | } |
342 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
343 | ||
bf66f18e PM |
344 | /* |
345 | * Force a quiescent state for RCU BH. | |
346 | */ | |
347 | void rcu_bh_force_quiescent_state(void) | |
348 | { | |
4cdfc175 | 349 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
350 | } |
351 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
352 | ||
afea227f PM |
353 | /* |
354 | * Show the state of the grace-period kthreads. | |
355 | */ | |
356 | void show_rcu_gp_kthreads(void) | |
357 | { | |
358 | struct rcu_state *rsp; | |
359 | ||
360 | for_each_rcu_flavor(rsp) { | |
361 | pr_info("%s: wait state: %d ->state: %#lx\n", | |
362 | rsp->name, rsp->gp_state, rsp->gp_kthread->state); | |
363 | /* sched_show_task(rsp->gp_kthread); */ | |
364 | } | |
365 | } | |
366 | EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); | |
367 | ||
4a298656 PM |
368 | /* |
369 | * Record the number of times rcutorture tests have been initiated and | |
370 | * terminated. This information allows the debugfs tracing stats to be | |
371 | * correlated to the rcutorture messages, even when the rcutorture module | |
372 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
373 | * store this state in rcutorture itself. | |
374 | */ | |
375 | void rcutorture_record_test_transition(void) | |
376 | { | |
377 | rcutorture_testseq++; | |
378 | rcutorture_vernum = 0; | |
379 | } | |
380 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
381 | ||
ad0dc7f9 PM |
382 | /* |
383 | * Send along grace-period-related data for rcutorture diagnostics. | |
384 | */ | |
385 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
386 | unsigned long *gpnum, unsigned long *completed) | |
387 | { | |
388 | struct rcu_state *rsp = NULL; | |
389 | ||
390 | switch (test_type) { | |
391 | case RCU_FLAVOR: | |
e534165b | 392 | rsp = rcu_state_p; |
ad0dc7f9 PM |
393 | break; |
394 | case RCU_BH_FLAVOR: | |
395 | rsp = &rcu_bh_state; | |
396 | break; | |
397 | case RCU_SCHED_FLAVOR: | |
398 | rsp = &rcu_sched_state; | |
399 | break; | |
400 | default: | |
401 | break; | |
402 | } | |
403 | if (rsp != NULL) { | |
404 | *flags = ACCESS_ONCE(rsp->gp_flags); | |
405 | *gpnum = ACCESS_ONCE(rsp->gpnum); | |
406 | *completed = ACCESS_ONCE(rsp->completed); | |
407 | return; | |
408 | } | |
409 | *flags = 0; | |
410 | *gpnum = 0; | |
411 | *completed = 0; | |
412 | } | |
413 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
414 | ||
4a298656 PM |
415 | /* |
416 | * Record the number of writer passes through the current rcutorture test. | |
417 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
418 | * messages. | |
419 | */ | |
420 | void rcutorture_record_progress(unsigned long vernum) | |
421 | { | |
422 | rcutorture_vernum++; | |
423 | } | |
424 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
425 | ||
bf66f18e PM |
426 | /* |
427 | * Force a quiescent state for RCU-sched. | |
428 | */ | |
429 | void rcu_sched_force_quiescent_state(void) | |
430 | { | |
4cdfc175 | 431 | force_quiescent_state(&rcu_sched_state); |
bf66f18e PM |
432 | } |
433 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
434 | ||
64db4cff PM |
435 | /* |
436 | * Does the CPU have callbacks ready to be invoked? | |
437 | */ | |
438 | static int | |
439 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
440 | { | |
3fbfbf7a PM |
441 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
442 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
443 | } |
444 | ||
365187fb PM |
445 | /* |
446 | * Return the root node of the specified rcu_state structure. | |
447 | */ | |
448 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
449 | { | |
450 | return &rsp->node[0]; | |
451 | } | |
452 | ||
453 | /* | |
454 | * Is there any need for future grace periods? | |
455 | * Interrupts must be disabled. If the caller does not hold the root | |
456 | * rnp_node structure's ->lock, the results are advisory only. | |
457 | */ | |
458 | static int rcu_future_needs_gp(struct rcu_state *rsp) | |
459 | { | |
460 | struct rcu_node *rnp = rcu_get_root(rsp); | |
461 | int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1; | |
462 | int *fp = &rnp->need_future_gp[idx]; | |
463 | ||
464 | return ACCESS_ONCE(*fp); | |
465 | } | |
466 | ||
64db4cff | 467 | /* |
dc35c893 PM |
468 | * Does the current CPU require a not-yet-started grace period? |
469 | * The caller must have disabled interrupts to prevent races with | |
470 | * normal callback registry. | |
64db4cff PM |
471 | */ |
472 | static int | |
473 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
474 | { | |
dc35c893 | 475 | int i; |
3fbfbf7a | 476 | |
dc35c893 PM |
477 | if (rcu_gp_in_progress(rsp)) |
478 | return 0; /* No, a grace period is already in progress. */ | |
365187fb | 479 | if (rcu_future_needs_gp(rsp)) |
34ed6246 | 480 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
481 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
482 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
483 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
484 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
485 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
486 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
487 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), | |
488 | rdp->nxtcompleted[i])) | |
489 | return 1; /* Yes, CBs for future grace period. */ | |
490 | return 0; /* No grace period needed. */ | |
64db4cff PM |
491 | } |
492 | ||
9b2e4f18 | 493 | /* |
adf5091e | 494 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
495 | * |
496 | * If the new value of the ->dynticks_nesting counter now is zero, | |
497 | * we really have entered idle, and must do the appropriate accounting. | |
498 | * The caller must have disabled interrupts. | |
499 | */ | |
28ced795 | 500 | static void rcu_eqs_enter_common(long long oldval, bool user) |
9b2e4f18 | 501 | { |
96d3fd0d PM |
502 | struct rcu_state *rsp; |
503 | struct rcu_data *rdp; | |
28ced795 | 504 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
96d3fd0d | 505 | |
f7f7bac9 | 506 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 507 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
508 | struct task_struct *idle __maybe_unused = |
509 | idle_task(smp_processor_id()); | |
0989cb46 | 510 | |
f7f7bac9 | 511 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 512 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
513 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
514 | current->pid, current->comm, | |
515 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 516 | } |
96d3fd0d PM |
517 | for_each_rcu_flavor(rsp) { |
518 | rdp = this_cpu_ptr(rsp->rda); | |
519 | do_nocb_deferred_wakeup(rdp); | |
520 | } | |
198bbf81 | 521 | rcu_prepare_for_idle(); |
9b2e4f18 | 522 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 523 | smp_mb__before_atomic(); /* See above. */ |
9b2e4f18 | 524 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 525 | smp_mb__after_atomic(); /* Force ordering with next sojourn. */ |
9b2e4f18 | 526 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
176f8f7a | 527 | rcu_dynticks_task_enter(); |
c44e2cdd PM |
528 | |
529 | /* | |
adf5091e | 530 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
531 | * in an RCU read-side critical section. |
532 | */ | |
533 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
534 | "Illegal idle entry in RCU read-side critical section."); | |
535 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
536 | "Illegal idle entry in RCU-bh read-side critical section."); | |
537 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
538 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 539 | } |
64db4cff | 540 | |
adf5091e FW |
541 | /* |
542 | * Enter an RCU extended quiescent state, which can be either the | |
543 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 544 | */ |
adf5091e | 545 | static void rcu_eqs_enter(bool user) |
64db4cff | 546 | { |
4145fa7f | 547 | long long oldval; |
64db4cff PM |
548 | struct rcu_dynticks *rdtp; |
549 | ||
c9d4b0af | 550 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 551 | oldval = rdtp->dynticks_nesting; |
29e37d81 | 552 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
3a592405 | 553 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { |
29e37d81 | 554 | rdtp->dynticks_nesting = 0; |
28ced795 | 555 | rcu_eqs_enter_common(oldval, user); |
3a592405 | 556 | } else { |
29e37d81 | 557 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; |
3a592405 | 558 | } |
64db4cff | 559 | } |
adf5091e FW |
560 | |
561 | /** | |
562 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
563 | * | |
564 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
565 | * read-side critical sections can occur. (Though RCU read-side | |
566 | * critical sections can occur in irq handlers in idle, a possibility | |
567 | * handled by irq_enter() and irq_exit().) | |
568 | * | |
569 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
570 | * the possibility of usermode upcalls having messed up our count | |
571 | * of interrupt nesting level during the prior busy period. | |
572 | */ | |
573 | void rcu_idle_enter(void) | |
574 | { | |
c5d900bf FW |
575 | unsigned long flags; |
576 | ||
577 | local_irq_save(flags); | |
cb349ca9 | 578 | rcu_eqs_enter(false); |
28ced795 | 579 | rcu_sysidle_enter(0); |
c5d900bf | 580 | local_irq_restore(flags); |
adf5091e | 581 | } |
8a2ecf47 | 582 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 583 | |
2b1d5024 | 584 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
585 | /** |
586 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
587 | * | |
588 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
589 | * is permitted between this call and rcu_user_exit(). This way the | |
590 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
591 | * when the CPU runs in userspace. | |
592 | */ | |
593 | void rcu_user_enter(void) | |
594 | { | |
91d1aa43 | 595 | rcu_eqs_enter(1); |
adf5091e | 596 | } |
2b1d5024 | 597 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 598 | |
9b2e4f18 PM |
599 | /** |
600 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
601 | * | |
602 | * Exit from an interrupt handler, which might possibly result in entering | |
603 | * idle mode, in other words, leaving the mode in which read-side critical | |
604 | * sections can occur. | |
64db4cff | 605 | * |
9b2e4f18 PM |
606 | * This code assumes that the idle loop never does anything that might |
607 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
608 | * architecture violates this assumption, RCU will give you what you | |
609 | * deserve, good and hard. But very infrequently and irreproducibly. | |
610 | * | |
611 | * Use things like work queues to work around this limitation. | |
612 | * | |
613 | * You have been warned. | |
64db4cff | 614 | */ |
9b2e4f18 | 615 | void rcu_irq_exit(void) |
64db4cff PM |
616 | { |
617 | unsigned long flags; | |
4145fa7f | 618 | long long oldval; |
64db4cff PM |
619 | struct rcu_dynticks *rdtp; |
620 | ||
621 | local_irq_save(flags); | |
c9d4b0af | 622 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 623 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
624 | rdtp->dynticks_nesting--; |
625 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 | 626 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 627 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 628 | else |
28ced795 CL |
629 | rcu_eqs_enter_common(oldval, true); |
630 | rcu_sysidle_enter(1); | |
9b2e4f18 PM |
631 | local_irq_restore(flags); |
632 | } | |
633 | ||
634 | /* | |
adf5091e | 635 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
636 | * |
637 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
638 | * we really have exited idle, and must do the appropriate accounting. | |
639 | * The caller must have disabled interrupts. | |
640 | */ | |
28ced795 | 641 | static void rcu_eqs_exit_common(long long oldval, int user) |
9b2e4f18 | 642 | { |
28ced795 CL |
643 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
644 | ||
176f8f7a | 645 | rcu_dynticks_task_exit(); |
4e857c58 | 646 | smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */ |
23b5c8fa PM |
647 | atomic_inc(&rdtp->dynticks); |
648 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
4e857c58 | 649 | smp_mb__after_atomic(); /* See above. */ |
23b5c8fa | 650 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); |
8fa7845d | 651 | rcu_cleanup_after_idle(); |
f7f7bac9 | 652 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 653 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
654 | struct task_struct *idle __maybe_unused = |
655 | idle_task(smp_processor_id()); | |
0989cb46 | 656 | |
f7f7bac9 | 657 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 658 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 659 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
660 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
661 | current->pid, current->comm, | |
662 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
663 | } |
664 | } | |
665 | ||
adf5091e FW |
666 | /* |
667 | * Exit an RCU extended quiescent state, which can be either the | |
668 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 669 | */ |
adf5091e | 670 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 671 | { |
9b2e4f18 PM |
672 | struct rcu_dynticks *rdtp; |
673 | long long oldval; | |
674 | ||
c9d4b0af | 675 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 | 676 | oldval = rdtp->dynticks_nesting; |
29e37d81 | 677 | WARN_ON_ONCE(oldval < 0); |
3a592405 | 678 | if (oldval & DYNTICK_TASK_NEST_MASK) { |
29e37d81 | 679 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; |
3a592405 | 680 | } else { |
29e37d81 | 681 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
28ced795 | 682 | rcu_eqs_exit_common(oldval, user); |
3a592405 | 683 | } |
9b2e4f18 | 684 | } |
adf5091e FW |
685 | |
686 | /** | |
687 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
688 | * | |
689 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
690 | * read-side critical sections can occur. | |
691 | * | |
692 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
693 | * allow for the possibility of usermode upcalls messing up our count | |
694 | * of interrupt nesting level during the busy period that is just | |
695 | * now starting. | |
696 | */ | |
697 | void rcu_idle_exit(void) | |
698 | { | |
c5d900bf FW |
699 | unsigned long flags; |
700 | ||
701 | local_irq_save(flags); | |
cb349ca9 | 702 | rcu_eqs_exit(false); |
28ced795 | 703 | rcu_sysidle_exit(0); |
c5d900bf | 704 | local_irq_restore(flags); |
adf5091e | 705 | } |
8a2ecf47 | 706 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 707 | |
2b1d5024 | 708 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
709 | /** |
710 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
711 | * | |
712 | * Exit RCU idle mode while entering the kernel because it can | |
713 | * run a RCU read side critical section anytime. | |
714 | */ | |
715 | void rcu_user_exit(void) | |
716 | { | |
91d1aa43 | 717 | rcu_eqs_exit(1); |
adf5091e | 718 | } |
2b1d5024 | 719 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 720 | |
9b2e4f18 PM |
721 | /** |
722 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
723 | * | |
724 | * Enter an interrupt handler, which might possibly result in exiting | |
725 | * idle mode, in other words, entering the mode in which read-side critical | |
726 | * sections can occur. | |
727 | * | |
728 | * Note that the Linux kernel is fully capable of entering an interrupt | |
729 | * handler that it never exits, for example when doing upcalls to | |
730 | * user mode! This code assumes that the idle loop never does upcalls to | |
731 | * user mode. If your architecture does do upcalls from the idle loop (or | |
732 | * does anything else that results in unbalanced calls to the irq_enter() | |
733 | * and irq_exit() functions), RCU will give you what you deserve, good | |
734 | * and hard. But very infrequently and irreproducibly. | |
735 | * | |
736 | * Use things like work queues to work around this limitation. | |
737 | * | |
738 | * You have been warned. | |
739 | */ | |
740 | void rcu_irq_enter(void) | |
741 | { | |
742 | unsigned long flags; | |
743 | struct rcu_dynticks *rdtp; | |
744 | long long oldval; | |
745 | ||
746 | local_irq_save(flags); | |
c9d4b0af | 747 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 PM |
748 | oldval = rdtp->dynticks_nesting; |
749 | rdtp->dynticks_nesting++; | |
750 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 | 751 | if (oldval) |
f7f7bac9 | 752 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 753 | else |
28ced795 CL |
754 | rcu_eqs_exit_common(oldval, true); |
755 | rcu_sysidle_exit(1); | |
64db4cff | 756 | local_irq_restore(flags); |
64db4cff PM |
757 | } |
758 | ||
759 | /** | |
760 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
761 | * | |
734d1680 PM |
762 | * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and |
763 | * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know | |
764 | * that the CPU is active. This implementation permits nested NMIs, as | |
765 | * long as the nesting level does not overflow an int. (You will probably | |
766 | * run out of stack space first.) | |
64db4cff PM |
767 | */ |
768 | void rcu_nmi_enter(void) | |
769 | { | |
c9d4b0af | 770 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
734d1680 | 771 | int incby = 2; |
64db4cff | 772 | |
734d1680 PM |
773 | /* Complain about underflow. */ |
774 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0); | |
775 | ||
776 | /* | |
777 | * If idle from RCU viewpoint, atomically increment ->dynticks | |
778 | * to mark non-idle and increment ->dynticks_nmi_nesting by one. | |
779 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means | |
780 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed | |
781 | * to be in the outermost NMI handler that interrupted an RCU-idle | |
782 | * period (observation due to Andy Lutomirski). | |
783 | */ | |
784 | if (!(atomic_read(&rdtp->dynticks) & 0x1)) { | |
785 | smp_mb__before_atomic(); /* Force delay from prior write. */ | |
786 | atomic_inc(&rdtp->dynticks); | |
787 | /* atomic_inc() before later RCU read-side crit sects */ | |
788 | smp_mb__after_atomic(); /* See above. */ | |
789 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
790 | incby = 1; | |
791 | } | |
792 | rdtp->dynticks_nmi_nesting += incby; | |
793 | barrier(); | |
64db4cff PM |
794 | } |
795 | ||
796 | /** | |
797 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
798 | * | |
734d1680 PM |
799 | * If we are returning from the outermost NMI handler that interrupted an |
800 | * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting | |
801 | * to let the RCU grace-period handling know that the CPU is back to | |
802 | * being RCU-idle. | |
64db4cff PM |
803 | */ |
804 | void rcu_nmi_exit(void) | |
805 | { | |
c9d4b0af | 806 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 807 | |
734d1680 PM |
808 | /* |
809 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. | |
810 | * (We are exiting an NMI handler, so RCU better be paying attention | |
811 | * to us!) | |
812 | */ | |
813 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0); | |
814 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
815 | ||
816 | /* | |
817 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so | |
818 | * leave it in non-RCU-idle state. | |
819 | */ | |
820 | if (rdtp->dynticks_nmi_nesting != 1) { | |
821 | rdtp->dynticks_nmi_nesting -= 2; | |
64db4cff | 822 | return; |
734d1680 PM |
823 | } |
824 | ||
825 | /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ | |
826 | rdtp->dynticks_nmi_nesting = 0; | |
23b5c8fa | 827 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 828 | smp_mb__before_atomic(); /* See above. */ |
23b5c8fa | 829 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 830 | smp_mb__after_atomic(); /* Force delay to next write. */ |
23b5c8fa | 831 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
64db4cff PM |
832 | } |
833 | ||
834 | /** | |
5c173eb8 PM |
835 | * __rcu_is_watching - are RCU read-side critical sections safe? |
836 | * | |
837 | * Return true if RCU is watching the running CPU, which means that | |
838 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
839 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
840 | * least disabled preemption. | |
841 | */ | |
9418fb20 | 842 | bool notrace __rcu_is_watching(void) |
5c173eb8 PM |
843 | { |
844 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
845 | } | |
846 | ||
847 | /** | |
848 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
64db4cff | 849 | * |
9b2e4f18 | 850 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 851 | * or NMI handler, return true. |
64db4cff | 852 | */ |
9418fb20 | 853 | bool notrace rcu_is_watching(void) |
64db4cff | 854 | { |
f534ed1f | 855 | bool ret; |
34240697 PM |
856 | |
857 | preempt_disable(); | |
5c173eb8 | 858 | ret = __rcu_is_watching(); |
34240697 PM |
859 | preempt_enable(); |
860 | return ret; | |
64db4cff | 861 | } |
5c173eb8 | 862 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 863 | |
62fde6ed | 864 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
865 | |
866 | /* | |
867 | * Is the current CPU online? Disable preemption to avoid false positives | |
868 | * that could otherwise happen due to the current CPU number being sampled, | |
869 | * this task being preempted, its old CPU being taken offline, resuming | |
870 | * on some other CPU, then determining that its old CPU is now offline. | |
871 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
872 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
873 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
874 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
875 | * offline to continue to use RCU for one jiffy after marking itself | |
876 | * offline in the cpu_online_mask. This leniency is necessary given the | |
877 | * non-atomic nature of the online and offline processing, for example, | |
878 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
879 | * notifiers. | |
880 | * | |
881 | * This is also why RCU internally marks CPUs online during the | |
882 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
883 | * |
884 | * Disable checking if in an NMI handler because we cannot safely report | |
885 | * errors from NMI handlers anyway. | |
886 | */ | |
887 | bool rcu_lockdep_current_cpu_online(void) | |
888 | { | |
2036d94a PM |
889 | struct rcu_data *rdp; |
890 | struct rcu_node *rnp; | |
c0d6d01b PM |
891 | bool ret; |
892 | ||
893 | if (in_nmi()) | |
f6f7ee9a | 894 | return true; |
c0d6d01b | 895 | preempt_disable(); |
c9d4b0af | 896 | rdp = this_cpu_ptr(&rcu_sched_data); |
2036d94a PM |
897 | rnp = rdp->mynode; |
898 | ret = (rdp->grpmask & rnp->qsmaskinit) || | |
c0d6d01b PM |
899 | !rcu_scheduler_fully_active; |
900 | preempt_enable(); | |
901 | return ret; | |
902 | } | |
903 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
904 | ||
62fde6ed | 905 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 906 | |
64db4cff | 907 | /** |
9b2e4f18 | 908 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 909 | * |
9b2e4f18 PM |
910 | * If the current CPU is idle or running at a first-level (not nested) |
911 | * interrupt from idle, return true. The caller must have at least | |
912 | * disabled preemption. | |
64db4cff | 913 | */ |
62e3cb14 | 914 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 915 | { |
c9d4b0af | 916 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
64db4cff PM |
917 | } |
918 | ||
64db4cff PM |
919 | /* |
920 | * Snapshot the specified CPU's dynticks counter so that we can later | |
921 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 922 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 923 | */ |
217af2a2 PM |
924 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
925 | bool *isidle, unsigned long *maxj) | |
64db4cff | 926 | { |
23b5c8fa | 927 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 928 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
7941dbde ACB |
929 | if ((rdp->dynticks_snap & 0x1) == 0) { |
930 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); | |
931 | return 1; | |
932 | } else { | |
933 | return 0; | |
934 | } | |
64db4cff PM |
935 | } |
936 | ||
6193c76a PM |
937 | /* |
938 | * This function really isn't for public consumption, but RCU is special in | |
939 | * that context switches can allow the state machine to make progress. | |
940 | */ | |
941 | extern void resched_cpu(int cpu); | |
942 | ||
64db4cff PM |
943 | /* |
944 | * Return true if the specified CPU has passed through a quiescent | |
945 | * state by virtue of being in or having passed through an dynticks | |
946 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 947 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 948 | */ |
217af2a2 PM |
949 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
950 | bool *isidle, unsigned long *maxj) | |
64db4cff | 951 | { |
7eb4f455 | 952 | unsigned int curr; |
4a81e832 | 953 | int *rcrmp; |
7eb4f455 | 954 | unsigned int snap; |
64db4cff | 955 | |
7eb4f455 PM |
956 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
957 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
958 | |
959 | /* | |
960 | * If the CPU passed through or entered a dynticks idle phase with | |
961 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
962 | * already acknowledged the request to pass through a quiescent | |
963 | * state. Either way, that CPU cannot possibly be in an RCU | |
964 | * read-side critical section that started before the beginning | |
965 | * of the current RCU grace period. | |
966 | */ | |
7eb4f455 | 967 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 968 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
969 | rdp->dynticks_fqs++; |
970 | return 1; | |
971 | } | |
972 | ||
a82dcc76 PM |
973 | /* |
974 | * Check for the CPU being offline, but only if the grace period | |
975 | * is old enough. We don't need to worry about the CPU changing | |
976 | * state: If we see it offline even once, it has been through a | |
977 | * quiescent state. | |
978 | * | |
979 | * The reason for insisting that the grace period be at least | |
980 | * one jiffy old is that CPUs that are not quite online and that | |
981 | * have just gone offline can still execute RCU read-side critical | |
982 | * sections. | |
983 | */ | |
984 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
985 | return 0; /* Grace period is not old enough. */ | |
986 | barrier(); | |
987 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 988 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
989 | rdp->offline_fqs++; |
990 | return 1; | |
991 | } | |
65d798f0 PM |
992 | |
993 | /* | |
4a81e832 PM |
994 | * A CPU running for an extended time within the kernel can |
995 | * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode, | |
996 | * even context-switching back and forth between a pair of | |
997 | * in-kernel CPU-bound tasks cannot advance grace periods. | |
998 | * So if the grace period is old enough, make the CPU pay attention. | |
999 | * Note that the unsynchronized assignments to the per-CPU | |
1000 | * rcu_sched_qs_mask variable are safe. Yes, setting of | |
1001 | * bits can be lost, but they will be set again on the next | |
1002 | * force-quiescent-state pass. So lost bit sets do not result | |
1003 | * in incorrect behavior, merely in a grace period lasting | |
1004 | * a few jiffies longer than it might otherwise. Because | |
1005 | * there are at most four threads involved, and because the | |
1006 | * updates are only once every few jiffies, the probability of | |
1007 | * lossage (and thus of slight grace-period extension) is | |
1008 | * quite low. | |
1009 | * | |
1010 | * Note that if the jiffies_till_sched_qs boot/sysfs parameter | |
1011 | * is set too high, we override with half of the RCU CPU stall | |
1012 | * warning delay. | |
6193c76a | 1013 | */ |
4a81e832 PM |
1014 | rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu); |
1015 | if (ULONG_CMP_GE(jiffies, | |
1016 | rdp->rsp->gp_start + jiffies_till_sched_qs) || | |
cb1e78cf | 1017 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { |
4a81e832 PM |
1018 | if (!(ACCESS_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) { |
1019 | ACCESS_ONCE(rdp->cond_resched_completed) = | |
1020 | ACCESS_ONCE(rdp->mynode->completed); | |
1021 | smp_mb(); /* ->cond_resched_completed before *rcrmp. */ | |
1022 | ACCESS_ONCE(*rcrmp) = | |
1023 | ACCESS_ONCE(*rcrmp) + rdp->rsp->flavor_mask; | |
1024 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1025 | rdp->rsp->jiffies_resched += 5; /* Enable beating. */ | |
1026 | } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { | |
1027 | /* Time to beat on that CPU again! */ | |
1028 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1029 | rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */ | |
1030 | } | |
6193c76a PM |
1031 | } |
1032 | ||
a82dcc76 | 1033 | return 0; |
64db4cff PM |
1034 | } |
1035 | ||
64db4cff PM |
1036 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
1037 | { | |
cb1e78cf | 1038 | unsigned long j = jiffies; |
6193c76a | 1039 | unsigned long j1; |
26cdfedf PM |
1040 | |
1041 | rsp->gp_start = j; | |
1042 | smp_wmb(); /* Record start time before stall time. */ | |
6193c76a | 1043 | j1 = rcu_jiffies_till_stall_check(); |
4fc5b755 | 1044 | ACCESS_ONCE(rsp->jiffies_stall) = j + j1; |
6193c76a | 1045 | rsp->jiffies_resched = j + j1 / 2; |
fc908ed3 | 1046 | rsp->n_force_qs_gpstart = ACCESS_ONCE(rsp->n_force_qs); |
64db4cff PM |
1047 | } |
1048 | ||
b637a328 | 1049 | /* |
bc1dce51 | 1050 | * Dump stacks of all tasks running on stalled CPUs. |
b637a328 PM |
1051 | */ |
1052 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
1053 | { | |
1054 | int cpu; | |
1055 | unsigned long flags; | |
1056 | struct rcu_node *rnp; | |
1057 | ||
1058 | rcu_for_each_leaf_node(rsp, rnp) { | |
1059 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1060 | if (rnp->qsmask != 0) { | |
1061 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1062 | if (rnp->qsmask & (1UL << cpu)) | |
1063 | dump_cpu_task(rnp->grplo + cpu); | |
1064 | } | |
1065 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1066 | } | |
1067 | } | |
1068 | ||
64db4cff PM |
1069 | static void print_other_cpu_stall(struct rcu_state *rsp) |
1070 | { | |
1071 | int cpu; | |
1072 | long delta; | |
1073 | unsigned long flags; | |
285fe294 | 1074 | int ndetected = 0; |
64db4cff | 1075 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 1076 | long totqlen = 0; |
64db4cff PM |
1077 | |
1078 | /* Only let one CPU complain about others per time interval. */ | |
1079 | ||
1304afb2 | 1080 | raw_spin_lock_irqsave(&rnp->lock, flags); |
4fc5b755 | 1081 | delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall); |
fc2219d4 | 1082 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 1083 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1084 | return; |
1085 | } | |
4fc5b755 | 1086 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 1087 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1088 | |
8cdd32a9 PM |
1089 | /* |
1090 | * OK, time to rat on our buddy... | |
1091 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1092 | * RCU CPU stall warnings. | |
1093 | */ | |
d7f3e207 | 1094 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 1095 | rsp->name); |
a858af28 | 1096 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 1097 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 1098 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 1099 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
1100 | if (rnp->qsmask != 0) { |
1101 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1102 | if (rnp->qsmask & (1UL << cpu)) { | |
1103 | print_cpu_stall_info(rsp, | |
1104 | rnp->grplo + cpu); | |
1105 | ndetected++; | |
1106 | } | |
1107 | } | |
3acd9eb3 | 1108 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1109 | } |
a858af28 PM |
1110 | |
1111 | /* | |
1112 | * Now rat on any tasks that got kicked up to the root rcu_node | |
1113 | * due to CPU offlining. | |
1114 | */ | |
1115 | rnp = rcu_get_root(rsp); | |
1116 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
285fe294 | 1117 | ndetected += rcu_print_task_stall(rnp); |
a858af28 PM |
1118 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1119 | ||
1120 | print_cpu_stall_info_end(); | |
53bb857c PM |
1121 | for_each_possible_cpu(cpu) |
1122 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e | 1123 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", |
eee05882 | 1124 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
83ebe63e | 1125 | (long)rsp->gpnum, (long)rsp->completed, totqlen); |
9bc8b558 | 1126 | if (ndetected == 0) |
d7f3e207 | 1127 | pr_err("INFO: Stall ended before state dump start\n"); |
bc1dce51 | 1128 | else |
b637a328 | 1129 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 1130 | |
4cdfc175 | 1131 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
1132 | |
1133 | rcu_print_detail_task_stall(rsp); | |
1134 | ||
4cdfc175 | 1135 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
1136 | } |
1137 | ||
1138 | static void print_cpu_stall(struct rcu_state *rsp) | |
1139 | { | |
53bb857c | 1140 | int cpu; |
64db4cff PM |
1141 | unsigned long flags; |
1142 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 1143 | long totqlen = 0; |
64db4cff | 1144 | |
8cdd32a9 PM |
1145 | /* |
1146 | * OK, time to rat on ourselves... | |
1147 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1148 | * RCU CPU stall warnings. | |
1149 | */ | |
d7f3e207 | 1150 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
1151 | print_cpu_stall_info_begin(); |
1152 | print_cpu_stall_info(rsp, smp_processor_id()); | |
1153 | print_cpu_stall_info_end(); | |
53bb857c PM |
1154 | for_each_possible_cpu(cpu) |
1155 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e PM |
1156 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", |
1157 | jiffies - rsp->gp_start, | |
1158 | (long)rsp->gpnum, (long)rsp->completed, totqlen); | |
bc1dce51 | 1159 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 1160 | |
1304afb2 | 1161 | raw_spin_lock_irqsave(&rnp->lock, flags); |
4fc5b755 IM |
1162 | if (ULONG_CMP_GE(jiffies, ACCESS_ONCE(rsp->jiffies_stall))) |
1163 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + | |
6bfc09e2 | 1164 | 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 1165 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 1166 | |
b021fe3e PZ |
1167 | /* |
1168 | * Attempt to revive the RCU machinery by forcing a context switch. | |
1169 | * | |
1170 | * A context switch would normally allow the RCU state machine to make | |
1171 | * progress and it could be we're stuck in kernel space without context | |
1172 | * switches for an entirely unreasonable amount of time. | |
1173 | */ | |
1174 | resched_cpu(smp_processor_id()); | |
64db4cff PM |
1175 | } |
1176 | ||
1177 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
1178 | { | |
26cdfedf PM |
1179 | unsigned long completed; |
1180 | unsigned long gpnum; | |
1181 | unsigned long gps; | |
bad6e139 PM |
1182 | unsigned long j; |
1183 | unsigned long js; | |
64db4cff PM |
1184 | struct rcu_node *rnp; |
1185 | ||
26cdfedf | 1186 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
c68de209 | 1187 | return; |
cb1e78cf | 1188 | j = jiffies; |
26cdfedf PM |
1189 | |
1190 | /* | |
1191 | * Lots of memory barriers to reject false positives. | |
1192 | * | |
1193 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
1194 | * then rsp->gp_start, and finally rsp->completed. These values | |
1195 | * are updated in the opposite order with memory barriers (or | |
1196 | * equivalent) during grace-period initialization and cleanup. | |
1197 | * Now, a false positive can occur if we get an new value of | |
1198 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
1199 | * the memory barriers, the only way that this can happen is if one | |
1200 | * grace period ends and another starts between these two fetches. | |
1201 | * Detect this by comparing rsp->completed with the previous fetch | |
1202 | * from rsp->gpnum. | |
1203 | * | |
1204 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
1205 | * and rsp->gp_start suffice to forestall false positives. | |
1206 | */ | |
1207 | gpnum = ACCESS_ONCE(rsp->gpnum); | |
1208 | smp_rmb(); /* Pick up ->gpnum first... */ | |
bad6e139 | 1209 | js = ACCESS_ONCE(rsp->jiffies_stall); |
26cdfedf PM |
1210 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
1211 | gps = ACCESS_ONCE(rsp->gp_start); | |
1212 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ | |
1213 | completed = ACCESS_ONCE(rsp->completed); | |
1214 | if (ULONG_CMP_GE(completed, gpnum) || | |
1215 | ULONG_CMP_LT(j, js) || | |
1216 | ULONG_CMP_GE(gps, js)) | |
1217 | return; /* No stall or GP completed since entering function. */ | |
64db4cff | 1218 | rnp = rdp->mynode; |
c96ea7cf | 1219 | if (rcu_gp_in_progress(rsp) && |
26cdfedf | 1220 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) { |
64db4cff PM |
1221 | |
1222 | /* We haven't checked in, so go dump stack. */ | |
1223 | print_cpu_stall(rsp); | |
1224 | ||
bad6e139 PM |
1225 | } else if (rcu_gp_in_progress(rsp) && |
1226 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 1227 | |
bad6e139 | 1228 | /* They had a few time units to dump stack, so complain. */ |
64db4cff PM |
1229 | print_other_cpu_stall(rsp); |
1230 | } | |
1231 | } | |
1232 | ||
53d84e00 PM |
1233 | /** |
1234 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1235 | * | |
1236 | * Set the stall-warning timeout way off into the future, thus preventing | |
1237 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1238 | * RCU grace periods. | |
1239 | * | |
1240 | * The caller must disable hard irqs. | |
1241 | */ | |
1242 | void rcu_cpu_stall_reset(void) | |
1243 | { | |
6ce75a23 PM |
1244 | struct rcu_state *rsp; |
1245 | ||
1246 | for_each_rcu_flavor(rsp) | |
4fc5b755 | 1247 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + ULONG_MAX / 2; |
53d84e00 PM |
1248 | } |
1249 | ||
3f5d3ea6 PM |
1250 | /* |
1251 | * Initialize the specified rcu_data structure's callback list to empty. | |
1252 | */ | |
1253 | static void init_callback_list(struct rcu_data *rdp) | |
1254 | { | |
1255 | int i; | |
1256 | ||
34ed6246 PM |
1257 | if (init_nocb_callback_list(rdp)) |
1258 | return; | |
3f5d3ea6 PM |
1259 | rdp->nxtlist = NULL; |
1260 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1261 | rdp->nxttail[i] = &rdp->nxtlist; | |
1262 | } | |
1263 | ||
dc35c893 PM |
1264 | /* |
1265 | * Determine the value that ->completed will have at the end of the | |
1266 | * next subsequent grace period. This is used to tag callbacks so that | |
1267 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1268 | * been dyntick-idle for an extended period with callbacks under the | |
1269 | * influence of RCU_FAST_NO_HZ. | |
1270 | * | |
1271 | * The caller must hold rnp->lock with interrupts disabled. | |
1272 | */ | |
1273 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1274 | struct rcu_node *rnp) | |
1275 | { | |
1276 | /* | |
1277 | * If RCU is idle, we just wait for the next grace period. | |
1278 | * But we can only be sure that RCU is idle if we are looking | |
1279 | * at the root rcu_node structure -- otherwise, a new grace | |
1280 | * period might have started, but just not yet gotten around | |
1281 | * to initializing the current non-root rcu_node structure. | |
1282 | */ | |
1283 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1284 | return rnp->completed + 1; | |
1285 | ||
1286 | /* | |
1287 | * Otherwise, wait for a possible partial grace period and | |
1288 | * then the subsequent full grace period. | |
1289 | */ | |
1290 | return rnp->completed + 2; | |
1291 | } | |
1292 | ||
0446be48 PM |
1293 | /* |
1294 | * Trace-event helper function for rcu_start_future_gp() and | |
1295 | * rcu_nocb_wait_gp(). | |
1296 | */ | |
1297 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1298 | unsigned long c, const char *s) |
0446be48 PM |
1299 | { |
1300 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1301 | rnp->completed, c, rnp->level, | |
1302 | rnp->grplo, rnp->grphi, s); | |
1303 | } | |
1304 | ||
1305 | /* | |
1306 | * Start some future grace period, as needed to handle newly arrived | |
1307 | * callbacks. The required future grace periods are recorded in each | |
48a7639c PM |
1308 | * rcu_node structure's ->need_future_gp field. Returns true if there |
1309 | * is reason to awaken the grace-period kthread. | |
0446be48 PM |
1310 | * |
1311 | * The caller must hold the specified rcu_node structure's ->lock. | |
1312 | */ | |
48a7639c PM |
1313 | static bool __maybe_unused |
1314 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
1315 | unsigned long *c_out) | |
0446be48 PM |
1316 | { |
1317 | unsigned long c; | |
1318 | int i; | |
48a7639c | 1319 | bool ret = false; |
0446be48 PM |
1320 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); |
1321 | ||
1322 | /* | |
1323 | * Pick up grace-period number for new callbacks. If this | |
1324 | * grace period is already marked as needed, return to the caller. | |
1325 | */ | |
1326 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1327 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1328 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1329 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
48a7639c | 1330 | goto out; |
0446be48 PM |
1331 | } |
1332 | ||
1333 | /* | |
1334 | * If either this rcu_node structure or the root rcu_node structure | |
1335 | * believe that a grace period is in progress, then we must wait | |
1336 | * for the one following, which is in "c". Because our request | |
1337 | * will be noticed at the end of the current grace period, we don't | |
48bd8e9b PK |
1338 | * need to explicitly start one. We only do the lockless check |
1339 | * of rnp_root's fields if the current rcu_node structure thinks | |
1340 | * there is no grace period in flight, and because we hold rnp->lock, | |
1341 | * the only possible change is when rnp_root's two fields are | |
1342 | * equal, in which case rnp_root->gpnum might be concurrently | |
1343 | * incremented. But that is OK, as it will just result in our | |
1344 | * doing some extra useless work. | |
0446be48 PM |
1345 | */ |
1346 | if (rnp->gpnum != rnp->completed || | |
48bd8e9b | 1347 | ACCESS_ONCE(rnp_root->gpnum) != ACCESS_ONCE(rnp_root->completed)) { |
0446be48 | 1348 | rnp->need_future_gp[c & 0x1]++; |
f7f7bac9 | 1349 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
48a7639c | 1350 | goto out; |
0446be48 PM |
1351 | } |
1352 | ||
1353 | /* | |
1354 | * There might be no grace period in progress. If we don't already | |
1355 | * hold it, acquire the root rcu_node structure's lock in order to | |
1356 | * start one (if needed). | |
1357 | */ | |
6303b9c8 | 1358 | if (rnp != rnp_root) { |
0446be48 | 1359 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 PM |
1360 | smp_mb__after_unlock_lock(); |
1361 | } | |
0446be48 PM |
1362 | |
1363 | /* | |
1364 | * Get a new grace-period number. If there really is no grace | |
1365 | * period in progress, it will be smaller than the one we obtained | |
1366 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1367 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1368 | */ | |
1369 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1370 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1371 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1372 | rdp->nxtcompleted[i] = c; | |
1373 | ||
1374 | /* | |
1375 | * If the needed for the required grace period is already | |
1376 | * recorded, trace and leave. | |
1377 | */ | |
1378 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1379 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1380 | goto unlock_out; |
1381 | } | |
1382 | ||
1383 | /* Record the need for the future grace period. */ | |
1384 | rnp_root->need_future_gp[c & 0x1]++; | |
1385 | ||
1386 | /* If a grace period is not already in progress, start one. */ | |
1387 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1388 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1389 | } else { |
f7f7bac9 | 1390 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
48a7639c | 1391 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1392 | } |
1393 | unlock_out: | |
1394 | if (rnp != rnp_root) | |
1395 | raw_spin_unlock(&rnp_root->lock); | |
48a7639c PM |
1396 | out: |
1397 | if (c_out != NULL) | |
1398 | *c_out = c; | |
1399 | return ret; | |
0446be48 PM |
1400 | } |
1401 | ||
1402 | /* | |
1403 | * Clean up any old requests for the just-ended grace period. Also return | |
1404 | * whether any additional grace periods have been requested. Also invoke | |
1405 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1406 | * waiting for this grace period to complete. | |
1407 | */ | |
1408 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1409 | { | |
1410 | int c = rnp->completed; | |
1411 | int needmore; | |
1412 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1413 | ||
1414 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1415 | rnp->need_future_gp[c & 0x1] = 0; | |
1416 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1417 | trace_rcu_future_gp(rnp, rdp, c, |
1418 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1419 | return needmore; |
1420 | } | |
1421 | ||
48a7639c PM |
1422 | /* |
1423 | * Awaken the grace-period kthread for the specified flavor of RCU. | |
1424 | * Don't do a self-awaken, and don't bother awakening when there is | |
1425 | * nothing for the grace-period kthread to do (as in several CPUs | |
1426 | * raced to awaken, and we lost), and finally don't try to awaken | |
1427 | * a kthread that has not yet been created. | |
1428 | */ | |
1429 | static void rcu_gp_kthread_wake(struct rcu_state *rsp) | |
1430 | { | |
1431 | if (current == rsp->gp_kthread || | |
1432 | !ACCESS_ONCE(rsp->gp_flags) || | |
1433 | !rsp->gp_kthread) | |
1434 | return; | |
1435 | wake_up(&rsp->gp_wq); | |
1436 | } | |
1437 | ||
dc35c893 PM |
1438 | /* |
1439 | * If there is room, assign a ->completed number to any callbacks on | |
1440 | * this CPU that have not already been assigned. Also accelerate any | |
1441 | * callbacks that were previously assigned a ->completed number that has | |
1442 | * since proven to be too conservative, which can happen if callbacks get | |
1443 | * assigned a ->completed number while RCU is idle, but with reference to | |
1444 | * a non-root rcu_node structure. This function is idempotent, so it does | |
48a7639c PM |
1445 | * not hurt to call it repeatedly. Returns an flag saying that we should |
1446 | * awaken the RCU grace-period kthread. | |
dc35c893 PM |
1447 | * |
1448 | * The caller must hold rnp->lock with interrupts disabled. | |
1449 | */ | |
48a7639c | 1450 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1451 | struct rcu_data *rdp) |
1452 | { | |
1453 | unsigned long c; | |
1454 | int i; | |
48a7639c | 1455 | bool ret; |
dc35c893 PM |
1456 | |
1457 | /* If the CPU has no callbacks, nothing to do. */ | |
1458 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1459 | return false; |
dc35c893 PM |
1460 | |
1461 | /* | |
1462 | * Starting from the sublist containing the callbacks most | |
1463 | * recently assigned a ->completed number and working down, find the | |
1464 | * first sublist that is not assignable to an upcoming grace period. | |
1465 | * Such a sublist has something in it (first two tests) and has | |
1466 | * a ->completed number assigned that will complete sooner than | |
1467 | * the ->completed number for newly arrived callbacks (last test). | |
1468 | * | |
1469 | * The key point is that any later sublist can be assigned the | |
1470 | * same ->completed number as the newly arrived callbacks, which | |
1471 | * means that the callbacks in any of these later sublist can be | |
1472 | * grouped into a single sublist, whether or not they have already | |
1473 | * been assigned a ->completed number. | |
1474 | */ | |
1475 | c = rcu_cbs_completed(rsp, rnp); | |
1476 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1477 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1478 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1479 | break; | |
1480 | ||
1481 | /* | |
1482 | * If there are no sublist for unassigned callbacks, leave. | |
1483 | * At the same time, advance "i" one sublist, so that "i" will | |
1484 | * index into the sublist where all the remaining callbacks should | |
1485 | * be grouped into. | |
1486 | */ | |
1487 | if (++i >= RCU_NEXT_TAIL) | |
48a7639c | 1488 | return false; |
dc35c893 PM |
1489 | |
1490 | /* | |
1491 | * Assign all subsequent callbacks' ->completed number to the next | |
1492 | * full grace period and group them all in the sublist initially | |
1493 | * indexed by "i". | |
1494 | */ | |
1495 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1496 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1497 | rdp->nxtcompleted[i] = c; | |
1498 | } | |
910ee45d | 1499 | /* Record any needed additional grace periods. */ |
48a7639c | 1500 | ret = rcu_start_future_gp(rnp, rdp, NULL); |
6d4b418c PM |
1501 | |
1502 | /* Trace depending on how much we were able to accelerate. */ | |
1503 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1504 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1505 | else |
f7f7bac9 | 1506 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
48a7639c | 1507 | return ret; |
dc35c893 PM |
1508 | } |
1509 | ||
1510 | /* | |
1511 | * Move any callbacks whose grace period has completed to the | |
1512 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1513 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1514 | * sublist. This function is idempotent, so it does not hurt to | |
1515 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1516 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1517 | * |
1518 | * The caller must hold rnp->lock with interrupts disabled. | |
1519 | */ | |
48a7639c | 1520 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1521 | struct rcu_data *rdp) |
1522 | { | |
1523 | int i, j; | |
1524 | ||
1525 | /* If the CPU has no callbacks, nothing to do. */ | |
1526 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1527 | return false; |
dc35c893 PM |
1528 | |
1529 | /* | |
1530 | * Find all callbacks whose ->completed numbers indicate that they | |
1531 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1532 | */ | |
1533 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1534 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1535 | break; | |
1536 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1537 | } | |
1538 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1539 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1540 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1541 | ||
1542 | /* Copy down callbacks to fill in empty sublists. */ | |
1543 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1544 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1545 | break; | |
1546 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1547 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1548 | } | |
1549 | ||
1550 | /* Classify any remaining callbacks. */ | |
48a7639c | 1551 | return rcu_accelerate_cbs(rsp, rnp, rdp); |
dc35c893 PM |
1552 | } |
1553 | ||
d09b62df | 1554 | /* |
ba9fbe95 PM |
1555 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1556 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1557 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1558 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1559 | */ |
48a7639c PM |
1560 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, |
1561 | struct rcu_data *rdp) | |
d09b62df | 1562 | { |
48a7639c PM |
1563 | bool ret; |
1564 | ||
ba9fbe95 | 1565 | /* Handle the ends of any preceding grace periods first. */ |
dc35c893 | 1566 | if (rdp->completed == rnp->completed) { |
d09b62df | 1567 | |
ba9fbe95 | 1568 | /* No grace period end, so just accelerate recent callbacks. */ |
48a7639c | 1569 | ret = rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1570 | |
dc35c893 PM |
1571 | } else { |
1572 | ||
1573 | /* Advance callbacks. */ | |
48a7639c | 1574 | ret = rcu_advance_cbs(rsp, rnp, rdp); |
d09b62df PM |
1575 | |
1576 | /* Remember that we saw this grace-period completion. */ | |
1577 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1578 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1579 | } |
398ebe60 | 1580 | |
6eaef633 PM |
1581 | if (rdp->gpnum != rnp->gpnum) { |
1582 | /* | |
1583 | * If the current grace period is waiting for this CPU, | |
1584 | * set up to detect a quiescent state, otherwise don't | |
1585 | * go looking for one. | |
1586 | */ | |
1587 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1588 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
6eaef633 PM |
1589 | rdp->passed_quiesce = 0; |
1590 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); | |
1591 | zero_cpu_stall_ticks(rdp); | |
1592 | } | |
48a7639c | 1593 | return ret; |
6eaef633 PM |
1594 | } |
1595 | ||
d34ea322 | 1596 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1597 | { |
1598 | unsigned long flags; | |
48a7639c | 1599 | bool needwake; |
6eaef633 PM |
1600 | struct rcu_node *rnp; |
1601 | ||
1602 | local_irq_save(flags); | |
1603 | rnp = rdp->mynode; | |
d34ea322 PM |
1604 | if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) && |
1605 | rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */ | |
6eaef633 PM |
1606 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
1607 | local_irq_restore(flags); | |
1608 | return; | |
1609 | } | |
6303b9c8 | 1610 | smp_mb__after_unlock_lock(); |
48a7639c | 1611 | needwake = __note_gp_changes(rsp, rnp, rdp); |
6eaef633 | 1612 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
48a7639c PM |
1613 | if (needwake) |
1614 | rcu_gp_kthread_wake(rsp); | |
6eaef633 PM |
1615 | } |
1616 | ||
b3dbec76 | 1617 | /* |
f7be8209 | 1618 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1619 | */ |
7fdefc10 | 1620 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 PM |
1621 | { |
1622 | struct rcu_data *rdp; | |
7fdefc10 | 1623 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1624 | |
eb75767b | 1625 | rcu_bind_gp_kthread(); |
7fdefc10 | 1626 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1627 | smp_mb__after_unlock_lock(); |
91dc9542 | 1628 | if (!ACCESS_ONCE(rsp->gp_flags)) { |
f7be8209 PM |
1629 | /* Spurious wakeup, tell caller to go back to sleep. */ |
1630 | raw_spin_unlock_irq(&rnp->lock); | |
1631 | return 0; | |
1632 | } | |
91dc9542 | 1633 | ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */ |
b3dbec76 | 1634 | |
f7be8209 PM |
1635 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1636 | /* | |
1637 | * Grace period already in progress, don't start another. | |
1638 | * Not supposed to be able to happen. | |
1639 | */ | |
7fdefc10 PM |
1640 | raw_spin_unlock_irq(&rnp->lock); |
1641 | return 0; | |
1642 | } | |
1643 | ||
7fdefc10 | 1644 | /* Advance to a new grace period and initialize state. */ |
26cdfedf | 1645 | record_gp_stall_check_time(rsp); |
765a3f4f PM |
1646 | /* Record GP times before starting GP, hence smp_store_release(). */ |
1647 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); | |
f7f7bac9 | 1648 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1649 | raw_spin_unlock_irq(&rnp->lock); |
1650 | ||
1651 | /* Exclude any concurrent CPU-hotplug operations. */ | |
a4fbe35a | 1652 | mutex_lock(&rsp->onoff_mutex); |
765a3f4f | 1653 | smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */ |
7fdefc10 PM |
1654 | |
1655 | /* | |
1656 | * Set the quiescent-state-needed bits in all the rcu_node | |
1657 | * structures for all currently online CPUs in breadth-first order, | |
1658 | * starting from the root rcu_node structure, relying on the layout | |
1659 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1660 | * will access only the leaves of the hierarchy, thus seeing that no | |
1661 | * grace period is in progress, at least until the corresponding | |
1662 | * leaf node has been initialized. In addition, we have excluded | |
1663 | * CPU-hotplug operations. | |
1664 | * | |
1665 | * The grace period cannot complete until the initialization | |
1666 | * process finishes, because this kthread handles both. | |
1667 | */ | |
1668 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
b3dbec76 | 1669 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1670 | smp_mb__after_unlock_lock(); |
b3dbec76 | 1671 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1672 | rcu_preempt_check_blocked_tasks(rnp); |
1673 | rnp->qsmask = rnp->qsmaskinit; | |
0446be48 | 1674 | ACCESS_ONCE(rnp->gpnum) = rsp->gpnum; |
25d30cf4 | 1675 | WARN_ON_ONCE(rnp->completed != rsp->completed); |
0446be48 | 1676 | ACCESS_ONCE(rnp->completed) = rsp->completed; |
7fdefc10 | 1677 | if (rnp == rdp->mynode) |
48a7639c | 1678 | (void)__note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1679 | rcu_preempt_boost_start_gp(rnp); |
1680 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1681 | rnp->level, rnp->grplo, | |
1682 | rnp->grphi, rnp->qsmask); | |
1683 | raw_spin_unlock_irq(&rnp->lock); | |
bde6c3aa | 1684 | cond_resched_rcu_qs(); |
7fdefc10 | 1685 | } |
b3dbec76 | 1686 | |
a4fbe35a | 1687 | mutex_unlock(&rsp->onoff_mutex); |
7fdefc10 PM |
1688 | return 1; |
1689 | } | |
b3dbec76 | 1690 | |
4cdfc175 PM |
1691 | /* |
1692 | * Do one round of quiescent-state forcing. | |
1693 | */ | |
01896f7e | 1694 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) |
4cdfc175 PM |
1695 | { |
1696 | int fqs_state = fqs_state_in; | |
217af2a2 PM |
1697 | bool isidle = false; |
1698 | unsigned long maxj; | |
4cdfc175 PM |
1699 | struct rcu_node *rnp = rcu_get_root(rsp); |
1700 | ||
1701 | rsp->n_force_qs++; | |
1702 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1703 | /* Collect dyntick-idle snapshots. */ | |
0edd1b17 | 1704 | if (is_sysidle_rcu_state(rsp)) { |
e02b2edf | 1705 | isidle = true; |
0edd1b17 PM |
1706 | maxj = jiffies - ULONG_MAX / 4; |
1707 | } | |
217af2a2 PM |
1708 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1709 | &isidle, &maxj); | |
0edd1b17 | 1710 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1711 | fqs_state = RCU_FORCE_QS; |
1712 | } else { | |
1713 | /* Handle dyntick-idle and offline CPUs. */ | |
e02b2edf | 1714 | isidle = false; |
217af2a2 | 1715 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1716 | } |
1717 | /* Clear flag to prevent immediate re-entry. */ | |
1718 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1719 | raw_spin_lock_irq(&rnp->lock); | |
6303b9c8 | 1720 | smp_mb__after_unlock_lock(); |
4de376a1 PK |
1721 | ACCESS_ONCE(rsp->gp_flags) = |
1722 | ACCESS_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS; | |
4cdfc175 PM |
1723 | raw_spin_unlock_irq(&rnp->lock); |
1724 | } | |
1725 | return fqs_state; | |
1726 | } | |
1727 | ||
7fdefc10 PM |
1728 | /* |
1729 | * Clean up after the old grace period. | |
1730 | */ | |
4cdfc175 | 1731 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1732 | { |
1733 | unsigned long gp_duration; | |
48a7639c | 1734 | bool needgp = false; |
dae6e64d | 1735 | int nocb = 0; |
7fdefc10 PM |
1736 | struct rcu_data *rdp; |
1737 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1738 | |
7fdefc10 | 1739 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1740 | smp_mb__after_unlock_lock(); |
7fdefc10 PM |
1741 | gp_duration = jiffies - rsp->gp_start; |
1742 | if (gp_duration > rsp->gp_max) | |
1743 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1744 | |
7fdefc10 PM |
1745 | /* |
1746 | * We know the grace period is complete, but to everyone else | |
1747 | * it appears to still be ongoing. But it is also the case | |
1748 | * that to everyone else it looks like there is nothing that | |
1749 | * they can do to advance the grace period. It is therefore | |
1750 | * safe for us to drop the lock in order to mark the grace | |
1751 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1752 | */ |
5d4b8659 | 1753 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1754 | |
5d4b8659 PM |
1755 | /* |
1756 | * Propagate new ->completed value to rcu_node structures so | |
1757 | * that other CPUs don't have to wait until the start of the next | |
1758 | * grace period to process their callbacks. This also avoids | |
1759 | * some nasty RCU grace-period initialization races by forcing | |
1760 | * the end of the current grace period to be completely recorded in | |
1761 | * all of the rcu_node structures before the beginning of the next | |
1762 | * grace period is recorded in any of the rcu_node structures. | |
1763 | */ | |
1764 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1765 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1766 | smp_mb__after_unlock_lock(); |
0446be48 | 1767 | ACCESS_ONCE(rnp->completed) = rsp->gpnum; |
b11cc576 PM |
1768 | rdp = this_cpu_ptr(rsp->rda); |
1769 | if (rnp == rdp->mynode) | |
48a7639c | 1770 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; |
78e4bc34 | 1771 | /* smp_mb() provided by prior unlock-lock pair. */ |
0446be48 | 1772 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 | 1773 | raw_spin_unlock_irq(&rnp->lock); |
bde6c3aa | 1774 | cond_resched_rcu_qs(); |
7fdefc10 | 1775 | } |
5d4b8659 PM |
1776 | rnp = rcu_get_root(rsp); |
1777 | raw_spin_lock_irq(&rnp->lock); | |
765a3f4f | 1778 | smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */ |
dae6e64d | 1779 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 | 1780 | |
765a3f4f PM |
1781 | /* Declare grace period done. */ |
1782 | ACCESS_ONCE(rsp->completed) = rsp->gpnum; | |
f7f7bac9 | 1783 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
7fdefc10 | 1784 | rsp->fqs_state = RCU_GP_IDLE; |
5d4b8659 | 1785 | rdp = this_cpu_ptr(rsp->rda); |
48a7639c PM |
1786 | /* Advance CBs to reduce false positives below. */ |
1787 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; | |
1788 | if (needgp || cpu_needs_another_gp(rsp, rdp)) { | |
91dc9542 | 1789 | ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1790 | trace_rcu_grace_period(rsp->name, |
1791 | ACCESS_ONCE(rsp->gpnum), | |
1792 | TPS("newreq")); | |
1793 | } | |
7fdefc10 | 1794 | raw_spin_unlock_irq(&rnp->lock); |
7fdefc10 PM |
1795 | } |
1796 | ||
1797 | /* | |
1798 | * Body of kthread that handles grace periods. | |
1799 | */ | |
1800 | static int __noreturn rcu_gp_kthread(void *arg) | |
1801 | { | |
4cdfc175 | 1802 | int fqs_state; |
88d6df61 | 1803 | int gf; |
d40011f6 | 1804 | unsigned long j; |
4cdfc175 | 1805 | int ret; |
7fdefc10 PM |
1806 | struct rcu_state *rsp = arg; |
1807 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1808 | ||
1809 | for (;;) { | |
1810 | ||
1811 | /* Handle grace-period start. */ | |
1812 | for (;;) { | |
63c4db78 PM |
1813 | trace_rcu_grace_period(rsp->name, |
1814 | ACCESS_ONCE(rsp->gpnum), | |
1815 | TPS("reqwait")); | |
afea227f | 1816 | rsp->gp_state = RCU_GP_WAIT_GPS; |
4cdfc175 | 1817 | wait_event_interruptible(rsp->gp_wq, |
591c6d17 | 1818 | ACCESS_ONCE(rsp->gp_flags) & |
4cdfc175 | 1819 | RCU_GP_FLAG_INIT); |
78e4bc34 | 1820 | /* Locking provides needed memory barrier. */ |
f7be8209 | 1821 | if (rcu_gp_init(rsp)) |
7fdefc10 | 1822 | break; |
bde6c3aa | 1823 | cond_resched_rcu_qs(); |
73a860cd | 1824 | WARN_ON(signal_pending(current)); |
63c4db78 PM |
1825 | trace_rcu_grace_period(rsp->name, |
1826 | ACCESS_ONCE(rsp->gpnum), | |
1827 | TPS("reqwaitsig")); | |
7fdefc10 | 1828 | } |
cabc49c1 | 1829 | |
4cdfc175 PM |
1830 | /* Handle quiescent-state forcing. */ |
1831 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
1832 | j = jiffies_till_first_fqs; |
1833 | if (j > HZ) { | |
1834 | j = HZ; | |
1835 | jiffies_till_first_fqs = HZ; | |
1836 | } | |
88d6df61 | 1837 | ret = 0; |
cabc49c1 | 1838 | for (;;) { |
88d6df61 PM |
1839 | if (!ret) |
1840 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 PM |
1841 | trace_rcu_grace_period(rsp->name, |
1842 | ACCESS_ONCE(rsp->gpnum), | |
1843 | TPS("fqswait")); | |
afea227f | 1844 | rsp->gp_state = RCU_GP_WAIT_FQS; |
4cdfc175 | 1845 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
88d6df61 PM |
1846 | ((gf = ACCESS_ONCE(rsp->gp_flags)) & |
1847 | RCU_GP_FLAG_FQS) || | |
4cdfc175 PM |
1848 | (!ACCESS_ONCE(rnp->qsmask) && |
1849 | !rcu_preempt_blocked_readers_cgp(rnp)), | |
d40011f6 | 1850 | j); |
78e4bc34 | 1851 | /* Locking provides needed memory barriers. */ |
4cdfc175 | 1852 | /* If grace period done, leave loop. */ |
cabc49c1 | 1853 | if (!ACCESS_ONCE(rnp->qsmask) && |
4cdfc175 | 1854 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 1855 | break; |
4cdfc175 | 1856 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
1857 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
1858 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 PM |
1859 | trace_rcu_grace_period(rsp->name, |
1860 | ACCESS_ONCE(rsp->gpnum), | |
1861 | TPS("fqsstart")); | |
4cdfc175 | 1862 | fqs_state = rcu_gp_fqs(rsp, fqs_state); |
63c4db78 PM |
1863 | trace_rcu_grace_period(rsp->name, |
1864 | ACCESS_ONCE(rsp->gpnum), | |
1865 | TPS("fqsend")); | |
bde6c3aa | 1866 | cond_resched_rcu_qs(); |
4cdfc175 PM |
1867 | } else { |
1868 | /* Deal with stray signal. */ | |
bde6c3aa | 1869 | cond_resched_rcu_qs(); |
73a860cd | 1870 | WARN_ON(signal_pending(current)); |
63c4db78 PM |
1871 | trace_rcu_grace_period(rsp->name, |
1872 | ACCESS_ONCE(rsp->gpnum), | |
1873 | TPS("fqswaitsig")); | |
4cdfc175 | 1874 | } |
d40011f6 PM |
1875 | j = jiffies_till_next_fqs; |
1876 | if (j > HZ) { | |
1877 | j = HZ; | |
1878 | jiffies_till_next_fqs = HZ; | |
1879 | } else if (j < 1) { | |
1880 | j = 1; | |
1881 | jiffies_till_next_fqs = 1; | |
1882 | } | |
cabc49c1 | 1883 | } |
4cdfc175 PM |
1884 | |
1885 | /* Handle grace-period end. */ | |
1886 | rcu_gp_cleanup(rsp); | |
b3dbec76 | 1887 | } |
b3dbec76 PM |
1888 | } |
1889 | ||
64db4cff PM |
1890 | /* |
1891 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
1892 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 1893 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
1894 | * |
1895 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
1896 | * invoke this function. This can happen when the dying CPU reports its | |
1897 | * quiescent state. | |
48a7639c PM |
1898 | * |
1899 | * Returns true if the grace-period kthread must be awakened. | |
64db4cff | 1900 | */ |
48a7639c | 1901 | static bool |
910ee45d PM |
1902 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
1903 | struct rcu_data *rdp) | |
64db4cff | 1904 | { |
b8462084 | 1905 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 1906 | /* |
b3dbec76 | 1907 | * Either we have not yet spawned the grace-period |
62da1921 PM |
1908 | * task, this CPU does not need another grace period, |
1909 | * or a grace period is already in progress. | |
b3dbec76 | 1910 | * Either way, don't start a new grace period. |
afe24b12 | 1911 | */ |
48a7639c | 1912 | return false; |
afe24b12 | 1913 | } |
91dc9542 | 1914 | ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1915 | trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum), |
1916 | TPS("newreq")); | |
62da1921 | 1917 | |
016a8d5b SR |
1918 | /* |
1919 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c | 1920 | * could cause possible deadlocks with the rq->lock. Defer |
48a7639c | 1921 | * the wakeup to our caller. |
016a8d5b | 1922 | */ |
48a7639c | 1923 | return true; |
64db4cff PM |
1924 | } |
1925 | ||
910ee45d PM |
1926 | /* |
1927 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
1928 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
1929 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
1930 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
1931 | * that is encountered beforehand. | |
48a7639c PM |
1932 | * |
1933 | * Returns true if the grace-period kthread needs to be awakened. | |
910ee45d | 1934 | */ |
48a7639c | 1935 | static bool rcu_start_gp(struct rcu_state *rsp) |
910ee45d PM |
1936 | { |
1937 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1938 | struct rcu_node *rnp = rcu_get_root(rsp); | |
48a7639c | 1939 | bool ret = false; |
910ee45d PM |
1940 | |
1941 | /* | |
1942 | * If there is no grace period in progress right now, any | |
1943 | * callbacks we have up to this point will be satisfied by the | |
1944 | * next grace period. Also, advancing the callbacks reduces the | |
1945 | * probability of false positives from cpu_needs_another_gp() | |
1946 | * resulting in pointless grace periods. So, advance callbacks | |
1947 | * then start the grace period! | |
1948 | */ | |
48a7639c PM |
1949 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; |
1950 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; | |
1951 | return ret; | |
910ee45d PM |
1952 | } |
1953 | ||
f41d911f | 1954 | /* |
d3f6bad3 PM |
1955 | * Report a full set of quiescent states to the specified rcu_state |
1956 | * data structure. This involves cleaning up after the prior grace | |
1957 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
1958 | * if one is needed. Note that the caller must hold rnp->lock, which |
1959 | * is released before return. | |
f41d911f | 1960 | */ |
d3f6bad3 | 1961 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 1962 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 1963 | { |
fc2219d4 | 1964 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cabc49c1 | 1965 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
2aa792e6 | 1966 | rcu_gp_kthread_wake(rsp); |
f41d911f PM |
1967 | } |
1968 | ||
64db4cff | 1969 | /* |
d3f6bad3 PM |
1970 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
1971 | * Allows quiescent states for a group of CPUs to be reported at one go | |
1972 | * to the specified rcu_node structure, though all the CPUs in the group | |
1973 | * must be represented by the same rcu_node structure (which need not be | |
1974 | * a leaf rcu_node structure, though it often will be). That structure's | |
1975 | * lock must be held upon entry, and it is released before return. | |
64db4cff PM |
1976 | */ |
1977 | static void | |
d3f6bad3 PM |
1978 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
1979 | struct rcu_node *rnp, unsigned long flags) | |
64db4cff PM |
1980 | __releases(rnp->lock) |
1981 | { | |
28ecd580 PM |
1982 | struct rcu_node *rnp_c; |
1983 | ||
64db4cff PM |
1984 | /* Walk up the rcu_node hierarchy. */ |
1985 | for (;;) { | |
1986 | if (!(rnp->qsmask & mask)) { | |
1987 | ||
1988 | /* Our bit has already been cleared, so done. */ | |
1304afb2 | 1989 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1990 | return; |
1991 | } | |
1992 | rnp->qsmask &= ~mask; | |
d4c08f2a PM |
1993 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
1994 | mask, rnp->qsmask, rnp->level, | |
1995 | rnp->grplo, rnp->grphi, | |
1996 | !!rnp->gp_tasks); | |
27f4d280 | 1997 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
1998 | |
1999 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 2000 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2001 | return; |
2002 | } | |
2003 | mask = rnp->grpmask; | |
2004 | if (rnp->parent == NULL) { | |
2005 | ||
2006 | /* No more levels. Exit loop holding root lock. */ | |
2007 | ||
2008 | break; | |
2009 | } | |
1304afb2 | 2010 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 2011 | rnp_c = rnp; |
64db4cff | 2012 | rnp = rnp->parent; |
1304afb2 | 2013 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2014 | smp_mb__after_unlock_lock(); |
28ecd580 | 2015 | WARN_ON_ONCE(rnp_c->qsmask); |
64db4cff PM |
2016 | } |
2017 | ||
2018 | /* | |
2019 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 2020 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 2021 | * to clean up and start the next grace period if one is needed. |
64db4cff | 2022 | */ |
d3f6bad3 | 2023 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
2024 | } |
2025 | ||
2026 | /* | |
d3f6bad3 PM |
2027 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
2028 | * structure. This must be either called from the specified CPU, or | |
2029 | * called when the specified CPU is known to be offline (and when it is | |
2030 | * also known that no other CPU is concurrently trying to help the offline | |
2031 | * CPU). The lastcomp argument is used to make sure we are still in the | |
2032 | * grace period of interest. We don't want to end the current grace period | |
2033 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
2034 | */ |
2035 | static void | |
d7d6a11e | 2036 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2037 | { |
2038 | unsigned long flags; | |
2039 | unsigned long mask; | |
48a7639c | 2040 | bool needwake; |
64db4cff PM |
2041 | struct rcu_node *rnp; |
2042 | ||
2043 | rnp = rdp->mynode; | |
1304afb2 | 2044 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2045 | smp_mb__after_unlock_lock(); |
d7d6a11e PM |
2046 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum || |
2047 | rnp->completed == rnp->gpnum) { | |
64db4cff PM |
2048 | |
2049 | /* | |
e4cc1f22 PM |
2050 | * The grace period in which this quiescent state was |
2051 | * recorded has ended, so don't report it upwards. | |
2052 | * We will instead need a new quiescent state that lies | |
2053 | * within the current grace period. | |
64db4cff | 2054 | */ |
e4cc1f22 | 2055 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
1304afb2 | 2056 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2057 | return; |
2058 | } | |
2059 | mask = rdp->grpmask; | |
2060 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 2061 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2062 | } else { |
2063 | rdp->qs_pending = 0; | |
2064 | ||
2065 | /* | |
2066 | * This GP can't end until cpu checks in, so all of our | |
2067 | * callbacks can be processed during the next GP. | |
2068 | */ | |
48a7639c | 2069 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 2070 | |
d3f6bad3 | 2071 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
48a7639c PM |
2072 | if (needwake) |
2073 | rcu_gp_kthread_wake(rsp); | |
64db4cff PM |
2074 | } |
2075 | } | |
2076 | ||
2077 | /* | |
2078 | * Check to see if there is a new grace period of which this CPU | |
2079 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2080 | * Otherwise, see if this CPU has just passed through its first | |
2081 | * quiescent state for this grace period, and record that fact if so. | |
2082 | */ | |
2083 | static void | |
2084 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
2085 | { | |
05eb552b PM |
2086 | /* Check for grace-period ends and beginnings. */ |
2087 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
2088 | |
2089 | /* | |
2090 | * Does this CPU still need to do its part for current grace period? | |
2091 | * If no, return and let the other CPUs do their part as well. | |
2092 | */ | |
2093 | if (!rdp->qs_pending) | |
2094 | return; | |
2095 | ||
2096 | /* | |
2097 | * Was there a quiescent state since the beginning of the grace | |
2098 | * period? If no, then exit and wait for the next call. | |
2099 | */ | |
e4cc1f22 | 2100 | if (!rdp->passed_quiesce) |
64db4cff PM |
2101 | return; |
2102 | ||
d3f6bad3 PM |
2103 | /* |
2104 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2105 | * judge of that). | |
2106 | */ | |
d7d6a11e | 2107 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
2108 | } |
2109 | ||
2110 | #ifdef CONFIG_HOTPLUG_CPU | |
2111 | ||
e74f4c45 | 2112 | /* |
b1420f1c PM |
2113 | * Send the specified CPU's RCU callbacks to the orphanage. The |
2114 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 2115 | * ->orphan_lock. |
e74f4c45 | 2116 | */ |
b1420f1c PM |
2117 | static void |
2118 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
2119 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 2120 | { |
3fbfbf7a | 2121 | /* No-CBs CPUs do not have orphanable callbacks. */ |
d1e43fa5 | 2122 | if (rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
2123 | return; |
2124 | ||
b1420f1c PM |
2125 | /* |
2126 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
2127 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
2128 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 2129 | */ |
a50c3af9 | 2130 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2131 | rsp->qlen_lazy += rdp->qlen_lazy; |
2132 | rsp->qlen += rdp->qlen; | |
2133 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 2134 | rdp->qlen_lazy = 0; |
1d1fb395 | 2135 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
2136 | } |
2137 | ||
2138 | /* | |
b1420f1c PM |
2139 | * Next, move those callbacks still needing a grace period to |
2140 | * the orphanage, where some other CPU will pick them up. | |
2141 | * Some of the callbacks might have gone partway through a grace | |
2142 | * period, but that is too bad. They get to start over because we | |
2143 | * cannot assume that grace periods are synchronized across CPUs. | |
2144 | * We don't bother updating the ->nxttail[] array yet, instead | |
2145 | * we just reset the whole thing later on. | |
a50c3af9 | 2146 | */ |
b1420f1c PM |
2147 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
2148 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2149 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2150 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
2151 | } |
2152 | ||
2153 | /* | |
b1420f1c PM |
2154 | * Then move the ready-to-invoke callbacks to the orphanage, |
2155 | * where some other CPU will pick them up. These will not be | |
2156 | * required to pass though another grace period: They are done. | |
a50c3af9 | 2157 | */ |
e5601400 | 2158 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2159 | *rsp->orphan_donetail = rdp->nxtlist; |
2160 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 2161 | } |
e74f4c45 | 2162 | |
b1420f1c | 2163 | /* Finally, initialize the rcu_data structure's list to empty. */ |
3f5d3ea6 | 2164 | init_callback_list(rdp); |
b1420f1c PM |
2165 | } |
2166 | ||
2167 | /* | |
2168 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 2169 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c | 2170 | */ |
96d3fd0d | 2171 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) |
b1420f1c PM |
2172 | { |
2173 | int i; | |
fa07a58f | 2174 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
b1420f1c | 2175 | |
3fbfbf7a | 2176 | /* No-CBs CPUs are handled specially. */ |
96d3fd0d | 2177 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) |
3fbfbf7a PM |
2178 | return; |
2179 | ||
b1420f1c PM |
2180 | /* Do the accounting first. */ |
2181 | rdp->qlen_lazy += rsp->qlen_lazy; | |
2182 | rdp->qlen += rsp->qlen; | |
2183 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
2184 | if (rsp->qlen_lazy != rsp->qlen) |
2185 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2186 | rsp->qlen_lazy = 0; |
2187 | rsp->qlen = 0; | |
2188 | ||
2189 | /* | |
2190 | * We do not need a memory barrier here because the only way we | |
2191 | * can get here if there is an rcu_barrier() in flight is if | |
2192 | * we are the task doing the rcu_barrier(). | |
2193 | */ | |
2194 | ||
2195 | /* First adopt the ready-to-invoke callbacks. */ | |
2196 | if (rsp->orphan_donelist != NULL) { | |
2197 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2198 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
2199 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
2200 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2201 | rdp->nxttail[i] = rsp->orphan_donetail; | |
2202 | rsp->orphan_donelist = NULL; | |
2203 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2204 | } | |
2205 | ||
2206 | /* And then adopt the callbacks that still need a grace period. */ | |
2207 | if (rsp->orphan_nxtlist != NULL) { | |
2208 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
2209 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
2210 | rsp->orphan_nxtlist = NULL; | |
2211 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2212 | } | |
2213 | } | |
2214 | ||
2215 | /* | |
2216 | * Trace the fact that this CPU is going offline. | |
2217 | */ | |
2218 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
2219 | { | |
2220 | RCU_TRACE(unsigned long mask); | |
2221 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
2222 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
2223 | ||
2224 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
2225 | trace_rcu_grace_period(rsp->name, |
2226 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 2227 | TPS("cpuofl")); |
64db4cff PM |
2228 | } |
2229 | ||
2230 | /* | |
e5601400 | 2231 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
2232 | * this fact from process context. Do the remainder of the cleanup, |
2233 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
2234 | * adopting them. There can only be one CPU hotplug operation at a time, |
2235 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 2236 | */ |
e5601400 | 2237 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 2238 | { |
2036d94a PM |
2239 | unsigned long flags; |
2240 | unsigned long mask; | |
2241 | int need_report = 0; | |
e5601400 | 2242 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 2243 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2244 | |
2036d94a | 2245 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2246 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 2247 | |
2036d94a | 2248 | /* Exclude any attempts to start a new grace period. */ |
a4fbe35a | 2249 | mutex_lock(&rsp->onoff_mutex); |
7b2e6011 | 2250 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
2036d94a | 2251 | |
b1420f1c PM |
2252 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
2253 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
96d3fd0d | 2254 | rcu_adopt_orphan_cbs(rsp, flags); |
b1420f1c | 2255 | |
2036d94a PM |
2256 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
2257 | mask = rdp->grpmask; /* rnp->grplo is constant. */ | |
2258 | do { | |
2259 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
6303b9c8 | 2260 | smp_mb__after_unlock_lock(); |
2036d94a PM |
2261 | rnp->qsmaskinit &= ~mask; |
2262 | if (rnp->qsmaskinit != 0) { | |
2263 | if (rnp != rdp->mynode) | |
2264 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2265 | break; | |
2266 | } | |
2267 | if (rnp == rdp->mynode) | |
2268 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); | |
2269 | else | |
2270 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2271 | mask = rnp->grpmask; | |
2272 | rnp = rnp->parent; | |
2273 | } while (rnp != NULL); | |
2274 | ||
2275 | /* | |
2276 | * We still hold the leaf rcu_node structure lock here, and | |
2277 | * irqs are still disabled. The reason for this subterfuge is | |
7b2e6011 | 2278 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock |
2036d94a PM |
2279 | * held leads to deadlock. |
2280 | */ | |
7b2e6011 | 2281 | raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */ |
2036d94a PM |
2282 | rnp = rdp->mynode; |
2283 | if (need_report & RCU_OFL_TASKS_NORM_GP) | |
2284 | rcu_report_unblock_qs_rnp(rnp, flags); | |
2285 | else | |
2286 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2287 | if (need_report & RCU_OFL_TASKS_EXP_GP) | |
2288 | rcu_report_exp_rnp(rsp, rnp, true); | |
cf01537e PM |
2289 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
2290 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2291 | cpu, rdp->qlen, rdp->nxtlist); | |
0d8ee37e PM |
2292 | init_callback_list(rdp); |
2293 | /* Disallow further callbacks on this CPU. */ | |
2294 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; | |
a4fbe35a | 2295 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
2296 | } |
2297 | ||
2298 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
2299 | ||
e5601400 | 2300 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
2301 | { |
2302 | } | |
2303 | ||
e5601400 | 2304 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
2305 | { |
2306 | } | |
2307 | ||
2308 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
2309 | ||
2310 | /* | |
2311 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2312 | * period. Thottle as specified by rdp->blimit. | |
2313 | */ | |
37c72e56 | 2314 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2315 | { |
2316 | unsigned long flags; | |
2317 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
2318 | long bl, count, count_lazy; |
2319 | int i; | |
64db4cff | 2320 | |
dc35c893 | 2321 | /* If no callbacks are ready, just return. */ |
29c00b4a | 2322 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 2323 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
2324 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
2325 | need_resched(), is_idle_task(current), | |
2326 | rcu_is_callbacks_kthread()); | |
64db4cff | 2327 | return; |
29c00b4a | 2328 | } |
64db4cff PM |
2329 | |
2330 | /* | |
2331 | * Extract the list of ready callbacks, disabling to prevent | |
2332 | * races with call_rcu() from interrupt handlers. | |
2333 | */ | |
2334 | local_irq_save(flags); | |
8146c4e2 | 2335 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2336 | bl = rdp->blimit; |
486e2593 | 2337 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2338 | list = rdp->nxtlist; |
2339 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2340 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2341 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2342 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2343 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2344 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2345 | local_irq_restore(flags); |
2346 | ||
2347 | /* Invoke callbacks. */ | |
486e2593 | 2348 | count = count_lazy = 0; |
64db4cff PM |
2349 | while (list) { |
2350 | next = list->next; | |
2351 | prefetch(next); | |
551d55a9 | 2352 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2353 | if (__rcu_reclaim(rsp->name, list)) |
2354 | count_lazy++; | |
64db4cff | 2355 | list = next; |
dff1672d PM |
2356 | /* Stop only if limit reached and CPU has something to do. */ |
2357 | if (++count >= bl && | |
2358 | (need_resched() || | |
2359 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2360 | break; |
2361 | } | |
2362 | ||
2363 | local_irq_save(flags); | |
4968c300 PM |
2364 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2365 | is_idle_task(current), | |
2366 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2367 | |
2368 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2369 | if (list != NULL) { |
2370 | *tail = rdp->nxtlist; | |
2371 | rdp->nxtlist = list; | |
b41772ab PM |
2372 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2373 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2374 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2375 | else |
2376 | break; | |
2377 | } | |
b1420f1c PM |
2378 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2379 | rdp->qlen_lazy -= count_lazy; | |
a792563b | 2380 | ACCESS_ONCE(rdp->qlen) = rdp->qlen - count; |
b1420f1c | 2381 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2382 | |
2383 | /* Reinstate batch limit if we have worked down the excess. */ | |
2384 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2385 | rdp->blimit = blimit; | |
2386 | ||
37c72e56 PM |
2387 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2388 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2389 | rdp->qlen_last_fqs_check = 0; | |
2390 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2391 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2392 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2393 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2394 | |
64db4cff PM |
2395 | local_irq_restore(flags); |
2396 | ||
e0f23060 | 2397 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2398 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2399 | invoke_rcu_core(); |
64db4cff PM |
2400 | } |
2401 | ||
2402 | /* | |
2403 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2404 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2405 | * Also schedule RCU core processing. |
64db4cff | 2406 | * |
9b2e4f18 | 2407 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2408 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2409 | * false, there is no point in invoking rcu_check_callbacks(). | |
2410 | */ | |
c3377c2d | 2411 | void rcu_check_callbacks(int user) |
64db4cff | 2412 | { |
f7f7bac9 | 2413 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2414 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2415 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2416 | |
2417 | /* | |
2418 | * Get here if this CPU took its interrupt from user | |
2419 | * mode or from the idle loop, and if this is not a | |
2420 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2421 | * a quiescent state, so note it. |
64db4cff PM |
2422 | * |
2423 | * No memory barrier is required here because both | |
d6714c22 PM |
2424 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2425 | * variables that other CPUs neither access nor modify, | |
2426 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2427 | */ |
2428 | ||
284a8c93 PM |
2429 | rcu_sched_qs(); |
2430 | rcu_bh_qs(); | |
64db4cff PM |
2431 | |
2432 | } else if (!in_softirq()) { | |
2433 | ||
2434 | /* | |
2435 | * Get here if this CPU did not take its interrupt from | |
2436 | * softirq, in other words, if it is not interrupting | |
2437 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2438 | * critical section, so note it. |
64db4cff PM |
2439 | */ |
2440 | ||
284a8c93 | 2441 | rcu_bh_qs(); |
64db4cff | 2442 | } |
86aea0e6 | 2443 | rcu_preempt_check_callbacks(); |
e3950ecd | 2444 | if (rcu_pending()) |
a46e0899 | 2445 | invoke_rcu_core(); |
8315f422 PM |
2446 | if (user) |
2447 | rcu_note_voluntary_context_switch(current); | |
f7f7bac9 | 2448 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2449 | } |
2450 | ||
64db4cff PM |
2451 | /* |
2452 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2453 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2454 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2455 | * | |
ee47eb9f | 2456 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2457 | */ |
217af2a2 PM |
2458 | static void force_qs_rnp(struct rcu_state *rsp, |
2459 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2460 | unsigned long *maxj), | |
2461 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2462 | { |
2463 | unsigned long bit; | |
2464 | int cpu; | |
2465 | unsigned long flags; | |
2466 | unsigned long mask; | |
a0b6c9a7 | 2467 | struct rcu_node *rnp; |
64db4cff | 2468 | |
a0b6c9a7 | 2469 | rcu_for_each_leaf_node(rsp, rnp) { |
bde6c3aa | 2470 | cond_resched_rcu_qs(); |
64db4cff | 2471 | mask = 0; |
1304afb2 | 2472 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2473 | smp_mb__after_unlock_lock(); |
ee47eb9f | 2474 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 2475 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 2476 | return; |
64db4cff | 2477 | } |
a0b6c9a7 | 2478 | if (rnp->qsmask == 0) { |
1217ed1b | 2479 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
64db4cff PM |
2480 | continue; |
2481 | } | |
a0b6c9a7 | 2482 | cpu = rnp->grplo; |
64db4cff | 2483 | bit = 1; |
a0b6c9a7 | 2484 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 PM |
2485 | if ((rnp->qsmask & bit) != 0) { |
2486 | if ((rnp->qsmaskinit & bit) != 0) | |
e02b2edf | 2487 | *isidle = false; |
0edd1b17 PM |
2488 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) |
2489 | mask |= bit; | |
2490 | } | |
64db4cff | 2491 | } |
45f014c5 | 2492 | if (mask != 0) { |
64db4cff | 2493 | |
d3f6bad3 PM |
2494 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
2495 | rcu_report_qs_rnp(mask, rsp, rnp, flags); | |
64db4cff PM |
2496 | continue; |
2497 | } | |
1304afb2 | 2498 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2499 | } |
27f4d280 | 2500 | rnp = rcu_get_root(rsp); |
1217ed1b PM |
2501 | if (rnp->qsmask == 0) { |
2502 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 2503 | smp_mb__after_unlock_lock(); |
1217ed1b PM |
2504 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ |
2505 | } | |
64db4cff PM |
2506 | } |
2507 | ||
2508 | /* | |
2509 | * Force quiescent states on reluctant CPUs, and also detect which | |
2510 | * CPUs are in dyntick-idle mode. | |
2511 | */ | |
4cdfc175 | 2512 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2513 | { |
2514 | unsigned long flags; | |
394f2769 PM |
2515 | bool ret; |
2516 | struct rcu_node *rnp; | |
2517 | struct rcu_node *rnp_old = NULL; | |
2518 | ||
2519 | /* Funnel through hierarchy to reduce memory contention. */ | |
d860d403 | 2520 | rnp = __this_cpu_read(rsp->rda->mynode); |
394f2769 PM |
2521 | for (; rnp != NULL; rnp = rnp->parent) { |
2522 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2523 | !raw_spin_trylock(&rnp->fqslock); | |
2524 | if (rnp_old != NULL) | |
2525 | raw_spin_unlock(&rnp_old->fqslock); | |
2526 | if (ret) { | |
a792563b | 2527 | rsp->n_force_qs_lh++; |
394f2769 PM |
2528 | return; |
2529 | } | |
2530 | rnp_old = rnp; | |
2531 | } | |
2532 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2533 | |
394f2769 PM |
2534 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2535 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
6303b9c8 | 2536 | smp_mb__after_unlock_lock(); |
394f2769 PM |
2537 | raw_spin_unlock(&rnp_old->fqslock); |
2538 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
a792563b | 2539 | rsp->n_force_qs_lh++; |
394f2769 | 2540 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2541 | return; /* Someone beat us to it. */ |
46a1e34e | 2542 | } |
4de376a1 PK |
2543 | ACCESS_ONCE(rsp->gp_flags) = |
2544 | ACCESS_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS; | |
394f2769 | 2545 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
2aa792e6 | 2546 | rcu_gp_kthread_wake(rsp); |
64db4cff PM |
2547 | } |
2548 | ||
64db4cff | 2549 | /* |
e0f23060 PM |
2550 | * This does the RCU core processing work for the specified rcu_state |
2551 | * and rcu_data structures. This may be called only from the CPU to | |
2552 | * whom the rdp belongs. | |
64db4cff PM |
2553 | */ |
2554 | static void | |
1bca8cf1 | 2555 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2556 | { |
2557 | unsigned long flags; | |
48a7639c | 2558 | bool needwake; |
fa07a58f | 2559 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
64db4cff | 2560 | |
2e597558 PM |
2561 | WARN_ON_ONCE(rdp->beenonline == 0); |
2562 | ||
64db4cff PM |
2563 | /* Update RCU state based on any recent quiescent states. */ |
2564 | rcu_check_quiescent_state(rsp, rdp); | |
2565 | ||
2566 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2567 | local_irq_save(flags); |
64db4cff | 2568 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2569 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
48a7639c | 2570 | needwake = rcu_start_gp(rsp); |
b8462084 | 2571 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
48a7639c PM |
2572 | if (needwake) |
2573 | rcu_gp_kthread_wake(rsp); | |
dc35c893 PM |
2574 | } else { |
2575 | local_irq_restore(flags); | |
64db4cff PM |
2576 | } |
2577 | ||
2578 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2579 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2580 | invoke_rcu_callbacks(rsp, rdp); |
96d3fd0d PM |
2581 | |
2582 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2583 | do_nocb_deferred_wakeup(rdp); | |
09223371 SL |
2584 | } |
2585 | ||
64db4cff | 2586 | /* |
e0f23060 | 2587 | * Do RCU core processing for the current CPU. |
64db4cff | 2588 | */ |
09223371 | 2589 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2590 | { |
6ce75a23 PM |
2591 | struct rcu_state *rsp; |
2592 | ||
bfa00b4c PM |
2593 | if (cpu_is_offline(smp_processor_id())) |
2594 | return; | |
f7f7bac9 | 2595 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2596 | for_each_rcu_flavor(rsp) |
2597 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2598 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2599 | } |
2600 | ||
a26ac245 | 2601 | /* |
e0f23060 PM |
2602 | * Schedule RCU callback invocation. If the specified type of RCU |
2603 | * does not support RCU priority boosting, just do a direct call, | |
2604 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
2605 | * are running on the current CPU with interrupts disabled, the | |
2606 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
a26ac245 | 2607 | */ |
a46e0899 | 2608 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2609 | { |
b0d30417 PM |
2610 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
2611 | return; | |
a46e0899 PM |
2612 | if (likely(!rsp->boost)) { |
2613 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2614 | return; |
2615 | } | |
a46e0899 | 2616 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2617 | } |
2618 | ||
a46e0899 | 2619 | static void invoke_rcu_core(void) |
09223371 | 2620 | { |
b0f74036 PM |
2621 | if (cpu_online(smp_processor_id())) |
2622 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2623 | } |
2624 | ||
29154c57 PM |
2625 | /* |
2626 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2627 | */ | |
2628 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2629 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2630 | { |
48a7639c PM |
2631 | bool needwake; |
2632 | ||
62fde6ed PM |
2633 | /* |
2634 | * If called from an extended quiescent state, invoke the RCU | |
2635 | * core in order to force a re-evaluation of RCU's idleness. | |
2636 | */ | |
5c173eb8 | 2637 | if (!rcu_is_watching() && cpu_online(smp_processor_id())) |
62fde6ed PM |
2638 | invoke_rcu_core(); |
2639 | ||
a16b7a69 | 2640 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2641 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2642 | return; |
64db4cff | 2643 | |
37c72e56 PM |
2644 | /* |
2645 | * Force the grace period if too many callbacks or too long waiting. | |
2646 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2647 | * if some other CPU has recently done so. Also, don't bother | |
2648 | * invoking force_quiescent_state() if the newly enqueued callback | |
2649 | * is the only one waiting for a grace period to complete. | |
2650 | */ | |
2655d57e | 2651 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2652 | |
2653 | /* Are we ignoring a completed grace period? */ | |
470716fc | 2654 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
2655 | |
2656 | /* Start a new grace period if one not already started. */ | |
2657 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
2658 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
2659 | ||
b8462084 | 2660 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 | 2661 | smp_mb__after_unlock_lock(); |
48a7639c | 2662 | needwake = rcu_start_gp(rsp); |
b8462084 | 2663 | raw_spin_unlock(&rnp_root->lock); |
48a7639c PM |
2664 | if (needwake) |
2665 | rcu_gp_kthread_wake(rsp); | |
b52573d2 PM |
2666 | } else { |
2667 | /* Give the grace period a kick. */ | |
2668 | rdp->blimit = LONG_MAX; | |
2669 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2670 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2671 | force_quiescent_state(rsp); |
b52573d2 PM |
2672 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2673 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2674 | } | |
4cdfc175 | 2675 | } |
29154c57 PM |
2676 | } |
2677 | ||
ae150184 PM |
2678 | /* |
2679 | * RCU callback function to leak a callback. | |
2680 | */ | |
2681 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2682 | { | |
2683 | } | |
2684 | ||
3fbfbf7a PM |
2685 | /* |
2686 | * Helper function for call_rcu() and friends. The cpu argument will | |
2687 | * normally be -1, indicating "currently running CPU". It may specify | |
2688 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2689 | * is expected to specify a CPU. | |
2690 | */ | |
64db4cff PM |
2691 | static void |
2692 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 2693 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
2694 | { |
2695 | unsigned long flags; | |
2696 | struct rcu_data *rdp; | |
2697 | ||
1146edcb | 2698 | WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */ |
ae150184 PM |
2699 | if (debug_rcu_head_queue(head)) { |
2700 | /* Probable double call_rcu(), so leak the callback. */ | |
2701 | ACCESS_ONCE(head->func) = rcu_leak_callback; | |
2702 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); | |
2703 | return; | |
2704 | } | |
64db4cff PM |
2705 | head->func = func; |
2706 | head->next = NULL; | |
2707 | ||
64db4cff PM |
2708 | /* |
2709 | * Opportunistically note grace-period endings and beginnings. | |
2710 | * Note that we might see a beginning right after we see an | |
2711 | * end, but never vice versa, since this CPU has to pass through | |
2712 | * a quiescent state betweentimes. | |
2713 | */ | |
2714 | local_irq_save(flags); | |
394f99a9 | 2715 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
2716 | |
2717 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
2718 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
2719 | int offline; | |
2720 | ||
2721 | if (cpu != -1) | |
2722 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
96d3fd0d | 2723 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); |
3fbfbf7a | 2724 | WARN_ON_ONCE(offline); |
0d8ee37e | 2725 | /* _call_rcu() is illegal on offline CPU; leak the callback. */ |
0d8ee37e PM |
2726 | local_irq_restore(flags); |
2727 | return; | |
2728 | } | |
a792563b | 2729 | ACCESS_ONCE(rdp->qlen) = rdp->qlen + 1; |
486e2593 PM |
2730 | if (lazy) |
2731 | rdp->qlen_lazy++; | |
c57afe80 PM |
2732 | else |
2733 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2734 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
2735 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
2736 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 2737 | |
d4c08f2a PM |
2738 | if (__is_kfree_rcu_offset((unsigned long)func)) |
2739 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 2740 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2741 | else |
486e2593 | 2742 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2743 | |
29154c57 PM |
2744 | /* Go handle any RCU core processing required. */ |
2745 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
2746 | local_irq_restore(flags); |
2747 | } | |
2748 | ||
2749 | /* | |
d6714c22 | 2750 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 2751 | */ |
d6714c22 | 2752 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 2753 | { |
3fbfbf7a | 2754 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 2755 | } |
d6714c22 | 2756 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
2757 | |
2758 | /* | |
486e2593 | 2759 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
2760 | */ |
2761 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
2762 | { | |
3fbfbf7a | 2763 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
2764 | } |
2765 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
2766 | ||
495aa969 ACB |
2767 | /* |
2768 | * Queue an RCU callback for lazy invocation after a grace period. | |
2769 | * This will likely be later named something like "call_rcu_lazy()", | |
2770 | * but this change will require some way of tagging the lazy RCU | |
2771 | * callbacks in the list of pending callbacks. Until then, this | |
2772 | * function may only be called from __kfree_rcu(). | |
2773 | */ | |
2774 | void kfree_call_rcu(struct rcu_head *head, | |
2775 | void (*func)(struct rcu_head *rcu)) | |
2776 | { | |
e534165b | 2777 | __call_rcu(head, func, rcu_state_p, -1, 1); |
495aa969 ACB |
2778 | } |
2779 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
2780 | ||
6d813391 PM |
2781 | /* |
2782 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
2783 | * any blocking grace-period wait automatically implies a grace period | |
2784 | * if there is only one CPU online at any point time during execution | |
2785 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
2786 | * occasionally incorrectly indicate that there are multiple CPUs online | |
2787 | * when there was in fact only one the whole time, as this just adds | |
2788 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
2789 | */ |
2790 | static inline int rcu_blocking_is_gp(void) | |
2791 | { | |
95f0c1de PM |
2792 | int ret; |
2793 | ||
6d813391 | 2794 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
2795 | preempt_disable(); |
2796 | ret = num_online_cpus() <= 1; | |
2797 | preempt_enable(); | |
2798 | return ret; | |
6d813391 PM |
2799 | } |
2800 | ||
6ebb237b PM |
2801 | /** |
2802 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
2803 | * | |
2804 | * Control will return to the caller some time after a full rcu-sched | |
2805 | * grace period has elapsed, in other words after all currently executing | |
2806 | * rcu-sched read-side critical sections have completed. These read-side | |
2807 | * critical sections are delimited by rcu_read_lock_sched() and | |
2808 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
2809 | * local_irq_disable(), and so on may be used in place of | |
2810 | * rcu_read_lock_sched(). | |
2811 | * | |
2812 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
2813 | * non-threaded hardware-interrupt handlers, in progress on entry will |
2814 | * have completed before this primitive returns. However, this does not | |
2815 | * guarantee that softirq handlers will have completed, since in some | |
2816 | * kernels, these handlers can run in process context, and can block. | |
2817 | * | |
2818 | * Note that this guarantee implies further memory-ordering guarantees. | |
2819 | * On systems with more than one CPU, when synchronize_sched() returns, | |
2820 | * each CPU is guaranteed to have executed a full memory barrier since the | |
2821 | * end of its last RCU-sched read-side critical section whose beginning | |
2822 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
2823 | * an RCU read-side critical section that extends beyond the return from | |
2824 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
2825 | * after the beginning of synchronize_sched() and before the beginning of | |
2826 | * that RCU read-side critical section. Note that these guarantees include | |
2827 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
2828 | * that are executing in the kernel. | |
2829 | * | |
2830 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
2831 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
2832 | * to have executed a full memory barrier during the execution of | |
2833 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
2834 | * again only if the system has more than one CPU). | |
6ebb237b PM |
2835 | * |
2836 | * This primitive provides the guarantees made by the (now removed) | |
2837 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
2838 | * guarantees that rcu_read_lock() sections will have completed. | |
2839 | * In "classic RCU", these two guarantees happen to be one and | |
2840 | * the same, but can differ in realtime RCU implementations. | |
2841 | */ | |
2842 | void synchronize_sched(void) | |
2843 | { | |
fe15d706 PM |
2844 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2845 | !lock_is_held(&rcu_lock_map) && | |
2846 | !lock_is_held(&rcu_sched_lock_map), | |
2847 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
2848 | if (rcu_blocking_is_gp()) |
2849 | return; | |
3705b88d AM |
2850 | if (rcu_expedited) |
2851 | synchronize_sched_expedited(); | |
2852 | else | |
2853 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
2854 | } |
2855 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
2856 | ||
2857 | /** | |
2858 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
2859 | * | |
2860 | * Control will return to the caller some time after a full rcu_bh grace | |
2861 | * period has elapsed, in other words after all currently executing rcu_bh | |
2862 | * read-side critical sections have completed. RCU read-side critical | |
2863 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
2864 | * and may be nested. | |
f0a0e6f2 PM |
2865 | * |
2866 | * See the description of synchronize_sched() for more detailed information | |
2867 | * on memory ordering guarantees. | |
6ebb237b PM |
2868 | */ |
2869 | void synchronize_rcu_bh(void) | |
2870 | { | |
fe15d706 PM |
2871 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2872 | !lock_is_held(&rcu_lock_map) && | |
2873 | !lock_is_held(&rcu_sched_lock_map), | |
2874 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
2875 | if (rcu_blocking_is_gp()) |
2876 | return; | |
3705b88d AM |
2877 | if (rcu_expedited) |
2878 | synchronize_rcu_bh_expedited(); | |
2879 | else | |
2880 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
2881 | } |
2882 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
2883 | ||
765a3f4f PM |
2884 | /** |
2885 | * get_state_synchronize_rcu - Snapshot current RCU state | |
2886 | * | |
2887 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
2888 | * to determine whether or not a full grace period has elapsed in the | |
2889 | * meantime. | |
2890 | */ | |
2891 | unsigned long get_state_synchronize_rcu(void) | |
2892 | { | |
2893 | /* | |
2894 | * Any prior manipulation of RCU-protected data must happen | |
2895 | * before the load from ->gpnum. | |
2896 | */ | |
2897 | smp_mb(); /* ^^^ */ | |
2898 | ||
2899 | /* | |
2900 | * Make sure this load happens before the purportedly | |
2901 | * time-consuming work between get_state_synchronize_rcu() | |
2902 | * and cond_synchronize_rcu(). | |
2903 | */ | |
e534165b | 2904 | return smp_load_acquire(&rcu_state_p->gpnum); |
765a3f4f PM |
2905 | } |
2906 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
2907 | ||
2908 | /** | |
2909 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
2910 | * | |
2911 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
2912 | * | |
2913 | * If a full RCU grace period has elapsed since the earlier call to | |
2914 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
2915 | * synchronize_rcu() to wait for a full grace period. | |
2916 | * | |
2917 | * Yes, this function does not take counter wrap into account. But | |
2918 | * counter wrap is harmless. If the counter wraps, we have waited for | |
2919 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
2920 | * so waiting for one additional grace period should be just fine. | |
2921 | */ | |
2922 | void cond_synchronize_rcu(unsigned long oldstate) | |
2923 | { | |
2924 | unsigned long newstate; | |
2925 | ||
2926 | /* | |
2927 | * Ensure that this load happens before any RCU-destructive | |
2928 | * actions the caller might carry out after we return. | |
2929 | */ | |
e534165b | 2930 | newstate = smp_load_acquire(&rcu_state_p->completed); |
765a3f4f PM |
2931 | if (ULONG_CMP_GE(oldstate, newstate)) |
2932 | synchronize_rcu(); | |
2933 | } | |
2934 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
2935 | ||
3d3b7db0 PM |
2936 | static int synchronize_sched_expedited_cpu_stop(void *data) |
2937 | { | |
2938 | /* | |
2939 | * There must be a full memory barrier on each affected CPU | |
2940 | * between the time that try_stop_cpus() is called and the | |
2941 | * time that it returns. | |
2942 | * | |
2943 | * In the current initial implementation of cpu_stop, the | |
2944 | * above condition is already met when the control reaches | |
2945 | * this point and the following smp_mb() is not strictly | |
2946 | * necessary. Do smp_mb() anyway for documentation and | |
2947 | * robustness against future implementation changes. | |
2948 | */ | |
2949 | smp_mb(); /* See above comment block. */ | |
2950 | return 0; | |
2951 | } | |
2952 | ||
236fefaf PM |
2953 | /** |
2954 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
2955 | * | |
2956 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
2957 | * approach to force the grace period to end quickly. This consumes | |
2958 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
2959 | * so is thus not recommended for any sort of common-case code. In fact, | |
2960 | * if you are using synchronize_sched_expedited() in a loop, please | |
2961 | * restructure your code to batch your updates, and then use a single | |
2962 | * synchronize_sched() instead. | |
3d3b7db0 | 2963 | * |
3d3b7db0 PM |
2964 | * This implementation can be thought of as an application of ticket |
2965 | * locking to RCU, with sync_sched_expedited_started and | |
2966 | * sync_sched_expedited_done taking on the roles of the halves | |
2967 | * of the ticket-lock word. Each task atomically increments | |
2968 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
2969 | * then attempts to stop all the CPUs. If this succeeds, then each | |
2970 | * CPU will have executed a context switch, resulting in an RCU-sched | |
2971 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
2972 | * update sync_sched_expedited_done to match our snapshot -- but | |
2973 | * only if someone else has not already advanced past our snapshot. | |
2974 | * | |
2975 | * On the other hand, if try_stop_cpus() fails, we check the value | |
2976 | * of sync_sched_expedited_done. If it has advanced past our | |
2977 | * initial snapshot, then someone else must have forced a grace period | |
2978 | * some time after we took our snapshot. In this case, our work is | |
2979 | * done for us, and we can simply return. Otherwise, we try again, | |
2980 | * but keep our initial snapshot for purposes of checking for someone | |
2981 | * doing our work for us. | |
2982 | * | |
2983 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
2984 | */ | |
2985 | void synchronize_sched_expedited(void) | |
2986 | { | |
e0775cef PM |
2987 | cpumask_var_t cm; |
2988 | bool cma = false; | |
2989 | int cpu; | |
1924bcb0 PM |
2990 | long firstsnap, s, snap; |
2991 | int trycount = 0; | |
40694d66 | 2992 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 2993 | |
1924bcb0 PM |
2994 | /* |
2995 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
2996 | * By allowing sync_sched_expedited_started to advance no more than | |
2997 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
2998 | * that more than 3.5 billion CPUs would be required to force a | |
2999 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
3000 | * course be required on a 64-bit system. | |
3001 | */ | |
40694d66 PM |
3002 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
3003 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 PM |
3004 | ULONG_MAX / 8)) { |
3005 | synchronize_sched(); | |
a30489c5 | 3006 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
3007 | return; |
3008 | } | |
3d3b7db0 | 3009 | |
1924bcb0 PM |
3010 | /* |
3011 | * Take a ticket. Note that atomic_inc_return() implies a | |
3012 | * full memory barrier. | |
3013 | */ | |
40694d66 | 3014 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 3015 | firstsnap = snap; |
dd56af42 PM |
3016 | if (!try_get_online_cpus()) { |
3017 | /* CPU hotplug operation in flight, fall back to normal GP. */ | |
3018 | wait_rcu_gp(call_rcu_sched); | |
3019 | atomic_long_inc(&rsp->expedited_normal); | |
3020 | return; | |
3021 | } | |
1cc85961 | 3022 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 | 3023 | |
e0775cef PM |
3024 | /* Offline CPUs, idle CPUs, and any CPU we run on are quiescent. */ |
3025 | cma = zalloc_cpumask_var(&cm, GFP_KERNEL); | |
3026 | if (cma) { | |
3027 | cpumask_copy(cm, cpu_online_mask); | |
3028 | cpumask_clear_cpu(raw_smp_processor_id(), cm); | |
3029 | for_each_cpu(cpu, cm) { | |
3030 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); | |
3031 | ||
3032 | if (!(atomic_add_return(0, &rdtp->dynticks) & 0x1)) | |
3033 | cpumask_clear_cpu(cpu, cm); | |
3034 | } | |
3035 | if (cpumask_weight(cm) == 0) | |
3036 | goto all_cpus_idle; | |
3037 | } | |
3038 | ||
3d3b7db0 PM |
3039 | /* |
3040 | * Each pass through the following loop attempts to force a | |
3041 | * context switch on each CPU. | |
3042 | */ | |
e0775cef | 3043 | while (try_stop_cpus(cma ? cm : cpu_online_mask, |
3d3b7db0 PM |
3044 | synchronize_sched_expedited_cpu_stop, |
3045 | NULL) == -EAGAIN) { | |
3046 | put_online_cpus(); | |
a30489c5 | 3047 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 3048 | |
1924bcb0 | 3049 | /* Check to see if someone else did our work for us. */ |
40694d66 | 3050 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3051 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 | 3052 | /* ensure test happens before caller kfree */ |
4e857c58 | 3053 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3054 | atomic_long_inc(&rsp->expedited_workdone1); |
e0775cef | 3055 | free_cpumask_var(cm); |
1924bcb0 PM |
3056 | return; |
3057 | } | |
3d3b7db0 PM |
3058 | |
3059 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 3060 | if (trycount++ < 10) { |
3d3b7db0 | 3061 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 3062 | } else { |
3705b88d | 3063 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 3064 | atomic_long_inc(&rsp->expedited_normal); |
e0775cef | 3065 | free_cpumask_var(cm); |
3d3b7db0 PM |
3066 | return; |
3067 | } | |
3068 | ||
1924bcb0 | 3069 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 3070 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3071 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 | 3072 | /* ensure test happens before caller kfree */ |
4e857c58 | 3073 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3074 | atomic_long_inc(&rsp->expedited_workdone2); |
e0775cef | 3075 | free_cpumask_var(cm); |
3d3b7db0 PM |
3076 | return; |
3077 | } | |
3078 | ||
3079 | /* | |
3080 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
3081 | * callers to piggyback on our grace period. We retry |
3082 | * after they started, so our grace period works for them, | |
3083 | * and they started after our first try, so their grace | |
3084 | * period works for us. | |
3d3b7db0 | 3085 | */ |
dd56af42 PM |
3086 | if (!try_get_online_cpus()) { |
3087 | /* CPU hotplug operation in flight, use normal GP. */ | |
3088 | wait_rcu_gp(call_rcu_sched); | |
3089 | atomic_long_inc(&rsp->expedited_normal); | |
e0775cef | 3090 | free_cpumask_var(cm); |
dd56af42 PM |
3091 | return; |
3092 | } | |
40694d66 | 3093 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
3094 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
3095 | } | |
a30489c5 | 3096 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 | 3097 | |
e0775cef PM |
3098 | all_cpus_idle: |
3099 | free_cpumask_var(cm); | |
3100 | ||
3d3b7db0 PM |
3101 | /* |
3102 | * Everyone up to our most recent fetch is covered by our grace | |
3103 | * period. Update the counter, but only if our work is still | |
3104 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 3105 | * than we did already did their update. |
3d3b7db0 PM |
3106 | */ |
3107 | do { | |
a30489c5 | 3108 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 3109 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3110 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 | 3111 | /* ensure test happens before caller kfree */ |
4e857c58 | 3112 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3113 | atomic_long_inc(&rsp->expedited_done_lost); |
3d3b7db0 PM |
3114 | break; |
3115 | } | |
40694d66 | 3116 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 3117 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
3118 | |
3119 | put_online_cpus(); | |
3120 | } | |
3121 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
3122 | ||
64db4cff PM |
3123 | /* |
3124 | * Check to see if there is any immediate RCU-related work to be done | |
3125 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
3126 | * The checks are in order of increasing expense: checks that can be | |
3127 | * carried out against CPU-local state are performed first. However, | |
3128 | * we must check for CPU stalls first, else we might not get a chance. | |
3129 | */ | |
3130 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
3131 | { | |
2f51f988 PM |
3132 | struct rcu_node *rnp = rdp->mynode; |
3133 | ||
64db4cff PM |
3134 | rdp->n_rcu_pending++; |
3135 | ||
3136 | /* Check for CPU stalls, if enabled. */ | |
3137 | check_cpu_stall(rsp, rdp); | |
3138 | ||
a096932f PM |
3139 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ |
3140 | if (rcu_nohz_full_cpu(rsp)) | |
3141 | return 0; | |
3142 | ||
64db4cff | 3143 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
5c51dd73 PM |
3144 | if (rcu_scheduler_fully_active && |
3145 | rdp->qs_pending && !rdp->passed_quiesce) { | |
d21670ac | 3146 | rdp->n_rp_qs_pending++; |
e4cc1f22 | 3147 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
d21670ac | 3148 | rdp->n_rp_report_qs++; |
64db4cff | 3149 | return 1; |
7ba5c840 | 3150 | } |
64db4cff PM |
3151 | |
3152 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
3153 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
3154 | rdp->n_rp_cb_ready++; | |
64db4cff | 3155 | return 1; |
7ba5c840 | 3156 | } |
64db4cff PM |
3157 | |
3158 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
3159 | if (cpu_needs_another_gp(rsp, rdp)) { |
3160 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 3161 | return 1; |
7ba5c840 | 3162 | } |
64db4cff PM |
3163 | |
3164 | /* Has another RCU grace period completed? */ | |
2f51f988 | 3165 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 3166 | rdp->n_rp_gp_completed++; |
64db4cff | 3167 | return 1; |
7ba5c840 | 3168 | } |
64db4cff PM |
3169 | |
3170 | /* Has a new RCU grace period started? */ | |
2f51f988 | 3171 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
7ba5c840 | 3172 | rdp->n_rp_gp_started++; |
64db4cff | 3173 | return 1; |
7ba5c840 | 3174 | } |
64db4cff | 3175 | |
96d3fd0d PM |
3176 | /* Does this CPU need a deferred NOCB wakeup? */ |
3177 | if (rcu_nocb_need_deferred_wakeup(rdp)) { | |
3178 | rdp->n_rp_nocb_defer_wakeup++; | |
3179 | return 1; | |
3180 | } | |
3181 | ||
64db4cff | 3182 | /* nothing to do */ |
7ba5c840 | 3183 | rdp->n_rp_need_nothing++; |
64db4cff PM |
3184 | return 0; |
3185 | } | |
3186 | ||
3187 | /* | |
3188 | * Check to see if there is any immediate RCU-related work to be done | |
3189 | * by the current CPU, returning 1 if so. This function is part of the | |
3190 | * RCU implementation; it is -not- an exported member of the RCU API. | |
3191 | */ | |
e3950ecd | 3192 | static int rcu_pending(void) |
64db4cff | 3193 | { |
6ce75a23 PM |
3194 | struct rcu_state *rsp; |
3195 | ||
3196 | for_each_rcu_flavor(rsp) | |
e3950ecd | 3197 | if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda))) |
6ce75a23 PM |
3198 | return 1; |
3199 | return 0; | |
64db4cff PM |
3200 | } |
3201 | ||
3202 | /* | |
c0f4dfd4 PM |
3203 | * Return true if the specified CPU has any callback. If all_lazy is |
3204 | * non-NULL, store an indication of whether all callbacks are lazy. | |
3205 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 3206 | */ |
aa6da514 | 3207 | static int __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy) |
64db4cff | 3208 | { |
c0f4dfd4 PM |
3209 | bool al = true; |
3210 | bool hc = false; | |
3211 | struct rcu_data *rdp; | |
6ce75a23 PM |
3212 | struct rcu_state *rsp; |
3213 | ||
c0f4dfd4 | 3214 | for_each_rcu_flavor(rsp) { |
aa6da514 | 3215 | rdp = this_cpu_ptr(rsp->rda); |
69c8d28c PM |
3216 | if (!rdp->nxtlist) |
3217 | continue; | |
3218 | hc = true; | |
3219 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
c0f4dfd4 | 3220 | al = false; |
69c8d28c PM |
3221 | break; |
3222 | } | |
c0f4dfd4 PM |
3223 | } |
3224 | if (all_lazy) | |
3225 | *all_lazy = al; | |
3226 | return hc; | |
64db4cff PM |
3227 | } |
3228 | ||
a83eff0a PM |
3229 | /* |
3230 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
3231 | * the compiler is expected to optimize this away. | |
3232 | */ | |
e66c33d5 | 3233 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
3234 | int cpu, unsigned long done) |
3235 | { | |
3236 | trace_rcu_barrier(rsp->name, s, cpu, | |
3237 | atomic_read(&rsp->barrier_cpu_count), done); | |
3238 | } | |
3239 | ||
b1420f1c PM |
3240 | /* |
3241 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
3242 | * up the task executing _rcu_barrier(). | |
3243 | */ | |
24ebbca8 | 3244 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3245 | { |
24ebbca8 PM |
3246 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
3247 | struct rcu_state *rsp = rdp->rsp; | |
3248 | ||
a83eff0a PM |
3249 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
3250 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 3251 | complete(&rsp->barrier_completion); |
a83eff0a PM |
3252 | } else { |
3253 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
3254 | } | |
d0ec774c PM |
3255 | } |
3256 | ||
3257 | /* | |
3258 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3259 | */ | |
3260 | static void rcu_barrier_func(void *type) | |
3261 | { | |
037b64ed | 3262 | struct rcu_state *rsp = type; |
fa07a58f | 3263 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
d0ec774c | 3264 | |
a83eff0a | 3265 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 3266 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 3267 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
3268 | } |
3269 | ||
d0ec774c PM |
3270 | /* |
3271 | * Orchestrate the specified type of RCU barrier, waiting for all | |
3272 | * RCU callbacks of the specified type to complete. | |
3273 | */ | |
037b64ed | 3274 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 3275 | { |
b1420f1c | 3276 | int cpu; |
b1420f1c | 3277 | struct rcu_data *rdp; |
cf3a9c48 PM |
3278 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
3279 | unsigned long snap_done; | |
b1420f1c | 3280 | |
a83eff0a | 3281 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 3282 | |
e74f4c45 | 3283 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 3284 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 3285 | |
cf3a9c48 PM |
3286 | /* |
3287 | * Ensure that all prior references, including to ->n_barrier_done, | |
3288 | * are ordered before the _rcu_barrier() machinery. | |
3289 | */ | |
3290 | smp_mb(); /* See above block comment. */ | |
3291 | ||
3292 | /* | |
3293 | * Recheck ->n_barrier_done to see if others did our work for us. | |
3294 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
3295 | * transition. The "if" expression below therefore rounds the old | |
3296 | * value up to the next even number and adds two before comparing. | |
3297 | */ | |
458fb381 | 3298 | snap_done = rsp->n_barrier_done; |
a83eff0a | 3299 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
458fb381 PM |
3300 | |
3301 | /* | |
3302 | * If the value in snap is odd, we needed to wait for the current | |
3303 | * rcu_barrier() to complete, then wait for the next one, in other | |
3304 | * words, we need the value of snap_done to be three larger than | |
3305 | * the value of snap. On the other hand, if the value in snap is | |
3306 | * even, we only had to wait for the next rcu_barrier() to complete, | |
3307 | * in other words, we need the value of snap_done to be only two | |
3308 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes | |
3309 | * this for us (thank you, Linus!). | |
3310 | */ | |
3311 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) { | |
a83eff0a | 3312 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
3313 | smp_mb(); /* caller's subsequent code after above check. */ |
3314 | mutex_unlock(&rsp->barrier_mutex); | |
3315 | return; | |
3316 | } | |
3317 | ||
3318 | /* | |
3319 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
3320 | * ACCESS_ONCE() to prevent the compiler from speculating | |
3321 | * the increment to precede the early-exit check. | |
3322 | */ | |
a792563b | 3323 | ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1; |
cf3a9c48 | 3324 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); |
a83eff0a | 3325 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 3326 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 3327 | |
d0ec774c | 3328 | /* |
b1420f1c PM |
3329 | * Initialize the count to one rather than to zero in order to |
3330 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
3331 | * (or preemption of this task). Exclude CPU-hotplug operations |
3332 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 3333 | */ |
7db74df8 | 3334 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 3335 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 3336 | get_online_cpus(); |
b1420f1c PM |
3337 | |
3338 | /* | |
1331e7a1 PM |
3339 | * Force each CPU with callbacks to register a new callback. |
3340 | * When that callback is invoked, we will know that all of the | |
3341 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 3342 | */ |
3fbfbf7a | 3343 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 3344 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 3345 | continue; |
b1420f1c | 3346 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 3347 | if (rcu_is_nocb_cpu(cpu)) { |
d7e29933 PM |
3348 | if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) { |
3349 | _rcu_barrier_trace(rsp, "OfflineNoCB", cpu, | |
3350 | rsp->n_barrier_done); | |
3351 | } else { | |
3352 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, | |
3353 | rsp->n_barrier_done); | |
3354 | atomic_inc(&rsp->barrier_cpu_count); | |
3355 | __call_rcu(&rdp->barrier_head, | |
3356 | rcu_barrier_callback, rsp, cpu, 0); | |
3357 | } | |
3fbfbf7a | 3358 | } else if (ACCESS_ONCE(rdp->qlen)) { |
a83eff0a PM |
3359 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
3360 | rsp->n_barrier_done); | |
037b64ed | 3361 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 3362 | } else { |
a83eff0a PM |
3363 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
3364 | rsp->n_barrier_done); | |
b1420f1c PM |
3365 | } |
3366 | } | |
1331e7a1 | 3367 | put_online_cpus(); |
b1420f1c PM |
3368 | |
3369 | /* | |
3370 | * Now that we have an rcu_barrier_callback() callback on each | |
3371 | * CPU, and thus each counted, remove the initial count. | |
3372 | */ | |
24ebbca8 | 3373 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 3374 | complete(&rsp->barrier_completion); |
b1420f1c | 3375 | |
cf3a9c48 PM |
3376 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
3377 | smp_mb(); /* Keep increment after above mechanism. */ | |
a792563b | 3378 | ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1; |
cf3a9c48 | 3379 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); |
a83eff0a | 3380 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
3381 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
3382 | ||
b1420f1c | 3383 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 3384 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
3385 | |
3386 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 3387 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 3388 | } |
d0ec774c PM |
3389 | |
3390 | /** | |
3391 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
3392 | */ | |
3393 | void rcu_barrier_bh(void) | |
3394 | { | |
037b64ed | 3395 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
3396 | } |
3397 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
3398 | ||
3399 | /** | |
3400 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
3401 | */ | |
3402 | void rcu_barrier_sched(void) | |
3403 | { | |
037b64ed | 3404 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
3405 | } |
3406 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
3407 | ||
64db4cff | 3408 | /* |
27569620 | 3409 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 3410 | */ |
27569620 PM |
3411 | static void __init |
3412 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
3413 | { |
3414 | unsigned long flags; | |
394f99a9 | 3415 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
3416 | struct rcu_node *rnp = rcu_get_root(rsp); |
3417 | ||
3418 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 3419 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 3420 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
3f5d3ea6 | 3421 | init_callback_list(rdp); |
486e2593 | 3422 | rdp->qlen_lazy = 0; |
1d1fb395 | 3423 | ACCESS_ONCE(rdp->qlen) = 0; |
27569620 | 3424 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 3425 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 3426 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 3427 | rdp->cpu = cpu; |
d4c08f2a | 3428 | rdp->rsp = rsp; |
3fbfbf7a | 3429 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 3430 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
3431 | } |
3432 | ||
3433 | /* | |
3434 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
3435 | * offline event can be happening at a given time. Note also that we | |
3436 | * can accept some slop in the rsp->completed access due to the fact | |
3437 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 3438 | */ |
49fb4c62 | 3439 | static void |
9b67122a | 3440 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp) |
64db4cff PM |
3441 | { |
3442 | unsigned long flags; | |
64db4cff | 3443 | unsigned long mask; |
394f99a9 | 3444 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
3445 | struct rcu_node *rnp = rcu_get_root(rsp); |
3446 | ||
a4fbe35a PM |
3447 | /* Exclude new grace periods. */ |
3448 | mutex_lock(&rsp->onoff_mutex); | |
3449 | ||
64db4cff | 3450 | /* Set up local state, ensuring consistent view of global state. */ |
1304afb2 | 3451 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 3452 | rdp->beenonline = 1; /* We have now been online. */ |
37c72e56 PM |
3453 | rdp->qlen_last_fqs_check = 0; |
3454 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 3455 | rdp->blimit = blimit; |
0d8ee37e | 3456 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ |
29e37d81 | 3457 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 3458 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
3459 | atomic_set(&rdp->dynticks->dynticks, |
3460 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 3461 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 3462 | |
64db4cff PM |
3463 | /* Add CPU to rcu_node bitmasks. */ |
3464 | rnp = rdp->mynode; | |
3465 | mask = rdp->grpmask; | |
3466 | do { | |
3467 | /* Exclude any attempts to start a new GP on small systems. */ | |
1304afb2 | 3468 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3469 | rnp->qsmaskinit |= mask; |
3470 | mask = rnp->grpmask; | |
d09b62df | 3471 | if (rnp == rdp->mynode) { |
06ae115a PM |
3472 | /* |
3473 | * If there is a grace period in progress, we will | |
3474 | * set up to wait for it next time we run the | |
3475 | * RCU core code. | |
3476 | */ | |
3477 | rdp->gpnum = rnp->completed; | |
d09b62df | 3478 | rdp->completed = rnp->completed; |
06ae115a PM |
3479 | rdp->passed_quiesce = 0; |
3480 | rdp->qs_pending = 0; | |
f7f7bac9 | 3481 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); |
d09b62df | 3482 | } |
1304afb2 | 3483 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3484 | rnp = rnp->parent; |
3485 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); | |
a4fbe35a | 3486 | local_irq_restore(flags); |
64db4cff | 3487 | |
a4fbe35a | 3488 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
3489 | } |
3490 | ||
49fb4c62 | 3491 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 3492 | { |
6ce75a23 PM |
3493 | struct rcu_state *rsp; |
3494 | ||
3495 | for_each_rcu_flavor(rsp) | |
9b67122a | 3496 | rcu_init_percpu_data(cpu, rsp); |
64db4cff PM |
3497 | } |
3498 | ||
3499 | /* | |
f41d911f | 3500 | * Handle CPU online/offline notification events. |
64db4cff | 3501 | */ |
49fb4c62 | 3502 | static int rcu_cpu_notify(struct notifier_block *self, |
9f680ab4 | 3503 | unsigned long action, void *hcpu) |
64db4cff PM |
3504 | { |
3505 | long cpu = (long)hcpu; | |
e534165b | 3506 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
a26ac245 | 3507 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 3508 | struct rcu_state *rsp; |
64db4cff | 3509 | |
f7f7bac9 | 3510 | trace_rcu_utilization(TPS("Start CPU hotplug")); |
64db4cff PM |
3511 | switch (action) { |
3512 | case CPU_UP_PREPARE: | |
3513 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
3514 | rcu_prepare_cpu(cpu); |
3515 | rcu_prepare_kthreads(cpu); | |
35ce7f29 | 3516 | rcu_spawn_all_nocb_kthreads(cpu); |
a26ac245 PM |
3517 | break; |
3518 | case CPU_ONLINE: | |
0f962a5e | 3519 | case CPU_DOWN_FAILED: |
5d01bbd1 | 3520 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
3521 | break; |
3522 | case CPU_DOWN_PREPARE: | |
34ed6246 | 3523 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 3524 | break; |
d0ec774c PM |
3525 | case CPU_DYING: |
3526 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
3527 | for_each_rcu_flavor(rsp) |
3528 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 3529 | break; |
64db4cff PM |
3530 | case CPU_DEAD: |
3531 | case CPU_DEAD_FROZEN: | |
3532 | case CPU_UP_CANCELED: | |
3533 | case CPU_UP_CANCELED_FROZEN: | |
776d6807 | 3534 | for_each_rcu_flavor(rsp) { |
6ce75a23 | 3535 | rcu_cleanup_dead_cpu(cpu, rsp); |
776d6807 PM |
3536 | do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu)); |
3537 | } | |
64db4cff PM |
3538 | break; |
3539 | default: | |
3540 | break; | |
3541 | } | |
f7f7bac9 | 3542 | trace_rcu_utilization(TPS("End CPU hotplug")); |
34ed6246 | 3543 | return NOTIFY_OK; |
64db4cff PM |
3544 | } |
3545 | ||
d1d74d14 BP |
3546 | static int rcu_pm_notify(struct notifier_block *self, |
3547 | unsigned long action, void *hcpu) | |
3548 | { | |
3549 | switch (action) { | |
3550 | case PM_HIBERNATION_PREPARE: | |
3551 | case PM_SUSPEND_PREPARE: | |
3552 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
3553 | rcu_expedited = 1; | |
3554 | break; | |
3555 | case PM_POST_HIBERNATION: | |
3556 | case PM_POST_SUSPEND: | |
3557 | rcu_expedited = 0; | |
3558 | break; | |
3559 | default: | |
3560 | break; | |
3561 | } | |
3562 | return NOTIFY_OK; | |
3563 | } | |
3564 | ||
b3dbec76 | 3565 | /* |
9386c0b7 | 3566 | * Spawn the kthreads that handle each RCU flavor's grace periods. |
b3dbec76 PM |
3567 | */ |
3568 | static int __init rcu_spawn_gp_kthread(void) | |
3569 | { | |
3570 | unsigned long flags; | |
3571 | struct rcu_node *rnp; | |
3572 | struct rcu_state *rsp; | |
3573 | struct task_struct *t; | |
3574 | ||
9386c0b7 | 3575 | rcu_scheduler_fully_active = 1; |
b3dbec76 | 3576 | for_each_rcu_flavor(rsp) { |
f170168b | 3577 | t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
3578 | BUG_ON(IS_ERR(t)); |
3579 | rnp = rcu_get_root(rsp); | |
3580 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3581 | rsp->gp_kthread = t; | |
3582 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3583 | } | |
35ce7f29 | 3584 | rcu_spawn_nocb_kthreads(); |
9386c0b7 | 3585 | rcu_spawn_boost_kthreads(); |
b3dbec76 PM |
3586 | return 0; |
3587 | } | |
3588 | early_initcall(rcu_spawn_gp_kthread); | |
3589 | ||
bbad9379 PM |
3590 | /* |
3591 | * This function is invoked towards the end of the scheduler's initialization | |
3592 | * process. Before this is called, the idle task might contain | |
3593 | * RCU read-side critical sections (during which time, this idle | |
3594 | * task is booting the system). After this function is called, the | |
3595 | * idle tasks are prohibited from containing RCU read-side critical | |
3596 | * sections. This function also enables RCU lockdep checking. | |
3597 | */ | |
3598 | void rcu_scheduler_starting(void) | |
3599 | { | |
3600 | WARN_ON(num_online_cpus() != 1); | |
3601 | WARN_ON(nr_context_switches() > 0); | |
3602 | rcu_scheduler_active = 1; | |
3603 | } | |
3604 | ||
64db4cff PM |
3605 | /* |
3606 | * Compute the per-level fanout, either using the exact fanout specified | |
3607 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
3608 | */ | |
3609 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
3610 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3611 | { | |
3612 | int i; | |
3613 | ||
04f34650 PM |
3614 | rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; |
3615 | for (i = rcu_num_lvls - 2; i >= 0; i--) | |
64db4cff PM |
3616 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
3617 | } | |
3618 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3619 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3620 | { | |
3621 | int ccur; | |
3622 | int cprv; | |
3623 | int i; | |
3624 | ||
4dbd6bb3 | 3625 | cprv = nr_cpu_ids; |
f885b7f2 | 3626 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3627 | ccur = rsp->levelcnt[i]; |
3628 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
3629 | cprv = ccur; | |
3630 | } | |
3631 | } | |
3632 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3633 | ||
3634 | /* | |
3635 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3636 | */ | |
394f99a9 LJ |
3637 | static void __init rcu_init_one(struct rcu_state *rsp, |
3638 | struct rcu_data __percpu *rda) | |
64db4cff | 3639 | { |
b4426b49 FF |
3640 | static const char * const buf[] = { |
3641 | "rcu_node_0", | |
3642 | "rcu_node_1", | |
3643 | "rcu_node_2", | |
3644 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ | |
3645 | static const char * const fqs[] = { | |
3646 | "rcu_node_fqs_0", | |
3647 | "rcu_node_fqs_1", | |
3648 | "rcu_node_fqs_2", | |
3649 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ | |
4a81e832 | 3650 | static u8 fl_mask = 0x1; |
64db4cff PM |
3651 | int cpustride = 1; |
3652 | int i; | |
3653 | int j; | |
3654 | struct rcu_node *rnp; | |
3655 | ||
b6407e86 PM |
3656 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
3657 | ||
4930521a PM |
3658 | /* Silence gcc 4.8 warning about array index out of range. */ |
3659 | if (rcu_num_lvls > RCU_NUM_LVLS) | |
3660 | panic("rcu_init_one: rcu_num_lvls overflow"); | |
3661 | ||
64db4cff PM |
3662 | /* Initialize the level-tracking arrays. */ |
3663 | ||
f885b7f2 PM |
3664 | for (i = 0; i < rcu_num_lvls; i++) |
3665 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
3666 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
3667 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
3668 | rcu_init_levelspread(rsp); | |
4a81e832 PM |
3669 | rsp->flavor_mask = fl_mask; |
3670 | fl_mask <<= 1; | |
64db4cff PM |
3671 | |
3672 | /* Initialize the elements themselves, starting from the leaves. */ | |
3673 | ||
f885b7f2 | 3674 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3675 | cpustride *= rsp->levelspread[i]; |
3676 | rnp = rsp->level[i]; | |
3677 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 3678 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
3679 | lockdep_set_class_and_name(&rnp->lock, |
3680 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
3681 | raw_spin_lock_init(&rnp->fqslock); |
3682 | lockdep_set_class_and_name(&rnp->fqslock, | |
3683 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
3684 | rnp->gpnum = rsp->gpnum; |
3685 | rnp->completed = rsp->completed; | |
64db4cff PM |
3686 | rnp->qsmask = 0; |
3687 | rnp->qsmaskinit = 0; | |
3688 | rnp->grplo = j * cpustride; | |
3689 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
3690 | if (rnp->grphi >= nr_cpu_ids) |
3691 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
3692 | if (i == 0) { |
3693 | rnp->grpnum = 0; | |
3694 | rnp->grpmask = 0; | |
3695 | rnp->parent = NULL; | |
3696 | } else { | |
3697 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
3698 | rnp->grpmask = 1UL << rnp->grpnum; | |
3699 | rnp->parent = rsp->level[i - 1] + | |
3700 | j / rsp->levelspread[i - 1]; | |
3701 | } | |
3702 | rnp->level = i; | |
12f5f524 | 3703 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 3704 | rcu_init_one_nocb(rnp); |
64db4cff PM |
3705 | } |
3706 | } | |
0c34029a | 3707 | |
394f99a9 | 3708 | rsp->rda = rda; |
b3dbec76 | 3709 | init_waitqueue_head(&rsp->gp_wq); |
f885b7f2 | 3710 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 3711 | for_each_possible_cpu(i) { |
4a90a068 | 3712 | while (i > rnp->grphi) |
0c34029a | 3713 | rnp++; |
394f99a9 | 3714 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
3715 | rcu_boot_init_percpu_data(i, rsp); |
3716 | } | |
6ce75a23 | 3717 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
3718 | } |
3719 | ||
f885b7f2 PM |
3720 | /* |
3721 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 3722 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
3723 | * the ->node array in the rcu_state structure. |
3724 | */ | |
3725 | static void __init rcu_init_geometry(void) | |
3726 | { | |
026ad283 | 3727 | ulong d; |
f885b7f2 PM |
3728 | int i; |
3729 | int j; | |
cca6f393 | 3730 | int n = nr_cpu_ids; |
f885b7f2 PM |
3731 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
3732 | ||
026ad283 PM |
3733 | /* |
3734 | * Initialize any unspecified boot parameters. | |
3735 | * The default values of jiffies_till_first_fqs and | |
3736 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
3737 | * value, which is a function of HZ, then adding one for each | |
3738 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
3739 | */ | |
3740 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
3741 | if (jiffies_till_first_fqs == ULONG_MAX) | |
3742 | jiffies_till_first_fqs = d; | |
3743 | if (jiffies_till_next_fqs == ULONG_MAX) | |
3744 | jiffies_till_next_fqs = d; | |
3745 | ||
f885b7f2 | 3746 | /* If the compile-time values are accurate, just leave. */ |
b17c7035 PM |
3747 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
3748 | nr_cpu_ids == NR_CPUS) | |
f885b7f2 | 3749 | return; |
39479098 PM |
3750 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", |
3751 | rcu_fanout_leaf, nr_cpu_ids); | |
f885b7f2 PM |
3752 | |
3753 | /* | |
3754 | * Compute number of nodes that can be handled an rcu_node tree | |
3755 | * with the given number of levels. Setting rcu_capacity[0] makes | |
3756 | * some of the arithmetic easier. | |
3757 | */ | |
3758 | rcu_capacity[0] = 1; | |
3759 | rcu_capacity[1] = rcu_fanout_leaf; | |
3760 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
3761 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
3762 | ||
3763 | /* | |
3764 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
3765 | * to increase the leaf-level fanout, not decrease it. Of course, | |
3766 | * the leaf-level fanout cannot exceed the number of bits in | |
3767 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
3768 | * the configured number of CPUs. Complain and fall back to the | |
3769 | * compile-time values if these limits are exceeded. | |
3770 | */ | |
3771 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
3772 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
3773 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
3774 | WARN_ON(1); | |
3775 | return; | |
3776 | } | |
3777 | ||
3778 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
3779 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
3780 | if (n <= rcu_capacity[i]) { | |
3781 | for (j = 0; j <= i; j++) | |
3782 | num_rcu_lvl[j] = | |
3783 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
3784 | rcu_num_lvls = i; | |
3785 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
3786 | num_rcu_lvl[j] = 0; | |
3787 | break; | |
3788 | } | |
3789 | ||
3790 | /* Calculate the total number of rcu_node structures. */ | |
3791 | rcu_num_nodes = 0; | |
3792 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
3793 | rcu_num_nodes += num_rcu_lvl[i]; | |
3794 | rcu_num_nodes -= n; | |
3795 | } | |
3796 | ||
9f680ab4 | 3797 | void __init rcu_init(void) |
64db4cff | 3798 | { |
017c4261 | 3799 | int cpu; |
9f680ab4 | 3800 | |
f41d911f | 3801 | rcu_bootup_announce(); |
f885b7f2 | 3802 | rcu_init_geometry(); |
394f99a9 | 3803 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); |
69c8d28c | 3804 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
f41d911f | 3805 | __rcu_init_preempt(); |
b5b39360 | 3806 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
3807 | |
3808 | /* | |
3809 | * We don't need protection against CPU-hotplug here because | |
3810 | * this is called early in boot, before either interrupts | |
3811 | * or the scheduler are operational. | |
3812 | */ | |
3813 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 3814 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
3815 | for_each_online_cpu(cpu) |
3816 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
aa23c6fb PK |
3817 | |
3818 | rcu_early_boot_tests(); | |
64db4cff PM |
3819 | } |
3820 | ||
4102adab | 3821 | #include "tree_plugin.h" |