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 | |
15 | * along with this program; if not, write to the Free Software | |
16 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. | |
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> | |
44 | #include <linux/percpu.h> | |
45 | #include <linux/notifier.h> | |
46 | #include <linux/cpu.h> | |
47 | #include <linux/mutex.h> | |
48 | #include <linux/time.h> | |
bbad9379 | 49 | #include <linux/kernel_stat.h> |
a26ac245 PM |
50 | #include <linux/wait.h> |
51 | #include <linux/kthread.h> | |
268bb0ce | 52 | #include <linux/prefetch.h> |
3d3b7db0 PM |
53 | #include <linux/delay.h> |
54 | #include <linux/stop_machine.h> | |
661a85dc | 55 | #include <linux/random.h> |
f7f7bac9 | 56 | #include <linux/ftrace_event.h> |
d1d74d14 | 57 | #include <linux/suspend.h> |
64db4cff | 58 | |
9f77da9f | 59 | #include "rcutree.h" |
29c00b4a PM |
60 | #include <trace/events/rcu.h> |
61 | ||
62 | #include "rcu.h" | |
9f77da9f | 63 | |
f7f7bac9 SRRH |
64 | /* |
65 | * Strings used in tracepoints need to be exported via the | |
66 | * tracing system such that tools like perf and trace-cmd can | |
67 | * translate the string address pointers to actual text. | |
68 | */ | |
69 | #define TPS(x) tracepoint_string(x) | |
70 | ||
64db4cff PM |
71 | /* Data structures. */ |
72 | ||
f885b7f2 | 73 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
394f2769 | 74 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; |
88b91c7c | 75 | |
f7f7bac9 SRRH |
76 | /* |
77 | * In order to export the rcu_state name to the tracing tools, it | |
78 | * needs to be added in the __tracepoint_string section. | |
79 | * This requires defining a separate variable tp_<sname>_varname | |
80 | * that points to the string being used, and this will allow | |
81 | * the tracing userspace tools to be able to decipher the string | |
82 | * address to the matching string. | |
83 | */ | |
a41bfeb2 | 84 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ |
f7f7bac9 SRRH |
85 | static char sname##_varname[] = #sname; \ |
86 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \ | |
a41bfeb2 | 87 | struct rcu_state sname##_state = { \ |
6c90cc7b | 88 | .level = { &sname##_state.node[0] }, \ |
037b64ed | 89 | .call = cr, \ |
af446b70 | 90 | .fqs_state = RCU_GP_IDLE, \ |
42c3533e PM |
91 | .gpnum = 0UL - 300UL, \ |
92 | .completed = 0UL - 300UL, \ | |
7b2e6011 | 93 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
6c90cc7b PM |
94 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
95 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 96 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
a4fbe35a | 97 | .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \ |
f7f7bac9 | 98 | .name = sname##_varname, \ |
a4889858 | 99 | .abbr = sabbr, \ |
a41bfeb2 SRRH |
100 | }; \ |
101 | DEFINE_PER_CPU(struct rcu_data, sname##_data) | |
64db4cff | 102 | |
a41bfeb2 SRRH |
103 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); |
104 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); | |
b1f77b05 | 105 | |
27f4d280 | 106 | static struct rcu_state *rcu_state; |
6ce75a23 | 107 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 108 | |
f885b7f2 PM |
109 | /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */ |
110 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF; | |
7e5c2dfb | 111 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 PM |
112 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
113 | static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */ | |
114 | NUM_RCU_LVL_0, | |
115 | NUM_RCU_LVL_1, | |
116 | NUM_RCU_LVL_2, | |
117 | NUM_RCU_LVL_3, | |
118 | NUM_RCU_LVL_4, | |
119 | }; | |
120 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ | |
121 | ||
b0d30417 PM |
122 | /* |
123 | * The rcu_scheduler_active variable transitions from zero to one just | |
124 | * before the first task is spawned. So when this variable is zero, RCU | |
125 | * can assume that there is but one task, allowing RCU to (for example) | |
b44f6656 | 126 | * optimize synchronize_sched() to a simple barrier(). When this variable |
b0d30417 PM |
127 | * is one, RCU must actually do all the hard work required to detect real |
128 | * grace periods. This variable is also used to suppress boot-time false | |
129 | * positives from lockdep-RCU error checking. | |
130 | */ | |
bbad9379 PM |
131 | int rcu_scheduler_active __read_mostly; |
132 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
133 | ||
b0d30417 PM |
134 | /* |
135 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
136 | * during the early_initcall() processing, which is after the scheduler | |
137 | * is capable of creating new tasks. So RCU processing (for example, | |
138 | * creating tasks for RCU priority boosting) must be delayed until after | |
139 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
140 | * currently delay invocation of any RCU callbacks until after this point. | |
141 | * | |
142 | * It might later prove better for people registering RCU callbacks during | |
143 | * early boot to take responsibility for these callbacks, but one step at | |
144 | * a time. | |
145 | */ | |
146 | static int rcu_scheduler_fully_active __read_mostly; | |
147 | ||
a46e0899 PM |
148 | #ifdef CONFIG_RCU_BOOST |
149 | ||
a26ac245 PM |
150 | /* |
151 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
152 | * handle all flavors of RCU. | |
153 | */ | |
154 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
d71df90e | 155 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); |
5ece5bab | 156 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); |
d71df90e | 157 | DEFINE_PER_CPU(char, rcu_cpu_has_work); |
a26ac245 | 158 | |
a46e0899 PM |
159 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
160 | ||
5d01bbd1 | 161 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
162 | static void invoke_rcu_core(void); |
163 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 164 | |
4a298656 PM |
165 | /* |
166 | * Track the rcutorture test sequence number and the update version | |
167 | * number within a given test. The rcutorture_testseq is incremented | |
168 | * on every rcutorture module load and unload, so has an odd value | |
169 | * when a test is running. The rcutorture_vernum is set to zero | |
170 | * when rcutorture starts and is incremented on each rcutorture update. | |
171 | * These variables enable correlating rcutorture output with the | |
172 | * RCU tracing information. | |
173 | */ | |
174 | unsigned long rcutorture_testseq; | |
175 | unsigned long rcutorture_vernum; | |
176 | ||
fc2219d4 PM |
177 | /* |
178 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
179 | * permit this function to be invoked without holding the root rcu_node | |
180 | * structure's ->lock, but of course results can be subject to change. | |
181 | */ | |
182 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
183 | { | |
184 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
185 | } | |
186 | ||
b1f77b05 | 187 | /* |
d6714c22 | 188 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 189 | * how many quiescent states passed, just if there was at least |
d6714c22 | 190 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 191 | * The caller must have disabled preemption. |
b1f77b05 | 192 | */ |
d6714c22 | 193 | void rcu_sched_qs(int cpu) |
b1f77b05 | 194 | { |
25502a6c | 195 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu); |
f41d911f | 196 | |
e4cc1f22 | 197 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 198 | trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 199 | rdp->passed_quiesce = 1; |
b1f77b05 IM |
200 | } |
201 | ||
d6714c22 | 202 | void rcu_bh_qs(int cpu) |
b1f77b05 | 203 | { |
25502a6c | 204 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu); |
f41d911f | 205 | |
e4cc1f22 | 206 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 207 | trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 208 | rdp->passed_quiesce = 1; |
b1f77b05 | 209 | } |
64db4cff | 210 | |
25502a6c PM |
211 | /* |
212 | * Note a context switch. This is a quiescent state for RCU-sched, | |
213 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 214 | * The caller must have disabled preemption. |
25502a6c PM |
215 | */ |
216 | void rcu_note_context_switch(int cpu) | |
217 | { | |
f7f7bac9 | 218 | trace_rcu_utilization(TPS("Start context switch")); |
25502a6c | 219 | rcu_sched_qs(cpu); |
cba6d0d6 | 220 | rcu_preempt_note_context_switch(cpu); |
f7f7bac9 | 221 | trace_rcu_utilization(TPS("End context switch")); |
25502a6c | 222 | } |
29ce8310 | 223 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 224 | |
90a4d2c0 | 225 | DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { |
29e37d81 | 226 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, |
23b5c8fa | 227 | .dynticks = ATOMIC_INIT(1), |
2333210b PM |
228 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
229 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
230 | .dynticks_idle = ATOMIC_INIT(1), | |
231 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
90a4d2c0 | 232 | }; |
64db4cff | 233 | |
878d7439 ED |
234 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
235 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
236 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 237 | |
878d7439 ED |
238 | module_param(blimit, long, 0444); |
239 | module_param(qhimark, long, 0444); | |
240 | module_param(qlowmark, long, 0444); | |
3d76c082 | 241 | |
026ad283 PM |
242 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
243 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
244 | |
245 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
246 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
247 | ||
910ee45d PM |
248 | static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
249 | struct rcu_data *rdp); | |
217af2a2 PM |
250 | static void force_qs_rnp(struct rcu_state *rsp, |
251 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
252 | unsigned long *maxj), | |
253 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 254 | static void force_quiescent_state(struct rcu_state *rsp); |
a157229c | 255 | static int rcu_pending(int cpu); |
64db4cff PM |
256 | |
257 | /* | |
d6714c22 | 258 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
64db4cff | 259 | */ |
d6714c22 | 260 | long rcu_batches_completed_sched(void) |
64db4cff | 261 | { |
d6714c22 | 262 | return rcu_sched_state.completed; |
64db4cff | 263 | } |
d6714c22 | 264 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
265 | |
266 | /* | |
267 | * Return the number of RCU BH batches processed thus far for debug & stats. | |
268 | */ | |
269 | long rcu_batches_completed_bh(void) | |
270 | { | |
271 | return rcu_bh_state.completed; | |
272 | } | |
273 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
274 | ||
bf66f18e PM |
275 | /* |
276 | * Force a quiescent state for RCU BH. | |
277 | */ | |
278 | void rcu_bh_force_quiescent_state(void) | |
279 | { | |
4cdfc175 | 280 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
281 | } |
282 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
283 | ||
4a298656 PM |
284 | /* |
285 | * Record the number of times rcutorture tests have been initiated and | |
286 | * terminated. This information allows the debugfs tracing stats to be | |
287 | * correlated to the rcutorture messages, even when the rcutorture module | |
288 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
289 | * store this state in rcutorture itself. | |
290 | */ | |
291 | void rcutorture_record_test_transition(void) | |
292 | { | |
293 | rcutorture_testseq++; | |
294 | rcutorture_vernum = 0; | |
295 | } | |
296 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
297 | ||
298 | /* | |
299 | * Record the number of writer passes through the current rcutorture test. | |
300 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
301 | * messages. | |
302 | */ | |
303 | void rcutorture_record_progress(unsigned long vernum) | |
304 | { | |
305 | rcutorture_vernum++; | |
306 | } | |
307 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
308 | ||
bf66f18e PM |
309 | /* |
310 | * Force a quiescent state for RCU-sched. | |
311 | */ | |
312 | void rcu_sched_force_quiescent_state(void) | |
313 | { | |
4cdfc175 | 314 | force_quiescent_state(&rcu_sched_state); |
bf66f18e PM |
315 | } |
316 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
317 | ||
64db4cff PM |
318 | /* |
319 | * Does the CPU have callbacks ready to be invoked? | |
320 | */ | |
321 | static int | |
322 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
323 | { | |
3fbfbf7a PM |
324 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
325 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
326 | } |
327 | ||
328 | /* | |
dc35c893 PM |
329 | * Does the current CPU require a not-yet-started grace period? |
330 | * The caller must have disabled interrupts to prevent races with | |
331 | * normal callback registry. | |
64db4cff PM |
332 | */ |
333 | static int | |
334 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
335 | { | |
dc35c893 | 336 | int i; |
3fbfbf7a | 337 | |
dc35c893 PM |
338 | if (rcu_gp_in_progress(rsp)) |
339 | return 0; /* No, a grace period is already in progress. */ | |
dae6e64d | 340 | if (rcu_nocb_needs_gp(rsp)) |
34ed6246 | 341 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
342 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
343 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
344 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
345 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
346 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
347 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
348 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), | |
349 | rdp->nxtcompleted[i])) | |
350 | return 1; /* Yes, CBs for future grace period. */ | |
351 | return 0; /* No grace period needed. */ | |
64db4cff PM |
352 | } |
353 | ||
354 | /* | |
355 | * Return the root node of the specified rcu_state structure. | |
356 | */ | |
357 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
358 | { | |
359 | return &rsp->node[0]; | |
360 | } | |
361 | ||
9b2e4f18 | 362 | /* |
adf5091e | 363 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
364 | * |
365 | * If the new value of the ->dynticks_nesting counter now is zero, | |
366 | * we really have entered idle, and must do the appropriate accounting. | |
367 | * The caller must have disabled interrupts. | |
368 | */ | |
adf5091e FW |
369 | static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval, |
370 | bool user) | |
9b2e4f18 | 371 | { |
f7f7bac9 | 372 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 373 | if (!user && !is_idle_task(current)) { |
0989cb46 PM |
374 | struct task_struct *idle = idle_task(smp_processor_id()); |
375 | ||
f7f7bac9 | 376 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 377 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
378 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
379 | current->pid, current->comm, | |
380 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 381 | } |
aea1b35e | 382 | rcu_prepare_for_idle(smp_processor_id()); |
9b2e4f18 PM |
383 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
384 | smp_mb__before_atomic_inc(); /* See above. */ | |
385 | atomic_inc(&rdtp->dynticks); | |
386 | smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */ | |
387 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
c44e2cdd PM |
388 | |
389 | /* | |
adf5091e | 390 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
391 | * in an RCU read-side critical section. |
392 | */ | |
393 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
394 | "Illegal idle entry in RCU read-side critical section."); | |
395 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
396 | "Illegal idle entry in RCU-bh read-side critical section."); | |
397 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
398 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 399 | } |
64db4cff | 400 | |
adf5091e FW |
401 | /* |
402 | * Enter an RCU extended quiescent state, which can be either the | |
403 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 404 | */ |
adf5091e | 405 | static void rcu_eqs_enter(bool user) |
64db4cff | 406 | { |
4145fa7f | 407 | long long oldval; |
64db4cff PM |
408 | struct rcu_dynticks *rdtp; |
409 | ||
64db4cff | 410 | rdtp = &__get_cpu_var(rcu_dynticks); |
4145fa7f | 411 | oldval = rdtp->dynticks_nesting; |
29e37d81 PM |
412 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
413 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) | |
414 | rdtp->dynticks_nesting = 0; | |
415 | else | |
416 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; | |
adf5091e | 417 | rcu_eqs_enter_common(rdtp, oldval, user); |
64db4cff | 418 | } |
adf5091e FW |
419 | |
420 | /** | |
421 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
422 | * | |
423 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
424 | * read-side critical sections can occur. (Though RCU read-side | |
425 | * critical sections can occur in irq handlers in idle, a possibility | |
426 | * handled by irq_enter() and irq_exit().) | |
427 | * | |
428 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
429 | * the possibility of usermode upcalls having messed up our count | |
430 | * of interrupt nesting level during the prior busy period. | |
431 | */ | |
432 | void rcu_idle_enter(void) | |
433 | { | |
c5d900bf FW |
434 | unsigned long flags; |
435 | ||
436 | local_irq_save(flags); | |
cb349ca9 | 437 | rcu_eqs_enter(false); |
eb348b89 | 438 | rcu_sysidle_enter(&__get_cpu_var(rcu_dynticks), 0); |
c5d900bf | 439 | local_irq_restore(flags); |
adf5091e | 440 | } |
8a2ecf47 | 441 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 442 | |
2b1d5024 | 443 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
444 | /** |
445 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
446 | * | |
447 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
448 | * is permitted between this call and rcu_user_exit(). This way the | |
449 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
450 | * when the CPU runs in userspace. | |
451 | */ | |
452 | void rcu_user_enter(void) | |
453 | { | |
91d1aa43 | 454 | rcu_eqs_enter(1); |
adf5091e | 455 | } |
2b1d5024 | 456 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 457 | |
9b2e4f18 PM |
458 | /** |
459 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
460 | * | |
461 | * Exit from an interrupt handler, which might possibly result in entering | |
462 | * idle mode, in other words, leaving the mode in which read-side critical | |
463 | * sections can occur. | |
64db4cff | 464 | * |
9b2e4f18 PM |
465 | * This code assumes that the idle loop never does anything that might |
466 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
467 | * architecture violates this assumption, RCU will give you what you | |
468 | * deserve, good and hard. But very infrequently and irreproducibly. | |
469 | * | |
470 | * Use things like work queues to work around this limitation. | |
471 | * | |
472 | * You have been warned. | |
64db4cff | 473 | */ |
9b2e4f18 | 474 | void rcu_irq_exit(void) |
64db4cff PM |
475 | { |
476 | unsigned long flags; | |
4145fa7f | 477 | long long oldval; |
64db4cff PM |
478 | struct rcu_dynticks *rdtp; |
479 | ||
480 | local_irq_save(flags); | |
481 | rdtp = &__get_cpu_var(rcu_dynticks); | |
4145fa7f | 482 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
483 | rdtp->dynticks_nesting--; |
484 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 | 485 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 486 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 487 | else |
cb349ca9 | 488 | rcu_eqs_enter_common(rdtp, oldval, true); |
eb348b89 | 489 | rcu_sysidle_enter(rdtp, 1); |
9b2e4f18 PM |
490 | local_irq_restore(flags); |
491 | } | |
492 | ||
493 | /* | |
adf5091e | 494 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
495 | * |
496 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
497 | * we really have exited idle, and must do the appropriate accounting. | |
498 | * The caller must have disabled interrupts. | |
499 | */ | |
adf5091e FW |
500 | static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval, |
501 | int user) | |
9b2e4f18 | 502 | { |
23b5c8fa PM |
503 | smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */ |
504 | atomic_inc(&rdtp->dynticks); | |
505 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
506 | smp_mb__after_atomic_inc(); /* See above. */ | |
507 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
7cb92499 | 508 | rcu_cleanup_after_idle(smp_processor_id()); |
f7f7bac9 | 509 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 510 | if (!user && !is_idle_task(current)) { |
0989cb46 PM |
511 | struct task_struct *idle = idle_task(smp_processor_id()); |
512 | ||
f7f7bac9 | 513 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 514 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 515 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
516 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
517 | current->pid, current->comm, | |
518 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
519 | } |
520 | } | |
521 | ||
adf5091e FW |
522 | /* |
523 | * Exit an RCU extended quiescent state, which can be either the | |
524 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 525 | */ |
adf5091e | 526 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 527 | { |
9b2e4f18 PM |
528 | struct rcu_dynticks *rdtp; |
529 | long long oldval; | |
530 | ||
9b2e4f18 PM |
531 | rdtp = &__get_cpu_var(rcu_dynticks); |
532 | oldval = rdtp->dynticks_nesting; | |
29e37d81 PM |
533 | WARN_ON_ONCE(oldval < 0); |
534 | if (oldval & DYNTICK_TASK_NEST_MASK) | |
535 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; | |
536 | else | |
537 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; | |
adf5091e | 538 | rcu_eqs_exit_common(rdtp, oldval, user); |
9b2e4f18 | 539 | } |
adf5091e FW |
540 | |
541 | /** | |
542 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
543 | * | |
544 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
545 | * read-side critical sections can occur. | |
546 | * | |
547 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
548 | * allow for the possibility of usermode upcalls messing up our count | |
549 | * of interrupt nesting level during the busy period that is just | |
550 | * now starting. | |
551 | */ | |
552 | void rcu_idle_exit(void) | |
553 | { | |
c5d900bf FW |
554 | unsigned long flags; |
555 | ||
556 | local_irq_save(flags); | |
cb349ca9 | 557 | rcu_eqs_exit(false); |
eb348b89 | 558 | rcu_sysidle_exit(&__get_cpu_var(rcu_dynticks), 0); |
c5d900bf | 559 | local_irq_restore(flags); |
adf5091e | 560 | } |
8a2ecf47 | 561 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 562 | |
2b1d5024 | 563 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
564 | /** |
565 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
566 | * | |
567 | * Exit RCU idle mode while entering the kernel because it can | |
568 | * run a RCU read side critical section anytime. | |
569 | */ | |
570 | void rcu_user_exit(void) | |
571 | { | |
91d1aa43 | 572 | rcu_eqs_exit(1); |
adf5091e | 573 | } |
2b1d5024 | 574 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 575 | |
9b2e4f18 PM |
576 | /** |
577 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
578 | * | |
579 | * Enter an interrupt handler, which might possibly result in exiting | |
580 | * idle mode, in other words, entering the mode in which read-side critical | |
581 | * sections can occur. | |
582 | * | |
583 | * Note that the Linux kernel is fully capable of entering an interrupt | |
584 | * handler that it never exits, for example when doing upcalls to | |
585 | * user mode! This code assumes that the idle loop never does upcalls to | |
586 | * user mode. If your architecture does do upcalls from the idle loop (or | |
587 | * does anything else that results in unbalanced calls to the irq_enter() | |
588 | * and irq_exit() functions), RCU will give you what you deserve, good | |
589 | * and hard. But very infrequently and irreproducibly. | |
590 | * | |
591 | * Use things like work queues to work around this limitation. | |
592 | * | |
593 | * You have been warned. | |
594 | */ | |
595 | void rcu_irq_enter(void) | |
596 | { | |
597 | unsigned long flags; | |
598 | struct rcu_dynticks *rdtp; | |
599 | long long oldval; | |
600 | ||
601 | local_irq_save(flags); | |
602 | rdtp = &__get_cpu_var(rcu_dynticks); | |
603 | oldval = rdtp->dynticks_nesting; | |
604 | rdtp->dynticks_nesting++; | |
605 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 | 606 | if (oldval) |
f7f7bac9 | 607 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 608 | else |
cb349ca9 | 609 | rcu_eqs_exit_common(rdtp, oldval, true); |
eb348b89 | 610 | rcu_sysidle_exit(rdtp, 1); |
64db4cff | 611 | local_irq_restore(flags); |
64db4cff PM |
612 | } |
613 | ||
614 | /** | |
615 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
616 | * | |
617 | * If the CPU was idle with dynamic ticks active, and there is no | |
618 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
619 | * RCU grace-period handling know that the CPU is active. | |
620 | */ | |
621 | void rcu_nmi_enter(void) | |
622 | { | |
623 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); | |
624 | ||
23b5c8fa PM |
625 | if (rdtp->dynticks_nmi_nesting == 0 && |
626 | (atomic_read(&rdtp->dynticks) & 0x1)) | |
64db4cff | 627 | return; |
23b5c8fa PM |
628 | rdtp->dynticks_nmi_nesting++; |
629 | smp_mb__before_atomic_inc(); /* Force delay from prior write. */ | |
630 | atomic_inc(&rdtp->dynticks); | |
631 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
632 | smp_mb__after_atomic_inc(); /* See above. */ | |
633 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
64db4cff PM |
634 | } |
635 | ||
636 | /** | |
637 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
638 | * | |
639 | * If the CPU was idle with dynamic ticks active, and there is no | |
640 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
641 | * RCU grace-period handling know that the CPU is no longer active. | |
642 | */ | |
643 | void rcu_nmi_exit(void) | |
644 | { | |
645 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); | |
646 | ||
23b5c8fa PM |
647 | if (rdtp->dynticks_nmi_nesting == 0 || |
648 | --rdtp->dynticks_nmi_nesting != 0) | |
64db4cff | 649 | return; |
23b5c8fa PM |
650 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
651 | smp_mb__before_atomic_inc(); /* See above. */ | |
652 | atomic_inc(&rdtp->dynticks); | |
653 | smp_mb__after_atomic_inc(); /* Force delay to next write. */ | |
654 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
64db4cff PM |
655 | } |
656 | ||
657 | /** | |
9b2e4f18 | 658 | * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle |
64db4cff | 659 | * |
9b2e4f18 | 660 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 661 | * or NMI handler, return true. |
64db4cff | 662 | */ |
9b2e4f18 | 663 | int rcu_is_cpu_idle(void) |
64db4cff | 664 | { |
34240697 PM |
665 | int ret; |
666 | ||
667 | preempt_disable(); | |
668 | ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0; | |
669 | preempt_enable(); | |
670 | return ret; | |
64db4cff | 671 | } |
e6b80a3b | 672 | EXPORT_SYMBOL(rcu_is_cpu_idle); |
64db4cff | 673 | |
62fde6ed | 674 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
675 | |
676 | /* | |
677 | * Is the current CPU online? Disable preemption to avoid false positives | |
678 | * that could otherwise happen due to the current CPU number being sampled, | |
679 | * this task being preempted, its old CPU being taken offline, resuming | |
680 | * on some other CPU, then determining that its old CPU is now offline. | |
681 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
682 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
683 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
684 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
685 | * offline to continue to use RCU for one jiffy after marking itself | |
686 | * offline in the cpu_online_mask. This leniency is necessary given the | |
687 | * non-atomic nature of the online and offline processing, for example, | |
688 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
689 | * notifiers. | |
690 | * | |
691 | * This is also why RCU internally marks CPUs online during the | |
692 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
693 | * |
694 | * Disable checking if in an NMI handler because we cannot safely report | |
695 | * errors from NMI handlers anyway. | |
696 | */ | |
697 | bool rcu_lockdep_current_cpu_online(void) | |
698 | { | |
2036d94a PM |
699 | struct rcu_data *rdp; |
700 | struct rcu_node *rnp; | |
c0d6d01b PM |
701 | bool ret; |
702 | ||
703 | if (in_nmi()) | |
704 | return 1; | |
705 | preempt_disable(); | |
2036d94a PM |
706 | rdp = &__get_cpu_var(rcu_sched_data); |
707 | rnp = rdp->mynode; | |
708 | ret = (rdp->grpmask & rnp->qsmaskinit) || | |
c0d6d01b PM |
709 | !rcu_scheduler_fully_active; |
710 | preempt_enable(); | |
711 | return ret; | |
712 | } | |
713 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
714 | ||
62fde6ed | 715 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 716 | |
64db4cff | 717 | /** |
9b2e4f18 | 718 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 719 | * |
9b2e4f18 PM |
720 | * If the current CPU is idle or running at a first-level (not nested) |
721 | * interrupt from idle, return true. The caller must have at least | |
722 | * disabled preemption. | |
64db4cff | 723 | */ |
62e3cb14 | 724 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 725 | { |
9b2e4f18 | 726 | return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1; |
64db4cff PM |
727 | } |
728 | ||
64db4cff PM |
729 | /* |
730 | * Snapshot the specified CPU's dynticks counter so that we can later | |
731 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 732 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 733 | */ |
217af2a2 PM |
734 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
735 | bool *isidle, unsigned long *maxj) | |
64db4cff | 736 | { |
23b5c8fa | 737 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 738 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
f0e7c19d | 739 | return (rdp->dynticks_snap & 0x1) == 0; |
64db4cff PM |
740 | } |
741 | ||
742 | /* | |
743 | * Return true if the specified CPU has passed through a quiescent | |
744 | * state by virtue of being in or having passed through an dynticks | |
745 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 746 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 747 | */ |
217af2a2 PM |
748 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
749 | bool *isidle, unsigned long *maxj) | |
64db4cff | 750 | { |
7eb4f455 PM |
751 | unsigned int curr; |
752 | unsigned int snap; | |
64db4cff | 753 | |
7eb4f455 PM |
754 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
755 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
756 | |
757 | /* | |
758 | * If the CPU passed through or entered a dynticks idle phase with | |
759 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
760 | * already acknowledged the request to pass through a quiescent | |
761 | * state. Either way, that CPU cannot possibly be in an RCU | |
762 | * read-side critical section that started before the beginning | |
763 | * of the current RCU grace period. | |
764 | */ | |
7eb4f455 | 765 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 766 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
767 | rdp->dynticks_fqs++; |
768 | return 1; | |
769 | } | |
770 | ||
a82dcc76 PM |
771 | /* |
772 | * Check for the CPU being offline, but only if the grace period | |
773 | * is old enough. We don't need to worry about the CPU changing | |
774 | * state: If we see it offline even once, it has been through a | |
775 | * quiescent state. | |
776 | * | |
777 | * The reason for insisting that the grace period be at least | |
778 | * one jiffy old is that CPUs that are not quite online and that | |
779 | * have just gone offline can still execute RCU read-side critical | |
780 | * sections. | |
781 | */ | |
782 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
783 | return 0; /* Grace period is not old enough. */ | |
784 | barrier(); | |
785 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 786 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
787 | rdp->offline_fqs++; |
788 | return 1; | |
789 | } | |
65d798f0 PM |
790 | |
791 | /* | |
792 | * There is a possibility that a CPU in adaptive-ticks state | |
793 | * might run in the kernel with the scheduling-clock tick disabled | |
794 | * for an extended time period. Invoke rcu_kick_nohz_cpu() to | |
795 | * force the CPU to restart the scheduling-clock tick in this | |
796 | * CPU is in this state. | |
797 | */ | |
798 | rcu_kick_nohz_cpu(rdp->cpu); | |
799 | ||
a82dcc76 | 800 | return 0; |
64db4cff PM |
801 | } |
802 | ||
64db4cff PM |
803 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
804 | { | |
805 | rsp->gp_start = jiffies; | |
6bfc09e2 | 806 | rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check(); |
64db4cff PM |
807 | } |
808 | ||
b637a328 PM |
809 | /* |
810 | * Dump stacks of all tasks running on stalled CPUs. This is a fallback | |
811 | * for architectures that do not implement trigger_all_cpu_backtrace(). | |
812 | * The NMI-triggered stack traces are more accurate because they are | |
813 | * printed by the target CPU. | |
814 | */ | |
815 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
816 | { | |
817 | int cpu; | |
818 | unsigned long flags; | |
819 | struct rcu_node *rnp; | |
820 | ||
821 | rcu_for_each_leaf_node(rsp, rnp) { | |
822 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
823 | if (rnp->qsmask != 0) { | |
824 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
825 | if (rnp->qsmask & (1UL << cpu)) | |
826 | dump_cpu_task(rnp->grplo + cpu); | |
827 | } | |
828 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
829 | } | |
830 | } | |
831 | ||
64db4cff PM |
832 | static void print_other_cpu_stall(struct rcu_state *rsp) |
833 | { | |
834 | int cpu; | |
835 | long delta; | |
836 | unsigned long flags; | |
285fe294 | 837 | int ndetected = 0; |
64db4cff | 838 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 839 | long totqlen = 0; |
64db4cff PM |
840 | |
841 | /* Only let one CPU complain about others per time interval. */ | |
842 | ||
1304afb2 | 843 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 844 | delta = jiffies - rsp->jiffies_stall; |
fc2219d4 | 845 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 846 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
847 | return; |
848 | } | |
6bfc09e2 | 849 | rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 850 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 851 | |
8cdd32a9 PM |
852 | /* |
853 | * OK, time to rat on our buddy... | |
854 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
855 | * RCU CPU stall warnings. | |
856 | */ | |
d7f3e207 | 857 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 858 | rsp->name); |
a858af28 | 859 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 860 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 861 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 862 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
863 | if (rnp->qsmask != 0) { |
864 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
865 | if (rnp->qsmask & (1UL << cpu)) { | |
866 | print_cpu_stall_info(rsp, | |
867 | rnp->grplo + cpu); | |
868 | ndetected++; | |
869 | } | |
870 | } | |
3acd9eb3 | 871 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 872 | } |
a858af28 PM |
873 | |
874 | /* | |
875 | * Now rat on any tasks that got kicked up to the root rcu_node | |
876 | * due to CPU offlining. | |
877 | */ | |
878 | rnp = rcu_get_root(rsp); | |
879 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
285fe294 | 880 | ndetected += rcu_print_task_stall(rnp); |
a858af28 PM |
881 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
882 | ||
883 | print_cpu_stall_info_end(); | |
53bb857c PM |
884 | for_each_possible_cpu(cpu) |
885 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
886 | pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n", | |
eee05882 | 887 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
53bb857c | 888 | rsp->gpnum, rsp->completed, totqlen); |
9bc8b558 | 889 | if (ndetected == 0) |
d7f3e207 | 890 | pr_err("INFO: Stall ended before state dump start\n"); |
9bc8b558 | 891 | else if (!trigger_all_cpu_backtrace()) |
b637a328 | 892 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 893 | |
4cdfc175 | 894 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
895 | |
896 | rcu_print_detail_task_stall(rsp); | |
897 | ||
4cdfc175 | 898 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
899 | } |
900 | ||
901 | static void print_cpu_stall(struct rcu_state *rsp) | |
902 | { | |
53bb857c | 903 | int cpu; |
64db4cff PM |
904 | unsigned long flags; |
905 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 906 | long totqlen = 0; |
64db4cff | 907 | |
8cdd32a9 PM |
908 | /* |
909 | * OK, time to rat on ourselves... | |
910 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
911 | * RCU CPU stall warnings. | |
912 | */ | |
d7f3e207 | 913 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
914 | print_cpu_stall_info_begin(); |
915 | print_cpu_stall_info(rsp, smp_processor_id()); | |
916 | print_cpu_stall_info_end(); | |
53bb857c PM |
917 | for_each_possible_cpu(cpu) |
918 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
919 | pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n", | |
920 | jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen); | |
4627e240 PM |
921 | if (!trigger_all_cpu_backtrace()) |
922 | dump_stack(); | |
c1dc0b9c | 923 | |
1304afb2 | 924 | raw_spin_lock_irqsave(&rnp->lock, flags); |
20133cfc | 925 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall)) |
13cfcca0 | 926 | rsp->jiffies_stall = jiffies + |
6bfc09e2 | 927 | 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 928 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 929 | |
64db4cff PM |
930 | set_need_resched(); /* kick ourselves to get things going. */ |
931 | } | |
932 | ||
933 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
934 | { | |
bad6e139 PM |
935 | unsigned long j; |
936 | unsigned long js; | |
64db4cff PM |
937 | struct rcu_node *rnp; |
938 | ||
742734ee | 939 | if (rcu_cpu_stall_suppress) |
c68de209 | 940 | return; |
bad6e139 PM |
941 | j = ACCESS_ONCE(jiffies); |
942 | js = ACCESS_ONCE(rsp->jiffies_stall); | |
64db4cff | 943 | rnp = rdp->mynode; |
c96ea7cf PM |
944 | if (rcu_gp_in_progress(rsp) && |
945 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) { | |
64db4cff PM |
946 | |
947 | /* We haven't checked in, so go dump stack. */ | |
948 | print_cpu_stall(rsp); | |
949 | ||
bad6e139 PM |
950 | } else if (rcu_gp_in_progress(rsp) && |
951 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 952 | |
bad6e139 | 953 | /* They had a few time units to dump stack, so complain. */ |
64db4cff PM |
954 | print_other_cpu_stall(rsp); |
955 | } | |
956 | } | |
957 | ||
53d84e00 PM |
958 | /** |
959 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
960 | * | |
961 | * Set the stall-warning timeout way off into the future, thus preventing | |
962 | * any RCU CPU stall-warning messages from appearing in the current set of | |
963 | * RCU grace periods. | |
964 | * | |
965 | * The caller must disable hard irqs. | |
966 | */ | |
967 | void rcu_cpu_stall_reset(void) | |
968 | { | |
6ce75a23 PM |
969 | struct rcu_state *rsp; |
970 | ||
971 | for_each_rcu_flavor(rsp) | |
972 | rsp->jiffies_stall = jiffies + ULONG_MAX / 2; | |
53d84e00 PM |
973 | } |
974 | ||
3f5d3ea6 PM |
975 | /* |
976 | * Initialize the specified rcu_data structure's callback list to empty. | |
977 | */ | |
978 | static void init_callback_list(struct rcu_data *rdp) | |
979 | { | |
980 | int i; | |
981 | ||
34ed6246 PM |
982 | if (init_nocb_callback_list(rdp)) |
983 | return; | |
3f5d3ea6 PM |
984 | rdp->nxtlist = NULL; |
985 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
986 | rdp->nxttail[i] = &rdp->nxtlist; | |
987 | } | |
988 | ||
dc35c893 PM |
989 | /* |
990 | * Determine the value that ->completed will have at the end of the | |
991 | * next subsequent grace period. This is used to tag callbacks so that | |
992 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
993 | * been dyntick-idle for an extended period with callbacks under the | |
994 | * influence of RCU_FAST_NO_HZ. | |
995 | * | |
996 | * The caller must hold rnp->lock with interrupts disabled. | |
997 | */ | |
998 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
999 | struct rcu_node *rnp) | |
1000 | { | |
1001 | /* | |
1002 | * If RCU is idle, we just wait for the next grace period. | |
1003 | * But we can only be sure that RCU is idle if we are looking | |
1004 | * at the root rcu_node structure -- otherwise, a new grace | |
1005 | * period might have started, but just not yet gotten around | |
1006 | * to initializing the current non-root rcu_node structure. | |
1007 | */ | |
1008 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1009 | return rnp->completed + 1; | |
1010 | ||
1011 | /* | |
1012 | * Otherwise, wait for a possible partial grace period and | |
1013 | * then the subsequent full grace period. | |
1014 | */ | |
1015 | return rnp->completed + 2; | |
1016 | } | |
1017 | ||
0446be48 PM |
1018 | /* |
1019 | * Trace-event helper function for rcu_start_future_gp() and | |
1020 | * rcu_nocb_wait_gp(). | |
1021 | */ | |
1022 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1023 | unsigned long c, const char *s) |
0446be48 PM |
1024 | { |
1025 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1026 | rnp->completed, c, rnp->level, | |
1027 | rnp->grplo, rnp->grphi, s); | |
1028 | } | |
1029 | ||
1030 | /* | |
1031 | * Start some future grace period, as needed to handle newly arrived | |
1032 | * callbacks. The required future grace periods are recorded in each | |
1033 | * rcu_node structure's ->need_future_gp field. | |
1034 | * | |
1035 | * The caller must hold the specified rcu_node structure's ->lock. | |
1036 | */ | |
1037 | static unsigned long __maybe_unused | |
1038 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp) | |
1039 | { | |
1040 | unsigned long c; | |
1041 | int i; | |
1042 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); | |
1043 | ||
1044 | /* | |
1045 | * Pick up grace-period number for new callbacks. If this | |
1046 | * grace period is already marked as needed, return to the caller. | |
1047 | */ | |
1048 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1049 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1050 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1051 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
0446be48 PM |
1052 | return c; |
1053 | } | |
1054 | ||
1055 | /* | |
1056 | * If either this rcu_node structure or the root rcu_node structure | |
1057 | * believe that a grace period is in progress, then we must wait | |
1058 | * for the one following, which is in "c". Because our request | |
1059 | * will be noticed at the end of the current grace period, we don't | |
1060 | * need to explicitly start one. | |
1061 | */ | |
1062 | if (rnp->gpnum != rnp->completed || | |
1063 | ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) { | |
1064 | rnp->need_future_gp[c & 0x1]++; | |
f7f7bac9 | 1065 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
0446be48 PM |
1066 | return c; |
1067 | } | |
1068 | ||
1069 | /* | |
1070 | * There might be no grace period in progress. If we don't already | |
1071 | * hold it, acquire the root rcu_node structure's lock in order to | |
1072 | * start one (if needed). | |
1073 | */ | |
1074 | if (rnp != rnp_root) | |
1075 | raw_spin_lock(&rnp_root->lock); | |
1076 | ||
1077 | /* | |
1078 | * Get a new grace-period number. If there really is no grace | |
1079 | * period in progress, it will be smaller than the one we obtained | |
1080 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1081 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1082 | */ | |
1083 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1084 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1085 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1086 | rdp->nxtcompleted[i] = c; | |
1087 | ||
1088 | /* | |
1089 | * If the needed for the required grace period is already | |
1090 | * recorded, trace and leave. | |
1091 | */ | |
1092 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1093 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1094 | goto unlock_out; |
1095 | } | |
1096 | ||
1097 | /* Record the need for the future grace period. */ | |
1098 | rnp_root->need_future_gp[c & 0x1]++; | |
1099 | ||
1100 | /* If a grace period is not already in progress, start one. */ | |
1101 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1102 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1103 | } else { |
f7f7bac9 | 1104 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
910ee45d | 1105 | rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1106 | } |
1107 | unlock_out: | |
1108 | if (rnp != rnp_root) | |
1109 | raw_spin_unlock(&rnp_root->lock); | |
1110 | return c; | |
1111 | } | |
1112 | ||
1113 | /* | |
1114 | * Clean up any old requests for the just-ended grace period. Also return | |
1115 | * whether any additional grace periods have been requested. Also invoke | |
1116 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1117 | * waiting for this grace period to complete. | |
1118 | */ | |
1119 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1120 | { | |
1121 | int c = rnp->completed; | |
1122 | int needmore; | |
1123 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1124 | ||
1125 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1126 | rnp->need_future_gp[c & 0x1] = 0; | |
1127 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1128 | trace_rcu_future_gp(rnp, rdp, c, |
1129 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1130 | return needmore; |
1131 | } | |
1132 | ||
dc35c893 PM |
1133 | /* |
1134 | * If there is room, assign a ->completed number to any callbacks on | |
1135 | * this CPU that have not already been assigned. Also accelerate any | |
1136 | * callbacks that were previously assigned a ->completed number that has | |
1137 | * since proven to be too conservative, which can happen if callbacks get | |
1138 | * assigned a ->completed number while RCU is idle, but with reference to | |
1139 | * a non-root rcu_node structure. This function is idempotent, so it does | |
1140 | * not hurt to call it repeatedly. | |
1141 | * | |
1142 | * The caller must hold rnp->lock with interrupts disabled. | |
1143 | */ | |
1144 | static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1145 | struct rcu_data *rdp) | |
1146 | { | |
1147 | unsigned long c; | |
1148 | int i; | |
1149 | ||
1150 | /* If the CPU has no callbacks, nothing to do. */ | |
1151 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1152 | return; | |
1153 | ||
1154 | /* | |
1155 | * Starting from the sublist containing the callbacks most | |
1156 | * recently assigned a ->completed number and working down, find the | |
1157 | * first sublist that is not assignable to an upcoming grace period. | |
1158 | * Such a sublist has something in it (first two tests) and has | |
1159 | * a ->completed number assigned that will complete sooner than | |
1160 | * the ->completed number for newly arrived callbacks (last test). | |
1161 | * | |
1162 | * The key point is that any later sublist can be assigned the | |
1163 | * same ->completed number as the newly arrived callbacks, which | |
1164 | * means that the callbacks in any of these later sublist can be | |
1165 | * grouped into a single sublist, whether or not they have already | |
1166 | * been assigned a ->completed number. | |
1167 | */ | |
1168 | c = rcu_cbs_completed(rsp, rnp); | |
1169 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1170 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1171 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1172 | break; | |
1173 | ||
1174 | /* | |
1175 | * If there are no sublist for unassigned callbacks, leave. | |
1176 | * At the same time, advance "i" one sublist, so that "i" will | |
1177 | * index into the sublist where all the remaining callbacks should | |
1178 | * be grouped into. | |
1179 | */ | |
1180 | if (++i >= RCU_NEXT_TAIL) | |
1181 | return; | |
1182 | ||
1183 | /* | |
1184 | * Assign all subsequent callbacks' ->completed number to the next | |
1185 | * full grace period and group them all in the sublist initially | |
1186 | * indexed by "i". | |
1187 | */ | |
1188 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1189 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1190 | rdp->nxtcompleted[i] = c; | |
1191 | } | |
910ee45d PM |
1192 | /* Record any needed additional grace periods. */ |
1193 | rcu_start_future_gp(rnp, rdp); | |
6d4b418c PM |
1194 | |
1195 | /* Trace depending on how much we were able to accelerate. */ | |
1196 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1197 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1198 | else |
f7f7bac9 | 1199 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
dc35c893 PM |
1200 | } |
1201 | ||
1202 | /* | |
1203 | * Move any callbacks whose grace period has completed to the | |
1204 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1205 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1206 | * sublist. This function is idempotent, so it does not hurt to | |
1207 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
1208 | * | |
1209 | * The caller must hold rnp->lock with interrupts disabled. | |
1210 | */ | |
1211 | static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1212 | struct rcu_data *rdp) | |
1213 | { | |
1214 | int i, j; | |
1215 | ||
1216 | /* If the CPU has no callbacks, nothing to do. */ | |
1217 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1218 | return; | |
1219 | ||
1220 | /* | |
1221 | * Find all callbacks whose ->completed numbers indicate that they | |
1222 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1223 | */ | |
1224 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1225 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1226 | break; | |
1227 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1228 | } | |
1229 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1230 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1231 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1232 | ||
1233 | /* Copy down callbacks to fill in empty sublists. */ | |
1234 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1235 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1236 | break; | |
1237 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1238 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1239 | } | |
1240 | ||
1241 | /* Classify any remaining callbacks. */ | |
1242 | rcu_accelerate_cbs(rsp, rnp, rdp); | |
1243 | } | |
1244 | ||
d09b62df | 1245 | /* |
ba9fbe95 PM |
1246 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1247 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1248 | * structure corresponding to the current CPU, and must have irqs disabled. | |
d09b62df | 1249 | */ |
ba9fbe95 | 1250 | static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) |
d09b62df | 1251 | { |
ba9fbe95 | 1252 | /* Handle the ends of any preceding grace periods first. */ |
dc35c893 | 1253 | if (rdp->completed == rnp->completed) { |
d09b62df | 1254 | |
ba9fbe95 | 1255 | /* No grace period end, so just accelerate recent callbacks. */ |
dc35c893 | 1256 | rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1257 | |
dc35c893 PM |
1258 | } else { |
1259 | ||
1260 | /* Advance callbacks. */ | |
1261 | rcu_advance_cbs(rsp, rnp, rdp); | |
d09b62df PM |
1262 | |
1263 | /* Remember that we saw this grace-period completion. */ | |
1264 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1265 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1266 | } |
398ebe60 | 1267 | |
6eaef633 PM |
1268 | if (rdp->gpnum != rnp->gpnum) { |
1269 | /* | |
1270 | * If the current grace period is waiting for this CPU, | |
1271 | * set up to detect a quiescent state, otherwise don't | |
1272 | * go looking for one. | |
1273 | */ | |
1274 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1275 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
6eaef633 PM |
1276 | rdp->passed_quiesce = 0; |
1277 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); | |
1278 | zero_cpu_stall_ticks(rdp); | |
1279 | } | |
1280 | } | |
1281 | ||
d34ea322 | 1282 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1283 | { |
1284 | unsigned long flags; | |
1285 | struct rcu_node *rnp; | |
1286 | ||
1287 | local_irq_save(flags); | |
1288 | rnp = rdp->mynode; | |
d34ea322 PM |
1289 | if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) && |
1290 | rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */ | |
6eaef633 PM |
1291 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
1292 | local_irq_restore(flags); | |
1293 | return; | |
1294 | } | |
d34ea322 | 1295 | __note_gp_changes(rsp, rnp, rdp); |
6eaef633 PM |
1296 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1297 | } | |
1298 | ||
b3dbec76 | 1299 | /* |
f7be8209 | 1300 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1301 | */ |
7fdefc10 | 1302 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 PM |
1303 | { |
1304 | struct rcu_data *rdp; | |
7fdefc10 | 1305 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1306 | |
eb75767b | 1307 | rcu_bind_gp_kthread(); |
7fdefc10 | 1308 | raw_spin_lock_irq(&rnp->lock); |
f7be8209 PM |
1309 | if (rsp->gp_flags == 0) { |
1310 | /* Spurious wakeup, tell caller to go back to sleep. */ | |
1311 | raw_spin_unlock_irq(&rnp->lock); | |
1312 | return 0; | |
1313 | } | |
4cdfc175 | 1314 | rsp->gp_flags = 0; /* Clear all flags: New grace period. */ |
b3dbec76 | 1315 | |
f7be8209 PM |
1316 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1317 | /* | |
1318 | * Grace period already in progress, don't start another. | |
1319 | * Not supposed to be able to happen. | |
1320 | */ | |
7fdefc10 PM |
1321 | raw_spin_unlock_irq(&rnp->lock); |
1322 | return 0; | |
1323 | } | |
1324 | ||
7fdefc10 PM |
1325 | /* Advance to a new grace period and initialize state. */ |
1326 | rsp->gpnum++; | |
f7f7bac9 | 1327 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1328 | record_gp_stall_check_time(rsp); |
1329 | raw_spin_unlock_irq(&rnp->lock); | |
1330 | ||
1331 | /* Exclude any concurrent CPU-hotplug operations. */ | |
a4fbe35a | 1332 | mutex_lock(&rsp->onoff_mutex); |
7fdefc10 PM |
1333 | |
1334 | /* | |
1335 | * Set the quiescent-state-needed bits in all the rcu_node | |
1336 | * structures for all currently online CPUs in breadth-first order, | |
1337 | * starting from the root rcu_node structure, relying on the layout | |
1338 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1339 | * will access only the leaves of the hierarchy, thus seeing that no | |
1340 | * grace period is in progress, at least until the corresponding | |
1341 | * leaf node has been initialized. In addition, we have excluded | |
1342 | * CPU-hotplug operations. | |
1343 | * | |
1344 | * The grace period cannot complete until the initialization | |
1345 | * process finishes, because this kthread handles both. | |
1346 | */ | |
1347 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
b3dbec76 | 1348 | raw_spin_lock_irq(&rnp->lock); |
b3dbec76 | 1349 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1350 | rcu_preempt_check_blocked_tasks(rnp); |
1351 | rnp->qsmask = rnp->qsmaskinit; | |
0446be48 | 1352 | ACCESS_ONCE(rnp->gpnum) = rsp->gpnum; |
25d30cf4 | 1353 | WARN_ON_ONCE(rnp->completed != rsp->completed); |
0446be48 | 1354 | ACCESS_ONCE(rnp->completed) = rsp->completed; |
7fdefc10 | 1355 | if (rnp == rdp->mynode) |
ce3d9c03 | 1356 | __note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1357 | rcu_preempt_boost_start_gp(rnp); |
1358 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1359 | rnp->level, rnp->grplo, | |
1360 | rnp->grphi, rnp->qsmask); | |
1361 | raw_spin_unlock_irq(&rnp->lock); | |
661a85dc | 1362 | #ifdef CONFIG_PROVE_RCU_DELAY |
971394f3 | 1363 | if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 && |
81e59494 | 1364 | system_state == SYSTEM_RUNNING) |
971394f3 | 1365 | udelay(200); |
661a85dc | 1366 | #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */ |
7fdefc10 PM |
1367 | cond_resched(); |
1368 | } | |
b3dbec76 | 1369 | |
a4fbe35a | 1370 | mutex_unlock(&rsp->onoff_mutex); |
7fdefc10 PM |
1371 | return 1; |
1372 | } | |
b3dbec76 | 1373 | |
4cdfc175 PM |
1374 | /* |
1375 | * Do one round of quiescent-state forcing. | |
1376 | */ | |
1377 | int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) | |
1378 | { | |
1379 | int fqs_state = fqs_state_in; | |
217af2a2 PM |
1380 | bool isidle = false; |
1381 | unsigned long maxj; | |
4cdfc175 PM |
1382 | struct rcu_node *rnp = rcu_get_root(rsp); |
1383 | ||
1384 | rsp->n_force_qs++; | |
1385 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1386 | /* Collect dyntick-idle snapshots. */ | |
0edd1b17 PM |
1387 | if (is_sysidle_rcu_state(rsp)) { |
1388 | isidle = 1; | |
1389 | maxj = jiffies - ULONG_MAX / 4; | |
1390 | } | |
217af2a2 PM |
1391 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1392 | &isidle, &maxj); | |
0edd1b17 | 1393 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1394 | fqs_state = RCU_FORCE_QS; |
1395 | } else { | |
1396 | /* Handle dyntick-idle and offline CPUs. */ | |
0edd1b17 | 1397 | isidle = 0; |
217af2a2 | 1398 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1399 | } |
1400 | /* Clear flag to prevent immediate re-entry. */ | |
1401 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1402 | raw_spin_lock_irq(&rnp->lock); | |
1403 | rsp->gp_flags &= ~RCU_GP_FLAG_FQS; | |
1404 | raw_spin_unlock_irq(&rnp->lock); | |
1405 | } | |
1406 | return fqs_state; | |
1407 | } | |
1408 | ||
7fdefc10 PM |
1409 | /* |
1410 | * Clean up after the old grace period. | |
1411 | */ | |
4cdfc175 | 1412 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1413 | { |
1414 | unsigned long gp_duration; | |
dae6e64d | 1415 | int nocb = 0; |
7fdefc10 PM |
1416 | struct rcu_data *rdp; |
1417 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1418 | |
7fdefc10 PM |
1419 | raw_spin_lock_irq(&rnp->lock); |
1420 | gp_duration = jiffies - rsp->gp_start; | |
1421 | if (gp_duration > rsp->gp_max) | |
1422 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1423 | |
7fdefc10 PM |
1424 | /* |
1425 | * We know the grace period is complete, but to everyone else | |
1426 | * it appears to still be ongoing. But it is also the case | |
1427 | * that to everyone else it looks like there is nothing that | |
1428 | * they can do to advance the grace period. It is therefore | |
1429 | * safe for us to drop the lock in order to mark the grace | |
1430 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1431 | */ |
5d4b8659 | 1432 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1433 | |
5d4b8659 PM |
1434 | /* |
1435 | * Propagate new ->completed value to rcu_node structures so | |
1436 | * that other CPUs don't have to wait until the start of the next | |
1437 | * grace period to process their callbacks. This also avoids | |
1438 | * some nasty RCU grace-period initialization races by forcing | |
1439 | * the end of the current grace period to be completely recorded in | |
1440 | * all of the rcu_node structures before the beginning of the next | |
1441 | * grace period is recorded in any of the rcu_node structures. | |
1442 | */ | |
1443 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1444 | raw_spin_lock_irq(&rnp->lock); |
0446be48 | 1445 | ACCESS_ONCE(rnp->completed) = rsp->gpnum; |
b11cc576 PM |
1446 | rdp = this_cpu_ptr(rsp->rda); |
1447 | if (rnp == rdp->mynode) | |
470716fc | 1448 | __note_gp_changes(rsp, rnp, rdp); |
0446be48 | 1449 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 PM |
1450 | raw_spin_unlock_irq(&rnp->lock); |
1451 | cond_resched(); | |
7fdefc10 | 1452 | } |
5d4b8659 PM |
1453 | rnp = rcu_get_root(rsp); |
1454 | raw_spin_lock_irq(&rnp->lock); | |
dae6e64d | 1455 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 PM |
1456 | |
1457 | rsp->completed = rsp->gpnum; /* Declare grace period done. */ | |
f7f7bac9 | 1458 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
7fdefc10 | 1459 | rsp->fqs_state = RCU_GP_IDLE; |
5d4b8659 | 1460 | rdp = this_cpu_ptr(rsp->rda); |
b11cc576 | 1461 | rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */ |
7fdefc10 PM |
1462 | if (cpu_needs_another_gp(rsp, rdp)) |
1463 | rsp->gp_flags = 1; | |
1464 | raw_spin_unlock_irq(&rnp->lock); | |
7fdefc10 PM |
1465 | } |
1466 | ||
1467 | /* | |
1468 | * Body of kthread that handles grace periods. | |
1469 | */ | |
1470 | static int __noreturn rcu_gp_kthread(void *arg) | |
1471 | { | |
4cdfc175 | 1472 | int fqs_state; |
88d6df61 | 1473 | int gf; |
d40011f6 | 1474 | unsigned long j; |
4cdfc175 | 1475 | int ret; |
7fdefc10 PM |
1476 | struct rcu_state *rsp = arg; |
1477 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1478 | ||
1479 | for (;;) { | |
1480 | ||
1481 | /* Handle grace-period start. */ | |
1482 | for (;;) { | |
63c4db78 PM |
1483 | trace_rcu_grace_period(rsp->name, |
1484 | ACCESS_ONCE(rsp->gpnum), | |
1485 | TPS("reqwait")); | |
4cdfc175 | 1486 | wait_event_interruptible(rsp->gp_wq, |
591c6d17 | 1487 | ACCESS_ONCE(rsp->gp_flags) & |
4cdfc175 | 1488 | RCU_GP_FLAG_INIT); |
f7be8209 | 1489 | if (rcu_gp_init(rsp)) |
7fdefc10 PM |
1490 | break; |
1491 | cond_resched(); | |
1492 | flush_signals(current); | |
63c4db78 PM |
1493 | trace_rcu_grace_period(rsp->name, |
1494 | ACCESS_ONCE(rsp->gpnum), | |
1495 | TPS("reqwaitsig")); | |
7fdefc10 | 1496 | } |
cabc49c1 | 1497 | |
4cdfc175 PM |
1498 | /* Handle quiescent-state forcing. */ |
1499 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
1500 | j = jiffies_till_first_fqs; |
1501 | if (j > HZ) { | |
1502 | j = HZ; | |
1503 | jiffies_till_first_fqs = HZ; | |
1504 | } | |
88d6df61 | 1505 | ret = 0; |
cabc49c1 | 1506 | for (;;) { |
88d6df61 PM |
1507 | if (!ret) |
1508 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 PM |
1509 | trace_rcu_grace_period(rsp->name, |
1510 | ACCESS_ONCE(rsp->gpnum), | |
1511 | TPS("fqswait")); | |
4cdfc175 | 1512 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
88d6df61 PM |
1513 | ((gf = ACCESS_ONCE(rsp->gp_flags)) & |
1514 | RCU_GP_FLAG_FQS) || | |
4cdfc175 PM |
1515 | (!ACCESS_ONCE(rnp->qsmask) && |
1516 | !rcu_preempt_blocked_readers_cgp(rnp)), | |
d40011f6 | 1517 | j); |
4cdfc175 | 1518 | /* If grace period done, leave loop. */ |
cabc49c1 | 1519 | if (!ACCESS_ONCE(rnp->qsmask) && |
4cdfc175 | 1520 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 1521 | break; |
4cdfc175 | 1522 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
1523 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
1524 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 PM |
1525 | trace_rcu_grace_period(rsp->name, |
1526 | ACCESS_ONCE(rsp->gpnum), | |
1527 | TPS("fqsstart")); | |
4cdfc175 | 1528 | fqs_state = rcu_gp_fqs(rsp, fqs_state); |
63c4db78 PM |
1529 | trace_rcu_grace_period(rsp->name, |
1530 | ACCESS_ONCE(rsp->gpnum), | |
1531 | TPS("fqsend")); | |
4cdfc175 PM |
1532 | cond_resched(); |
1533 | } else { | |
1534 | /* Deal with stray signal. */ | |
1535 | cond_resched(); | |
1536 | flush_signals(current); | |
63c4db78 PM |
1537 | trace_rcu_grace_period(rsp->name, |
1538 | ACCESS_ONCE(rsp->gpnum), | |
1539 | TPS("fqswaitsig")); | |
4cdfc175 | 1540 | } |
d40011f6 PM |
1541 | j = jiffies_till_next_fqs; |
1542 | if (j > HZ) { | |
1543 | j = HZ; | |
1544 | jiffies_till_next_fqs = HZ; | |
1545 | } else if (j < 1) { | |
1546 | j = 1; | |
1547 | jiffies_till_next_fqs = 1; | |
1548 | } | |
cabc49c1 | 1549 | } |
4cdfc175 PM |
1550 | |
1551 | /* Handle grace-period end. */ | |
1552 | rcu_gp_cleanup(rsp); | |
b3dbec76 | 1553 | } |
b3dbec76 PM |
1554 | } |
1555 | ||
016a8d5b SR |
1556 | static void rsp_wakeup(struct irq_work *work) |
1557 | { | |
1558 | struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work); | |
1559 | ||
1560 | /* Wake up rcu_gp_kthread() to start the grace period. */ | |
1561 | wake_up(&rsp->gp_wq); | |
1562 | } | |
1563 | ||
64db4cff PM |
1564 | /* |
1565 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
1566 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 1567 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
1568 | * |
1569 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
1570 | * invoke this function. This can happen when the dying CPU reports its | |
1571 | * quiescent state. | |
64db4cff PM |
1572 | */ |
1573 | static void | |
910ee45d PM |
1574 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
1575 | struct rcu_data *rdp) | |
64db4cff | 1576 | { |
b8462084 | 1577 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 1578 | /* |
b3dbec76 | 1579 | * Either we have not yet spawned the grace-period |
62da1921 PM |
1580 | * task, this CPU does not need another grace period, |
1581 | * or a grace period is already in progress. | |
b3dbec76 | 1582 | * Either way, don't start a new grace period. |
afe24b12 | 1583 | */ |
afe24b12 PM |
1584 | return; |
1585 | } | |
4cdfc175 | 1586 | rsp->gp_flags = RCU_GP_FLAG_INIT; |
62da1921 | 1587 | |
016a8d5b SR |
1588 | /* |
1589 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c PM |
1590 | * could cause possible deadlocks with the rq->lock. Defer |
1591 | * the wakeup to interrupt context. And don't bother waking | |
1592 | * up the running kthread. | |
016a8d5b | 1593 | */ |
1eafd31c PM |
1594 | if (current != rsp->gp_kthread) |
1595 | irq_work_queue(&rsp->wakeup_work); | |
64db4cff PM |
1596 | } |
1597 | ||
910ee45d PM |
1598 | /* |
1599 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
1600 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
1601 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
1602 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
1603 | * that is encountered beforehand. | |
1604 | */ | |
1605 | static void | |
1606 | rcu_start_gp(struct rcu_state *rsp) | |
1607 | { | |
1608 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1609 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1610 | ||
1611 | /* | |
1612 | * If there is no grace period in progress right now, any | |
1613 | * callbacks we have up to this point will be satisfied by the | |
1614 | * next grace period. Also, advancing the callbacks reduces the | |
1615 | * probability of false positives from cpu_needs_another_gp() | |
1616 | * resulting in pointless grace periods. So, advance callbacks | |
1617 | * then start the grace period! | |
1618 | */ | |
1619 | rcu_advance_cbs(rsp, rnp, rdp); | |
1620 | rcu_start_gp_advanced(rsp, rnp, rdp); | |
1621 | } | |
1622 | ||
f41d911f | 1623 | /* |
d3f6bad3 PM |
1624 | * Report a full set of quiescent states to the specified rcu_state |
1625 | * data structure. This involves cleaning up after the prior grace | |
1626 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
1627 | * if one is needed. Note that the caller must hold rnp->lock, which |
1628 | * is released before return. | |
f41d911f | 1629 | */ |
d3f6bad3 | 1630 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 1631 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 1632 | { |
fc2219d4 | 1633 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cabc49c1 PM |
1634 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
1635 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ | |
f41d911f PM |
1636 | } |
1637 | ||
64db4cff | 1638 | /* |
d3f6bad3 PM |
1639 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
1640 | * Allows quiescent states for a group of CPUs to be reported at one go | |
1641 | * to the specified rcu_node structure, though all the CPUs in the group | |
1642 | * must be represented by the same rcu_node structure (which need not be | |
1643 | * a leaf rcu_node structure, though it often will be). That structure's | |
1644 | * lock must be held upon entry, and it is released before return. | |
64db4cff PM |
1645 | */ |
1646 | static void | |
d3f6bad3 PM |
1647 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
1648 | struct rcu_node *rnp, unsigned long flags) | |
64db4cff PM |
1649 | __releases(rnp->lock) |
1650 | { | |
28ecd580 PM |
1651 | struct rcu_node *rnp_c; |
1652 | ||
64db4cff PM |
1653 | /* Walk up the rcu_node hierarchy. */ |
1654 | for (;;) { | |
1655 | if (!(rnp->qsmask & mask)) { | |
1656 | ||
1657 | /* Our bit has already been cleared, so done. */ | |
1304afb2 | 1658 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1659 | return; |
1660 | } | |
1661 | rnp->qsmask &= ~mask; | |
d4c08f2a PM |
1662 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
1663 | mask, rnp->qsmask, rnp->level, | |
1664 | rnp->grplo, rnp->grphi, | |
1665 | !!rnp->gp_tasks); | |
27f4d280 | 1666 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
1667 | |
1668 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 1669 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1670 | return; |
1671 | } | |
1672 | mask = rnp->grpmask; | |
1673 | if (rnp->parent == NULL) { | |
1674 | ||
1675 | /* No more levels. Exit loop holding root lock. */ | |
1676 | ||
1677 | break; | |
1678 | } | |
1304afb2 | 1679 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 1680 | rnp_c = rnp; |
64db4cff | 1681 | rnp = rnp->parent; |
1304afb2 | 1682 | raw_spin_lock_irqsave(&rnp->lock, flags); |
28ecd580 | 1683 | WARN_ON_ONCE(rnp_c->qsmask); |
64db4cff PM |
1684 | } |
1685 | ||
1686 | /* | |
1687 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 1688 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 1689 | * to clean up and start the next grace period if one is needed. |
64db4cff | 1690 | */ |
d3f6bad3 | 1691 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
1692 | } |
1693 | ||
1694 | /* | |
d3f6bad3 PM |
1695 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
1696 | * structure. This must be either called from the specified CPU, or | |
1697 | * called when the specified CPU is known to be offline (and when it is | |
1698 | * also known that no other CPU is concurrently trying to help the offline | |
1699 | * CPU). The lastcomp argument is used to make sure we are still in the | |
1700 | * grace period of interest. We don't want to end the current grace period | |
1701 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
1702 | */ |
1703 | static void | |
d7d6a11e | 1704 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1705 | { |
1706 | unsigned long flags; | |
1707 | unsigned long mask; | |
1708 | struct rcu_node *rnp; | |
1709 | ||
1710 | rnp = rdp->mynode; | |
1304afb2 | 1711 | raw_spin_lock_irqsave(&rnp->lock, flags); |
d7d6a11e PM |
1712 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum || |
1713 | rnp->completed == rnp->gpnum) { | |
64db4cff PM |
1714 | |
1715 | /* | |
e4cc1f22 PM |
1716 | * The grace period in which this quiescent state was |
1717 | * recorded has ended, so don't report it upwards. | |
1718 | * We will instead need a new quiescent state that lies | |
1719 | * within the current grace period. | |
64db4cff | 1720 | */ |
e4cc1f22 | 1721 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
1304afb2 | 1722 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1723 | return; |
1724 | } | |
1725 | mask = rdp->grpmask; | |
1726 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 1727 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1728 | } else { |
1729 | rdp->qs_pending = 0; | |
1730 | ||
1731 | /* | |
1732 | * This GP can't end until cpu checks in, so all of our | |
1733 | * callbacks can be processed during the next GP. | |
1734 | */ | |
dc35c893 | 1735 | rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 1736 | |
d3f6bad3 | 1737 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
64db4cff PM |
1738 | } |
1739 | } | |
1740 | ||
1741 | /* | |
1742 | * Check to see if there is a new grace period of which this CPU | |
1743 | * is not yet aware, and if so, set up local rcu_data state for it. | |
1744 | * Otherwise, see if this CPU has just passed through its first | |
1745 | * quiescent state for this grace period, and record that fact if so. | |
1746 | */ | |
1747 | static void | |
1748 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
1749 | { | |
05eb552b PM |
1750 | /* Check for grace-period ends and beginnings. */ |
1751 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
1752 | |
1753 | /* | |
1754 | * Does this CPU still need to do its part for current grace period? | |
1755 | * If no, return and let the other CPUs do their part as well. | |
1756 | */ | |
1757 | if (!rdp->qs_pending) | |
1758 | return; | |
1759 | ||
1760 | /* | |
1761 | * Was there a quiescent state since the beginning of the grace | |
1762 | * period? If no, then exit and wait for the next call. | |
1763 | */ | |
e4cc1f22 | 1764 | if (!rdp->passed_quiesce) |
64db4cff PM |
1765 | return; |
1766 | ||
d3f6bad3 PM |
1767 | /* |
1768 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
1769 | * judge of that). | |
1770 | */ | |
d7d6a11e | 1771 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
1772 | } |
1773 | ||
1774 | #ifdef CONFIG_HOTPLUG_CPU | |
1775 | ||
e74f4c45 | 1776 | /* |
b1420f1c PM |
1777 | * Send the specified CPU's RCU callbacks to the orphanage. The |
1778 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 1779 | * ->orphan_lock. |
e74f4c45 | 1780 | */ |
b1420f1c PM |
1781 | static void |
1782 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
1783 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 1784 | { |
3fbfbf7a | 1785 | /* No-CBs CPUs do not have orphanable callbacks. */ |
d1e43fa5 | 1786 | if (rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
1787 | return; |
1788 | ||
b1420f1c PM |
1789 | /* |
1790 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
1791 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
1792 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 1793 | */ |
a50c3af9 | 1794 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1795 | rsp->qlen_lazy += rdp->qlen_lazy; |
1796 | rsp->qlen += rdp->qlen; | |
1797 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 1798 | rdp->qlen_lazy = 0; |
1d1fb395 | 1799 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
1800 | } |
1801 | ||
1802 | /* | |
b1420f1c PM |
1803 | * Next, move those callbacks still needing a grace period to |
1804 | * the orphanage, where some other CPU will pick them up. | |
1805 | * Some of the callbacks might have gone partway through a grace | |
1806 | * period, but that is too bad. They get to start over because we | |
1807 | * cannot assume that grace periods are synchronized across CPUs. | |
1808 | * We don't bother updating the ->nxttail[] array yet, instead | |
1809 | * we just reset the whole thing later on. | |
a50c3af9 | 1810 | */ |
b1420f1c PM |
1811 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
1812 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1813 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
1814 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
1815 | } |
1816 | ||
1817 | /* | |
b1420f1c PM |
1818 | * Then move the ready-to-invoke callbacks to the orphanage, |
1819 | * where some other CPU will pick them up. These will not be | |
1820 | * required to pass though another grace period: They are done. | |
a50c3af9 | 1821 | */ |
e5601400 | 1822 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1823 | *rsp->orphan_donetail = rdp->nxtlist; |
1824 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 1825 | } |
e74f4c45 | 1826 | |
b1420f1c | 1827 | /* Finally, initialize the rcu_data structure's list to empty. */ |
3f5d3ea6 | 1828 | init_callback_list(rdp); |
b1420f1c PM |
1829 | } |
1830 | ||
1831 | /* | |
1832 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 1833 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c PM |
1834 | */ |
1835 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp) | |
1836 | { | |
1837 | int i; | |
1838 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); | |
1839 | ||
3fbfbf7a PM |
1840 | /* No-CBs CPUs are handled specially. */ |
1841 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp)) | |
1842 | return; | |
1843 | ||
b1420f1c PM |
1844 | /* Do the accounting first. */ |
1845 | rdp->qlen_lazy += rsp->qlen_lazy; | |
1846 | rdp->qlen += rsp->qlen; | |
1847 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
1848 | if (rsp->qlen_lazy != rsp->qlen) |
1849 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
1850 | rsp->qlen_lazy = 0; |
1851 | rsp->qlen = 0; | |
1852 | ||
1853 | /* | |
1854 | * We do not need a memory barrier here because the only way we | |
1855 | * can get here if there is an rcu_barrier() in flight is if | |
1856 | * we are the task doing the rcu_barrier(). | |
1857 | */ | |
1858 | ||
1859 | /* First adopt the ready-to-invoke callbacks. */ | |
1860 | if (rsp->orphan_donelist != NULL) { | |
1861 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1862 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
1863 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
1864 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
1865 | rdp->nxttail[i] = rsp->orphan_donetail; | |
1866 | rsp->orphan_donelist = NULL; | |
1867 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
1868 | } | |
1869 | ||
1870 | /* And then adopt the callbacks that still need a grace period. */ | |
1871 | if (rsp->orphan_nxtlist != NULL) { | |
1872 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
1873 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
1874 | rsp->orphan_nxtlist = NULL; | |
1875 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
1876 | } | |
1877 | } | |
1878 | ||
1879 | /* | |
1880 | * Trace the fact that this CPU is going offline. | |
1881 | */ | |
1882 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
1883 | { | |
1884 | RCU_TRACE(unsigned long mask); | |
1885 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
1886 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
1887 | ||
1888 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
1889 | trace_rcu_grace_period(rsp->name, |
1890 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 1891 | TPS("cpuofl")); |
64db4cff PM |
1892 | } |
1893 | ||
1894 | /* | |
e5601400 | 1895 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
1896 | * this fact from process context. Do the remainder of the cleanup, |
1897 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
1898 | * adopting them. There can only be one CPU hotplug operation at a time, |
1899 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 1900 | */ |
e5601400 | 1901 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 1902 | { |
2036d94a PM |
1903 | unsigned long flags; |
1904 | unsigned long mask; | |
1905 | int need_report = 0; | |
e5601400 | 1906 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 1907 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 1908 | |
2036d94a | 1909 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 1910 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 1911 | |
b1420f1c | 1912 | /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */ |
2036d94a PM |
1913 | |
1914 | /* Exclude any attempts to start a new grace period. */ | |
a4fbe35a | 1915 | mutex_lock(&rsp->onoff_mutex); |
7b2e6011 | 1916 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
2036d94a | 1917 | |
b1420f1c PM |
1918 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
1919 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
1920 | rcu_adopt_orphan_cbs(rsp); | |
1921 | ||
2036d94a PM |
1922 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
1923 | mask = rdp->grpmask; /* rnp->grplo is constant. */ | |
1924 | do { | |
1925 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
1926 | rnp->qsmaskinit &= ~mask; | |
1927 | if (rnp->qsmaskinit != 0) { | |
1928 | if (rnp != rdp->mynode) | |
1929 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1930 | break; | |
1931 | } | |
1932 | if (rnp == rdp->mynode) | |
1933 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); | |
1934 | else | |
1935 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1936 | mask = rnp->grpmask; | |
1937 | rnp = rnp->parent; | |
1938 | } while (rnp != NULL); | |
1939 | ||
1940 | /* | |
1941 | * We still hold the leaf rcu_node structure lock here, and | |
1942 | * irqs are still disabled. The reason for this subterfuge is | |
7b2e6011 | 1943 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock |
2036d94a PM |
1944 | * held leads to deadlock. |
1945 | */ | |
7b2e6011 | 1946 | raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */ |
2036d94a PM |
1947 | rnp = rdp->mynode; |
1948 | if (need_report & RCU_OFL_TASKS_NORM_GP) | |
1949 | rcu_report_unblock_qs_rnp(rnp, flags); | |
1950 | else | |
1951 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1952 | if (need_report & RCU_OFL_TASKS_EXP_GP) | |
1953 | rcu_report_exp_rnp(rsp, rnp, true); | |
cf01537e PM |
1954 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
1955 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
1956 | cpu, rdp->qlen, rdp->nxtlist); | |
0d8ee37e PM |
1957 | init_callback_list(rdp); |
1958 | /* Disallow further callbacks on this CPU. */ | |
1959 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; | |
a4fbe35a | 1960 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
1961 | } |
1962 | ||
1963 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1964 | ||
e5601400 | 1965 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
1966 | { |
1967 | } | |
1968 | ||
e5601400 | 1969 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
1970 | { |
1971 | } | |
1972 | ||
1973 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
1974 | ||
1975 | /* | |
1976 | * Invoke any RCU callbacks that have made it to the end of their grace | |
1977 | * period. Thottle as specified by rdp->blimit. | |
1978 | */ | |
37c72e56 | 1979 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1980 | { |
1981 | unsigned long flags; | |
1982 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
1983 | long bl, count, count_lazy; |
1984 | int i; | |
64db4cff | 1985 | |
dc35c893 | 1986 | /* If no callbacks are ready, just return. */ |
29c00b4a | 1987 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 1988 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
1989 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
1990 | need_resched(), is_idle_task(current), | |
1991 | rcu_is_callbacks_kthread()); | |
64db4cff | 1992 | return; |
29c00b4a | 1993 | } |
64db4cff PM |
1994 | |
1995 | /* | |
1996 | * Extract the list of ready callbacks, disabling to prevent | |
1997 | * races with call_rcu() from interrupt handlers. | |
1998 | */ | |
1999 | local_irq_save(flags); | |
8146c4e2 | 2000 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2001 | bl = rdp->blimit; |
486e2593 | 2002 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2003 | list = rdp->nxtlist; |
2004 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2005 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2006 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2007 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2008 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2009 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2010 | local_irq_restore(flags); |
2011 | ||
2012 | /* Invoke callbacks. */ | |
486e2593 | 2013 | count = count_lazy = 0; |
64db4cff PM |
2014 | while (list) { |
2015 | next = list->next; | |
2016 | prefetch(next); | |
551d55a9 | 2017 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2018 | if (__rcu_reclaim(rsp->name, list)) |
2019 | count_lazy++; | |
64db4cff | 2020 | list = next; |
dff1672d PM |
2021 | /* Stop only if limit reached and CPU has something to do. */ |
2022 | if (++count >= bl && | |
2023 | (need_resched() || | |
2024 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2025 | break; |
2026 | } | |
2027 | ||
2028 | local_irq_save(flags); | |
4968c300 PM |
2029 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2030 | is_idle_task(current), | |
2031 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2032 | |
2033 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2034 | if (list != NULL) { |
2035 | *tail = rdp->nxtlist; | |
2036 | rdp->nxtlist = list; | |
b41772ab PM |
2037 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2038 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2039 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2040 | else |
2041 | break; | |
2042 | } | |
b1420f1c PM |
2043 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2044 | rdp->qlen_lazy -= count_lazy; | |
1d1fb395 | 2045 | ACCESS_ONCE(rdp->qlen) -= count; |
b1420f1c | 2046 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2047 | |
2048 | /* Reinstate batch limit if we have worked down the excess. */ | |
2049 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2050 | rdp->blimit = blimit; | |
2051 | ||
37c72e56 PM |
2052 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2053 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2054 | rdp->qlen_last_fqs_check = 0; | |
2055 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2056 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2057 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2058 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2059 | |
64db4cff PM |
2060 | local_irq_restore(flags); |
2061 | ||
e0f23060 | 2062 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2063 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2064 | invoke_rcu_core(); |
64db4cff PM |
2065 | } |
2066 | ||
2067 | /* | |
2068 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2069 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2070 | * Also schedule RCU core processing. |
64db4cff | 2071 | * |
9b2e4f18 | 2072 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2073 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2074 | * false, there is no point in invoking rcu_check_callbacks(). | |
2075 | */ | |
2076 | void rcu_check_callbacks(int cpu, int user) | |
2077 | { | |
f7f7bac9 | 2078 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2079 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2080 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2081 | |
2082 | /* | |
2083 | * Get here if this CPU took its interrupt from user | |
2084 | * mode or from the idle loop, and if this is not a | |
2085 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2086 | * a quiescent state, so note it. |
64db4cff PM |
2087 | * |
2088 | * No memory barrier is required here because both | |
d6714c22 PM |
2089 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2090 | * variables that other CPUs neither access nor modify, | |
2091 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2092 | */ |
2093 | ||
d6714c22 PM |
2094 | rcu_sched_qs(cpu); |
2095 | rcu_bh_qs(cpu); | |
64db4cff PM |
2096 | |
2097 | } else if (!in_softirq()) { | |
2098 | ||
2099 | /* | |
2100 | * Get here if this CPU did not take its interrupt from | |
2101 | * softirq, in other words, if it is not interrupting | |
2102 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2103 | * critical section, so note it. |
64db4cff PM |
2104 | */ |
2105 | ||
d6714c22 | 2106 | rcu_bh_qs(cpu); |
64db4cff | 2107 | } |
f41d911f | 2108 | rcu_preempt_check_callbacks(cpu); |
d21670ac | 2109 | if (rcu_pending(cpu)) |
a46e0899 | 2110 | invoke_rcu_core(); |
f7f7bac9 | 2111 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2112 | } |
2113 | ||
64db4cff PM |
2114 | /* |
2115 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2116 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2117 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2118 | * | |
ee47eb9f | 2119 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2120 | */ |
217af2a2 PM |
2121 | static void force_qs_rnp(struct rcu_state *rsp, |
2122 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2123 | unsigned long *maxj), | |
2124 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2125 | { |
2126 | unsigned long bit; | |
2127 | int cpu; | |
2128 | unsigned long flags; | |
2129 | unsigned long mask; | |
a0b6c9a7 | 2130 | struct rcu_node *rnp; |
64db4cff | 2131 | |
a0b6c9a7 | 2132 | rcu_for_each_leaf_node(rsp, rnp) { |
b4be093f | 2133 | cond_resched(); |
64db4cff | 2134 | mask = 0; |
1304afb2 | 2135 | raw_spin_lock_irqsave(&rnp->lock, flags); |
ee47eb9f | 2136 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 2137 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 2138 | return; |
64db4cff | 2139 | } |
a0b6c9a7 | 2140 | if (rnp->qsmask == 0) { |
1217ed1b | 2141 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
64db4cff PM |
2142 | continue; |
2143 | } | |
a0b6c9a7 | 2144 | cpu = rnp->grplo; |
64db4cff | 2145 | bit = 1; |
a0b6c9a7 | 2146 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 PM |
2147 | if ((rnp->qsmask & bit) != 0) { |
2148 | if ((rnp->qsmaskinit & bit) != 0) | |
2149 | *isidle = 0; | |
2150 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) | |
2151 | mask |= bit; | |
2152 | } | |
64db4cff | 2153 | } |
45f014c5 | 2154 | if (mask != 0) { |
64db4cff | 2155 | |
d3f6bad3 PM |
2156 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
2157 | rcu_report_qs_rnp(mask, rsp, rnp, flags); | |
64db4cff PM |
2158 | continue; |
2159 | } | |
1304afb2 | 2160 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2161 | } |
27f4d280 | 2162 | rnp = rcu_get_root(rsp); |
1217ed1b PM |
2163 | if (rnp->qsmask == 0) { |
2164 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2165 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ | |
2166 | } | |
64db4cff PM |
2167 | } |
2168 | ||
2169 | /* | |
2170 | * Force quiescent states on reluctant CPUs, and also detect which | |
2171 | * CPUs are in dyntick-idle mode. | |
2172 | */ | |
4cdfc175 | 2173 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2174 | { |
2175 | unsigned long flags; | |
394f2769 PM |
2176 | bool ret; |
2177 | struct rcu_node *rnp; | |
2178 | struct rcu_node *rnp_old = NULL; | |
2179 | ||
2180 | /* Funnel through hierarchy to reduce memory contention. */ | |
2181 | rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode; | |
2182 | for (; rnp != NULL; rnp = rnp->parent) { | |
2183 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2184 | !raw_spin_trylock(&rnp->fqslock); | |
2185 | if (rnp_old != NULL) | |
2186 | raw_spin_unlock(&rnp_old->fqslock); | |
2187 | if (ret) { | |
2188 | rsp->n_force_qs_lh++; | |
2189 | return; | |
2190 | } | |
2191 | rnp_old = rnp; | |
2192 | } | |
2193 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2194 | |
394f2769 PM |
2195 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2196 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
2197 | raw_spin_unlock(&rnp_old->fqslock); | |
2198 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
2199 | rsp->n_force_qs_lh++; | |
2200 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); | |
4cdfc175 | 2201 | return; /* Someone beat us to it. */ |
46a1e34e | 2202 | } |
4cdfc175 | 2203 | rsp->gp_flags |= RCU_GP_FLAG_FQS; |
394f2769 | 2204 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2205 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ |
64db4cff PM |
2206 | } |
2207 | ||
64db4cff | 2208 | /* |
e0f23060 PM |
2209 | * This does the RCU core processing work for the specified rcu_state |
2210 | * and rcu_data structures. This may be called only from the CPU to | |
2211 | * whom the rdp belongs. | |
64db4cff PM |
2212 | */ |
2213 | static void | |
1bca8cf1 | 2214 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2215 | { |
2216 | unsigned long flags; | |
1bca8cf1 | 2217 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
64db4cff | 2218 | |
2e597558 PM |
2219 | WARN_ON_ONCE(rdp->beenonline == 0); |
2220 | ||
64db4cff PM |
2221 | /* Update RCU state based on any recent quiescent states. */ |
2222 | rcu_check_quiescent_state(rsp, rdp); | |
2223 | ||
2224 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2225 | local_irq_save(flags); |
64db4cff | 2226 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2227 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
b8462084 PM |
2228 | rcu_start_gp(rsp); |
2229 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); | |
dc35c893 PM |
2230 | } else { |
2231 | local_irq_restore(flags); | |
64db4cff PM |
2232 | } |
2233 | ||
2234 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2235 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2236 | invoke_rcu_callbacks(rsp, rdp); |
09223371 SL |
2237 | } |
2238 | ||
64db4cff | 2239 | /* |
e0f23060 | 2240 | * Do RCU core processing for the current CPU. |
64db4cff | 2241 | */ |
09223371 | 2242 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2243 | { |
6ce75a23 PM |
2244 | struct rcu_state *rsp; |
2245 | ||
bfa00b4c PM |
2246 | if (cpu_is_offline(smp_processor_id())) |
2247 | return; | |
f7f7bac9 | 2248 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2249 | for_each_rcu_flavor(rsp) |
2250 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2251 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2252 | } |
2253 | ||
a26ac245 | 2254 | /* |
e0f23060 PM |
2255 | * Schedule RCU callback invocation. If the specified type of RCU |
2256 | * does not support RCU priority boosting, just do a direct call, | |
2257 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
2258 | * are running on the current CPU with interrupts disabled, the | |
2259 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
a26ac245 | 2260 | */ |
a46e0899 | 2261 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2262 | { |
b0d30417 PM |
2263 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
2264 | return; | |
a46e0899 PM |
2265 | if (likely(!rsp->boost)) { |
2266 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2267 | return; |
2268 | } | |
a46e0899 | 2269 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2270 | } |
2271 | ||
a46e0899 | 2272 | static void invoke_rcu_core(void) |
09223371 | 2273 | { |
b0f74036 PM |
2274 | if (cpu_online(smp_processor_id())) |
2275 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2276 | } |
2277 | ||
29154c57 PM |
2278 | /* |
2279 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2280 | */ | |
2281 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2282 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2283 | { |
62fde6ed PM |
2284 | /* |
2285 | * If called from an extended quiescent state, invoke the RCU | |
2286 | * core in order to force a re-evaluation of RCU's idleness. | |
2287 | */ | |
a16b7a69 | 2288 | if (rcu_is_cpu_idle() && cpu_online(smp_processor_id())) |
62fde6ed PM |
2289 | invoke_rcu_core(); |
2290 | ||
a16b7a69 | 2291 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2292 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2293 | return; |
64db4cff | 2294 | |
37c72e56 PM |
2295 | /* |
2296 | * Force the grace period if too many callbacks or too long waiting. | |
2297 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2298 | * if some other CPU has recently done so. Also, don't bother | |
2299 | * invoking force_quiescent_state() if the newly enqueued callback | |
2300 | * is the only one waiting for a grace period to complete. | |
2301 | */ | |
2655d57e | 2302 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2303 | |
2304 | /* Are we ignoring a completed grace period? */ | |
470716fc | 2305 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
2306 | |
2307 | /* Start a new grace period if one not already started. */ | |
2308 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
2309 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
2310 | ||
b8462084 PM |
2311 | raw_spin_lock(&rnp_root->lock); |
2312 | rcu_start_gp(rsp); | |
2313 | raw_spin_unlock(&rnp_root->lock); | |
b52573d2 PM |
2314 | } else { |
2315 | /* Give the grace period a kick. */ | |
2316 | rdp->blimit = LONG_MAX; | |
2317 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2318 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2319 | force_quiescent_state(rsp); |
b52573d2 PM |
2320 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2321 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2322 | } | |
4cdfc175 | 2323 | } |
29154c57 PM |
2324 | } |
2325 | ||
ae150184 PM |
2326 | /* |
2327 | * RCU callback function to leak a callback. | |
2328 | */ | |
2329 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2330 | { | |
2331 | } | |
2332 | ||
3fbfbf7a PM |
2333 | /* |
2334 | * Helper function for call_rcu() and friends. The cpu argument will | |
2335 | * normally be -1, indicating "currently running CPU". It may specify | |
2336 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2337 | * is expected to specify a CPU. | |
2338 | */ | |
64db4cff PM |
2339 | static void |
2340 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 2341 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
2342 | { |
2343 | unsigned long flags; | |
2344 | struct rcu_data *rdp; | |
2345 | ||
0bb7b59d | 2346 | WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */ |
ae150184 PM |
2347 | if (debug_rcu_head_queue(head)) { |
2348 | /* Probable double call_rcu(), so leak the callback. */ | |
2349 | ACCESS_ONCE(head->func) = rcu_leak_callback; | |
2350 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); | |
2351 | return; | |
2352 | } | |
64db4cff PM |
2353 | head->func = func; |
2354 | head->next = NULL; | |
2355 | ||
64db4cff PM |
2356 | /* |
2357 | * Opportunistically note grace-period endings and beginnings. | |
2358 | * Note that we might see a beginning right after we see an | |
2359 | * end, but never vice versa, since this CPU has to pass through | |
2360 | * a quiescent state betweentimes. | |
2361 | */ | |
2362 | local_irq_save(flags); | |
394f99a9 | 2363 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
2364 | |
2365 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
2366 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
2367 | int offline; | |
2368 | ||
2369 | if (cpu != -1) | |
2370 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
2371 | offline = !__call_rcu_nocb(rdp, head, lazy); | |
2372 | WARN_ON_ONCE(offline); | |
0d8ee37e | 2373 | /* _call_rcu() is illegal on offline CPU; leak the callback. */ |
0d8ee37e PM |
2374 | local_irq_restore(flags); |
2375 | return; | |
2376 | } | |
29154c57 | 2377 | ACCESS_ONCE(rdp->qlen)++; |
486e2593 PM |
2378 | if (lazy) |
2379 | rdp->qlen_lazy++; | |
c57afe80 PM |
2380 | else |
2381 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2382 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
2383 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
2384 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 2385 | |
d4c08f2a PM |
2386 | if (__is_kfree_rcu_offset((unsigned long)func)) |
2387 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 2388 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2389 | else |
486e2593 | 2390 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2391 | |
29154c57 PM |
2392 | /* Go handle any RCU core processing required. */ |
2393 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
2394 | local_irq_restore(flags); |
2395 | } | |
2396 | ||
2397 | /* | |
d6714c22 | 2398 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 2399 | */ |
d6714c22 | 2400 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 2401 | { |
3fbfbf7a | 2402 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 2403 | } |
d6714c22 | 2404 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
2405 | |
2406 | /* | |
486e2593 | 2407 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
2408 | */ |
2409 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
2410 | { | |
3fbfbf7a | 2411 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
2412 | } |
2413 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
2414 | ||
6d813391 PM |
2415 | /* |
2416 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
2417 | * any blocking grace-period wait automatically implies a grace period | |
2418 | * if there is only one CPU online at any point time during execution | |
2419 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
2420 | * occasionally incorrectly indicate that there are multiple CPUs online | |
2421 | * when there was in fact only one the whole time, as this just adds | |
2422 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
2423 | */ |
2424 | static inline int rcu_blocking_is_gp(void) | |
2425 | { | |
95f0c1de PM |
2426 | int ret; |
2427 | ||
6d813391 | 2428 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
2429 | preempt_disable(); |
2430 | ret = num_online_cpus() <= 1; | |
2431 | preempt_enable(); | |
2432 | return ret; | |
6d813391 PM |
2433 | } |
2434 | ||
6ebb237b PM |
2435 | /** |
2436 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
2437 | * | |
2438 | * Control will return to the caller some time after a full rcu-sched | |
2439 | * grace period has elapsed, in other words after all currently executing | |
2440 | * rcu-sched read-side critical sections have completed. These read-side | |
2441 | * critical sections are delimited by rcu_read_lock_sched() and | |
2442 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
2443 | * local_irq_disable(), and so on may be used in place of | |
2444 | * rcu_read_lock_sched(). | |
2445 | * | |
2446 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
2447 | * non-threaded hardware-interrupt handlers, in progress on entry will |
2448 | * have completed before this primitive returns. However, this does not | |
2449 | * guarantee that softirq handlers will have completed, since in some | |
2450 | * kernels, these handlers can run in process context, and can block. | |
2451 | * | |
2452 | * Note that this guarantee implies further memory-ordering guarantees. | |
2453 | * On systems with more than one CPU, when synchronize_sched() returns, | |
2454 | * each CPU is guaranteed to have executed a full memory barrier since the | |
2455 | * end of its last RCU-sched read-side critical section whose beginning | |
2456 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
2457 | * an RCU read-side critical section that extends beyond the return from | |
2458 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
2459 | * after the beginning of synchronize_sched() and before the beginning of | |
2460 | * that RCU read-side critical section. Note that these guarantees include | |
2461 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
2462 | * that are executing in the kernel. | |
2463 | * | |
2464 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
2465 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
2466 | * to have executed a full memory barrier during the execution of | |
2467 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
2468 | * again only if the system has more than one CPU). | |
6ebb237b PM |
2469 | * |
2470 | * This primitive provides the guarantees made by the (now removed) | |
2471 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
2472 | * guarantees that rcu_read_lock() sections will have completed. | |
2473 | * In "classic RCU", these two guarantees happen to be one and | |
2474 | * the same, but can differ in realtime RCU implementations. | |
2475 | */ | |
2476 | void synchronize_sched(void) | |
2477 | { | |
fe15d706 PM |
2478 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2479 | !lock_is_held(&rcu_lock_map) && | |
2480 | !lock_is_held(&rcu_sched_lock_map), | |
2481 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
2482 | if (rcu_blocking_is_gp()) |
2483 | return; | |
3705b88d AM |
2484 | if (rcu_expedited) |
2485 | synchronize_sched_expedited(); | |
2486 | else | |
2487 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
2488 | } |
2489 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
2490 | ||
2491 | /** | |
2492 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
2493 | * | |
2494 | * Control will return to the caller some time after a full rcu_bh grace | |
2495 | * period has elapsed, in other words after all currently executing rcu_bh | |
2496 | * read-side critical sections have completed. RCU read-side critical | |
2497 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
2498 | * and may be nested. | |
f0a0e6f2 PM |
2499 | * |
2500 | * See the description of synchronize_sched() for more detailed information | |
2501 | * on memory ordering guarantees. | |
6ebb237b PM |
2502 | */ |
2503 | void synchronize_rcu_bh(void) | |
2504 | { | |
fe15d706 PM |
2505 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2506 | !lock_is_held(&rcu_lock_map) && | |
2507 | !lock_is_held(&rcu_sched_lock_map), | |
2508 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
2509 | if (rcu_blocking_is_gp()) |
2510 | return; | |
3705b88d AM |
2511 | if (rcu_expedited) |
2512 | synchronize_rcu_bh_expedited(); | |
2513 | else | |
2514 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
2515 | } |
2516 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
2517 | ||
3d3b7db0 PM |
2518 | static int synchronize_sched_expedited_cpu_stop(void *data) |
2519 | { | |
2520 | /* | |
2521 | * There must be a full memory barrier on each affected CPU | |
2522 | * between the time that try_stop_cpus() is called and the | |
2523 | * time that it returns. | |
2524 | * | |
2525 | * In the current initial implementation of cpu_stop, the | |
2526 | * above condition is already met when the control reaches | |
2527 | * this point and the following smp_mb() is not strictly | |
2528 | * necessary. Do smp_mb() anyway for documentation and | |
2529 | * robustness against future implementation changes. | |
2530 | */ | |
2531 | smp_mb(); /* See above comment block. */ | |
2532 | return 0; | |
2533 | } | |
2534 | ||
236fefaf PM |
2535 | /** |
2536 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
2537 | * | |
2538 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
2539 | * approach to force the grace period to end quickly. This consumes | |
2540 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
2541 | * so is thus not recommended for any sort of common-case code. In fact, | |
2542 | * if you are using synchronize_sched_expedited() in a loop, please | |
2543 | * restructure your code to batch your updates, and then use a single | |
2544 | * synchronize_sched() instead. | |
3d3b7db0 | 2545 | * |
236fefaf PM |
2546 | * Note that it is illegal to call this function while holding any lock |
2547 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal | |
2548 | * to call this function from a CPU-hotplug notifier. Failing to observe | |
2549 | * these restriction will result in deadlock. | |
3d3b7db0 PM |
2550 | * |
2551 | * This implementation can be thought of as an application of ticket | |
2552 | * locking to RCU, with sync_sched_expedited_started and | |
2553 | * sync_sched_expedited_done taking on the roles of the halves | |
2554 | * of the ticket-lock word. Each task atomically increments | |
2555 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
2556 | * then attempts to stop all the CPUs. If this succeeds, then each | |
2557 | * CPU will have executed a context switch, resulting in an RCU-sched | |
2558 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
2559 | * update sync_sched_expedited_done to match our snapshot -- but | |
2560 | * only if someone else has not already advanced past our snapshot. | |
2561 | * | |
2562 | * On the other hand, if try_stop_cpus() fails, we check the value | |
2563 | * of sync_sched_expedited_done. If it has advanced past our | |
2564 | * initial snapshot, then someone else must have forced a grace period | |
2565 | * some time after we took our snapshot. In this case, our work is | |
2566 | * done for us, and we can simply return. Otherwise, we try again, | |
2567 | * but keep our initial snapshot for purposes of checking for someone | |
2568 | * doing our work for us. | |
2569 | * | |
2570 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
2571 | */ | |
2572 | void synchronize_sched_expedited(void) | |
2573 | { | |
1924bcb0 PM |
2574 | long firstsnap, s, snap; |
2575 | int trycount = 0; | |
40694d66 | 2576 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 2577 | |
1924bcb0 PM |
2578 | /* |
2579 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
2580 | * By allowing sync_sched_expedited_started to advance no more than | |
2581 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
2582 | * that more than 3.5 billion CPUs would be required to force a | |
2583 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
2584 | * course be required on a 64-bit system. | |
2585 | */ | |
40694d66 PM |
2586 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
2587 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 PM |
2588 | ULONG_MAX / 8)) { |
2589 | synchronize_sched(); | |
a30489c5 | 2590 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
2591 | return; |
2592 | } | |
3d3b7db0 | 2593 | |
1924bcb0 PM |
2594 | /* |
2595 | * Take a ticket. Note that atomic_inc_return() implies a | |
2596 | * full memory barrier. | |
2597 | */ | |
40694d66 | 2598 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 2599 | firstsnap = snap; |
3d3b7db0 | 2600 | get_online_cpus(); |
1cc85961 | 2601 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 PM |
2602 | |
2603 | /* | |
2604 | * Each pass through the following loop attempts to force a | |
2605 | * context switch on each CPU. | |
2606 | */ | |
2607 | while (try_stop_cpus(cpu_online_mask, | |
2608 | synchronize_sched_expedited_cpu_stop, | |
2609 | NULL) == -EAGAIN) { | |
2610 | put_online_cpus(); | |
a30489c5 | 2611 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 2612 | |
1924bcb0 | 2613 | /* Check to see if someone else did our work for us. */ |
40694d66 | 2614 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2615 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2616 | /* ensure test happens before caller kfree */ |
2617 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2618 | atomic_long_inc(&rsp->expedited_workdone1); | |
1924bcb0 PM |
2619 | return; |
2620 | } | |
3d3b7db0 PM |
2621 | |
2622 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 2623 | if (trycount++ < 10) { |
3d3b7db0 | 2624 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 2625 | } else { |
3705b88d | 2626 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 2627 | atomic_long_inc(&rsp->expedited_normal); |
3d3b7db0 PM |
2628 | return; |
2629 | } | |
2630 | ||
1924bcb0 | 2631 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 2632 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2633 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2634 | /* ensure test happens before caller kfree */ |
2635 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2636 | atomic_long_inc(&rsp->expedited_workdone2); | |
3d3b7db0 PM |
2637 | return; |
2638 | } | |
2639 | ||
2640 | /* | |
2641 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
2642 | * callers to piggyback on our grace period. We retry |
2643 | * after they started, so our grace period works for them, | |
2644 | * and they started after our first try, so their grace | |
2645 | * period works for us. | |
3d3b7db0 PM |
2646 | */ |
2647 | get_online_cpus(); | |
40694d66 | 2648 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
2649 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
2650 | } | |
a30489c5 | 2651 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 PM |
2652 | |
2653 | /* | |
2654 | * Everyone up to our most recent fetch is covered by our grace | |
2655 | * period. Update the counter, but only if our work is still | |
2656 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 2657 | * than we did already did their update. |
3d3b7db0 PM |
2658 | */ |
2659 | do { | |
a30489c5 | 2660 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 2661 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2662 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 PM |
2663 | /* ensure test happens before caller kfree */ |
2664 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2665 | atomic_long_inc(&rsp->expedited_done_lost); | |
3d3b7db0 PM |
2666 | break; |
2667 | } | |
40694d66 | 2668 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 2669 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
2670 | |
2671 | put_online_cpus(); | |
2672 | } | |
2673 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
2674 | ||
64db4cff PM |
2675 | /* |
2676 | * Check to see if there is any immediate RCU-related work to be done | |
2677 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
2678 | * The checks are in order of increasing expense: checks that can be | |
2679 | * carried out against CPU-local state are performed first. However, | |
2680 | * we must check for CPU stalls first, else we might not get a chance. | |
2681 | */ | |
2682 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
2683 | { | |
2f51f988 PM |
2684 | struct rcu_node *rnp = rdp->mynode; |
2685 | ||
64db4cff PM |
2686 | rdp->n_rcu_pending++; |
2687 | ||
2688 | /* Check for CPU stalls, if enabled. */ | |
2689 | check_cpu_stall(rsp, rdp); | |
2690 | ||
2691 | /* Is the RCU core waiting for a quiescent state from this CPU? */ | |
5c51dd73 PM |
2692 | if (rcu_scheduler_fully_active && |
2693 | rdp->qs_pending && !rdp->passed_quiesce) { | |
d21670ac | 2694 | rdp->n_rp_qs_pending++; |
e4cc1f22 | 2695 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
d21670ac | 2696 | rdp->n_rp_report_qs++; |
64db4cff | 2697 | return 1; |
7ba5c840 | 2698 | } |
64db4cff PM |
2699 | |
2700 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
2701 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
2702 | rdp->n_rp_cb_ready++; | |
64db4cff | 2703 | return 1; |
7ba5c840 | 2704 | } |
64db4cff PM |
2705 | |
2706 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
2707 | if (cpu_needs_another_gp(rsp, rdp)) { |
2708 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 2709 | return 1; |
7ba5c840 | 2710 | } |
64db4cff PM |
2711 | |
2712 | /* Has another RCU grace period completed? */ | |
2f51f988 | 2713 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 2714 | rdp->n_rp_gp_completed++; |
64db4cff | 2715 | return 1; |
7ba5c840 | 2716 | } |
64db4cff PM |
2717 | |
2718 | /* Has a new RCU grace period started? */ | |
2f51f988 | 2719 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
7ba5c840 | 2720 | rdp->n_rp_gp_started++; |
64db4cff | 2721 | return 1; |
7ba5c840 | 2722 | } |
64db4cff | 2723 | |
64db4cff | 2724 | /* nothing to do */ |
7ba5c840 | 2725 | rdp->n_rp_need_nothing++; |
64db4cff PM |
2726 | return 0; |
2727 | } | |
2728 | ||
2729 | /* | |
2730 | * Check to see if there is any immediate RCU-related work to be done | |
2731 | * by the current CPU, returning 1 if so. This function is part of the | |
2732 | * RCU implementation; it is -not- an exported member of the RCU API. | |
2733 | */ | |
a157229c | 2734 | static int rcu_pending(int cpu) |
64db4cff | 2735 | { |
6ce75a23 PM |
2736 | struct rcu_state *rsp; |
2737 | ||
2738 | for_each_rcu_flavor(rsp) | |
2739 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu))) | |
2740 | return 1; | |
2741 | return 0; | |
64db4cff PM |
2742 | } |
2743 | ||
2744 | /* | |
c0f4dfd4 PM |
2745 | * Return true if the specified CPU has any callback. If all_lazy is |
2746 | * non-NULL, store an indication of whether all callbacks are lazy. | |
2747 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 2748 | */ |
c0f4dfd4 | 2749 | static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy) |
64db4cff | 2750 | { |
c0f4dfd4 PM |
2751 | bool al = true; |
2752 | bool hc = false; | |
2753 | struct rcu_data *rdp; | |
6ce75a23 PM |
2754 | struct rcu_state *rsp; |
2755 | ||
c0f4dfd4 PM |
2756 | for_each_rcu_flavor(rsp) { |
2757 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
2758 | if (rdp->qlen != rdp->qlen_lazy) | |
2759 | al = false; | |
2760 | if (rdp->nxtlist) | |
2761 | hc = true; | |
2762 | } | |
2763 | if (all_lazy) | |
2764 | *all_lazy = al; | |
2765 | return hc; | |
64db4cff PM |
2766 | } |
2767 | ||
a83eff0a PM |
2768 | /* |
2769 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
2770 | * the compiler is expected to optimize this away. | |
2771 | */ | |
e66c33d5 | 2772 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
2773 | int cpu, unsigned long done) |
2774 | { | |
2775 | trace_rcu_barrier(rsp->name, s, cpu, | |
2776 | atomic_read(&rsp->barrier_cpu_count), done); | |
2777 | } | |
2778 | ||
b1420f1c PM |
2779 | /* |
2780 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
2781 | * up the task executing _rcu_barrier(). | |
2782 | */ | |
24ebbca8 | 2783 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 2784 | { |
24ebbca8 PM |
2785 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
2786 | struct rcu_state *rsp = rdp->rsp; | |
2787 | ||
a83eff0a PM |
2788 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
2789 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 2790 | complete(&rsp->barrier_completion); |
a83eff0a PM |
2791 | } else { |
2792 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
2793 | } | |
d0ec774c PM |
2794 | } |
2795 | ||
2796 | /* | |
2797 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
2798 | */ | |
2799 | static void rcu_barrier_func(void *type) | |
2800 | { | |
037b64ed | 2801 | struct rcu_state *rsp = type; |
06668efa | 2802 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
d0ec774c | 2803 | |
a83eff0a | 2804 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 2805 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 2806 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
2807 | } |
2808 | ||
d0ec774c PM |
2809 | /* |
2810 | * Orchestrate the specified type of RCU barrier, waiting for all | |
2811 | * RCU callbacks of the specified type to complete. | |
2812 | */ | |
037b64ed | 2813 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 2814 | { |
b1420f1c | 2815 | int cpu; |
b1420f1c | 2816 | struct rcu_data *rdp; |
cf3a9c48 PM |
2817 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
2818 | unsigned long snap_done; | |
b1420f1c | 2819 | |
a83eff0a | 2820 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 2821 | |
e74f4c45 | 2822 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 2823 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 2824 | |
cf3a9c48 PM |
2825 | /* |
2826 | * Ensure that all prior references, including to ->n_barrier_done, | |
2827 | * are ordered before the _rcu_barrier() machinery. | |
2828 | */ | |
2829 | smp_mb(); /* See above block comment. */ | |
2830 | ||
2831 | /* | |
2832 | * Recheck ->n_barrier_done to see if others did our work for us. | |
2833 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
2834 | * transition. The "if" expression below therefore rounds the old | |
2835 | * value up to the next even number and adds two before comparing. | |
2836 | */ | |
458fb381 | 2837 | snap_done = rsp->n_barrier_done; |
a83eff0a | 2838 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
458fb381 PM |
2839 | |
2840 | /* | |
2841 | * If the value in snap is odd, we needed to wait for the current | |
2842 | * rcu_barrier() to complete, then wait for the next one, in other | |
2843 | * words, we need the value of snap_done to be three larger than | |
2844 | * the value of snap. On the other hand, if the value in snap is | |
2845 | * even, we only had to wait for the next rcu_barrier() to complete, | |
2846 | * in other words, we need the value of snap_done to be only two | |
2847 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes | |
2848 | * this for us (thank you, Linus!). | |
2849 | */ | |
2850 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) { | |
a83eff0a | 2851 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
2852 | smp_mb(); /* caller's subsequent code after above check. */ |
2853 | mutex_unlock(&rsp->barrier_mutex); | |
2854 | return; | |
2855 | } | |
2856 | ||
2857 | /* | |
2858 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
2859 | * ACCESS_ONCE() to prevent the compiler from speculating | |
2860 | * the increment to precede the early-exit check. | |
2861 | */ | |
2862 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2863 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); | |
a83eff0a | 2864 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 2865 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 2866 | |
d0ec774c | 2867 | /* |
b1420f1c PM |
2868 | * Initialize the count to one rather than to zero in order to |
2869 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
2870 | * (or preemption of this task). Exclude CPU-hotplug operations |
2871 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 2872 | */ |
7db74df8 | 2873 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 2874 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 2875 | get_online_cpus(); |
b1420f1c PM |
2876 | |
2877 | /* | |
1331e7a1 PM |
2878 | * Force each CPU with callbacks to register a new callback. |
2879 | * When that callback is invoked, we will know that all of the | |
2880 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 2881 | */ |
3fbfbf7a | 2882 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 2883 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 2884 | continue; |
b1420f1c | 2885 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 2886 | if (rcu_is_nocb_cpu(cpu)) { |
3fbfbf7a PM |
2887 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, |
2888 | rsp->n_barrier_done); | |
2889 | atomic_inc(&rsp->barrier_cpu_count); | |
2890 | __call_rcu(&rdp->barrier_head, rcu_barrier_callback, | |
2891 | rsp, cpu, 0); | |
2892 | } else if (ACCESS_ONCE(rdp->qlen)) { | |
a83eff0a PM |
2893 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
2894 | rsp->n_barrier_done); | |
037b64ed | 2895 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 2896 | } else { |
a83eff0a PM |
2897 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
2898 | rsp->n_barrier_done); | |
b1420f1c PM |
2899 | } |
2900 | } | |
1331e7a1 | 2901 | put_online_cpus(); |
b1420f1c PM |
2902 | |
2903 | /* | |
2904 | * Now that we have an rcu_barrier_callback() callback on each | |
2905 | * CPU, and thus each counted, remove the initial count. | |
2906 | */ | |
24ebbca8 | 2907 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 2908 | complete(&rsp->barrier_completion); |
b1420f1c | 2909 | |
cf3a9c48 PM |
2910 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
2911 | smp_mb(); /* Keep increment after above mechanism. */ | |
2912 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2913 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); | |
a83eff0a | 2914 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
2915 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
2916 | ||
b1420f1c | 2917 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 2918 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
2919 | |
2920 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 2921 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 2922 | } |
d0ec774c PM |
2923 | |
2924 | /** | |
2925 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
2926 | */ | |
2927 | void rcu_barrier_bh(void) | |
2928 | { | |
037b64ed | 2929 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
2930 | } |
2931 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
2932 | ||
2933 | /** | |
2934 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
2935 | */ | |
2936 | void rcu_barrier_sched(void) | |
2937 | { | |
037b64ed | 2938 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
2939 | } |
2940 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
2941 | ||
64db4cff | 2942 | /* |
27569620 | 2943 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 2944 | */ |
27569620 PM |
2945 | static void __init |
2946 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
2947 | { |
2948 | unsigned long flags; | |
394f99a9 | 2949 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
2950 | struct rcu_node *rnp = rcu_get_root(rsp); |
2951 | ||
2952 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 2953 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 2954 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
3f5d3ea6 | 2955 | init_callback_list(rdp); |
486e2593 | 2956 | rdp->qlen_lazy = 0; |
1d1fb395 | 2957 | ACCESS_ONCE(rdp->qlen) = 0; |
27569620 | 2958 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 2959 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 2960 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 2961 | rdp->cpu = cpu; |
d4c08f2a | 2962 | rdp->rsp = rsp; |
3fbfbf7a | 2963 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 2964 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
2965 | } |
2966 | ||
2967 | /* | |
2968 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
2969 | * offline event can be happening at a given time. Note also that we | |
2970 | * can accept some slop in the rsp->completed access due to the fact | |
2971 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 2972 | */ |
49fb4c62 | 2973 | static void |
6cc68793 | 2974 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible) |
64db4cff PM |
2975 | { |
2976 | unsigned long flags; | |
64db4cff | 2977 | unsigned long mask; |
394f99a9 | 2978 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
2979 | struct rcu_node *rnp = rcu_get_root(rsp); |
2980 | ||
a4fbe35a PM |
2981 | /* Exclude new grace periods. */ |
2982 | mutex_lock(&rsp->onoff_mutex); | |
2983 | ||
64db4cff | 2984 | /* Set up local state, ensuring consistent view of global state. */ |
1304afb2 | 2985 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 2986 | rdp->beenonline = 1; /* We have now been online. */ |
6cc68793 | 2987 | rdp->preemptible = preemptible; |
37c72e56 PM |
2988 | rdp->qlen_last_fqs_check = 0; |
2989 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 2990 | rdp->blimit = blimit; |
0d8ee37e | 2991 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ |
29e37d81 | 2992 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 2993 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
2994 | atomic_set(&rdp->dynticks->dynticks, |
2995 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 2996 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 2997 | |
64db4cff PM |
2998 | /* Add CPU to rcu_node bitmasks. */ |
2999 | rnp = rdp->mynode; | |
3000 | mask = rdp->grpmask; | |
3001 | do { | |
3002 | /* Exclude any attempts to start a new GP on small systems. */ | |
1304afb2 | 3003 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3004 | rnp->qsmaskinit |= mask; |
3005 | mask = rnp->grpmask; | |
d09b62df | 3006 | if (rnp == rdp->mynode) { |
06ae115a PM |
3007 | /* |
3008 | * If there is a grace period in progress, we will | |
3009 | * set up to wait for it next time we run the | |
3010 | * RCU core code. | |
3011 | */ | |
3012 | rdp->gpnum = rnp->completed; | |
d09b62df | 3013 | rdp->completed = rnp->completed; |
06ae115a PM |
3014 | rdp->passed_quiesce = 0; |
3015 | rdp->qs_pending = 0; | |
f7f7bac9 | 3016 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); |
d09b62df | 3017 | } |
1304afb2 | 3018 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3019 | rnp = rnp->parent; |
3020 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); | |
a4fbe35a | 3021 | local_irq_restore(flags); |
64db4cff | 3022 | |
a4fbe35a | 3023 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
3024 | } |
3025 | ||
49fb4c62 | 3026 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 3027 | { |
6ce75a23 PM |
3028 | struct rcu_state *rsp; |
3029 | ||
3030 | for_each_rcu_flavor(rsp) | |
3031 | rcu_init_percpu_data(cpu, rsp, | |
3032 | strcmp(rsp->name, "rcu_preempt") == 0); | |
64db4cff PM |
3033 | } |
3034 | ||
3035 | /* | |
f41d911f | 3036 | * Handle CPU online/offline notification events. |
64db4cff | 3037 | */ |
49fb4c62 | 3038 | static int rcu_cpu_notify(struct notifier_block *self, |
9f680ab4 | 3039 | unsigned long action, void *hcpu) |
64db4cff PM |
3040 | { |
3041 | long cpu = (long)hcpu; | |
27f4d280 | 3042 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); |
a26ac245 | 3043 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 3044 | struct rcu_state *rsp; |
64db4cff | 3045 | |
f7f7bac9 | 3046 | trace_rcu_utilization(TPS("Start CPU hotplug")); |
64db4cff PM |
3047 | switch (action) { |
3048 | case CPU_UP_PREPARE: | |
3049 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
3050 | rcu_prepare_cpu(cpu); |
3051 | rcu_prepare_kthreads(cpu); | |
a26ac245 PM |
3052 | break; |
3053 | case CPU_ONLINE: | |
0f962a5e | 3054 | case CPU_DOWN_FAILED: |
5d01bbd1 | 3055 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
3056 | break; |
3057 | case CPU_DOWN_PREPARE: | |
34ed6246 | 3058 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 3059 | break; |
d0ec774c PM |
3060 | case CPU_DYING: |
3061 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
3062 | for_each_rcu_flavor(rsp) |
3063 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 3064 | break; |
64db4cff PM |
3065 | case CPU_DEAD: |
3066 | case CPU_DEAD_FROZEN: | |
3067 | case CPU_UP_CANCELED: | |
3068 | case CPU_UP_CANCELED_FROZEN: | |
6ce75a23 PM |
3069 | for_each_rcu_flavor(rsp) |
3070 | rcu_cleanup_dead_cpu(cpu, rsp); | |
64db4cff PM |
3071 | break; |
3072 | default: | |
3073 | break; | |
3074 | } | |
f7f7bac9 | 3075 | trace_rcu_utilization(TPS("End CPU hotplug")); |
34ed6246 | 3076 | return NOTIFY_OK; |
64db4cff PM |
3077 | } |
3078 | ||
d1d74d14 BP |
3079 | static int rcu_pm_notify(struct notifier_block *self, |
3080 | unsigned long action, void *hcpu) | |
3081 | { | |
3082 | switch (action) { | |
3083 | case PM_HIBERNATION_PREPARE: | |
3084 | case PM_SUSPEND_PREPARE: | |
3085 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
3086 | rcu_expedited = 1; | |
3087 | break; | |
3088 | case PM_POST_HIBERNATION: | |
3089 | case PM_POST_SUSPEND: | |
3090 | rcu_expedited = 0; | |
3091 | break; | |
3092 | default: | |
3093 | break; | |
3094 | } | |
3095 | return NOTIFY_OK; | |
3096 | } | |
3097 | ||
b3dbec76 PM |
3098 | /* |
3099 | * Spawn the kthread that handles this RCU flavor's grace periods. | |
3100 | */ | |
3101 | static int __init rcu_spawn_gp_kthread(void) | |
3102 | { | |
3103 | unsigned long flags; | |
3104 | struct rcu_node *rnp; | |
3105 | struct rcu_state *rsp; | |
3106 | struct task_struct *t; | |
3107 | ||
3108 | for_each_rcu_flavor(rsp) { | |
f170168b | 3109 | t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
3110 | BUG_ON(IS_ERR(t)); |
3111 | rnp = rcu_get_root(rsp); | |
3112 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3113 | rsp->gp_kthread = t; | |
3114 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3fbfbf7a | 3115 | rcu_spawn_nocb_kthreads(rsp); |
b3dbec76 PM |
3116 | } |
3117 | return 0; | |
3118 | } | |
3119 | early_initcall(rcu_spawn_gp_kthread); | |
3120 | ||
bbad9379 PM |
3121 | /* |
3122 | * This function is invoked towards the end of the scheduler's initialization | |
3123 | * process. Before this is called, the idle task might contain | |
3124 | * RCU read-side critical sections (during which time, this idle | |
3125 | * task is booting the system). After this function is called, the | |
3126 | * idle tasks are prohibited from containing RCU read-side critical | |
3127 | * sections. This function also enables RCU lockdep checking. | |
3128 | */ | |
3129 | void rcu_scheduler_starting(void) | |
3130 | { | |
3131 | WARN_ON(num_online_cpus() != 1); | |
3132 | WARN_ON(nr_context_switches() > 0); | |
3133 | rcu_scheduler_active = 1; | |
3134 | } | |
3135 | ||
64db4cff PM |
3136 | /* |
3137 | * Compute the per-level fanout, either using the exact fanout specified | |
3138 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
3139 | */ | |
3140 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
3141 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3142 | { | |
3143 | int i; | |
3144 | ||
f885b7f2 | 3145 | for (i = rcu_num_lvls - 1; i > 0; i--) |
64db4cff | 3146 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
f885b7f2 | 3147 | rsp->levelspread[0] = rcu_fanout_leaf; |
64db4cff PM |
3148 | } |
3149 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3150 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3151 | { | |
3152 | int ccur; | |
3153 | int cprv; | |
3154 | int i; | |
3155 | ||
4dbd6bb3 | 3156 | cprv = nr_cpu_ids; |
f885b7f2 | 3157 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3158 | ccur = rsp->levelcnt[i]; |
3159 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
3160 | cprv = ccur; | |
3161 | } | |
3162 | } | |
3163 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3164 | ||
3165 | /* | |
3166 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3167 | */ | |
394f99a9 LJ |
3168 | static void __init rcu_init_one(struct rcu_state *rsp, |
3169 | struct rcu_data __percpu *rda) | |
64db4cff | 3170 | { |
394f2769 PM |
3171 | static char *buf[] = { "rcu_node_0", |
3172 | "rcu_node_1", | |
3173 | "rcu_node_2", | |
3174 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ | |
3175 | static char *fqs[] = { "rcu_node_fqs_0", | |
3176 | "rcu_node_fqs_1", | |
3177 | "rcu_node_fqs_2", | |
3178 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ | |
64db4cff PM |
3179 | int cpustride = 1; |
3180 | int i; | |
3181 | int j; | |
3182 | struct rcu_node *rnp; | |
3183 | ||
b6407e86 PM |
3184 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
3185 | ||
4930521a PM |
3186 | /* Silence gcc 4.8 warning about array index out of range. */ |
3187 | if (rcu_num_lvls > RCU_NUM_LVLS) | |
3188 | panic("rcu_init_one: rcu_num_lvls overflow"); | |
3189 | ||
64db4cff PM |
3190 | /* Initialize the level-tracking arrays. */ |
3191 | ||
f885b7f2 PM |
3192 | for (i = 0; i < rcu_num_lvls; i++) |
3193 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
3194 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
3195 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
3196 | rcu_init_levelspread(rsp); | |
3197 | ||
3198 | /* Initialize the elements themselves, starting from the leaves. */ | |
3199 | ||
f885b7f2 | 3200 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3201 | cpustride *= rsp->levelspread[i]; |
3202 | rnp = rsp->level[i]; | |
3203 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 3204 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
3205 | lockdep_set_class_and_name(&rnp->lock, |
3206 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
3207 | raw_spin_lock_init(&rnp->fqslock); |
3208 | lockdep_set_class_and_name(&rnp->fqslock, | |
3209 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
3210 | rnp->gpnum = rsp->gpnum; |
3211 | rnp->completed = rsp->completed; | |
64db4cff PM |
3212 | rnp->qsmask = 0; |
3213 | rnp->qsmaskinit = 0; | |
3214 | rnp->grplo = j * cpustride; | |
3215 | rnp->grphi = (j + 1) * cpustride - 1; | |
3216 | if (rnp->grphi >= NR_CPUS) | |
3217 | rnp->grphi = NR_CPUS - 1; | |
3218 | if (i == 0) { | |
3219 | rnp->grpnum = 0; | |
3220 | rnp->grpmask = 0; | |
3221 | rnp->parent = NULL; | |
3222 | } else { | |
3223 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
3224 | rnp->grpmask = 1UL << rnp->grpnum; | |
3225 | rnp->parent = rsp->level[i - 1] + | |
3226 | j / rsp->levelspread[i - 1]; | |
3227 | } | |
3228 | rnp->level = i; | |
12f5f524 | 3229 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 3230 | rcu_init_one_nocb(rnp); |
64db4cff PM |
3231 | } |
3232 | } | |
0c34029a | 3233 | |
394f99a9 | 3234 | rsp->rda = rda; |
b3dbec76 | 3235 | init_waitqueue_head(&rsp->gp_wq); |
016a8d5b | 3236 | init_irq_work(&rsp->wakeup_work, rsp_wakeup); |
f885b7f2 | 3237 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 3238 | for_each_possible_cpu(i) { |
4a90a068 | 3239 | while (i > rnp->grphi) |
0c34029a | 3240 | rnp++; |
394f99a9 | 3241 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
3242 | rcu_boot_init_percpu_data(i, rsp); |
3243 | } | |
6ce75a23 | 3244 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
3245 | } |
3246 | ||
f885b7f2 PM |
3247 | /* |
3248 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
3249 | * replace the definitions in rcutree.h because those are needed to size | |
3250 | * the ->node array in the rcu_state structure. | |
3251 | */ | |
3252 | static void __init rcu_init_geometry(void) | |
3253 | { | |
026ad283 | 3254 | ulong d; |
f885b7f2 PM |
3255 | int i; |
3256 | int j; | |
cca6f393 | 3257 | int n = nr_cpu_ids; |
f885b7f2 PM |
3258 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
3259 | ||
026ad283 PM |
3260 | /* |
3261 | * Initialize any unspecified boot parameters. | |
3262 | * The default values of jiffies_till_first_fqs and | |
3263 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
3264 | * value, which is a function of HZ, then adding one for each | |
3265 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
3266 | */ | |
3267 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
3268 | if (jiffies_till_first_fqs == ULONG_MAX) | |
3269 | jiffies_till_first_fqs = d; | |
3270 | if (jiffies_till_next_fqs == ULONG_MAX) | |
3271 | jiffies_till_next_fqs = d; | |
3272 | ||
f885b7f2 | 3273 | /* If the compile-time values are accurate, just leave. */ |
b17c7035 PM |
3274 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
3275 | nr_cpu_ids == NR_CPUS) | |
f885b7f2 PM |
3276 | return; |
3277 | ||
3278 | /* | |
3279 | * Compute number of nodes that can be handled an rcu_node tree | |
3280 | * with the given number of levels. Setting rcu_capacity[0] makes | |
3281 | * some of the arithmetic easier. | |
3282 | */ | |
3283 | rcu_capacity[0] = 1; | |
3284 | rcu_capacity[1] = rcu_fanout_leaf; | |
3285 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
3286 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
3287 | ||
3288 | /* | |
3289 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
3290 | * to increase the leaf-level fanout, not decrease it. Of course, | |
3291 | * the leaf-level fanout cannot exceed the number of bits in | |
3292 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
3293 | * the configured number of CPUs. Complain and fall back to the | |
3294 | * compile-time values if these limits are exceeded. | |
3295 | */ | |
3296 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
3297 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
3298 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
3299 | WARN_ON(1); | |
3300 | return; | |
3301 | } | |
3302 | ||
3303 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
3304 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
3305 | if (n <= rcu_capacity[i]) { | |
3306 | for (j = 0; j <= i; j++) | |
3307 | num_rcu_lvl[j] = | |
3308 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
3309 | rcu_num_lvls = i; | |
3310 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
3311 | num_rcu_lvl[j] = 0; | |
3312 | break; | |
3313 | } | |
3314 | ||
3315 | /* Calculate the total number of rcu_node structures. */ | |
3316 | rcu_num_nodes = 0; | |
3317 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
3318 | rcu_num_nodes += num_rcu_lvl[i]; | |
3319 | rcu_num_nodes -= n; | |
3320 | } | |
3321 | ||
9f680ab4 | 3322 | void __init rcu_init(void) |
64db4cff | 3323 | { |
017c4261 | 3324 | int cpu; |
9f680ab4 | 3325 | |
f41d911f | 3326 | rcu_bootup_announce(); |
f885b7f2 | 3327 | rcu_init_geometry(); |
394f99a9 LJ |
3328 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
3329 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); | |
f41d911f | 3330 | __rcu_init_preempt(); |
b5b39360 | 3331 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
3332 | |
3333 | /* | |
3334 | * We don't need protection against CPU-hotplug here because | |
3335 | * this is called early in boot, before either interrupts | |
3336 | * or the scheduler are operational. | |
3337 | */ | |
3338 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 3339 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
3340 | for_each_online_cpu(cpu) |
3341 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
3342 | } |
3343 | ||
1eba8f84 | 3344 | #include "rcutree_plugin.h" |