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 | */ | |
a41bfeb2 | 82 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ |
f7f7bac9 SRRH |
83 | static char sname##_varname[] = #sname; \ |
84 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \ | |
a41bfeb2 | 85 | struct rcu_state sname##_state = { \ |
6c90cc7b | 86 | .level = { &sname##_state.node[0] }, \ |
037b64ed | 87 | .call = cr, \ |
af446b70 | 88 | .fqs_state = RCU_GP_IDLE, \ |
42c3533e PM |
89 | .gpnum = 0UL - 300UL, \ |
90 | .completed = 0UL - 300UL, \ | |
7b2e6011 | 91 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
6c90cc7b PM |
92 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
93 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 94 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
a4fbe35a | 95 | .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \ |
f7f7bac9 | 96 | .name = sname##_varname, \ |
a4889858 | 97 | .abbr = sabbr, \ |
a41bfeb2 SRRH |
98 | }; \ |
99 | DEFINE_PER_CPU(struct rcu_data, sname##_data) | |
64db4cff | 100 | |
a41bfeb2 SRRH |
101 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); |
102 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); | |
b1f77b05 | 103 | |
27f4d280 | 104 | static struct rcu_state *rcu_state; |
6ce75a23 | 105 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 106 | |
f885b7f2 PM |
107 | /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */ |
108 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF; | |
7e5c2dfb | 109 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 PM |
110 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
111 | static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */ | |
112 | NUM_RCU_LVL_0, | |
113 | NUM_RCU_LVL_1, | |
114 | NUM_RCU_LVL_2, | |
115 | NUM_RCU_LVL_3, | |
116 | NUM_RCU_LVL_4, | |
117 | }; | |
118 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ | |
119 | ||
b0d30417 PM |
120 | /* |
121 | * The rcu_scheduler_active variable transitions from zero to one just | |
122 | * before the first task is spawned. So when this variable is zero, RCU | |
123 | * can assume that there is but one task, allowing RCU to (for example) | |
b44f6656 | 124 | * optimize synchronize_sched() to a simple barrier(). When this variable |
b0d30417 PM |
125 | * is one, RCU must actually do all the hard work required to detect real |
126 | * grace periods. This variable is also used to suppress boot-time false | |
127 | * positives from lockdep-RCU error checking. | |
128 | */ | |
bbad9379 PM |
129 | int rcu_scheduler_active __read_mostly; |
130 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
131 | ||
b0d30417 PM |
132 | /* |
133 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
134 | * during the early_initcall() processing, which is after the scheduler | |
135 | * is capable of creating new tasks. So RCU processing (for example, | |
136 | * creating tasks for RCU priority boosting) must be delayed until after | |
137 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
138 | * currently delay invocation of any RCU callbacks until after this point. | |
139 | * | |
140 | * It might later prove better for people registering RCU callbacks during | |
141 | * early boot to take responsibility for these callbacks, but one step at | |
142 | * a time. | |
143 | */ | |
144 | static int rcu_scheduler_fully_active __read_mostly; | |
145 | ||
a46e0899 PM |
146 | #ifdef CONFIG_RCU_BOOST |
147 | ||
a26ac245 PM |
148 | /* |
149 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
150 | * handle all flavors of RCU. | |
151 | */ | |
152 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
d71df90e | 153 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); |
5ece5bab | 154 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); |
d71df90e | 155 | DEFINE_PER_CPU(char, rcu_cpu_has_work); |
a26ac245 | 156 | |
a46e0899 PM |
157 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
158 | ||
5d01bbd1 | 159 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
160 | static void invoke_rcu_core(void); |
161 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 162 | |
4a298656 PM |
163 | /* |
164 | * Track the rcutorture test sequence number and the update version | |
165 | * number within a given test. The rcutorture_testseq is incremented | |
166 | * on every rcutorture module load and unload, so has an odd value | |
167 | * when a test is running. The rcutorture_vernum is set to zero | |
168 | * when rcutorture starts and is incremented on each rcutorture update. | |
169 | * These variables enable correlating rcutorture output with the | |
170 | * RCU tracing information. | |
171 | */ | |
172 | unsigned long rcutorture_testseq; | |
173 | unsigned long rcutorture_vernum; | |
174 | ||
fc2219d4 PM |
175 | /* |
176 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
177 | * permit this function to be invoked without holding the root rcu_node | |
178 | * structure's ->lock, but of course results can be subject to change. | |
179 | */ | |
180 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
181 | { | |
182 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
183 | } | |
184 | ||
b1f77b05 | 185 | /* |
d6714c22 | 186 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 187 | * how many quiescent states passed, just if there was at least |
d6714c22 | 188 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 189 | * The caller must have disabled preemption. |
b1f77b05 | 190 | */ |
d6714c22 | 191 | void rcu_sched_qs(int cpu) |
b1f77b05 | 192 | { |
25502a6c | 193 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu); |
f41d911f | 194 | |
e4cc1f22 | 195 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 196 | trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 197 | rdp->passed_quiesce = 1; |
b1f77b05 IM |
198 | } |
199 | ||
d6714c22 | 200 | void rcu_bh_qs(int cpu) |
b1f77b05 | 201 | { |
25502a6c | 202 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu); |
f41d911f | 203 | |
e4cc1f22 | 204 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 205 | trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 206 | rdp->passed_quiesce = 1; |
b1f77b05 | 207 | } |
64db4cff | 208 | |
25502a6c PM |
209 | /* |
210 | * Note a context switch. This is a quiescent state for RCU-sched, | |
211 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 212 | * The caller must have disabled preemption. |
25502a6c PM |
213 | */ |
214 | void rcu_note_context_switch(int cpu) | |
215 | { | |
f7f7bac9 | 216 | trace_rcu_utilization(TPS("Start context switch")); |
25502a6c | 217 | rcu_sched_qs(cpu); |
cba6d0d6 | 218 | rcu_preempt_note_context_switch(cpu); |
f7f7bac9 | 219 | trace_rcu_utilization(TPS("End context switch")); |
25502a6c | 220 | } |
29ce8310 | 221 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 222 | |
01896f7e | 223 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { |
29e37d81 | 224 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, |
23b5c8fa | 225 | .dynticks = ATOMIC_INIT(1), |
2333210b PM |
226 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
227 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
228 | .dynticks_idle = ATOMIC_INIT(1), | |
229 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
90a4d2c0 | 230 | }; |
64db4cff | 231 | |
878d7439 ED |
232 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
233 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
234 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 235 | |
878d7439 ED |
236 | module_param(blimit, long, 0444); |
237 | module_param(qhimark, long, 0444); | |
238 | module_param(qlowmark, long, 0444); | |
3d76c082 | 239 | |
026ad283 PM |
240 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
241 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
242 | |
243 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
244 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
245 | ||
910ee45d PM |
246 | static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
247 | struct rcu_data *rdp); | |
217af2a2 PM |
248 | static void force_qs_rnp(struct rcu_state *rsp, |
249 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
250 | unsigned long *maxj), | |
251 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 252 | static void force_quiescent_state(struct rcu_state *rsp); |
a157229c | 253 | static int rcu_pending(int cpu); |
64db4cff PM |
254 | |
255 | /* | |
d6714c22 | 256 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
64db4cff | 257 | */ |
d6714c22 | 258 | long rcu_batches_completed_sched(void) |
64db4cff | 259 | { |
d6714c22 | 260 | return rcu_sched_state.completed; |
64db4cff | 261 | } |
d6714c22 | 262 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
263 | |
264 | /* | |
265 | * Return the number of RCU BH batches processed thus far for debug & stats. | |
266 | */ | |
267 | long rcu_batches_completed_bh(void) | |
268 | { | |
269 | return rcu_bh_state.completed; | |
270 | } | |
271 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
272 | ||
bf66f18e PM |
273 | /* |
274 | * Force a quiescent state for RCU BH. | |
275 | */ | |
276 | void rcu_bh_force_quiescent_state(void) | |
277 | { | |
4cdfc175 | 278 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
279 | } |
280 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
281 | ||
4a298656 PM |
282 | /* |
283 | * Record the number of times rcutorture tests have been initiated and | |
284 | * terminated. This information allows the debugfs tracing stats to be | |
285 | * correlated to the rcutorture messages, even when the rcutorture module | |
286 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
287 | * store this state in rcutorture itself. | |
288 | */ | |
289 | void rcutorture_record_test_transition(void) | |
290 | { | |
291 | rcutorture_testseq++; | |
292 | rcutorture_vernum = 0; | |
293 | } | |
294 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
295 | ||
ad0dc7f9 PM |
296 | /* |
297 | * Send along grace-period-related data for rcutorture diagnostics. | |
298 | */ | |
299 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
300 | unsigned long *gpnum, unsigned long *completed) | |
301 | { | |
302 | struct rcu_state *rsp = NULL; | |
303 | ||
304 | switch (test_type) { | |
305 | case RCU_FLAVOR: | |
306 | rsp = rcu_state; | |
307 | break; | |
308 | case RCU_BH_FLAVOR: | |
309 | rsp = &rcu_bh_state; | |
310 | break; | |
311 | case RCU_SCHED_FLAVOR: | |
312 | rsp = &rcu_sched_state; | |
313 | break; | |
314 | default: | |
315 | break; | |
316 | } | |
317 | if (rsp != NULL) { | |
318 | *flags = ACCESS_ONCE(rsp->gp_flags); | |
319 | *gpnum = ACCESS_ONCE(rsp->gpnum); | |
320 | *completed = ACCESS_ONCE(rsp->completed); | |
321 | return; | |
322 | } | |
323 | *flags = 0; | |
324 | *gpnum = 0; | |
325 | *completed = 0; | |
326 | } | |
327 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
328 | ||
4a298656 PM |
329 | /* |
330 | * Record the number of writer passes through the current rcutorture test. | |
331 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
332 | * messages. | |
333 | */ | |
334 | void rcutorture_record_progress(unsigned long vernum) | |
335 | { | |
336 | rcutorture_vernum++; | |
337 | } | |
338 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
339 | ||
bf66f18e PM |
340 | /* |
341 | * Force a quiescent state for RCU-sched. | |
342 | */ | |
343 | void rcu_sched_force_quiescent_state(void) | |
344 | { | |
4cdfc175 | 345 | force_quiescent_state(&rcu_sched_state); |
bf66f18e PM |
346 | } |
347 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
348 | ||
64db4cff PM |
349 | /* |
350 | * Does the CPU have callbacks ready to be invoked? | |
351 | */ | |
352 | static int | |
353 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
354 | { | |
3fbfbf7a PM |
355 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
356 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
357 | } |
358 | ||
359 | /* | |
dc35c893 PM |
360 | * Does the current CPU require a not-yet-started grace period? |
361 | * The caller must have disabled interrupts to prevent races with | |
362 | * normal callback registry. | |
64db4cff PM |
363 | */ |
364 | static int | |
365 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
366 | { | |
dc35c893 | 367 | int i; |
3fbfbf7a | 368 | |
dc35c893 PM |
369 | if (rcu_gp_in_progress(rsp)) |
370 | return 0; /* No, a grace period is already in progress. */ | |
dae6e64d | 371 | if (rcu_nocb_needs_gp(rsp)) |
34ed6246 | 372 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
373 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
374 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
375 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
376 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
377 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
378 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
379 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), | |
380 | rdp->nxtcompleted[i])) | |
381 | return 1; /* Yes, CBs for future grace period. */ | |
382 | return 0; /* No grace period needed. */ | |
64db4cff PM |
383 | } |
384 | ||
385 | /* | |
386 | * Return the root node of the specified rcu_state structure. | |
387 | */ | |
388 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
389 | { | |
390 | return &rsp->node[0]; | |
391 | } | |
392 | ||
9b2e4f18 | 393 | /* |
adf5091e | 394 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
395 | * |
396 | * If the new value of the ->dynticks_nesting counter now is zero, | |
397 | * we really have entered idle, and must do the appropriate accounting. | |
398 | * The caller must have disabled interrupts. | |
399 | */ | |
adf5091e FW |
400 | static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval, |
401 | bool user) | |
9b2e4f18 | 402 | { |
96d3fd0d PM |
403 | struct rcu_state *rsp; |
404 | struct rcu_data *rdp; | |
405 | ||
f7f7bac9 | 406 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 407 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
408 | struct task_struct *idle __maybe_unused = |
409 | idle_task(smp_processor_id()); | |
0989cb46 | 410 | |
f7f7bac9 | 411 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 412 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
413 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
414 | current->pid, current->comm, | |
415 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 416 | } |
96d3fd0d PM |
417 | for_each_rcu_flavor(rsp) { |
418 | rdp = this_cpu_ptr(rsp->rda); | |
419 | do_nocb_deferred_wakeup(rdp); | |
420 | } | |
aea1b35e | 421 | rcu_prepare_for_idle(smp_processor_id()); |
9b2e4f18 PM |
422 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
423 | smp_mb__before_atomic_inc(); /* See above. */ | |
424 | atomic_inc(&rdtp->dynticks); | |
425 | smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */ | |
426 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
c44e2cdd PM |
427 | |
428 | /* | |
adf5091e | 429 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
430 | * in an RCU read-side critical section. |
431 | */ | |
432 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
433 | "Illegal idle entry in RCU read-side critical section."); | |
434 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
435 | "Illegal idle entry in RCU-bh read-side critical section."); | |
436 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
437 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 438 | } |
64db4cff | 439 | |
adf5091e FW |
440 | /* |
441 | * Enter an RCU extended quiescent state, which can be either the | |
442 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 443 | */ |
adf5091e | 444 | static void rcu_eqs_enter(bool user) |
64db4cff | 445 | { |
4145fa7f | 446 | long long oldval; |
64db4cff PM |
447 | struct rcu_dynticks *rdtp; |
448 | ||
c9d4b0af | 449 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 450 | oldval = rdtp->dynticks_nesting; |
29e37d81 | 451 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
3a592405 | 452 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { |
29e37d81 | 453 | rdtp->dynticks_nesting = 0; |
3a592405 PM |
454 | rcu_eqs_enter_common(rdtp, oldval, user); |
455 | } else { | |
29e37d81 | 456 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; |
3a592405 | 457 | } |
64db4cff | 458 | } |
adf5091e FW |
459 | |
460 | /** | |
461 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
462 | * | |
463 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
464 | * read-side critical sections can occur. (Though RCU read-side | |
465 | * critical sections can occur in irq handlers in idle, a possibility | |
466 | * handled by irq_enter() and irq_exit().) | |
467 | * | |
468 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
469 | * the possibility of usermode upcalls having messed up our count | |
470 | * of interrupt nesting level during the prior busy period. | |
471 | */ | |
472 | void rcu_idle_enter(void) | |
473 | { | |
c5d900bf FW |
474 | unsigned long flags; |
475 | ||
476 | local_irq_save(flags); | |
cb349ca9 | 477 | rcu_eqs_enter(false); |
c9d4b0af | 478 | rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0); |
c5d900bf | 479 | local_irq_restore(flags); |
adf5091e | 480 | } |
8a2ecf47 | 481 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 482 | |
2b1d5024 | 483 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
484 | /** |
485 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
486 | * | |
487 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
488 | * is permitted between this call and rcu_user_exit(). This way the | |
489 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
490 | * when the CPU runs in userspace. | |
491 | */ | |
492 | void rcu_user_enter(void) | |
493 | { | |
91d1aa43 | 494 | rcu_eqs_enter(1); |
adf5091e | 495 | } |
2b1d5024 | 496 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 497 | |
9b2e4f18 PM |
498 | /** |
499 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
500 | * | |
501 | * Exit from an interrupt handler, which might possibly result in entering | |
502 | * idle mode, in other words, leaving the mode in which read-side critical | |
503 | * sections can occur. | |
64db4cff | 504 | * |
9b2e4f18 PM |
505 | * This code assumes that the idle loop never does anything that might |
506 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
507 | * architecture violates this assumption, RCU will give you what you | |
508 | * deserve, good and hard. But very infrequently and irreproducibly. | |
509 | * | |
510 | * Use things like work queues to work around this limitation. | |
511 | * | |
512 | * You have been warned. | |
64db4cff | 513 | */ |
9b2e4f18 | 514 | void rcu_irq_exit(void) |
64db4cff PM |
515 | { |
516 | unsigned long flags; | |
4145fa7f | 517 | long long oldval; |
64db4cff PM |
518 | struct rcu_dynticks *rdtp; |
519 | ||
520 | local_irq_save(flags); | |
c9d4b0af | 521 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 522 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
523 | rdtp->dynticks_nesting--; |
524 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 | 525 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 526 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 527 | else |
cb349ca9 | 528 | rcu_eqs_enter_common(rdtp, oldval, true); |
eb348b89 | 529 | rcu_sysidle_enter(rdtp, 1); |
9b2e4f18 PM |
530 | local_irq_restore(flags); |
531 | } | |
532 | ||
533 | /* | |
adf5091e | 534 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
535 | * |
536 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
537 | * we really have exited idle, and must do the appropriate accounting. | |
538 | * The caller must have disabled interrupts. | |
539 | */ | |
adf5091e FW |
540 | static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval, |
541 | int user) | |
9b2e4f18 | 542 | { |
23b5c8fa PM |
543 | smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */ |
544 | atomic_inc(&rdtp->dynticks); | |
545 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
546 | smp_mb__after_atomic_inc(); /* See above. */ | |
547 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
7cb92499 | 548 | rcu_cleanup_after_idle(smp_processor_id()); |
f7f7bac9 | 549 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 550 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
551 | struct task_struct *idle __maybe_unused = |
552 | idle_task(smp_processor_id()); | |
0989cb46 | 553 | |
f7f7bac9 | 554 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 555 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 556 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
557 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
558 | current->pid, current->comm, | |
559 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
560 | } |
561 | } | |
562 | ||
adf5091e FW |
563 | /* |
564 | * Exit an RCU extended quiescent state, which can be either the | |
565 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 566 | */ |
adf5091e | 567 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 568 | { |
9b2e4f18 PM |
569 | struct rcu_dynticks *rdtp; |
570 | long long oldval; | |
571 | ||
c9d4b0af | 572 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 | 573 | oldval = rdtp->dynticks_nesting; |
29e37d81 | 574 | WARN_ON_ONCE(oldval < 0); |
3a592405 | 575 | if (oldval & DYNTICK_TASK_NEST_MASK) { |
29e37d81 | 576 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; |
3a592405 | 577 | } else { |
29e37d81 | 578 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
3a592405 PM |
579 | rcu_eqs_exit_common(rdtp, oldval, user); |
580 | } | |
9b2e4f18 | 581 | } |
adf5091e FW |
582 | |
583 | /** | |
584 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
585 | * | |
586 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
587 | * read-side critical sections can occur. | |
588 | * | |
589 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
590 | * allow for the possibility of usermode upcalls messing up our count | |
591 | * of interrupt nesting level during the busy period that is just | |
592 | * now starting. | |
593 | */ | |
594 | void rcu_idle_exit(void) | |
595 | { | |
c5d900bf FW |
596 | unsigned long flags; |
597 | ||
598 | local_irq_save(flags); | |
cb349ca9 | 599 | rcu_eqs_exit(false); |
c9d4b0af | 600 | rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0); |
c5d900bf | 601 | local_irq_restore(flags); |
adf5091e | 602 | } |
8a2ecf47 | 603 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 604 | |
2b1d5024 | 605 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
606 | /** |
607 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
608 | * | |
609 | * Exit RCU idle mode while entering the kernel because it can | |
610 | * run a RCU read side critical section anytime. | |
611 | */ | |
612 | void rcu_user_exit(void) | |
613 | { | |
91d1aa43 | 614 | rcu_eqs_exit(1); |
adf5091e | 615 | } |
2b1d5024 | 616 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 617 | |
9b2e4f18 PM |
618 | /** |
619 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
620 | * | |
621 | * Enter an interrupt handler, which might possibly result in exiting | |
622 | * idle mode, in other words, entering the mode in which read-side critical | |
623 | * sections can occur. | |
624 | * | |
625 | * Note that the Linux kernel is fully capable of entering an interrupt | |
626 | * handler that it never exits, for example when doing upcalls to | |
627 | * user mode! This code assumes that the idle loop never does upcalls to | |
628 | * user mode. If your architecture does do upcalls from the idle loop (or | |
629 | * does anything else that results in unbalanced calls to the irq_enter() | |
630 | * and irq_exit() functions), RCU will give you what you deserve, good | |
631 | * and hard. But very infrequently and irreproducibly. | |
632 | * | |
633 | * Use things like work queues to work around this limitation. | |
634 | * | |
635 | * You have been warned. | |
636 | */ | |
637 | void rcu_irq_enter(void) | |
638 | { | |
639 | unsigned long flags; | |
640 | struct rcu_dynticks *rdtp; | |
641 | long long oldval; | |
642 | ||
643 | local_irq_save(flags); | |
c9d4b0af | 644 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 PM |
645 | oldval = rdtp->dynticks_nesting; |
646 | rdtp->dynticks_nesting++; | |
647 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 | 648 | if (oldval) |
f7f7bac9 | 649 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 650 | else |
cb349ca9 | 651 | rcu_eqs_exit_common(rdtp, oldval, true); |
eb348b89 | 652 | rcu_sysidle_exit(rdtp, 1); |
64db4cff | 653 | local_irq_restore(flags); |
64db4cff PM |
654 | } |
655 | ||
656 | /** | |
657 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
658 | * | |
659 | * If the CPU was idle with dynamic ticks active, and there is no | |
660 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
661 | * RCU grace-period handling know that the CPU is active. | |
662 | */ | |
663 | void rcu_nmi_enter(void) | |
664 | { | |
c9d4b0af | 665 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 666 | |
23b5c8fa PM |
667 | if (rdtp->dynticks_nmi_nesting == 0 && |
668 | (atomic_read(&rdtp->dynticks) & 0x1)) | |
64db4cff | 669 | return; |
23b5c8fa PM |
670 | rdtp->dynticks_nmi_nesting++; |
671 | smp_mb__before_atomic_inc(); /* Force delay from prior write. */ | |
672 | atomic_inc(&rdtp->dynticks); | |
673 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
674 | smp_mb__after_atomic_inc(); /* See above. */ | |
675 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
64db4cff PM |
676 | } |
677 | ||
678 | /** | |
679 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
680 | * | |
681 | * If the CPU was idle with dynamic ticks active, and there is no | |
682 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
683 | * RCU grace-period handling know that the CPU is no longer active. | |
684 | */ | |
685 | void rcu_nmi_exit(void) | |
686 | { | |
c9d4b0af | 687 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 688 | |
23b5c8fa PM |
689 | if (rdtp->dynticks_nmi_nesting == 0 || |
690 | --rdtp->dynticks_nmi_nesting != 0) | |
64db4cff | 691 | return; |
23b5c8fa PM |
692 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
693 | smp_mb__before_atomic_inc(); /* See above. */ | |
694 | atomic_inc(&rdtp->dynticks); | |
695 | smp_mb__after_atomic_inc(); /* Force delay to next write. */ | |
696 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
64db4cff PM |
697 | } |
698 | ||
699 | /** | |
5c173eb8 PM |
700 | * __rcu_is_watching - are RCU read-side critical sections safe? |
701 | * | |
702 | * Return true if RCU is watching the running CPU, which means that | |
703 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
704 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
705 | * least disabled preemption. | |
706 | */ | |
9418fb20 | 707 | bool notrace __rcu_is_watching(void) |
5c173eb8 PM |
708 | { |
709 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
710 | } | |
711 | ||
712 | /** | |
713 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
64db4cff | 714 | * |
9b2e4f18 | 715 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 716 | * or NMI handler, return true. |
64db4cff | 717 | */ |
9418fb20 | 718 | bool notrace rcu_is_watching(void) |
64db4cff | 719 | { |
34240697 PM |
720 | int ret; |
721 | ||
722 | preempt_disable(); | |
5c173eb8 | 723 | ret = __rcu_is_watching(); |
34240697 PM |
724 | preempt_enable(); |
725 | return ret; | |
64db4cff | 726 | } |
5c173eb8 | 727 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 728 | |
62fde6ed | 729 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
730 | |
731 | /* | |
732 | * Is the current CPU online? Disable preemption to avoid false positives | |
733 | * that could otherwise happen due to the current CPU number being sampled, | |
734 | * this task being preempted, its old CPU being taken offline, resuming | |
735 | * on some other CPU, then determining that its old CPU is now offline. | |
736 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
737 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
738 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
739 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
740 | * offline to continue to use RCU for one jiffy after marking itself | |
741 | * offline in the cpu_online_mask. This leniency is necessary given the | |
742 | * non-atomic nature of the online and offline processing, for example, | |
743 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
744 | * notifiers. | |
745 | * | |
746 | * This is also why RCU internally marks CPUs online during the | |
747 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
748 | * |
749 | * Disable checking if in an NMI handler because we cannot safely report | |
750 | * errors from NMI handlers anyway. | |
751 | */ | |
752 | bool rcu_lockdep_current_cpu_online(void) | |
753 | { | |
2036d94a PM |
754 | struct rcu_data *rdp; |
755 | struct rcu_node *rnp; | |
c0d6d01b PM |
756 | bool ret; |
757 | ||
758 | if (in_nmi()) | |
f6f7ee9a | 759 | return true; |
c0d6d01b | 760 | preempt_disable(); |
c9d4b0af | 761 | rdp = this_cpu_ptr(&rcu_sched_data); |
2036d94a PM |
762 | rnp = rdp->mynode; |
763 | ret = (rdp->grpmask & rnp->qsmaskinit) || | |
c0d6d01b PM |
764 | !rcu_scheduler_fully_active; |
765 | preempt_enable(); | |
766 | return ret; | |
767 | } | |
768 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
769 | ||
62fde6ed | 770 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 771 | |
64db4cff | 772 | /** |
9b2e4f18 | 773 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 774 | * |
9b2e4f18 PM |
775 | * If the current CPU is idle or running at a first-level (not nested) |
776 | * interrupt from idle, return true. The caller must have at least | |
777 | * disabled preemption. | |
64db4cff | 778 | */ |
62e3cb14 | 779 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 780 | { |
c9d4b0af | 781 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
64db4cff PM |
782 | } |
783 | ||
64db4cff PM |
784 | /* |
785 | * Snapshot the specified CPU's dynticks counter so that we can later | |
786 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 787 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 788 | */ |
217af2a2 PM |
789 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
790 | bool *isidle, unsigned long *maxj) | |
64db4cff | 791 | { |
23b5c8fa | 792 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 793 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
f0e7c19d | 794 | return (rdp->dynticks_snap & 0x1) == 0; |
64db4cff PM |
795 | } |
796 | ||
6193c76a PM |
797 | /* |
798 | * This function really isn't for public consumption, but RCU is special in | |
799 | * that context switches can allow the state machine to make progress. | |
800 | */ | |
801 | extern void resched_cpu(int cpu); | |
802 | ||
64db4cff PM |
803 | /* |
804 | * Return true if the specified CPU has passed through a quiescent | |
805 | * state by virtue of being in or having passed through an dynticks | |
806 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 807 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 808 | */ |
217af2a2 PM |
809 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
810 | bool *isidle, unsigned long *maxj) | |
64db4cff | 811 | { |
7eb4f455 PM |
812 | unsigned int curr; |
813 | unsigned int snap; | |
64db4cff | 814 | |
7eb4f455 PM |
815 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
816 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
817 | |
818 | /* | |
819 | * If the CPU passed through or entered a dynticks idle phase with | |
820 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
821 | * already acknowledged the request to pass through a quiescent | |
822 | * state. Either way, that CPU cannot possibly be in an RCU | |
823 | * read-side critical section that started before the beginning | |
824 | * of the current RCU grace period. | |
825 | */ | |
7eb4f455 | 826 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 827 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
828 | rdp->dynticks_fqs++; |
829 | return 1; | |
830 | } | |
831 | ||
a82dcc76 PM |
832 | /* |
833 | * Check for the CPU being offline, but only if the grace period | |
834 | * is old enough. We don't need to worry about the CPU changing | |
835 | * state: If we see it offline even once, it has been through a | |
836 | * quiescent state. | |
837 | * | |
838 | * The reason for insisting that the grace period be at least | |
839 | * one jiffy old is that CPUs that are not quite online and that | |
840 | * have just gone offline can still execute RCU read-side critical | |
841 | * sections. | |
842 | */ | |
843 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
844 | return 0; /* Grace period is not old enough. */ | |
845 | barrier(); | |
846 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 847 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
848 | rdp->offline_fqs++; |
849 | return 1; | |
850 | } | |
65d798f0 PM |
851 | |
852 | /* | |
853 | * There is a possibility that a CPU in adaptive-ticks state | |
854 | * might run in the kernel with the scheduling-clock tick disabled | |
855 | * for an extended time period. Invoke rcu_kick_nohz_cpu() to | |
856 | * force the CPU to restart the scheduling-clock tick in this | |
857 | * CPU is in this state. | |
858 | */ | |
859 | rcu_kick_nohz_cpu(rdp->cpu); | |
860 | ||
6193c76a PM |
861 | /* |
862 | * Alternatively, the CPU might be running in the kernel | |
863 | * for an extended period of time without a quiescent state. | |
864 | * Attempt to force the CPU through the scheduler to gain the | |
865 | * needed quiescent state, but only if the grace period has gone | |
866 | * on for an uncommonly long time. If there are many stuck CPUs, | |
867 | * we will beat on the first one until it gets unstuck, then move | |
868 | * to the next. Only do this for the primary flavor of RCU. | |
869 | */ | |
870 | if (rdp->rsp == rcu_state && | |
cb1e78cf | 871 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { |
6193c76a PM |
872 | rdp->rsp->jiffies_resched += 5; |
873 | resched_cpu(rdp->cpu); | |
874 | } | |
875 | ||
a82dcc76 | 876 | return 0; |
64db4cff PM |
877 | } |
878 | ||
64db4cff PM |
879 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
880 | { | |
cb1e78cf | 881 | unsigned long j = jiffies; |
6193c76a | 882 | unsigned long j1; |
26cdfedf PM |
883 | |
884 | rsp->gp_start = j; | |
885 | smp_wmb(); /* Record start time before stall time. */ | |
6193c76a PM |
886 | j1 = rcu_jiffies_till_stall_check(); |
887 | rsp->jiffies_stall = j + j1; | |
888 | rsp->jiffies_resched = j + j1 / 2; | |
64db4cff PM |
889 | } |
890 | ||
b637a328 PM |
891 | /* |
892 | * Dump stacks of all tasks running on stalled CPUs. This is a fallback | |
893 | * for architectures that do not implement trigger_all_cpu_backtrace(). | |
894 | * The NMI-triggered stack traces are more accurate because they are | |
895 | * printed by the target CPU. | |
896 | */ | |
897 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
898 | { | |
899 | int cpu; | |
900 | unsigned long flags; | |
901 | struct rcu_node *rnp; | |
902 | ||
903 | rcu_for_each_leaf_node(rsp, rnp) { | |
904 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
905 | if (rnp->qsmask != 0) { | |
906 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
907 | if (rnp->qsmask & (1UL << cpu)) | |
908 | dump_cpu_task(rnp->grplo + cpu); | |
909 | } | |
910 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
911 | } | |
912 | } | |
913 | ||
64db4cff PM |
914 | static void print_other_cpu_stall(struct rcu_state *rsp) |
915 | { | |
916 | int cpu; | |
917 | long delta; | |
918 | unsigned long flags; | |
285fe294 | 919 | int ndetected = 0; |
64db4cff | 920 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 921 | long totqlen = 0; |
64db4cff PM |
922 | |
923 | /* Only let one CPU complain about others per time interval. */ | |
924 | ||
1304afb2 | 925 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 926 | delta = jiffies - rsp->jiffies_stall; |
fc2219d4 | 927 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 928 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
929 | return; |
930 | } | |
6bfc09e2 | 931 | rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 932 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 933 | |
8cdd32a9 PM |
934 | /* |
935 | * OK, time to rat on our buddy... | |
936 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
937 | * RCU CPU stall warnings. | |
938 | */ | |
d7f3e207 | 939 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 940 | rsp->name); |
a858af28 | 941 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 942 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 943 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 944 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
945 | if (rnp->qsmask != 0) { |
946 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
947 | if (rnp->qsmask & (1UL << cpu)) { | |
948 | print_cpu_stall_info(rsp, | |
949 | rnp->grplo + cpu); | |
950 | ndetected++; | |
951 | } | |
952 | } | |
3acd9eb3 | 953 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 954 | } |
a858af28 PM |
955 | |
956 | /* | |
957 | * Now rat on any tasks that got kicked up to the root rcu_node | |
958 | * due to CPU offlining. | |
959 | */ | |
960 | rnp = rcu_get_root(rsp); | |
961 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
285fe294 | 962 | ndetected += rcu_print_task_stall(rnp); |
a858af28 PM |
963 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
964 | ||
965 | print_cpu_stall_info_end(); | |
53bb857c PM |
966 | for_each_possible_cpu(cpu) |
967 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
968 | pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n", | |
eee05882 | 969 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
53bb857c | 970 | rsp->gpnum, rsp->completed, totqlen); |
9bc8b558 | 971 | if (ndetected == 0) |
d7f3e207 | 972 | pr_err("INFO: Stall ended before state dump start\n"); |
9bc8b558 | 973 | else if (!trigger_all_cpu_backtrace()) |
b637a328 | 974 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 975 | |
4cdfc175 | 976 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
977 | |
978 | rcu_print_detail_task_stall(rsp); | |
979 | ||
4cdfc175 | 980 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
981 | } |
982 | ||
b021fe3e PZ |
983 | /* |
984 | * This function really isn't for public consumption, but RCU is special in | |
985 | * that context switches can allow the state machine to make progress. | |
986 | */ | |
987 | extern void resched_cpu(int cpu); | |
988 | ||
64db4cff PM |
989 | static void print_cpu_stall(struct rcu_state *rsp) |
990 | { | |
53bb857c | 991 | int cpu; |
64db4cff PM |
992 | unsigned long flags; |
993 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 994 | long totqlen = 0; |
64db4cff | 995 | |
8cdd32a9 PM |
996 | /* |
997 | * OK, time to rat on ourselves... | |
998 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
999 | * RCU CPU stall warnings. | |
1000 | */ | |
d7f3e207 | 1001 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
1002 | print_cpu_stall_info_begin(); |
1003 | print_cpu_stall_info(rsp, smp_processor_id()); | |
1004 | print_cpu_stall_info_end(); | |
53bb857c PM |
1005 | for_each_possible_cpu(cpu) |
1006 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
1007 | pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n", | |
1008 | jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen); | |
4627e240 PM |
1009 | if (!trigger_all_cpu_backtrace()) |
1010 | dump_stack(); | |
c1dc0b9c | 1011 | |
1304afb2 | 1012 | raw_spin_lock_irqsave(&rnp->lock, flags); |
20133cfc | 1013 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall)) |
13cfcca0 | 1014 | rsp->jiffies_stall = jiffies + |
6bfc09e2 | 1015 | 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 1016 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 1017 | |
b021fe3e PZ |
1018 | /* |
1019 | * Attempt to revive the RCU machinery by forcing a context switch. | |
1020 | * | |
1021 | * A context switch would normally allow the RCU state machine to make | |
1022 | * progress and it could be we're stuck in kernel space without context | |
1023 | * switches for an entirely unreasonable amount of time. | |
1024 | */ | |
1025 | resched_cpu(smp_processor_id()); | |
64db4cff PM |
1026 | } |
1027 | ||
1028 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
1029 | { | |
26cdfedf PM |
1030 | unsigned long completed; |
1031 | unsigned long gpnum; | |
1032 | unsigned long gps; | |
bad6e139 PM |
1033 | unsigned long j; |
1034 | unsigned long js; | |
64db4cff PM |
1035 | struct rcu_node *rnp; |
1036 | ||
26cdfedf | 1037 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
c68de209 | 1038 | return; |
cb1e78cf | 1039 | j = jiffies; |
26cdfedf PM |
1040 | |
1041 | /* | |
1042 | * Lots of memory barriers to reject false positives. | |
1043 | * | |
1044 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
1045 | * then rsp->gp_start, and finally rsp->completed. These values | |
1046 | * are updated in the opposite order with memory barriers (or | |
1047 | * equivalent) during grace-period initialization and cleanup. | |
1048 | * Now, a false positive can occur if we get an new value of | |
1049 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
1050 | * the memory barriers, the only way that this can happen is if one | |
1051 | * grace period ends and another starts between these two fetches. | |
1052 | * Detect this by comparing rsp->completed with the previous fetch | |
1053 | * from rsp->gpnum. | |
1054 | * | |
1055 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
1056 | * and rsp->gp_start suffice to forestall false positives. | |
1057 | */ | |
1058 | gpnum = ACCESS_ONCE(rsp->gpnum); | |
1059 | smp_rmb(); /* Pick up ->gpnum first... */ | |
bad6e139 | 1060 | js = ACCESS_ONCE(rsp->jiffies_stall); |
26cdfedf PM |
1061 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
1062 | gps = ACCESS_ONCE(rsp->gp_start); | |
1063 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ | |
1064 | completed = ACCESS_ONCE(rsp->completed); | |
1065 | if (ULONG_CMP_GE(completed, gpnum) || | |
1066 | ULONG_CMP_LT(j, js) || | |
1067 | ULONG_CMP_GE(gps, js)) | |
1068 | return; /* No stall or GP completed since entering function. */ | |
64db4cff | 1069 | rnp = rdp->mynode; |
c96ea7cf | 1070 | if (rcu_gp_in_progress(rsp) && |
26cdfedf | 1071 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) { |
64db4cff PM |
1072 | |
1073 | /* We haven't checked in, so go dump stack. */ | |
1074 | print_cpu_stall(rsp); | |
1075 | ||
bad6e139 PM |
1076 | } else if (rcu_gp_in_progress(rsp) && |
1077 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 1078 | |
bad6e139 | 1079 | /* They had a few time units to dump stack, so complain. */ |
64db4cff PM |
1080 | print_other_cpu_stall(rsp); |
1081 | } | |
1082 | } | |
1083 | ||
53d84e00 PM |
1084 | /** |
1085 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1086 | * | |
1087 | * Set the stall-warning timeout way off into the future, thus preventing | |
1088 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1089 | * RCU grace periods. | |
1090 | * | |
1091 | * The caller must disable hard irqs. | |
1092 | */ | |
1093 | void rcu_cpu_stall_reset(void) | |
1094 | { | |
6ce75a23 PM |
1095 | struct rcu_state *rsp; |
1096 | ||
1097 | for_each_rcu_flavor(rsp) | |
1098 | rsp->jiffies_stall = jiffies + ULONG_MAX / 2; | |
53d84e00 PM |
1099 | } |
1100 | ||
3f5d3ea6 PM |
1101 | /* |
1102 | * Initialize the specified rcu_data structure's callback list to empty. | |
1103 | */ | |
1104 | static void init_callback_list(struct rcu_data *rdp) | |
1105 | { | |
1106 | int i; | |
1107 | ||
34ed6246 PM |
1108 | if (init_nocb_callback_list(rdp)) |
1109 | return; | |
3f5d3ea6 PM |
1110 | rdp->nxtlist = NULL; |
1111 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1112 | rdp->nxttail[i] = &rdp->nxtlist; | |
1113 | } | |
1114 | ||
dc35c893 PM |
1115 | /* |
1116 | * Determine the value that ->completed will have at the end of the | |
1117 | * next subsequent grace period. This is used to tag callbacks so that | |
1118 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1119 | * been dyntick-idle for an extended period with callbacks under the | |
1120 | * influence of RCU_FAST_NO_HZ. | |
1121 | * | |
1122 | * The caller must hold rnp->lock with interrupts disabled. | |
1123 | */ | |
1124 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1125 | struct rcu_node *rnp) | |
1126 | { | |
1127 | /* | |
1128 | * If RCU is idle, we just wait for the next grace period. | |
1129 | * But we can only be sure that RCU is idle if we are looking | |
1130 | * at the root rcu_node structure -- otherwise, a new grace | |
1131 | * period might have started, but just not yet gotten around | |
1132 | * to initializing the current non-root rcu_node structure. | |
1133 | */ | |
1134 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1135 | return rnp->completed + 1; | |
1136 | ||
1137 | /* | |
1138 | * Otherwise, wait for a possible partial grace period and | |
1139 | * then the subsequent full grace period. | |
1140 | */ | |
1141 | return rnp->completed + 2; | |
1142 | } | |
1143 | ||
0446be48 PM |
1144 | /* |
1145 | * Trace-event helper function for rcu_start_future_gp() and | |
1146 | * rcu_nocb_wait_gp(). | |
1147 | */ | |
1148 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1149 | unsigned long c, const char *s) |
0446be48 PM |
1150 | { |
1151 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1152 | rnp->completed, c, rnp->level, | |
1153 | rnp->grplo, rnp->grphi, s); | |
1154 | } | |
1155 | ||
1156 | /* | |
1157 | * Start some future grace period, as needed to handle newly arrived | |
1158 | * callbacks. The required future grace periods are recorded in each | |
1159 | * rcu_node structure's ->need_future_gp field. | |
1160 | * | |
1161 | * The caller must hold the specified rcu_node structure's ->lock. | |
1162 | */ | |
1163 | static unsigned long __maybe_unused | |
1164 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp) | |
1165 | { | |
1166 | unsigned long c; | |
1167 | int i; | |
1168 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); | |
1169 | ||
1170 | /* | |
1171 | * Pick up grace-period number for new callbacks. If this | |
1172 | * grace period is already marked as needed, return to the caller. | |
1173 | */ | |
1174 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1175 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1176 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1177 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
0446be48 PM |
1178 | return c; |
1179 | } | |
1180 | ||
1181 | /* | |
1182 | * If either this rcu_node structure or the root rcu_node structure | |
1183 | * believe that a grace period is in progress, then we must wait | |
1184 | * for the one following, which is in "c". Because our request | |
1185 | * will be noticed at the end of the current grace period, we don't | |
1186 | * need to explicitly start one. | |
1187 | */ | |
1188 | if (rnp->gpnum != rnp->completed || | |
1189 | ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) { | |
1190 | rnp->need_future_gp[c & 0x1]++; | |
f7f7bac9 | 1191 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
0446be48 PM |
1192 | return c; |
1193 | } | |
1194 | ||
1195 | /* | |
1196 | * There might be no grace period in progress. If we don't already | |
1197 | * hold it, acquire the root rcu_node structure's lock in order to | |
1198 | * start one (if needed). | |
1199 | */ | |
6303b9c8 | 1200 | if (rnp != rnp_root) { |
0446be48 | 1201 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 PM |
1202 | smp_mb__after_unlock_lock(); |
1203 | } | |
0446be48 PM |
1204 | |
1205 | /* | |
1206 | * Get a new grace-period number. If there really is no grace | |
1207 | * period in progress, it will be smaller than the one we obtained | |
1208 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1209 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1210 | */ | |
1211 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1212 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1213 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1214 | rdp->nxtcompleted[i] = c; | |
1215 | ||
1216 | /* | |
1217 | * If the needed for the required grace period is already | |
1218 | * recorded, trace and leave. | |
1219 | */ | |
1220 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1221 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1222 | goto unlock_out; |
1223 | } | |
1224 | ||
1225 | /* Record the need for the future grace period. */ | |
1226 | rnp_root->need_future_gp[c & 0x1]++; | |
1227 | ||
1228 | /* If a grace period is not already in progress, start one. */ | |
1229 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1230 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1231 | } else { |
f7f7bac9 | 1232 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
910ee45d | 1233 | rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1234 | } |
1235 | unlock_out: | |
1236 | if (rnp != rnp_root) | |
1237 | raw_spin_unlock(&rnp_root->lock); | |
1238 | return c; | |
1239 | } | |
1240 | ||
1241 | /* | |
1242 | * Clean up any old requests for the just-ended grace period. Also return | |
1243 | * whether any additional grace periods have been requested. Also invoke | |
1244 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1245 | * waiting for this grace period to complete. | |
1246 | */ | |
1247 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1248 | { | |
1249 | int c = rnp->completed; | |
1250 | int needmore; | |
1251 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1252 | ||
1253 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1254 | rnp->need_future_gp[c & 0x1] = 0; | |
1255 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1256 | trace_rcu_future_gp(rnp, rdp, c, |
1257 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1258 | return needmore; |
1259 | } | |
1260 | ||
dc35c893 PM |
1261 | /* |
1262 | * If there is room, assign a ->completed number to any callbacks on | |
1263 | * this CPU that have not already been assigned. Also accelerate any | |
1264 | * callbacks that were previously assigned a ->completed number that has | |
1265 | * since proven to be too conservative, which can happen if callbacks get | |
1266 | * assigned a ->completed number while RCU is idle, but with reference to | |
1267 | * a non-root rcu_node structure. This function is idempotent, so it does | |
1268 | * not hurt to call it repeatedly. | |
1269 | * | |
1270 | * The caller must hold rnp->lock with interrupts disabled. | |
1271 | */ | |
1272 | static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1273 | struct rcu_data *rdp) | |
1274 | { | |
1275 | unsigned long c; | |
1276 | int i; | |
1277 | ||
1278 | /* If the CPU has no callbacks, nothing to do. */ | |
1279 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1280 | return; | |
1281 | ||
1282 | /* | |
1283 | * Starting from the sublist containing the callbacks most | |
1284 | * recently assigned a ->completed number and working down, find the | |
1285 | * first sublist that is not assignable to an upcoming grace period. | |
1286 | * Such a sublist has something in it (first two tests) and has | |
1287 | * a ->completed number assigned that will complete sooner than | |
1288 | * the ->completed number for newly arrived callbacks (last test). | |
1289 | * | |
1290 | * The key point is that any later sublist can be assigned the | |
1291 | * same ->completed number as the newly arrived callbacks, which | |
1292 | * means that the callbacks in any of these later sublist can be | |
1293 | * grouped into a single sublist, whether or not they have already | |
1294 | * been assigned a ->completed number. | |
1295 | */ | |
1296 | c = rcu_cbs_completed(rsp, rnp); | |
1297 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1298 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1299 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1300 | break; | |
1301 | ||
1302 | /* | |
1303 | * If there are no sublist for unassigned callbacks, leave. | |
1304 | * At the same time, advance "i" one sublist, so that "i" will | |
1305 | * index into the sublist where all the remaining callbacks should | |
1306 | * be grouped into. | |
1307 | */ | |
1308 | if (++i >= RCU_NEXT_TAIL) | |
1309 | return; | |
1310 | ||
1311 | /* | |
1312 | * Assign all subsequent callbacks' ->completed number to the next | |
1313 | * full grace period and group them all in the sublist initially | |
1314 | * indexed by "i". | |
1315 | */ | |
1316 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1317 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1318 | rdp->nxtcompleted[i] = c; | |
1319 | } | |
910ee45d PM |
1320 | /* Record any needed additional grace periods. */ |
1321 | rcu_start_future_gp(rnp, rdp); | |
6d4b418c PM |
1322 | |
1323 | /* Trace depending on how much we were able to accelerate. */ | |
1324 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1325 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1326 | else |
f7f7bac9 | 1327 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
dc35c893 PM |
1328 | } |
1329 | ||
1330 | /* | |
1331 | * Move any callbacks whose grace period has completed to the | |
1332 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1333 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1334 | * sublist. This function is idempotent, so it does not hurt to | |
1335 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
1336 | * | |
1337 | * The caller must hold rnp->lock with interrupts disabled. | |
1338 | */ | |
1339 | static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1340 | struct rcu_data *rdp) | |
1341 | { | |
1342 | int i, j; | |
1343 | ||
1344 | /* If the CPU has no callbacks, nothing to do. */ | |
1345 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1346 | return; | |
1347 | ||
1348 | /* | |
1349 | * Find all callbacks whose ->completed numbers indicate that they | |
1350 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1351 | */ | |
1352 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1353 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1354 | break; | |
1355 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1356 | } | |
1357 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1358 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1359 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1360 | ||
1361 | /* Copy down callbacks to fill in empty sublists. */ | |
1362 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1363 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1364 | break; | |
1365 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1366 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1367 | } | |
1368 | ||
1369 | /* Classify any remaining callbacks. */ | |
1370 | rcu_accelerate_cbs(rsp, rnp, rdp); | |
1371 | } | |
1372 | ||
d09b62df | 1373 | /* |
ba9fbe95 PM |
1374 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1375 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1376 | * structure corresponding to the current CPU, and must have irqs disabled. | |
d09b62df | 1377 | */ |
ba9fbe95 | 1378 | static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) |
d09b62df | 1379 | { |
ba9fbe95 | 1380 | /* Handle the ends of any preceding grace periods first. */ |
dc35c893 | 1381 | if (rdp->completed == rnp->completed) { |
d09b62df | 1382 | |
ba9fbe95 | 1383 | /* No grace period end, so just accelerate recent callbacks. */ |
dc35c893 | 1384 | rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1385 | |
dc35c893 PM |
1386 | } else { |
1387 | ||
1388 | /* Advance callbacks. */ | |
1389 | rcu_advance_cbs(rsp, rnp, rdp); | |
d09b62df PM |
1390 | |
1391 | /* Remember that we saw this grace-period completion. */ | |
1392 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1393 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1394 | } |
398ebe60 | 1395 | |
6eaef633 PM |
1396 | if (rdp->gpnum != rnp->gpnum) { |
1397 | /* | |
1398 | * If the current grace period is waiting for this CPU, | |
1399 | * set up to detect a quiescent state, otherwise don't | |
1400 | * go looking for one. | |
1401 | */ | |
1402 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1403 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
6eaef633 PM |
1404 | rdp->passed_quiesce = 0; |
1405 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); | |
1406 | zero_cpu_stall_ticks(rdp); | |
1407 | } | |
1408 | } | |
1409 | ||
d34ea322 | 1410 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1411 | { |
1412 | unsigned long flags; | |
1413 | struct rcu_node *rnp; | |
1414 | ||
1415 | local_irq_save(flags); | |
1416 | rnp = rdp->mynode; | |
d34ea322 PM |
1417 | if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) && |
1418 | rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */ | |
6eaef633 PM |
1419 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
1420 | local_irq_restore(flags); | |
1421 | return; | |
1422 | } | |
6303b9c8 | 1423 | smp_mb__after_unlock_lock(); |
d34ea322 | 1424 | __note_gp_changes(rsp, rnp, rdp); |
6eaef633 PM |
1425 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1426 | } | |
1427 | ||
b3dbec76 | 1428 | /* |
f7be8209 | 1429 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1430 | */ |
7fdefc10 | 1431 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 PM |
1432 | { |
1433 | struct rcu_data *rdp; | |
7fdefc10 | 1434 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1435 | |
eb75767b | 1436 | rcu_bind_gp_kthread(); |
7fdefc10 | 1437 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1438 | smp_mb__after_unlock_lock(); |
f7be8209 PM |
1439 | if (rsp->gp_flags == 0) { |
1440 | /* Spurious wakeup, tell caller to go back to sleep. */ | |
1441 | raw_spin_unlock_irq(&rnp->lock); | |
1442 | return 0; | |
1443 | } | |
4cdfc175 | 1444 | rsp->gp_flags = 0; /* Clear all flags: New grace period. */ |
b3dbec76 | 1445 | |
f7be8209 PM |
1446 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1447 | /* | |
1448 | * Grace period already in progress, don't start another. | |
1449 | * Not supposed to be able to happen. | |
1450 | */ | |
7fdefc10 PM |
1451 | raw_spin_unlock_irq(&rnp->lock); |
1452 | return 0; | |
1453 | } | |
1454 | ||
7fdefc10 | 1455 | /* Advance to a new grace period and initialize state. */ |
26cdfedf | 1456 | record_gp_stall_check_time(rsp); |
765a3f4f PM |
1457 | /* Record GP times before starting GP, hence smp_store_release(). */ |
1458 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); | |
f7f7bac9 | 1459 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1460 | raw_spin_unlock_irq(&rnp->lock); |
1461 | ||
1462 | /* Exclude any concurrent CPU-hotplug operations. */ | |
a4fbe35a | 1463 | mutex_lock(&rsp->onoff_mutex); |
765a3f4f | 1464 | smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */ |
7fdefc10 PM |
1465 | |
1466 | /* | |
1467 | * Set the quiescent-state-needed bits in all the rcu_node | |
1468 | * structures for all currently online CPUs in breadth-first order, | |
1469 | * starting from the root rcu_node structure, relying on the layout | |
1470 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1471 | * will access only the leaves of the hierarchy, thus seeing that no | |
1472 | * grace period is in progress, at least until the corresponding | |
1473 | * leaf node has been initialized. In addition, we have excluded | |
1474 | * CPU-hotplug operations. | |
1475 | * | |
1476 | * The grace period cannot complete until the initialization | |
1477 | * process finishes, because this kthread handles both. | |
1478 | */ | |
1479 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
b3dbec76 | 1480 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1481 | smp_mb__after_unlock_lock(); |
b3dbec76 | 1482 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1483 | rcu_preempt_check_blocked_tasks(rnp); |
1484 | rnp->qsmask = rnp->qsmaskinit; | |
0446be48 | 1485 | ACCESS_ONCE(rnp->gpnum) = rsp->gpnum; |
25d30cf4 | 1486 | WARN_ON_ONCE(rnp->completed != rsp->completed); |
0446be48 | 1487 | ACCESS_ONCE(rnp->completed) = rsp->completed; |
7fdefc10 | 1488 | if (rnp == rdp->mynode) |
ce3d9c03 | 1489 | __note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1490 | rcu_preempt_boost_start_gp(rnp); |
1491 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1492 | rnp->level, rnp->grplo, | |
1493 | rnp->grphi, rnp->qsmask); | |
1494 | raw_spin_unlock_irq(&rnp->lock); | |
661a85dc | 1495 | #ifdef CONFIG_PROVE_RCU_DELAY |
971394f3 | 1496 | if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 && |
81e59494 | 1497 | system_state == SYSTEM_RUNNING) |
971394f3 | 1498 | udelay(200); |
661a85dc | 1499 | #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */ |
7fdefc10 PM |
1500 | cond_resched(); |
1501 | } | |
b3dbec76 | 1502 | |
a4fbe35a | 1503 | mutex_unlock(&rsp->onoff_mutex); |
7fdefc10 PM |
1504 | return 1; |
1505 | } | |
b3dbec76 | 1506 | |
4cdfc175 PM |
1507 | /* |
1508 | * Do one round of quiescent-state forcing. | |
1509 | */ | |
01896f7e | 1510 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) |
4cdfc175 PM |
1511 | { |
1512 | int fqs_state = fqs_state_in; | |
217af2a2 PM |
1513 | bool isidle = false; |
1514 | unsigned long maxj; | |
4cdfc175 PM |
1515 | struct rcu_node *rnp = rcu_get_root(rsp); |
1516 | ||
1517 | rsp->n_force_qs++; | |
1518 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1519 | /* Collect dyntick-idle snapshots. */ | |
0edd1b17 PM |
1520 | if (is_sysidle_rcu_state(rsp)) { |
1521 | isidle = 1; | |
1522 | maxj = jiffies - ULONG_MAX / 4; | |
1523 | } | |
217af2a2 PM |
1524 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1525 | &isidle, &maxj); | |
0edd1b17 | 1526 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1527 | fqs_state = RCU_FORCE_QS; |
1528 | } else { | |
1529 | /* Handle dyntick-idle and offline CPUs. */ | |
0edd1b17 | 1530 | isidle = 0; |
217af2a2 | 1531 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1532 | } |
1533 | /* Clear flag to prevent immediate re-entry. */ | |
1534 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1535 | raw_spin_lock_irq(&rnp->lock); | |
6303b9c8 | 1536 | smp_mb__after_unlock_lock(); |
4cdfc175 PM |
1537 | rsp->gp_flags &= ~RCU_GP_FLAG_FQS; |
1538 | raw_spin_unlock_irq(&rnp->lock); | |
1539 | } | |
1540 | return fqs_state; | |
1541 | } | |
1542 | ||
7fdefc10 PM |
1543 | /* |
1544 | * Clean up after the old grace period. | |
1545 | */ | |
4cdfc175 | 1546 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1547 | { |
1548 | unsigned long gp_duration; | |
dae6e64d | 1549 | int nocb = 0; |
7fdefc10 PM |
1550 | struct rcu_data *rdp; |
1551 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1552 | |
7fdefc10 | 1553 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1554 | smp_mb__after_unlock_lock(); |
7fdefc10 PM |
1555 | gp_duration = jiffies - rsp->gp_start; |
1556 | if (gp_duration > rsp->gp_max) | |
1557 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1558 | |
7fdefc10 PM |
1559 | /* |
1560 | * We know the grace period is complete, but to everyone else | |
1561 | * it appears to still be ongoing. But it is also the case | |
1562 | * that to everyone else it looks like there is nothing that | |
1563 | * they can do to advance the grace period. It is therefore | |
1564 | * safe for us to drop the lock in order to mark the grace | |
1565 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1566 | */ |
5d4b8659 | 1567 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1568 | |
5d4b8659 PM |
1569 | /* |
1570 | * Propagate new ->completed value to rcu_node structures so | |
1571 | * that other CPUs don't have to wait until the start of the next | |
1572 | * grace period to process their callbacks. This also avoids | |
1573 | * some nasty RCU grace-period initialization races by forcing | |
1574 | * the end of the current grace period to be completely recorded in | |
1575 | * all of the rcu_node structures before the beginning of the next | |
1576 | * grace period is recorded in any of the rcu_node structures. | |
1577 | */ | |
1578 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1579 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1580 | smp_mb__after_unlock_lock(); |
0446be48 | 1581 | ACCESS_ONCE(rnp->completed) = rsp->gpnum; |
b11cc576 PM |
1582 | rdp = this_cpu_ptr(rsp->rda); |
1583 | if (rnp == rdp->mynode) | |
470716fc | 1584 | __note_gp_changes(rsp, rnp, rdp); |
78e4bc34 | 1585 | /* smp_mb() provided by prior unlock-lock pair. */ |
0446be48 | 1586 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 PM |
1587 | raw_spin_unlock_irq(&rnp->lock); |
1588 | cond_resched(); | |
7fdefc10 | 1589 | } |
5d4b8659 PM |
1590 | rnp = rcu_get_root(rsp); |
1591 | raw_spin_lock_irq(&rnp->lock); | |
765a3f4f | 1592 | smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */ |
dae6e64d | 1593 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 | 1594 | |
765a3f4f PM |
1595 | /* Declare grace period done. */ |
1596 | ACCESS_ONCE(rsp->completed) = rsp->gpnum; | |
f7f7bac9 | 1597 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
7fdefc10 | 1598 | rsp->fqs_state = RCU_GP_IDLE; |
5d4b8659 | 1599 | rdp = this_cpu_ptr(rsp->rda); |
b11cc576 | 1600 | rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */ |
bb311ecc | 1601 | if (cpu_needs_another_gp(rsp, rdp)) { |
b3f2d025 | 1602 | rsp->gp_flags = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1603 | trace_rcu_grace_period(rsp->name, |
1604 | ACCESS_ONCE(rsp->gpnum), | |
1605 | TPS("newreq")); | |
1606 | } | |
7fdefc10 | 1607 | raw_spin_unlock_irq(&rnp->lock); |
7fdefc10 PM |
1608 | } |
1609 | ||
1610 | /* | |
1611 | * Body of kthread that handles grace periods. | |
1612 | */ | |
1613 | static int __noreturn rcu_gp_kthread(void *arg) | |
1614 | { | |
4cdfc175 | 1615 | int fqs_state; |
88d6df61 | 1616 | int gf; |
d40011f6 | 1617 | unsigned long j; |
4cdfc175 | 1618 | int ret; |
7fdefc10 PM |
1619 | struct rcu_state *rsp = arg; |
1620 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1621 | ||
1622 | for (;;) { | |
1623 | ||
1624 | /* Handle grace-period start. */ | |
1625 | for (;;) { | |
63c4db78 PM |
1626 | trace_rcu_grace_period(rsp->name, |
1627 | ACCESS_ONCE(rsp->gpnum), | |
1628 | TPS("reqwait")); | |
4cdfc175 | 1629 | wait_event_interruptible(rsp->gp_wq, |
591c6d17 | 1630 | ACCESS_ONCE(rsp->gp_flags) & |
4cdfc175 | 1631 | RCU_GP_FLAG_INIT); |
78e4bc34 | 1632 | /* Locking provides needed memory barrier. */ |
f7be8209 | 1633 | if (rcu_gp_init(rsp)) |
7fdefc10 PM |
1634 | break; |
1635 | cond_resched(); | |
1636 | flush_signals(current); | |
63c4db78 PM |
1637 | trace_rcu_grace_period(rsp->name, |
1638 | ACCESS_ONCE(rsp->gpnum), | |
1639 | TPS("reqwaitsig")); | |
7fdefc10 | 1640 | } |
cabc49c1 | 1641 | |
4cdfc175 PM |
1642 | /* Handle quiescent-state forcing. */ |
1643 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
1644 | j = jiffies_till_first_fqs; |
1645 | if (j > HZ) { | |
1646 | j = HZ; | |
1647 | jiffies_till_first_fqs = HZ; | |
1648 | } | |
88d6df61 | 1649 | ret = 0; |
cabc49c1 | 1650 | for (;;) { |
88d6df61 PM |
1651 | if (!ret) |
1652 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 PM |
1653 | trace_rcu_grace_period(rsp->name, |
1654 | ACCESS_ONCE(rsp->gpnum), | |
1655 | TPS("fqswait")); | |
4cdfc175 | 1656 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
88d6df61 PM |
1657 | ((gf = ACCESS_ONCE(rsp->gp_flags)) & |
1658 | RCU_GP_FLAG_FQS) || | |
4cdfc175 PM |
1659 | (!ACCESS_ONCE(rnp->qsmask) && |
1660 | !rcu_preempt_blocked_readers_cgp(rnp)), | |
d40011f6 | 1661 | j); |
78e4bc34 | 1662 | /* Locking provides needed memory barriers. */ |
4cdfc175 | 1663 | /* If grace period done, leave loop. */ |
cabc49c1 | 1664 | if (!ACCESS_ONCE(rnp->qsmask) && |
4cdfc175 | 1665 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 1666 | break; |
4cdfc175 | 1667 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
1668 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
1669 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 PM |
1670 | trace_rcu_grace_period(rsp->name, |
1671 | ACCESS_ONCE(rsp->gpnum), | |
1672 | TPS("fqsstart")); | |
4cdfc175 | 1673 | fqs_state = rcu_gp_fqs(rsp, fqs_state); |
63c4db78 PM |
1674 | trace_rcu_grace_period(rsp->name, |
1675 | ACCESS_ONCE(rsp->gpnum), | |
1676 | TPS("fqsend")); | |
4cdfc175 PM |
1677 | cond_resched(); |
1678 | } else { | |
1679 | /* Deal with stray signal. */ | |
1680 | cond_resched(); | |
1681 | flush_signals(current); | |
63c4db78 PM |
1682 | trace_rcu_grace_period(rsp->name, |
1683 | ACCESS_ONCE(rsp->gpnum), | |
1684 | TPS("fqswaitsig")); | |
4cdfc175 | 1685 | } |
d40011f6 PM |
1686 | j = jiffies_till_next_fqs; |
1687 | if (j > HZ) { | |
1688 | j = HZ; | |
1689 | jiffies_till_next_fqs = HZ; | |
1690 | } else if (j < 1) { | |
1691 | j = 1; | |
1692 | jiffies_till_next_fqs = 1; | |
1693 | } | |
cabc49c1 | 1694 | } |
4cdfc175 PM |
1695 | |
1696 | /* Handle grace-period end. */ | |
1697 | rcu_gp_cleanup(rsp); | |
b3dbec76 | 1698 | } |
b3dbec76 PM |
1699 | } |
1700 | ||
016a8d5b SR |
1701 | static void rsp_wakeup(struct irq_work *work) |
1702 | { | |
1703 | struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work); | |
1704 | ||
1705 | /* Wake up rcu_gp_kthread() to start the grace period. */ | |
1706 | wake_up(&rsp->gp_wq); | |
1707 | } | |
1708 | ||
64db4cff PM |
1709 | /* |
1710 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
1711 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 1712 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
1713 | * |
1714 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
1715 | * invoke this function. This can happen when the dying CPU reports its | |
1716 | * quiescent state. | |
64db4cff PM |
1717 | */ |
1718 | static void | |
910ee45d PM |
1719 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
1720 | struct rcu_data *rdp) | |
64db4cff | 1721 | { |
b8462084 | 1722 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 1723 | /* |
b3dbec76 | 1724 | * Either we have not yet spawned the grace-period |
62da1921 PM |
1725 | * task, this CPU does not need another grace period, |
1726 | * or a grace period is already in progress. | |
b3dbec76 | 1727 | * Either way, don't start a new grace period. |
afe24b12 | 1728 | */ |
afe24b12 PM |
1729 | return; |
1730 | } | |
4cdfc175 | 1731 | rsp->gp_flags = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1732 | trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum), |
1733 | TPS("newreq")); | |
62da1921 | 1734 | |
016a8d5b SR |
1735 | /* |
1736 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c PM |
1737 | * could cause possible deadlocks with the rq->lock. Defer |
1738 | * the wakeup to interrupt context. And don't bother waking | |
1739 | * up the running kthread. | |
016a8d5b | 1740 | */ |
1eafd31c PM |
1741 | if (current != rsp->gp_kthread) |
1742 | irq_work_queue(&rsp->wakeup_work); | |
64db4cff PM |
1743 | } |
1744 | ||
910ee45d PM |
1745 | /* |
1746 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
1747 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
1748 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
1749 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
1750 | * that is encountered beforehand. | |
1751 | */ | |
1752 | static void | |
1753 | rcu_start_gp(struct rcu_state *rsp) | |
1754 | { | |
1755 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1756 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1757 | ||
1758 | /* | |
1759 | * If there is no grace period in progress right now, any | |
1760 | * callbacks we have up to this point will be satisfied by the | |
1761 | * next grace period. Also, advancing the callbacks reduces the | |
1762 | * probability of false positives from cpu_needs_another_gp() | |
1763 | * resulting in pointless grace periods. So, advance callbacks | |
1764 | * then start the grace period! | |
1765 | */ | |
1766 | rcu_advance_cbs(rsp, rnp, rdp); | |
1767 | rcu_start_gp_advanced(rsp, rnp, rdp); | |
1768 | } | |
1769 | ||
f41d911f | 1770 | /* |
d3f6bad3 PM |
1771 | * Report a full set of quiescent states to the specified rcu_state |
1772 | * data structure. This involves cleaning up after the prior grace | |
1773 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
1774 | * if one is needed. Note that the caller must hold rnp->lock, which |
1775 | * is released before return. | |
f41d911f | 1776 | */ |
d3f6bad3 | 1777 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 1778 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 1779 | { |
fc2219d4 | 1780 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cabc49c1 PM |
1781 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
1782 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ | |
f41d911f PM |
1783 | } |
1784 | ||
64db4cff | 1785 | /* |
d3f6bad3 PM |
1786 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
1787 | * Allows quiescent states for a group of CPUs to be reported at one go | |
1788 | * to the specified rcu_node structure, though all the CPUs in the group | |
1789 | * must be represented by the same rcu_node structure (which need not be | |
1790 | * a leaf rcu_node structure, though it often will be). That structure's | |
1791 | * lock must be held upon entry, and it is released before return. | |
64db4cff PM |
1792 | */ |
1793 | static void | |
d3f6bad3 PM |
1794 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
1795 | struct rcu_node *rnp, unsigned long flags) | |
64db4cff PM |
1796 | __releases(rnp->lock) |
1797 | { | |
28ecd580 PM |
1798 | struct rcu_node *rnp_c; |
1799 | ||
64db4cff PM |
1800 | /* Walk up the rcu_node hierarchy. */ |
1801 | for (;;) { | |
1802 | if (!(rnp->qsmask & mask)) { | |
1803 | ||
1804 | /* Our bit has already been cleared, so done. */ | |
1304afb2 | 1805 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1806 | return; |
1807 | } | |
1808 | rnp->qsmask &= ~mask; | |
d4c08f2a PM |
1809 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
1810 | mask, rnp->qsmask, rnp->level, | |
1811 | rnp->grplo, rnp->grphi, | |
1812 | !!rnp->gp_tasks); | |
27f4d280 | 1813 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
1814 | |
1815 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 1816 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1817 | return; |
1818 | } | |
1819 | mask = rnp->grpmask; | |
1820 | if (rnp->parent == NULL) { | |
1821 | ||
1822 | /* No more levels. Exit loop holding root lock. */ | |
1823 | ||
1824 | break; | |
1825 | } | |
1304afb2 | 1826 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 1827 | rnp_c = rnp; |
64db4cff | 1828 | rnp = rnp->parent; |
1304afb2 | 1829 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 1830 | smp_mb__after_unlock_lock(); |
28ecd580 | 1831 | WARN_ON_ONCE(rnp_c->qsmask); |
64db4cff PM |
1832 | } |
1833 | ||
1834 | /* | |
1835 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 1836 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 1837 | * to clean up and start the next grace period if one is needed. |
64db4cff | 1838 | */ |
d3f6bad3 | 1839 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
1840 | } |
1841 | ||
1842 | /* | |
d3f6bad3 PM |
1843 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
1844 | * structure. This must be either called from the specified CPU, or | |
1845 | * called when the specified CPU is known to be offline (and when it is | |
1846 | * also known that no other CPU is concurrently trying to help the offline | |
1847 | * CPU). The lastcomp argument is used to make sure we are still in the | |
1848 | * grace period of interest. We don't want to end the current grace period | |
1849 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
1850 | */ |
1851 | static void | |
d7d6a11e | 1852 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1853 | { |
1854 | unsigned long flags; | |
1855 | unsigned long mask; | |
1856 | struct rcu_node *rnp; | |
1857 | ||
1858 | rnp = rdp->mynode; | |
1304afb2 | 1859 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 1860 | smp_mb__after_unlock_lock(); |
d7d6a11e PM |
1861 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum || |
1862 | rnp->completed == rnp->gpnum) { | |
64db4cff PM |
1863 | |
1864 | /* | |
e4cc1f22 PM |
1865 | * The grace period in which this quiescent state was |
1866 | * recorded has ended, so don't report it upwards. | |
1867 | * We will instead need a new quiescent state that lies | |
1868 | * within the current grace period. | |
64db4cff | 1869 | */ |
e4cc1f22 | 1870 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
1304afb2 | 1871 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1872 | return; |
1873 | } | |
1874 | mask = rdp->grpmask; | |
1875 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 1876 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1877 | } else { |
1878 | rdp->qs_pending = 0; | |
1879 | ||
1880 | /* | |
1881 | * This GP can't end until cpu checks in, so all of our | |
1882 | * callbacks can be processed during the next GP. | |
1883 | */ | |
dc35c893 | 1884 | rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 1885 | |
d3f6bad3 | 1886 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
64db4cff PM |
1887 | } |
1888 | } | |
1889 | ||
1890 | /* | |
1891 | * Check to see if there is a new grace period of which this CPU | |
1892 | * is not yet aware, and if so, set up local rcu_data state for it. | |
1893 | * Otherwise, see if this CPU has just passed through its first | |
1894 | * quiescent state for this grace period, and record that fact if so. | |
1895 | */ | |
1896 | static void | |
1897 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
1898 | { | |
05eb552b PM |
1899 | /* Check for grace-period ends and beginnings. */ |
1900 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
1901 | |
1902 | /* | |
1903 | * Does this CPU still need to do its part for current grace period? | |
1904 | * If no, return and let the other CPUs do their part as well. | |
1905 | */ | |
1906 | if (!rdp->qs_pending) | |
1907 | return; | |
1908 | ||
1909 | /* | |
1910 | * Was there a quiescent state since the beginning of the grace | |
1911 | * period? If no, then exit and wait for the next call. | |
1912 | */ | |
e4cc1f22 | 1913 | if (!rdp->passed_quiesce) |
64db4cff PM |
1914 | return; |
1915 | ||
d3f6bad3 PM |
1916 | /* |
1917 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
1918 | * judge of that). | |
1919 | */ | |
d7d6a11e | 1920 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
1921 | } |
1922 | ||
1923 | #ifdef CONFIG_HOTPLUG_CPU | |
1924 | ||
e74f4c45 | 1925 | /* |
b1420f1c PM |
1926 | * Send the specified CPU's RCU callbacks to the orphanage. The |
1927 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 1928 | * ->orphan_lock. |
e74f4c45 | 1929 | */ |
b1420f1c PM |
1930 | static void |
1931 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
1932 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 1933 | { |
3fbfbf7a | 1934 | /* No-CBs CPUs do not have orphanable callbacks. */ |
d1e43fa5 | 1935 | if (rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
1936 | return; |
1937 | ||
b1420f1c PM |
1938 | /* |
1939 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
1940 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
1941 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 1942 | */ |
a50c3af9 | 1943 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1944 | rsp->qlen_lazy += rdp->qlen_lazy; |
1945 | rsp->qlen += rdp->qlen; | |
1946 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 1947 | rdp->qlen_lazy = 0; |
1d1fb395 | 1948 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
1949 | } |
1950 | ||
1951 | /* | |
b1420f1c PM |
1952 | * Next, move those callbacks still needing a grace period to |
1953 | * the orphanage, where some other CPU will pick them up. | |
1954 | * Some of the callbacks might have gone partway through a grace | |
1955 | * period, but that is too bad. They get to start over because we | |
1956 | * cannot assume that grace periods are synchronized across CPUs. | |
1957 | * We don't bother updating the ->nxttail[] array yet, instead | |
1958 | * we just reset the whole thing later on. | |
a50c3af9 | 1959 | */ |
b1420f1c PM |
1960 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
1961 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1962 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
1963 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
1964 | } |
1965 | ||
1966 | /* | |
b1420f1c PM |
1967 | * Then move the ready-to-invoke callbacks to the orphanage, |
1968 | * where some other CPU will pick them up. These will not be | |
1969 | * required to pass though another grace period: They are done. | |
a50c3af9 | 1970 | */ |
e5601400 | 1971 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1972 | *rsp->orphan_donetail = rdp->nxtlist; |
1973 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 1974 | } |
e74f4c45 | 1975 | |
b1420f1c | 1976 | /* Finally, initialize the rcu_data structure's list to empty. */ |
3f5d3ea6 | 1977 | init_callback_list(rdp); |
b1420f1c PM |
1978 | } |
1979 | ||
1980 | /* | |
1981 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 1982 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c | 1983 | */ |
96d3fd0d | 1984 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) |
b1420f1c PM |
1985 | { |
1986 | int i; | |
1987 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); | |
1988 | ||
3fbfbf7a | 1989 | /* No-CBs CPUs are handled specially. */ |
96d3fd0d | 1990 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) |
3fbfbf7a PM |
1991 | return; |
1992 | ||
b1420f1c PM |
1993 | /* Do the accounting first. */ |
1994 | rdp->qlen_lazy += rsp->qlen_lazy; | |
1995 | rdp->qlen += rsp->qlen; | |
1996 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
1997 | if (rsp->qlen_lazy != rsp->qlen) |
1998 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
1999 | rsp->qlen_lazy = 0; |
2000 | rsp->qlen = 0; | |
2001 | ||
2002 | /* | |
2003 | * We do not need a memory barrier here because the only way we | |
2004 | * can get here if there is an rcu_barrier() in flight is if | |
2005 | * we are the task doing the rcu_barrier(). | |
2006 | */ | |
2007 | ||
2008 | /* First adopt the ready-to-invoke callbacks. */ | |
2009 | if (rsp->orphan_donelist != NULL) { | |
2010 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2011 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
2012 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
2013 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2014 | rdp->nxttail[i] = rsp->orphan_donetail; | |
2015 | rsp->orphan_donelist = NULL; | |
2016 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2017 | } | |
2018 | ||
2019 | /* And then adopt the callbacks that still need a grace period. */ | |
2020 | if (rsp->orphan_nxtlist != NULL) { | |
2021 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
2022 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
2023 | rsp->orphan_nxtlist = NULL; | |
2024 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2025 | } | |
2026 | } | |
2027 | ||
2028 | /* | |
2029 | * Trace the fact that this CPU is going offline. | |
2030 | */ | |
2031 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
2032 | { | |
2033 | RCU_TRACE(unsigned long mask); | |
2034 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
2035 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
2036 | ||
2037 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
2038 | trace_rcu_grace_period(rsp->name, |
2039 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 2040 | TPS("cpuofl")); |
64db4cff PM |
2041 | } |
2042 | ||
2043 | /* | |
e5601400 | 2044 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
2045 | * this fact from process context. Do the remainder of the cleanup, |
2046 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
2047 | * adopting them. There can only be one CPU hotplug operation at a time, |
2048 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 2049 | */ |
e5601400 | 2050 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 2051 | { |
2036d94a PM |
2052 | unsigned long flags; |
2053 | unsigned long mask; | |
2054 | int need_report = 0; | |
e5601400 | 2055 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 2056 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2057 | |
2036d94a | 2058 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2059 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 2060 | |
b1420f1c | 2061 | /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */ |
2036d94a PM |
2062 | |
2063 | /* Exclude any attempts to start a new grace period. */ | |
a4fbe35a | 2064 | mutex_lock(&rsp->onoff_mutex); |
7b2e6011 | 2065 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
2036d94a | 2066 | |
b1420f1c PM |
2067 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
2068 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
96d3fd0d | 2069 | rcu_adopt_orphan_cbs(rsp, flags); |
b1420f1c | 2070 | |
2036d94a PM |
2071 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
2072 | mask = rdp->grpmask; /* rnp->grplo is constant. */ | |
2073 | do { | |
2074 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
6303b9c8 | 2075 | smp_mb__after_unlock_lock(); |
2036d94a PM |
2076 | rnp->qsmaskinit &= ~mask; |
2077 | if (rnp->qsmaskinit != 0) { | |
2078 | if (rnp != rdp->mynode) | |
2079 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2080 | break; | |
2081 | } | |
2082 | if (rnp == rdp->mynode) | |
2083 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); | |
2084 | else | |
2085 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2086 | mask = rnp->grpmask; | |
2087 | rnp = rnp->parent; | |
2088 | } while (rnp != NULL); | |
2089 | ||
2090 | /* | |
2091 | * We still hold the leaf rcu_node structure lock here, and | |
2092 | * irqs are still disabled. The reason for this subterfuge is | |
7b2e6011 | 2093 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock |
2036d94a PM |
2094 | * held leads to deadlock. |
2095 | */ | |
7b2e6011 | 2096 | raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */ |
2036d94a PM |
2097 | rnp = rdp->mynode; |
2098 | if (need_report & RCU_OFL_TASKS_NORM_GP) | |
2099 | rcu_report_unblock_qs_rnp(rnp, flags); | |
2100 | else | |
2101 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2102 | if (need_report & RCU_OFL_TASKS_EXP_GP) | |
2103 | rcu_report_exp_rnp(rsp, rnp, true); | |
cf01537e PM |
2104 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
2105 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2106 | cpu, rdp->qlen, rdp->nxtlist); | |
0d8ee37e PM |
2107 | init_callback_list(rdp); |
2108 | /* Disallow further callbacks on this CPU. */ | |
2109 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; | |
a4fbe35a | 2110 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
2111 | } |
2112 | ||
2113 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
2114 | ||
e5601400 | 2115 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
2116 | { |
2117 | } | |
2118 | ||
e5601400 | 2119 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
2120 | { |
2121 | } | |
2122 | ||
2123 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
2124 | ||
2125 | /* | |
2126 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2127 | * period. Thottle as specified by rdp->blimit. | |
2128 | */ | |
37c72e56 | 2129 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2130 | { |
2131 | unsigned long flags; | |
2132 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
2133 | long bl, count, count_lazy; |
2134 | int i; | |
64db4cff | 2135 | |
dc35c893 | 2136 | /* If no callbacks are ready, just return. */ |
29c00b4a | 2137 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 2138 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
2139 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
2140 | need_resched(), is_idle_task(current), | |
2141 | rcu_is_callbacks_kthread()); | |
64db4cff | 2142 | return; |
29c00b4a | 2143 | } |
64db4cff PM |
2144 | |
2145 | /* | |
2146 | * Extract the list of ready callbacks, disabling to prevent | |
2147 | * races with call_rcu() from interrupt handlers. | |
2148 | */ | |
2149 | local_irq_save(flags); | |
8146c4e2 | 2150 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2151 | bl = rdp->blimit; |
486e2593 | 2152 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2153 | list = rdp->nxtlist; |
2154 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2155 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2156 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2157 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2158 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2159 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2160 | local_irq_restore(flags); |
2161 | ||
2162 | /* Invoke callbacks. */ | |
486e2593 | 2163 | count = count_lazy = 0; |
64db4cff PM |
2164 | while (list) { |
2165 | next = list->next; | |
2166 | prefetch(next); | |
551d55a9 | 2167 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2168 | if (__rcu_reclaim(rsp->name, list)) |
2169 | count_lazy++; | |
64db4cff | 2170 | list = next; |
dff1672d PM |
2171 | /* Stop only if limit reached and CPU has something to do. */ |
2172 | if (++count >= bl && | |
2173 | (need_resched() || | |
2174 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2175 | break; |
2176 | } | |
2177 | ||
2178 | local_irq_save(flags); | |
4968c300 PM |
2179 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2180 | is_idle_task(current), | |
2181 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2182 | |
2183 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2184 | if (list != NULL) { |
2185 | *tail = rdp->nxtlist; | |
2186 | rdp->nxtlist = list; | |
b41772ab PM |
2187 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2188 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2189 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2190 | else |
2191 | break; | |
2192 | } | |
b1420f1c PM |
2193 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2194 | rdp->qlen_lazy -= count_lazy; | |
1d1fb395 | 2195 | ACCESS_ONCE(rdp->qlen) -= count; |
b1420f1c | 2196 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2197 | |
2198 | /* Reinstate batch limit if we have worked down the excess. */ | |
2199 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2200 | rdp->blimit = blimit; | |
2201 | ||
37c72e56 PM |
2202 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2203 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2204 | rdp->qlen_last_fqs_check = 0; | |
2205 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2206 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2207 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2208 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2209 | |
64db4cff PM |
2210 | local_irq_restore(flags); |
2211 | ||
e0f23060 | 2212 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2213 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2214 | invoke_rcu_core(); |
64db4cff PM |
2215 | } |
2216 | ||
2217 | /* | |
2218 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2219 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2220 | * Also schedule RCU core processing. |
64db4cff | 2221 | * |
9b2e4f18 | 2222 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2223 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2224 | * false, there is no point in invoking rcu_check_callbacks(). | |
2225 | */ | |
2226 | void rcu_check_callbacks(int cpu, int user) | |
2227 | { | |
f7f7bac9 | 2228 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2229 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2230 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2231 | |
2232 | /* | |
2233 | * Get here if this CPU took its interrupt from user | |
2234 | * mode or from the idle loop, and if this is not a | |
2235 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2236 | * a quiescent state, so note it. |
64db4cff PM |
2237 | * |
2238 | * No memory barrier is required here because both | |
d6714c22 PM |
2239 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2240 | * variables that other CPUs neither access nor modify, | |
2241 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2242 | */ |
2243 | ||
d6714c22 PM |
2244 | rcu_sched_qs(cpu); |
2245 | rcu_bh_qs(cpu); | |
64db4cff PM |
2246 | |
2247 | } else if (!in_softirq()) { | |
2248 | ||
2249 | /* | |
2250 | * Get here if this CPU did not take its interrupt from | |
2251 | * softirq, in other words, if it is not interrupting | |
2252 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2253 | * critical section, so note it. |
64db4cff PM |
2254 | */ |
2255 | ||
d6714c22 | 2256 | rcu_bh_qs(cpu); |
64db4cff | 2257 | } |
f41d911f | 2258 | rcu_preempt_check_callbacks(cpu); |
d21670ac | 2259 | if (rcu_pending(cpu)) |
a46e0899 | 2260 | invoke_rcu_core(); |
f7f7bac9 | 2261 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2262 | } |
2263 | ||
64db4cff PM |
2264 | /* |
2265 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2266 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2267 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2268 | * | |
ee47eb9f | 2269 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2270 | */ |
217af2a2 PM |
2271 | static void force_qs_rnp(struct rcu_state *rsp, |
2272 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2273 | unsigned long *maxj), | |
2274 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2275 | { |
2276 | unsigned long bit; | |
2277 | int cpu; | |
2278 | unsigned long flags; | |
2279 | unsigned long mask; | |
a0b6c9a7 | 2280 | struct rcu_node *rnp; |
64db4cff | 2281 | |
a0b6c9a7 | 2282 | rcu_for_each_leaf_node(rsp, rnp) { |
b4be093f | 2283 | cond_resched(); |
64db4cff | 2284 | mask = 0; |
1304afb2 | 2285 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2286 | smp_mb__after_unlock_lock(); |
ee47eb9f | 2287 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 2288 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 2289 | return; |
64db4cff | 2290 | } |
a0b6c9a7 | 2291 | if (rnp->qsmask == 0) { |
1217ed1b | 2292 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
64db4cff PM |
2293 | continue; |
2294 | } | |
a0b6c9a7 | 2295 | cpu = rnp->grplo; |
64db4cff | 2296 | bit = 1; |
a0b6c9a7 | 2297 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 PM |
2298 | if ((rnp->qsmask & bit) != 0) { |
2299 | if ((rnp->qsmaskinit & bit) != 0) | |
2300 | *isidle = 0; | |
2301 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) | |
2302 | mask |= bit; | |
2303 | } | |
64db4cff | 2304 | } |
45f014c5 | 2305 | if (mask != 0) { |
64db4cff | 2306 | |
d3f6bad3 PM |
2307 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
2308 | rcu_report_qs_rnp(mask, rsp, rnp, flags); | |
64db4cff PM |
2309 | continue; |
2310 | } | |
1304afb2 | 2311 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2312 | } |
27f4d280 | 2313 | rnp = rcu_get_root(rsp); |
1217ed1b PM |
2314 | if (rnp->qsmask == 0) { |
2315 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 2316 | smp_mb__after_unlock_lock(); |
1217ed1b PM |
2317 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ |
2318 | } | |
64db4cff PM |
2319 | } |
2320 | ||
2321 | /* | |
2322 | * Force quiescent states on reluctant CPUs, and also detect which | |
2323 | * CPUs are in dyntick-idle mode. | |
2324 | */ | |
4cdfc175 | 2325 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2326 | { |
2327 | unsigned long flags; | |
394f2769 PM |
2328 | bool ret; |
2329 | struct rcu_node *rnp; | |
2330 | struct rcu_node *rnp_old = NULL; | |
2331 | ||
2332 | /* Funnel through hierarchy to reduce memory contention. */ | |
2333 | rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode; | |
2334 | for (; rnp != NULL; rnp = rnp->parent) { | |
2335 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2336 | !raw_spin_trylock(&rnp->fqslock); | |
2337 | if (rnp_old != NULL) | |
2338 | raw_spin_unlock(&rnp_old->fqslock); | |
2339 | if (ret) { | |
3660c281 | 2340 | ACCESS_ONCE(rsp->n_force_qs_lh)++; |
394f2769 PM |
2341 | return; |
2342 | } | |
2343 | rnp_old = rnp; | |
2344 | } | |
2345 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2346 | |
394f2769 PM |
2347 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2348 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
6303b9c8 | 2349 | smp_mb__after_unlock_lock(); |
394f2769 PM |
2350 | raw_spin_unlock(&rnp_old->fqslock); |
2351 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
3660c281 | 2352 | ACCESS_ONCE(rsp->n_force_qs_lh)++; |
394f2769 | 2353 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2354 | return; /* Someone beat us to it. */ |
46a1e34e | 2355 | } |
4cdfc175 | 2356 | rsp->gp_flags |= RCU_GP_FLAG_FQS; |
394f2769 | 2357 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2358 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ |
64db4cff PM |
2359 | } |
2360 | ||
64db4cff | 2361 | /* |
e0f23060 PM |
2362 | * This does the RCU core processing work for the specified rcu_state |
2363 | * and rcu_data structures. This may be called only from the CPU to | |
2364 | * whom the rdp belongs. | |
64db4cff PM |
2365 | */ |
2366 | static void | |
1bca8cf1 | 2367 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2368 | { |
2369 | unsigned long flags; | |
1bca8cf1 | 2370 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
64db4cff | 2371 | |
2e597558 PM |
2372 | WARN_ON_ONCE(rdp->beenonline == 0); |
2373 | ||
64db4cff PM |
2374 | /* Update RCU state based on any recent quiescent states. */ |
2375 | rcu_check_quiescent_state(rsp, rdp); | |
2376 | ||
2377 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2378 | local_irq_save(flags); |
64db4cff | 2379 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2380 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
b8462084 PM |
2381 | rcu_start_gp(rsp); |
2382 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); | |
dc35c893 PM |
2383 | } else { |
2384 | local_irq_restore(flags); | |
64db4cff PM |
2385 | } |
2386 | ||
2387 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2388 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2389 | invoke_rcu_callbacks(rsp, rdp); |
96d3fd0d PM |
2390 | |
2391 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2392 | do_nocb_deferred_wakeup(rdp); | |
09223371 SL |
2393 | } |
2394 | ||
64db4cff | 2395 | /* |
e0f23060 | 2396 | * Do RCU core processing for the current CPU. |
64db4cff | 2397 | */ |
09223371 | 2398 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2399 | { |
6ce75a23 PM |
2400 | struct rcu_state *rsp; |
2401 | ||
bfa00b4c PM |
2402 | if (cpu_is_offline(smp_processor_id())) |
2403 | return; | |
f7f7bac9 | 2404 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2405 | for_each_rcu_flavor(rsp) |
2406 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2407 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2408 | } |
2409 | ||
a26ac245 | 2410 | /* |
e0f23060 PM |
2411 | * Schedule RCU callback invocation. If the specified type of RCU |
2412 | * does not support RCU priority boosting, just do a direct call, | |
2413 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
2414 | * are running on the current CPU with interrupts disabled, the | |
2415 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
a26ac245 | 2416 | */ |
a46e0899 | 2417 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2418 | { |
b0d30417 PM |
2419 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
2420 | return; | |
a46e0899 PM |
2421 | if (likely(!rsp->boost)) { |
2422 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2423 | return; |
2424 | } | |
a46e0899 | 2425 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2426 | } |
2427 | ||
a46e0899 | 2428 | static void invoke_rcu_core(void) |
09223371 | 2429 | { |
b0f74036 PM |
2430 | if (cpu_online(smp_processor_id())) |
2431 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2432 | } |
2433 | ||
29154c57 PM |
2434 | /* |
2435 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2436 | */ | |
2437 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2438 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2439 | { |
62fde6ed PM |
2440 | /* |
2441 | * If called from an extended quiescent state, invoke the RCU | |
2442 | * core in order to force a re-evaluation of RCU's idleness. | |
2443 | */ | |
5c173eb8 | 2444 | if (!rcu_is_watching() && cpu_online(smp_processor_id())) |
62fde6ed PM |
2445 | invoke_rcu_core(); |
2446 | ||
a16b7a69 | 2447 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2448 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2449 | return; |
64db4cff | 2450 | |
37c72e56 PM |
2451 | /* |
2452 | * Force the grace period if too many callbacks or too long waiting. | |
2453 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2454 | * if some other CPU has recently done so. Also, don't bother | |
2455 | * invoking force_quiescent_state() if the newly enqueued callback | |
2456 | * is the only one waiting for a grace period to complete. | |
2457 | */ | |
2655d57e | 2458 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2459 | |
2460 | /* Are we ignoring a completed grace period? */ | |
470716fc | 2461 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
2462 | |
2463 | /* Start a new grace period if one not already started. */ | |
2464 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
2465 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
2466 | ||
b8462084 | 2467 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 | 2468 | smp_mb__after_unlock_lock(); |
b8462084 PM |
2469 | rcu_start_gp(rsp); |
2470 | raw_spin_unlock(&rnp_root->lock); | |
b52573d2 PM |
2471 | } else { |
2472 | /* Give the grace period a kick. */ | |
2473 | rdp->blimit = LONG_MAX; | |
2474 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2475 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2476 | force_quiescent_state(rsp); |
b52573d2 PM |
2477 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2478 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2479 | } | |
4cdfc175 | 2480 | } |
29154c57 PM |
2481 | } |
2482 | ||
ae150184 PM |
2483 | /* |
2484 | * RCU callback function to leak a callback. | |
2485 | */ | |
2486 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2487 | { | |
2488 | } | |
2489 | ||
3fbfbf7a PM |
2490 | /* |
2491 | * Helper function for call_rcu() and friends. The cpu argument will | |
2492 | * normally be -1, indicating "currently running CPU". It may specify | |
2493 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2494 | * is expected to specify a CPU. | |
2495 | */ | |
64db4cff PM |
2496 | static void |
2497 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 2498 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
2499 | { |
2500 | unsigned long flags; | |
2501 | struct rcu_data *rdp; | |
2502 | ||
0bb7b59d | 2503 | WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */ |
ae150184 PM |
2504 | if (debug_rcu_head_queue(head)) { |
2505 | /* Probable double call_rcu(), so leak the callback. */ | |
2506 | ACCESS_ONCE(head->func) = rcu_leak_callback; | |
2507 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); | |
2508 | return; | |
2509 | } | |
64db4cff PM |
2510 | head->func = func; |
2511 | head->next = NULL; | |
2512 | ||
64db4cff PM |
2513 | /* |
2514 | * Opportunistically note grace-period endings and beginnings. | |
2515 | * Note that we might see a beginning right after we see an | |
2516 | * end, but never vice versa, since this CPU has to pass through | |
2517 | * a quiescent state betweentimes. | |
2518 | */ | |
2519 | local_irq_save(flags); | |
394f99a9 | 2520 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
2521 | |
2522 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
2523 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
2524 | int offline; | |
2525 | ||
2526 | if (cpu != -1) | |
2527 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
96d3fd0d | 2528 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); |
3fbfbf7a | 2529 | WARN_ON_ONCE(offline); |
0d8ee37e | 2530 | /* _call_rcu() is illegal on offline CPU; leak the callback. */ |
0d8ee37e PM |
2531 | local_irq_restore(flags); |
2532 | return; | |
2533 | } | |
29154c57 | 2534 | ACCESS_ONCE(rdp->qlen)++; |
486e2593 PM |
2535 | if (lazy) |
2536 | rdp->qlen_lazy++; | |
c57afe80 PM |
2537 | else |
2538 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2539 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
2540 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
2541 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 2542 | |
d4c08f2a PM |
2543 | if (__is_kfree_rcu_offset((unsigned long)func)) |
2544 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 2545 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2546 | else |
486e2593 | 2547 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2548 | |
29154c57 PM |
2549 | /* Go handle any RCU core processing required. */ |
2550 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
2551 | local_irq_restore(flags); |
2552 | } | |
2553 | ||
2554 | /* | |
d6714c22 | 2555 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 2556 | */ |
d6714c22 | 2557 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 2558 | { |
3fbfbf7a | 2559 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 2560 | } |
d6714c22 | 2561 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
2562 | |
2563 | /* | |
486e2593 | 2564 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
2565 | */ |
2566 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
2567 | { | |
3fbfbf7a | 2568 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
2569 | } |
2570 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
2571 | ||
6d813391 PM |
2572 | /* |
2573 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
2574 | * any blocking grace-period wait automatically implies a grace period | |
2575 | * if there is only one CPU online at any point time during execution | |
2576 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
2577 | * occasionally incorrectly indicate that there are multiple CPUs online | |
2578 | * when there was in fact only one the whole time, as this just adds | |
2579 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
2580 | */ |
2581 | static inline int rcu_blocking_is_gp(void) | |
2582 | { | |
95f0c1de PM |
2583 | int ret; |
2584 | ||
6d813391 | 2585 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
2586 | preempt_disable(); |
2587 | ret = num_online_cpus() <= 1; | |
2588 | preempt_enable(); | |
2589 | return ret; | |
6d813391 PM |
2590 | } |
2591 | ||
6ebb237b PM |
2592 | /** |
2593 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
2594 | * | |
2595 | * Control will return to the caller some time after a full rcu-sched | |
2596 | * grace period has elapsed, in other words after all currently executing | |
2597 | * rcu-sched read-side critical sections have completed. These read-side | |
2598 | * critical sections are delimited by rcu_read_lock_sched() and | |
2599 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
2600 | * local_irq_disable(), and so on may be used in place of | |
2601 | * rcu_read_lock_sched(). | |
2602 | * | |
2603 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
2604 | * non-threaded hardware-interrupt handlers, in progress on entry will |
2605 | * have completed before this primitive returns. However, this does not | |
2606 | * guarantee that softirq handlers will have completed, since in some | |
2607 | * kernels, these handlers can run in process context, and can block. | |
2608 | * | |
2609 | * Note that this guarantee implies further memory-ordering guarantees. | |
2610 | * On systems with more than one CPU, when synchronize_sched() returns, | |
2611 | * each CPU is guaranteed to have executed a full memory barrier since the | |
2612 | * end of its last RCU-sched read-side critical section whose beginning | |
2613 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
2614 | * an RCU read-side critical section that extends beyond the return from | |
2615 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
2616 | * after the beginning of synchronize_sched() and before the beginning of | |
2617 | * that RCU read-side critical section. Note that these guarantees include | |
2618 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
2619 | * that are executing in the kernel. | |
2620 | * | |
2621 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
2622 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
2623 | * to have executed a full memory barrier during the execution of | |
2624 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
2625 | * again only if the system has more than one CPU). | |
6ebb237b PM |
2626 | * |
2627 | * This primitive provides the guarantees made by the (now removed) | |
2628 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
2629 | * guarantees that rcu_read_lock() sections will have completed. | |
2630 | * In "classic RCU", these two guarantees happen to be one and | |
2631 | * the same, but can differ in realtime RCU implementations. | |
2632 | */ | |
2633 | void synchronize_sched(void) | |
2634 | { | |
fe15d706 PM |
2635 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2636 | !lock_is_held(&rcu_lock_map) && | |
2637 | !lock_is_held(&rcu_sched_lock_map), | |
2638 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
2639 | if (rcu_blocking_is_gp()) |
2640 | return; | |
3705b88d AM |
2641 | if (rcu_expedited) |
2642 | synchronize_sched_expedited(); | |
2643 | else | |
2644 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
2645 | } |
2646 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
2647 | ||
2648 | /** | |
2649 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
2650 | * | |
2651 | * Control will return to the caller some time after a full rcu_bh grace | |
2652 | * period has elapsed, in other words after all currently executing rcu_bh | |
2653 | * read-side critical sections have completed. RCU read-side critical | |
2654 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
2655 | * and may be nested. | |
f0a0e6f2 PM |
2656 | * |
2657 | * See the description of synchronize_sched() for more detailed information | |
2658 | * on memory ordering guarantees. | |
6ebb237b PM |
2659 | */ |
2660 | void synchronize_rcu_bh(void) | |
2661 | { | |
fe15d706 PM |
2662 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2663 | !lock_is_held(&rcu_lock_map) && | |
2664 | !lock_is_held(&rcu_sched_lock_map), | |
2665 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
2666 | if (rcu_blocking_is_gp()) |
2667 | return; | |
3705b88d AM |
2668 | if (rcu_expedited) |
2669 | synchronize_rcu_bh_expedited(); | |
2670 | else | |
2671 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
2672 | } |
2673 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
2674 | ||
765a3f4f PM |
2675 | /** |
2676 | * get_state_synchronize_rcu - Snapshot current RCU state | |
2677 | * | |
2678 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
2679 | * to determine whether or not a full grace period has elapsed in the | |
2680 | * meantime. | |
2681 | */ | |
2682 | unsigned long get_state_synchronize_rcu(void) | |
2683 | { | |
2684 | /* | |
2685 | * Any prior manipulation of RCU-protected data must happen | |
2686 | * before the load from ->gpnum. | |
2687 | */ | |
2688 | smp_mb(); /* ^^^ */ | |
2689 | ||
2690 | /* | |
2691 | * Make sure this load happens before the purportedly | |
2692 | * time-consuming work between get_state_synchronize_rcu() | |
2693 | * and cond_synchronize_rcu(). | |
2694 | */ | |
2695 | return smp_load_acquire(&rcu_state->gpnum); | |
2696 | } | |
2697 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
2698 | ||
2699 | /** | |
2700 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
2701 | * | |
2702 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
2703 | * | |
2704 | * If a full RCU grace period has elapsed since the earlier call to | |
2705 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
2706 | * synchronize_rcu() to wait for a full grace period. | |
2707 | * | |
2708 | * Yes, this function does not take counter wrap into account. But | |
2709 | * counter wrap is harmless. If the counter wraps, we have waited for | |
2710 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
2711 | * so waiting for one additional grace period should be just fine. | |
2712 | */ | |
2713 | void cond_synchronize_rcu(unsigned long oldstate) | |
2714 | { | |
2715 | unsigned long newstate; | |
2716 | ||
2717 | /* | |
2718 | * Ensure that this load happens before any RCU-destructive | |
2719 | * actions the caller might carry out after we return. | |
2720 | */ | |
2721 | newstate = smp_load_acquire(&rcu_state->completed); | |
2722 | if (ULONG_CMP_GE(oldstate, newstate)) | |
2723 | synchronize_rcu(); | |
2724 | } | |
2725 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
2726 | ||
3d3b7db0 PM |
2727 | static int synchronize_sched_expedited_cpu_stop(void *data) |
2728 | { | |
2729 | /* | |
2730 | * There must be a full memory barrier on each affected CPU | |
2731 | * between the time that try_stop_cpus() is called and the | |
2732 | * time that it returns. | |
2733 | * | |
2734 | * In the current initial implementation of cpu_stop, the | |
2735 | * above condition is already met when the control reaches | |
2736 | * this point and the following smp_mb() is not strictly | |
2737 | * necessary. Do smp_mb() anyway for documentation and | |
2738 | * robustness against future implementation changes. | |
2739 | */ | |
2740 | smp_mb(); /* See above comment block. */ | |
2741 | return 0; | |
2742 | } | |
2743 | ||
236fefaf PM |
2744 | /** |
2745 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
2746 | * | |
2747 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
2748 | * approach to force the grace period to end quickly. This consumes | |
2749 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
2750 | * so is thus not recommended for any sort of common-case code. In fact, | |
2751 | * if you are using synchronize_sched_expedited() in a loop, please | |
2752 | * restructure your code to batch your updates, and then use a single | |
2753 | * synchronize_sched() instead. | |
3d3b7db0 | 2754 | * |
236fefaf PM |
2755 | * Note that it is illegal to call this function while holding any lock |
2756 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal | |
2757 | * to call this function from a CPU-hotplug notifier. Failing to observe | |
2758 | * these restriction will result in deadlock. | |
3d3b7db0 PM |
2759 | * |
2760 | * This implementation can be thought of as an application of ticket | |
2761 | * locking to RCU, with sync_sched_expedited_started and | |
2762 | * sync_sched_expedited_done taking on the roles of the halves | |
2763 | * of the ticket-lock word. Each task atomically increments | |
2764 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
2765 | * then attempts to stop all the CPUs. If this succeeds, then each | |
2766 | * CPU will have executed a context switch, resulting in an RCU-sched | |
2767 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
2768 | * update sync_sched_expedited_done to match our snapshot -- but | |
2769 | * only if someone else has not already advanced past our snapshot. | |
2770 | * | |
2771 | * On the other hand, if try_stop_cpus() fails, we check the value | |
2772 | * of sync_sched_expedited_done. If it has advanced past our | |
2773 | * initial snapshot, then someone else must have forced a grace period | |
2774 | * some time after we took our snapshot. In this case, our work is | |
2775 | * done for us, and we can simply return. Otherwise, we try again, | |
2776 | * but keep our initial snapshot for purposes of checking for someone | |
2777 | * doing our work for us. | |
2778 | * | |
2779 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
2780 | */ | |
2781 | void synchronize_sched_expedited(void) | |
2782 | { | |
1924bcb0 PM |
2783 | long firstsnap, s, snap; |
2784 | int trycount = 0; | |
40694d66 | 2785 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 2786 | |
1924bcb0 PM |
2787 | /* |
2788 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
2789 | * By allowing sync_sched_expedited_started to advance no more than | |
2790 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
2791 | * that more than 3.5 billion CPUs would be required to force a | |
2792 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
2793 | * course be required on a 64-bit system. | |
2794 | */ | |
40694d66 PM |
2795 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
2796 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 PM |
2797 | ULONG_MAX / 8)) { |
2798 | synchronize_sched(); | |
a30489c5 | 2799 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
2800 | return; |
2801 | } | |
3d3b7db0 | 2802 | |
1924bcb0 PM |
2803 | /* |
2804 | * Take a ticket. Note that atomic_inc_return() implies a | |
2805 | * full memory barrier. | |
2806 | */ | |
40694d66 | 2807 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 2808 | firstsnap = snap; |
3d3b7db0 | 2809 | get_online_cpus(); |
1cc85961 | 2810 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 PM |
2811 | |
2812 | /* | |
2813 | * Each pass through the following loop attempts to force a | |
2814 | * context switch on each CPU. | |
2815 | */ | |
2816 | while (try_stop_cpus(cpu_online_mask, | |
2817 | synchronize_sched_expedited_cpu_stop, | |
2818 | NULL) == -EAGAIN) { | |
2819 | put_online_cpus(); | |
a30489c5 | 2820 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 2821 | |
1924bcb0 | 2822 | /* Check to see if someone else did our work for us. */ |
40694d66 | 2823 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2824 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2825 | /* ensure test happens before caller kfree */ |
2826 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2827 | atomic_long_inc(&rsp->expedited_workdone1); | |
1924bcb0 PM |
2828 | return; |
2829 | } | |
3d3b7db0 PM |
2830 | |
2831 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 2832 | if (trycount++ < 10) { |
3d3b7db0 | 2833 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 2834 | } else { |
3705b88d | 2835 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 2836 | atomic_long_inc(&rsp->expedited_normal); |
3d3b7db0 PM |
2837 | return; |
2838 | } | |
2839 | ||
1924bcb0 | 2840 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 2841 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2842 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2843 | /* ensure test happens before caller kfree */ |
2844 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2845 | atomic_long_inc(&rsp->expedited_workdone2); | |
3d3b7db0 PM |
2846 | return; |
2847 | } | |
2848 | ||
2849 | /* | |
2850 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
2851 | * callers to piggyback on our grace period. We retry |
2852 | * after they started, so our grace period works for them, | |
2853 | * and they started after our first try, so their grace | |
2854 | * period works for us. | |
3d3b7db0 PM |
2855 | */ |
2856 | get_online_cpus(); | |
40694d66 | 2857 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
2858 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
2859 | } | |
a30489c5 | 2860 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 PM |
2861 | |
2862 | /* | |
2863 | * Everyone up to our most recent fetch is covered by our grace | |
2864 | * period. Update the counter, but only if our work is still | |
2865 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 2866 | * than we did already did their update. |
3d3b7db0 PM |
2867 | */ |
2868 | do { | |
a30489c5 | 2869 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 2870 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2871 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 PM |
2872 | /* ensure test happens before caller kfree */ |
2873 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2874 | atomic_long_inc(&rsp->expedited_done_lost); | |
3d3b7db0 PM |
2875 | break; |
2876 | } | |
40694d66 | 2877 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 2878 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
2879 | |
2880 | put_online_cpus(); | |
2881 | } | |
2882 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
2883 | ||
64db4cff PM |
2884 | /* |
2885 | * Check to see if there is any immediate RCU-related work to be done | |
2886 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
2887 | * The checks are in order of increasing expense: checks that can be | |
2888 | * carried out against CPU-local state are performed first. However, | |
2889 | * we must check for CPU stalls first, else we might not get a chance. | |
2890 | */ | |
2891 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
2892 | { | |
2f51f988 PM |
2893 | struct rcu_node *rnp = rdp->mynode; |
2894 | ||
64db4cff PM |
2895 | rdp->n_rcu_pending++; |
2896 | ||
2897 | /* Check for CPU stalls, if enabled. */ | |
2898 | check_cpu_stall(rsp, rdp); | |
2899 | ||
a096932f PM |
2900 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ |
2901 | if (rcu_nohz_full_cpu(rsp)) | |
2902 | return 0; | |
2903 | ||
64db4cff | 2904 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
5c51dd73 PM |
2905 | if (rcu_scheduler_fully_active && |
2906 | rdp->qs_pending && !rdp->passed_quiesce) { | |
d21670ac | 2907 | rdp->n_rp_qs_pending++; |
e4cc1f22 | 2908 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
d21670ac | 2909 | rdp->n_rp_report_qs++; |
64db4cff | 2910 | return 1; |
7ba5c840 | 2911 | } |
64db4cff PM |
2912 | |
2913 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
2914 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
2915 | rdp->n_rp_cb_ready++; | |
64db4cff | 2916 | return 1; |
7ba5c840 | 2917 | } |
64db4cff PM |
2918 | |
2919 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
2920 | if (cpu_needs_another_gp(rsp, rdp)) { |
2921 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 2922 | return 1; |
7ba5c840 | 2923 | } |
64db4cff PM |
2924 | |
2925 | /* Has another RCU grace period completed? */ | |
2f51f988 | 2926 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 2927 | rdp->n_rp_gp_completed++; |
64db4cff | 2928 | return 1; |
7ba5c840 | 2929 | } |
64db4cff PM |
2930 | |
2931 | /* Has a new RCU grace period started? */ | |
2f51f988 | 2932 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
7ba5c840 | 2933 | rdp->n_rp_gp_started++; |
64db4cff | 2934 | return 1; |
7ba5c840 | 2935 | } |
64db4cff | 2936 | |
96d3fd0d PM |
2937 | /* Does this CPU need a deferred NOCB wakeup? */ |
2938 | if (rcu_nocb_need_deferred_wakeup(rdp)) { | |
2939 | rdp->n_rp_nocb_defer_wakeup++; | |
2940 | return 1; | |
2941 | } | |
2942 | ||
64db4cff | 2943 | /* nothing to do */ |
7ba5c840 | 2944 | rdp->n_rp_need_nothing++; |
64db4cff PM |
2945 | return 0; |
2946 | } | |
2947 | ||
2948 | /* | |
2949 | * Check to see if there is any immediate RCU-related work to be done | |
2950 | * by the current CPU, returning 1 if so. This function is part of the | |
2951 | * RCU implementation; it is -not- an exported member of the RCU API. | |
2952 | */ | |
a157229c | 2953 | static int rcu_pending(int cpu) |
64db4cff | 2954 | { |
6ce75a23 PM |
2955 | struct rcu_state *rsp; |
2956 | ||
2957 | for_each_rcu_flavor(rsp) | |
2958 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu))) | |
2959 | return 1; | |
2960 | return 0; | |
64db4cff PM |
2961 | } |
2962 | ||
2963 | /* | |
c0f4dfd4 PM |
2964 | * Return true if the specified CPU has any callback. If all_lazy is |
2965 | * non-NULL, store an indication of whether all callbacks are lazy. | |
2966 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 2967 | */ |
ffa83fb5 | 2968 | static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy) |
64db4cff | 2969 | { |
c0f4dfd4 PM |
2970 | bool al = true; |
2971 | bool hc = false; | |
2972 | struct rcu_data *rdp; | |
6ce75a23 PM |
2973 | struct rcu_state *rsp; |
2974 | ||
c0f4dfd4 PM |
2975 | for_each_rcu_flavor(rsp) { |
2976 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
69c8d28c PM |
2977 | if (!rdp->nxtlist) |
2978 | continue; | |
2979 | hc = true; | |
2980 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
c0f4dfd4 | 2981 | al = false; |
69c8d28c PM |
2982 | break; |
2983 | } | |
c0f4dfd4 PM |
2984 | } |
2985 | if (all_lazy) | |
2986 | *all_lazy = al; | |
2987 | return hc; | |
64db4cff PM |
2988 | } |
2989 | ||
a83eff0a PM |
2990 | /* |
2991 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
2992 | * the compiler is expected to optimize this away. | |
2993 | */ | |
e66c33d5 | 2994 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
2995 | int cpu, unsigned long done) |
2996 | { | |
2997 | trace_rcu_barrier(rsp->name, s, cpu, | |
2998 | atomic_read(&rsp->barrier_cpu_count), done); | |
2999 | } | |
3000 | ||
b1420f1c PM |
3001 | /* |
3002 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
3003 | * up the task executing _rcu_barrier(). | |
3004 | */ | |
24ebbca8 | 3005 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3006 | { |
24ebbca8 PM |
3007 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
3008 | struct rcu_state *rsp = rdp->rsp; | |
3009 | ||
a83eff0a PM |
3010 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
3011 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 3012 | complete(&rsp->barrier_completion); |
a83eff0a PM |
3013 | } else { |
3014 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
3015 | } | |
d0ec774c PM |
3016 | } |
3017 | ||
3018 | /* | |
3019 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3020 | */ | |
3021 | static void rcu_barrier_func(void *type) | |
3022 | { | |
037b64ed | 3023 | struct rcu_state *rsp = type; |
06668efa | 3024 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
d0ec774c | 3025 | |
a83eff0a | 3026 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 3027 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 3028 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
3029 | } |
3030 | ||
d0ec774c PM |
3031 | /* |
3032 | * Orchestrate the specified type of RCU barrier, waiting for all | |
3033 | * RCU callbacks of the specified type to complete. | |
3034 | */ | |
037b64ed | 3035 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 3036 | { |
b1420f1c | 3037 | int cpu; |
b1420f1c | 3038 | struct rcu_data *rdp; |
cf3a9c48 PM |
3039 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
3040 | unsigned long snap_done; | |
b1420f1c | 3041 | |
a83eff0a | 3042 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 3043 | |
e74f4c45 | 3044 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 3045 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 3046 | |
cf3a9c48 PM |
3047 | /* |
3048 | * Ensure that all prior references, including to ->n_barrier_done, | |
3049 | * are ordered before the _rcu_barrier() machinery. | |
3050 | */ | |
3051 | smp_mb(); /* See above block comment. */ | |
3052 | ||
3053 | /* | |
3054 | * Recheck ->n_barrier_done to see if others did our work for us. | |
3055 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
3056 | * transition. The "if" expression below therefore rounds the old | |
3057 | * value up to the next even number and adds two before comparing. | |
3058 | */ | |
458fb381 | 3059 | snap_done = rsp->n_barrier_done; |
a83eff0a | 3060 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
458fb381 PM |
3061 | |
3062 | /* | |
3063 | * If the value in snap is odd, we needed to wait for the current | |
3064 | * rcu_barrier() to complete, then wait for the next one, in other | |
3065 | * words, we need the value of snap_done to be three larger than | |
3066 | * the value of snap. On the other hand, if the value in snap is | |
3067 | * even, we only had to wait for the next rcu_barrier() to complete, | |
3068 | * in other words, we need the value of snap_done to be only two | |
3069 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes | |
3070 | * this for us (thank you, Linus!). | |
3071 | */ | |
3072 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) { | |
a83eff0a | 3073 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
3074 | smp_mb(); /* caller's subsequent code after above check. */ |
3075 | mutex_unlock(&rsp->barrier_mutex); | |
3076 | return; | |
3077 | } | |
3078 | ||
3079 | /* | |
3080 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
3081 | * ACCESS_ONCE() to prevent the compiler from speculating | |
3082 | * the increment to precede the early-exit check. | |
3083 | */ | |
3084 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
3085 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); | |
a83eff0a | 3086 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 3087 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 3088 | |
d0ec774c | 3089 | /* |
b1420f1c PM |
3090 | * Initialize the count to one rather than to zero in order to |
3091 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
3092 | * (or preemption of this task). Exclude CPU-hotplug operations |
3093 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 3094 | */ |
7db74df8 | 3095 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 3096 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 3097 | get_online_cpus(); |
b1420f1c PM |
3098 | |
3099 | /* | |
1331e7a1 PM |
3100 | * Force each CPU with callbacks to register a new callback. |
3101 | * When that callback is invoked, we will know that all of the | |
3102 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 3103 | */ |
3fbfbf7a | 3104 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 3105 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 3106 | continue; |
b1420f1c | 3107 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 3108 | if (rcu_is_nocb_cpu(cpu)) { |
3fbfbf7a PM |
3109 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, |
3110 | rsp->n_barrier_done); | |
3111 | atomic_inc(&rsp->barrier_cpu_count); | |
3112 | __call_rcu(&rdp->barrier_head, rcu_barrier_callback, | |
3113 | rsp, cpu, 0); | |
3114 | } else if (ACCESS_ONCE(rdp->qlen)) { | |
a83eff0a PM |
3115 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
3116 | rsp->n_barrier_done); | |
037b64ed | 3117 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 3118 | } else { |
a83eff0a PM |
3119 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
3120 | rsp->n_barrier_done); | |
b1420f1c PM |
3121 | } |
3122 | } | |
1331e7a1 | 3123 | put_online_cpus(); |
b1420f1c PM |
3124 | |
3125 | /* | |
3126 | * Now that we have an rcu_barrier_callback() callback on each | |
3127 | * CPU, and thus each counted, remove the initial count. | |
3128 | */ | |
24ebbca8 | 3129 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 3130 | complete(&rsp->barrier_completion); |
b1420f1c | 3131 | |
cf3a9c48 PM |
3132 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
3133 | smp_mb(); /* Keep increment after above mechanism. */ | |
3134 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
3135 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); | |
a83eff0a | 3136 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
3137 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
3138 | ||
b1420f1c | 3139 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 3140 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
3141 | |
3142 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 3143 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 3144 | } |
d0ec774c PM |
3145 | |
3146 | /** | |
3147 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
3148 | */ | |
3149 | void rcu_barrier_bh(void) | |
3150 | { | |
037b64ed | 3151 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
3152 | } |
3153 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
3154 | ||
3155 | /** | |
3156 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
3157 | */ | |
3158 | void rcu_barrier_sched(void) | |
3159 | { | |
037b64ed | 3160 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
3161 | } |
3162 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
3163 | ||
64db4cff | 3164 | /* |
27569620 | 3165 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 3166 | */ |
27569620 PM |
3167 | static void __init |
3168 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
3169 | { |
3170 | unsigned long flags; | |
394f99a9 | 3171 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
3172 | struct rcu_node *rnp = rcu_get_root(rsp); |
3173 | ||
3174 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 3175 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 3176 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
3f5d3ea6 | 3177 | init_callback_list(rdp); |
486e2593 | 3178 | rdp->qlen_lazy = 0; |
1d1fb395 | 3179 | ACCESS_ONCE(rdp->qlen) = 0; |
27569620 | 3180 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 3181 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 3182 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 3183 | rdp->cpu = cpu; |
d4c08f2a | 3184 | rdp->rsp = rsp; |
3fbfbf7a | 3185 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 3186 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
3187 | } |
3188 | ||
3189 | /* | |
3190 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
3191 | * offline event can be happening at a given time. Note also that we | |
3192 | * can accept some slop in the rsp->completed access due to the fact | |
3193 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 3194 | */ |
49fb4c62 | 3195 | static void |
6cc68793 | 3196 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible) |
64db4cff PM |
3197 | { |
3198 | unsigned long flags; | |
64db4cff | 3199 | unsigned long mask; |
394f99a9 | 3200 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
3201 | struct rcu_node *rnp = rcu_get_root(rsp); |
3202 | ||
a4fbe35a PM |
3203 | /* Exclude new grace periods. */ |
3204 | mutex_lock(&rsp->onoff_mutex); | |
3205 | ||
64db4cff | 3206 | /* Set up local state, ensuring consistent view of global state. */ |
1304afb2 | 3207 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 3208 | rdp->beenonline = 1; /* We have now been online. */ |
6cc68793 | 3209 | rdp->preemptible = preemptible; |
37c72e56 PM |
3210 | rdp->qlen_last_fqs_check = 0; |
3211 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 3212 | rdp->blimit = blimit; |
0d8ee37e | 3213 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ |
29e37d81 | 3214 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 3215 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
3216 | atomic_set(&rdp->dynticks->dynticks, |
3217 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 3218 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 3219 | |
64db4cff PM |
3220 | /* Add CPU to rcu_node bitmasks. */ |
3221 | rnp = rdp->mynode; | |
3222 | mask = rdp->grpmask; | |
3223 | do { | |
3224 | /* Exclude any attempts to start a new GP on small systems. */ | |
1304afb2 | 3225 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3226 | rnp->qsmaskinit |= mask; |
3227 | mask = rnp->grpmask; | |
d09b62df | 3228 | if (rnp == rdp->mynode) { |
06ae115a PM |
3229 | /* |
3230 | * If there is a grace period in progress, we will | |
3231 | * set up to wait for it next time we run the | |
3232 | * RCU core code. | |
3233 | */ | |
3234 | rdp->gpnum = rnp->completed; | |
d09b62df | 3235 | rdp->completed = rnp->completed; |
06ae115a PM |
3236 | rdp->passed_quiesce = 0; |
3237 | rdp->qs_pending = 0; | |
f7f7bac9 | 3238 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); |
d09b62df | 3239 | } |
1304afb2 | 3240 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3241 | rnp = rnp->parent; |
3242 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); | |
a4fbe35a | 3243 | local_irq_restore(flags); |
64db4cff | 3244 | |
a4fbe35a | 3245 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
3246 | } |
3247 | ||
49fb4c62 | 3248 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 3249 | { |
6ce75a23 PM |
3250 | struct rcu_state *rsp; |
3251 | ||
3252 | for_each_rcu_flavor(rsp) | |
3253 | rcu_init_percpu_data(cpu, rsp, | |
3254 | strcmp(rsp->name, "rcu_preempt") == 0); | |
64db4cff PM |
3255 | } |
3256 | ||
3257 | /* | |
f41d911f | 3258 | * Handle CPU online/offline notification events. |
64db4cff | 3259 | */ |
49fb4c62 | 3260 | static int rcu_cpu_notify(struct notifier_block *self, |
9f680ab4 | 3261 | unsigned long action, void *hcpu) |
64db4cff PM |
3262 | { |
3263 | long cpu = (long)hcpu; | |
27f4d280 | 3264 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); |
a26ac245 | 3265 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 3266 | struct rcu_state *rsp; |
64db4cff | 3267 | |
f7f7bac9 | 3268 | trace_rcu_utilization(TPS("Start CPU hotplug")); |
64db4cff PM |
3269 | switch (action) { |
3270 | case CPU_UP_PREPARE: | |
3271 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
3272 | rcu_prepare_cpu(cpu); |
3273 | rcu_prepare_kthreads(cpu); | |
a26ac245 PM |
3274 | break; |
3275 | case CPU_ONLINE: | |
0f962a5e | 3276 | case CPU_DOWN_FAILED: |
5d01bbd1 | 3277 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
3278 | break; |
3279 | case CPU_DOWN_PREPARE: | |
34ed6246 | 3280 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 3281 | break; |
d0ec774c PM |
3282 | case CPU_DYING: |
3283 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
3284 | for_each_rcu_flavor(rsp) |
3285 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 3286 | break; |
64db4cff PM |
3287 | case CPU_DEAD: |
3288 | case CPU_DEAD_FROZEN: | |
3289 | case CPU_UP_CANCELED: | |
3290 | case CPU_UP_CANCELED_FROZEN: | |
6ce75a23 PM |
3291 | for_each_rcu_flavor(rsp) |
3292 | rcu_cleanup_dead_cpu(cpu, rsp); | |
64db4cff PM |
3293 | break; |
3294 | default: | |
3295 | break; | |
3296 | } | |
f7f7bac9 | 3297 | trace_rcu_utilization(TPS("End CPU hotplug")); |
34ed6246 | 3298 | return NOTIFY_OK; |
64db4cff PM |
3299 | } |
3300 | ||
d1d74d14 BP |
3301 | static int rcu_pm_notify(struct notifier_block *self, |
3302 | unsigned long action, void *hcpu) | |
3303 | { | |
3304 | switch (action) { | |
3305 | case PM_HIBERNATION_PREPARE: | |
3306 | case PM_SUSPEND_PREPARE: | |
3307 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
3308 | rcu_expedited = 1; | |
3309 | break; | |
3310 | case PM_POST_HIBERNATION: | |
3311 | case PM_POST_SUSPEND: | |
3312 | rcu_expedited = 0; | |
3313 | break; | |
3314 | default: | |
3315 | break; | |
3316 | } | |
3317 | return NOTIFY_OK; | |
3318 | } | |
3319 | ||
b3dbec76 PM |
3320 | /* |
3321 | * Spawn the kthread that handles this RCU flavor's grace periods. | |
3322 | */ | |
3323 | static int __init rcu_spawn_gp_kthread(void) | |
3324 | { | |
3325 | unsigned long flags; | |
3326 | struct rcu_node *rnp; | |
3327 | struct rcu_state *rsp; | |
3328 | struct task_struct *t; | |
3329 | ||
3330 | for_each_rcu_flavor(rsp) { | |
f170168b | 3331 | t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
3332 | BUG_ON(IS_ERR(t)); |
3333 | rnp = rcu_get_root(rsp); | |
3334 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3335 | rsp->gp_kthread = t; | |
3336 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3fbfbf7a | 3337 | rcu_spawn_nocb_kthreads(rsp); |
b3dbec76 PM |
3338 | } |
3339 | return 0; | |
3340 | } | |
3341 | early_initcall(rcu_spawn_gp_kthread); | |
3342 | ||
bbad9379 PM |
3343 | /* |
3344 | * This function is invoked towards the end of the scheduler's initialization | |
3345 | * process. Before this is called, the idle task might contain | |
3346 | * RCU read-side critical sections (during which time, this idle | |
3347 | * task is booting the system). After this function is called, the | |
3348 | * idle tasks are prohibited from containing RCU read-side critical | |
3349 | * sections. This function also enables RCU lockdep checking. | |
3350 | */ | |
3351 | void rcu_scheduler_starting(void) | |
3352 | { | |
3353 | WARN_ON(num_online_cpus() != 1); | |
3354 | WARN_ON(nr_context_switches() > 0); | |
3355 | rcu_scheduler_active = 1; | |
3356 | } | |
3357 | ||
64db4cff PM |
3358 | /* |
3359 | * Compute the per-level fanout, either using the exact fanout specified | |
3360 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
3361 | */ | |
3362 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
3363 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3364 | { | |
3365 | int i; | |
3366 | ||
04f34650 PM |
3367 | rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; |
3368 | for (i = rcu_num_lvls - 2; i >= 0; i--) | |
64db4cff PM |
3369 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
3370 | } | |
3371 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3372 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3373 | { | |
3374 | int ccur; | |
3375 | int cprv; | |
3376 | int i; | |
3377 | ||
4dbd6bb3 | 3378 | cprv = nr_cpu_ids; |
f885b7f2 | 3379 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3380 | ccur = rsp->levelcnt[i]; |
3381 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
3382 | cprv = ccur; | |
3383 | } | |
3384 | } | |
3385 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3386 | ||
3387 | /* | |
3388 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3389 | */ | |
394f99a9 LJ |
3390 | static void __init rcu_init_one(struct rcu_state *rsp, |
3391 | struct rcu_data __percpu *rda) | |
64db4cff | 3392 | { |
394f2769 PM |
3393 | static char *buf[] = { "rcu_node_0", |
3394 | "rcu_node_1", | |
3395 | "rcu_node_2", | |
3396 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ | |
3397 | static char *fqs[] = { "rcu_node_fqs_0", | |
3398 | "rcu_node_fqs_1", | |
3399 | "rcu_node_fqs_2", | |
3400 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ | |
64db4cff PM |
3401 | int cpustride = 1; |
3402 | int i; | |
3403 | int j; | |
3404 | struct rcu_node *rnp; | |
3405 | ||
b6407e86 PM |
3406 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
3407 | ||
4930521a PM |
3408 | /* Silence gcc 4.8 warning about array index out of range. */ |
3409 | if (rcu_num_lvls > RCU_NUM_LVLS) | |
3410 | panic("rcu_init_one: rcu_num_lvls overflow"); | |
3411 | ||
64db4cff PM |
3412 | /* Initialize the level-tracking arrays. */ |
3413 | ||
f885b7f2 PM |
3414 | for (i = 0; i < rcu_num_lvls; i++) |
3415 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
3416 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
3417 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
3418 | rcu_init_levelspread(rsp); | |
3419 | ||
3420 | /* Initialize the elements themselves, starting from the leaves. */ | |
3421 | ||
f885b7f2 | 3422 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3423 | cpustride *= rsp->levelspread[i]; |
3424 | rnp = rsp->level[i]; | |
3425 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 3426 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
3427 | lockdep_set_class_and_name(&rnp->lock, |
3428 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
3429 | raw_spin_lock_init(&rnp->fqslock); |
3430 | lockdep_set_class_and_name(&rnp->fqslock, | |
3431 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
3432 | rnp->gpnum = rsp->gpnum; |
3433 | rnp->completed = rsp->completed; | |
64db4cff PM |
3434 | rnp->qsmask = 0; |
3435 | rnp->qsmaskinit = 0; | |
3436 | rnp->grplo = j * cpustride; | |
3437 | rnp->grphi = (j + 1) * cpustride - 1; | |
3438 | if (rnp->grphi >= NR_CPUS) | |
3439 | rnp->grphi = NR_CPUS - 1; | |
3440 | if (i == 0) { | |
3441 | rnp->grpnum = 0; | |
3442 | rnp->grpmask = 0; | |
3443 | rnp->parent = NULL; | |
3444 | } else { | |
3445 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
3446 | rnp->grpmask = 1UL << rnp->grpnum; | |
3447 | rnp->parent = rsp->level[i - 1] + | |
3448 | j / rsp->levelspread[i - 1]; | |
3449 | } | |
3450 | rnp->level = i; | |
12f5f524 | 3451 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 3452 | rcu_init_one_nocb(rnp); |
64db4cff PM |
3453 | } |
3454 | } | |
0c34029a | 3455 | |
394f99a9 | 3456 | rsp->rda = rda; |
b3dbec76 | 3457 | init_waitqueue_head(&rsp->gp_wq); |
016a8d5b | 3458 | init_irq_work(&rsp->wakeup_work, rsp_wakeup); |
f885b7f2 | 3459 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 3460 | for_each_possible_cpu(i) { |
4a90a068 | 3461 | while (i > rnp->grphi) |
0c34029a | 3462 | rnp++; |
394f99a9 | 3463 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
3464 | rcu_boot_init_percpu_data(i, rsp); |
3465 | } | |
6ce75a23 | 3466 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
3467 | } |
3468 | ||
f885b7f2 PM |
3469 | /* |
3470 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 3471 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
3472 | * the ->node array in the rcu_state structure. |
3473 | */ | |
3474 | static void __init rcu_init_geometry(void) | |
3475 | { | |
026ad283 | 3476 | ulong d; |
f885b7f2 PM |
3477 | int i; |
3478 | int j; | |
cca6f393 | 3479 | int n = nr_cpu_ids; |
f885b7f2 PM |
3480 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
3481 | ||
026ad283 PM |
3482 | /* |
3483 | * Initialize any unspecified boot parameters. | |
3484 | * The default values of jiffies_till_first_fqs and | |
3485 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
3486 | * value, which is a function of HZ, then adding one for each | |
3487 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
3488 | */ | |
3489 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
3490 | if (jiffies_till_first_fqs == ULONG_MAX) | |
3491 | jiffies_till_first_fqs = d; | |
3492 | if (jiffies_till_next_fqs == ULONG_MAX) | |
3493 | jiffies_till_next_fqs = d; | |
3494 | ||
f885b7f2 | 3495 | /* If the compile-time values are accurate, just leave. */ |
b17c7035 PM |
3496 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
3497 | nr_cpu_ids == NR_CPUS) | |
f885b7f2 | 3498 | return; |
39479098 PM |
3499 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", |
3500 | rcu_fanout_leaf, nr_cpu_ids); | |
f885b7f2 PM |
3501 | |
3502 | /* | |
3503 | * Compute number of nodes that can be handled an rcu_node tree | |
3504 | * with the given number of levels. Setting rcu_capacity[0] makes | |
3505 | * some of the arithmetic easier. | |
3506 | */ | |
3507 | rcu_capacity[0] = 1; | |
3508 | rcu_capacity[1] = rcu_fanout_leaf; | |
3509 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
3510 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
3511 | ||
3512 | /* | |
3513 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
3514 | * to increase the leaf-level fanout, not decrease it. Of course, | |
3515 | * the leaf-level fanout cannot exceed the number of bits in | |
3516 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
3517 | * the configured number of CPUs. Complain and fall back to the | |
3518 | * compile-time values if these limits are exceeded. | |
3519 | */ | |
3520 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
3521 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
3522 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
3523 | WARN_ON(1); | |
3524 | return; | |
3525 | } | |
3526 | ||
3527 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
3528 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
3529 | if (n <= rcu_capacity[i]) { | |
3530 | for (j = 0; j <= i; j++) | |
3531 | num_rcu_lvl[j] = | |
3532 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
3533 | rcu_num_lvls = i; | |
3534 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
3535 | num_rcu_lvl[j] = 0; | |
3536 | break; | |
3537 | } | |
3538 | ||
3539 | /* Calculate the total number of rcu_node structures. */ | |
3540 | rcu_num_nodes = 0; | |
3541 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
3542 | rcu_num_nodes += num_rcu_lvl[i]; | |
3543 | rcu_num_nodes -= n; | |
3544 | } | |
3545 | ||
9f680ab4 | 3546 | void __init rcu_init(void) |
64db4cff | 3547 | { |
017c4261 | 3548 | int cpu; |
9f680ab4 | 3549 | |
f41d911f | 3550 | rcu_bootup_announce(); |
f885b7f2 | 3551 | rcu_init_geometry(); |
394f99a9 | 3552 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); |
69c8d28c | 3553 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
f41d911f | 3554 | __rcu_init_preempt(); |
b5b39360 | 3555 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
3556 | |
3557 | /* | |
3558 | * We don't need protection against CPU-hotplug here because | |
3559 | * this is called early in boot, before either interrupts | |
3560 | * or the scheduler are operational. | |
3561 | */ | |
3562 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 3563 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
3564 | for_each_online_cpu(cpu) |
3565 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
3566 | } |
3567 | ||
4102adab | 3568 | #include "tree_plugin.h" |