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 SRRH |
107 | }; \ |
108 | DEFINE_PER_CPU(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 | ||
a46e0899 PM |
155 | #ifdef CONFIG_RCU_BOOST |
156 | ||
a26ac245 PM |
157 | /* |
158 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
159 | * handle all flavors of RCU. | |
160 | */ | |
161 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
d71df90e | 162 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); |
5ece5bab | 163 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); |
d71df90e | 164 | DEFINE_PER_CPU(char, rcu_cpu_has_work); |
a26ac245 | 165 | |
a46e0899 PM |
166 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
167 | ||
5d01bbd1 | 168 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
169 | static void invoke_rcu_core(void); |
170 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 171 | |
4a298656 PM |
172 | /* |
173 | * Track the rcutorture test sequence number and the update version | |
174 | * number within a given test. The rcutorture_testseq is incremented | |
175 | * on every rcutorture module load and unload, so has an odd value | |
176 | * when a test is running. The rcutorture_vernum is set to zero | |
177 | * when rcutorture starts and is incremented on each rcutorture update. | |
178 | * These variables enable correlating rcutorture output with the | |
179 | * RCU tracing information. | |
180 | */ | |
181 | unsigned long rcutorture_testseq; | |
182 | unsigned long rcutorture_vernum; | |
183 | ||
fc2219d4 PM |
184 | /* |
185 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
186 | * permit this function to be invoked without holding the root rcu_node | |
187 | * structure's ->lock, but of course results can be subject to change. | |
188 | */ | |
189 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
190 | { | |
191 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
192 | } | |
193 | ||
b1f77b05 | 194 | /* |
d6714c22 | 195 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 196 | * how many quiescent states passed, just if there was at least |
d6714c22 | 197 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 198 | * The caller must have disabled preemption. |
b1f77b05 | 199 | */ |
d6714c22 | 200 | void rcu_sched_qs(int cpu) |
b1f77b05 | 201 | { |
25502a6c | 202 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu); |
f41d911f | 203 | |
e4cc1f22 | 204 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 205 | trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 206 | rdp->passed_quiesce = 1; |
b1f77b05 IM |
207 | } |
208 | ||
d6714c22 | 209 | void rcu_bh_qs(int cpu) |
b1f77b05 | 210 | { |
25502a6c | 211 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu); |
f41d911f | 212 | |
e4cc1f22 | 213 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 214 | trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 215 | rdp->passed_quiesce = 1; |
b1f77b05 | 216 | } |
64db4cff | 217 | |
4a81e832 PM |
218 | static DEFINE_PER_CPU(int, rcu_sched_qs_mask); |
219 | ||
220 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { | |
221 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, | |
222 | .dynticks = ATOMIC_INIT(1), | |
223 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
224 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
225 | .dynticks_idle = ATOMIC_INIT(1), | |
226 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
227 | }; | |
228 | ||
229 | /* | |
230 | * Let the RCU core know that this CPU has gone through the scheduler, | |
231 | * which is a quiescent state. This is called when the need for a | |
232 | * quiescent state is urgent, so we burn an atomic operation and full | |
233 | * memory barriers to let the RCU core know about it, regardless of what | |
234 | * this CPU might (or might not) do in the near future. | |
235 | * | |
236 | * We inform the RCU core by emulating a zero-duration dyntick-idle | |
237 | * period, which we in turn do by incrementing the ->dynticks counter | |
238 | * by two. | |
239 | */ | |
240 | static void rcu_momentary_dyntick_idle(void) | |
241 | { | |
242 | unsigned long flags; | |
243 | struct rcu_data *rdp; | |
244 | struct rcu_dynticks *rdtp; | |
245 | int resched_mask; | |
246 | struct rcu_state *rsp; | |
247 | ||
248 | local_irq_save(flags); | |
249 | ||
250 | /* | |
251 | * Yes, we can lose flag-setting operations. This is OK, because | |
252 | * the flag will be set again after some delay. | |
253 | */ | |
254 | resched_mask = raw_cpu_read(rcu_sched_qs_mask); | |
255 | raw_cpu_write(rcu_sched_qs_mask, 0); | |
256 | ||
257 | /* Find the flavor that needs a quiescent state. */ | |
258 | for_each_rcu_flavor(rsp) { | |
259 | rdp = raw_cpu_ptr(rsp->rda); | |
260 | if (!(resched_mask & rsp->flavor_mask)) | |
261 | continue; | |
262 | smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */ | |
263 | if (ACCESS_ONCE(rdp->mynode->completed) != | |
264 | ACCESS_ONCE(rdp->cond_resched_completed)) | |
265 | continue; | |
266 | ||
267 | /* | |
268 | * Pretend to be momentarily idle for the quiescent state. | |
269 | * This allows the grace-period kthread to record the | |
270 | * quiescent state, with no need for this CPU to do anything | |
271 | * further. | |
272 | */ | |
273 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
274 | smp_mb__before_atomic(); /* Earlier stuff before QS. */ | |
275 | atomic_add(2, &rdtp->dynticks); /* QS. */ | |
276 | smp_mb__after_atomic(); /* Later stuff after QS. */ | |
277 | break; | |
278 | } | |
279 | local_irq_restore(flags); | |
280 | } | |
281 | ||
25502a6c PM |
282 | /* |
283 | * Note a context switch. This is a quiescent state for RCU-sched, | |
284 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 285 | * The caller must have disabled preemption. |
25502a6c PM |
286 | */ |
287 | void rcu_note_context_switch(int cpu) | |
288 | { | |
f7f7bac9 | 289 | trace_rcu_utilization(TPS("Start context switch")); |
25502a6c | 290 | rcu_sched_qs(cpu); |
cba6d0d6 | 291 | rcu_preempt_note_context_switch(cpu); |
4a81e832 PM |
292 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) |
293 | rcu_momentary_dyntick_idle(); | |
f7f7bac9 | 294 | trace_rcu_utilization(TPS("End context switch")); |
25502a6c | 295 | } |
29ce8310 | 296 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 297 | |
878d7439 ED |
298 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
299 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
300 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 301 | |
878d7439 ED |
302 | module_param(blimit, long, 0444); |
303 | module_param(qhimark, long, 0444); | |
304 | module_param(qlowmark, long, 0444); | |
3d76c082 | 305 | |
026ad283 PM |
306 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
307 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
308 | |
309 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
310 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
311 | ||
4a81e832 PM |
312 | /* |
313 | * How long the grace period must be before we start recruiting | |
314 | * quiescent-state help from rcu_note_context_switch(). | |
315 | */ | |
316 | static ulong jiffies_till_sched_qs = HZ / 20; | |
317 | module_param(jiffies_till_sched_qs, ulong, 0644); | |
318 | ||
48a7639c | 319 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
910ee45d | 320 | struct rcu_data *rdp); |
217af2a2 PM |
321 | static void force_qs_rnp(struct rcu_state *rsp, |
322 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
323 | unsigned long *maxj), | |
324 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 325 | static void force_quiescent_state(struct rcu_state *rsp); |
a157229c | 326 | static int rcu_pending(int cpu); |
64db4cff PM |
327 | |
328 | /* | |
d6714c22 | 329 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
64db4cff | 330 | */ |
d6714c22 | 331 | long rcu_batches_completed_sched(void) |
64db4cff | 332 | { |
d6714c22 | 333 | return rcu_sched_state.completed; |
64db4cff | 334 | } |
d6714c22 | 335 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
336 | |
337 | /* | |
338 | * Return the number of RCU BH batches processed thus far for debug & stats. | |
339 | */ | |
340 | long rcu_batches_completed_bh(void) | |
341 | { | |
342 | return rcu_bh_state.completed; | |
343 | } | |
344 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
345 | ||
a381d757 ACB |
346 | /* |
347 | * Force a quiescent state. | |
348 | */ | |
349 | void rcu_force_quiescent_state(void) | |
350 | { | |
e534165b | 351 | force_quiescent_state(rcu_state_p); |
a381d757 ACB |
352 | } |
353 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
354 | ||
bf66f18e PM |
355 | /* |
356 | * Force a quiescent state for RCU BH. | |
357 | */ | |
358 | void rcu_bh_force_quiescent_state(void) | |
359 | { | |
4cdfc175 | 360 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
361 | } |
362 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
363 | ||
afea227f PM |
364 | /* |
365 | * Show the state of the grace-period kthreads. | |
366 | */ | |
367 | void show_rcu_gp_kthreads(void) | |
368 | { | |
369 | struct rcu_state *rsp; | |
370 | ||
371 | for_each_rcu_flavor(rsp) { | |
372 | pr_info("%s: wait state: %d ->state: %#lx\n", | |
373 | rsp->name, rsp->gp_state, rsp->gp_kthread->state); | |
374 | /* sched_show_task(rsp->gp_kthread); */ | |
375 | } | |
376 | } | |
377 | EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); | |
378 | ||
4a298656 PM |
379 | /* |
380 | * Record the number of times rcutorture tests have been initiated and | |
381 | * terminated. This information allows the debugfs tracing stats to be | |
382 | * correlated to the rcutorture messages, even when the rcutorture module | |
383 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
384 | * store this state in rcutorture itself. | |
385 | */ | |
386 | void rcutorture_record_test_transition(void) | |
387 | { | |
388 | rcutorture_testseq++; | |
389 | rcutorture_vernum = 0; | |
390 | } | |
391 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
392 | ||
ad0dc7f9 PM |
393 | /* |
394 | * Send along grace-period-related data for rcutorture diagnostics. | |
395 | */ | |
396 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
397 | unsigned long *gpnum, unsigned long *completed) | |
398 | { | |
399 | struct rcu_state *rsp = NULL; | |
400 | ||
401 | switch (test_type) { | |
402 | case RCU_FLAVOR: | |
e534165b | 403 | rsp = rcu_state_p; |
ad0dc7f9 PM |
404 | break; |
405 | case RCU_BH_FLAVOR: | |
406 | rsp = &rcu_bh_state; | |
407 | break; | |
408 | case RCU_SCHED_FLAVOR: | |
409 | rsp = &rcu_sched_state; | |
410 | break; | |
411 | default: | |
412 | break; | |
413 | } | |
414 | if (rsp != NULL) { | |
415 | *flags = ACCESS_ONCE(rsp->gp_flags); | |
416 | *gpnum = ACCESS_ONCE(rsp->gpnum); | |
417 | *completed = ACCESS_ONCE(rsp->completed); | |
418 | return; | |
419 | } | |
420 | *flags = 0; | |
421 | *gpnum = 0; | |
422 | *completed = 0; | |
423 | } | |
424 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
425 | ||
4a298656 PM |
426 | /* |
427 | * Record the number of writer passes through the current rcutorture test. | |
428 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
429 | * messages. | |
430 | */ | |
431 | void rcutorture_record_progress(unsigned long vernum) | |
432 | { | |
433 | rcutorture_vernum++; | |
434 | } | |
435 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
436 | ||
bf66f18e PM |
437 | /* |
438 | * Force a quiescent state for RCU-sched. | |
439 | */ | |
440 | void rcu_sched_force_quiescent_state(void) | |
441 | { | |
4cdfc175 | 442 | force_quiescent_state(&rcu_sched_state); |
bf66f18e PM |
443 | } |
444 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
445 | ||
64db4cff PM |
446 | /* |
447 | * Does the CPU have callbacks ready to be invoked? | |
448 | */ | |
449 | static int | |
450 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
451 | { | |
3fbfbf7a PM |
452 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
453 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
454 | } |
455 | ||
365187fb PM |
456 | /* |
457 | * Return the root node of the specified rcu_state structure. | |
458 | */ | |
459 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
460 | { | |
461 | return &rsp->node[0]; | |
462 | } | |
463 | ||
464 | /* | |
465 | * Is there any need for future grace periods? | |
466 | * Interrupts must be disabled. If the caller does not hold the root | |
467 | * rnp_node structure's ->lock, the results are advisory only. | |
468 | */ | |
469 | static int rcu_future_needs_gp(struct rcu_state *rsp) | |
470 | { | |
471 | struct rcu_node *rnp = rcu_get_root(rsp); | |
472 | int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1; | |
473 | int *fp = &rnp->need_future_gp[idx]; | |
474 | ||
475 | return ACCESS_ONCE(*fp); | |
476 | } | |
477 | ||
64db4cff | 478 | /* |
dc35c893 PM |
479 | * Does the current CPU require a not-yet-started grace period? |
480 | * The caller must have disabled interrupts to prevent races with | |
481 | * normal callback registry. | |
64db4cff PM |
482 | */ |
483 | static int | |
484 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
485 | { | |
dc35c893 | 486 | int i; |
3fbfbf7a | 487 | |
dc35c893 PM |
488 | if (rcu_gp_in_progress(rsp)) |
489 | return 0; /* No, a grace period is already in progress. */ | |
365187fb | 490 | if (rcu_future_needs_gp(rsp)) |
34ed6246 | 491 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
492 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
493 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
494 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
495 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
496 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
497 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
498 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), | |
499 | rdp->nxtcompleted[i])) | |
500 | return 1; /* Yes, CBs for future grace period. */ | |
501 | return 0; /* No grace period needed. */ | |
64db4cff PM |
502 | } |
503 | ||
9b2e4f18 | 504 | /* |
adf5091e | 505 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
506 | * |
507 | * If the new value of the ->dynticks_nesting counter now is zero, | |
508 | * we really have entered idle, and must do the appropriate accounting. | |
509 | * The caller must have disabled interrupts. | |
510 | */ | |
adf5091e FW |
511 | static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval, |
512 | bool user) | |
9b2e4f18 | 513 | { |
96d3fd0d PM |
514 | struct rcu_state *rsp; |
515 | struct rcu_data *rdp; | |
516 | ||
f7f7bac9 | 517 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 518 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
519 | struct task_struct *idle __maybe_unused = |
520 | idle_task(smp_processor_id()); | |
0989cb46 | 521 | |
f7f7bac9 | 522 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 523 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
524 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
525 | current->pid, current->comm, | |
526 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 527 | } |
96d3fd0d PM |
528 | for_each_rcu_flavor(rsp) { |
529 | rdp = this_cpu_ptr(rsp->rda); | |
530 | do_nocb_deferred_wakeup(rdp); | |
531 | } | |
aea1b35e | 532 | rcu_prepare_for_idle(smp_processor_id()); |
9b2e4f18 | 533 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 534 | smp_mb__before_atomic(); /* See above. */ |
9b2e4f18 | 535 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 536 | smp_mb__after_atomic(); /* Force ordering with next sojourn. */ |
9b2e4f18 | 537 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
c44e2cdd PM |
538 | |
539 | /* | |
adf5091e | 540 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
541 | * in an RCU read-side critical section. |
542 | */ | |
543 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
544 | "Illegal idle entry in RCU read-side critical section."); | |
545 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
546 | "Illegal idle entry in RCU-bh read-side critical section."); | |
547 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
548 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 549 | } |
64db4cff | 550 | |
adf5091e FW |
551 | /* |
552 | * Enter an RCU extended quiescent state, which can be either the | |
553 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 554 | */ |
adf5091e | 555 | static void rcu_eqs_enter(bool user) |
64db4cff | 556 | { |
4145fa7f | 557 | long long oldval; |
64db4cff PM |
558 | struct rcu_dynticks *rdtp; |
559 | ||
c9d4b0af | 560 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 561 | oldval = rdtp->dynticks_nesting; |
29e37d81 | 562 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
3a592405 | 563 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { |
29e37d81 | 564 | rdtp->dynticks_nesting = 0; |
3a592405 PM |
565 | rcu_eqs_enter_common(rdtp, oldval, user); |
566 | } else { | |
29e37d81 | 567 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; |
3a592405 | 568 | } |
64db4cff | 569 | } |
adf5091e FW |
570 | |
571 | /** | |
572 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
573 | * | |
574 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
575 | * read-side critical sections can occur. (Though RCU read-side | |
576 | * critical sections can occur in irq handlers in idle, a possibility | |
577 | * handled by irq_enter() and irq_exit().) | |
578 | * | |
579 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
580 | * the possibility of usermode upcalls having messed up our count | |
581 | * of interrupt nesting level during the prior busy period. | |
582 | */ | |
583 | void rcu_idle_enter(void) | |
584 | { | |
c5d900bf FW |
585 | unsigned long flags; |
586 | ||
587 | local_irq_save(flags); | |
cb349ca9 | 588 | rcu_eqs_enter(false); |
c9d4b0af | 589 | rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0); |
c5d900bf | 590 | local_irq_restore(flags); |
adf5091e | 591 | } |
8a2ecf47 | 592 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 593 | |
2b1d5024 | 594 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
595 | /** |
596 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
597 | * | |
598 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
599 | * is permitted between this call and rcu_user_exit(). This way the | |
600 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
601 | * when the CPU runs in userspace. | |
602 | */ | |
603 | void rcu_user_enter(void) | |
604 | { | |
91d1aa43 | 605 | rcu_eqs_enter(1); |
adf5091e | 606 | } |
2b1d5024 | 607 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 608 | |
9b2e4f18 PM |
609 | /** |
610 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
611 | * | |
612 | * Exit from an interrupt handler, which might possibly result in entering | |
613 | * idle mode, in other words, leaving the mode in which read-side critical | |
614 | * sections can occur. | |
64db4cff | 615 | * |
9b2e4f18 PM |
616 | * This code assumes that the idle loop never does anything that might |
617 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
618 | * architecture violates this assumption, RCU will give you what you | |
619 | * deserve, good and hard. But very infrequently and irreproducibly. | |
620 | * | |
621 | * Use things like work queues to work around this limitation. | |
622 | * | |
623 | * You have been warned. | |
64db4cff | 624 | */ |
9b2e4f18 | 625 | void rcu_irq_exit(void) |
64db4cff PM |
626 | { |
627 | unsigned long flags; | |
4145fa7f | 628 | long long oldval; |
64db4cff PM |
629 | struct rcu_dynticks *rdtp; |
630 | ||
631 | local_irq_save(flags); | |
c9d4b0af | 632 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 633 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
634 | rdtp->dynticks_nesting--; |
635 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 | 636 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 637 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 638 | else |
cb349ca9 | 639 | rcu_eqs_enter_common(rdtp, oldval, true); |
eb348b89 | 640 | rcu_sysidle_enter(rdtp, 1); |
9b2e4f18 PM |
641 | local_irq_restore(flags); |
642 | } | |
643 | ||
644 | /* | |
adf5091e | 645 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
646 | * |
647 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
648 | * we really have exited idle, and must do the appropriate accounting. | |
649 | * The caller must have disabled interrupts. | |
650 | */ | |
adf5091e FW |
651 | static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval, |
652 | int user) | |
9b2e4f18 | 653 | { |
4e857c58 | 654 | smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */ |
23b5c8fa PM |
655 | atomic_inc(&rdtp->dynticks); |
656 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
4e857c58 | 657 | smp_mb__after_atomic(); /* See above. */ |
23b5c8fa | 658 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); |
7cb92499 | 659 | rcu_cleanup_after_idle(smp_processor_id()); |
f7f7bac9 | 660 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 661 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
662 | struct task_struct *idle __maybe_unused = |
663 | idle_task(smp_processor_id()); | |
0989cb46 | 664 | |
f7f7bac9 | 665 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 666 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 667 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
668 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
669 | current->pid, current->comm, | |
670 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
671 | } |
672 | } | |
673 | ||
adf5091e FW |
674 | /* |
675 | * Exit an RCU extended quiescent state, which can be either the | |
676 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 677 | */ |
adf5091e | 678 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 679 | { |
9b2e4f18 PM |
680 | struct rcu_dynticks *rdtp; |
681 | long long oldval; | |
682 | ||
c9d4b0af | 683 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 | 684 | oldval = rdtp->dynticks_nesting; |
29e37d81 | 685 | WARN_ON_ONCE(oldval < 0); |
3a592405 | 686 | if (oldval & DYNTICK_TASK_NEST_MASK) { |
29e37d81 | 687 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; |
3a592405 | 688 | } else { |
29e37d81 | 689 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
3a592405 PM |
690 | rcu_eqs_exit_common(rdtp, oldval, user); |
691 | } | |
9b2e4f18 | 692 | } |
adf5091e FW |
693 | |
694 | /** | |
695 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
696 | * | |
697 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
698 | * read-side critical sections can occur. | |
699 | * | |
700 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
701 | * allow for the possibility of usermode upcalls messing up our count | |
702 | * of interrupt nesting level during the busy period that is just | |
703 | * now starting. | |
704 | */ | |
705 | void rcu_idle_exit(void) | |
706 | { | |
c5d900bf FW |
707 | unsigned long flags; |
708 | ||
709 | local_irq_save(flags); | |
cb349ca9 | 710 | rcu_eqs_exit(false); |
c9d4b0af | 711 | rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0); |
c5d900bf | 712 | local_irq_restore(flags); |
adf5091e | 713 | } |
8a2ecf47 | 714 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 715 | |
2b1d5024 | 716 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
717 | /** |
718 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
719 | * | |
720 | * Exit RCU idle mode while entering the kernel because it can | |
721 | * run a RCU read side critical section anytime. | |
722 | */ | |
723 | void rcu_user_exit(void) | |
724 | { | |
91d1aa43 | 725 | rcu_eqs_exit(1); |
adf5091e | 726 | } |
2b1d5024 | 727 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 728 | |
9b2e4f18 PM |
729 | /** |
730 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
731 | * | |
732 | * Enter an interrupt handler, which might possibly result in exiting | |
733 | * idle mode, in other words, entering the mode in which read-side critical | |
734 | * sections can occur. | |
735 | * | |
736 | * Note that the Linux kernel is fully capable of entering an interrupt | |
737 | * handler that it never exits, for example when doing upcalls to | |
738 | * user mode! This code assumes that the idle loop never does upcalls to | |
739 | * user mode. If your architecture does do upcalls from the idle loop (or | |
740 | * does anything else that results in unbalanced calls to the irq_enter() | |
741 | * and irq_exit() functions), RCU will give you what you deserve, good | |
742 | * and hard. But very infrequently and irreproducibly. | |
743 | * | |
744 | * Use things like work queues to work around this limitation. | |
745 | * | |
746 | * You have been warned. | |
747 | */ | |
748 | void rcu_irq_enter(void) | |
749 | { | |
750 | unsigned long flags; | |
751 | struct rcu_dynticks *rdtp; | |
752 | long long oldval; | |
753 | ||
754 | local_irq_save(flags); | |
c9d4b0af | 755 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 PM |
756 | oldval = rdtp->dynticks_nesting; |
757 | rdtp->dynticks_nesting++; | |
758 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 | 759 | if (oldval) |
f7f7bac9 | 760 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 761 | else |
cb349ca9 | 762 | rcu_eqs_exit_common(rdtp, oldval, true); |
eb348b89 | 763 | rcu_sysidle_exit(rdtp, 1); |
64db4cff | 764 | local_irq_restore(flags); |
64db4cff PM |
765 | } |
766 | ||
767 | /** | |
768 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
769 | * | |
770 | * If the CPU was idle with dynamic ticks active, and there is no | |
771 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
772 | * RCU grace-period handling know that the CPU is active. | |
773 | */ | |
774 | void rcu_nmi_enter(void) | |
775 | { | |
c9d4b0af | 776 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 777 | |
23b5c8fa PM |
778 | if (rdtp->dynticks_nmi_nesting == 0 && |
779 | (atomic_read(&rdtp->dynticks) & 0x1)) | |
64db4cff | 780 | return; |
23b5c8fa | 781 | rdtp->dynticks_nmi_nesting++; |
4e857c58 | 782 | smp_mb__before_atomic(); /* Force delay from prior write. */ |
23b5c8fa PM |
783 | atomic_inc(&rdtp->dynticks); |
784 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
4e857c58 | 785 | smp_mb__after_atomic(); /* See above. */ |
23b5c8fa | 786 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); |
64db4cff PM |
787 | } |
788 | ||
789 | /** | |
790 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
791 | * | |
792 | * If the CPU was idle with dynamic ticks active, and there is no | |
793 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
794 | * RCU grace-period handling know that the CPU is no longer active. | |
795 | */ | |
796 | void rcu_nmi_exit(void) | |
797 | { | |
c9d4b0af | 798 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 799 | |
23b5c8fa PM |
800 | if (rdtp->dynticks_nmi_nesting == 0 || |
801 | --rdtp->dynticks_nmi_nesting != 0) | |
64db4cff | 802 | return; |
23b5c8fa | 803 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 804 | smp_mb__before_atomic(); /* See above. */ |
23b5c8fa | 805 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 806 | smp_mb__after_atomic(); /* Force delay to next write. */ |
23b5c8fa | 807 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
64db4cff PM |
808 | } |
809 | ||
810 | /** | |
5c173eb8 PM |
811 | * __rcu_is_watching - are RCU read-side critical sections safe? |
812 | * | |
813 | * Return true if RCU is watching the running CPU, which means that | |
814 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
815 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
816 | * least disabled preemption. | |
817 | */ | |
9418fb20 | 818 | bool notrace __rcu_is_watching(void) |
5c173eb8 PM |
819 | { |
820 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
821 | } | |
822 | ||
823 | /** | |
824 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
64db4cff | 825 | * |
9b2e4f18 | 826 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 827 | * or NMI handler, return true. |
64db4cff | 828 | */ |
9418fb20 | 829 | bool notrace rcu_is_watching(void) |
64db4cff | 830 | { |
f534ed1f | 831 | bool ret; |
34240697 PM |
832 | |
833 | preempt_disable(); | |
5c173eb8 | 834 | ret = __rcu_is_watching(); |
34240697 PM |
835 | preempt_enable(); |
836 | return ret; | |
64db4cff | 837 | } |
5c173eb8 | 838 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 839 | |
62fde6ed | 840 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
841 | |
842 | /* | |
843 | * Is the current CPU online? Disable preemption to avoid false positives | |
844 | * that could otherwise happen due to the current CPU number being sampled, | |
845 | * this task being preempted, its old CPU being taken offline, resuming | |
846 | * on some other CPU, then determining that its old CPU is now offline. | |
847 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
848 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
849 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
850 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
851 | * offline to continue to use RCU for one jiffy after marking itself | |
852 | * offline in the cpu_online_mask. This leniency is necessary given the | |
853 | * non-atomic nature of the online and offline processing, for example, | |
854 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
855 | * notifiers. | |
856 | * | |
857 | * This is also why RCU internally marks CPUs online during the | |
858 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
859 | * |
860 | * Disable checking if in an NMI handler because we cannot safely report | |
861 | * errors from NMI handlers anyway. | |
862 | */ | |
863 | bool rcu_lockdep_current_cpu_online(void) | |
864 | { | |
2036d94a PM |
865 | struct rcu_data *rdp; |
866 | struct rcu_node *rnp; | |
c0d6d01b PM |
867 | bool ret; |
868 | ||
869 | if (in_nmi()) | |
f6f7ee9a | 870 | return true; |
c0d6d01b | 871 | preempt_disable(); |
c9d4b0af | 872 | rdp = this_cpu_ptr(&rcu_sched_data); |
2036d94a PM |
873 | rnp = rdp->mynode; |
874 | ret = (rdp->grpmask & rnp->qsmaskinit) || | |
c0d6d01b PM |
875 | !rcu_scheduler_fully_active; |
876 | preempt_enable(); | |
877 | return ret; | |
878 | } | |
879 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
880 | ||
62fde6ed | 881 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 882 | |
64db4cff | 883 | /** |
9b2e4f18 | 884 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 885 | * |
9b2e4f18 PM |
886 | * If the current CPU is idle or running at a first-level (not nested) |
887 | * interrupt from idle, return true. The caller must have at least | |
888 | * disabled preemption. | |
64db4cff | 889 | */ |
62e3cb14 | 890 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 891 | { |
c9d4b0af | 892 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
64db4cff PM |
893 | } |
894 | ||
64db4cff PM |
895 | /* |
896 | * Snapshot the specified CPU's dynticks counter so that we can later | |
897 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 898 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 899 | */ |
217af2a2 PM |
900 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
901 | bool *isidle, unsigned long *maxj) | |
64db4cff | 902 | { |
23b5c8fa | 903 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 904 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
7941dbde ACB |
905 | if ((rdp->dynticks_snap & 0x1) == 0) { |
906 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); | |
907 | return 1; | |
908 | } else { | |
909 | return 0; | |
910 | } | |
64db4cff PM |
911 | } |
912 | ||
6193c76a PM |
913 | /* |
914 | * This function really isn't for public consumption, but RCU is special in | |
915 | * that context switches can allow the state machine to make progress. | |
916 | */ | |
917 | extern void resched_cpu(int cpu); | |
918 | ||
64db4cff PM |
919 | /* |
920 | * Return true if the specified CPU has passed through a quiescent | |
921 | * state by virtue of being in or having passed through an dynticks | |
922 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 923 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 924 | */ |
217af2a2 PM |
925 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
926 | bool *isidle, unsigned long *maxj) | |
64db4cff | 927 | { |
7eb4f455 | 928 | unsigned int curr; |
4a81e832 | 929 | int *rcrmp; |
7eb4f455 | 930 | unsigned int snap; |
64db4cff | 931 | |
7eb4f455 PM |
932 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
933 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
934 | |
935 | /* | |
936 | * If the CPU passed through or entered a dynticks idle phase with | |
937 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
938 | * already acknowledged the request to pass through a quiescent | |
939 | * state. Either way, that CPU cannot possibly be in an RCU | |
940 | * read-side critical section that started before the beginning | |
941 | * of the current RCU grace period. | |
942 | */ | |
7eb4f455 | 943 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 944 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
945 | rdp->dynticks_fqs++; |
946 | return 1; | |
947 | } | |
948 | ||
a82dcc76 PM |
949 | /* |
950 | * Check for the CPU being offline, but only if the grace period | |
951 | * is old enough. We don't need to worry about the CPU changing | |
952 | * state: If we see it offline even once, it has been through a | |
953 | * quiescent state. | |
954 | * | |
955 | * The reason for insisting that the grace period be at least | |
956 | * one jiffy old is that CPUs that are not quite online and that | |
957 | * have just gone offline can still execute RCU read-side critical | |
958 | * sections. | |
959 | */ | |
960 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
961 | return 0; /* Grace period is not old enough. */ | |
962 | barrier(); | |
963 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 964 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
965 | rdp->offline_fqs++; |
966 | return 1; | |
967 | } | |
65d798f0 PM |
968 | |
969 | /* | |
4a81e832 PM |
970 | * A CPU running for an extended time within the kernel can |
971 | * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode, | |
972 | * even context-switching back and forth between a pair of | |
973 | * in-kernel CPU-bound tasks cannot advance grace periods. | |
974 | * So if the grace period is old enough, make the CPU pay attention. | |
975 | * Note that the unsynchronized assignments to the per-CPU | |
976 | * rcu_sched_qs_mask variable are safe. Yes, setting of | |
977 | * bits can be lost, but they will be set again on the next | |
978 | * force-quiescent-state pass. So lost bit sets do not result | |
979 | * in incorrect behavior, merely in a grace period lasting | |
980 | * a few jiffies longer than it might otherwise. Because | |
981 | * there are at most four threads involved, and because the | |
982 | * updates are only once every few jiffies, the probability of | |
983 | * lossage (and thus of slight grace-period extension) is | |
984 | * quite low. | |
985 | * | |
986 | * Note that if the jiffies_till_sched_qs boot/sysfs parameter | |
987 | * is set too high, we override with half of the RCU CPU stall | |
988 | * warning delay. | |
6193c76a | 989 | */ |
4a81e832 PM |
990 | rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu); |
991 | if (ULONG_CMP_GE(jiffies, | |
992 | rdp->rsp->gp_start + jiffies_till_sched_qs) || | |
cb1e78cf | 993 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { |
4a81e832 PM |
994 | if (!(ACCESS_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) { |
995 | ACCESS_ONCE(rdp->cond_resched_completed) = | |
996 | ACCESS_ONCE(rdp->mynode->completed); | |
997 | smp_mb(); /* ->cond_resched_completed before *rcrmp. */ | |
998 | ACCESS_ONCE(*rcrmp) = | |
999 | ACCESS_ONCE(*rcrmp) + rdp->rsp->flavor_mask; | |
1000 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1001 | rdp->rsp->jiffies_resched += 5; /* Enable beating. */ | |
1002 | } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { | |
1003 | /* Time to beat on that CPU again! */ | |
1004 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1005 | rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */ | |
1006 | } | |
6193c76a PM |
1007 | } |
1008 | ||
a82dcc76 | 1009 | return 0; |
64db4cff PM |
1010 | } |
1011 | ||
64db4cff PM |
1012 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
1013 | { | |
cb1e78cf | 1014 | unsigned long j = jiffies; |
6193c76a | 1015 | unsigned long j1; |
26cdfedf PM |
1016 | |
1017 | rsp->gp_start = j; | |
1018 | smp_wmb(); /* Record start time before stall time. */ | |
6193c76a | 1019 | j1 = rcu_jiffies_till_stall_check(); |
4fc5b755 | 1020 | ACCESS_ONCE(rsp->jiffies_stall) = j + j1; |
6193c76a | 1021 | rsp->jiffies_resched = j + j1 / 2; |
64db4cff PM |
1022 | } |
1023 | ||
b637a328 | 1024 | /* |
bc1dce51 | 1025 | * Dump stacks of all tasks running on stalled CPUs. |
b637a328 PM |
1026 | */ |
1027 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
1028 | { | |
1029 | int cpu; | |
1030 | unsigned long flags; | |
1031 | struct rcu_node *rnp; | |
1032 | ||
1033 | rcu_for_each_leaf_node(rsp, rnp) { | |
1034 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1035 | if (rnp->qsmask != 0) { | |
1036 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1037 | if (rnp->qsmask & (1UL << cpu)) | |
1038 | dump_cpu_task(rnp->grplo + cpu); | |
1039 | } | |
1040 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1041 | } | |
1042 | } | |
1043 | ||
64db4cff PM |
1044 | static void print_other_cpu_stall(struct rcu_state *rsp) |
1045 | { | |
1046 | int cpu; | |
1047 | long delta; | |
1048 | unsigned long flags; | |
285fe294 | 1049 | int ndetected = 0; |
64db4cff | 1050 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 1051 | long totqlen = 0; |
64db4cff PM |
1052 | |
1053 | /* Only let one CPU complain about others per time interval. */ | |
1054 | ||
1304afb2 | 1055 | raw_spin_lock_irqsave(&rnp->lock, flags); |
4fc5b755 | 1056 | delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall); |
fc2219d4 | 1057 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 1058 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1059 | return; |
1060 | } | |
4fc5b755 | 1061 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 1062 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1063 | |
8cdd32a9 PM |
1064 | /* |
1065 | * OK, time to rat on our buddy... | |
1066 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1067 | * RCU CPU stall warnings. | |
1068 | */ | |
d7f3e207 | 1069 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 1070 | rsp->name); |
a858af28 | 1071 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 1072 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 1073 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 1074 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
1075 | if (rnp->qsmask != 0) { |
1076 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1077 | if (rnp->qsmask & (1UL << cpu)) { | |
1078 | print_cpu_stall_info(rsp, | |
1079 | rnp->grplo + cpu); | |
1080 | ndetected++; | |
1081 | } | |
1082 | } | |
3acd9eb3 | 1083 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1084 | } |
a858af28 PM |
1085 | |
1086 | /* | |
1087 | * Now rat on any tasks that got kicked up to the root rcu_node | |
1088 | * due to CPU offlining. | |
1089 | */ | |
1090 | rnp = rcu_get_root(rsp); | |
1091 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
285fe294 | 1092 | ndetected += rcu_print_task_stall(rnp); |
a858af28 PM |
1093 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1094 | ||
1095 | print_cpu_stall_info_end(); | |
53bb857c PM |
1096 | for_each_possible_cpu(cpu) |
1097 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e | 1098 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", |
eee05882 | 1099 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
83ebe63e | 1100 | (long)rsp->gpnum, (long)rsp->completed, totqlen); |
9bc8b558 | 1101 | if (ndetected == 0) |
d7f3e207 | 1102 | pr_err("INFO: Stall ended before state dump start\n"); |
bc1dce51 | 1103 | else |
b637a328 | 1104 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 1105 | |
4cdfc175 | 1106 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
1107 | |
1108 | rcu_print_detail_task_stall(rsp); | |
1109 | ||
4cdfc175 | 1110 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
1111 | } |
1112 | ||
1113 | static void print_cpu_stall(struct rcu_state *rsp) | |
1114 | { | |
53bb857c | 1115 | int cpu; |
64db4cff PM |
1116 | unsigned long flags; |
1117 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 1118 | long totqlen = 0; |
64db4cff | 1119 | |
8cdd32a9 PM |
1120 | /* |
1121 | * OK, time to rat on ourselves... | |
1122 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1123 | * RCU CPU stall warnings. | |
1124 | */ | |
d7f3e207 | 1125 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
1126 | print_cpu_stall_info_begin(); |
1127 | print_cpu_stall_info(rsp, smp_processor_id()); | |
1128 | print_cpu_stall_info_end(); | |
53bb857c PM |
1129 | for_each_possible_cpu(cpu) |
1130 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e PM |
1131 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", |
1132 | jiffies - rsp->gp_start, | |
1133 | (long)rsp->gpnum, (long)rsp->completed, totqlen); | |
bc1dce51 | 1134 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 1135 | |
1304afb2 | 1136 | raw_spin_lock_irqsave(&rnp->lock, flags); |
4fc5b755 IM |
1137 | if (ULONG_CMP_GE(jiffies, ACCESS_ONCE(rsp->jiffies_stall))) |
1138 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + | |
6bfc09e2 | 1139 | 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 1140 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 1141 | |
b021fe3e PZ |
1142 | /* |
1143 | * Attempt to revive the RCU machinery by forcing a context switch. | |
1144 | * | |
1145 | * A context switch would normally allow the RCU state machine to make | |
1146 | * progress and it could be we're stuck in kernel space without context | |
1147 | * switches for an entirely unreasonable amount of time. | |
1148 | */ | |
1149 | resched_cpu(smp_processor_id()); | |
64db4cff PM |
1150 | } |
1151 | ||
1152 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
1153 | { | |
26cdfedf PM |
1154 | unsigned long completed; |
1155 | unsigned long gpnum; | |
1156 | unsigned long gps; | |
bad6e139 PM |
1157 | unsigned long j; |
1158 | unsigned long js; | |
64db4cff PM |
1159 | struct rcu_node *rnp; |
1160 | ||
26cdfedf | 1161 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
c68de209 | 1162 | return; |
cb1e78cf | 1163 | j = jiffies; |
26cdfedf PM |
1164 | |
1165 | /* | |
1166 | * Lots of memory barriers to reject false positives. | |
1167 | * | |
1168 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
1169 | * then rsp->gp_start, and finally rsp->completed. These values | |
1170 | * are updated in the opposite order with memory barriers (or | |
1171 | * equivalent) during grace-period initialization and cleanup. | |
1172 | * Now, a false positive can occur if we get an new value of | |
1173 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
1174 | * the memory barriers, the only way that this can happen is if one | |
1175 | * grace period ends and another starts between these two fetches. | |
1176 | * Detect this by comparing rsp->completed with the previous fetch | |
1177 | * from rsp->gpnum. | |
1178 | * | |
1179 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
1180 | * and rsp->gp_start suffice to forestall false positives. | |
1181 | */ | |
1182 | gpnum = ACCESS_ONCE(rsp->gpnum); | |
1183 | smp_rmb(); /* Pick up ->gpnum first... */ | |
bad6e139 | 1184 | js = ACCESS_ONCE(rsp->jiffies_stall); |
26cdfedf PM |
1185 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
1186 | gps = ACCESS_ONCE(rsp->gp_start); | |
1187 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ | |
1188 | completed = ACCESS_ONCE(rsp->completed); | |
1189 | if (ULONG_CMP_GE(completed, gpnum) || | |
1190 | ULONG_CMP_LT(j, js) || | |
1191 | ULONG_CMP_GE(gps, js)) | |
1192 | return; /* No stall or GP completed since entering function. */ | |
64db4cff | 1193 | rnp = rdp->mynode; |
c96ea7cf | 1194 | if (rcu_gp_in_progress(rsp) && |
26cdfedf | 1195 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) { |
64db4cff PM |
1196 | |
1197 | /* We haven't checked in, so go dump stack. */ | |
1198 | print_cpu_stall(rsp); | |
1199 | ||
bad6e139 PM |
1200 | } else if (rcu_gp_in_progress(rsp) && |
1201 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 1202 | |
bad6e139 | 1203 | /* They had a few time units to dump stack, so complain. */ |
64db4cff PM |
1204 | print_other_cpu_stall(rsp); |
1205 | } | |
1206 | } | |
1207 | ||
53d84e00 PM |
1208 | /** |
1209 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1210 | * | |
1211 | * Set the stall-warning timeout way off into the future, thus preventing | |
1212 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1213 | * RCU grace periods. | |
1214 | * | |
1215 | * The caller must disable hard irqs. | |
1216 | */ | |
1217 | void rcu_cpu_stall_reset(void) | |
1218 | { | |
6ce75a23 PM |
1219 | struct rcu_state *rsp; |
1220 | ||
1221 | for_each_rcu_flavor(rsp) | |
4fc5b755 | 1222 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + ULONG_MAX / 2; |
53d84e00 PM |
1223 | } |
1224 | ||
3f5d3ea6 PM |
1225 | /* |
1226 | * Initialize the specified rcu_data structure's callback list to empty. | |
1227 | */ | |
1228 | static void init_callback_list(struct rcu_data *rdp) | |
1229 | { | |
1230 | int i; | |
1231 | ||
34ed6246 PM |
1232 | if (init_nocb_callback_list(rdp)) |
1233 | return; | |
3f5d3ea6 PM |
1234 | rdp->nxtlist = NULL; |
1235 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1236 | rdp->nxttail[i] = &rdp->nxtlist; | |
1237 | } | |
1238 | ||
dc35c893 PM |
1239 | /* |
1240 | * Determine the value that ->completed will have at the end of the | |
1241 | * next subsequent grace period. This is used to tag callbacks so that | |
1242 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1243 | * been dyntick-idle for an extended period with callbacks under the | |
1244 | * influence of RCU_FAST_NO_HZ. | |
1245 | * | |
1246 | * The caller must hold rnp->lock with interrupts disabled. | |
1247 | */ | |
1248 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1249 | struct rcu_node *rnp) | |
1250 | { | |
1251 | /* | |
1252 | * If RCU is idle, we just wait for the next grace period. | |
1253 | * But we can only be sure that RCU is idle if we are looking | |
1254 | * at the root rcu_node structure -- otherwise, a new grace | |
1255 | * period might have started, but just not yet gotten around | |
1256 | * to initializing the current non-root rcu_node structure. | |
1257 | */ | |
1258 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1259 | return rnp->completed + 1; | |
1260 | ||
1261 | /* | |
1262 | * Otherwise, wait for a possible partial grace period and | |
1263 | * then the subsequent full grace period. | |
1264 | */ | |
1265 | return rnp->completed + 2; | |
1266 | } | |
1267 | ||
0446be48 PM |
1268 | /* |
1269 | * Trace-event helper function for rcu_start_future_gp() and | |
1270 | * rcu_nocb_wait_gp(). | |
1271 | */ | |
1272 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1273 | unsigned long c, const char *s) |
0446be48 PM |
1274 | { |
1275 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1276 | rnp->completed, c, rnp->level, | |
1277 | rnp->grplo, rnp->grphi, s); | |
1278 | } | |
1279 | ||
1280 | /* | |
1281 | * Start some future grace period, as needed to handle newly arrived | |
1282 | * callbacks. The required future grace periods are recorded in each | |
48a7639c PM |
1283 | * rcu_node structure's ->need_future_gp field. Returns true if there |
1284 | * is reason to awaken the grace-period kthread. | |
0446be48 PM |
1285 | * |
1286 | * The caller must hold the specified rcu_node structure's ->lock. | |
1287 | */ | |
48a7639c PM |
1288 | static bool __maybe_unused |
1289 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
1290 | unsigned long *c_out) | |
0446be48 PM |
1291 | { |
1292 | unsigned long c; | |
1293 | int i; | |
48a7639c | 1294 | bool ret = false; |
0446be48 PM |
1295 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); |
1296 | ||
1297 | /* | |
1298 | * Pick up grace-period number for new callbacks. If this | |
1299 | * grace period is already marked as needed, return to the caller. | |
1300 | */ | |
1301 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1302 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1303 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1304 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
48a7639c | 1305 | goto out; |
0446be48 PM |
1306 | } |
1307 | ||
1308 | /* | |
1309 | * If either this rcu_node structure or the root rcu_node structure | |
1310 | * believe that a grace period is in progress, then we must wait | |
1311 | * for the one following, which is in "c". Because our request | |
1312 | * will be noticed at the end of the current grace period, we don't | |
48bd8e9b PK |
1313 | * need to explicitly start one. We only do the lockless check |
1314 | * of rnp_root's fields if the current rcu_node structure thinks | |
1315 | * there is no grace period in flight, and because we hold rnp->lock, | |
1316 | * the only possible change is when rnp_root's two fields are | |
1317 | * equal, in which case rnp_root->gpnum might be concurrently | |
1318 | * incremented. But that is OK, as it will just result in our | |
1319 | * doing some extra useless work. | |
0446be48 PM |
1320 | */ |
1321 | if (rnp->gpnum != rnp->completed || | |
48bd8e9b | 1322 | ACCESS_ONCE(rnp_root->gpnum) != ACCESS_ONCE(rnp_root->completed)) { |
0446be48 | 1323 | rnp->need_future_gp[c & 0x1]++; |
f7f7bac9 | 1324 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
48a7639c | 1325 | goto out; |
0446be48 PM |
1326 | } |
1327 | ||
1328 | /* | |
1329 | * There might be no grace period in progress. If we don't already | |
1330 | * hold it, acquire the root rcu_node structure's lock in order to | |
1331 | * start one (if needed). | |
1332 | */ | |
6303b9c8 | 1333 | if (rnp != rnp_root) { |
0446be48 | 1334 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 PM |
1335 | smp_mb__after_unlock_lock(); |
1336 | } | |
0446be48 PM |
1337 | |
1338 | /* | |
1339 | * Get a new grace-period number. If there really is no grace | |
1340 | * period in progress, it will be smaller than the one we obtained | |
1341 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1342 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1343 | */ | |
1344 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1345 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1346 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1347 | rdp->nxtcompleted[i] = c; | |
1348 | ||
1349 | /* | |
1350 | * If the needed for the required grace period is already | |
1351 | * recorded, trace and leave. | |
1352 | */ | |
1353 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1354 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1355 | goto unlock_out; |
1356 | } | |
1357 | ||
1358 | /* Record the need for the future grace period. */ | |
1359 | rnp_root->need_future_gp[c & 0x1]++; | |
1360 | ||
1361 | /* If a grace period is not already in progress, start one. */ | |
1362 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1363 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1364 | } else { |
f7f7bac9 | 1365 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
48a7639c | 1366 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1367 | } |
1368 | unlock_out: | |
1369 | if (rnp != rnp_root) | |
1370 | raw_spin_unlock(&rnp_root->lock); | |
48a7639c PM |
1371 | out: |
1372 | if (c_out != NULL) | |
1373 | *c_out = c; | |
1374 | return ret; | |
0446be48 PM |
1375 | } |
1376 | ||
1377 | /* | |
1378 | * Clean up any old requests for the just-ended grace period. Also return | |
1379 | * whether any additional grace periods have been requested. Also invoke | |
1380 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1381 | * waiting for this grace period to complete. | |
1382 | */ | |
1383 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1384 | { | |
1385 | int c = rnp->completed; | |
1386 | int needmore; | |
1387 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1388 | ||
1389 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1390 | rnp->need_future_gp[c & 0x1] = 0; | |
1391 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1392 | trace_rcu_future_gp(rnp, rdp, c, |
1393 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1394 | return needmore; |
1395 | } | |
1396 | ||
48a7639c PM |
1397 | /* |
1398 | * Awaken the grace-period kthread for the specified flavor of RCU. | |
1399 | * Don't do a self-awaken, and don't bother awakening when there is | |
1400 | * nothing for the grace-period kthread to do (as in several CPUs | |
1401 | * raced to awaken, and we lost), and finally don't try to awaken | |
1402 | * a kthread that has not yet been created. | |
1403 | */ | |
1404 | static void rcu_gp_kthread_wake(struct rcu_state *rsp) | |
1405 | { | |
1406 | if (current == rsp->gp_kthread || | |
1407 | !ACCESS_ONCE(rsp->gp_flags) || | |
1408 | !rsp->gp_kthread) | |
1409 | return; | |
1410 | wake_up(&rsp->gp_wq); | |
1411 | } | |
1412 | ||
dc35c893 PM |
1413 | /* |
1414 | * If there is room, assign a ->completed number to any callbacks on | |
1415 | * this CPU that have not already been assigned. Also accelerate any | |
1416 | * callbacks that were previously assigned a ->completed number that has | |
1417 | * since proven to be too conservative, which can happen if callbacks get | |
1418 | * assigned a ->completed number while RCU is idle, but with reference to | |
1419 | * a non-root rcu_node structure. This function is idempotent, so it does | |
48a7639c PM |
1420 | * not hurt to call it repeatedly. Returns an flag saying that we should |
1421 | * awaken the RCU grace-period kthread. | |
dc35c893 PM |
1422 | * |
1423 | * The caller must hold rnp->lock with interrupts disabled. | |
1424 | */ | |
48a7639c | 1425 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1426 | struct rcu_data *rdp) |
1427 | { | |
1428 | unsigned long c; | |
1429 | int i; | |
48a7639c | 1430 | bool ret; |
dc35c893 PM |
1431 | |
1432 | /* If the CPU has no callbacks, nothing to do. */ | |
1433 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1434 | return false; |
dc35c893 PM |
1435 | |
1436 | /* | |
1437 | * Starting from the sublist containing the callbacks most | |
1438 | * recently assigned a ->completed number and working down, find the | |
1439 | * first sublist that is not assignable to an upcoming grace period. | |
1440 | * Such a sublist has something in it (first two tests) and has | |
1441 | * a ->completed number assigned that will complete sooner than | |
1442 | * the ->completed number for newly arrived callbacks (last test). | |
1443 | * | |
1444 | * The key point is that any later sublist can be assigned the | |
1445 | * same ->completed number as the newly arrived callbacks, which | |
1446 | * means that the callbacks in any of these later sublist can be | |
1447 | * grouped into a single sublist, whether or not they have already | |
1448 | * been assigned a ->completed number. | |
1449 | */ | |
1450 | c = rcu_cbs_completed(rsp, rnp); | |
1451 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1452 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1453 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1454 | break; | |
1455 | ||
1456 | /* | |
1457 | * If there are no sublist for unassigned callbacks, leave. | |
1458 | * At the same time, advance "i" one sublist, so that "i" will | |
1459 | * index into the sublist where all the remaining callbacks should | |
1460 | * be grouped into. | |
1461 | */ | |
1462 | if (++i >= RCU_NEXT_TAIL) | |
48a7639c | 1463 | return false; |
dc35c893 PM |
1464 | |
1465 | /* | |
1466 | * Assign all subsequent callbacks' ->completed number to the next | |
1467 | * full grace period and group them all in the sublist initially | |
1468 | * indexed by "i". | |
1469 | */ | |
1470 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1471 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1472 | rdp->nxtcompleted[i] = c; | |
1473 | } | |
910ee45d | 1474 | /* Record any needed additional grace periods. */ |
48a7639c | 1475 | ret = rcu_start_future_gp(rnp, rdp, NULL); |
6d4b418c PM |
1476 | |
1477 | /* Trace depending on how much we were able to accelerate. */ | |
1478 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1479 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1480 | else |
f7f7bac9 | 1481 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
48a7639c | 1482 | return ret; |
dc35c893 PM |
1483 | } |
1484 | ||
1485 | /* | |
1486 | * Move any callbacks whose grace period has completed to the | |
1487 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1488 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1489 | * sublist. This function is idempotent, so it does not hurt to | |
1490 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1491 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1492 | * |
1493 | * The caller must hold rnp->lock with interrupts disabled. | |
1494 | */ | |
48a7639c | 1495 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1496 | struct rcu_data *rdp) |
1497 | { | |
1498 | int i, j; | |
1499 | ||
1500 | /* If the CPU has no callbacks, nothing to do. */ | |
1501 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1502 | return false; |
dc35c893 PM |
1503 | |
1504 | /* | |
1505 | * Find all callbacks whose ->completed numbers indicate that they | |
1506 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1507 | */ | |
1508 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1509 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1510 | break; | |
1511 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1512 | } | |
1513 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1514 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1515 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1516 | ||
1517 | /* Copy down callbacks to fill in empty sublists. */ | |
1518 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1519 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1520 | break; | |
1521 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1522 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1523 | } | |
1524 | ||
1525 | /* Classify any remaining callbacks. */ | |
48a7639c | 1526 | return rcu_accelerate_cbs(rsp, rnp, rdp); |
dc35c893 PM |
1527 | } |
1528 | ||
d09b62df | 1529 | /* |
ba9fbe95 PM |
1530 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1531 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1532 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1533 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1534 | */ |
48a7639c PM |
1535 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, |
1536 | struct rcu_data *rdp) | |
d09b62df | 1537 | { |
48a7639c PM |
1538 | bool ret; |
1539 | ||
ba9fbe95 | 1540 | /* Handle the ends of any preceding grace periods first. */ |
dc35c893 | 1541 | if (rdp->completed == rnp->completed) { |
d09b62df | 1542 | |
ba9fbe95 | 1543 | /* No grace period end, so just accelerate recent callbacks. */ |
48a7639c | 1544 | ret = rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1545 | |
dc35c893 PM |
1546 | } else { |
1547 | ||
1548 | /* Advance callbacks. */ | |
48a7639c | 1549 | ret = rcu_advance_cbs(rsp, rnp, rdp); |
d09b62df PM |
1550 | |
1551 | /* Remember that we saw this grace-period completion. */ | |
1552 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1553 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1554 | } |
398ebe60 | 1555 | |
6eaef633 PM |
1556 | if (rdp->gpnum != rnp->gpnum) { |
1557 | /* | |
1558 | * If the current grace period is waiting for this CPU, | |
1559 | * set up to detect a quiescent state, otherwise don't | |
1560 | * go looking for one. | |
1561 | */ | |
1562 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1563 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
6eaef633 PM |
1564 | rdp->passed_quiesce = 0; |
1565 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); | |
1566 | zero_cpu_stall_ticks(rdp); | |
1567 | } | |
48a7639c | 1568 | return ret; |
6eaef633 PM |
1569 | } |
1570 | ||
d34ea322 | 1571 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1572 | { |
1573 | unsigned long flags; | |
48a7639c | 1574 | bool needwake; |
6eaef633 PM |
1575 | struct rcu_node *rnp; |
1576 | ||
1577 | local_irq_save(flags); | |
1578 | rnp = rdp->mynode; | |
d34ea322 PM |
1579 | if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) && |
1580 | rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */ | |
6eaef633 PM |
1581 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
1582 | local_irq_restore(flags); | |
1583 | return; | |
1584 | } | |
6303b9c8 | 1585 | smp_mb__after_unlock_lock(); |
48a7639c | 1586 | needwake = __note_gp_changes(rsp, rnp, rdp); |
6eaef633 | 1587 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
48a7639c PM |
1588 | if (needwake) |
1589 | rcu_gp_kthread_wake(rsp); | |
6eaef633 PM |
1590 | } |
1591 | ||
b3dbec76 | 1592 | /* |
f7be8209 | 1593 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1594 | */ |
7fdefc10 | 1595 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 PM |
1596 | { |
1597 | struct rcu_data *rdp; | |
7fdefc10 | 1598 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1599 | |
eb75767b | 1600 | rcu_bind_gp_kthread(); |
7fdefc10 | 1601 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1602 | smp_mb__after_unlock_lock(); |
91dc9542 | 1603 | if (!ACCESS_ONCE(rsp->gp_flags)) { |
f7be8209 PM |
1604 | /* Spurious wakeup, tell caller to go back to sleep. */ |
1605 | raw_spin_unlock_irq(&rnp->lock); | |
1606 | return 0; | |
1607 | } | |
91dc9542 | 1608 | ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */ |
b3dbec76 | 1609 | |
f7be8209 PM |
1610 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1611 | /* | |
1612 | * Grace period already in progress, don't start another. | |
1613 | * Not supposed to be able to happen. | |
1614 | */ | |
7fdefc10 PM |
1615 | raw_spin_unlock_irq(&rnp->lock); |
1616 | return 0; | |
1617 | } | |
1618 | ||
7fdefc10 | 1619 | /* Advance to a new grace period and initialize state. */ |
26cdfedf | 1620 | record_gp_stall_check_time(rsp); |
765a3f4f PM |
1621 | /* Record GP times before starting GP, hence smp_store_release(). */ |
1622 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); | |
f7f7bac9 | 1623 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1624 | raw_spin_unlock_irq(&rnp->lock); |
1625 | ||
1626 | /* Exclude any concurrent CPU-hotplug operations. */ | |
a4fbe35a | 1627 | mutex_lock(&rsp->onoff_mutex); |
765a3f4f | 1628 | smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */ |
7fdefc10 PM |
1629 | |
1630 | /* | |
1631 | * Set the quiescent-state-needed bits in all the rcu_node | |
1632 | * structures for all currently online CPUs in breadth-first order, | |
1633 | * starting from the root rcu_node structure, relying on the layout | |
1634 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1635 | * will access only the leaves of the hierarchy, thus seeing that no | |
1636 | * grace period is in progress, at least until the corresponding | |
1637 | * leaf node has been initialized. In addition, we have excluded | |
1638 | * CPU-hotplug operations. | |
1639 | * | |
1640 | * The grace period cannot complete until the initialization | |
1641 | * process finishes, because this kthread handles both. | |
1642 | */ | |
1643 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
b3dbec76 | 1644 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1645 | smp_mb__after_unlock_lock(); |
b3dbec76 | 1646 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1647 | rcu_preempt_check_blocked_tasks(rnp); |
1648 | rnp->qsmask = rnp->qsmaskinit; | |
0446be48 | 1649 | ACCESS_ONCE(rnp->gpnum) = rsp->gpnum; |
25d30cf4 | 1650 | WARN_ON_ONCE(rnp->completed != rsp->completed); |
0446be48 | 1651 | ACCESS_ONCE(rnp->completed) = rsp->completed; |
7fdefc10 | 1652 | if (rnp == rdp->mynode) |
48a7639c | 1653 | (void)__note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1654 | rcu_preempt_boost_start_gp(rnp); |
1655 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1656 | rnp->level, rnp->grplo, | |
1657 | rnp->grphi, rnp->qsmask); | |
1658 | raw_spin_unlock_irq(&rnp->lock); | |
1659 | cond_resched(); | |
1660 | } | |
b3dbec76 | 1661 | |
a4fbe35a | 1662 | mutex_unlock(&rsp->onoff_mutex); |
7fdefc10 PM |
1663 | return 1; |
1664 | } | |
b3dbec76 | 1665 | |
4cdfc175 PM |
1666 | /* |
1667 | * Do one round of quiescent-state forcing. | |
1668 | */ | |
01896f7e | 1669 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) |
4cdfc175 PM |
1670 | { |
1671 | int fqs_state = fqs_state_in; | |
217af2a2 PM |
1672 | bool isidle = false; |
1673 | unsigned long maxj; | |
4cdfc175 PM |
1674 | struct rcu_node *rnp = rcu_get_root(rsp); |
1675 | ||
1676 | rsp->n_force_qs++; | |
1677 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1678 | /* Collect dyntick-idle snapshots. */ | |
0edd1b17 | 1679 | if (is_sysidle_rcu_state(rsp)) { |
e02b2edf | 1680 | isidle = true; |
0edd1b17 PM |
1681 | maxj = jiffies - ULONG_MAX / 4; |
1682 | } | |
217af2a2 PM |
1683 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1684 | &isidle, &maxj); | |
0edd1b17 | 1685 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1686 | fqs_state = RCU_FORCE_QS; |
1687 | } else { | |
1688 | /* Handle dyntick-idle and offline CPUs. */ | |
e02b2edf | 1689 | isidle = false; |
217af2a2 | 1690 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1691 | } |
1692 | /* Clear flag to prevent immediate re-entry. */ | |
1693 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1694 | raw_spin_lock_irq(&rnp->lock); | |
6303b9c8 | 1695 | smp_mb__after_unlock_lock(); |
4de376a1 PK |
1696 | ACCESS_ONCE(rsp->gp_flags) = |
1697 | ACCESS_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS; | |
4cdfc175 PM |
1698 | raw_spin_unlock_irq(&rnp->lock); |
1699 | } | |
1700 | return fqs_state; | |
1701 | } | |
1702 | ||
7fdefc10 PM |
1703 | /* |
1704 | * Clean up after the old grace period. | |
1705 | */ | |
4cdfc175 | 1706 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1707 | { |
1708 | unsigned long gp_duration; | |
48a7639c | 1709 | bool needgp = false; |
dae6e64d | 1710 | int nocb = 0; |
7fdefc10 PM |
1711 | struct rcu_data *rdp; |
1712 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1713 | |
7fdefc10 | 1714 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1715 | smp_mb__after_unlock_lock(); |
7fdefc10 PM |
1716 | gp_duration = jiffies - rsp->gp_start; |
1717 | if (gp_duration > rsp->gp_max) | |
1718 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1719 | |
7fdefc10 PM |
1720 | /* |
1721 | * We know the grace period is complete, but to everyone else | |
1722 | * it appears to still be ongoing. But it is also the case | |
1723 | * that to everyone else it looks like there is nothing that | |
1724 | * they can do to advance the grace period. It is therefore | |
1725 | * safe for us to drop the lock in order to mark the grace | |
1726 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1727 | */ |
5d4b8659 | 1728 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1729 | |
5d4b8659 PM |
1730 | /* |
1731 | * Propagate new ->completed value to rcu_node structures so | |
1732 | * that other CPUs don't have to wait until the start of the next | |
1733 | * grace period to process their callbacks. This also avoids | |
1734 | * some nasty RCU grace-period initialization races by forcing | |
1735 | * the end of the current grace period to be completely recorded in | |
1736 | * all of the rcu_node structures before the beginning of the next | |
1737 | * grace period is recorded in any of the rcu_node structures. | |
1738 | */ | |
1739 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1740 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1741 | smp_mb__after_unlock_lock(); |
0446be48 | 1742 | ACCESS_ONCE(rnp->completed) = rsp->gpnum; |
b11cc576 PM |
1743 | rdp = this_cpu_ptr(rsp->rda); |
1744 | if (rnp == rdp->mynode) | |
48a7639c | 1745 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; |
78e4bc34 | 1746 | /* smp_mb() provided by prior unlock-lock pair. */ |
0446be48 | 1747 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 PM |
1748 | raw_spin_unlock_irq(&rnp->lock); |
1749 | cond_resched(); | |
7fdefc10 | 1750 | } |
5d4b8659 PM |
1751 | rnp = rcu_get_root(rsp); |
1752 | raw_spin_lock_irq(&rnp->lock); | |
765a3f4f | 1753 | smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */ |
dae6e64d | 1754 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 | 1755 | |
765a3f4f PM |
1756 | /* Declare grace period done. */ |
1757 | ACCESS_ONCE(rsp->completed) = rsp->gpnum; | |
f7f7bac9 | 1758 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
7fdefc10 | 1759 | rsp->fqs_state = RCU_GP_IDLE; |
5d4b8659 | 1760 | rdp = this_cpu_ptr(rsp->rda); |
48a7639c PM |
1761 | /* Advance CBs to reduce false positives below. */ |
1762 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; | |
1763 | if (needgp || cpu_needs_another_gp(rsp, rdp)) { | |
91dc9542 | 1764 | ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1765 | trace_rcu_grace_period(rsp->name, |
1766 | ACCESS_ONCE(rsp->gpnum), | |
1767 | TPS("newreq")); | |
1768 | } | |
7fdefc10 | 1769 | raw_spin_unlock_irq(&rnp->lock); |
7fdefc10 PM |
1770 | } |
1771 | ||
1772 | /* | |
1773 | * Body of kthread that handles grace periods. | |
1774 | */ | |
1775 | static int __noreturn rcu_gp_kthread(void *arg) | |
1776 | { | |
4cdfc175 | 1777 | int fqs_state; |
88d6df61 | 1778 | int gf; |
d40011f6 | 1779 | unsigned long j; |
4cdfc175 | 1780 | int ret; |
7fdefc10 PM |
1781 | struct rcu_state *rsp = arg; |
1782 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1783 | ||
1784 | for (;;) { | |
1785 | ||
1786 | /* Handle grace-period start. */ | |
1787 | for (;;) { | |
63c4db78 PM |
1788 | trace_rcu_grace_period(rsp->name, |
1789 | ACCESS_ONCE(rsp->gpnum), | |
1790 | TPS("reqwait")); | |
afea227f | 1791 | rsp->gp_state = RCU_GP_WAIT_GPS; |
4cdfc175 | 1792 | wait_event_interruptible(rsp->gp_wq, |
591c6d17 | 1793 | ACCESS_ONCE(rsp->gp_flags) & |
4cdfc175 | 1794 | RCU_GP_FLAG_INIT); |
78e4bc34 | 1795 | /* Locking provides needed memory barrier. */ |
f7be8209 | 1796 | if (rcu_gp_init(rsp)) |
7fdefc10 PM |
1797 | break; |
1798 | cond_resched(); | |
73a860cd | 1799 | WARN_ON(signal_pending(current)); |
63c4db78 PM |
1800 | trace_rcu_grace_period(rsp->name, |
1801 | ACCESS_ONCE(rsp->gpnum), | |
1802 | TPS("reqwaitsig")); | |
7fdefc10 | 1803 | } |
cabc49c1 | 1804 | |
4cdfc175 PM |
1805 | /* Handle quiescent-state forcing. */ |
1806 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
1807 | j = jiffies_till_first_fqs; |
1808 | if (j > HZ) { | |
1809 | j = HZ; | |
1810 | jiffies_till_first_fqs = HZ; | |
1811 | } | |
88d6df61 | 1812 | ret = 0; |
cabc49c1 | 1813 | for (;;) { |
88d6df61 PM |
1814 | if (!ret) |
1815 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 PM |
1816 | trace_rcu_grace_period(rsp->name, |
1817 | ACCESS_ONCE(rsp->gpnum), | |
1818 | TPS("fqswait")); | |
afea227f | 1819 | rsp->gp_state = RCU_GP_WAIT_FQS; |
4cdfc175 | 1820 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
88d6df61 PM |
1821 | ((gf = ACCESS_ONCE(rsp->gp_flags)) & |
1822 | RCU_GP_FLAG_FQS) || | |
4cdfc175 PM |
1823 | (!ACCESS_ONCE(rnp->qsmask) && |
1824 | !rcu_preempt_blocked_readers_cgp(rnp)), | |
d40011f6 | 1825 | j); |
78e4bc34 | 1826 | /* Locking provides needed memory barriers. */ |
4cdfc175 | 1827 | /* If grace period done, leave loop. */ |
cabc49c1 | 1828 | if (!ACCESS_ONCE(rnp->qsmask) && |
4cdfc175 | 1829 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 1830 | break; |
4cdfc175 | 1831 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
1832 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
1833 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 PM |
1834 | trace_rcu_grace_period(rsp->name, |
1835 | ACCESS_ONCE(rsp->gpnum), | |
1836 | TPS("fqsstart")); | |
4cdfc175 | 1837 | fqs_state = rcu_gp_fqs(rsp, fqs_state); |
63c4db78 PM |
1838 | trace_rcu_grace_period(rsp->name, |
1839 | ACCESS_ONCE(rsp->gpnum), | |
1840 | TPS("fqsend")); | |
4cdfc175 PM |
1841 | cond_resched(); |
1842 | } else { | |
1843 | /* Deal with stray signal. */ | |
1844 | cond_resched(); | |
73a860cd | 1845 | WARN_ON(signal_pending(current)); |
63c4db78 PM |
1846 | trace_rcu_grace_period(rsp->name, |
1847 | ACCESS_ONCE(rsp->gpnum), | |
1848 | TPS("fqswaitsig")); | |
4cdfc175 | 1849 | } |
d40011f6 PM |
1850 | j = jiffies_till_next_fqs; |
1851 | if (j > HZ) { | |
1852 | j = HZ; | |
1853 | jiffies_till_next_fqs = HZ; | |
1854 | } else if (j < 1) { | |
1855 | j = 1; | |
1856 | jiffies_till_next_fqs = 1; | |
1857 | } | |
cabc49c1 | 1858 | } |
4cdfc175 PM |
1859 | |
1860 | /* Handle grace-period end. */ | |
1861 | rcu_gp_cleanup(rsp); | |
b3dbec76 | 1862 | } |
b3dbec76 PM |
1863 | } |
1864 | ||
64db4cff PM |
1865 | /* |
1866 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
1867 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 1868 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
1869 | * |
1870 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
1871 | * invoke this function. This can happen when the dying CPU reports its | |
1872 | * quiescent state. | |
48a7639c PM |
1873 | * |
1874 | * Returns true if the grace-period kthread must be awakened. | |
64db4cff | 1875 | */ |
48a7639c | 1876 | static bool |
910ee45d PM |
1877 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
1878 | struct rcu_data *rdp) | |
64db4cff | 1879 | { |
b8462084 | 1880 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 1881 | /* |
b3dbec76 | 1882 | * Either we have not yet spawned the grace-period |
62da1921 PM |
1883 | * task, this CPU does not need another grace period, |
1884 | * or a grace period is already in progress. | |
b3dbec76 | 1885 | * Either way, don't start a new grace period. |
afe24b12 | 1886 | */ |
48a7639c | 1887 | return false; |
afe24b12 | 1888 | } |
91dc9542 | 1889 | ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1890 | trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum), |
1891 | TPS("newreq")); | |
62da1921 | 1892 | |
016a8d5b SR |
1893 | /* |
1894 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c | 1895 | * could cause possible deadlocks with the rq->lock. Defer |
48a7639c | 1896 | * the wakeup to our caller. |
016a8d5b | 1897 | */ |
48a7639c | 1898 | return true; |
64db4cff PM |
1899 | } |
1900 | ||
910ee45d PM |
1901 | /* |
1902 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
1903 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
1904 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
1905 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
1906 | * that is encountered beforehand. | |
48a7639c PM |
1907 | * |
1908 | * Returns true if the grace-period kthread needs to be awakened. | |
910ee45d | 1909 | */ |
48a7639c | 1910 | static bool rcu_start_gp(struct rcu_state *rsp) |
910ee45d PM |
1911 | { |
1912 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1913 | struct rcu_node *rnp = rcu_get_root(rsp); | |
48a7639c | 1914 | bool ret = false; |
910ee45d PM |
1915 | |
1916 | /* | |
1917 | * If there is no grace period in progress right now, any | |
1918 | * callbacks we have up to this point will be satisfied by the | |
1919 | * next grace period. Also, advancing the callbacks reduces the | |
1920 | * probability of false positives from cpu_needs_another_gp() | |
1921 | * resulting in pointless grace periods. So, advance callbacks | |
1922 | * then start the grace period! | |
1923 | */ | |
48a7639c PM |
1924 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; |
1925 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; | |
1926 | return ret; | |
910ee45d PM |
1927 | } |
1928 | ||
f41d911f | 1929 | /* |
d3f6bad3 PM |
1930 | * Report a full set of quiescent states to the specified rcu_state |
1931 | * data structure. This involves cleaning up after the prior grace | |
1932 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
1933 | * if one is needed. Note that the caller must hold rnp->lock, which |
1934 | * is released before return. | |
f41d911f | 1935 | */ |
d3f6bad3 | 1936 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 1937 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 1938 | { |
fc2219d4 | 1939 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cabc49c1 | 1940 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
2aa792e6 | 1941 | rcu_gp_kthread_wake(rsp); |
f41d911f PM |
1942 | } |
1943 | ||
64db4cff | 1944 | /* |
d3f6bad3 PM |
1945 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
1946 | * Allows quiescent states for a group of CPUs to be reported at one go | |
1947 | * to the specified rcu_node structure, though all the CPUs in the group | |
1948 | * must be represented by the same rcu_node structure (which need not be | |
1949 | * a leaf rcu_node structure, though it often will be). That structure's | |
1950 | * lock must be held upon entry, and it is released before return. | |
64db4cff PM |
1951 | */ |
1952 | static void | |
d3f6bad3 PM |
1953 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
1954 | struct rcu_node *rnp, unsigned long flags) | |
64db4cff PM |
1955 | __releases(rnp->lock) |
1956 | { | |
28ecd580 PM |
1957 | struct rcu_node *rnp_c; |
1958 | ||
64db4cff PM |
1959 | /* Walk up the rcu_node hierarchy. */ |
1960 | for (;;) { | |
1961 | if (!(rnp->qsmask & mask)) { | |
1962 | ||
1963 | /* Our bit has already been cleared, so done. */ | |
1304afb2 | 1964 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1965 | return; |
1966 | } | |
1967 | rnp->qsmask &= ~mask; | |
d4c08f2a PM |
1968 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
1969 | mask, rnp->qsmask, rnp->level, | |
1970 | rnp->grplo, rnp->grphi, | |
1971 | !!rnp->gp_tasks); | |
27f4d280 | 1972 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
1973 | |
1974 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 1975 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1976 | return; |
1977 | } | |
1978 | mask = rnp->grpmask; | |
1979 | if (rnp->parent == NULL) { | |
1980 | ||
1981 | /* No more levels. Exit loop holding root lock. */ | |
1982 | ||
1983 | break; | |
1984 | } | |
1304afb2 | 1985 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 1986 | rnp_c = rnp; |
64db4cff | 1987 | rnp = rnp->parent; |
1304afb2 | 1988 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 1989 | smp_mb__after_unlock_lock(); |
28ecd580 | 1990 | WARN_ON_ONCE(rnp_c->qsmask); |
64db4cff PM |
1991 | } |
1992 | ||
1993 | /* | |
1994 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 1995 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 1996 | * to clean up and start the next grace period if one is needed. |
64db4cff | 1997 | */ |
d3f6bad3 | 1998 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
1999 | } |
2000 | ||
2001 | /* | |
d3f6bad3 PM |
2002 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
2003 | * structure. This must be either called from the specified CPU, or | |
2004 | * called when the specified CPU is known to be offline (and when it is | |
2005 | * also known that no other CPU is concurrently trying to help the offline | |
2006 | * CPU). The lastcomp argument is used to make sure we are still in the | |
2007 | * grace period of interest. We don't want to end the current grace period | |
2008 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
2009 | */ |
2010 | static void | |
d7d6a11e | 2011 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2012 | { |
2013 | unsigned long flags; | |
2014 | unsigned long mask; | |
48a7639c | 2015 | bool needwake; |
64db4cff PM |
2016 | struct rcu_node *rnp; |
2017 | ||
2018 | rnp = rdp->mynode; | |
1304afb2 | 2019 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2020 | smp_mb__after_unlock_lock(); |
d7d6a11e PM |
2021 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum || |
2022 | rnp->completed == rnp->gpnum) { | |
64db4cff PM |
2023 | |
2024 | /* | |
e4cc1f22 PM |
2025 | * The grace period in which this quiescent state was |
2026 | * recorded has ended, so don't report it upwards. | |
2027 | * We will instead need a new quiescent state that lies | |
2028 | * within the current grace period. | |
64db4cff | 2029 | */ |
e4cc1f22 | 2030 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
1304afb2 | 2031 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2032 | return; |
2033 | } | |
2034 | mask = rdp->grpmask; | |
2035 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 2036 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2037 | } else { |
2038 | rdp->qs_pending = 0; | |
2039 | ||
2040 | /* | |
2041 | * This GP can't end until cpu checks in, so all of our | |
2042 | * callbacks can be processed during the next GP. | |
2043 | */ | |
48a7639c | 2044 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 2045 | |
d3f6bad3 | 2046 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
48a7639c PM |
2047 | if (needwake) |
2048 | rcu_gp_kthread_wake(rsp); | |
64db4cff PM |
2049 | } |
2050 | } | |
2051 | ||
2052 | /* | |
2053 | * Check to see if there is a new grace period of which this CPU | |
2054 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2055 | * Otherwise, see if this CPU has just passed through its first | |
2056 | * quiescent state for this grace period, and record that fact if so. | |
2057 | */ | |
2058 | static void | |
2059 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
2060 | { | |
05eb552b PM |
2061 | /* Check for grace-period ends and beginnings. */ |
2062 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
2063 | |
2064 | /* | |
2065 | * Does this CPU still need to do its part for current grace period? | |
2066 | * If no, return and let the other CPUs do their part as well. | |
2067 | */ | |
2068 | if (!rdp->qs_pending) | |
2069 | return; | |
2070 | ||
2071 | /* | |
2072 | * Was there a quiescent state since the beginning of the grace | |
2073 | * period? If no, then exit and wait for the next call. | |
2074 | */ | |
e4cc1f22 | 2075 | if (!rdp->passed_quiesce) |
64db4cff PM |
2076 | return; |
2077 | ||
d3f6bad3 PM |
2078 | /* |
2079 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2080 | * judge of that). | |
2081 | */ | |
d7d6a11e | 2082 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
2083 | } |
2084 | ||
2085 | #ifdef CONFIG_HOTPLUG_CPU | |
2086 | ||
e74f4c45 | 2087 | /* |
b1420f1c PM |
2088 | * Send the specified CPU's RCU callbacks to the orphanage. The |
2089 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 2090 | * ->orphan_lock. |
e74f4c45 | 2091 | */ |
b1420f1c PM |
2092 | static void |
2093 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
2094 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 2095 | { |
3fbfbf7a | 2096 | /* No-CBs CPUs do not have orphanable callbacks. */ |
d1e43fa5 | 2097 | if (rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
2098 | return; |
2099 | ||
b1420f1c PM |
2100 | /* |
2101 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
2102 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
2103 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 2104 | */ |
a50c3af9 | 2105 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2106 | rsp->qlen_lazy += rdp->qlen_lazy; |
2107 | rsp->qlen += rdp->qlen; | |
2108 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 2109 | rdp->qlen_lazy = 0; |
1d1fb395 | 2110 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
2111 | } |
2112 | ||
2113 | /* | |
b1420f1c PM |
2114 | * Next, move those callbacks still needing a grace period to |
2115 | * the orphanage, where some other CPU will pick them up. | |
2116 | * Some of the callbacks might have gone partway through a grace | |
2117 | * period, but that is too bad. They get to start over because we | |
2118 | * cannot assume that grace periods are synchronized across CPUs. | |
2119 | * We don't bother updating the ->nxttail[] array yet, instead | |
2120 | * we just reset the whole thing later on. | |
a50c3af9 | 2121 | */ |
b1420f1c PM |
2122 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
2123 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2124 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2125 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
2126 | } |
2127 | ||
2128 | /* | |
b1420f1c PM |
2129 | * Then move the ready-to-invoke callbacks to the orphanage, |
2130 | * where some other CPU will pick them up. These will not be | |
2131 | * required to pass though another grace period: They are done. | |
a50c3af9 | 2132 | */ |
e5601400 | 2133 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2134 | *rsp->orphan_donetail = rdp->nxtlist; |
2135 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 2136 | } |
e74f4c45 | 2137 | |
b1420f1c | 2138 | /* Finally, initialize the rcu_data structure's list to empty. */ |
3f5d3ea6 | 2139 | init_callback_list(rdp); |
b1420f1c PM |
2140 | } |
2141 | ||
2142 | /* | |
2143 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 2144 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c | 2145 | */ |
96d3fd0d | 2146 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) |
b1420f1c PM |
2147 | { |
2148 | int i; | |
fa07a58f | 2149 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
b1420f1c | 2150 | |
3fbfbf7a | 2151 | /* No-CBs CPUs are handled specially. */ |
96d3fd0d | 2152 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) |
3fbfbf7a PM |
2153 | return; |
2154 | ||
b1420f1c PM |
2155 | /* Do the accounting first. */ |
2156 | rdp->qlen_lazy += rsp->qlen_lazy; | |
2157 | rdp->qlen += rsp->qlen; | |
2158 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
2159 | if (rsp->qlen_lazy != rsp->qlen) |
2160 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2161 | rsp->qlen_lazy = 0; |
2162 | rsp->qlen = 0; | |
2163 | ||
2164 | /* | |
2165 | * We do not need a memory barrier here because the only way we | |
2166 | * can get here if there is an rcu_barrier() in flight is if | |
2167 | * we are the task doing the rcu_barrier(). | |
2168 | */ | |
2169 | ||
2170 | /* First adopt the ready-to-invoke callbacks. */ | |
2171 | if (rsp->orphan_donelist != NULL) { | |
2172 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2173 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
2174 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
2175 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2176 | rdp->nxttail[i] = rsp->orphan_donetail; | |
2177 | rsp->orphan_donelist = NULL; | |
2178 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2179 | } | |
2180 | ||
2181 | /* And then adopt the callbacks that still need a grace period. */ | |
2182 | if (rsp->orphan_nxtlist != NULL) { | |
2183 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
2184 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
2185 | rsp->orphan_nxtlist = NULL; | |
2186 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2187 | } | |
2188 | } | |
2189 | ||
2190 | /* | |
2191 | * Trace the fact that this CPU is going offline. | |
2192 | */ | |
2193 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
2194 | { | |
2195 | RCU_TRACE(unsigned long mask); | |
2196 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
2197 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
2198 | ||
2199 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
2200 | trace_rcu_grace_period(rsp->name, |
2201 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 2202 | TPS("cpuofl")); |
64db4cff PM |
2203 | } |
2204 | ||
2205 | /* | |
e5601400 | 2206 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
2207 | * this fact from process context. Do the remainder of the cleanup, |
2208 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
2209 | * adopting them. There can only be one CPU hotplug operation at a time, |
2210 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 2211 | */ |
e5601400 | 2212 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 2213 | { |
2036d94a PM |
2214 | unsigned long flags; |
2215 | unsigned long mask; | |
2216 | int need_report = 0; | |
e5601400 | 2217 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 2218 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2219 | |
2036d94a | 2220 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2221 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 2222 | |
2036d94a | 2223 | /* Exclude any attempts to start a new grace period. */ |
a4fbe35a | 2224 | mutex_lock(&rsp->onoff_mutex); |
7b2e6011 | 2225 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
2036d94a | 2226 | |
b1420f1c PM |
2227 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
2228 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
96d3fd0d | 2229 | rcu_adopt_orphan_cbs(rsp, flags); |
b1420f1c | 2230 | |
2036d94a PM |
2231 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
2232 | mask = rdp->grpmask; /* rnp->grplo is constant. */ | |
2233 | do { | |
2234 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
6303b9c8 | 2235 | smp_mb__after_unlock_lock(); |
2036d94a PM |
2236 | rnp->qsmaskinit &= ~mask; |
2237 | if (rnp->qsmaskinit != 0) { | |
2238 | if (rnp != rdp->mynode) | |
2239 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2240 | break; | |
2241 | } | |
2242 | if (rnp == rdp->mynode) | |
2243 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); | |
2244 | else | |
2245 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2246 | mask = rnp->grpmask; | |
2247 | rnp = rnp->parent; | |
2248 | } while (rnp != NULL); | |
2249 | ||
2250 | /* | |
2251 | * We still hold the leaf rcu_node structure lock here, and | |
2252 | * irqs are still disabled. The reason for this subterfuge is | |
7b2e6011 | 2253 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock |
2036d94a PM |
2254 | * held leads to deadlock. |
2255 | */ | |
7b2e6011 | 2256 | raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */ |
2036d94a PM |
2257 | rnp = rdp->mynode; |
2258 | if (need_report & RCU_OFL_TASKS_NORM_GP) | |
2259 | rcu_report_unblock_qs_rnp(rnp, flags); | |
2260 | else | |
2261 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2262 | if (need_report & RCU_OFL_TASKS_EXP_GP) | |
2263 | rcu_report_exp_rnp(rsp, rnp, true); | |
cf01537e PM |
2264 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
2265 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2266 | cpu, rdp->qlen, rdp->nxtlist); | |
0d8ee37e PM |
2267 | init_callback_list(rdp); |
2268 | /* Disallow further callbacks on this CPU. */ | |
2269 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; | |
a4fbe35a | 2270 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
2271 | } |
2272 | ||
2273 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
2274 | ||
e5601400 | 2275 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
2276 | { |
2277 | } | |
2278 | ||
e5601400 | 2279 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
2280 | { |
2281 | } | |
2282 | ||
2283 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
2284 | ||
2285 | /* | |
2286 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2287 | * period. Thottle as specified by rdp->blimit. | |
2288 | */ | |
37c72e56 | 2289 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2290 | { |
2291 | unsigned long flags; | |
2292 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
2293 | long bl, count, count_lazy; |
2294 | int i; | |
64db4cff | 2295 | |
dc35c893 | 2296 | /* If no callbacks are ready, just return. */ |
29c00b4a | 2297 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 2298 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
2299 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
2300 | need_resched(), is_idle_task(current), | |
2301 | rcu_is_callbacks_kthread()); | |
64db4cff | 2302 | return; |
29c00b4a | 2303 | } |
64db4cff PM |
2304 | |
2305 | /* | |
2306 | * Extract the list of ready callbacks, disabling to prevent | |
2307 | * races with call_rcu() from interrupt handlers. | |
2308 | */ | |
2309 | local_irq_save(flags); | |
8146c4e2 | 2310 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2311 | bl = rdp->blimit; |
486e2593 | 2312 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2313 | list = rdp->nxtlist; |
2314 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2315 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2316 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2317 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2318 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2319 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2320 | local_irq_restore(flags); |
2321 | ||
2322 | /* Invoke callbacks. */ | |
486e2593 | 2323 | count = count_lazy = 0; |
64db4cff PM |
2324 | while (list) { |
2325 | next = list->next; | |
2326 | prefetch(next); | |
551d55a9 | 2327 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2328 | if (__rcu_reclaim(rsp->name, list)) |
2329 | count_lazy++; | |
64db4cff | 2330 | list = next; |
dff1672d PM |
2331 | /* Stop only if limit reached and CPU has something to do. */ |
2332 | if (++count >= bl && | |
2333 | (need_resched() || | |
2334 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2335 | break; |
2336 | } | |
2337 | ||
2338 | local_irq_save(flags); | |
4968c300 PM |
2339 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2340 | is_idle_task(current), | |
2341 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2342 | |
2343 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2344 | if (list != NULL) { |
2345 | *tail = rdp->nxtlist; | |
2346 | rdp->nxtlist = list; | |
b41772ab PM |
2347 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2348 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2349 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2350 | else |
2351 | break; | |
2352 | } | |
b1420f1c PM |
2353 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2354 | rdp->qlen_lazy -= count_lazy; | |
a792563b | 2355 | ACCESS_ONCE(rdp->qlen) = rdp->qlen - count; |
b1420f1c | 2356 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2357 | |
2358 | /* Reinstate batch limit if we have worked down the excess. */ | |
2359 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2360 | rdp->blimit = blimit; | |
2361 | ||
37c72e56 PM |
2362 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2363 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2364 | rdp->qlen_last_fqs_check = 0; | |
2365 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2366 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2367 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2368 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2369 | |
64db4cff PM |
2370 | local_irq_restore(flags); |
2371 | ||
e0f23060 | 2372 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2373 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2374 | invoke_rcu_core(); |
64db4cff PM |
2375 | } |
2376 | ||
2377 | /* | |
2378 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2379 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2380 | * Also schedule RCU core processing. |
64db4cff | 2381 | * |
9b2e4f18 | 2382 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2383 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2384 | * false, there is no point in invoking rcu_check_callbacks(). | |
2385 | */ | |
2386 | void rcu_check_callbacks(int cpu, int user) | |
2387 | { | |
f7f7bac9 | 2388 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2389 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2390 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2391 | |
2392 | /* | |
2393 | * Get here if this CPU took its interrupt from user | |
2394 | * mode or from the idle loop, and if this is not a | |
2395 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2396 | * a quiescent state, so note it. |
64db4cff PM |
2397 | * |
2398 | * No memory barrier is required here because both | |
d6714c22 PM |
2399 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2400 | * variables that other CPUs neither access nor modify, | |
2401 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2402 | */ |
2403 | ||
d6714c22 PM |
2404 | rcu_sched_qs(cpu); |
2405 | rcu_bh_qs(cpu); | |
64db4cff PM |
2406 | |
2407 | } else if (!in_softirq()) { | |
2408 | ||
2409 | /* | |
2410 | * Get here if this CPU did not take its interrupt from | |
2411 | * softirq, in other words, if it is not interrupting | |
2412 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2413 | * critical section, so note it. |
64db4cff PM |
2414 | */ |
2415 | ||
d6714c22 | 2416 | rcu_bh_qs(cpu); |
64db4cff | 2417 | } |
f41d911f | 2418 | rcu_preempt_check_callbacks(cpu); |
d21670ac | 2419 | if (rcu_pending(cpu)) |
a46e0899 | 2420 | invoke_rcu_core(); |
f7f7bac9 | 2421 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2422 | } |
2423 | ||
64db4cff PM |
2424 | /* |
2425 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2426 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2427 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2428 | * | |
ee47eb9f | 2429 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2430 | */ |
217af2a2 PM |
2431 | static void force_qs_rnp(struct rcu_state *rsp, |
2432 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2433 | unsigned long *maxj), | |
2434 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2435 | { |
2436 | unsigned long bit; | |
2437 | int cpu; | |
2438 | unsigned long flags; | |
2439 | unsigned long mask; | |
a0b6c9a7 | 2440 | struct rcu_node *rnp; |
64db4cff | 2441 | |
a0b6c9a7 | 2442 | rcu_for_each_leaf_node(rsp, rnp) { |
b4be093f | 2443 | cond_resched(); |
64db4cff | 2444 | mask = 0; |
1304afb2 | 2445 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2446 | smp_mb__after_unlock_lock(); |
ee47eb9f | 2447 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 2448 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 2449 | return; |
64db4cff | 2450 | } |
a0b6c9a7 | 2451 | if (rnp->qsmask == 0) { |
1217ed1b | 2452 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
64db4cff PM |
2453 | continue; |
2454 | } | |
a0b6c9a7 | 2455 | cpu = rnp->grplo; |
64db4cff | 2456 | bit = 1; |
a0b6c9a7 | 2457 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 PM |
2458 | if ((rnp->qsmask & bit) != 0) { |
2459 | if ((rnp->qsmaskinit & bit) != 0) | |
e02b2edf | 2460 | *isidle = false; |
0edd1b17 PM |
2461 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) |
2462 | mask |= bit; | |
2463 | } | |
64db4cff | 2464 | } |
45f014c5 | 2465 | if (mask != 0) { |
64db4cff | 2466 | |
d3f6bad3 PM |
2467 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
2468 | rcu_report_qs_rnp(mask, rsp, rnp, flags); | |
64db4cff PM |
2469 | continue; |
2470 | } | |
1304afb2 | 2471 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2472 | } |
27f4d280 | 2473 | rnp = rcu_get_root(rsp); |
1217ed1b PM |
2474 | if (rnp->qsmask == 0) { |
2475 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 2476 | smp_mb__after_unlock_lock(); |
1217ed1b PM |
2477 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ |
2478 | } | |
64db4cff PM |
2479 | } |
2480 | ||
2481 | /* | |
2482 | * Force quiescent states on reluctant CPUs, and also detect which | |
2483 | * CPUs are in dyntick-idle mode. | |
2484 | */ | |
4cdfc175 | 2485 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2486 | { |
2487 | unsigned long flags; | |
394f2769 PM |
2488 | bool ret; |
2489 | struct rcu_node *rnp; | |
2490 | struct rcu_node *rnp_old = NULL; | |
2491 | ||
2492 | /* Funnel through hierarchy to reduce memory contention. */ | |
d860d403 | 2493 | rnp = __this_cpu_read(rsp->rda->mynode); |
394f2769 PM |
2494 | for (; rnp != NULL; rnp = rnp->parent) { |
2495 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2496 | !raw_spin_trylock(&rnp->fqslock); | |
2497 | if (rnp_old != NULL) | |
2498 | raw_spin_unlock(&rnp_old->fqslock); | |
2499 | if (ret) { | |
a792563b | 2500 | rsp->n_force_qs_lh++; |
394f2769 PM |
2501 | return; |
2502 | } | |
2503 | rnp_old = rnp; | |
2504 | } | |
2505 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2506 | |
394f2769 PM |
2507 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2508 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
6303b9c8 | 2509 | smp_mb__after_unlock_lock(); |
394f2769 PM |
2510 | raw_spin_unlock(&rnp_old->fqslock); |
2511 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
a792563b | 2512 | rsp->n_force_qs_lh++; |
394f2769 | 2513 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2514 | return; /* Someone beat us to it. */ |
46a1e34e | 2515 | } |
4de376a1 PK |
2516 | ACCESS_ONCE(rsp->gp_flags) = |
2517 | ACCESS_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS; | |
394f2769 | 2518 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
2aa792e6 | 2519 | rcu_gp_kthread_wake(rsp); |
64db4cff PM |
2520 | } |
2521 | ||
64db4cff | 2522 | /* |
e0f23060 PM |
2523 | * This does the RCU core processing work for the specified rcu_state |
2524 | * and rcu_data structures. This may be called only from the CPU to | |
2525 | * whom the rdp belongs. | |
64db4cff PM |
2526 | */ |
2527 | static void | |
1bca8cf1 | 2528 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2529 | { |
2530 | unsigned long flags; | |
48a7639c | 2531 | bool needwake; |
fa07a58f | 2532 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
64db4cff | 2533 | |
2e597558 PM |
2534 | WARN_ON_ONCE(rdp->beenonline == 0); |
2535 | ||
64db4cff PM |
2536 | /* Update RCU state based on any recent quiescent states. */ |
2537 | rcu_check_quiescent_state(rsp, rdp); | |
2538 | ||
2539 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2540 | local_irq_save(flags); |
64db4cff | 2541 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2542 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
48a7639c | 2543 | needwake = rcu_start_gp(rsp); |
b8462084 | 2544 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
48a7639c PM |
2545 | if (needwake) |
2546 | rcu_gp_kthread_wake(rsp); | |
dc35c893 PM |
2547 | } else { |
2548 | local_irq_restore(flags); | |
64db4cff PM |
2549 | } |
2550 | ||
2551 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2552 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2553 | invoke_rcu_callbacks(rsp, rdp); |
96d3fd0d PM |
2554 | |
2555 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2556 | do_nocb_deferred_wakeup(rdp); | |
09223371 SL |
2557 | } |
2558 | ||
64db4cff | 2559 | /* |
e0f23060 | 2560 | * Do RCU core processing for the current CPU. |
64db4cff | 2561 | */ |
09223371 | 2562 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2563 | { |
6ce75a23 PM |
2564 | struct rcu_state *rsp; |
2565 | ||
bfa00b4c PM |
2566 | if (cpu_is_offline(smp_processor_id())) |
2567 | return; | |
f7f7bac9 | 2568 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2569 | for_each_rcu_flavor(rsp) |
2570 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2571 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2572 | } |
2573 | ||
a26ac245 | 2574 | /* |
e0f23060 PM |
2575 | * Schedule RCU callback invocation. If the specified type of RCU |
2576 | * does not support RCU priority boosting, just do a direct call, | |
2577 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
2578 | * are running on the current CPU with interrupts disabled, the | |
2579 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
a26ac245 | 2580 | */ |
a46e0899 | 2581 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2582 | { |
b0d30417 PM |
2583 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
2584 | return; | |
a46e0899 PM |
2585 | if (likely(!rsp->boost)) { |
2586 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2587 | return; |
2588 | } | |
a46e0899 | 2589 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2590 | } |
2591 | ||
a46e0899 | 2592 | static void invoke_rcu_core(void) |
09223371 | 2593 | { |
b0f74036 PM |
2594 | if (cpu_online(smp_processor_id())) |
2595 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2596 | } |
2597 | ||
29154c57 PM |
2598 | /* |
2599 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2600 | */ | |
2601 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2602 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2603 | { |
48a7639c PM |
2604 | bool needwake; |
2605 | ||
62fde6ed PM |
2606 | /* |
2607 | * If called from an extended quiescent state, invoke the RCU | |
2608 | * core in order to force a re-evaluation of RCU's idleness. | |
2609 | */ | |
5c173eb8 | 2610 | if (!rcu_is_watching() && cpu_online(smp_processor_id())) |
62fde6ed PM |
2611 | invoke_rcu_core(); |
2612 | ||
a16b7a69 | 2613 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2614 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2615 | return; |
64db4cff | 2616 | |
37c72e56 PM |
2617 | /* |
2618 | * Force the grace period if too many callbacks or too long waiting. | |
2619 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2620 | * if some other CPU has recently done so. Also, don't bother | |
2621 | * invoking force_quiescent_state() if the newly enqueued callback | |
2622 | * is the only one waiting for a grace period to complete. | |
2623 | */ | |
2655d57e | 2624 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2625 | |
2626 | /* Are we ignoring a completed grace period? */ | |
470716fc | 2627 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
2628 | |
2629 | /* Start a new grace period if one not already started. */ | |
2630 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
2631 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
2632 | ||
b8462084 | 2633 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 | 2634 | smp_mb__after_unlock_lock(); |
48a7639c | 2635 | needwake = rcu_start_gp(rsp); |
b8462084 | 2636 | raw_spin_unlock(&rnp_root->lock); |
48a7639c PM |
2637 | if (needwake) |
2638 | rcu_gp_kthread_wake(rsp); | |
b52573d2 PM |
2639 | } else { |
2640 | /* Give the grace period a kick. */ | |
2641 | rdp->blimit = LONG_MAX; | |
2642 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2643 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2644 | force_quiescent_state(rsp); |
b52573d2 PM |
2645 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2646 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2647 | } | |
4cdfc175 | 2648 | } |
29154c57 PM |
2649 | } |
2650 | ||
ae150184 PM |
2651 | /* |
2652 | * RCU callback function to leak a callback. | |
2653 | */ | |
2654 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2655 | { | |
2656 | } | |
2657 | ||
3fbfbf7a PM |
2658 | /* |
2659 | * Helper function for call_rcu() and friends. The cpu argument will | |
2660 | * normally be -1, indicating "currently running CPU". It may specify | |
2661 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2662 | * is expected to specify a CPU. | |
2663 | */ | |
64db4cff PM |
2664 | static void |
2665 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 2666 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
2667 | { |
2668 | unsigned long flags; | |
2669 | struct rcu_data *rdp; | |
2670 | ||
1146edcb | 2671 | WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */ |
ae150184 PM |
2672 | if (debug_rcu_head_queue(head)) { |
2673 | /* Probable double call_rcu(), so leak the callback. */ | |
2674 | ACCESS_ONCE(head->func) = rcu_leak_callback; | |
2675 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); | |
2676 | return; | |
2677 | } | |
64db4cff PM |
2678 | head->func = func; |
2679 | head->next = NULL; | |
2680 | ||
64db4cff PM |
2681 | /* |
2682 | * Opportunistically note grace-period endings and beginnings. | |
2683 | * Note that we might see a beginning right after we see an | |
2684 | * end, but never vice versa, since this CPU has to pass through | |
2685 | * a quiescent state betweentimes. | |
2686 | */ | |
2687 | local_irq_save(flags); | |
394f99a9 | 2688 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
2689 | |
2690 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
2691 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
2692 | int offline; | |
2693 | ||
2694 | if (cpu != -1) | |
2695 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
96d3fd0d | 2696 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); |
3fbfbf7a | 2697 | WARN_ON_ONCE(offline); |
0d8ee37e | 2698 | /* _call_rcu() is illegal on offline CPU; leak the callback. */ |
0d8ee37e PM |
2699 | local_irq_restore(flags); |
2700 | return; | |
2701 | } | |
a792563b | 2702 | ACCESS_ONCE(rdp->qlen) = rdp->qlen + 1; |
486e2593 PM |
2703 | if (lazy) |
2704 | rdp->qlen_lazy++; | |
c57afe80 PM |
2705 | else |
2706 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2707 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
2708 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
2709 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 2710 | |
d4c08f2a PM |
2711 | if (__is_kfree_rcu_offset((unsigned long)func)) |
2712 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 2713 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2714 | else |
486e2593 | 2715 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2716 | |
29154c57 PM |
2717 | /* Go handle any RCU core processing required. */ |
2718 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
2719 | local_irq_restore(flags); |
2720 | } | |
2721 | ||
2722 | /* | |
d6714c22 | 2723 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 2724 | */ |
d6714c22 | 2725 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 2726 | { |
3fbfbf7a | 2727 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 2728 | } |
d6714c22 | 2729 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
2730 | |
2731 | /* | |
486e2593 | 2732 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
2733 | */ |
2734 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
2735 | { | |
3fbfbf7a | 2736 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
2737 | } |
2738 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
2739 | ||
495aa969 ACB |
2740 | /* |
2741 | * Queue an RCU callback for lazy invocation after a grace period. | |
2742 | * This will likely be later named something like "call_rcu_lazy()", | |
2743 | * but this change will require some way of tagging the lazy RCU | |
2744 | * callbacks in the list of pending callbacks. Until then, this | |
2745 | * function may only be called from __kfree_rcu(). | |
2746 | */ | |
2747 | void kfree_call_rcu(struct rcu_head *head, | |
2748 | void (*func)(struct rcu_head *rcu)) | |
2749 | { | |
e534165b | 2750 | __call_rcu(head, func, rcu_state_p, -1, 1); |
495aa969 ACB |
2751 | } |
2752 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
2753 | ||
6d813391 PM |
2754 | /* |
2755 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
2756 | * any blocking grace-period wait automatically implies a grace period | |
2757 | * if there is only one CPU online at any point time during execution | |
2758 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
2759 | * occasionally incorrectly indicate that there are multiple CPUs online | |
2760 | * when there was in fact only one the whole time, as this just adds | |
2761 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
2762 | */ |
2763 | static inline int rcu_blocking_is_gp(void) | |
2764 | { | |
95f0c1de PM |
2765 | int ret; |
2766 | ||
6d813391 | 2767 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
2768 | preempt_disable(); |
2769 | ret = num_online_cpus() <= 1; | |
2770 | preempt_enable(); | |
2771 | return ret; | |
6d813391 PM |
2772 | } |
2773 | ||
6ebb237b PM |
2774 | /** |
2775 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
2776 | * | |
2777 | * Control will return to the caller some time after a full rcu-sched | |
2778 | * grace period has elapsed, in other words after all currently executing | |
2779 | * rcu-sched read-side critical sections have completed. These read-side | |
2780 | * critical sections are delimited by rcu_read_lock_sched() and | |
2781 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
2782 | * local_irq_disable(), and so on may be used in place of | |
2783 | * rcu_read_lock_sched(). | |
2784 | * | |
2785 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
2786 | * non-threaded hardware-interrupt handlers, in progress on entry will |
2787 | * have completed before this primitive returns. However, this does not | |
2788 | * guarantee that softirq handlers will have completed, since in some | |
2789 | * kernels, these handlers can run in process context, and can block. | |
2790 | * | |
2791 | * Note that this guarantee implies further memory-ordering guarantees. | |
2792 | * On systems with more than one CPU, when synchronize_sched() returns, | |
2793 | * each CPU is guaranteed to have executed a full memory barrier since the | |
2794 | * end of its last RCU-sched read-side critical section whose beginning | |
2795 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
2796 | * an RCU read-side critical section that extends beyond the return from | |
2797 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
2798 | * after the beginning of synchronize_sched() and before the beginning of | |
2799 | * that RCU read-side critical section. Note that these guarantees include | |
2800 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
2801 | * that are executing in the kernel. | |
2802 | * | |
2803 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
2804 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
2805 | * to have executed a full memory barrier during the execution of | |
2806 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
2807 | * again only if the system has more than one CPU). | |
6ebb237b PM |
2808 | * |
2809 | * This primitive provides the guarantees made by the (now removed) | |
2810 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
2811 | * guarantees that rcu_read_lock() sections will have completed. | |
2812 | * In "classic RCU", these two guarantees happen to be one and | |
2813 | * the same, but can differ in realtime RCU implementations. | |
2814 | */ | |
2815 | void synchronize_sched(void) | |
2816 | { | |
fe15d706 PM |
2817 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2818 | !lock_is_held(&rcu_lock_map) && | |
2819 | !lock_is_held(&rcu_sched_lock_map), | |
2820 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
2821 | if (rcu_blocking_is_gp()) |
2822 | return; | |
3705b88d AM |
2823 | if (rcu_expedited) |
2824 | synchronize_sched_expedited(); | |
2825 | else | |
2826 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
2827 | } |
2828 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
2829 | ||
2830 | /** | |
2831 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
2832 | * | |
2833 | * Control will return to the caller some time after a full rcu_bh grace | |
2834 | * period has elapsed, in other words after all currently executing rcu_bh | |
2835 | * read-side critical sections have completed. RCU read-side critical | |
2836 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
2837 | * and may be nested. | |
f0a0e6f2 PM |
2838 | * |
2839 | * See the description of synchronize_sched() for more detailed information | |
2840 | * on memory ordering guarantees. | |
6ebb237b PM |
2841 | */ |
2842 | void synchronize_rcu_bh(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_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
2848 | if (rcu_blocking_is_gp()) |
2849 | return; | |
3705b88d AM |
2850 | if (rcu_expedited) |
2851 | synchronize_rcu_bh_expedited(); | |
2852 | else | |
2853 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
2854 | } |
2855 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
2856 | ||
765a3f4f PM |
2857 | /** |
2858 | * get_state_synchronize_rcu - Snapshot current RCU state | |
2859 | * | |
2860 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
2861 | * to determine whether or not a full grace period has elapsed in the | |
2862 | * meantime. | |
2863 | */ | |
2864 | unsigned long get_state_synchronize_rcu(void) | |
2865 | { | |
2866 | /* | |
2867 | * Any prior manipulation of RCU-protected data must happen | |
2868 | * before the load from ->gpnum. | |
2869 | */ | |
2870 | smp_mb(); /* ^^^ */ | |
2871 | ||
2872 | /* | |
2873 | * Make sure this load happens before the purportedly | |
2874 | * time-consuming work between get_state_synchronize_rcu() | |
2875 | * and cond_synchronize_rcu(). | |
2876 | */ | |
e534165b | 2877 | return smp_load_acquire(&rcu_state_p->gpnum); |
765a3f4f PM |
2878 | } |
2879 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
2880 | ||
2881 | /** | |
2882 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
2883 | * | |
2884 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
2885 | * | |
2886 | * If a full RCU grace period has elapsed since the earlier call to | |
2887 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
2888 | * synchronize_rcu() to wait for a full grace period. | |
2889 | * | |
2890 | * Yes, this function does not take counter wrap into account. But | |
2891 | * counter wrap is harmless. If the counter wraps, we have waited for | |
2892 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
2893 | * so waiting for one additional grace period should be just fine. | |
2894 | */ | |
2895 | void cond_synchronize_rcu(unsigned long oldstate) | |
2896 | { | |
2897 | unsigned long newstate; | |
2898 | ||
2899 | /* | |
2900 | * Ensure that this load happens before any RCU-destructive | |
2901 | * actions the caller might carry out after we return. | |
2902 | */ | |
e534165b | 2903 | newstate = smp_load_acquire(&rcu_state_p->completed); |
765a3f4f PM |
2904 | if (ULONG_CMP_GE(oldstate, newstate)) |
2905 | synchronize_rcu(); | |
2906 | } | |
2907 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
2908 | ||
3d3b7db0 PM |
2909 | static int synchronize_sched_expedited_cpu_stop(void *data) |
2910 | { | |
2911 | /* | |
2912 | * There must be a full memory barrier on each affected CPU | |
2913 | * between the time that try_stop_cpus() is called and the | |
2914 | * time that it returns. | |
2915 | * | |
2916 | * In the current initial implementation of cpu_stop, the | |
2917 | * above condition is already met when the control reaches | |
2918 | * this point and the following smp_mb() is not strictly | |
2919 | * necessary. Do smp_mb() anyway for documentation and | |
2920 | * robustness against future implementation changes. | |
2921 | */ | |
2922 | smp_mb(); /* See above comment block. */ | |
2923 | return 0; | |
2924 | } | |
2925 | ||
236fefaf PM |
2926 | /** |
2927 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
2928 | * | |
2929 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
2930 | * approach to force the grace period to end quickly. This consumes | |
2931 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
2932 | * so is thus not recommended for any sort of common-case code. In fact, | |
2933 | * if you are using synchronize_sched_expedited() in a loop, please | |
2934 | * restructure your code to batch your updates, and then use a single | |
2935 | * synchronize_sched() instead. | |
3d3b7db0 | 2936 | * |
236fefaf PM |
2937 | * Note that it is illegal to call this function while holding any lock |
2938 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal | |
2939 | * to call this function from a CPU-hotplug notifier. Failing to observe | |
2940 | * these restriction will result in deadlock. | |
3d3b7db0 PM |
2941 | * |
2942 | * This implementation can be thought of as an application of ticket | |
2943 | * locking to RCU, with sync_sched_expedited_started and | |
2944 | * sync_sched_expedited_done taking on the roles of the halves | |
2945 | * of the ticket-lock word. Each task atomically increments | |
2946 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
2947 | * then attempts to stop all the CPUs. If this succeeds, then each | |
2948 | * CPU will have executed a context switch, resulting in an RCU-sched | |
2949 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
2950 | * update sync_sched_expedited_done to match our snapshot -- but | |
2951 | * only if someone else has not already advanced past our snapshot. | |
2952 | * | |
2953 | * On the other hand, if try_stop_cpus() fails, we check the value | |
2954 | * of sync_sched_expedited_done. If it has advanced past our | |
2955 | * initial snapshot, then someone else must have forced a grace period | |
2956 | * some time after we took our snapshot. In this case, our work is | |
2957 | * done for us, and we can simply return. Otherwise, we try again, | |
2958 | * but keep our initial snapshot for purposes of checking for someone | |
2959 | * doing our work for us. | |
2960 | * | |
2961 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
2962 | */ | |
2963 | void synchronize_sched_expedited(void) | |
2964 | { | |
1924bcb0 PM |
2965 | long firstsnap, s, snap; |
2966 | int trycount = 0; | |
40694d66 | 2967 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 2968 | |
1924bcb0 PM |
2969 | /* |
2970 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
2971 | * By allowing sync_sched_expedited_started to advance no more than | |
2972 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
2973 | * that more than 3.5 billion CPUs would be required to force a | |
2974 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
2975 | * course be required on a 64-bit system. | |
2976 | */ | |
40694d66 PM |
2977 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
2978 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 PM |
2979 | ULONG_MAX / 8)) { |
2980 | synchronize_sched(); | |
a30489c5 | 2981 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
2982 | return; |
2983 | } | |
3d3b7db0 | 2984 | |
1924bcb0 PM |
2985 | /* |
2986 | * Take a ticket. Note that atomic_inc_return() implies a | |
2987 | * full memory barrier. | |
2988 | */ | |
40694d66 | 2989 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 2990 | firstsnap = snap; |
3d3b7db0 | 2991 | get_online_cpus(); |
1cc85961 | 2992 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 PM |
2993 | |
2994 | /* | |
2995 | * Each pass through the following loop attempts to force a | |
2996 | * context switch on each CPU. | |
2997 | */ | |
2998 | while (try_stop_cpus(cpu_online_mask, | |
2999 | synchronize_sched_expedited_cpu_stop, | |
3000 | NULL) == -EAGAIN) { | |
3001 | put_online_cpus(); | |
a30489c5 | 3002 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 3003 | |
1924bcb0 | 3004 | /* Check to see if someone else did our work for us. */ |
40694d66 | 3005 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3006 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 | 3007 | /* ensure test happens before caller kfree */ |
4e857c58 | 3008 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3009 | atomic_long_inc(&rsp->expedited_workdone1); |
1924bcb0 PM |
3010 | return; |
3011 | } | |
3d3b7db0 PM |
3012 | |
3013 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 3014 | if (trycount++ < 10) { |
3d3b7db0 | 3015 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 3016 | } else { |
3705b88d | 3017 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 3018 | atomic_long_inc(&rsp->expedited_normal); |
3d3b7db0 PM |
3019 | return; |
3020 | } | |
3021 | ||
1924bcb0 | 3022 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 3023 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3024 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 | 3025 | /* ensure test happens before caller kfree */ |
4e857c58 | 3026 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3027 | atomic_long_inc(&rsp->expedited_workdone2); |
3d3b7db0 PM |
3028 | return; |
3029 | } | |
3030 | ||
3031 | /* | |
3032 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
3033 | * callers to piggyback on our grace period. We retry |
3034 | * after they started, so our grace period works for them, | |
3035 | * and they started after our first try, so their grace | |
3036 | * period works for us. | |
3d3b7db0 PM |
3037 | */ |
3038 | get_online_cpus(); | |
40694d66 | 3039 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
3040 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
3041 | } | |
a30489c5 | 3042 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 PM |
3043 | |
3044 | /* | |
3045 | * Everyone up to our most recent fetch is covered by our grace | |
3046 | * period. Update the counter, but only if our work is still | |
3047 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 3048 | * than we did already did their update. |
3d3b7db0 PM |
3049 | */ |
3050 | do { | |
a30489c5 | 3051 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 3052 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 3053 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 | 3054 | /* ensure test happens before caller kfree */ |
4e857c58 | 3055 | smp_mb__before_atomic(); /* ^^^ */ |
a30489c5 | 3056 | atomic_long_inc(&rsp->expedited_done_lost); |
3d3b7db0 PM |
3057 | break; |
3058 | } | |
40694d66 | 3059 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 3060 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
3061 | |
3062 | put_online_cpus(); | |
3063 | } | |
3064 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
3065 | ||
64db4cff PM |
3066 | /* |
3067 | * Check to see if there is any immediate RCU-related work to be done | |
3068 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
3069 | * The checks are in order of increasing expense: checks that can be | |
3070 | * carried out against CPU-local state are performed first. However, | |
3071 | * we must check for CPU stalls first, else we might not get a chance. | |
3072 | */ | |
3073 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
3074 | { | |
2f51f988 PM |
3075 | struct rcu_node *rnp = rdp->mynode; |
3076 | ||
64db4cff PM |
3077 | rdp->n_rcu_pending++; |
3078 | ||
3079 | /* Check for CPU stalls, if enabled. */ | |
3080 | check_cpu_stall(rsp, rdp); | |
3081 | ||
a096932f PM |
3082 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ |
3083 | if (rcu_nohz_full_cpu(rsp)) | |
3084 | return 0; | |
3085 | ||
64db4cff | 3086 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
5c51dd73 PM |
3087 | if (rcu_scheduler_fully_active && |
3088 | rdp->qs_pending && !rdp->passed_quiesce) { | |
d21670ac | 3089 | rdp->n_rp_qs_pending++; |
e4cc1f22 | 3090 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
d21670ac | 3091 | rdp->n_rp_report_qs++; |
64db4cff | 3092 | return 1; |
7ba5c840 | 3093 | } |
64db4cff PM |
3094 | |
3095 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
3096 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
3097 | rdp->n_rp_cb_ready++; | |
64db4cff | 3098 | return 1; |
7ba5c840 | 3099 | } |
64db4cff PM |
3100 | |
3101 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
3102 | if (cpu_needs_another_gp(rsp, rdp)) { |
3103 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 3104 | return 1; |
7ba5c840 | 3105 | } |
64db4cff PM |
3106 | |
3107 | /* Has another RCU grace period completed? */ | |
2f51f988 | 3108 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 3109 | rdp->n_rp_gp_completed++; |
64db4cff | 3110 | return 1; |
7ba5c840 | 3111 | } |
64db4cff PM |
3112 | |
3113 | /* Has a new RCU grace period started? */ | |
2f51f988 | 3114 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
7ba5c840 | 3115 | rdp->n_rp_gp_started++; |
64db4cff | 3116 | return 1; |
7ba5c840 | 3117 | } |
64db4cff | 3118 | |
96d3fd0d PM |
3119 | /* Does this CPU need a deferred NOCB wakeup? */ |
3120 | if (rcu_nocb_need_deferred_wakeup(rdp)) { | |
3121 | rdp->n_rp_nocb_defer_wakeup++; | |
3122 | return 1; | |
3123 | } | |
3124 | ||
64db4cff | 3125 | /* nothing to do */ |
7ba5c840 | 3126 | rdp->n_rp_need_nothing++; |
64db4cff PM |
3127 | return 0; |
3128 | } | |
3129 | ||
3130 | /* | |
3131 | * Check to see if there is any immediate RCU-related work to be done | |
3132 | * by the current CPU, returning 1 if so. This function is part of the | |
3133 | * RCU implementation; it is -not- an exported member of the RCU API. | |
3134 | */ | |
a157229c | 3135 | static int rcu_pending(int cpu) |
64db4cff | 3136 | { |
6ce75a23 PM |
3137 | struct rcu_state *rsp; |
3138 | ||
3139 | for_each_rcu_flavor(rsp) | |
3140 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu))) | |
3141 | return 1; | |
3142 | return 0; | |
64db4cff PM |
3143 | } |
3144 | ||
3145 | /* | |
c0f4dfd4 PM |
3146 | * Return true if the specified CPU has any callback. If all_lazy is |
3147 | * non-NULL, store an indication of whether all callbacks are lazy. | |
3148 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 3149 | */ |
ffa83fb5 | 3150 | static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy) |
64db4cff | 3151 | { |
c0f4dfd4 PM |
3152 | bool al = true; |
3153 | bool hc = false; | |
3154 | struct rcu_data *rdp; | |
6ce75a23 PM |
3155 | struct rcu_state *rsp; |
3156 | ||
c0f4dfd4 PM |
3157 | for_each_rcu_flavor(rsp) { |
3158 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
69c8d28c PM |
3159 | if (!rdp->nxtlist) |
3160 | continue; | |
3161 | hc = true; | |
3162 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
c0f4dfd4 | 3163 | al = false; |
69c8d28c PM |
3164 | break; |
3165 | } | |
c0f4dfd4 PM |
3166 | } |
3167 | if (all_lazy) | |
3168 | *all_lazy = al; | |
3169 | return hc; | |
64db4cff PM |
3170 | } |
3171 | ||
a83eff0a PM |
3172 | /* |
3173 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
3174 | * the compiler is expected to optimize this away. | |
3175 | */ | |
e66c33d5 | 3176 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
3177 | int cpu, unsigned long done) |
3178 | { | |
3179 | trace_rcu_barrier(rsp->name, s, cpu, | |
3180 | atomic_read(&rsp->barrier_cpu_count), done); | |
3181 | } | |
3182 | ||
b1420f1c PM |
3183 | /* |
3184 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
3185 | * up the task executing _rcu_barrier(). | |
3186 | */ | |
24ebbca8 | 3187 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3188 | { |
24ebbca8 PM |
3189 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
3190 | struct rcu_state *rsp = rdp->rsp; | |
3191 | ||
a83eff0a PM |
3192 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
3193 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 3194 | complete(&rsp->barrier_completion); |
a83eff0a PM |
3195 | } else { |
3196 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
3197 | } | |
d0ec774c PM |
3198 | } |
3199 | ||
3200 | /* | |
3201 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3202 | */ | |
3203 | static void rcu_barrier_func(void *type) | |
3204 | { | |
037b64ed | 3205 | struct rcu_state *rsp = type; |
fa07a58f | 3206 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
d0ec774c | 3207 | |
a83eff0a | 3208 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 3209 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 3210 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
3211 | } |
3212 | ||
d0ec774c PM |
3213 | /* |
3214 | * Orchestrate the specified type of RCU barrier, waiting for all | |
3215 | * RCU callbacks of the specified type to complete. | |
3216 | */ | |
037b64ed | 3217 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 3218 | { |
b1420f1c | 3219 | int cpu; |
b1420f1c | 3220 | struct rcu_data *rdp; |
cf3a9c48 PM |
3221 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
3222 | unsigned long snap_done; | |
b1420f1c | 3223 | |
a83eff0a | 3224 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 3225 | |
e74f4c45 | 3226 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 3227 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 3228 | |
cf3a9c48 PM |
3229 | /* |
3230 | * Ensure that all prior references, including to ->n_barrier_done, | |
3231 | * are ordered before the _rcu_barrier() machinery. | |
3232 | */ | |
3233 | smp_mb(); /* See above block comment. */ | |
3234 | ||
3235 | /* | |
3236 | * Recheck ->n_barrier_done to see if others did our work for us. | |
3237 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
3238 | * transition. The "if" expression below therefore rounds the old | |
3239 | * value up to the next even number and adds two before comparing. | |
3240 | */ | |
458fb381 | 3241 | snap_done = rsp->n_barrier_done; |
a83eff0a | 3242 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
458fb381 PM |
3243 | |
3244 | /* | |
3245 | * If the value in snap is odd, we needed to wait for the current | |
3246 | * rcu_barrier() to complete, then wait for the next one, in other | |
3247 | * words, we need the value of snap_done to be three larger than | |
3248 | * the value of snap. On the other hand, if the value in snap is | |
3249 | * even, we only had to wait for the next rcu_barrier() to complete, | |
3250 | * in other words, we need the value of snap_done to be only two | |
3251 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes | |
3252 | * this for us (thank you, Linus!). | |
3253 | */ | |
3254 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) { | |
a83eff0a | 3255 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
3256 | smp_mb(); /* caller's subsequent code after above check. */ |
3257 | mutex_unlock(&rsp->barrier_mutex); | |
3258 | return; | |
3259 | } | |
3260 | ||
3261 | /* | |
3262 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
3263 | * ACCESS_ONCE() to prevent the compiler from speculating | |
3264 | * the increment to precede the early-exit check. | |
3265 | */ | |
a792563b | 3266 | ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1; |
cf3a9c48 | 3267 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); |
a83eff0a | 3268 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 3269 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 3270 | |
d0ec774c | 3271 | /* |
b1420f1c PM |
3272 | * Initialize the count to one rather than to zero in order to |
3273 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
3274 | * (or preemption of this task). Exclude CPU-hotplug operations |
3275 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 3276 | */ |
7db74df8 | 3277 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 3278 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 3279 | get_online_cpus(); |
b1420f1c PM |
3280 | |
3281 | /* | |
1331e7a1 PM |
3282 | * Force each CPU with callbacks to register a new callback. |
3283 | * When that callback is invoked, we will know that all of the | |
3284 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 3285 | */ |
3fbfbf7a | 3286 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 3287 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 3288 | continue; |
b1420f1c | 3289 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 3290 | if (rcu_is_nocb_cpu(cpu)) { |
3fbfbf7a PM |
3291 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, |
3292 | rsp->n_barrier_done); | |
3293 | atomic_inc(&rsp->barrier_cpu_count); | |
3294 | __call_rcu(&rdp->barrier_head, rcu_barrier_callback, | |
3295 | rsp, cpu, 0); | |
3296 | } else if (ACCESS_ONCE(rdp->qlen)) { | |
a83eff0a PM |
3297 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
3298 | rsp->n_barrier_done); | |
037b64ed | 3299 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 3300 | } else { |
a83eff0a PM |
3301 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
3302 | rsp->n_barrier_done); | |
b1420f1c PM |
3303 | } |
3304 | } | |
1331e7a1 | 3305 | put_online_cpus(); |
b1420f1c PM |
3306 | |
3307 | /* | |
3308 | * Now that we have an rcu_barrier_callback() callback on each | |
3309 | * CPU, and thus each counted, remove the initial count. | |
3310 | */ | |
24ebbca8 | 3311 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 3312 | complete(&rsp->barrier_completion); |
b1420f1c | 3313 | |
cf3a9c48 PM |
3314 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
3315 | smp_mb(); /* Keep increment after above mechanism. */ | |
a792563b | 3316 | ACCESS_ONCE(rsp->n_barrier_done) = rsp->n_barrier_done + 1; |
cf3a9c48 | 3317 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); |
a83eff0a | 3318 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
3319 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
3320 | ||
b1420f1c | 3321 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 3322 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
3323 | |
3324 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 3325 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 3326 | } |
d0ec774c PM |
3327 | |
3328 | /** | |
3329 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
3330 | */ | |
3331 | void rcu_barrier_bh(void) | |
3332 | { | |
037b64ed | 3333 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
3334 | } |
3335 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
3336 | ||
3337 | /** | |
3338 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
3339 | */ | |
3340 | void rcu_barrier_sched(void) | |
3341 | { | |
037b64ed | 3342 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
3343 | } |
3344 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
3345 | ||
64db4cff | 3346 | /* |
27569620 | 3347 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 3348 | */ |
27569620 PM |
3349 | static void __init |
3350 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
3351 | { |
3352 | unsigned long flags; | |
394f99a9 | 3353 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
3354 | struct rcu_node *rnp = rcu_get_root(rsp); |
3355 | ||
3356 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 3357 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 3358 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
3f5d3ea6 | 3359 | init_callback_list(rdp); |
486e2593 | 3360 | rdp->qlen_lazy = 0; |
1d1fb395 | 3361 | ACCESS_ONCE(rdp->qlen) = 0; |
27569620 | 3362 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 3363 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 3364 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 3365 | rdp->cpu = cpu; |
d4c08f2a | 3366 | rdp->rsp = rsp; |
3fbfbf7a | 3367 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 3368 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
3369 | } |
3370 | ||
3371 | /* | |
3372 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
3373 | * offline event can be happening at a given time. Note also that we | |
3374 | * can accept some slop in the rsp->completed access due to the fact | |
3375 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 3376 | */ |
49fb4c62 | 3377 | static void |
9b67122a | 3378 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp) |
64db4cff PM |
3379 | { |
3380 | unsigned long flags; | |
64db4cff | 3381 | unsigned long mask; |
394f99a9 | 3382 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
3383 | struct rcu_node *rnp = rcu_get_root(rsp); |
3384 | ||
a4fbe35a PM |
3385 | /* Exclude new grace periods. */ |
3386 | mutex_lock(&rsp->onoff_mutex); | |
3387 | ||
64db4cff | 3388 | /* Set up local state, ensuring consistent view of global state. */ |
1304afb2 | 3389 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 3390 | rdp->beenonline = 1; /* We have now been online. */ |
37c72e56 PM |
3391 | rdp->qlen_last_fqs_check = 0; |
3392 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 3393 | rdp->blimit = blimit; |
0d8ee37e | 3394 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ |
29e37d81 | 3395 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 3396 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
3397 | atomic_set(&rdp->dynticks->dynticks, |
3398 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 3399 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 3400 | |
64db4cff PM |
3401 | /* Add CPU to rcu_node bitmasks. */ |
3402 | rnp = rdp->mynode; | |
3403 | mask = rdp->grpmask; | |
3404 | do { | |
3405 | /* Exclude any attempts to start a new GP on small systems. */ | |
1304afb2 | 3406 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3407 | rnp->qsmaskinit |= mask; |
3408 | mask = rnp->grpmask; | |
d09b62df | 3409 | if (rnp == rdp->mynode) { |
06ae115a PM |
3410 | /* |
3411 | * If there is a grace period in progress, we will | |
3412 | * set up to wait for it next time we run the | |
3413 | * RCU core code. | |
3414 | */ | |
3415 | rdp->gpnum = rnp->completed; | |
d09b62df | 3416 | rdp->completed = rnp->completed; |
06ae115a PM |
3417 | rdp->passed_quiesce = 0; |
3418 | rdp->qs_pending = 0; | |
f7f7bac9 | 3419 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); |
d09b62df | 3420 | } |
1304afb2 | 3421 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3422 | rnp = rnp->parent; |
3423 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); | |
a4fbe35a | 3424 | local_irq_restore(flags); |
64db4cff | 3425 | |
a4fbe35a | 3426 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
3427 | } |
3428 | ||
49fb4c62 | 3429 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 3430 | { |
6ce75a23 PM |
3431 | struct rcu_state *rsp; |
3432 | ||
3433 | for_each_rcu_flavor(rsp) | |
9b67122a | 3434 | rcu_init_percpu_data(cpu, rsp); |
64db4cff PM |
3435 | } |
3436 | ||
3437 | /* | |
f41d911f | 3438 | * Handle CPU online/offline notification events. |
64db4cff | 3439 | */ |
49fb4c62 | 3440 | static int rcu_cpu_notify(struct notifier_block *self, |
9f680ab4 | 3441 | unsigned long action, void *hcpu) |
64db4cff PM |
3442 | { |
3443 | long cpu = (long)hcpu; | |
e534165b | 3444 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
a26ac245 | 3445 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 3446 | struct rcu_state *rsp; |
64db4cff | 3447 | |
f7f7bac9 | 3448 | trace_rcu_utilization(TPS("Start CPU hotplug")); |
64db4cff PM |
3449 | switch (action) { |
3450 | case CPU_UP_PREPARE: | |
3451 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
3452 | rcu_prepare_cpu(cpu); |
3453 | rcu_prepare_kthreads(cpu); | |
a26ac245 PM |
3454 | break; |
3455 | case CPU_ONLINE: | |
0f962a5e | 3456 | case CPU_DOWN_FAILED: |
5d01bbd1 | 3457 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
3458 | break; |
3459 | case CPU_DOWN_PREPARE: | |
34ed6246 | 3460 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 3461 | break; |
d0ec774c PM |
3462 | case CPU_DYING: |
3463 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
3464 | for_each_rcu_flavor(rsp) |
3465 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 3466 | break; |
64db4cff PM |
3467 | case CPU_DEAD: |
3468 | case CPU_DEAD_FROZEN: | |
3469 | case CPU_UP_CANCELED: | |
3470 | case CPU_UP_CANCELED_FROZEN: | |
6ce75a23 PM |
3471 | for_each_rcu_flavor(rsp) |
3472 | rcu_cleanup_dead_cpu(cpu, rsp); | |
64db4cff PM |
3473 | break; |
3474 | default: | |
3475 | break; | |
3476 | } | |
f7f7bac9 | 3477 | trace_rcu_utilization(TPS("End CPU hotplug")); |
34ed6246 | 3478 | return NOTIFY_OK; |
64db4cff PM |
3479 | } |
3480 | ||
d1d74d14 BP |
3481 | static int rcu_pm_notify(struct notifier_block *self, |
3482 | unsigned long action, void *hcpu) | |
3483 | { | |
3484 | switch (action) { | |
3485 | case PM_HIBERNATION_PREPARE: | |
3486 | case PM_SUSPEND_PREPARE: | |
3487 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
3488 | rcu_expedited = 1; | |
3489 | break; | |
3490 | case PM_POST_HIBERNATION: | |
3491 | case PM_POST_SUSPEND: | |
3492 | rcu_expedited = 0; | |
3493 | break; | |
3494 | default: | |
3495 | break; | |
3496 | } | |
3497 | return NOTIFY_OK; | |
3498 | } | |
3499 | ||
b3dbec76 PM |
3500 | /* |
3501 | * Spawn the kthread that handles this RCU flavor's grace periods. | |
3502 | */ | |
3503 | static int __init rcu_spawn_gp_kthread(void) | |
3504 | { | |
3505 | unsigned long flags; | |
3506 | struct rcu_node *rnp; | |
3507 | struct rcu_state *rsp; | |
3508 | struct task_struct *t; | |
3509 | ||
3510 | for_each_rcu_flavor(rsp) { | |
f170168b | 3511 | t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
3512 | BUG_ON(IS_ERR(t)); |
3513 | rnp = rcu_get_root(rsp); | |
3514 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3515 | rsp->gp_kthread = t; | |
3516 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3fbfbf7a | 3517 | rcu_spawn_nocb_kthreads(rsp); |
b3dbec76 PM |
3518 | } |
3519 | return 0; | |
3520 | } | |
3521 | early_initcall(rcu_spawn_gp_kthread); | |
3522 | ||
bbad9379 PM |
3523 | /* |
3524 | * This function is invoked towards the end of the scheduler's initialization | |
3525 | * process. Before this is called, the idle task might contain | |
3526 | * RCU read-side critical sections (during which time, this idle | |
3527 | * task is booting the system). After this function is called, the | |
3528 | * idle tasks are prohibited from containing RCU read-side critical | |
3529 | * sections. This function also enables RCU lockdep checking. | |
3530 | */ | |
3531 | void rcu_scheduler_starting(void) | |
3532 | { | |
3533 | WARN_ON(num_online_cpus() != 1); | |
3534 | WARN_ON(nr_context_switches() > 0); | |
3535 | rcu_scheduler_active = 1; | |
3536 | } | |
3537 | ||
64db4cff PM |
3538 | /* |
3539 | * Compute the per-level fanout, either using the exact fanout specified | |
3540 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
3541 | */ | |
3542 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
3543 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3544 | { | |
3545 | int i; | |
3546 | ||
04f34650 PM |
3547 | rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; |
3548 | for (i = rcu_num_lvls - 2; i >= 0; i--) | |
64db4cff PM |
3549 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
3550 | } | |
3551 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3552 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3553 | { | |
3554 | int ccur; | |
3555 | int cprv; | |
3556 | int i; | |
3557 | ||
4dbd6bb3 | 3558 | cprv = nr_cpu_ids; |
f885b7f2 | 3559 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3560 | ccur = rsp->levelcnt[i]; |
3561 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
3562 | cprv = ccur; | |
3563 | } | |
3564 | } | |
3565 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3566 | ||
3567 | /* | |
3568 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3569 | */ | |
394f99a9 LJ |
3570 | static void __init rcu_init_one(struct rcu_state *rsp, |
3571 | struct rcu_data __percpu *rda) | |
64db4cff | 3572 | { |
b4426b49 FF |
3573 | static const char * const buf[] = { |
3574 | "rcu_node_0", | |
3575 | "rcu_node_1", | |
3576 | "rcu_node_2", | |
3577 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ | |
3578 | static const char * const fqs[] = { | |
3579 | "rcu_node_fqs_0", | |
3580 | "rcu_node_fqs_1", | |
3581 | "rcu_node_fqs_2", | |
3582 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ | |
4a81e832 | 3583 | static u8 fl_mask = 0x1; |
64db4cff PM |
3584 | int cpustride = 1; |
3585 | int i; | |
3586 | int j; | |
3587 | struct rcu_node *rnp; | |
3588 | ||
b6407e86 PM |
3589 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
3590 | ||
4930521a PM |
3591 | /* Silence gcc 4.8 warning about array index out of range. */ |
3592 | if (rcu_num_lvls > RCU_NUM_LVLS) | |
3593 | panic("rcu_init_one: rcu_num_lvls overflow"); | |
3594 | ||
64db4cff PM |
3595 | /* Initialize the level-tracking arrays. */ |
3596 | ||
f885b7f2 PM |
3597 | for (i = 0; i < rcu_num_lvls; i++) |
3598 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
3599 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
3600 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
3601 | rcu_init_levelspread(rsp); | |
4a81e832 PM |
3602 | rsp->flavor_mask = fl_mask; |
3603 | fl_mask <<= 1; | |
64db4cff PM |
3604 | |
3605 | /* Initialize the elements themselves, starting from the leaves. */ | |
3606 | ||
f885b7f2 | 3607 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3608 | cpustride *= rsp->levelspread[i]; |
3609 | rnp = rsp->level[i]; | |
3610 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 3611 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
3612 | lockdep_set_class_and_name(&rnp->lock, |
3613 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
3614 | raw_spin_lock_init(&rnp->fqslock); |
3615 | lockdep_set_class_and_name(&rnp->fqslock, | |
3616 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
3617 | rnp->gpnum = rsp->gpnum; |
3618 | rnp->completed = rsp->completed; | |
64db4cff PM |
3619 | rnp->qsmask = 0; |
3620 | rnp->qsmaskinit = 0; | |
3621 | rnp->grplo = j * cpustride; | |
3622 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
3623 | if (rnp->grphi >= nr_cpu_ids) |
3624 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
3625 | if (i == 0) { |
3626 | rnp->grpnum = 0; | |
3627 | rnp->grpmask = 0; | |
3628 | rnp->parent = NULL; | |
3629 | } else { | |
3630 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
3631 | rnp->grpmask = 1UL << rnp->grpnum; | |
3632 | rnp->parent = rsp->level[i - 1] + | |
3633 | j / rsp->levelspread[i - 1]; | |
3634 | } | |
3635 | rnp->level = i; | |
12f5f524 | 3636 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 3637 | rcu_init_one_nocb(rnp); |
64db4cff PM |
3638 | } |
3639 | } | |
0c34029a | 3640 | |
394f99a9 | 3641 | rsp->rda = rda; |
b3dbec76 | 3642 | init_waitqueue_head(&rsp->gp_wq); |
f885b7f2 | 3643 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 3644 | for_each_possible_cpu(i) { |
4a90a068 | 3645 | while (i > rnp->grphi) |
0c34029a | 3646 | rnp++; |
394f99a9 | 3647 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
3648 | rcu_boot_init_percpu_data(i, rsp); |
3649 | } | |
6ce75a23 | 3650 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
3651 | } |
3652 | ||
f885b7f2 PM |
3653 | /* |
3654 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 3655 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
3656 | * the ->node array in the rcu_state structure. |
3657 | */ | |
3658 | static void __init rcu_init_geometry(void) | |
3659 | { | |
026ad283 | 3660 | ulong d; |
f885b7f2 PM |
3661 | int i; |
3662 | int j; | |
cca6f393 | 3663 | int n = nr_cpu_ids; |
f885b7f2 PM |
3664 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
3665 | ||
026ad283 PM |
3666 | /* |
3667 | * Initialize any unspecified boot parameters. | |
3668 | * The default values of jiffies_till_first_fqs and | |
3669 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
3670 | * value, which is a function of HZ, then adding one for each | |
3671 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
3672 | */ | |
3673 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
3674 | if (jiffies_till_first_fqs == ULONG_MAX) | |
3675 | jiffies_till_first_fqs = d; | |
3676 | if (jiffies_till_next_fqs == ULONG_MAX) | |
3677 | jiffies_till_next_fqs = d; | |
3678 | ||
f885b7f2 | 3679 | /* If the compile-time values are accurate, just leave. */ |
b17c7035 PM |
3680 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
3681 | nr_cpu_ids == NR_CPUS) | |
f885b7f2 | 3682 | return; |
39479098 PM |
3683 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", |
3684 | rcu_fanout_leaf, nr_cpu_ids); | |
f885b7f2 PM |
3685 | |
3686 | /* | |
3687 | * Compute number of nodes that can be handled an rcu_node tree | |
3688 | * with the given number of levels. Setting rcu_capacity[0] makes | |
3689 | * some of the arithmetic easier. | |
3690 | */ | |
3691 | rcu_capacity[0] = 1; | |
3692 | rcu_capacity[1] = rcu_fanout_leaf; | |
3693 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
3694 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
3695 | ||
3696 | /* | |
3697 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
3698 | * to increase the leaf-level fanout, not decrease it. Of course, | |
3699 | * the leaf-level fanout cannot exceed the number of bits in | |
3700 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
3701 | * the configured number of CPUs. Complain and fall back to the | |
3702 | * compile-time values if these limits are exceeded. | |
3703 | */ | |
3704 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
3705 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
3706 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
3707 | WARN_ON(1); | |
3708 | return; | |
3709 | } | |
3710 | ||
3711 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
3712 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
3713 | if (n <= rcu_capacity[i]) { | |
3714 | for (j = 0; j <= i; j++) | |
3715 | num_rcu_lvl[j] = | |
3716 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
3717 | rcu_num_lvls = i; | |
3718 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
3719 | num_rcu_lvl[j] = 0; | |
3720 | break; | |
3721 | } | |
3722 | ||
3723 | /* Calculate the total number of rcu_node structures. */ | |
3724 | rcu_num_nodes = 0; | |
3725 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
3726 | rcu_num_nodes += num_rcu_lvl[i]; | |
3727 | rcu_num_nodes -= n; | |
3728 | } | |
3729 | ||
9f680ab4 | 3730 | void __init rcu_init(void) |
64db4cff | 3731 | { |
017c4261 | 3732 | int cpu; |
9f680ab4 | 3733 | |
f41d911f | 3734 | rcu_bootup_announce(); |
f885b7f2 | 3735 | rcu_init_geometry(); |
394f99a9 | 3736 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); |
69c8d28c | 3737 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
f41d911f | 3738 | __rcu_init_preempt(); |
b5b39360 | 3739 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
3740 | |
3741 | /* | |
3742 | * We don't need protection against CPU-hotplug here because | |
3743 | * this is called early in boot, before either interrupts | |
3744 | * or the scheduler are operational. | |
3745 | */ | |
3746 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 3747 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
3748 | for_each_online_cpu(cpu) |
3749 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
3750 | } |
3751 | ||
4102adab | 3752 | #include "tree_plugin.h" |