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> | |
64db4cff | 55 | |
9f77da9f | 56 | #include "rcutree.h" |
29c00b4a PM |
57 | #include <trace/events/rcu.h> |
58 | ||
59 | #include "rcu.h" | |
9f77da9f | 60 | |
64db4cff PM |
61 | /* Data structures. */ |
62 | ||
f885b7f2 | 63 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
88b91c7c | 64 | |
037b64ed | 65 | #define RCU_STATE_INITIALIZER(sname, cr) { \ |
6c90cc7b | 66 | .level = { &sname##_state.node[0] }, \ |
037b64ed | 67 | .call = cr, \ |
af446b70 | 68 | .fqs_state = RCU_GP_IDLE, \ |
64db4cff PM |
69 | .gpnum = -300, \ |
70 | .completed = -300, \ | |
6c90cc7b PM |
71 | .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.onofflock), \ |
72 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ | |
73 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 74 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
6c90cc7b | 75 | .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.fqslock), \ |
6c90cc7b | 76 | .name = #sname, \ |
64db4cff PM |
77 | } |
78 | ||
037b64ed PM |
79 | struct rcu_state rcu_sched_state = |
80 | RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched); | |
d6714c22 | 81 | DEFINE_PER_CPU(struct rcu_data, rcu_sched_data); |
64db4cff | 82 | |
037b64ed | 83 | struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh); |
6258c4fb | 84 | DEFINE_PER_CPU(struct rcu_data, rcu_bh_data); |
b1f77b05 | 85 | |
27f4d280 | 86 | static struct rcu_state *rcu_state; |
6ce75a23 | 87 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 88 | |
f885b7f2 PM |
89 | /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */ |
90 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF; | |
91 | module_param(rcu_fanout_leaf, int, 0); | |
92 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; | |
93 | static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */ | |
94 | NUM_RCU_LVL_0, | |
95 | NUM_RCU_LVL_1, | |
96 | NUM_RCU_LVL_2, | |
97 | NUM_RCU_LVL_3, | |
98 | NUM_RCU_LVL_4, | |
99 | }; | |
100 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ | |
101 | ||
b0d30417 PM |
102 | /* |
103 | * The rcu_scheduler_active variable transitions from zero to one just | |
104 | * before the first task is spawned. So when this variable is zero, RCU | |
105 | * can assume that there is but one task, allowing RCU to (for example) | |
106 | * optimized synchronize_sched() to a simple barrier(). When this variable | |
107 | * is one, RCU must actually do all the hard work required to detect real | |
108 | * grace periods. This variable is also used to suppress boot-time false | |
109 | * positives from lockdep-RCU error checking. | |
110 | */ | |
bbad9379 PM |
111 | int rcu_scheduler_active __read_mostly; |
112 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
113 | ||
b0d30417 PM |
114 | /* |
115 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
116 | * during the early_initcall() processing, which is after the scheduler | |
117 | * is capable of creating new tasks. So RCU processing (for example, | |
118 | * creating tasks for RCU priority boosting) must be delayed until after | |
119 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
120 | * currently delay invocation of any RCU callbacks until after this point. | |
121 | * | |
122 | * It might later prove better for people registering RCU callbacks during | |
123 | * early boot to take responsibility for these callbacks, but one step at | |
124 | * a time. | |
125 | */ | |
126 | static int rcu_scheduler_fully_active __read_mostly; | |
127 | ||
a46e0899 PM |
128 | #ifdef CONFIG_RCU_BOOST |
129 | ||
a26ac245 PM |
130 | /* |
131 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
132 | * handle all flavors of RCU. | |
133 | */ | |
134 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
d71df90e | 135 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); |
15ba0ba8 | 136 | DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu); |
5ece5bab | 137 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); |
d71df90e | 138 | DEFINE_PER_CPU(char, rcu_cpu_has_work); |
a26ac245 | 139 | |
a46e0899 PM |
140 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
141 | ||
0f962a5e | 142 | static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
143 | static void invoke_rcu_core(void); |
144 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 145 | |
4a298656 PM |
146 | /* |
147 | * Track the rcutorture test sequence number and the update version | |
148 | * number within a given test. The rcutorture_testseq is incremented | |
149 | * on every rcutorture module load and unload, so has an odd value | |
150 | * when a test is running. The rcutorture_vernum is set to zero | |
151 | * when rcutorture starts and is incremented on each rcutorture update. | |
152 | * These variables enable correlating rcutorture output with the | |
153 | * RCU tracing information. | |
154 | */ | |
155 | unsigned long rcutorture_testseq; | |
156 | unsigned long rcutorture_vernum; | |
157 | ||
fc2219d4 PM |
158 | /* |
159 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
160 | * permit this function to be invoked without holding the root rcu_node | |
161 | * structure's ->lock, but of course results can be subject to change. | |
162 | */ | |
163 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
164 | { | |
165 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
166 | } | |
167 | ||
b1f77b05 | 168 | /* |
d6714c22 | 169 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 170 | * how many quiescent states passed, just if there was at least |
d6714c22 | 171 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 172 | * The caller must have disabled preemption. |
b1f77b05 | 173 | */ |
d6714c22 | 174 | void rcu_sched_qs(int cpu) |
b1f77b05 | 175 | { |
25502a6c | 176 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu); |
f41d911f | 177 | |
e4cc1f22 | 178 | rdp->passed_quiesce_gpnum = rdp->gpnum; |
c3422bea | 179 | barrier(); |
e4cc1f22 | 180 | if (rdp->passed_quiesce == 0) |
d4c08f2a | 181 | trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs"); |
e4cc1f22 | 182 | rdp->passed_quiesce = 1; |
b1f77b05 IM |
183 | } |
184 | ||
d6714c22 | 185 | void rcu_bh_qs(int cpu) |
b1f77b05 | 186 | { |
25502a6c | 187 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu); |
f41d911f | 188 | |
e4cc1f22 | 189 | rdp->passed_quiesce_gpnum = rdp->gpnum; |
c3422bea | 190 | barrier(); |
e4cc1f22 | 191 | if (rdp->passed_quiesce == 0) |
d4c08f2a | 192 | trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs"); |
e4cc1f22 | 193 | rdp->passed_quiesce = 1; |
b1f77b05 | 194 | } |
64db4cff | 195 | |
25502a6c PM |
196 | /* |
197 | * Note a context switch. This is a quiescent state for RCU-sched, | |
198 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 199 | * The caller must have disabled preemption. |
25502a6c PM |
200 | */ |
201 | void rcu_note_context_switch(int cpu) | |
202 | { | |
300df91c | 203 | trace_rcu_utilization("Start context switch"); |
25502a6c | 204 | rcu_sched_qs(cpu); |
cba6d0d6 | 205 | rcu_preempt_note_context_switch(cpu); |
300df91c | 206 | trace_rcu_utilization("End context switch"); |
25502a6c | 207 | } |
29ce8310 | 208 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 209 | |
90a4d2c0 | 210 | DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { |
29e37d81 | 211 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, |
23b5c8fa | 212 | .dynticks = ATOMIC_INIT(1), |
90a4d2c0 | 213 | }; |
64db4cff | 214 | |
e0f23060 | 215 | static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
64db4cff PM |
216 | static int qhimark = 10000; /* If this many pending, ignore blimit. */ |
217 | static int qlowmark = 100; /* Once only this many pending, use blimit. */ | |
218 | ||
3d76c082 PM |
219 | module_param(blimit, int, 0); |
220 | module_param(qhimark, int, 0); | |
221 | module_param(qlowmark, int, 0); | |
222 | ||
13cfcca0 PM |
223 | int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */ |
224 | int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT; | |
225 | ||
f2e0dd70 | 226 | module_param(rcu_cpu_stall_suppress, int, 0644); |
13cfcca0 | 227 | module_param(rcu_cpu_stall_timeout, int, 0644); |
742734ee | 228 | |
64db4cff | 229 | static void force_quiescent_state(struct rcu_state *rsp, int relaxed); |
a157229c | 230 | static int rcu_pending(int cpu); |
64db4cff PM |
231 | |
232 | /* | |
d6714c22 | 233 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
64db4cff | 234 | */ |
d6714c22 | 235 | long rcu_batches_completed_sched(void) |
64db4cff | 236 | { |
d6714c22 | 237 | return rcu_sched_state.completed; |
64db4cff | 238 | } |
d6714c22 | 239 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
240 | |
241 | /* | |
242 | * Return the number of RCU BH batches processed thus far for debug & stats. | |
243 | */ | |
244 | long rcu_batches_completed_bh(void) | |
245 | { | |
246 | return rcu_bh_state.completed; | |
247 | } | |
248 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
249 | ||
bf66f18e PM |
250 | /* |
251 | * Force a quiescent state for RCU BH. | |
252 | */ | |
253 | void rcu_bh_force_quiescent_state(void) | |
254 | { | |
255 | force_quiescent_state(&rcu_bh_state, 0); | |
256 | } | |
257 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
258 | ||
4a298656 PM |
259 | /* |
260 | * Record the number of times rcutorture tests have been initiated and | |
261 | * terminated. This information allows the debugfs tracing stats to be | |
262 | * correlated to the rcutorture messages, even when the rcutorture module | |
263 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
264 | * store this state in rcutorture itself. | |
265 | */ | |
266 | void rcutorture_record_test_transition(void) | |
267 | { | |
268 | rcutorture_testseq++; | |
269 | rcutorture_vernum = 0; | |
270 | } | |
271 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
272 | ||
273 | /* | |
274 | * Record the number of writer passes through the current rcutorture test. | |
275 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
276 | * messages. | |
277 | */ | |
278 | void rcutorture_record_progress(unsigned long vernum) | |
279 | { | |
280 | rcutorture_vernum++; | |
281 | } | |
282 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
283 | ||
bf66f18e PM |
284 | /* |
285 | * Force a quiescent state for RCU-sched. | |
286 | */ | |
287 | void rcu_sched_force_quiescent_state(void) | |
288 | { | |
289 | force_quiescent_state(&rcu_sched_state, 0); | |
290 | } | |
291 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
292 | ||
64db4cff PM |
293 | /* |
294 | * Does the CPU have callbacks ready to be invoked? | |
295 | */ | |
296 | static int | |
297 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
298 | { | |
299 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL]; | |
300 | } | |
301 | ||
302 | /* | |
303 | * Does the current CPU require a yet-as-unscheduled grace period? | |
304 | */ | |
305 | static int | |
306 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
307 | { | |
a10d206e PM |
308 | return *rdp->nxttail[RCU_DONE_TAIL + |
309 | ACCESS_ONCE(rsp->completed) != rdp->completed] && | |
310 | !rcu_gp_in_progress(rsp); | |
64db4cff PM |
311 | } |
312 | ||
313 | /* | |
314 | * Return the root node of the specified rcu_state structure. | |
315 | */ | |
316 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
317 | { | |
318 | return &rsp->node[0]; | |
319 | } | |
320 | ||
9b2e4f18 PM |
321 | /* |
322 | * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle | |
323 | * | |
324 | * If the new value of the ->dynticks_nesting counter now is zero, | |
325 | * we really have entered idle, and must do the appropriate accounting. | |
326 | * The caller must have disabled interrupts. | |
327 | */ | |
4145fa7f | 328 | static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval) |
9b2e4f18 | 329 | { |
facc4e15 | 330 | trace_rcu_dyntick("Start", oldval, 0); |
99745b6a | 331 | if (!is_idle_task(current)) { |
0989cb46 PM |
332 | struct task_struct *idle = idle_task(smp_processor_id()); |
333 | ||
facc4e15 | 334 | trace_rcu_dyntick("Error on entry: not idle task", oldval, 0); |
bf1304e9 | 335 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
336 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
337 | current->pid, current->comm, | |
338 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 339 | } |
aea1b35e | 340 | rcu_prepare_for_idle(smp_processor_id()); |
9b2e4f18 PM |
341 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
342 | smp_mb__before_atomic_inc(); /* See above. */ | |
343 | atomic_inc(&rdtp->dynticks); | |
344 | smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */ | |
345 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
c44e2cdd PM |
346 | |
347 | /* | |
348 | * The idle task is not permitted to enter the idle loop while | |
349 | * in an RCU read-side critical section. | |
350 | */ | |
351 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
352 | "Illegal idle entry in RCU read-side critical section."); | |
353 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
354 | "Illegal idle entry in RCU-bh read-side critical section."); | |
355 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
356 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 357 | } |
64db4cff PM |
358 | |
359 | /** | |
9b2e4f18 | 360 | * rcu_idle_enter - inform RCU that current CPU is entering idle |
64db4cff | 361 | * |
9b2e4f18 | 362 | * Enter idle mode, in other words, -leave- the mode in which RCU |
64db4cff | 363 | * read-side critical sections can occur. (Though RCU read-side |
9b2e4f18 PM |
364 | * critical sections can occur in irq handlers in idle, a possibility |
365 | * handled by irq_enter() and irq_exit().) | |
366 | * | |
367 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
368 | * the possibility of usermode upcalls having messed up our count | |
369 | * of interrupt nesting level during the prior busy period. | |
64db4cff | 370 | */ |
9b2e4f18 | 371 | void rcu_idle_enter(void) |
64db4cff PM |
372 | { |
373 | unsigned long flags; | |
4145fa7f | 374 | long long oldval; |
64db4cff PM |
375 | struct rcu_dynticks *rdtp; |
376 | ||
64db4cff PM |
377 | local_irq_save(flags); |
378 | rdtp = &__get_cpu_var(rcu_dynticks); | |
4145fa7f | 379 | oldval = rdtp->dynticks_nesting; |
29e37d81 PM |
380 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
381 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) | |
382 | rdtp->dynticks_nesting = 0; | |
383 | else | |
384 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; | |
4145fa7f | 385 | rcu_idle_enter_common(rdtp, oldval); |
64db4cff PM |
386 | local_irq_restore(flags); |
387 | } | |
8a2ecf47 | 388 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 389 | |
9b2e4f18 PM |
390 | /** |
391 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
392 | * | |
393 | * Exit from an interrupt handler, which might possibly result in entering | |
394 | * idle mode, in other words, leaving the mode in which read-side critical | |
395 | * sections can occur. | |
64db4cff | 396 | * |
9b2e4f18 PM |
397 | * This code assumes that the idle loop never does anything that might |
398 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
399 | * architecture violates this assumption, RCU will give you what you | |
400 | * deserve, good and hard. But very infrequently and irreproducibly. | |
401 | * | |
402 | * Use things like work queues to work around this limitation. | |
403 | * | |
404 | * You have been warned. | |
64db4cff | 405 | */ |
9b2e4f18 | 406 | void rcu_irq_exit(void) |
64db4cff PM |
407 | { |
408 | unsigned long flags; | |
4145fa7f | 409 | long long oldval; |
64db4cff PM |
410 | struct rcu_dynticks *rdtp; |
411 | ||
412 | local_irq_save(flags); | |
413 | rdtp = &__get_cpu_var(rcu_dynticks); | |
4145fa7f | 414 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
415 | rdtp->dynticks_nesting--; |
416 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 FW |
417 | if (rdtp->dynticks_nesting) |
418 | trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting); | |
419 | else | |
420 | rcu_idle_enter_common(rdtp, oldval); | |
9b2e4f18 PM |
421 | local_irq_restore(flags); |
422 | } | |
423 | ||
424 | /* | |
425 | * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle | |
426 | * | |
427 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
428 | * we really have exited idle, and must do the appropriate accounting. | |
429 | * The caller must have disabled interrupts. | |
430 | */ | |
431 | static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval) | |
432 | { | |
23b5c8fa PM |
433 | smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */ |
434 | atomic_inc(&rdtp->dynticks); | |
435 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
436 | smp_mb__after_atomic_inc(); /* See above. */ | |
437 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
7cb92499 | 438 | rcu_cleanup_after_idle(smp_processor_id()); |
4145fa7f | 439 | trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting); |
99745b6a | 440 | if (!is_idle_task(current)) { |
0989cb46 PM |
441 | struct task_struct *idle = idle_task(smp_processor_id()); |
442 | ||
4145fa7f PM |
443 | trace_rcu_dyntick("Error on exit: not idle task", |
444 | oldval, rdtp->dynticks_nesting); | |
bf1304e9 | 445 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
446 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
447 | current->pid, current->comm, | |
448 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
449 | } |
450 | } | |
451 | ||
452 | /** | |
453 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
454 | * | |
455 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
456 | * read-side critical sections can occur. | |
457 | * | |
29e37d81 | 458 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to |
4145fa7f | 459 | * allow for the possibility of usermode upcalls messing up our count |
9b2e4f18 PM |
460 | * of interrupt nesting level during the busy period that is just |
461 | * now starting. | |
462 | */ | |
463 | void rcu_idle_exit(void) | |
464 | { | |
465 | unsigned long flags; | |
466 | struct rcu_dynticks *rdtp; | |
467 | long long oldval; | |
468 | ||
469 | local_irq_save(flags); | |
470 | rdtp = &__get_cpu_var(rcu_dynticks); | |
471 | oldval = rdtp->dynticks_nesting; | |
29e37d81 PM |
472 | WARN_ON_ONCE(oldval < 0); |
473 | if (oldval & DYNTICK_TASK_NEST_MASK) | |
474 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; | |
475 | else | |
476 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; | |
9b2e4f18 PM |
477 | rcu_idle_exit_common(rdtp, oldval); |
478 | local_irq_restore(flags); | |
479 | } | |
8a2ecf47 | 480 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 PM |
481 | |
482 | /** | |
483 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
484 | * | |
485 | * Enter an interrupt handler, which might possibly result in exiting | |
486 | * idle mode, in other words, entering the mode in which read-side critical | |
487 | * sections can occur. | |
488 | * | |
489 | * Note that the Linux kernel is fully capable of entering an interrupt | |
490 | * handler that it never exits, for example when doing upcalls to | |
491 | * user mode! This code assumes that the idle loop never does upcalls to | |
492 | * user mode. If your architecture does do upcalls from the idle loop (or | |
493 | * does anything else that results in unbalanced calls to the irq_enter() | |
494 | * and irq_exit() functions), RCU will give you what you deserve, good | |
495 | * and hard. But very infrequently and irreproducibly. | |
496 | * | |
497 | * Use things like work queues to work around this limitation. | |
498 | * | |
499 | * You have been warned. | |
500 | */ | |
501 | void rcu_irq_enter(void) | |
502 | { | |
503 | unsigned long flags; | |
504 | struct rcu_dynticks *rdtp; | |
505 | long long oldval; | |
506 | ||
507 | local_irq_save(flags); | |
508 | rdtp = &__get_cpu_var(rcu_dynticks); | |
509 | oldval = rdtp->dynticks_nesting; | |
510 | rdtp->dynticks_nesting++; | |
511 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 FW |
512 | if (oldval) |
513 | trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting); | |
514 | else | |
515 | rcu_idle_exit_common(rdtp, oldval); | |
64db4cff | 516 | local_irq_restore(flags); |
64db4cff PM |
517 | } |
518 | ||
519 | /** | |
520 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
521 | * | |
522 | * If the CPU was idle with dynamic ticks active, and there is no | |
523 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
524 | * RCU grace-period handling know that the CPU is active. | |
525 | */ | |
526 | void rcu_nmi_enter(void) | |
527 | { | |
528 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); | |
529 | ||
23b5c8fa PM |
530 | if (rdtp->dynticks_nmi_nesting == 0 && |
531 | (atomic_read(&rdtp->dynticks) & 0x1)) | |
64db4cff | 532 | return; |
23b5c8fa PM |
533 | rdtp->dynticks_nmi_nesting++; |
534 | smp_mb__before_atomic_inc(); /* Force delay from prior write. */ | |
535 | atomic_inc(&rdtp->dynticks); | |
536 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
537 | smp_mb__after_atomic_inc(); /* See above. */ | |
538 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
64db4cff PM |
539 | } |
540 | ||
541 | /** | |
542 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
543 | * | |
544 | * If the CPU was idle with dynamic ticks active, and there is no | |
545 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
546 | * RCU grace-period handling know that the CPU is no longer active. | |
547 | */ | |
548 | void rcu_nmi_exit(void) | |
549 | { | |
550 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); | |
551 | ||
23b5c8fa PM |
552 | if (rdtp->dynticks_nmi_nesting == 0 || |
553 | --rdtp->dynticks_nmi_nesting != 0) | |
64db4cff | 554 | return; |
23b5c8fa PM |
555 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
556 | smp_mb__before_atomic_inc(); /* See above. */ | |
557 | atomic_inc(&rdtp->dynticks); | |
558 | smp_mb__after_atomic_inc(); /* Force delay to next write. */ | |
559 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
64db4cff PM |
560 | } |
561 | ||
562 | /** | |
9b2e4f18 | 563 | * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle |
64db4cff | 564 | * |
9b2e4f18 | 565 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 566 | * or NMI handler, return true. |
64db4cff | 567 | */ |
9b2e4f18 | 568 | int rcu_is_cpu_idle(void) |
64db4cff | 569 | { |
34240697 PM |
570 | int ret; |
571 | ||
572 | preempt_disable(); | |
573 | ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0; | |
574 | preempt_enable(); | |
575 | return ret; | |
64db4cff | 576 | } |
e6b80a3b | 577 | EXPORT_SYMBOL(rcu_is_cpu_idle); |
64db4cff | 578 | |
62fde6ed | 579 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
580 | |
581 | /* | |
582 | * Is the current CPU online? Disable preemption to avoid false positives | |
583 | * that could otherwise happen due to the current CPU number being sampled, | |
584 | * this task being preempted, its old CPU being taken offline, resuming | |
585 | * on some other CPU, then determining that its old CPU is now offline. | |
586 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
587 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
588 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
589 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
590 | * offline to continue to use RCU for one jiffy after marking itself | |
591 | * offline in the cpu_online_mask. This leniency is necessary given the | |
592 | * non-atomic nature of the online and offline processing, for example, | |
593 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
594 | * notifiers. | |
595 | * | |
596 | * This is also why RCU internally marks CPUs online during the | |
597 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
598 | * |
599 | * Disable checking if in an NMI handler because we cannot safely report | |
600 | * errors from NMI handlers anyway. | |
601 | */ | |
602 | bool rcu_lockdep_current_cpu_online(void) | |
603 | { | |
2036d94a PM |
604 | struct rcu_data *rdp; |
605 | struct rcu_node *rnp; | |
c0d6d01b PM |
606 | bool ret; |
607 | ||
608 | if (in_nmi()) | |
609 | return 1; | |
610 | preempt_disable(); | |
2036d94a PM |
611 | rdp = &__get_cpu_var(rcu_sched_data); |
612 | rnp = rdp->mynode; | |
613 | ret = (rdp->grpmask & rnp->qsmaskinit) || | |
c0d6d01b PM |
614 | !rcu_scheduler_fully_active; |
615 | preempt_enable(); | |
616 | return ret; | |
617 | } | |
618 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
619 | ||
62fde6ed | 620 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 621 | |
64db4cff | 622 | /** |
9b2e4f18 | 623 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 624 | * |
9b2e4f18 PM |
625 | * If the current CPU is idle or running at a first-level (not nested) |
626 | * interrupt from idle, return true. The caller must have at least | |
627 | * disabled preemption. | |
64db4cff | 628 | */ |
9b2e4f18 | 629 | int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 630 | { |
9b2e4f18 | 631 | return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1; |
64db4cff PM |
632 | } |
633 | ||
64db4cff PM |
634 | /* |
635 | * Snapshot the specified CPU's dynticks counter so that we can later | |
636 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 637 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff PM |
638 | */ |
639 | static int dyntick_save_progress_counter(struct rcu_data *rdp) | |
640 | { | |
23b5c8fa | 641 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
f0e7c19d | 642 | return (rdp->dynticks_snap & 0x1) == 0; |
64db4cff PM |
643 | } |
644 | ||
645 | /* | |
646 | * Return true if the specified CPU has passed through a quiescent | |
647 | * state by virtue of being in or having passed through an dynticks | |
648 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 649 | * for this same CPU, or by virtue of having been offline. |
64db4cff PM |
650 | */ |
651 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp) | |
652 | { | |
7eb4f455 PM |
653 | unsigned int curr; |
654 | unsigned int snap; | |
64db4cff | 655 | |
7eb4f455 PM |
656 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
657 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
658 | |
659 | /* | |
660 | * If the CPU passed through or entered a dynticks idle phase with | |
661 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
662 | * already acknowledged the request to pass through a quiescent | |
663 | * state. Either way, that CPU cannot possibly be in an RCU | |
664 | * read-side critical section that started before the beginning | |
665 | * of the current RCU grace period. | |
666 | */ | |
7eb4f455 | 667 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
d4c08f2a | 668 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti"); |
64db4cff PM |
669 | rdp->dynticks_fqs++; |
670 | return 1; | |
671 | } | |
672 | ||
a82dcc76 PM |
673 | /* |
674 | * Check for the CPU being offline, but only if the grace period | |
675 | * is old enough. We don't need to worry about the CPU changing | |
676 | * state: If we see it offline even once, it has been through a | |
677 | * quiescent state. | |
678 | * | |
679 | * The reason for insisting that the grace period be at least | |
680 | * one jiffy old is that CPUs that are not quite online and that | |
681 | * have just gone offline can still execute RCU read-side critical | |
682 | * sections. | |
683 | */ | |
684 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
685 | return 0; /* Grace period is not old enough. */ | |
686 | barrier(); | |
687 | if (cpu_is_offline(rdp->cpu)) { | |
688 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl"); | |
689 | rdp->offline_fqs++; | |
690 | return 1; | |
691 | } | |
692 | return 0; | |
64db4cff PM |
693 | } |
694 | ||
13cfcca0 PM |
695 | static int jiffies_till_stall_check(void) |
696 | { | |
697 | int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout); | |
698 | ||
699 | /* | |
700 | * Limit check must be consistent with the Kconfig limits | |
701 | * for CONFIG_RCU_CPU_STALL_TIMEOUT. | |
702 | */ | |
703 | if (till_stall_check < 3) { | |
704 | ACCESS_ONCE(rcu_cpu_stall_timeout) = 3; | |
705 | till_stall_check = 3; | |
706 | } else if (till_stall_check > 300) { | |
707 | ACCESS_ONCE(rcu_cpu_stall_timeout) = 300; | |
708 | till_stall_check = 300; | |
709 | } | |
710 | return till_stall_check * HZ + RCU_STALL_DELAY_DELTA; | |
711 | } | |
712 | ||
64db4cff PM |
713 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
714 | { | |
715 | rsp->gp_start = jiffies; | |
13cfcca0 | 716 | rsp->jiffies_stall = jiffies + jiffies_till_stall_check(); |
64db4cff PM |
717 | } |
718 | ||
719 | static void print_other_cpu_stall(struct rcu_state *rsp) | |
720 | { | |
721 | int cpu; | |
722 | long delta; | |
723 | unsigned long flags; | |
285fe294 | 724 | int ndetected = 0; |
64db4cff | 725 | struct rcu_node *rnp = rcu_get_root(rsp); |
64db4cff PM |
726 | |
727 | /* Only let one CPU complain about others per time interval. */ | |
728 | ||
1304afb2 | 729 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 730 | delta = jiffies - rsp->jiffies_stall; |
fc2219d4 | 731 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 732 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
733 | return; |
734 | } | |
13cfcca0 | 735 | rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3; |
1304afb2 | 736 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 737 | |
8cdd32a9 PM |
738 | /* |
739 | * OK, time to rat on our buddy... | |
740 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
741 | * RCU CPU stall warnings. | |
742 | */ | |
a858af28 | 743 | printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 744 | rsp->name); |
a858af28 | 745 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 746 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 747 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 748 | ndetected += rcu_print_task_stall(rnp); |
3acd9eb3 | 749 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
a0b6c9a7 | 750 | if (rnp->qsmask == 0) |
64db4cff | 751 | continue; |
a0b6c9a7 | 752 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) |
9bc8b558 | 753 | if (rnp->qsmask & (1UL << cpu)) { |
a858af28 | 754 | print_cpu_stall_info(rsp, rnp->grplo + cpu); |
9bc8b558 PM |
755 | ndetected++; |
756 | } | |
64db4cff | 757 | } |
a858af28 PM |
758 | |
759 | /* | |
760 | * Now rat on any tasks that got kicked up to the root rcu_node | |
761 | * due to CPU offlining. | |
762 | */ | |
763 | rnp = rcu_get_root(rsp); | |
764 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
285fe294 | 765 | ndetected += rcu_print_task_stall(rnp); |
a858af28 PM |
766 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
767 | ||
768 | print_cpu_stall_info_end(); | |
769 | printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n", | |
64db4cff | 770 | smp_processor_id(), (long)(jiffies - rsp->gp_start)); |
9bc8b558 PM |
771 | if (ndetected == 0) |
772 | printk(KERN_ERR "INFO: Stall ended before state dump start\n"); | |
773 | else if (!trigger_all_cpu_backtrace()) | |
4627e240 | 774 | dump_stack(); |
c1dc0b9c | 775 | |
1ed509a2 PM |
776 | /* If so configured, complain about tasks blocking the grace period. */ |
777 | ||
778 | rcu_print_detail_task_stall(rsp); | |
779 | ||
64db4cff PM |
780 | force_quiescent_state(rsp, 0); /* Kick them all. */ |
781 | } | |
782 | ||
783 | static void print_cpu_stall(struct rcu_state *rsp) | |
784 | { | |
785 | unsigned long flags; | |
786 | struct rcu_node *rnp = rcu_get_root(rsp); | |
787 | ||
8cdd32a9 PM |
788 | /* |
789 | * OK, time to rat on ourselves... | |
790 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
791 | * RCU CPU stall warnings. | |
792 | */ | |
a858af28 PM |
793 | printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name); |
794 | print_cpu_stall_info_begin(); | |
795 | print_cpu_stall_info(rsp, smp_processor_id()); | |
796 | print_cpu_stall_info_end(); | |
797 | printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start); | |
4627e240 PM |
798 | if (!trigger_all_cpu_backtrace()) |
799 | dump_stack(); | |
c1dc0b9c | 800 | |
1304afb2 | 801 | raw_spin_lock_irqsave(&rnp->lock, flags); |
20133cfc | 802 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall)) |
13cfcca0 PM |
803 | rsp->jiffies_stall = jiffies + |
804 | 3 * jiffies_till_stall_check() + 3; | |
1304afb2 | 805 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 806 | |
64db4cff PM |
807 | set_need_resched(); /* kick ourselves to get things going. */ |
808 | } | |
809 | ||
810 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
811 | { | |
bad6e139 PM |
812 | unsigned long j; |
813 | unsigned long js; | |
64db4cff PM |
814 | struct rcu_node *rnp; |
815 | ||
742734ee | 816 | if (rcu_cpu_stall_suppress) |
c68de209 | 817 | return; |
bad6e139 PM |
818 | j = ACCESS_ONCE(jiffies); |
819 | js = ACCESS_ONCE(rsp->jiffies_stall); | |
64db4cff | 820 | rnp = rdp->mynode; |
bad6e139 | 821 | if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) { |
64db4cff PM |
822 | |
823 | /* We haven't checked in, so go dump stack. */ | |
824 | print_cpu_stall(rsp); | |
825 | ||
bad6e139 PM |
826 | } else if (rcu_gp_in_progress(rsp) && |
827 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 828 | |
bad6e139 | 829 | /* They had a few time units to dump stack, so complain. */ |
64db4cff PM |
830 | print_other_cpu_stall(rsp); |
831 | } | |
832 | } | |
833 | ||
c68de209 PM |
834 | static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr) |
835 | { | |
742734ee | 836 | rcu_cpu_stall_suppress = 1; |
c68de209 PM |
837 | return NOTIFY_DONE; |
838 | } | |
839 | ||
53d84e00 PM |
840 | /** |
841 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
842 | * | |
843 | * Set the stall-warning timeout way off into the future, thus preventing | |
844 | * any RCU CPU stall-warning messages from appearing in the current set of | |
845 | * RCU grace periods. | |
846 | * | |
847 | * The caller must disable hard irqs. | |
848 | */ | |
849 | void rcu_cpu_stall_reset(void) | |
850 | { | |
6ce75a23 PM |
851 | struct rcu_state *rsp; |
852 | ||
853 | for_each_rcu_flavor(rsp) | |
854 | rsp->jiffies_stall = jiffies + ULONG_MAX / 2; | |
53d84e00 PM |
855 | } |
856 | ||
c68de209 PM |
857 | static struct notifier_block rcu_panic_block = { |
858 | .notifier_call = rcu_panic, | |
859 | }; | |
860 | ||
861 | static void __init check_cpu_stall_init(void) | |
862 | { | |
863 | atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block); | |
864 | } | |
865 | ||
64db4cff PM |
866 | /* |
867 | * Update CPU-local rcu_data state to record the newly noticed grace period. | |
868 | * This is used both when we started the grace period and when we notice | |
9160306e PM |
869 | * that someone else started the grace period. The caller must hold the |
870 | * ->lock of the leaf rcu_node structure corresponding to the current CPU, | |
871 | * and must have irqs disabled. | |
64db4cff | 872 | */ |
9160306e PM |
873 | static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) |
874 | { | |
875 | if (rdp->gpnum != rnp->gpnum) { | |
121dfc4b PM |
876 | /* |
877 | * If the current grace period is waiting for this CPU, | |
878 | * set up to detect a quiescent state, otherwise don't | |
879 | * go looking for one. | |
880 | */ | |
9160306e | 881 | rdp->gpnum = rnp->gpnum; |
d4c08f2a | 882 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart"); |
121dfc4b PM |
883 | if (rnp->qsmask & rdp->grpmask) { |
884 | rdp->qs_pending = 1; | |
e4cc1f22 | 885 | rdp->passed_quiesce = 0; |
c701d5d9 | 886 | } else { |
121dfc4b | 887 | rdp->qs_pending = 0; |
c701d5d9 | 888 | } |
a858af28 | 889 | zero_cpu_stall_ticks(rdp); |
9160306e PM |
890 | } |
891 | } | |
892 | ||
64db4cff PM |
893 | static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp) |
894 | { | |
9160306e PM |
895 | unsigned long flags; |
896 | struct rcu_node *rnp; | |
897 | ||
898 | local_irq_save(flags); | |
899 | rnp = rdp->mynode; | |
900 | if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */ | |
1304afb2 | 901 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
9160306e PM |
902 | local_irq_restore(flags); |
903 | return; | |
904 | } | |
905 | __note_new_gpnum(rsp, rnp, rdp); | |
1304afb2 | 906 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
907 | } |
908 | ||
909 | /* | |
910 | * Did someone else start a new RCU grace period start since we last | |
911 | * checked? Update local state appropriately if so. Must be called | |
912 | * on the CPU corresponding to rdp. | |
913 | */ | |
914 | static int | |
915 | check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp) | |
916 | { | |
917 | unsigned long flags; | |
918 | int ret = 0; | |
919 | ||
920 | local_irq_save(flags); | |
921 | if (rdp->gpnum != rsp->gpnum) { | |
922 | note_new_gpnum(rsp, rdp); | |
923 | ret = 1; | |
924 | } | |
925 | local_irq_restore(flags); | |
926 | return ret; | |
927 | } | |
928 | ||
3f5d3ea6 PM |
929 | /* |
930 | * Initialize the specified rcu_data structure's callback list to empty. | |
931 | */ | |
932 | static void init_callback_list(struct rcu_data *rdp) | |
933 | { | |
934 | int i; | |
935 | ||
936 | rdp->nxtlist = NULL; | |
937 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
938 | rdp->nxttail[i] = &rdp->nxtlist; | |
939 | } | |
940 | ||
d09b62df PM |
941 | /* |
942 | * Advance this CPU's callbacks, but only if the current grace period | |
943 | * has ended. This may be called only from the CPU to whom the rdp | |
944 | * belongs. In addition, the corresponding leaf rcu_node structure's | |
945 | * ->lock must be held by the caller, with irqs disabled. | |
946 | */ | |
947 | static void | |
948 | __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) | |
949 | { | |
950 | /* Did another grace period end? */ | |
951 | if (rdp->completed != rnp->completed) { | |
952 | ||
953 | /* Advance callbacks. No harm if list empty. */ | |
954 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL]; | |
955 | rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL]; | |
956 | rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; | |
957 | ||
958 | /* Remember that we saw this grace-period completion. */ | |
959 | rdp->completed = rnp->completed; | |
d4c08f2a | 960 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend"); |
20377f32 | 961 | |
5ff8e6f0 FW |
962 | /* |
963 | * If we were in an extended quiescent state, we may have | |
121dfc4b | 964 | * missed some grace periods that others CPUs handled on |
5ff8e6f0 | 965 | * our behalf. Catch up with this state to avoid noting |
121dfc4b PM |
966 | * spurious new grace periods. If another grace period |
967 | * has started, then rnp->gpnum will have advanced, so | |
968 | * we will detect this later on. | |
5ff8e6f0 | 969 | */ |
121dfc4b | 970 | if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) |
5ff8e6f0 FW |
971 | rdp->gpnum = rdp->completed; |
972 | ||
20377f32 | 973 | /* |
121dfc4b PM |
974 | * If RCU does not need a quiescent state from this CPU, |
975 | * then make sure that this CPU doesn't go looking for one. | |
20377f32 | 976 | */ |
121dfc4b | 977 | if ((rnp->qsmask & rdp->grpmask) == 0) |
20377f32 | 978 | rdp->qs_pending = 0; |
d09b62df PM |
979 | } |
980 | } | |
981 | ||
982 | /* | |
983 | * Advance this CPU's callbacks, but only if the current grace period | |
984 | * has ended. This may be called only from the CPU to whom the rdp | |
985 | * belongs. | |
986 | */ | |
987 | static void | |
988 | rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp) | |
989 | { | |
990 | unsigned long flags; | |
991 | struct rcu_node *rnp; | |
992 | ||
993 | local_irq_save(flags); | |
994 | rnp = rdp->mynode; | |
995 | if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */ | |
1304afb2 | 996 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
d09b62df PM |
997 | local_irq_restore(flags); |
998 | return; | |
999 | } | |
1000 | __rcu_process_gp_end(rsp, rnp, rdp); | |
1304afb2 | 1001 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
d09b62df PM |
1002 | } |
1003 | ||
1004 | /* | |
1005 | * Do per-CPU grace-period initialization for running CPU. The caller | |
1006 | * must hold the lock of the leaf rcu_node structure corresponding to | |
1007 | * this CPU. | |
1008 | */ | |
1009 | static void | |
1010 | rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) | |
1011 | { | |
1012 | /* Prior grace period ended, so advance callbacks for current CPU. */ | |
1013 | __rcu_process_gp_end(rsp, rnp, rdp); | |
1014 | ||
1015 | /* | |
1016 | * Because this CPU just now started the new grace period, we know | |
1017 | * that all of its callbacks will be covered by this upcoming grace | |
1018 | * period, even the ones that were registered arbitrarily recently. | |
1019 | * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL. | |
1020 | * | |
1021 | * Other CPUs cannot be sure exactly when the grace period started. | |
1022 | * Therefore, their recently registered callbacks must pass through | |
1023 | * an additional RCU_NEXT_READY stage, so that they will be handled | |
1024 | * by the next RCU grace period. | |
1025 | */ | |
1026 | rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1027 | rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; | |
9160306e PM |
1028 | |
1029 | /* Set state so that this CPU will detect the next quiescent state. */ | |
1030 | __note_new_gpnum(rsp, rnp, rdp); | |
d09b62df PM |
1031 | } |
1032 | ||
64db4cff PM |
1033 | /* |
1034 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
1035 | * in preparation for detecting the next grace period. The caller must hold | |
1036 | * the root node's ->lock, which is released before return. Hard irqs must | |
1037 | * be disabled. | |
e5601400 PM |
1038 | * |
1039 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
1040 | * invoke this function. This can happen when the dying CPU reports its | |
1041 | * quiescent state. | |
64db4cff PM |
1042 | */ |
1043 | static void | |
1044 | rcu_start_gp(struct rcu_state *rsp, unsigned long flags) | |
1045 | __releases(rcu_get_root(rsp)->lock) | |
1046 | { | |
394f99a9 | 1047 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); |
64db4cff | 1048 | struct rcu_node *rnp = rcu_get_root(rsp); |
64db4cff | 1049 | |
037067a1 | 1050 | if (!rcu_scheduler_fully_active || |
afe24b12 PM |
1051 | !cpu_needs_another_gp(rsp, rdp)) { |
1052 | /* | |
1053 | * Either the scheduler hasn't yet spawned the first | |
1054 | * non-idle task or this CPU does not need another | |
1055 | * grace period. Either way, don't start a new grace | |
1056 | * period. | |
1057 | */ | |
1058 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1059 | return; | |
1060 | } | |
b32e9eb6 | 1061 | |
afe24b12 | 1062 | if (rsp->fqs_active) { |
b32e9eb6 | 1063 | /* |
afe24b12 PM |
1064 | * This CPU needs a grace period, but force_quiescent_state() |
1065 | * is running. Tell it to start one on this CPU's behalf. | |
b32e9eb6 | 1066 | */ |
afe24b12 PM |
1067 | rsp->fqs_need_gp = 1; |
1068 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff PM |
1069 | return; |
1070 | } | |
1071 | ||
1072 | /* Advance to a new grace period and initialize state. */ | |
1073 | rsp->gpnum++; | |
d4c08f2a | 1074 | trace_rcu_grace_period(rsp->name, rsp->gpnum, "start"); |
af446b70 PM |
1075 | WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT); |
1076 | rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */ | |
64db4cff | 1077 | rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS; |
64db4cff | 1078 | record_gp_stall_check_time(rsp); |
1304afb2 | 1079 | raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */ |
64db4cff | 1080 | |
64db4cff | 1081 | /* Exclude any concurrent CPU-hotplug operations. */ |
1304afb2 | 1082 | raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */ |
64db4cff PM |
1083 | |
1084 | /* | |
b835db1f PM |
1085 | * Set the quiescent-state-needed bits in all the rcu_node |
1086 | * structures for all currently online CPUs in breadth-first | |
1087 | * order, starting from the root rcu_node structure. This | |
1088 | * operation relies on the layout of the hierarchy within the | |
1089 | * rsp->node[] array. Note that other CPUs will access only | |
1090 | * the leaves of the hierarchy, which still indicate that no | |
1091 | * grace period is in progress, at least until the corresponding | |
1092 | * leaf node has been initialized. In addition, we have excluded | |
1093 | * CPU-hotplug operations. | |
64db4cff PM |
1094 | * |
1095 | * Note that the grace period cannot complete until we finish | |
1096 | * the initialization process, as there will be at least one | |
1097 | * qsmask bit set in the root node until that time, namely the | |
b835db1f PM |
1098 | * one corresponding to this CPU, due to the fact that we have |
1099 | * irqs disabled. | |
64db4cff | 1100 | */ |
a0b6c9a7 | 1101 | rcu_for_each_node_breadth_first(rsp, rnp) { |
1304afb2 | 1102 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
b0e165c0 | 1103 | rcu_preempt_check_blocked_tasks(rnp); |
49e29126 | 1104 | rnp->qsmask = rnp->qsmaskinit; |
de078d87 | 1105 | rnp->gpnum = rsp->gpnum; |
d09b62df PM |
1106 | rnp->completed = rsp->completed; |
1107 | if (rnp == rdp->mynode) | |
1108 | rcu_start_gp_per_cpu(rsp, rnp, rdp); | |
27f4d280 | 1109 | rcu_preempt_boost_start_gp(rnp); |
d4c08f2a PM |
1110 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, |
1111 | rnp->level, rnp->grplo, | |
1112 | rnp->grphi, rnp->qsmask); | |
1304afb2 | 1113 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff PM |
1114 | } |
1115 | ||
83f5b01f | 1116 | rnp = rcu_get_root(rsp); |
1304afb2 | 1117 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
af446b70 | 1118 | rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */ |
1304afb2 PM |
1119 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
1120 | raw_spin_unlock_irqrestore(&rsp->onofflock, flags); | |
64db4cff PM |
1121 | } |
1122 | ||
f41d911f | 1123 | /* |
d3f6bad3 PM |
1124 | * Report a full set of quiescent states to the specified rcu_state |
1125 | * data structure. This involves cleaning up after the prior grace | |
1126 | * period and letting rcu_start_gp() start up the next grace period | |
1127 | * if one is needed. Note that the caller must hold rnp->lock, as | |
1128 | * required by rcu_start_gp(), which will release it. | |
f41d911f | 1129 | */ |
d3f6bad3 | 1130 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 1131 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 1132 | { |
15ba0ba8 | 1133 | unsigned long gp_duration; |
afe24b12 PM |
1134 | struct rcu_node *rnp = rcu_get_root(rsp); |
1135 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
15ba0ba8 | 1136 | |
fc2219d4 | 1137 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
0bbcc529 PM |
1138 | |
1139 | /* | |
1140 | * Ensure that all grace-period and pre-grace-period activity | |
1141 | * is seen before the assignment to rsp->completed. | |
1142 | */ | |
1143 | smp_mb(); /* See above block comment. */ | |
15ba0ba8 PM |
1144 | gp_duration = jiffies - rsp->gp_start; |
1145 | if (gp_duration > rsp->gp_max) | |
1146 | rsp->gp_max = gp_duration; | |
afe24b12 PM |
1147 | |
1148 | /* | |
1149 | * We know the grace period is complete, but to everyone else | |
1150 | * it appears to still be ongoing. But it is also the case | |
1151 | * that to everyone else it looks like there is nothing that | |
1152 | * they can do to advance the grace period. It is therefore | |
1153 | * safe for us to drop the lock in order to mark the grace | |
1154 | * period as completed in all of the rcu_node structures. | |
1155 | * | |
1156 | * But if this CPU needs another grace period, it will take | |
1157 | * care of this while initializing the next grace period. | |
1158 | * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL | |
1159 | * because the callbacks have not yet been advanced: Those | |
1160 | * callbacks are waiting on the grace period that just now | |
1161 | * completed. | |
1162 | */ | |
1163 | if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) { | |
1164 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1165 | ||
1166 | /* | |
1167 | * Propagate new ->completed value to rcu_node structures | |
1168 | * so that other CPUs don't have to wait until the start | |
1169 | * of the next grace period to process their callbacks. | |
1170 | */ | |
1171 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
1172 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
1173 | rnp->completed = rsp->gpnum; | |
1174 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1175 | } | |
1176 | rnp = rcu_get_root(rsp); | |
1177 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
1178 | } | |
1179 | ||
1180 | rsp->completed = rsp->gpnum; /* Declare the grace period complete. */ | |
d4c08f2a | 1181 | trace_rcu_grace_period(rsp->name, rsp->completed, "end"); |
af446b70 | 1182 | rsp->fqs_state = RCU_GP_IDLE; |
f41d911f PM |
1183 | rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */ |
1184 | } | |
1185 | ||
64db4cff | 1186 | /* |
d3f6bad3 PM |
1187 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
1188 | * Allows quiescent states for a group of CPUs to be reported at one go | |
1189 | * to the specified rcu_node structure, though all the CPUs in the group | |
1190 | * must be represented by the same rcu_node structure (which need not be | |
1191 | * a leaf rcu_node structure, though it often will be). That structure's | |
1192 | * lock must be held upon entry, and it is released before return. | |
64db4cff PM |
1193 | */ |
1194 | static void | |
d3f6bad3 PM |
1195 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
1196 | struct rcu_node *rnp, unsigned long flags) | |
64db4cff PM |
1197 | __releases(rnp->lock) |
1198 | { | |
28ecd580 PM |
1199 | struct rcu_node *rnp_c; |
1200 | ||
64db4cff PM |
1201 | /* Walk up the rcu_node hierarchy. */ |
1202 | for (;;) { | |
1203 | if (!(rnp->qsmask & mask)) { | |
1204 | ||
1205 | /* Our bit has already been cleared, so done. */ | |
1304afb2 | 1206 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1207 | return; |
1208 | } | |
1209 | rnp->qsmask &= ~mask; | |
d4c08f2a PM |
1210 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
1211 | mask, rnp->qsmask, rnp->level, | |
1212 | rnp->grplo, rnp->grphi, | |
1213 | !!rnp->gp_tasks); | |
27f4d280 | 1214 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
1215 | |
1216 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 1217 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1218 | return; |
1219 | } | |
1220 | mask = rnp->grpmask; | |
1221 | if (rnp->parent == NULL) { | |
1222 | ||
1223 | /* No more levels. Exit loop holding root lock. */ | |
1224 | ||
1225 | break; | |
1226 | } | |
1304afb2 | 1227 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 1228 | rnp_c = rnp; |
64db4cff | 1229 | rnp = rnp->parent; |
1304afb2 | 1230 | raw_spin_lock_irqsave(&rnp->lock, flags); |
28ecd580 | 1231 | WARN_ON_ONCE(rnp_c->qsmask); |
64db4cff PM |
1232 | } |
1233 | ||
1234 | /* | |
1235 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 1236 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 1237 | * to clean up and start the next grace period if one is needed. |
64db4cff | 1238 | */ |
d3f6bad3 | 1239 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
1240 | } |
1241 | ||
1242 | /* | |
d3f6bad3 PM |
1243 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
1244 | * structure. This must be either called from the specified CPU, or | |
1245 | * called when the specified CPU is known to be offline (and when it is | |
1246 | * also known that no other CPU is concurrently trying to help the offline | |
1247 | * CPU). The lastcomp argument is used to make sure we are still in the | |
1248 | * grace period of interest. We don't want to end the current grace period | |
1249 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
1250 | */ |
1251 | static void | |
e4cc1f22 | 1252 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp) |
64db4cff PM |
1253 | { |
1254 | unsigned long flags; | |
1255 | unsigned long mask; | |
1256 | struct rcu_node *rnp; | |
1257 | ||
1258 | rnp = rdp->mynode; | |
1304afb2 | 1259 | raw_spin_lock_irqsave(&rnp->lock, flags); |
e4cc1f22 | 1260 | if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) { |
64db4cff PM |
1261 | |
1262 | /* | |
e4cc1f22 PM |
1263 | * The grace period in which this quiescent state was |
1264 | * recorded has ended, so don't report it upwards. | |
1265 | * We will instead need a new quiescent state that lies | |
1266 | * within the current grace period. | |
64db4cff | 1267 | */ |
e4cc1f22 | 1268 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
1304afb2 | 1269 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1270 | return; |
1271 | } | |
1272 | mask = rdp->grpmask; | |
1273 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 1274 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1275 | } else { |
1276 | rdp->qs_pending = 0; | |
1277 | ||
1278 | /* | |
1279 | * This GP can't end until cpu checks in, so all of our | |
1280 | * callbacks can be processed during the next GP. | |
1281 | */ | |
64db4cff PM |
1282 | rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL]; |
1283 | ||
d3f6bad3 | 1284 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
64db4cff PM |
1285 | } |
1286 | } | |
1287 | ||
1288 | /* | |
1289 | * Check to see if there is a new grace period of which this CPU | |
1290 | * is not yet aware, and if so, set up local rcu_data state for it. | |
1291 | * Otherwise, see if this CPU has just passed through its first | |
1292 | * quiescent state for this grace period, and record that fact if so. | |
1293 | */ | |
1294 | static void | |
1295 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
1296 | { | |
1297 | /* If there is now a new grace period, record and return. */ | |
1298 | if (check_for_new_grace_period(rsp, rdp)) | |
1299 | return; | |
1300 | ||
1301 | /* | |
1302 | * Does this CPU still need to do its part for current grace period? | |
1303 | * If no, return and let the other CPUs do their part as well. | |
1304 | */ | |
1305 | if (!rdp->qs_pending) | |
1306 | return; | |
1307 | ||
1308 | /* | |
1309 | * Was there a quiescent state since the beginning of the grace | |
1310 | * period? If no, then exit and wait for the next call. | |
1311 | */ | |
e4cc1f22 | 1312 | if (!rdp->passed_quiesce) |
64db4cff PM |
1313 | return; |
1314 | ||
d3f6bad3 PM |
1315 | /* |
1316 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
1317 | * judge of that). | |
1318 | */ | |
e4cc1f22 | 1319 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum); |
64db4cff PM |
1320 | } |
1321 | ||
1322 | #ifdef CONFIG_HOTPLUG_CPU | |
1323 | ||
e74f4c45 | 1324 | /* |
b1420f1c PM |
1325 | * Send the specified CPU's RCU callbacks to the orphanage. The |
1326 | * specified CPU must be offline, and the caller must hold the | |
1327 | * ->onofflock. | |
e74f4c45 | 1328 | */ |
b1420f1c PM |
1329 | static void |
1330 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
1331 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 1332 | { |
b1420f1c PM |
1333 | /* |
1334 | * Orphan the callbacks. First adjust the counts. This is safe | |
1335 | * because ->onofflock excludes _rcu_barrier()'s adoption of | |
1336 | * the callbacks, thus no memory barrier is required. | |
1337 | */ | |
a50c3af9 | 1338 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1339 | rsp->qlen_lazy += rdp->qlen_lazy; |
1340 | rsp->qlen += rdp->qlen; | |
1341 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 1342 | rdp->qlen_lazy = 0; |
1d1fb395 | 1343 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
1344 | } |
1345 | ||
1346 | /* | |
b1420f1c PM |
1347 | * Next, move those callbacks still needing a grace period to |
1348 | * the orphanage, where some other CPU will pick them up. | |
1349 | * Some of the callbacks might have gone partway through a grace | |
1350 | * period, but that is too bad. They get to start over because we | |
1351 | * cannot assume that grace periods are synchronized across CPUs. | |
1352 | * We don't bother updating the ->nxttail[] array yet, instead | |
1353 | * we just reset the whole thing later on. | |
a50c3af9 | 1354 | */ |
b1420f1c PM |
1355 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
1356 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1357 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
1358 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
1359 | } |
1360 | ||
1361 | /* | |
b1420f1c PM |
1362 | * Then move the ready-to-invoke callbacks to the orphanage, |
1363 | * where some other CPU will pick them up. These will not be | |
1364 | * required to pass though another grace period: They are done. | |
a50c3af9 | 1365 | */ |
e5601400 | 1366 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1367 | *rsp->orphan_donetail = rdp->nxtlist; |
1368 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 1369 | } |
e74f4c45 | 1370 | |
b1420f1c | 1371 | /* Finally, initialize the rcu_data structure's list to empty. */ |
3f5d3ea6 | 1372 | init_callback_list(rdp); |
b1420f1c PM |
1373 | } |
1374 | ||
1375 | /* | |
1376 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
1377 | * orphanage. The caller must hold the ->onofflock. | |
1378 | */ | |
1379 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp) | |
1380 | { | |
1381 | int i; | |
1382 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); | |
1383 | ||
a50c3af9 | 1384 | /* |
b1420f1c PM |
1385 | * If there is an rcu_barrier() operation in progress, then |
1386 | * only the task doing that operation is permitted to adopt | |
1387 | * callbacks. To do otherwise breaks rcu_barrier() and friends | |
1388 | * by causing them to fail to wait for the callbacks in the | |
1389 | * orphanage. | |
a50c3af9 | 1390 | */ |
b1420f1c PM |
1391 | if (rsp->rcu_barrier_in_progress && |
1392 | rsp->rcu_barrier_in_progress != current) | |
1393 | return; | |
1394 | ||
1395 | /* Do the accounting first. */ | |
1396 | rdp->qlen_lazy += rsp->qlen_lazy; | |
1397 | rdp->qlen += rsp->qlen; | |
1398 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
1399 | if (rsp->qlen_lazy != rsp->qlen) |
1400 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
1401 | rsp->qlen_lazy = 0; |
1402 | rsp->qlen = 0; | |
1403 | ||
1404 | /* | |
1405 | * We do not need a memory barrier here because the only way we | |
1406 | * can get here if there is an rcu_barrier() in flight is if | |
1407 | * we are the task doing the rcu_barrier(). | |
1408 | */ | |
1409 | ||
1410 | /* First adopt the ready-to-invoke callbacks. */ | |
1411 | if (rsp->orphan_donelist != NULL) { | |
1412 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1413 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
1414 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
1415 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
1416 | rdp->nxttail[i] = rsp->orphan_donetail; | |
1417 | rsp->orphan_donelist = NULL; | |
1418 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
1419 | } | |
1420 | ||
1421 | /* And then adopt the callbacks that still need a grace period. */ | |
1422 | if (rsp->orphan_nxtlist != NULL) { | |
1423 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
1424 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
1425 | rsp->orphan_nxtlist = NULL; | |
1426 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
1427 | } | |
1428 | } | |
1429 | ||
1430 | /* | |
1431 | * Trace the fact that this CPU is going offline. | |
1432 | */ | |
1433 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
1434 | { | |
1435 | RCU_TRACE(unsigned long mask); | |
1436 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
1437 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
1438 | ||
1439 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
1440 | trace_rcu_grace_period(rsp->name, |
1441 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
1442 | "cpuofl"); | |
64db4cff PM |
1443 | } |
1444 | ||
1445 | /* | |
e5601400 | 1446 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
1447 | * this fact from process context. Do the remainder of the cleanup, |
1448 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1449 | * adopting them, if there is no _rcu_barrier() instance running. | |
e5601400 PM |
1450 | * There can only be one CPU hotplug operation at a time, so no other |
1451 | * CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 1452 | */ |
e5601400 | 1453 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 1454 | { |
2036d94a PM |
1455 | unsigned long flags; |
1456 | unsigned long mask; | |
1457 | int need_report = 0; | |
e5601400 | 1458 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 1459 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 1460 | |
2036d94a | 1461 | /* Adjust any no-longer-needed kthreads. */ |
e5601400 PM |
1462 | rcu_stop_cpu_kthread(cpu); |
1463 | rcu_node_kthread_setaffinity(rnp, -1); | |
2036d94a | 1464 | |
b1420f1c | 1465 | /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */ |
2036d94a PM |
1466 | |
1467 | /* Exclude any attempts to start a new grace period. */ | |
1468 | raw_spin_lock_irqsave(&rsp->onofflock, flags); | |
1469 | ||
b1420f1c PM |
1470 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
1471 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
1472 | rcu_adopt_orphan_cbs(rsp); | |
1473 | ||
2036d94a PM |
1474 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
1475 | mask = rdp->grpmask; /* rnp->grplo is constant. */ | |
1476 | do { | |
1477 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
1478 | rnp->qsmaskinit &= ~mask; | |
1479 | if (rnp->qsmaskinit != 0) { | |
1480 | if (rnp != rdp->mynode) | |
1481 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1482 | break; | |
1483 | } | |
1484 | if (rnp == rdp->mynode) | |
1485 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); | |
1486 | else | |
1487 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1488 | mask = rnp->grpmask; | |
1489 | rnp = rnp->parent; | |
1490 | } while (rnp != NULL); | |
1491 | ||
1492 | /* | |
1493 | * We still hold the leaf rcu_node structure lock here, and | |
1494 | * irqs are still disabled. The reason for this subterfuge is | |
1495 | * because invoking rcu_report_unblock_qs_rnp() with ->onofflock | |
1496 | * held leads to deadlock. | |
1497 | */ | |
1498 | raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */ | |
1499 | rnp = rdp->mynode; | |
1500 | if (need_report & RCU_OFL_TASKS_NORM_GP) | |
1501 | rcu_report_unblock_qs_rnp(rnp, flags); | |
1502 | else | |
1503 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1504 | if (need_report & RCU_OFL_TASKS_EXP_GP) | |
1505 | rcu_report_exp_rnp(rsp, rnp, true); | |
cf01537e PM |
1506 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
1507 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
1508 | cpu, rdp->qlen, rdp->nxtlist); | |
64db4cff PM |
1509 | } |
1510 | ||
1511 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1512 | ||
b1420f1c PM |
1513 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp) |
1514 | { | |
1515 | } | |
1516 | ||
e5601400 | 1517 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
1518 | { |
1519 | } | |
1520 | ||
e5601400 | 1521 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
1522 | { |
1523 | } | |
1524 | ||
1525 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
1526 | ||
1527 | /* | |
1528 | * Invoke any RCU callbacks that have made it to the end of their grace | |
1529 | * period. Thottle as specified by rdp->blimit. | |
1530 | */ | |
37c72e56 | 1531 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1532 | { |
1533 | unsigned long flags; | |
1534 | struct rcu_head *next, *list, **tail; | |
b41772ab | 1535 | int bl, count, count_lazy, i; |
64db4cff PM |
1536 | |
1537 | /* If no callbacks are ready, just return.*/ | |
29c00b4a | 1538 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 1539 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
1540 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
1541 | need_resched(), is_idle_task(current), | |
1542 | rcu_is_callbacks_kthread()); | |
64db4cff | 1543 | return; |
29c00b4a | 1544 | } |
64db4cff PM |
1545 | |
1546 | /* | |
1547 | * Extract the list of ready callbacks, disabling to prevent | |
1548 | * races with call_rcu() from interrupt handlers. | |
1549 | */ | |
1550 | local_irq_save(flags); | |
8146c4e2 | 1551 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 1552 | bl = rdp->blimit; |
486e2593 | 1553 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
1554 | list = rdp->nxtlist; |
1555 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
1556 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
1557 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
1558 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
1559 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
1560 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
1561 | local_irq_restore(flags); |
1562 | ||
1563 | /* Invoke callbacks. */ | |
486e2593 | 1564 | count = count_lazy = 0; |
64db4cff PM |
1565 | while (list) { |
1566 | next = list->next; | |
1567 | prefetch(next); | |
551d55a9 | 1568 | debug_rcu_head_unqueue(list); |
486e2593 PM |
1569 | if (__rcu_reclaim(rsp->name, list)) |
1570 | count_lazy++; | |
64db4cff | 1571 | list = next; |
dff1672d PM |
1572 | /* Stop only if limit reached and CPU has something to do. */ |
1573 | if (++count >= bl && | |
1574 | (need_resched() || | |
1575 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
1576 | break; |
1577 | } | |
1578 | ||
1579 | local_irq_save(flags); | |
4968c300 PM |
1580 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
1581 | is_idle_task(current), | |
1582 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
1583 | |
1584 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
1585 | if (list != NULL) { |
1586 | *tail = rdp->nxtlist; | |
1587 | rdp->nxtlist = list; | |
b41772ab PM |
1588 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
1589 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
1590 | rdp->nxttail[i] = tail; | |
64db4cff PM |
1591 | else |
1592 | break; | |
1593 | } | |
b1420f1c PM |
1594 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
1595 | rdp->qlen_lazy -= count_lazy; | |
1d1fb395 | 1596 | ACCESS_ONCE(rdp->qlen) -= count; |
b1420f1c | 1597 | rdp->n_cbs_invoked += count; |
64db4cff PM |
1598 | |
1599 | /* Reinstate batch limit if we have worked down the excess. */ | |
1600 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
1601 | rdp->blimit = blimit; | |
1602 | ||
37c72e56 PM |
1603 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
1604 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
1605 | rdp->qlen_last_fqs_check = 0; | |
1606 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
1607 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
1608 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 1609 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 1610 | |
64db4cff PM |
1611 | local_irq_restore(flags); |
1612 | ||
e0f23060 | 1613 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 1614 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 1615 | invoke_rcu_core(); |
64db4cff PM |
1616 | } |
1617 | ||
1618 | /* | |
1619 | * Check to see if this CPU is in a non-context-switch quiescent state | |
1620 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 1621 | * Also schedule RCU core processing. |
64db4cff | 1622 | * |
9b2e4f18 | 1623 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
1624 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
1625 | * false, there is no point in invoking rcu_check_callbacks(). | |
1626 | */ | |
1627 | void rcu_check_callbacks(int cpu, int user) | |
1628 | { | |
300df91c | 1629 | trace_rcu_utilization("Start scheduler-tick"); |
a858af28 | 1630 | increment_cpu_stall_ticks(); |
9b2e4f18 | 1631 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
1632 | |
1633 | /* | |
1634 | * Get here if this CPU took its interrupt from user | |
1635 | * mode or from the idle loop, and if this is not a | |
1636 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 1637 | * a quiescent state, so note it. |
64db4cff PM |
1638 | * |
1639 | * No memory barrier is required here because both | |
d6714c22 PM |
1640 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
1641 | * variables that other CPUs neither access nor modify, | |
1642 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
1643 | */ |
1644 | ||
d6714c22 PM |
1645 | rcu_sched_qs(cpu); |
1646 | rcu_bh_qs(cpu); | |
64db4cff PM |
1647 | |
1648 | } else if (!in_softirq()) { | |
1649 | ||
1650 | /* | |
1651 | * Get here if this CPU did not take its interrupt from | |
1652 | * softirq, in other words, if it is not interrupting | |
1653 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 1654 | * critical section, so note it. |
64db4cff PM |
1655 | */ |
1656 | ||
d6714c22 | 1657 | rcu_bh_qs(cpu); |
64db4cff | 1658 | } |
f41d911f | 1659 | rcu_preempt_check_callbacks(cpu); |
d21670ac | 1660 | if (rcu_pending(cpu)) |
a46e0899 | 1661 | invoke_rcu_core(); |
300df91c | 1662 | trace_rcu_utilization("End scheduler-tick"); |
64db4cff PM |
1663 | } |
1664 | ||
64db4cff PM |
1665 | /* |
1666 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
1667 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
1668 | * Also initiate boosting for any threads blocked on the root rcu_node. |
1669 | * | |
ee47eb9f | 1670 | * The caller must have suppressed start of new grace periods. |
64db4cff | 1671 | */ |
45f014c5 | 1672 | static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *)) |
64db4cff PM |
1673 | { |
1674 | unsigned long bit; | |
1675 | int cpu; | |
1676 | unsigned long flags; | |
1677 | unsigned long mask; | |
a0b6c9a7 | 1678 | struct rcu_node *rnp; |
64db4cff | 1679 | |
a0b6c9a7 | 1680 | rcu_for_each_leaf_node(rsp, rnp) { |
64db4cff | 1681 | mask = 0; |
1304afb2 | 1682 | raw_spin_lock_irqsave(&rnp->lock, flags); |
ee47eb9f | 1683 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 1684 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 1685 | return; |
64db4cff | 1686 | } |
a0b6c9a7 | 1687 | if (rnp->qsmask == 0) { |
1217ed1b | 1688 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
64db4cff PM |
1689 | continue; |
1690 | } | |
a0b6c9a7 | 1691 | cpu = rnp->grplo; |
64db4cff | 1692 | bit = 1; |
a0b6c9a7 | 1693 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
394f99a9 LJ |
1694 | if ((rnp->qsmask & bit) != 0 && |
1695 | f(per_cpu_ptr(rsp->rda, cpu))) | |
64db4cff PM |
1696 | mask |= bit; |
1697 | } | |
45f014c5 | 1698 | if (mask != 0) { |
64db4cff | 1699 | |
d3f6bad3 PM |
1700 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
1701 | rcu_report_qs_rnp(mask, rsp, rnp, flags); | |
64db4cff PM |
1702 | continue; |
1703 | } | |
1304afb2 | 1704 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1705 | } |
27f4d280 | 1706 | rnp = rcu_get_root(rsp); |
1217ed1b PM |
1707 | if (rnp->qsmask == 0) { |
1708 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1709 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ | |
1710 | } | |
64db4cff PM |
1711 | } |
1712 | ||
1713 | /* | |
1714 | * Force quiescent states on reluctant CPUs, and also detect which | |
1715 | * CPUs are in dyntick-idle mode. | |
1716 | */ | |
1717 | static void force_quiescent_state(struct rcu_state *rsp, int relaxed) | |
1718 | { | |
1719 | unsigned long flags; | |
64db4cff | 1720 | struct rcu_node *rnp = rcu_get_root(rsp); |
64db4cff | 1721 | |
300df91c PM |
1722 | trace_rcu_utilization("Start fqs"); |
1723 | if (!rcu_gp_in_progress(rsp)) { | |
1724 | trace_rcu_utilization("End fqs"); | |
64db4cff | 1725 | return; /* No grace period in progress, nothing to force. */ |
300df91c | 1726 | } |
1304afb2 | 1727 | if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) { |
64db4cff | 1728 | rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */ |
300df91c | 1729 | trace_rcu_utilization("End fqs"); |
64db4cff PM |
1730 | return; /* Someone else is already on the job. */ |
1731 | } | |
20133cfc | 1732 | if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies)) |
f96e9232 | 1733 | goto unlock_fqs_ret; /* no emergency and done recently. */ |
64db4cff | 1734 | rsp->n_force_qs++; |
1304afb2 | 1735 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
64db4cff | 1736 | rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS; |
560d4bc0 | 1737 | if(!rcu_gp_in_progress(rsp)) { |
64db4cff | 1738 | rsp->n_force_qs_ngp++; |
1304afb2 | 1739 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
f96e9232 | 1740 | goto unlock_fqs_ret; /* no GP in progress, time updated. */ |
64db4cff | 1741 | } |
07079d53 | 1742 | rsp->fqs_active = 1; |
af446b70 | 1743 | switch (rsp->fqs_state) { |
83f5b01f | 1744 | case RCU_GP_IDLE: |
64db4cff PM |
1745 | case RCU_GP_INIT: |
1746 | ||
83f5b01f | 1747 | break; /* grace period idle or initializing, ignore. */ |
64db4cff PM |
1748 | |
1749 | case RCU_SAVE_DYNTICK: | |
64db4cff | 1750 | |
f261414f LJ |
1751 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
1752 | ||
64db4cff | 1753 | /* Record dyntick-idle state. */ |
45f014c5 | 1754 | force_qs_rnp(rsp, dyntick_save_progress_counter); |
1304afb2 | 1755 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
ee47eb9f | 1756 | if (rcu_gp_in_progress(rsp)) |
af446b70 | 1757 | rsp->fqs_state = RCU_FORCE_QS; |
ee47eb9f | 1758 | break; |
64db4cff PM |
1759 | |
1760 | case RCU_FORCE_QS: | |
1761 | ||
1762 | /* Check dyntick-idle state, send IPI to laggarts. */ | |
1304afb2 | 1763 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
45f014c5 | 1764 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs); |
64db4cff PM |
1765 | |
1766 | /* Leave state in case more forcing is required. */ | |
1767 | ||
1304afb2 | 1768 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
f96e9232 | 1769 | break; |
64db4cff | 1770 | } |
07079d53 | 1771 | rsp->fqs_active = 0; |
46a1e34e | 1772 | if (rsp->fqs_need_gp) { |
1304afb2 | 1773 | raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */ |
46a1e34e PM |
1774 | rsp->fqs_need_gp = 0; |
1775 | rcu_start_gp(rsp, flags); /* releases rnp->lock */ | |
300df91c | 1776 | trace_rcu_utilization("End fqs"); |
46a1e34e PM |
1777 | return; |
1778 | } | |
1304afb2 | 1779 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
f96e9232 | 1780 | unlock_fqs_ret: |
1304afb2 | 1781 | raw_spin_unlock_irqrestore(&rsp->fqslock, flags); |
300df91c | 1782 | trace_rcu_utilization("End fqs"); |
64db4cff PM |
1783 | } |
1784 | ||
64db4cff | 1785 | /* |
e0f23060 PM |
1786 | * This does the RCU core processing work for the specified rcu_state |
1787 | * and rcu_data structures. This may be called only from the CPU to | |
1788 | * whom the rdp belongs. | |
64db4cff PM |
1789 | */ |
1790 | static void | |
1bca8cf1 | 1791 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
1792 | { |
1793 | unsigned long flags; | |
1bca8cf1 | 1794 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
64db4cff | 1795 | |
2e597558 PM |
1796 | WARN_ON_ONCE(rdp->beenonline == 0); |
1797 | ||
64db4cff PM |
1798 | /* |
1799 | * If an RCU GP has gone long enough, go check for dyntick | |
1800 | * idle CPUs and, if needed, send resched IPIs. | |
1801 | */ | |
20133cfc | 1802 | if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) |
64db4cff PM |
1803 | force_quiescent_state(rsp, 1); |
1804 | ||
1805 | /* | |
1806 | * Advance callbacks in response to end of earlier grace | |
1807 | * period that some other CPU ended. | |
1808 | */ | |
1809 | rcu_process_gp_end(rsp, rdp); | |
1810 | ||
1811 | /* Update RCU state based on any recent quiescent states. */ | |
1812 | rcu_check_quiescent_state(rsp, rdp); | |
1813 | ||
1814 | /* Does this CPU require a not-yet-started grace period? */ | |
1815 | if (cpu_needs_another_gp(rsp, rdp)) { | |
1304afb2 | 1816 | raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags); |
64db4cff PM |
1817 | rcu_start_gp(rsp, flags); /* releases above lock */ |
1818 | } | |
1819 | ||
1820 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 1821 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 1822 | invoke_rcu_callbacks(rsp, rdp); |
09223371 SL |
1823 | } |
1824 | ||
64db4cff | 1825 | /* |
e0f23060 | 1826 | * Do RCU core processing for the current CPU. |
64db4cff | 1827 | */ |
09223371 | 1828 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 1829 | { |
6ce75a23 PM |
1830 | struct rcu_state *rsp; |
1831 | ||
300df91c | 1832 | trace_rcu_utilization("Start RCU core"); |
6ce75a23 PM |
1833 | for_each_rcu_flavor(rsp) |
1834 | __rcu_process_callbacks(rsp); | |
300df91c | 1835 | trace_rcu_utilization("End RCU core"); |
64db4cff PM |
1836 | } |
1837 | ||
a26ac245 | 1838 | /* |
e0f23060 PM |
1839 | * Schedule RCU callback invocation. If the specified type of RCU |
1840 | * does not support RCU priority boosting, just do a direct call, | |
1841 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
1842 | * are running on the current CPU with interrupts disabled, the | |
1843 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
a26ac245 | 1844 | */ |
a46e0899 | 1845 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 1846 | { |
b0d30417 PM |
1847 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
1848 | return; | |
a46e0899 PM |
1849 | if (likely(!rsp->boost)) { |
1850 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
1851 | return; |
1852 | } | |
a46e0899 | 1853 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
1854 | } |
1855 | ||
a46e0899 | 1856 | static void invoke_rcu_core(void) |
09223371 SL |
1857 | { |
1858 | raise_softirq(RCU_SOFTIRQ); | |
1859 | } | |
1860 | ||
29154c57 PM |
1861 | /* |
1862 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
1863 | */ | |
1864 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
1865 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 1866 | { |
62fde6ed PM |
1867 | /* |
1868 | * If called from an extended quiescent state, invoke the RCU | |
1869 | * core in order to force a re-evaluation of RCU's idleness. | |
1870 | */ | |
a16b7a69 | 1871 | if (rcu_is_cpu_idle() && cpu_online(smp_processor_id())) |
62fde6ed PM |
1872 | invoke_rcu_core(); |
1873 | ||
a16b7a69 | 1874 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 1875 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 1876 | return; |
64db4cff | 1877 | |
37c72e56 PM |
1878 | /* |
1879 | * Force the grace period if too many callbacks or too long waiting. | |
1880 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
1881 | * if some other CPU has recently done so. Also, don't bother | |
1882 | * invoking force_quiescent_state() if the newly enqueued callback | |
1883 | * is the only one waiting for a grace period to complete. | |
1884 | */ | |
2655d57e | 1885 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
1886 | |
1887 | /* Are we ignoring a completed grace period? */ | |
1888 | rcu_process_gp_end(rsp, rdp); | |
1889 | check_for_new_grace_period(rsp, rdp); | |
1890 | ||
1891 | /* Start a new grace period if one not already started. */ | |
1892 | if (!rcu_gp_in_progress(rsp)) { | |
1893 | unsigned long nestflag; | |
1894 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
1895 | ||
1896 | raw_spin_lock_irqsave(&rnp_root->lock, nestflag); | |
1897 | rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */ | |
1898 | } else { | |
1899 | /* Give the grace period a kick. */ | |
1900 | rdp->blimit = LONG_MAX; | |
1901 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
1902 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
1903 | force_quiescent_state(rsp, 0); | |
1904 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
1905 | rdp->qlen_last_fqs_check = rdp->qlen; | |
1906 | } | |
20133cfc | 1907 | } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) |
64db4cff | 1908 | force_quiescent_state(rsp, 1); |
29154c57 PM |
1909 | } |
1910 | ||
64db4cff PM |
1911 | static void |
1912 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
486e2593 | 1913 | struct rcu_state *rsp, bool lazy) |
64db4cff PM |
1914 | { |
1915 | unsigned long flags; | |
1916 | struct rcu_data *rdp; | |
1917 | ||
0bb7b59d | 1918 | WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */ |
551d55a9 | 1919 | debug_rcu_head_queue(head); |
64db4cff PM |
1920 | head->func = func; |
1921 | head->next = NULL; | |
1922 | ||
1923 | smp_mb(); /* Ensure RCU update seen before callback registry. */ | |
1924 | ||
1925 | /* | |
1926 | * Opportunistically note grace-period endings and beginnings. | |
1927 | * Note that we might see a beginning right after we see an | |
1928 | * end, but never vice versa, since this CPU has to pass through | |
1929 | * a quiescent state betweentimes. | |
1930 | */ | |
1931 | local_irq_save(flags); | |
394f99a9 | 1932 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
1933 | |
1934 | /* Add the callback to our list. */ | |
29154c57 | 1935 | ACCESS_ONCE(rdp->qlen)++; |
486e2593 PM |
1936 | if (lazy) |
1937 | rdp->qlen_lazy++; | |
c57afe80 PM |
1938 | else |
1939 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
1940 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
1941 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
1942 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 1943 | |
d4c08f2a PM |
1944 | if (__is_kfree_rcu_offset((unsigned long)func)) |
1945 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 1946 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 1947 | else |
486e2593 | 1948 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 1949 | |
29154c57 PM |
1950 | /* Go handle any RCU core processing required. */ |
1951 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
1952 | local_irq_restore(flags); |
1953 | } | |
1954 | ||
1955 | /* | |
d6714c22 | 1956 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 1957 | */ |
d6714c22 | 1958 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 1959 | { |
486e2593 | 1960 | __call_rcu(head, func, &rcu_sched_state, 0); |
64db4cff | 1961 | } |
d6714c22 | 1962 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
1963 | |
1964 | /* | |
486e2593 | 1965 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
1966 | */ |
1967 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
1968 | { | |
486e2593 | 1969 | __call_rcu(head, func, &rcu_bh_state, 0); |
64db4cff PM |
1970 | } |
1971 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
1972 | ||
6d813391 PM |
1973 | /* |
1974 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
1975 | * any blocking grace-period wait automatically implies a grace period | |
1976 | * if there is only one CPU online at any point time during execution | |
1977 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
1978 | * occasionally incorrectly indicate that there are multiple CPUs online | |
1979 | * when there was in fact only one the whole time, as this just adds | |
1980 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
1981 | */ |
1982 | static inline int rcu_blocking_is_gp(void) | |
1983 | { | |
95f0c1de PM |
1984 | int ret; |
1985 | ||
6d813391 | 1986 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
1987 | preempt_disable(); |
1988 | ret = num_online_cpus() <= 1; | |
1989 | preempt_enable(); | |
1990 | return ret; | |
6d813391 PM |
1991 | } |
1992 | ||
6ebb237b PM |
1993 | /** |
1994 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
1995 | * | |
1996 | * Control will return to the caller some time after a full rcu-sched | |
1997 | * grace period has elapsed, in other words after all currently executing | |
1998 | * rcu-sched read-side critical sections have completed. These read-side | |
1999 | * critical sections are delimited by rcu_read_lock_sched() and | |
2000 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
2001 | * local_irq_disable(), and so on may be used in place of | |
2002 | * rcu_read_lock_sched(). | |
2003 | * | |
2004 | * This means that all preempt_disable code sequences, including NMI and | |
2005 | * hardware-interrupt handlers, in progress on entry will have completed | |
2006 | * before this primitive returns. However, this does not guarantee that | |
2007 | * softirq handlers will have completed, since in some kernels, these | |
2008 | * handlers can run in process context, and can block. | |
2009 | * | |
2010 | * This primitive provides the guarantees made by the (now removed) | |
2011 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
2012 | * guarantees that rcu_read_lock() sections will have completed. | |
2013 | * In "classic RCU", these two guarantees happen to be one and | |
2014 | * the same, but can differ in realtime RCU implementations. | |
2015 | */ | |
2016 | void synchronize_sched(void) | |
2017 | { | |
fe15d706 PM |
2018 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2019 | !lock_is_held(&rcu_lock_map) && | |
2020 | !lock_is_held(&rcu_sched_lock_map), | |
2021 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
2022 | if (rcu_blocking_is_gp()) |
2023 | return; | |
2c42818e | 2024 | wait_rcu_gp(call_rcu_sched); |
6ebb237b PM |
2025 | } |
2026 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
2027 | ||
2028 | /** | |
2029 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
2030 | * | |
2031 | * Control will return to the caller some time after a full rcu_bh grace | |
2032 | * period has elapsed, in other words after all currently executing rcu_bh | |
2033 | * read-side critical sections have completed. RCU read-side critical | |
2034 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
2035 | * and may be nested. | |
2036 | */ | |
2037 | void synchronize_rcu_bh(void) | |
2038 | { | |
fe15d706 PM |
2039 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2040 | !lock_is_held(&rcu_lock_map) && | |
2041 | !lock_is_held(&rcu_sched_lock_map), | |
2042 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
2043 | if (rcu_blocking_is_gp()) |
2044 | return; | |
2c42818e | 2045 | wait_rcu_gp(call_rcu_bh); |
6ebb237b PM |
2046 | } |
2047 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
2048 | ||
3d3b7db0 PM |
2049 | static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0); |
2050 | static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0); | |
2051 | ||
2052 | static int synchronize_sched_expedited_cpu_stop(void *data) | |
2053 | { | |
2054 | /* | |
2055 | * There must be a full memory barrier on each affected CPU | |
2056 | * between the time that try_stop_cpus() is called and the | |
2057 | * time that it returns. | |
2058 | * | |
2059 | * In the current initial implementation of cpu_stop, the | |
2060 | * above condition is already met when the control reaches | |
2061 | * this point and the following smp_mb() is not strictly | |
2062 | * necessary. Do smp_mb() anyway for documentation and | |
2063 | * robustness against future implementation changes. | |
2064 | */ | |
2065 | smp_mb(); /* See above comment block. */ | |
2066 | return 0; | |
2067 | } | |
2068 | ||
236fefaf PM |
2069 | /** |
2070 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
2071 | * | |
2072 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
2073 | * approach to force the grace period to end quickly. This consumes | |
2074 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
2075 | * so is thus not recommended for any sort of common-case code. In fact, | |
2076 | * if you are using synchronize_sched_expedited() in a loop, please | |
2077 | * restructure your code to batch your updates, and then use a single | |
2078 | * synchronize_sched() instead. | |
3d3b7db0 | 2079 | * |
236fefaf PM |
2080 | * Note that it is illegal to call this function while holding any lock |
2081 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal | |
2082 | * to call this function from a CPU-hotplug notifier. Failing to observe | |
2083 | * these restriction will result in deadlock. | |
3d3b7db0 PM |
2084 | * |
2085 | * This implementation can be thought of as an application of ticket | |
2086 | * locking to RCU, with sync_sched_expedited_started and | |
2087 | * sync_sched_expedited_done taking on the roles of the halves | |
2088 | * of the ticket-lock word. Each task atomically increments | |
2089 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
2090 | * then attempts to stop all the CPUs. If this succeeds, then each | |
2091 | * CPU will have executed a context switch, resulting in an RCU-sched | |
2092 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
2093 | * update sync_sched_expedited_done to match our snapshot -- but | |
2094 | * only if someone else has not already advanced past our snapshot. | |
2095 | * | |
2096 | * On the other hand, if try_stop_cpus() fails, we check the value | |
2097 | * of sync_sched_expedited_done. If it has advanced past our | |
2098 | * initial snapshot, then someone else must have forced a grace period | |
2099 | * some time after we took our snapshot. In this case, our work is | |
2100 | * done for us, and we can simply return. Otherwise, we try again, | |
2101 | * but keep our initial snapshot for purposes of checking for someone | |
2102 | * doing our work for us. | |
2103 | * | |
2104 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
2105 | */ | |
2106 | void synchronize_sched_expedited(void) | |
2107 | { | |
2108 | int firstsnap, s, snap, trycount = 0; | |
2109 | ||
2110 | /* Note that atomic_inc_return() implies full memory barrier. */ | |
2111 | firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started); | |
2112 | get_online_cpus(); | |
1cc85961 | 2113 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 PM |
2114 | |
2115 | /* | |
2116 | * Each pass through the following loop attempts to force a | |
2117 | * context switch on each CPU. | |
2118 | */ | |
2119 | while (try_stop_cpus(cpu_online_mask, | |
2120 | synchronize_sched_expedited_cpu_stop, | |
2121 | NULL) == -EAGAIN) { | |
2122 | put_online_cpus(); | |
2123 | ||
2124 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 2125 | if (trycount++ < 10) { |
3d3b7db0 | 2126 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 2127 | } else { |
3d3b7db0 PM |
2128 | synchronize_sched(); |
2129 | return; | |
2130 | } | |
2131 | ||
2132 | /* Check to see if someone else did our work for us. */ | |
2133 | s = atomic_read(&sync_sched_expedited_done); | |
2134 | if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) { | |
2135 | smp_mb(); /* ensure test happens before caller kfree */ | |
2136 | return; | |
2137 | } | |
2138 | ||
2139 | /* | |
2140 | * Refetching sync_sched_expedited_started allows later | |
2141 | * callers to piggyback on our grace period. We subtract | |
2142 | * 1 to get the same token that the last incrementer got. | |
2143 | * We retry after they started, so our grace period works | |
2144 | * for them, and they started after our first try, so their | |
2145 | * grace period works for us. | |
2146 | */ | |
2147 | get_online_cpus(); | |
2148 | snap = atomic_read(&sync_sched_expedited_started); | |
2149 | smp_mb(); /* ensure read is before try_stop_cpus(). */ | |
2150 | } | |
2151 | ||
2152 | /* | |
2153 | * Everyone up to our most recent fetch is covered by our grace | |
2154 | * period. Update the counter, but only if our work is still | |
2155 | * relevant -- which it won't be if someone who started later | |
2156 | * than we did beat us to the punch. | |
2157 | */ | |
2158 | do { | |
2159 | s = atomic_read(&sync_sched_expedited_done); | |
2160 | if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) { | |
2161 | smp_mb(); /* ensure test happens before caller kfree */ | |
2162 | break; | |
2163 | } | |
2164 | } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s); | |
2165 | ||
2166 | put_online_cpus(); | |
2167 | } | |
2168 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
2169 | ||
64db4cff PM |
2170 | /* |
2171 | * Check to see if there is any immediate RCU-related work to be done | |
2172 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
2173 | * The checks are in order of increasing expense: checks that can be | |
2174 | * carried out against CPU-local state are performed first. However, | |
2175 | * we must check for CPU stalls first, else we might not get a chance. | |
2176 | */ | |
2177 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
2178 | { | |
2f51f988 PM |
2179 | struct rcu_node *rnp = rdp->mynode; |
2180 | ||
64db4cff PM |
2181 | rdp->n_rcu_pending++; |
2182 | ||
2183 | /* Check for CPU stalls, if enabled. */ | |
2184 | check_cpu_stall(rsp, rdp); | |
2185 | ||
2186 | /* Is the RCU core waiting for a quiescent state from this CPU? */ | |
5c51dd73 PM |
2187 | if (rcu_scheduler_fully_active && |
2188 | rdp->qs_pending && !rdp->passed_quiesce) { | |
d25eb944 PM |
2189 | |
2190 | /* | |
2191 | * If force_quiescent_state() coming soon and this CPU | |
2192 | * needs a quiescent state, and this is either RCU-sched | |
2193 | * or RCU-bh, force a local reschedule. | |
2194 | */ | |
d21670ac | 2195 | rdp->n_rp_qs_pending++; |
6cc68793 | 2196 | if (!rdp->preemptible && |
d25eb944 PM |
2197 | ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1, |
2198 | jiffies)) | |
2199 | set_need_resched(); | |
e4cc1f22 | 2200 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
d21670ac | 2201 | rdp->n_rp_report_qs++; |
64db4cff | 2202 | return 1; |
7ba5c840 | 2203 | } |
64db4cff PM |
2204 | |
2205 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
2206 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
2207 | rdp->n_rp_cb_ready++; | |
64db4cff | 2208 | return 1; |
7ba5c840 | 2209 | } |
64db4cff PM |
2210 | |
2211 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
2212 | if (cpu_needs_another_gp(rsp, rdp)) { |
2213 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 2214 | return 1; |
7ba5c840 | 2215 | } |
64db4cff PM |
2216 | |
2217 | /* Has another RCU grace period completed? */ | |
2f51f988 | 2218 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 2219 | rdp->n_rp_gp_completed++; |
64db4cff | 2220 | return 1; |
7ba5c840 | 2221 | } |
64db4cff PM |
2222 | |
2223 | /* Has a new RCU grace period started? */ | |
2f51f988 | 2224 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
7ba5c840 | 2225 | rdp->n_rp_gp_started++; |
64db4cff | 2226 | return 1; |
7ba5c840 | 2227 | } |
64db4cff PM |
2228 | |
2229 | /* Has an RCU GP gone long enough to send resched IPIs &c? */ | |
fc2219d4 | 2230 | if (rcu_gp_in_progress(rsp) && |
20133cfc | 2231 | ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) { |
7ba5c840 | 2232 | rdp->n_rp_need_fqs++; |
64db4cff | 2233 | return 1; |
7ba5c840 | 2234 | } |
64db4cff PM |
2235 | |
2236 | /* nothing to do */ | |
7ba5c840 | 2237 | rdp->n_rp_need_nothing++; |
64db4cff PM |
2238 | return 0; |
2239 | } | |
2240 | ||
2241 | /* | |
2242 | * Check to see if there is any immediate RCU-related work to be done | |
2243 | * by the current CPU, returning 1 if so. This function is part of the | |
2244 | * RCU implementation; it is -not- an exported member of the RCU API. | |
2245 | */ | |
a157229c | 2246 | static int rcu_pending(int cpu) |
64db4cff | 2247 | { |
6ce75a23 PM |
2248 | struct rcu_state *rsp; |
2249 | ||
2250 | for_each_rcu_flavor(rsp) | |
2251 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu))) | |
2252 | return 1; | |
2253 | return 0; | |
64db4cff PM |
2254 | } |
2255 | ||
2256 | /* | |
2257 | * Check to see if any future RCU-related work will need to be done | |
2258 | * by the current CPU, even if none need be done immediately, returning | |
8bd93a2c | 2259 | * 1 if so. |
64db4cff | 2260 | */ |
aea1b35e | 2261 | static int rcu_cpu_has_callbacks(int cpu) |
64db4cff | 2262 | { |
6ce75a23 PM |
2263 | struct rcu_state *rsp; |
2264 | ||
64db4cff | 2265 | /* RCU callbacks either ready or pending? */ |
6ce75a23 PM |
2266 | for_each_rcu_flavor(rsp) |
2267 | if (per_cpu_ptr(rsp->rda, cpu)->nxtlist) | |
2268 | return 1; | |
2269 | return 0; | |
64db4cff PM |
2270 | } |
2271 | ||
a83eff0a PM |
2272 | /* |
2273 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
2274 | * the compiler is expected to optimize this away. | |
2275 | */ | |
2276 | static void _rcu_barrier_trace(struct rcu_state *rsp, char *s, | |
2277 | int cpu, unsigned long done) | |
2278 | { | |
2279 | trace_rcu_barrier(rsp->name, s, cpu, | |
2280 | atomic_read(&rsp->barrier_cpu_count), done); | |
2281 | } | |
2282 | ||
b1420f1c PM |
2283 | /* |
2284 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
2285 | * up the task executing _rcu_barrier(). | |
2286 | */ | |
24ebbca8 | 2287 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 2288 | { |
24ebbca8 PM |
2289 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
2290 | struct rcu_state *rsp = rdp->rsp; | |
2291 | ||
a83eff0a PM |
2292 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
2293 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 2294 | complete(&rsp->barrier_completion); |
a83eff0a PM |
2295 | } else { |
2296 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
2297 | } | |
d0ec774c PM |
2298 | } |
2299 | ||
2300 | /* | |
2301 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
2302 | */ | |
2303 | static void rcu_barrier_func(void *type) | |
2304 | { | |
037b64ed | 2305 | struct rcu_state *rsp = type; |
06668efa | 2306 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
d0ec774c | 2307 | |
a83eff0a | 2308 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 2309 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 2310 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
2311 | } |
2312 | ||
d0ec774c PM |
2313 | /* |
2314 | * Orchestrate the specified type of RCU barrier, waiting for all | |
2315 | * RCU callbacks of the specified type to complete. | |
2316 | */ | |
037b64ed | 2317 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 2318 | { |
b1420f1c PM |
2319 | int cpu; |
2320 | unsigned long flags; | |
2321 | struct rcu_data *rdp; | |
24ebbca8 | 2322 | struct rcu_data rd; |
cf3a9c48 PM |
2323 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
2324 | unsigned long snap_done; | |
b1420f1c | 2325 | |
24ebbca8 | 2326 | init_rcu_head_on_stack(&rd.barrier_head); |
a83eff0a | 2327 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 2328 | |
e74f4c45 | 2329 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 2330 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 2331 | |
cf3a9c48 PM |
2332 | /* |
2333 | * Ensure that all prior references, including to ->n_barrier_done, | |
2334 | * are ordered before the _rcu_barrier() machinery. | |
2335 | */ | |
2336 | smp_mb(); /* See above block comment. */ | |
2337 | ||
2338 | /* | |
2339 | * Recheck ->n_barrier_done to see if others did our work for us. | |
2340 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
2341 | * transition. The "if" expression below therefore rounds the old | |
2342 | * value up to the next even number and adds two before comparing. | |
2343 | */ | |
2344 | snap_done = ACCESS_ONCE(rsp->n_barrier_done); | |
a83eff0a | 2345 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
cf3a9c48 | 2346 | if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) { |
a83eff0a | 2347 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
2348 | smp_mb(); /* caller's subsequent code after above check. */ |
2349 | mutex_unlock(&rsp->barrier_mutex); | |
2350 | return; | |
2351 | } | |
2352 | ||
2353 | /* | |
2354 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
2355 | * ACCESS_ONCE() to prevent the compiler from speculating | |
2356 | * the increment to precede the early-exit check. | |
2357 | */ | |
2358 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2359 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); | |
a83eff0a | 2360 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 2361 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 2362 | |
d0ec774c | 2363 | /* |
b1420f1c PM |
2364 | * Initialize the count to one rather than to zero in order to |
2365 | * avoid a too-soon return to zero in case of a short grace period | |
2366 | * (or preemption of this task). Also flag this task as doing | |
2367 | * an rcu_barrier(). This will prevent anyone else from adopting | |
2368 | * orphaned callbacks, which could cause otherwise failure if a | |
2369 | * CPU went offline and quickly came back online. To see this, | |
2370 | * consider the following sequence of events: | |
2371 | * | |
2372 | * 1. We cause CPU 0 to post an rcu_barrier_callback() callback. | |
2373 | * 2. CPU 1 goes offline, orphaning its callbacks. | |
2374 | * 3. CPU 0 adopts CPU 1's orphaned callbacks. | |
2375 | * 4. CPU 1 comes back online. | |
2376 | * 5. We cause CPU 1 to post an rcu_barrier_callback() callback. | |
2377 | * 6. Both rcu_barrier_callback() callbacks are invoked, awakening | |
2378 | * us -- but before CPU 1's orphaned callbacks are invoked!!! | |
d0ec774c | 2379 | */ |
7db74df8 | 2380 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 2381 | atomic_set(&rsp->barrier_cpu_count, 1); |
b1420f1c PM |
2382 | raw_spin_lock_irqsave(&rsp->onofflock, flags); |
2383 | rsp->rcu_barrier_in_progress = current; | |
2384 | raw_spin_unlock_irqrestore(&rsp->onofflock, flags); | |
2385 | ||
2386 | /* | |
2387 | * Force every CPU with callbacks to register a new callback | |
2388 | * that will tell us when all the preceding callbacks have | |
2389 | * been invoked. If an offline CPU has callbacks, wait for | |
2390 | * it to either come back online or to finish orphaning those | |
2391 | * callbacks. | |
2392 | */ | |
2393 | for_each_possible_cpu(cpu) { | |
2394 | preempt_disable(); | |
2395 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
2396 | if (cpu_is_offline(cpu)) { | |
a83eff0a PM |
2397 | _rcu_barrier_trace(rsp, "Offline", cpu, |
2398 | rsp->n_barrier_done); | |
b1420f1c PM |
2399 | preempt_enable(); |
2400 | while (cpu_is_offline(cpu) && ACCESS_ONCE(rdp->qlen)) | |
2401 | schedule_timeout_interruptible(1); | |
2402 | } else if (ACCESS_ONCE(rdp->qlen)) { | |
a83eff0a PM |
2403 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
2404 | rsp->n_barrier_done); | |
037b64ed | 2405 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c PM |
2406 | preempt_enable(); |
2407 | } else { | |
a83eff0a PM |
2408 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
2409 | rsp->n_barrier_done); | |
b1420f1c PM |
2410 | preempt_enable(); |
2411 | } | |
2412 | } | |
2413 | ||
2414 | /* | |
2415 | * Now that all online CPUs have rcu_barrier_callback() callbacks | |
2416 | * posted, we can adopt all of the orphaned callbacks and place | |
2417 | * an rcu_barrier_callback() callback after them. When that is done, | |
2418 | * we are guaranteed to have an rcu_barrier_callback() callback | |
2419 | * following every callback that could possibly have been | |
2420 | * registered before _rcu_barrier() was called. | |
2421 | */ | |
2422 | raw_spin_lock_irqsave(&rsp->onofflock, flags); | |
2423 | rcu_adopt_orphan_cbs(rsp); | |
2424 | rsp->rcu_barrier_in_progress = NULL; | |
2425 | raw_spin_unlock_irqrestore(&rsp->onofflock, flags); | |
24ebbca8 | 2426 | atomic_inc(&rsp->barrier_cpu_count); |
b1420f1c | 2427 | smp_mb__after_atomic_inc(); /* Ensure atomic_inc() before callback. */ |
24ebbca8 PM |
2428 | rd.rsp = rsp; |
2429 | rsp->call(&rd.barrier_head, rcu_barrier_callback); | |
b1420f1c PM |
2430 | |
2431 | /* | |
2432 | * Now that we have an rcu_barrier_callback() callback on each | |
2433 | * CPU, and thus each counted, remove the initial count. | |
2434 | */ | |
24ebbca8 | 2435 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 2436 | complete(&rsp->barrier_completion); |
b1420f1c | 2437 | |
cf3a9c48 PM |
2438 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
2439 | smp_mb(); /* Keep increment after above mechanism. */ | |
2440 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2441 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); | |
a83eff0a | 2442 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
2443 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
2444 | ||
b1420f1c | 2445 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 2446 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
2447 | |
2448 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 2449 | mutex_unlock(&rsp->barrier_mutex); |
b1420f1c | 2450 | |
24ebbca8 | 2451 | destroy_rcu_head_on_stack(&rd.barrier_head); |
d0ec774c | 2452 | } |
d0ec774c PM |
2453 | |
2454 | /** | |
2455 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
2456 | */ | |
2457 | void rcu_barrier_bh(void) | |
2458 | { | |
037b64ed | 2459 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
2460 | } |
2461 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
2462 | ||
2463 | /** | |
2464 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
2465 | */ | |
2466 | void rcu_barrier_sched(void) | |
2467 | { | |
037b64ed | 2468 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
2469 | } |
2470 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
2471 | ||
64db4cff | 2472 | /* |
27569620 | 2473 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 2474 | */ |
27569620 PM |
2475 | static void __init |
2476 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
2477 | { |
2478 | unsigned long flags; | |
394f99a9 | 2479 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
2480 | struct rcu_node *rnp = rcu_get_root(rsp); |
2481 | ||
2482 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 2483 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 2484 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
3f5d3ea6 | 2485 | init_callback_list(rdp); |
486e2593 | 2486 | rdp->qlen_lazy = 0; |
1d1fb395 | 2487 | ACCESS_ONCE(rdp->qlen) = 0; |
27569620 | 2488 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 2489 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 2490 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 2491 | rdp->cpu = cpu; |
d4c08f2a | 2492 | rdp->rsp = rsp; |
1304afb2 | 2493 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
2494 | } |
2495 | ||
2496 | /* | |
2497 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
2498 | * offline event can be happening at a given time. Note also that we | |
2499 | * can accept some slop in the rsp->completed access due to the fact | |
2500 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 2501 | */ |
e4fa4c97 | 2502 | static void __cpuinit |
6cc68793 | 2503 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible) |
64db4cff PM |
2504 | { |
2505 | unsigned long flags; | |
64db4cff | 2506 | unsigned long mask; |
394f99a9 | 2507 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
2508 | struct rcu_node *rnp = rcu_get_root(rsp); |
2509 | ||
2510 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 2511 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 2512 | rdp->beenonline = 1; /* We have now been online. */ |
6cc68793 | 2513 | rdp->preemptible = preemptible; |
37c72e56 PM |
2514 | rdp->qlen_last_fqs_check = 0; |
2515 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 2516 | rdp->blimit = blimit; |
29e37d81 | 2517 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
c92b131b PM |
2518 | atomic_set(&rdp->dynticks->dynticks, |
2519 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
7cb92499 | 2520 | rcu_prepare_for_idle_init(cpu); |
1304afb2 | 2521 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff PM |
2522 | |
2523 | /* | |
2524 | * A new grace period might start here. If so, we won't be part | |
2525 | * of it, but that is OK, as we are currently in a quiescent state. | |
2526 | */ | |
2527 | ||
2528 | /* Exclude any attempts to start a new GP on large systems. */ | |
1304afb2 | 2529 | raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */ |
64db4cff PM |
2530 | |
2531 | /* Add CPU to rcu_node bitmasks. */ | |
2532 | rnp = rdp->mynode; | |
2533 | mask = rdp->grpmask; | |
2534 | do { | |
2535 | /* Exclude any attempts to start a new GP on small systems. */ | |
1304afb2 | 2536 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
2537 | rnp->qsmaskinit |= mask; |
2538 | mask = rnp->grpmask; | |
d09b62df | 2539 | if (rnp == rdp->mynode) { |
06ae115a PM |
2540 | /* |
2541 | * If there is a grace period in progress, we will | |
2542 | * set up to wait for it next time we run the | |
2543 | * RCU core code. | |
2544 | */ | |
2545 | rdp->gpnum = rnp->completed; | |
d09b62df | 2546 | rdp->completed = rnp->completed; |
06ae115a PM |
2547 | rdp->passed_quiesce = 0; |
2548 | rdp->qs_pending = 0; | |
e4cc1f22 | 2549 | rdp->passed_quiesce_gpnum = rnp->gpnum - 1; |
d4c08f2a | 2550 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl"); |
d09b62df | 2551 | } |
1304afb2 | 2552 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
2553 | rnp = rnp->parent; |
2554 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); | |
2555 | ||
1304afb2 | 2556 | raw_spin_unlock_irqrestore(&rsp->onofflock, flags); |
64db4cff PM |
2557 | } |
2558 | ||
d72bce0e | 2559 | static void __cpuinit rcu_prepare_cpu(int cpu) |
64db4cff | 2560 | { |
6ce75a23 PM |
2561 | struct rcu_state *rsp; |
2562 | ||
2563 | for_each_rcu_flavor(rsp) | |
2564 | rcu_init_percpu_data(cpu, rsp, | |
2565 | strcmp(rsp->name, "rcu_preempt") == 0); | |
64db4cff PM |
2566 | } |
2567 | ||
2568 | /* | |
f41d911f | 2569 | * Handle CPU online/offline notification events. |
64db4cff | 2570 | */ |
9f680ab4 PM |
2571 | static int __cpuinit rcu_cpu_notify(struct notifier_block *self, |
2572 | unsigned long action, void *hcpu) | |
64db4cff PM |
2573 | { |
2574 | long cpu = (long)hcpu; | |
27f4d280 | 2575 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); |
a26ac245 | 2576 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 2577 | struct rcu_state *rsp; |
64db4cff | 2578 | |
300df91c | 2579 | trace_rcu_utilization("Start CPU hotplug"); |
64db4cff PM |
2580 | switch (action) { |
2581 | case CPU_UP_PREPARE: | |
2582 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
2583 | rcu_prepare_cpu(cpu); |
2584 | rcu_prepare_kthreads(cpu); | |
a26ac245 PM |
2585 | break; |
2586 | case CPU_ONLINE: | |
0f962a5e PM |
2587 | case CPU_DOWN_FAILED: |
2588 | rcu_node_kthread_setaffinity(rnp, -1); | |
e3995a25 | 2589 | rcu_cpu_kthread_setrt(cpu, 1); |
0f962a5e PM |
2590 | break; |
2591 | case CPU_DOWN_PREPARE: | |
2592 | rcu_node_kthread_setaffinity(rnp, cpu); | |
e3995a25 | 2593 | rcu_cpu_kthread_setrt(cpu, 0); |
64db4cff | 2594 | break; |
d0ec774c PM |
2595 | case CPU_DYING: |
2596 | case CPU_DYING_FROZEN: | |
2597 | /* | |
2d999e03 PM |
2598 | * The whole machine is "stopped" except this CPU, so we can |
2599 | * touch any data without introducing corruption. We send the | |
2600 | * dying CPU's callbacks to an arbitrarily chosen online CPU. | |
d0ec774c | 2601 | */ |
6ce75a23 PM |
2602 | for_each_rcu_flavor(rsp) |
2603 | rcu_cleanup_dying_cpu(rsp); | |
7cb92499 | 2604 | rcu_cleanup_after_idle(cpu); |
d0ec774c | 2605 | break; |
64db4cff PM |
2606 | case CPU_DEAD: |
2607 | case CPU_DEAD_FROZEN: | |
2608 | case CPU_UP_CANCELED: | |
2609 | case CPU_UP_CANCELED_FROZEN: | |
6ce75a23 PM |
2610 | for_each_rcu_flavor(rsp) |
2611 | rcu_cleanup_dead_cpu(cpu, rsp); | |
64db4cff PM |
2612 | break; |
2613 | default: | |
2614 | break; | |
2615 | } | |
300df91c | 2616 | trace_rcu_utilization("End CPU hotplug"); |
64db4cff PM |
2617 | return NOTIFY_OK; |
2618 | } | |
2619 | ||
bbad9379 PM |
2620 | /* |
2621 | * This function is invoked towards the end of the scheduler's initialization | |
2622 | * process. Before this is called, the idle task might contain | |
2623 | * RCU read-side critical sections (during which time, this idle | |
2624 | * task is booting the system). After this function is called, the | |
2625 | * idle tasks are prohibited from containing RCU read-side critical | |
2626 | * sections. This function also enables RCU lockdep checking. | |
2627 | */ | |
2628 | void rcu_scheduler_starting(void) | |
2629 | { | |
2630 | WARN_ON(num_online_cpus() != 1); | |
2631 | WARN_ON(nr_context_switches() > 0); | |
2632 | rcu_scheduler_active = 1; | |
2633 | } | |
2634 | ||
64db4cff PM |
2635 | /* |
2636 | * Compute the per-level fanout, either using the exact fanout specified | |
2637 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
2638 | */ | |
2639 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
2640 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
2641 | { | |
2642 | int i; | |
2643 | ||
f885b7f2 | 2644 | for (i = rcu_num_lvls - 1; i > 0; i--) |
64db4cff | 2645 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
f885b7f2 | 2646 | rsp->levelspread[0] = rcu_fanout_leaf; |
64db4cff PM |
2647 | } |
2648 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
2649 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
2650 | { | |
2651 | int ccur; | |
2652 | int cprv; | |
2653 | int i; | |
2654 | ||
2655 | cprv = NR_CPUS; | |
f885b7f2 | 2656 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
2657 | ccur = rsp->levelcnt[i]; |
2658 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
2659 | cprv = ccur; | |
2660 | } | |
2661 | } | |
2662 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
2663 | ||
2664 | /* | |
2665 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
2666 | */ | |
394f99a9 LJ |
2667 | static void __init rcu_init_one(struct rcu_state *rsp, |
2668 | struct rcu_data __percpu *rda) | |
64db4cff | 2669 | { |
b6407e86 PM |
2670 | static char *buf[] = { "rcu_node_level_0", |
2671 | "rcu_node_level_1", | |
2672 | "rcu_node_level_2", | |
2673 | "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */ | |
64db4cff PM |
2674 | int cpustride = 1; |
2675 | int i; | |
2676 | int j; | |
2677 | struct rcu_node *rnp; | |
2678 | ||
b6407e86 PM |
2679 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
2680 | ||
64db4cff PM |
2681 | /* Initialize the level-tracking arrays. */ |
2682 | ||
f885b7f2 PM |
2683 | for (i = 0; i < rcu_num_lvls; i++) |
2684 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
2685 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
2686 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
2687 | rcu_init_levelspread(rsp); | |
2688 | ||
2689 | /* Initialize the elements themselves, starting from the leaves. */ | |
2690 | ||
f885b7f2 | 2691 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
2692 | cpustride *= rsp->levelspread[i]; |
2693 | rnp = rsp->level[i]; | |
2694 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 2695 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
2696 | lockdep_set_class_and_name(&rnp->lock, |
2697 | &rcu_node_class[i], buf[i]); | |
f41d911f | 2698 | rnp->gpnum = 0; |
64db4cff PM |
2699 | rnp->qsmask = 0; |
2700 | rnp->qsmaskinit = 0; | |
2701 | rnp->grplo = j * cpustride; | |
2702 | rnp->grphi = (j + 1) * cpustride - 1; | |
2703 | if (rnp->grphi >= NR_CPUS) | |
2704 | rnp->grphi = NR_CPUS - 1; | |
2705 | if (i == 0) { | |
2706 | rnp->grpnum = 0; | |
2707 | rnp->grpmask = 0; | |
2708 | rnp->parent = NULL; | |
2709 | } else { | |
2710 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
2711 | rnp->grpmask = 1UL << rnp->grpnum; | |
2712 | rnp->parent = rsp->level[i - 1] + | |
2713 | j / rsp->levelspread[i - 1]; | |
2714 | } | |
2715 | rnp->level = i; | |
12f5f524 | 2716 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
64db4cff PM |
2717 | } |
2718 | } | |
0c34029a | 2719 | |
394f99a9 | 2720 | rsp->rda = rda; |
f885b7f2 | 2721 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 2722 | for_each_possible_cpu(i) { |
4a90a068 | 2723 | while (i > rnp->grphi) |
0c34029a | 2724 | rnp++; |
394f99a9 | 2725 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
2726 | rcu_boot_init_percpu_data(i, rsp); |
2727 | } | |
6ce75a23 | 2728 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
2729 | } |
2730 | ||
f885b7f2 PM |
2731 | /* |
2732 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
2733 | * replace the definitions in rcutree.h because those are needed to size | |
2734 | * the ->node array in the rcu_state structure. | |
2735 | */ | |
2736 | static void __init rcu_init_geometry(void) | |
2737 | { | |
2738 | int i; | |
2739 | int j; | |
cca6f393 | 2740 | int n = nr_cpu_ids; |
f885b7f2 PM |
2741 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
2742 | ||
2743 | /* If the compile-time values are accurate, just leave. */ | |
2744 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF) | |
2745 | return; | |
2746 | ||
2747 | /* | |
2748 | * Compute number of nodes that can be handled an rcu_node tree | |
2749 | * with the given number of levels. Setting rcu_capacity[0] makes | |
2750 | * some of the arithmetic easier. | |
2751 | */ | |
2752 | rcu_capacity[0] = 1; | |
2753 | rcu_capacity[1] = rcu_fanout_leaf; | |
2754 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
2755 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
2756 | ||
2757 | /* | |
2758 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
2759 | * to increase the leaf-level fanout, not decrease it. Of course, | |
2760 | * the leaf-level fanout cannot exceed the number of bits in | |
2761 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
2762 | * the configured number of CPUs. Complain and fall back to the | |
2763 | * compile-time values if these limits are exceeded. | |
2764 | */ | |
2765 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
2766 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
2767 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
2768 | WARN_ON(1); | |
2769 | return; | |
2770 | } | |
2771 | ||
2772 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
2773 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
2774 | if (n <= rcu_capacity[i]) { | |
2775 | for (j = 0; j <= i; j++) | |
2776 | num_rcu_lvl[j] = | |
2777 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
2778 | rcu_num_lvls = i; | |
2779 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
2780 | num_rcu_lvl[j] = 0; | |
2781 | break; | |
2782 | } | |
2783 | ||
2784 | /* Calculate the total number of rcu_node structures. */ | |
2785 | rcu_num_nodes = 0; | |
2786 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
2787 | rcu_num_nodes += num_rcu_lvl[i]; | |
2788 | rcu_num_nodes -= n; | |
2789 | } | |
2790 | ||
9f680ab4 | 2791 | void __init rcu_init(void) |
64db4cff | 2792 | { |
017c4261 | 2793 | int cpu; |
9f680ab4 | 2794 | |
f41d911f | 2795 | rcu_bootup_announce(); |
f885b7f2 | 2796 | rcu_init_geometry(); |
394f99a9 LJ |
2797 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
2798 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); | |
f41d911f | 2799 | __rcu_init_preempt(); |
09223371 | 2800 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
2801 | |
2802 | /* | |
2803 | * We don't need protection against CPU-hotplug here because | |
2804 | * this is called early in boot, before either interrupts | |
2805 | * or the scheduler are operational. | |
2806 | */ | |
2807 | cpu_notifier(rcu_cpu_notify, 0); | |
017c4261 PM |
2808 | for_each_online_cpu(cpu) |
2809 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
c68de209 | 2810 | check_cpu_stall_init(); |
64db4cff PM |
2811 | } |
2812 | ||
1eba8f84 | 2813 | #include "rcutree_plugin.h" |