Commit | Line | Data |
---|---|---|
64db4cff PM |
1 | /* |
2 | * Read-Copy Update mechanism for mutual exclusion | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public License | |
15 | * along with this program; if not, write to the Free Software | |
16 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. | |
17 | * | |
18 | * Copyright IBM Corporation, 2008 | |
19 | * | |
20 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> | |
21 | * Manfred Spraul <manfred@colorfullife.com> | |
22 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version | |
23 | * | |
24 | * Based on the original work by Paul McKenney <paulmck@us.ibm.com> | |
25 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. | |
26 | * | |
27 | * For detailed explanation of Read-Copy Update mechanism see - | |
a71fca58 | 28 | * Documentation/RCU |
64db4cff PM |
29 | */ |
30 | #include <linux/types.h> | |
31 | #include <linux/kernel.h> | |
32 | #include <linux/init.h> | |
33 | #include <linux/spinlock.h> | |
34 | #include <linux/smp.h> | |
35 | #include <linux/rcupdate.h> | |
36 | #include <linux/interrupt.h> | |
37 | #include <linux/sched.h> | |
c1dc0b9c | 38 | #include <linux/nmi.h> |
8826f3b0 | 39 | #include <linux/atomic.h> |
64db4cff | 40 | #include <linux/bitops.h> |
9984de1a | 41 | #include <linux/export.h> |
64db4cff PM |
42 | #include <linux/completion.h> |
43 | #include <linux/moduleparam.h> | |
44 | #include <linux/percpu.h> | |
45 | #include <linux/notifier.h> | |
46 | #include <linux/cpu.h> | |
47 | #include <linux/mutex.h> | |
48 | #include <linux/time.h> | |
bbad9379 | 49 | #include <linux/kernel_stat.h> |
a26ac245 PM |
50 | #include <linux/wait.h> |
51 | #include <linux/kthread.h> | |
268bb0ce | 52 | #include <linux/prefetch.h> |
3d3b7db0 PM |
53 | #include <linux/delay.h> |
54 | #include <linux/stop_machine.h> | |
661a85dc | 55 | #include <linux/random.h> |
64db4cff | 56 | |
9f77da9f | 57 | #include "rcutree.h" |
29c00b4a PM |
58 | #include <trace/events/rcu.h> |
59 | ||
60 | #include "rcu.h" | |
9f77da9f | 61 | |
64db4cff PM |
62 | /* Data structures. */ |
63 | ||
f885b7f2 | 64 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
394f2769 | 65 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; |
88b91c7c | 66 | |
037b64ed | 67 | #define RCU_STATE_INITIALIZER(sname, cr) { \ |
6c90cc7b | 68 | .level = { &sname##_state.node[0] }, \ |
037b64ed | 69 | .call = cr, \ |
af446b70 | 70 | .fqs_state = RCU_GP_IDLE, \ |
42c3533e PM |
71 | .gpnum = 0UL - 300UL, \ |
72 | .completed = 0UL - 300UL, \ | |
7b2e6011 | 73 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
6c90cc7b PM |
74 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
75 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 76 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
a4fbe35a | 77 | .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \ |
6c90cc7b | 78 | .name = #sname, \ |
64db4cff PM |
79 | } |
80 | ||
037b64ed PM |
81 | struct rcu_state rcu_sched_state = |
82 | RCU_STATE_INITIALIZER(rcu_sched, call_rcu_sched); | |
d6714c22 | 83 | DEFINE_PER_CPU(struct rcu_data, rcu_sched_data); |
64db4cff | 84 | |
037b64ed | 85 | struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, call_rcu_bh); |
6258c4fb | 86 | DEFINE_PER_CPU(struct rcu_data, rcu_bh_data); |
b1f77b05 | 87 | |
27f4d280 | 88 | static struct rcu_state *rcu_state; |
6ce75a23 | 89 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 90 | |
f885b7f2 PM |
91 | /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */ |
92 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF; | |
7e5c2dfb | 93 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 PM |
94 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
95 | static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */ | |
96 | NUM_RCU_LVL_0, | |
97 | NUM_RCU_LVL_1, | |
98 | NUM_RCU_LVL_2, | |
99 | NUM_RCU_LVL_3, | |
100 | NUM_RCU_LVL_4, | |
101 | }; | |
102 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ | |
103 | ||
b0d30417 PM |
104 | /* |
105 | * The rcu_scheduler_active variable transitions from zero to one just | |
106 | * before the first task is spawned. So when this variable is zero, RCU | |
107 | * can assume that there is but one task, allowing RCU to (for example) | |
108 | * optimized synchronize_sched() to a simple barrier(). When this variable | |
109 | * is one, RCU must actually do all the hard work required to detect real | |
110 | * grace periods. This variable is also used to suppress boot-time false | |
111 | * positives from lockdep-RCU error checking. | |
112 | */ | |
bbad9379 PM |
113 | int rcu_scheduler_active __read_mostly; |
114 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
115 | ||
b0d30417 PM |
116 | /* |
117 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
118 | * during the early_initcall() processing, which is after the scheduler | |
119 | * is capable of creating new tasks. So RCU processing (for example, | |
120 | * creating tasks for RCU priority boosting) must be delayed until after | |
121 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
122 | * currently delay invocation of any RCU callbacks until after this point. | |
123 | * | |
124 | * It might later prove better for people registering RCU callbacks during | |
125 | * early boot to take responsibility for these callbacks, but one step at | |
126 | * a time. | |
127 | */ | |
128 | static int rcu_scheduler_fully_active __read_mostly; | |
129 | ||
a46e0899 PM |
130 | #ifdef CONFIG_RCU_BOOST |
131 | ||
a26ac245 PM |
132 | /* |
133 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
134 | * handle all flavors of RCU. | |
135 | */ | |
136 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
d71df90e | 137 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); |
5ece5bab | 138 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); |
d71df90e | 139 | DEFINE_PER_CPU(char, rcu_cpu_has_work); |
a26ac245 | 140 | |
a46e0899 PM |
141 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
142 | ||
5d01bbd1 | 143 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
144 | static void invoke_rcu_core(void); |
145 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 146 | |
4a298656 PM |
147 | /* |
148 | * Track the rcutorture test sequence number and the update version | |
149 | * number within a given test. The rcutorture_testseq is incremented | |
150 | * on every rcutorture module load and unload, so has an odd value | |
151 | * when a test is running. The rcutorture_vernum is set to zero | |
152 | * when rcutorture starts and is incremented on each rcutorture update. | |
153 | * These variables enable correlating rcutorture output with the | |
154 | * RCU tracing information. | |
155 | */ | |
156 | unsigned long rcutorture_testseq; | |
157 | unsigned long rcutorture_vernum; | |
158 | ||
fc2219d4 PM |
159 | /* |
160 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
161 | * permit this function to be invoked without holding the root rcu_node | |
162 | * structure's ->lock, but of course results can be subject to change. | |
163 | */ | |
164 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
165 | { | |
166 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
167 | } | |
168 | ||
b1f77b05 | 169 | /* |
d6714c22 | 170 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 171 | * how many quiescent states passed, just if there was at least |
d6714c22 | 172 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 173 | * The caller must have disabled preemption. |
b1f77b05 | 174 | */ |
d6714c22 | 175 | void rcu_sched_qs(int cpu) |
b1f77b05 | 176 | { |
25502a6c | 177 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu); |
f41d911f | 178 | |
e4cc1f22 | 179 | if (rdp->passed_quiesce == 0) |
d4c08f2a | 180 | trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs"); |
e4cc1f22 | 181 | rdp->passed_quiesce = 1; |
b1f77b05 IM |
182 | } |
183 | ||
d6714c22 | 184 | void rcu_bh_qs(int cpu) |
b1f77b05 | 185 | { |
25502a6c | 186 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu); |
f41d911f | 187 | |
e4cc1f22 | 188 | if (rdp->passed_quiesce == 0) |
d4c08f2a | 189 | trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs"); |
e4cc1f22 | 190 | rdp->passed_quiesce = 1; |
b1f77b05 | 191 | } |
64db4cff | 192 | |
25502a6c PM |
193 | /* |
194 | * Note a context switch. This is a quiescent state for RCU-sched, | |
195 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 196 | * The caller must have disabled preemption. |
25502a6c PM |
197 | */ |
198 | void rcu_note_context_switch(int cpu) | |
199 | { | |
300df91c | 200 | trace_rcu_utilization("Start context switch"); |
25502a6c | 201 | rcu_sched_qs(cpu); |
cba6d0d6 | 202 | rcu_preempt_note_context_switch(cpu); |
300df91c | 203 | trace_rcu_utilization("End context switch"); |
25502a6c | 204 | } |
29ce8310 | 205 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 206 | |
90a4d2c0 | 207 | DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { |
29e37d81 | 208 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, |
23b5c8fa | 209 | .dynticks = ATOMIC_INIT(1), |
90a4d2c0 | 210 | }; |
64db4cff | 211 | |
878d7439 ED |
212 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
213 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
214 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 215 | |
878d7439 ED |
216 | module_param(blimit, long, 0444); |
217 | module_param(qhimark, long, 0444); | |
218 | module_param(qlowmark, long, 0444); | |
3d76c082 | 219 | |
13cfcca0 PM |
220 | int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */ |
221 | int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT; | |
222 | ||
f2e0dd70 | 223 | module_param(rcu_cpu_stall_suppress, int, 0644); |
13cfcca0 | 224 | module_param(rcu_cpu_stall_timeout, int, 0644); |
742734ee | 225 | |
d40011f6 PM |
226 | static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS; |
227 | static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS; | |
228 | ||
229 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
230 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
231 | ||
4cdfc175 PM |
232 | static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *)); |
233 | static void force_quiescent_state(struct rcu_state *rsp); | |
a157229c | 234 | static int rcu_pending(int cpu); |
64db4cff PM |
235 | |
236 | /* | |
d6714c22 | 237 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
64db4cff | 238 | */ |
d6714c22 | 239 | long rcu_batches_completed_sched(void) |
64db4cff | 240 | { |
d6714c22 | 241 | return rcu_sched_state.completed; |
64db4cff | 242 | } |
d6714c22 | 243 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
244 | |
245 | /* | |
246 | * Return the number of RCU BH batches processed thus far for debug & stats. | |
247 | */ | |
248 | long rcu_batches_completed_bh(void) | |
249 | { | |
250 | return rcu_bh_state.completed; | |
251 | } | |
252 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
253 | ||
bf66f18e PM |
254 | /* |
255 | * Force a quiescent state for RCU BH. | |
256 | */ | |
257 | void rcu_bh_force_quiescent_state(void) | |
258 | { | |
4cdfc175 | 259 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
260 | } |
261 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
262 | ||
4a298656 PM |
263 | /* |
264 | * Record the number of times rcutorture tests have been initiated and | |
265 | * terminated. This information allows the debugfs tracing stats to be | |
266 | * correlated to the rcutorture messages, even when the rcutorture module | |
267 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
268 | * store this state in rcutorture itself. | |
269 | */ | |
270 | void rcutorture_record_test_transition(void) | |
271 | { | |
272 | rcutorture_testseq++; | |
273 | rcutorture_vernum = 0; | |
274 | } | |
275 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
276 | ||
277 | /* | |
278 | * Record the number of writer passes through the current rcutorture test. | |
279 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
280 | * messages. | |
281 | */ | |
282 | void rcutorture_record_progress(unsigned long vernum) | |
283 | { | |
284 | rcutorture_vernum++; | |
285 | } | |
286 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
287 | ||
bf66f18e PM |
288 | /* |
289 | * Force a quiescent state for RCU-sched. | |
290 | */ | |
291 | void rcu_sched_force_quiescent_state(void) | |
292 | { | |
4cdfc175 | 293 | force_quiescent_state(&rcu_sched_state); |
bf66f18e PM |
294 | } |
295 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
296 | ||
64db4cff PM |
297 | /* |
298 | * Does the CPU have callbacks ready to be invoked? | |
299 | */ | |
300 | static int | |
301 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
302 | { | |
3fbfbf7a PM |
303 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
304 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
305 | } |
306 | ||
307 | /* | |
dc35c893 PM |
308 | * Does the current CPU require a not-yet-started grace period? |
309 | * The caller must have disabled interrupts to prevent races with | |
310 | * normal callback registry. | |
64db4cff PM |
311 | */ |
312 | static int | |
313 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
314 | { | |
dc35c893 | 315 | int i; |
3fbfbf7a | 316 | |
dc35c893 PM |
317 | if (rcu_gp_in_progress(rsp)) |
318 | return 0; /* No, a grace period is already in progress. */ | |
319 | if (!rdp->nxttail[RCU_NEXT_TAIL]) | |
320 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
321 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
322 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
323 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
324 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
325 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), | |
326 | rdp->nxtcompleted[i])) | |
327 | return 1; /* Yes, CBs for future grace period. */ | |
328 | return 0; /* No grace period needed. */ | |
64db4cff PM |
329 | } |
330 | ||
331 | /* | |
332 | * Return the root node of the specified rcu_state structure. | |
333 | */ | |
334 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
335 | { | |
336 | return &rsp->node[0]; | |
337 | } | |
338 | ||
9b2e4f18 | 339 | /* |
adf5091e | 340 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
341 | * |
342 | * If the new value of the ->dynticks_nesting counter now is zero, | |
343 | * we really have entered idle, and must do the appropriate accounting. | |
344 | * The caller must have disabled interrupts. | |
345 | */ | |
adf5091e FW |
346 | static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval, |
347 | bool user) | |
9b2e4f18 | 348 | { |
facc4e15 | 349 | trace_rcu_dyntick("Start", oldval, 0); |
cb349ca9 | 350 | if (!user && !is_idle_task(current)) { |
0989cb46 PM |
351 | struct task_struct *idle = idle_task(smp_processor_id()); |
352 | ||
facc4e15 | 353 | trace_rcu_dyntick("Error on entry: not idle task", oldval, 0); |
bf1304e9 | 354 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
355 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
356 | current->pid, current->comm, | |
357 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 358 | } |
aea1b35e | 359 | rcu_prepare_for_idle(smp_processor_id()); |
9b2e4f18 PM |
360 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
361 | smp_mb__before_atomic_inc(); /* See above. */ | |
362 | atomic_inc(&rdtp->dynticks); | |
363 | smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */ | |
364 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
c44e2cdd PM |
365 | |
366 | /* | |
adf5091e | 367 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
368 | * in an RCU read-side critical section. |
369 | */ | |
370 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
371 | "Illegal idle entry in RCU read-side critical section."); | |
372 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
373 | "Illegal idle entry in RCU-bh read-side critical section."); | |
374 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
375 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 376 | } |
64db4cff | 377 | |
adf5091e FW |
378 | /* |
379 | * Enter an RCU extended quiescent state, which can be either the | |
380 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 381 | */ |
adf5091e | 382 | static void rcu_eqs_enter(bool user) |
64db4cff | 383 | { |
4145fa7f | 384 | long long oldval; |
64db4cff PM |
385 | struct rcu_dynticks *rdtp; |
386 | ||
64db4cff | 387 | rdtp = &__get_cpu_var(rcu_dynticks); |
4145fa7f | 388 | oldval = rdtp->dynticks_nesting; |
29e37d81 PM |
389 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
390 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) | |
391 | rdtp->dynticks_nesting = 0; | |
392 | else | |
393 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; | |
adf5091e | 394 | rcu_eqs_enter_common(rdtp, oldval, user); |
64db4cff | 395 | } |
adf5091e FW |
396 | |
397 | /** | |
398 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
399 | * | |
400 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
401 | * read-side critical sections can occur. (Though RCU read-side | |
402 | * critical sections can occur in irq handlers in idle, a possibility | |
403 | * handled by irq_enter() and irq_exit().) | |
404 | * | |
405 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
406 | * the possibility of usermode upcalls having messed up our count | |
407 | * of interrupt nesting level during the prior busy period. | |
408 | */ | |
409 | void rcu_idle_enter(void) | |
410 | { | |
c5d900bf FW |
411 | unsigned long flags; |
412 | ||
413 | local_irq_save(flags); | |
cb349ca9 | 414 | rcu_eqs_enter(false); |
c5d900bf | 415 | local_irq_restore(flags); |
adf5091e | 416 | } |
8a2ecf47 | 417 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 418 | |
2b1d5024 | 419 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
420 | /** |
421 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
422 | * | |
423 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
424 | * is permitted between this call and rcu_user_exit(). This way the | |
425 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
426 | * when the CPU runs in userspace. | |
427 | */ | |
428 | void rcu_user_enter(void) | |
429 | { | |
91d1aa43 | 430 | rcu_eqs_enter(1); |
adf5091e FW |
431 | } |
432 | ||
19dd1591 FW |
433 | /** |
434 | * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace | |
435 | * after the current irq returns. | |
436 | * | |
437 | * This is similar to rcu_user_enter() but in the context of a non-nesting | |
438 | * irq. After this call, RCU enters into idle mode when the interrupt | |
439 | * returns. | |
440 | */ | |
441 | void rcu_user_enter_after_irq(void) | |
442 | { | |
443 | unsigned long flags; | |
444 | struct rcu_dynticks *rdtp; | |
445 | ||
446 | local_irq_save(flags); | |
447 | rdtp = &__get_cpu_var(rcu_dynticks); | |
448 | /* Ensure this irq is interrupting a non-idle RCU state. */ | |
449 | WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK)); | |
450 | rdtp->dynticks_nesting = 1; | |
451 | local_irq_restore(flags); | |
452 | } | |
2b1d5024 | 453 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 454 | |
9b2e4f18 PM |
455 | /** |
456 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
457 | * | |
458 | * Exit from an interrupt handler, which might possibly result in entering | |
459 | * idle mode, in other words, leaving the mode in which read-side critical | |
460 | * sections can occur. | |
64db4cff | 461 | * |
9b2e4f18 PM |
462 | * This code assumes that the idle loop never does anything that might |
463 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
464 | * architecture violates this assumption, RCU will give you what you | |
465 | * deserve, good and hard. But very infrequently and irreproducibly. | |
466 | * | |
467 | * Use things like work queues to work around this limitation. | |
468 | * | |
469 | * You have been warned. | |
64db4cff | 470 | */ |
9b2e4f18 | 471 | void rcu_irq_exit(void) |
64db4cff PM |
472 | { |
473 | unsigned long flags; | |
4145fa7f | 474 | long long oldval; |
64db4cff PM |
475 | struct rcu_dynticks *rdtp; |
476 | ||
477 | local_irq_save(flags); | |
478 | rdtp = &__get_cpu_var(rcu_dynticks); | |
4145fa7f | 479 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
480 | rdtp->dynticks_nesting--; |
481 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 FW |
482 | if (rdtp->dynticks_nesting) |
483 | trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting); | |
484 | else | |
cb349ca9 | 485 | rcu_eqs_enter_common(rdtp, oldval, true); |
9b2e4f18 PM |
486 | local_irq_restore(flags); |
487 | } | |
488 | ||
489 | /* | |
adf5091e | 490 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
491 | * |
492 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
493 | * we really have exited idle, and must do the appropriate accounting. | |
494 | * The caller must have disabled interrupts. | |
495 | */ | |
adf5091e FW |
496 | static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval, |
497 | int user) | |
9b2e4f18 | 498 | { |
23b5c8fa PM |
499 | smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */ |
500 | atomic_inc(&rdtp->dynticks); | |
501 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
502 | smp_mb__after_atomic_inc(); /* See above. */ | |
503 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
7cb92499 | 504 | rcu_cleanup_after_idle(smp_processor_id()); |
4145fa7f | 505 | trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting); |
cb349ca9 | 506 | if (!user && !is_idle_task(current)) { |
0989cb46 PM |
507 | struct task_struct *idle = idle_task(smp_processor_id()); |
508 | ||
4145fa7f PM |
509 | trace_rcu_dyntick("Error on exit: not idle task", |
510 | oldval, rdtp->dynticks_nesting); | |
bf1304e9 | 511 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
512 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
513 | current->pid, current->comm, | |
514 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
515 | } |
516 | } | |
517 | ||
adf5091e FW |
518 | /* |
519 | * Exit an RCU extended quiescent state, which can be either the | |
520 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 521 | */ |
adf5091e | 522 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 523 | { |
9b2e4f18 PM |
524 | struct rcu_dynticks *rdtp; |
525 | long long oldval; | |
526 | ||
9b2e4f18 PM |
527 | rdtp = &__get_cpu_var(rcu_dynticks); |
528 | oldval = rdtp->dynticks_nesting; | |
29e37d81 PM |
529 | WARN_ON_ONCE(oldval < 0); |
530 | if (oldval & DYNTICK_TASK_NEST_MASK) | |
531 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; | |
532 | else | |
533 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; | |
adf5091e | 534 | rcu_eqs_exit_common(rdtp, oldval, user); |
9b2e4f18 | 535 | } |
adf5091e FW |
536 | |
537 | /** | |
538 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
539 | * | |
540 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
541 | * read-side critical sections can occur. | |
542 | * | |
543 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
544 | * allow for the possibility of usermode upcalls messing up our count | |
545 | * of interrupt nesting level during the busy period that is just | |
546 | * now starting. | |
547 | */ | |
548 | void rcu_idle_exit(void) | |
549 | { | |
c5d900bf FW |
550 | unsigned long flags; |
551 | ||
552 | local_irq_save(flags); | |
cb349ca9 | 553 | rcu_eqs_exit(false); |
c5d900bf | 554 | local_irq_restore(flags); |
adf5091e | 555 | } |
8a2ecf47 | 556 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 557 | |
2b1d5024 | 558 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
559 | /** |
560 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
561 | * | |
562 | * Exit RCU idle mode while entering the kernel because it can | |
563 | * run a RCU read side critical section anytime. | |
564 | */ | |
565 | void rcu_user_exit(void) | |
566 | { | |
91d1aa43 | 567 | rcu_eqs_exit(1); |
adf5091e FW |
568 | } |
569 | ||
19dd1591 FW |
570 | /** |
571 | * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace | |
572 | * idle mode after the current non-nesting irq returns. | |
573 | * | |
574 | * This is similar to rcu_user_exit() but in the context of an irq. | |
575 | * This is called when the irq has interrupted a userspace RCU idle mode | |
576 | * context. When the current non-nesting interrupt returns after this call, | |
577 | * the CPU won't restore the RCU idle mode. | |
578 | */ | |
579 | void rcu_user_exit_after_irq(void) | |
580 | { | |
581 | unsigned long flags; | |
582 | struct rcu_dynticks *rdtp; | |
583 | ||
584 | local_irq_save(flags); | |
585 | rdtp = &__get_cpu_var(rcu_dynticks); | |
586 | /* Ensure we are interrupting an RCU idle mode. */ | |
587 | WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK); | |
588 | rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE; | |
589 | local_irq_restore(flags); | |
590 | } | |
2b1d5024 | 591 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 592 | |
9b2e4f18 PM |
593 | /** |
594 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
595 | * | |
596 | * Enter an interrupt handler, which might possibly result in exiting | |
597 | * idle mode, in other words, entering the mode in which read-side critical | |
598 | * sections can occur. | |
599 | * | |
600 | * Note that the Linux kernel is fully capable of entering an interrupt | |
601 | * handler that it never exits, for example when doing upcalls to | |
602 | * user mode! This code assumes that the idle loop never does upcalls to | |
603 | * user mode. If your architecture does do upcalls from the idle loop (or | |
604 | * does anything else that results in unbalanced calls to the irq_enter() | |
605 | * and irq_exit() functions), RCU will give you what you deserve, good | |
606 | * and hard. But very infrequently and irreproducibly. | |
607 | * | |
608 | * Use things like work queues to work around this limitation. | |
609 | * | |
610 | * You have been warned. | |
611 | */ | |
612 | void rcu_irq_enter(void) | |
613 | { | |
614 | unsigned long flags; | |
615 | struct rcu_dynticks *rdtp; | |
616 | long long oldval; | |
617 | ||
618 | local_irq_save(flags); | |
619 | rdtp = &__get_cpu_var(rcu_dynticks); | |
620 | oldval = rdtp->dynticks_nesting; | |
621 | rdtp->dynticks_nesting++; | |
622 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 FW |
623 | if (oldval) |
624 | trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting); | |
625 | else | |
cb349ca9 | 626 | rcu_eqs_exit_common(rdtp, oldval, true); |
64db4cff | 627 | local_irq_restore(flags); |
64db4cff PM |
628 | } |
629 | ||
630 | /** | |
631 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
632 | * | |
633 | * If the CPU was idle with dynamic ticks active, and there is no | |
634 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
635 | * RCU grace-period handling know that the CPU is active. | |
636 | */ | |
637 | void rcu_nmi_enter(void) | |
638 | { | |
639 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); | |
640 | ||
23b5c8fa PM |
641 | if (rdtp->dynticks_nmi_nesting == 0 && |
642 | (atomic_read(&rdtp->dynticks) & 0x1)) | |
64db4cff | 643 | return; |
23b5c8fa PM |
644 | rdtp->dynticks_nmi_nesting++; |
645 | smp_mb__before_atomic_inc(); /* Force delay from prior write. */ | |
646 | atomic_inc(&rdtp->dynticks); | |
647 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
648 | smp_mb__after_atomic_inc(); /* See above. */ | |
649 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
64db4cff PM |
650 | } |
651 | ||
652 | /** | |
653 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
654 | * | |
655 | * If the CPU was idle with dynamic ticks active, and there is no | |
656 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
657 | * RCU grace-period handling know that the CPU is no longer active. | |
658 | */ | |
659 | void rcu_nmi_exit(void) | |
660 | { | |
661 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); | |
662 | ||
23b5c8fa PM |
663 | if (rdtp->dynticks_nmi_nesting == 0 || |
664 | --rdtp->dynticks_nmi_nesting != 0) | |
64db4cff | 665 | return; |
23b5c8fa PM |
666 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
667 | smp_mb__before_atomic_inc(); /* See above. */ | |
668 | atomic_inc(&rdtp->dynticks); | |
669 | smp_mb__after_atomic_inc(); /* Force delay to next write. */ | |
670 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
64db4cff PM |
671 | } |
672 | ||
673 | /** | |
9b2e4f18 | 674 | * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle |
64db4cff | 675 | * |
9b2e4f18 | 676 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 677 | * or NMI handler, return true. |
64db4cff | 678 | */ |
9b2e4f18 | 679 | int rcu_is_cpu_idle(void) |
64db4cff | 680 | { |
34240697 PM |
681 | int ret; |
682 | ||
683 | preempt_disable(); | |
684 | ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0; | |
685 | preempt_enable(); | |
686 | return ret; | |
64db4cff | 687 | } |
e6b80a3b | 688 | EXPORT_SYMBOL(rcu_is_cpu_idle); |
64db4cff | 689 | |
62fde6ed | 690 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
691 | |
692 | /* | |
693 | * Is the current CPU online? Disable preemption to avoid false positives | |
694 | * that could otherwise happen due to the current CPU number being sampled, | |
695 | * this task being preempted, its old CPU being taken offline, resuming | |
696 | * on some other CPU, then determining that its old CPU is now offline. | |
697 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
698 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
699 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
700 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
701 | * offline to continue to use RCU for one jiffy after marking itself | |
702 | * offline in the cpu_online_mask. This leniency is necessary given the | |
703 | * non-atomic nature of the online and offline processing, for example, | |
704 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
705 | * notifiers. | |
706 | * | |
707 | * This is also why RCU internally marks CPUs online during the | |
708 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
709 | * |
710 | * Disable checking if in an NMI handler because we cannot safely report | |
711 | * errors from NMI handlers anyway. | |
712 | */ | |
713 | bool rcu_lockdep_current_cpu_online(void) | |
714 | { | |
2036d94a PM |
715 | struct rcu_data *rdp; |
716 | struct rcu_node *rnp; | |
c0d6d01b PM |
717 | bool ret; |
718 | ||
719 | if (in_nmi()) | |
720 | return 1; | |
721 | preempt_disable(); | |
2036d94a PM |
722 | rdp = &__get_cpu_var(rcu_sched_data); |
723 | rnp = rdp->mynode; | |
724 | ret = (rdp->grpmask & rnp->qsmaskinit) || | |
c0d6d01b PM |
725 | !rcu_scheduler_fully_active; |
726 | preempt_enable(); | |
727 | return ret; | |
728 | } | |
729 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
730 | ||
62fde6ed | 731 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 732 | |
64db4cff | 733 | /** |
9b2e4f18 | 734 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 735 | * |
9b2e4f18 PM |
736 | * If the current CPU is idle or running at a first-level (not nested) |
737 | * interrupt from idle, return true. The caller must have at least | |
738 | * disabled preemption. | |
64db4cff | 739 | */ |
9b2e4f18 | 740 | int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 741 | { |
9b2e4f18 | 742 | return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1; |
64db4cff PM |
743 | } |
744 | ||
64db4cff PM |
745 | /* |
746 | * Snapshot the specified CPU's dynticks counter so that we can later | |
747 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 748 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff PM |
749 | */ |
750 | static int dyntick_save_progress_counter(struct rcu_data *rdp) | |
751 | { | |
23b5c8fa | 752 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
f0e7c19d | 753 | return (rdp->dynticks_snap & 0x1) == 0; |
64db4cff PM |
754 | } |
755 | ||
756 | /* | |
757 | * Return true if the specified CPU has passed through a quiescent | |
758 | * state by virtue of being in or having passed through an dynticks | |
759 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 760 | * for this same CPU, or by virtue of having been offline. |
64db4cff PM |
761 | */ |
762 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp) | |
763 | { | |
7eb4f455 PM |
764 | unsigned int curr; |
765 | unsigned int snap; | |
64db4cff | 766 | |
7eb4f455 PM |
767 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
768 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
769 | |
770 | /* | |
771 | * If the CPU passed through or entered a dynticks idle phase with | |
772 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
773 | * already acknowledged the request to pass through a quiescent | |
774 | * state. Either way, that CPU cannot possibly be in an RCU | |
775 | * read-side critical section that started before the beginning | |
776 | * of the current RCU grace period. | |
777 | */ | |
7eb4f455 | 778 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
d4c08f2a | 779 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti"); |
64db4cff PM |
780 | rdp->dynticks_fqs++; |
781 | return 1; | |
782 | } | |
783 | ||
a82dcc76 PM |
784 | /* |
785 | * Check for the CPU being offline, but only if the grace period | |
786 | * is old enough. We don't need to worry about the CPU changing | |
787 | * state: If we see it offline even once, it has been through a | |
788 | * quiescent state. | |
789 | * | |
790 | * The reason for insisting that the grace period be at least | |
791 | * one jiffy old is that CPUs that are not quite online and that | |
792 | * have just gone offline can still execute RCU read-side critical | |
793 | * sections. | |
794 | */ | |
795 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
796 | return 0; /* Grace period is not old enough. */ | |
797 | barrier(); | |
798 | if (cpu_is_offline(rdp->cpu)) { | |
799 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl"); | |
800 | rdp->offline_fqs++; | |
801 | return 1; | |
802 | } | |
803 | return 0; | |
64db4cff PM |
804 | } |
805 | ||
13cfcca0 PM |
806 | static int jiffies_till_stall_check(void) |
807 | { | |
808 | int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout); | |
809 | ||
810 | /* | |
811 | * Limit check must be consistent with the Kconfig limits | |
812 | * for CONFIG_RCU_CPU_STALL_TIMEOUT. | |
813 | */ | |
814 | if (till_stall_check < 3) { | |
815 | ACCESS_ONCE(rcu_cpu_stall_timeout) = 3; | |
816 | till_stall_check = 3; | |
817 | } else if (till_stall_check > 300) { | |
818 | ACCESS_ONCE(rcu_cpu_stall_timeout) = 300; | |
819 | till_stall_check = 300; | |
820 | } | |
821 | return till_stall_check * HZ + RCU_STALL_DELAY_DELTA; | |
822 | } | |
823 | ||
64db4cff PM |
824 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
825 | { | |
826 | rsp->gp_start = jiffies; | |
13cfcca0 | 827 | rsp->jiffies_stall = jiffies + jiffies_till_stall_check(); |
64db4cff PM |
828 | } |
829 | ||
b637a328 PM |
830 | /* |
831 | * Dump stacks of all tasks running on stalled CPUs. This is a fallback | |
832 | * for architectures that do not implement trigger_all_cpu_backtrace(). | |
833 | * The NMI-triggered stack traces are more accurate because they are | |
834 | * printed by the target CPU. | |
835 | */ | |
836 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
837 | { | |
838 | int cpu; | |
839 | unsigned long flags; | |
840 | struct rcu_node *rnp; | |
841 | ||
842 | rcu_for_each_leaf_node(rsp, rnp) { | |
843 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
844 | if (rnp->qsmask != 0) { | |
845 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
846 | if (rnp->qsmask & (1UL << cpu)) | |
847 | dump_cpu_task(rnp->grplo + cpu); | |
848 | } | |
849 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
850 | } | |
851 | } | |
852 | ||
64db4cff PM |
853 | static void print_other_cpu_stall(struct rcu_state *rsp) |
854 | { | |
855 | int cpu; | |
856 | long delta; | |
857 | unsigned long flags; | |
285fe294 | 858 | int ndetected = 0; |
64db4cff | 859 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 860 | long totqlen = 0; |
64db4cff PM |
861 | |
862 | /* Only let one CPU complain about others per time interval. */ | |
863 | ||
1304afb2 | 864 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 865 | delta = jiffies - rsp->jiffies_stall; |
fc2219d4 | 866 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 867 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
868 | return; |
869 | } | |
13cfcca0 | 870 | rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3; |
1304afb2 | 871 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 872 | |
8cdd32a9 PM |
873 | /* |
874 | * OK, time to rat on our buddy... | |
875 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
876 | * RCU CPU stall warnings. | |
877 | */ | |
a858af28 | 878 | printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 879 | rsp->name); |
a858af28 | 880 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 881 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 882 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 883 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
884 | if (rnp->qsmask != 0) { |
885 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
886 | if (rnp->qsmask & (1UL << cpu)) { | |
887 | print_cpu_stall_info(rsp, | |
888 | rnp->grplo + cpu); | |
889 | ndetected++; | |
890 | } | |
891 | } | |
3acd9eb3 | 892 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 893 | } |
a858af28 PM |
894 | |
895 | /* | |
896 | * Now rat on any tasks that got kicked up to the root rcu_node | |
897 | * due to CPU offlining. | |
898 | */ | |
899 | rnp = rcu_get_root(rsp); | |
900 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
285fe294 | 901 | ndetected += rcu_print_task_stall(rnp); |
a858af28 PM |
902 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
903 | ||
904 | print_cpu_stall_info_end(); | |
53bb857c PM |
905 | for_each_possible_cpu(cpu) |
906 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
907 | pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n", | |
eee05882 | 908 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
53bb857c | 909 | rsp->gpnum, rsp->completed, totqlen); |
9bc8b558 PM |
910 | if (ndetected == 0) |
911 | printk(KERN_ERR "INFO: Stall ended before state dump start\n"); | |
912 | else if (!trigger_all_cpu_backtrace()) | |
b637a328 | 913 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 914 | |
4cdfc175 | 915 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
916 | |
917 | rcu_print_detail_task_stall(rsp); | |
918 | ||
4cdfc175 | 919 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
920 | } |
921 | ||
922 | static void print_cpu_stall(struct rcu_state *rsp) | |
923 | { | |
53bb857c | 924 | int cpu; |
64db4cff PM |
925 | unsigned long flags; |
926 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 927 | long totqlen = 0; |
64db4cff | 928 | |
8cdd32a9 PM |
929 | /* |
930 | * OK, time to rat on ourselves... | |
931 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
932 | * RCU CPU stall warnings. | |
933 | */ | |
a858af28 PM |
934 | printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name); |
935 | print_cpu_stall_info_begin(); | |
936 | print_cpu_stall_info(rsp, smp_processor_id()); | |
937 | print_cpu_stall_info_end(); | |
53bb857c PM |
938 | for_each_possible_cpu(cpu) |
939 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
940 | pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n", | |
941 | jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen); | |
4627e240 PM |
942 | if (!trigger_all_cpu_backtrace()) |
943 | dump_stack(); | |
c1dc0b9c | 944 | |
1304afb2 | 945 | raw_spin_lock_irqsave(&rnp->lock, flags); |
20133cfc | 946 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall)) |
13cfcca0 PM |
947 | rsp->jiffies_stall = jiffies + |
948 | 3 * jiffies_till_stall_check() + 3; | |
1304afb2 | 949 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 950 | |
64db4cff PM |
951 | set_need_resched(); /* kick ourselves to get things going. */ |
952 | } | |
953 | ||
954 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
955 | { | |
bad6e139 PM |
956 | unsigned long j; |
957 | unsigned long js; | |
64db4cff PM |
958 | struct rcu_node *rnp; |
959 | ||
742734ee | 960 | if (rcu_cpu_stall_suppress) |
c68de209 | 961 | return; |
bad6e139 PM |
962 | j = ACCESS_ONCE(jiffies); |
963 | js = ACCESS_ONCE(rsp->jiffies_stall); | |
64db4cff | 964 | rnp = rdp->mynode; |
c96ea7cf PM |
965 | if (rcu_gp_in_progress(rsp) && |
966 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) { | |
64db4cff PM |
967 | |
968 | /* We haven't checked in, so go dump stack. */ | |
969 | print_cpu_stall(rsp); | |
970 | ||
bad6e139 PM |
971 | } else if (rcu_gp_in_progress(rsp) && |
972 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 973 | |
bad6e139 | 974 | /* They had a few time units to dump stack, so complain. */ |
64db4cff PM |
975 | print_other_cpu_stall(rsp); |
976 | } | |
977 | } | |
978 | ||
c68de209 PM |
979 | static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr) |
980 | { | |
742734ee | 981 | rcu_cpu_stall_suppress = 1; |
c68de209 PM |
982 | return NOTIFY_DONE; |
983 | } | |
984 | ||
53d84e00 PM |
985 | /** |
986 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
987 | * | |
988 | * Set the stall-warning timeout way off into the future, thus preventing | |
989 | * any RCU CPU stall-warning messages from appearing in the current set of | |
990 | * RCU grace periods. | |
991 | * | |
992 | * The caller must disable hard irqs. | |
993 | */ | |
994 | void rcu_cpu_stall_reset(void) | |
995 | { | |
6ce75a23 PM |
996 | struct rcu_state *rsp; |
997 | ||
998 | for_each_rcu_flavor(rsp) | |
999 | rsp->jiffies_stall = jiffies + ULONG_MAX / 2; | |
53d84e00 PM |
1000 | } |
1001 | ||
c68de209 PM |
1002 | static struct notifier_block rcu_panic_block = { |
1003 | .notifier_call = rcu_panic, | |
1004 | }; | |
1005 | ||
1006 | static void __init check_cpu_stall_init(void) | |
1007 | { | |
1008 | atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block); | |
1009 | } | |
1010 | ||
64db4cff PM |
1011 | /* |
1012 | * Update CPU-local rcu_data state to record the newly noticed grace period. | |
1013 | * This is used both when we started the grace period and when we notice | |
9160306e PM |
1014 | * that someone else started the grace period. The caller must hold the |
1015 | * ->lock of the leaf rcu_node structure corresponding to the current CPU, | |
1016 | * and must have irqs disabled. | |
64db4cff | 1017 | */ |
9160306e PM |
1018 | static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) |
1019 | { | |
1020 | if (rdp->gpnum != rnp->gpnum) { | |
121dfc4b PM |
1021 | /* |
1022 | * If the current grace period is waiting for this CPU, | |
1023 | * set up to detect a quiescent state, otherwise don't | |
1024 | * go looking for one. | |
1025 | */ | |
9160306e | 1026 | rdp->gpnum = rnp->gpnum; |
d4c08f2a | 1027 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart"); |
d7d6a11e PM |
1028 | rdp->passed_quiesce = 0; |
1029 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); | |
a858af28 | 1030 | zero_cpu_stall_ticks(rdp); |
9160306e PM |
1031 | } |
1032 | } | |
1033 | ||
64db4cff PM |
1034 | static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp) |
1035 | { | |
9160306e PM |
1036 | unsigned long flags; |
1037 | struct rcu_node *rnp; | |
1038 | ||
1039 | local_irq_save(flags); | |
1040 | rnp = rdp->mynode; | |
1041 | if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */ | |
1304afb2 | 1042 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
9160306e PM |
1043 | local_irq_restore(flags); |
1044 | return; | |
1045 | } | |
1046 | __note_new_gpnum(rsp, rnp, rdp); | |
1304afb2 | 1047 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1048 | } |
1049 | ||
1050 | /* | |
1051 | * Did someone else start a new RCU grace period start since we last | |
1052 | * checked? Update local state appropriately if so. Must be called | |
1053 | * on the CPU corresponding to rdp. | |
1054 | */ | |
1055 | static int | |
1056 | check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp) | |
1057 | { | |
1058 | unsigned long flags; | |
1059 | int ret = 0; | |
1060 | ||
1061 | local_irq_save(flags); | |
1062 | if (rdp->gpnum != rsp->gpnum) { | |
1063 | note_new_gpnum(rsp, rdp); | |
1064 | ret = 1; | |
1065 | } | |
1066 | local_irq_restore(flags); | |
1067 | return ret; | |
1068 | } | |
1069 | ||
3f5d3ea6 PM |
1070 | /* |
1071 | * Initialize the specified rcu_data structure's callback list to empty. | |
1072 | */ | |
1073 | static void init_callback_list(struct rcu_data *rdp) | |
1074 | { | |
1075 | int i; | |
1076 | ||
1077 | rdp->nxtlist = NULL; | |
1078 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1079 | rdp->nxttail[i] = &rdp->nxtlist; | |
3fbfbf7a | 1080 | init_nocb_callback_list(rdp); |
3f5d3ea6 PM |
1081 | } |
1082 | ||
dc35c893 PM |
1083 | /* |
1084 | * Determine the value that ->completed will have at the end of the | |
1085 | * next subsequent grace period. This is used to tag callbacks so that | |
1086 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1087 | * been dyntick-idle for an extended period with callbacks under the | |
1088 | * influence of RCU_FAST_NO_HZ. | |
1089 | * | |
1090 | * The caller must hold rnp->lock with interrupts disabled. | |
1091 | */ | |
1092 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1093 | struct rcu_node *rnp) | |
1094 | { | |
1095 | /* | |
1096 | * If RCU is idle, we just wait for the next grace period. | |
1097 | * But we can only be sure that RCU is idle if we are looking | |
1098 | * at the root rcu_node structure -- otherwise, a new grace | |
1099 | * period might have started, but just not yet gotten around | |
1100 | * to initializing the current non-root rcu_node structure. | |
1101 | */ | |
1102 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1103 | return rnp->completed + 1; | |
1104 | ||
1105 | /* | |
1106 | * Otherwise, wait for a possible partial grace period and | |
1107 | * then the subsequent full grace period. | |
1108 | */ | |
1109 | return rnp->completed + 2; | |
1110 | } | |
1111 | ||
1112 | /* | |
1113 | * If there is room, assign a ->completed number to any callbacks on | |
1114 | * this CPU that have not already been assigned. Also accelerate any | |
1115 | * callbacks that were previously assigned a ->completed number that has | |
1116 | * since proven to be too conservative, which can happen if callbacks get | |
1117 | * assigned a ->completed number while RCU is idle, but with reference to | |
1118 | * a non-root rcu_node structure. This function is idempotent, so it does | |
1119 | * not hurt to call it repeatedly. | |
1120 | * | |
1121 | * The caller must hold rnp->lock with interrupts disabled. | |
1122 | */ | |
1123 | static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1124 | struct rcu_data *rdp) | |
1125 | { | |
1126 | unsigned long c; | |
1127 | int i; | |
1128 | ||
1129 | /* If the CPU has no callbacks, nothing to do. */ | |
1130 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1131 | return; | |
1132 | ||
1133 | /* | |
1134 | * Starting from the sublist containing the callbacks most | |
1135 | * recently assigned a ->completed number and working down, find the | |
1136 | * first sublist that is not assignable to an upcoming grace period. | |
1137 | * Such a sublist has something in it (first two tests) and has | |
1138 | * a ->completed number assigned that will complete sooner than | |
1139 | * the ->completed number for newly arrived callbacks (last test). | |
1140 | * | |
1141 | * The key point is that any later sublist can be assigned the | |
1142 | * same ->completed number as the newly arrived callbacks, which | |
1143 | * means that the callbacks in any of these later sublist can be | |
1144 | * grouped into a single sublist, whether or not they have already | |
1145 | * been assigned a ->completed number. | |
1146 | */ | |
1147 | c = rcu_cbs_completed(rsp, rnp); | |
1148 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1149 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1150 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1151 | break; | |
1152 | ||
1153 | /* | |
1154 | * If there are no sublist for unassigned callbacks, leave. | |
1155 | * At the same time, advance "i" one sublist, so that "i" will | |
1156 | * index into the sublist where all the remaining callbacks should | |
1157 | * be grouped into. | |
1158 | */ | |
1159 | if (++i >= RCU_NEXT_TAIL) | |
1160 | return; | |
1161 | ||
1162 | /* | |
1163 | * Assign all subsequent callbacks' ->completed number to the next | |
1164 | * full grace period and group them all in the sublist initially | |
1165 | * indexed by "i". | |
1166 | */ | |
1167 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1168 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1169 | rdp->nxtcompleted[i] = c; | |
1170 | } | |
1171 | } | |
1172 | ||
1173 | /* | |
1174 | * Move any callbacks whose grace period has completed to the | |
1175 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1176 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1177 | * sublist. This function is idempotent, so it does not hurt to | |
1178 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
1179 | * | |
1180 | * The caller must hold rnp->lock with interrupts disabled. | |
1181 | */ | |
1182 | static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1183 | struct rcu_data *rdp) | |
1184 | { | |
1185 | int i, j; | |
1186 | ||
1187 | /* If the CPU has no callbacks, nothing to do. */ | |
1188 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1189 | return; | |
1190 | ||
1191 | /* | |
1192 | * Find all callbacks whose ->completed numbers indicate that they | |
1193 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1194 | */ | |
1195 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1196 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1197 | break; | |
1198 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1199 | } | |
1200 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1201 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1202 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1203 | ||
1204 | /* Copy down callbacks to fill in empty sublists. */ | |
1205 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1206 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1207 | break; | |
1208 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1209 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1210 | } | |
1211 | ||
1212 | /* Classify any remaining callbacks. */ | |
1213 | rcu_accelerate_cbs(rsp, rnp, rdp); | |
1214 | } | |
1215 | ||
d09b62df PM |
1216 | /* |
1217 | * Advance this CPU's callbacks, but only if the current grace period | |
1218 | * has ended. This may be called only from the CPU to whom the rdp | |
1219 | * belongs. In addition, the corresponding leaf rcu_node structure's | |
1220 | * ->lock must be held by the caller, with irqs disabled. | |
1221 | */ | |
1222 | static void | |
1223 | __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) | |
1224 | { | |
1225 | /* Did another grace period end? */ | |
dc35c893 PM |
1226 | if (rdp->completed == rnp->completed) { |
1227 | ||
1228 | /* No, so just accelerate recent callbacks. */ | |
1229 | rcu_accelerate_cbs(rsp, rnp, rdp); | |
d09b62df | 1230 | |
dc35c893 PM |
1231 | } else { |
1232 | ||
1233 | /* Advance callbacks. */ | |
1234 | rcu_advance_cbs(rsp, rnp, rdp); | |
d09b62df PM |
1235 | |
1236 | /* Remember that we saw this grace-period completion. */ | |
1237 | rdp->completed = rnp->completed; | |
d4c08f2a | 1238 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend"); |
20377f32 | 1239 | |
5ff8e6f0 FW |
1240 | /* |
1241 | * If we were in an extended quiescent state, we may have | |
121dfc4b | 1242 | * missed some grace periods that others CPUs handled on |
5ff8e6f0 | 1243 | * our behalf. Catch up with this state to avoid noting |
121dfc4b PM |
1244 | * spurious new grace periods. If another grace period |
1245 | * has started, then rnp->gpnum will have advanced, so | |
d7d6a11e PM |
1246 | * we will detect this later on. Of course, any quiescent |
1247 | * states we found for the old GP are now invalid. | |
5ff8e6f0 | 1248 | */ |
d7d6a11e | 1249 | if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) { |
5ff8e6f0 | 1250 | rdp->gpnum = rdp->completed; |
d7d6a11e PM |
1251 | rdp->passed_quiesce = 0; |
1252 | } | |
5ff8e6f0 | 1253 | |
20377f32 | 1254 | /* |
121dfc4b PM |
1255 | * If RCU does not need a quiescent state from this CPU, |
1256 | * then make sure that this CPU doesn't go looking for one. | |
20377f32 | 1257 | */ |
121dfc4b | 1258 | if ((rnp->qsmask & rdp->grpmask) == 0) |
20377f32 | 1259 | rdp->qs_pending = 0; |
d09b62df PM |
1260 | } |
1261 | } | |
1262 | ||
1263 | /* | |
1264 | * Advance this CPU's callbacks, but only if the current grace period | |
1265 | * has ended. This may be called only from the CPU to whom the rdp | |
1266 | * belongs. | |
1267 | */ | |
1268 | static void | |
1269 | rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp) | |
1270 | { | |
1271 | unsigned long flags; | |
1272 | struct rcu_node *rnp; | |
1273 | ||
1274 | local_irq_save(flags); | |
1275 | rnp = rdp->mynode; | |
1276 | if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */ | |
1304afb2 | 1277 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
d09b62df PM |
1278 | local_irq_restore(flags); |
1279 | return; | |
1280 | } | |
1281 | __rcu_process_gp_end(rsp, rnp, rdp); | |
1304afb2 | 1282 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
d09b62df PM |
1283 | } |
1284 | ||
1285 | /* | |
1286 | * Do per-CPU grace-period initialization for running CPU. The caller | |
1287 | * must hold the lock of the leaf rcu_node structure corresponding to | |
1288 | * this CPU. | |
1289 | */ | |
1290 | static void | |
1291 | rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) | |
1292 | { | |
1293 | /* Prior grace period ended, so advance callbacks for current CPU. */ | |
1294 | __rcu_process_gp_end(rsp, rnp, rdp); | |
1295 | ||
9160306e PM |
1296 | /* Set state so that this CPU will detect the next quiescent state. */ |
1297 | __note_new_gpnum(rsp, rnp, rdp); | |
d09b62df PM |
1298 | } |
1299 | ||
b3dbec76 | 1300 | /* |
7fdefc10 | 1301 | * Initialize a new grace period. |
b3dbec76 | 1302 | */ |
7fdefc10 | 1303 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 PM |
1304 | { |
1305 | struct rcu_data *rdp; | |
7fdefc10 | 1306 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1307 | |
7fdefc10 | 1308 | raw_spin_lock_irq(&rnp->lock); |
4cdfc175 | 1309 | rsp->gp_flags = 0; /* Clear all flags: New grace period. */ |
b3dbec76 | 1310 | |
7fdefc10 PM |
1311 | if (rcu_gp_in_progress(rsp)) { |
1312 | /* Grace period already in progress, don't start another. */ | |
1313 | raw_spin_unlock_irq(&rnp->lock); | |
1314 | return 0; | |
1315 | } | |
1316 | ||
7fdefc10 PM |
1317 | /* Advance to a new grace period and initialize state. */ |
1318 | rsp->gpnum++; | |
1319 | trace_rcu_grace_period(rsp->name, rsp->gpnum, "start"); | |
7fdefc10 PM |
1320 | record_gp_stall_check_time(rsp); |
1321 | raw_spin_unlock_irq(&rnp->lock); | |
1322 | ||
1323 | /* Exclude any concurrent CPU-hotplug operations. */ | |
a4fbe35a | 1324 | mutex_lock(&rsp->onoff_mutex); |
7fdefc10 PM |
1325 | |
1326 | /* | |
1327 | * Set the quiescent-state-needed bits in all the rcu_node | |
1328 | * structures for all currently online CPUs in breadth-first order, | |
1329 | * starting from the root rcu_node structure, relying on the layout | |
1330 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1331 | * will access only the leaves of the hierarchy, thus seeing that no | |
1332 | * grace period is in progress, at least until the corresponding | |
1333 | * leaf node has been initialized. In addition, we have excluded | |
1334 | * CPU-hotplug operations. | |
1335 | * | |
1336 | * The grace period cannot complete until the initialization | |
1337 | * process finishes, because this kthread handles both. | |
1338 | */ | |
1339 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
b3dbec76 | 1340 | raw_spin_lock_irq(&rnp->lock); |
b3dbec76 | 1341 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1342 | rcu_preempt_check_blocked_tasks(rnp); |
1343 | rnp->qsmask = rnp->qsmaskinit; | |
1344 | rnp->gpnum = rsp->gpnum; | |
25d30cf4 | 1345 | WARN_ON_ONCE(rnp->completed != rsp->completed); |
7fdefc10 PM |
1346 | rnp->completed = rsp->completed; |
1347 | if (rnp == rdp->mynode) | |
1348 | rcu_start_gp_per_cpu(rsp, rnp, rdp); | |
1349 | rcu_preempt_boost_start_gp(rnp); | |
1350 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1351 | rnp->level, rnp->grplo, | |
1352 | rnp->grphi, rnp->qsmask); | |
1353 | raw_spin_unlock_irq(&rnp->lock); | |
661a85dc PM |
1354 | #ifdef CONFIG_PROVE_RCU_DELAY |
1355 | if ((random32() % (rcu_num_nodes * 8)) == 0) | |
1356 | schedule_timeout_uninterruptible(2); | |
1357 | #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */ | |
7fdefc10 PM |
1358 | cond_resched(); |
1359 | } | |
b3dbec76 | 1360 | |
a4fbe35a | 1361 | mutex_unlock(&rsp->onoff_mutex); |
7fdefc10 PM |
1362 | return 1; |
1363 | } | |
b3dbec76 | 1364 | |
4cdfc175 PM |
1365 | /* |
1366 | * Do one round of quiescent-state forcing. | |
1367 | */ | |
1368 | int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) | |
1369 | { | |
1370 | int fqs_state = fqs_state_in; | |
1371 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1372 | ||
1373 | rsp->n_force_qs++; | |
1374 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1375 | /* Collect dyntick-idle snapshots. */ | |
1376 | force_qs_rnp(rsp, dyntick_save_progress_counter); | |
1377 | fqs_state = RCU_FORCE_QS; | |
1378 | } else { | |
1379 | /* Handle dyntick-idle and offline CPUs. */ | |
1380 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs); | |
1381 | } | |
1382 | /* Clear flag to prevent immediate re-entry. */ | |
1383 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1384 | raw_spin_lock_irq(&rnp->lock); | |
1385 | rsp->gp_flags &= ~RCU_GP_FLAG_FQS; | |
1386 | raw_spin_unlock_irq(&rnp->lock); | |
1387 | } | |
1388 | return fqs_state; | |
1389 | } | |
1390 | ||
7fdefc10 PM |
1391 | /* |
1392 | * Clean up after the old grace period. | |
1393 | */ | |
4cdfc175 | 1394 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1395 | { |
1396 | unsigned long gp_duration; | |
1397 | struct rcu_data *rdp; | |
1398 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1399 | |
7fdefc10 PM |
1400 | raw_spin_lock_irq(&rnp->lock); |
1401 | gp_duration = jiffies - rsp->gp_start; | |
1402 | if (gp_duration > rsp->gp_max) | |
1403 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1404 | |
7fdefc10 PM |
1405 | /* |
1406 | * We know the grace period is complete, but to everyone else | |
1407 | * it appears to still be ongoing. But it is also the case | |
1408 | * that to everyone else it looks like there is nothing that | |
1409 | * they can do to advance the grace period. It is therefore | |
1410 | * safe for us to drop the lock in order to mark the grace | |
1411 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1412 | */ |
5d4b8659 | 1413 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1414 | |
5d4b8659 PM |
1415 | /* |
1416 | * Propagate new ->completed value to rcu_node structures so | |
1417 | * that other CPUs don't have to wait until the start of the next | |
1418 | * grace period to process their callbacks. This also avoids | |
1419 | * some nasty RCU grace-period initialization races by forcing | |
1420 | * the end of the current grace period to be completely recorded in | |
1421 | * all of the rcu_node structures before the beginning of the next | |
1422 | * grace period is recorded in any of the rcu_node structures. | |
1423 | */ | |
1424 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1425 | raw_spin_lock_irq(&rnp->lock); |
5d4b8659 PM |
1426 | rnp->completed = rsp->gpnum; |
1427 | raw_spin_unlock_irq(&rnp->lock); | |
1428 | cond_resched(); | |
7fdefc10 | 1429 | } |
5d4b8659 PM |
1430 | rnp = rcu_get_root(rsp); |
1431 | raw_spin_lock_irq(&rnp->lock); | |
7fdefc10 PM |
1432 | |
1433 | rsp->completed = rsp->gpnum; /* Declare grace period done. */ | |
1434 | trace_rcu_grace_period(rsp->name, rsp->completed, "end"); | |
1435 | rsp->fqs_state = RCU_GP_IDLE; | |
5d4b8659 | 1436 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1437 | if (cpu_needs_another_gp(rsp, rdp)) |
1438 | rsp->gp_flags = 1; | |
1439 | raw_spin_unlock_irq(&rnp->lock); | |
7fdefc10 PM |
1440 | } |
1441 | ||
1442 | /* | |
1443 | * Body of kthread that handles grace periods. | |
1444 | */ | |
1445 | static int __noreturn rcu_gp_kthread(void *arg) | |
1446 | { | |
4cdfc175 | 1447 | int fqs_state; |
d40011f6 | 1448 | unsigned long j; |
4cdfc175 | 1449 | int ret; |
7fdefc10 PM |
1450 | struct rcu_state *rsp = arg; |
1451 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1452 | ||
1453 | for (;;) { | |
1454 | ||
1455 | /* Handle grace-period start. */ | |
1456 | for (;;) { | |
4cdfc175 PM |
1457 | wait_event_interruptible(rsp->gp_wq, |
1458 | rsp->gp_flags & | |
1459 | RCU_GP_FLAG_INIT); | |
1460 | if ((rsp->gp_flags & RCU_GP_FLAG_INIT) && | |
1461 | rcu_gp_init(rsp)) | |
7fdefc10 PM |
1462 | break; |
1463 | cond_resched(); | |
1464 | flush_signals(current); | |
1465 | } | |
cabc49c1 | 1466 | |
4cdfc175 PM |
1467 | /* Handle quiescent-state forcing. */ |
1468 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
1469 | j = jiffies_till_first_fqs; |
1470 | if (j > HZ) { | |
1471 | j = HZ; | |
1472 | jiffies_till_first_fqs = HZ; | |
1473 | } | |
cabc49c1 | 1474 | for (;;) { |
d40011f6 | 1475 | rsp->jiffies_force_qs = jiffies + j; |
4cdfc175 PM |
1476 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
1477 | (rsp->gp_flags & RCU_GP_FLAG_FQS) || | |
1478 | (!ACCESS_ONCE(rnp->qsmask) && | |
1479 | !rcu_preempt_blocked_readers_cgp(rnp)), | |
d40011f6 | 1480 | j); |
4cdfc175 | 1481 | /* If grace period done, leave loop. */ |
cabc49c1 | 1482 | if (!ACCESS_ONCE(rnp->qsmask) && |
4cdfc175 | 1483 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 1484 | break; |
4cdfc175 PM |
1485 | /* If time for quiescent-state forcing, do it. */ |
1486 | if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) { | |
1487 | fqs_state = rcu_gp_fqs(rsp, fqs_state); | |
1488 | cond_resched(); | |
1489 | } else { | |
1490 | /* Deal with stray signal. */ | |
1491 | cond_resched(); | |
1492 | flush_signals(current); | |
1493 | } | |
d40011f6 PM |
1494 | j = jiffies_till_next_fqs; |
1495 | if (j > HZ) { | |
1496 | j = HZ; | |
1497 | jiffies_till_next_fqs = HZ; | |
1498 | } else if (j < 1) { | |
1499 | j = 1; | |
1500 | jiffies_till_next_fqs = 1; | |
1501 | } | |
cabc49c1 | 1502 | } |
4cdfc175 PM |
1503 | |
1504 | /* Handle grace-period end. */ | |
1505 | rcu_gp_cleanup(rsp); | |
b3dbec76 | 1506 | } |
b3dbec76 PM |
1507 | } |
1508 | ||
64db4cff PM |
1509 | /* |
1510 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
1511 | * in preparation for detecting the next grace period. The caller must hold | |
1512 | * the root node's ->lock, which is released before return. Hard irqs must | |
1513 | * be disabled. | |
e5601400 PM |
1514 | * |
1515 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
1516 | * invoke this function. This can happen when the dying CPU reports its | |
1517 | * quiescent state. | |
64db4cff PM |
1518 | */ |
1519 | static void | |
1520 | rcu_start_gp(struct rcu_state *rsp, unsigned long flags) | |
1521 | __releases(rcu_get_root(rsp)->lock) | |
1522 | { | |
394f99a9 | 1523 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); |
64db4cff | 1524 | struct rcu_node *rnp = rcu_get_root(rsp); |
64db4cff | 1525 | |
b3dbec76 | 1526 | if (!rsp->gp_kthread || |
afe24b12 PM |
1527 | !cpu_needs_another_gp(rsp, rdp)) { |
1528 | /* | |
b3dbec76 | 1529 | * Either we have not yet spawned the grace-period |
62da1921 PM |
1530 | * task, this CPU does not need another grace period, |
1531 | * or a grace period is already in progress. | |
b3dbec76 | 1532 | * Either way, don't start a new grace period. |
afe24b12 PM |
1533 | */ |
1534 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1535 | return; | |
1536 | } | |
b32e9eb6 | 1537 | |
62da1921 PM |
1538 | /* |
1539 | * Because there is no grace period in progress right now, | |
1540 | * any callbacks we have up to this point will be satisfied | |
dc35c893 PM |
1541 | * by the next grace period. So this is a good place to |
1542 | * assign a grace period number to recently posted callbacks. | |
62da1921 | 1543 | */ |
dc35c893 | 1544 | rcu_accelerate_cbs(rsp, rnp, rdp); |
62da1921 | 1545 | |
4cdfc175 | 1546 | rsp->gp_flags = RCU_GP_FLAG_INIT; |
62da1921 PM |
1547 | raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */ |
1548 | ||
1549 | /* Ensure that CPU is aware of completion of last grace period. */ | |
1550 | rcu_process_gp_end(rsp, rdp); | |
1551 | local_irq_restore(flags); | |
1552 | ||
1553 | /* Wake up rcu_gp_kthread() to start the grace period. */ | |
b3dbec76 | 1554 | wake_up(&rsp->gp_wq); |
64db4cff PM |
1555 | } |
1556 | ||
f41d911f | 1557 | /* |
d3f6bad3 PM |
1558 | * Report a full set of quiescent states to the specified rcu_state |
1559 | * data structure. This involves cleaning up after the prior grace | |
1560 | * period and letting rcu_start_gp() start up the next grace period | |
1561 | * if one is needed. Note that the caller must hold rnp->lock, as | |
1562 | * required by rcu_start_gp(), which will release it. | |
f41d911f | 1563 | */ |
d3f6bad3 | 1564 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 1565 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 1566 | { |
fc2219d4 | 1567 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cabc49c1 PM |
1568 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
1569 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ | |
f41d911f PM |
1570 | } |
1571 | ||
64db4cff | 1572 | /* |
d3f6bad3 PM |
1573 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
1574 | * Allows quiescent states for a group of CPUs to be reported at one go | |
1575 | * to the specified rcu_node structure, though all the CPUs in the group | |
1576 | * must be represented by the same rcu_node structure (which need not be | |
1577 | * a leaf rcu_node structure, though it often will be). That structure's | |
1578 | * lock must be held upon entry, and it is released before return. | |
64db4cff PM |
1579 | */ |
1580 | static void | |
d3f6bad3 PM |
1581 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
1582 | struct rcu_node *rnp, unsigned long flags) | |
64db4cff PM |
1583 | __releases(rnp->lock) |
1584 | { | |
28ecd580 PM |
1585 | struct rcu_node *rnp_c; |
1586 | ||
64db4cff PM |
1587 | /* Walk up the rcu_node hierarchy. */ |
1588 | for (;;) { | |
1589 | if (!(rnp->qsmask & mask)) { | |
1590 | ||
1591 | /* Our bit has already been cleared, so done. */ | |
1304afb2 | 1592 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1593 | return; |
1594 | } | |
1595 | rnp->qsmask &= ~mask; | |
d4c08f2a PM |
1596 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
1597 | mask, rnp->qsmask, rnp->level, | |
1598 | rnp->grplo, rnp->grphi, | |
1599 | !!rnp->gp_tasks); | |
27f4d280 | 1600 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
1601 | |
1602 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 1603 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1604 | return; |
1605 | } | |
1606 | mask = rnp->grpmask; | |
1607 | if (rnp->parent == NULL) { | |
1608 | ||
1609 | /* No more levels. Exit loop holding root lock. */ | |
1610 | ||
1611 | break; | |
1612 | } | |
1304afb2 | 1613 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 1614 | rnp_c = rnp; |
64db4cff | 1615 | rnp = rnp->parent; |
1304afb2 | 1616 | raw_spin_lock_irqsave(&rnp->lock, flags); |
28ecd580 | 1617 | WARN_ON_ONCE(rnp_c->qsmask); |
64db4cff PM |
1618 | } |
1619 | ||
1620 | /* | |
1621 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 1622 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 1623 | * to clean up and start the next grace period if one is needed. |
64db4cff | 1624 | */ |
d3f6bad3 | 1625 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
1626 | } |
1627 | ||
1628 | /* | |
d3f6bad3 PM |
1629 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
1630 | * structure. This must be either called from the specified CPU, or | |
1631 | * called when the specified CPU is known to be offline (and when it is | |
1632 | * also known that no other CPU is concurrently trying to help the offline | |
1633 | * CPU). The lastcomp argument is used to make sure we are still in the | |
1634 | * grace period of interest. We don't want to end the current grace period | |
1635 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
1636 | */ |
1637 | static void | |
d7d6a11e | 1638 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1639 | { |
1640 | unsigned long flags; | |
1641 | unsigned long mask; | |
1642 | struct rcu_node *rnp; | |
1643 | ||
1644 | rnp = rdp->mynode; | |
1304afb2 | 1645 | raw_spin_lock_irqsave(&rnp->lock, flags); |
d7d6a11e PM |
1646 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum || |
1647 | rnp->completed == rnp->gpnum) { | |
64db4cff PM |
1648 | |
1649 | /* | |
e4cc1f22 PM |
1650 | * The grace period in which this quiescent state was |
1651 | * recorded has ended, so don't report it upwards. | |
1652 | * We will instead need a new quiescent state that lies | |
1653 | * within the current grace period. | |
64db4cff | 1654 | */ |
e4cc1f22 | 1655 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
1304afb2 | 1656 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1657 | return; |
1658 | } | |
1659 | mask = rdp->grpmask; | |
1660 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 1661 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1662 | } else { |
1663 | rdp->qs_pending = 0; | |
1664 | ||
1665 | /* | |
1666 | * This GP can't end until cpu checks in, so all of our | |
1667 | * callbacks can be processed during the next GP. | |
1668 | */ | |
dc35c893 | 1669 | rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 1670 | |
d3f6bad3 | 1671 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
64db4cff PM |
1672 | } |
1673 | } | |
1674 | ||
1675 | /* | |
1676 | * Check to see if there is a new grace period of which this CPU | |
1677 | * is not yet aware, and if so, set up local rcu_data state for it. | |
1678 | * Otherwise, see if this CPU has just passed through its first | |
1679 | * quiescent state for this grace period, and record that fact if so. | |
1680 | */ | |
1681 | static void | |
1682 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
1683 | { | |
1684 | /* If there is now a new grace period, record and return. */ | |
1685 | if (check_for_new_grace_period(rsp, rdp)) | |
1686 | return; | |
1687 | ||
1688 | /* | |
1689 | * Does this CPU still need to do its part for current grace period? | |
1690 | * If no, return and let the other CPUs do their part as well. | |
1691 | */ | |
1692 | if (!rdp->qs_pending) | |
1693 | return; | |
1694 | ||
1695 | /* | |
1696 | * Was there a quiescent state since the beginning of the grace | |
1697 | * period? If no, then exit and wait for the next call. | |
1698 | */ | |
e4cc1f22 | 1699 | if (!rdp->passed_quiesce) |
64db4cff PM |
1700 | return; |
1701 | ||
d3f6bad3 PM |
1702 | /* |
1703 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
1704 | * judge of that). | |
1705 | */ | |
d7d6a11e | 1706 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
1707 | } |
1708 | ||
1709 | #ifdef CONFIG_HOTPLUG_CPU | |
1710 | ||
e74f4c45 | 1711 | /* |
b1420f1c PM |
1712 | * Send the specified CPU's RCU callbacks to the orphanage. The |
1713 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 1714 | * ->orphan_lock. |
e74f4c45 | 1715 | */ |
b1420f1c PM |
1716 | static void |
1717 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
1718 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 1719 | { |
3fbfbf7a PM |
1720 | /* No-CBs CPUs do not have orphanable callbacks. */ |
1721 | if (is_nocb_cpu(rdp->cpu)) | |
1722 | return; | |
1723 | ||
b1420f1c PM |
1724 | /* |
1725 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
1726 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
1727 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 1728 | */ |
a50c3af9 | 1729 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1730 | rsp->qlen_lazy += rdp->qlen_lazy; |
1731 | rsp->qlen += rdp->qlen; | |
1732 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 1733 | rdp->qlen_lazy = 0; |
1d1fb395 | 1734 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
1735 | } |
1736 | ||
1737 | /* | |
b1420f1c PM |
1738 | * Next, move those callbacks still needing a grace period to |
1739 | * the orphanage, where some other CPU will pick them up. | |
1740 | * Some of the callbacks might have gone partway through a grace | |
1741 | * period, but that is too bad. They get to start over because we | |
1742 | * cannot assume that grace periods are synchronized across CPUs. | |
1743 | * We don't bother updating the ->nxttail[] array yet, instead | |
1744 | * we just reset the whole thing later on. | |
a50c3af9 | 1745 | */ |
b1420f1c PM |
1746 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
1747 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1748 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
1749 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
1750 | } |
1751 | ||
1752 | /* | |
b1420f1c PM |
1753 | * Then move the ready-to-invoke callbacks to the orphanage, |
1754 | * where some other CPU will pick them up. These will not be | |
1755 | * required to pass though another grace period: They are done. | |
a50c3af9 | 1756 | */ |
e5601400 | 1757 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1758 | *rsp->orphan_donetail = rdp->nxtlist; |
1759 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 1760 | } |
e74f4c45 | 1761 | |
b1420f1c | 1762 | /* Finally, initialize the rcu_data structure's list to empty. */ |
3f5d3ea6 | 1763 | init_callback_list(rdp); |
b1420f1c PM |
1764 | } |
1765 | ||
1766 | /* | |
1767 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 1768 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c PM |
1769 | */ |
1770 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp) | |
1771 | { | |
1772 | int i; | |
1773 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); | |
1774 | ||
3fbfbf7a PM |
1775 | /* No-CBs CPUs are handled specially. */ |
1776 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp)) | |
1777 | return; | |
1778 | ||
b1420f1c PM |
1779 | /* Do the accounting first. */ |
1780 | rdp->qlen_lazy += rsp->qlen_lazy; | |
1781 | rdp->qlen += rsp->qlen; | |
1782 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
1783 | if (rsp->qlen_lazy != rsp->qlen) |
1784 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
1785 | rsp->qlen_lazy = 0; |
1786 | rsp->qlen = 0; | |
1787 | ||
1788 | /* | |
1789 | * We do not need a memory barrier here because the only way we | |
1790 | * can get here if there is an rcu_barrier() in flight is if | |
1791 | * we are the task doing the rcu_barrier(). | |
1792 | */ | |
1793 | ||
1794 | /* First adopt the ready-to-invoke callbacks. */ | |
1795 | if (rsp->orphan_donelist != NULL) { | |
1796 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1797 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
1798 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
1799 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
1800 | rdp->nxttail[i] = rsp->orphan_donetail; | |
1801 | rsp->orphan_donelist = NULL; | |
1802 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
1803 | } | |
1804 | ||
1805 | /* And then adopt the callbacks that still need a grace period. */ | |
1806 | if (rsp->orphan_nxtlist != NULL) { | |
1807 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
1808 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
1809 | rsp->orphan_nxtlist = NULL; | |
1810 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
1811 | } | |
1812 | } | |
1813 | ||
1814 | /* | |
1815 | * Trace the fact that this CPU is going offline. | |
1816 | */ | |
1817 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
1818 | { | |
1819 | RCU_TRACE(unsigned long mask); | |
1820 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
1821 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
1822 | ||
1823 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
1824 | trace_rcu_grace_period(rsp->name, |
1825 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
1826 | "cpuofl"); | |
64db4cff PM |
1827 | } |
1828 | ||
1829 | /* | |
e5601400 | 1830 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
1831 | * this fact from process context. Do the remainder of the cleanup, |
1832 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
1833 | * adopting them. There can only be one CPU hotplug operation at a time, |
1834 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 1835 | */ |
e5601400 | 1836 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 1837 | { |
2036d94a PM |
1838 | unsigned long flags; |
1839 | unsigned long mask; | |
1840 | int need_report = 0; | |
e5601400 | 1841 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 1842 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 1843 | |
2036d94a | 1844 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 1845 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 1846 | |
b1420f1c | 1847 | /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */ |
2036d94a PM |
1848 | |
1849 | /* Exclude any attempts to start a new grace period. */ | |
a4fbe35a | 1850 | mutex_lock(&rsp->onoff_mutex); |
7b2e6011 | 1851 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
2036d94a | 1852 | |
b1420f1c PM |
1853 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
1854 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
1855 | rcu_adopt_orphan_cbs(rsp); | |
1856 | ||
2036d94a PM |
1857 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
1858 | mask = rdp->grpmask; /* rnp->grplo is constant. */ | |
1859 | do { | |
1860 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
1861 | rnp->qsmaskinit &= ~mask; | |
1862 | if (rnp->qsmaskinit != 0) { | |
1863 | if (rnp != rdp->mynode) | |
1864 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1865 | break; | |
1866 | } | |
1867 | if (rnp == rdp->mynode) | |
1868 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); | |
1869 | else | |
1870 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1871 | mask = rnp->grpmask; | |
1872 | rnp = rnp->parent; | |
1873 | } while (rnp != NULL); | |
1874 | ||
1875 | /* | |
1876 | * We still hold the leaf rcu_node structure lock here, and | |
1877 | * irqs are still disabled. The reason for this subterfuge is | |
7b2e6011 | 1878 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock |
2036d94a PM |
1879 | * held leads to deadlock. |
1880 | */ | |
7b2e6011 | 1881 | raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */ |
2036d94a PM |
1882 | rnp = rdp->mynode; |
1883 | if (need_report & RCU_OFL_TASKS_NORM_GP) | |
1884 | rcu_report_unblock_qs_rnp(rnp, flags); | |
1885 | else | |
1886 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1887 | if (need_report & RCU_OFL_TASKS_EXP_GP) | |
1888 | rcu_report_exp_rnp(rsp, rnp, true); | |
cf01537e PM |
1889 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
1890 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
1891 | cpu, rdp->qlen, rdp->nxtlist); | |
0d8ee37e PM |
1892 | init_callback_list(rdp); |
1893 | /* Disallow further callbacks on this CPU. */ | |
1894 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; | |
a4fbe35a | 1895 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
1896 | } |
1897 | ||
1898 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1899 | ||
e5601400 | 1900 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
1901 | { |
1902 | } | |
1903 | ||
e5601400 | 1904 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
1905 | { |
1906 | } | |
1907 | ||
1908 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
1909 | ||
1910 | /* | |
1911 | * Invoke any RCU callbacks that have made it to the end of their grace | |
1912 | * period. Thottle as specified by rdp->blimit. | |
1913 | */ | |
37c72e56 | 1914 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1915 | { |
1916 | unsigned long flags; | |
1917 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
1918 | long bl, count, count_lazy; |
1919 | int i; | |
64db4cff | 1920 | |
dc35c893 | 1921 | /* If no callbacks are ready, just return. */ |
29c00b4a | 1922 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 1923 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
1924 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
1925 | need_resched(), is_idle_task(current), | |
1926 | rcu_is_callbacks_kthread()); | |
64db4cff | 1927 | return; |
29c00b4a | 1928 | } |
64db4cff PM |
1929 | |
1930 | /* | |
1931 | * Extract the list of ready callbacks, disabling to prevent | |
1932 | * races with call_rcu() from interrupt handlers. | |
1933 | */ | |
1934 | local_irq_save(flags); | |
8146c4e2 | 1935 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 1936 | bl = rdp->blimit; |
486e2593 | 1937 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
1938 | list = rdp->nxtlist; |
1939 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
1940 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
1941 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
1942 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
1943 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
1944 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
1945 | local_irq_restore(flags); |
1946 | ||
1947 | /* Invoke callbacks. */ | |
486e2593 | 1948 | count = count_lazy = 0; |
64db4cff PM |
1949 | while (list) { |
1950 | next = list->next; | |
1951 | prefetch(next); | |
551d55a9 | 1952 | debug_rcu_head_unqueue(list); |
486e2593 PM |
1953 | if (__rcu_reclaim(rsp->name, list)) |
1954 | count_lazy++; | |
64db4cff | 1955 | list = next; |
dff1672d PM |
1956 | /* Stop only if limit reached and CPU has something to do. */ |
1957 | if (++count >= bl && | |
1958 | (need_resched() || | |
1959 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
1960 | break; |
1961 | } | |
1962 | ||
1963 | local_irq_save(flags); | |
4968c300 PM |
1964 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
1965 | is_idle_task(current), | |
1966 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
1967 | |
1968 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
1969 | if (list != NULL) { |
1970 | *tail = rdp->nxtlist; | |
1971 | rdp->nxtlist = list; | |
b41772ab PM |
1972 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
1973 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
1974 | rdp->nxttail[i] = tail; | |
64db4cff PM |
1975 | else |
1976 | break; | |
1977 | } | |
b1420f1c PM |
1978 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
1979 | rdp->qlen_lazy -= count_lazy; | |
1d1fb395 | 1980 | ACCESS_ONCE(rdp->qlen) -= count; |
b1420f1c | 1981 | rdp->n_cbs_invoked += count; |
64db4cff PM |
1982 | |
1983 | /* Reinstate batch limit if we have worked down the excess. */ | |
1984 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
1985 | rdp->blimit = blimit; | |
1986 | ||
37c72e56 PM |
1987 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
1988 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
1989 | rdp->qlen_last_fqs_check = 0; | |
1990 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
1991 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
1992 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 1993 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 1994 | |
64db4cff PM |
1995 | local_irq_restore(flags); |
1996 | ||
e0f23060 | 1997 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 1998 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 1999 | invoke_rcu_core(); |
64db4cff PM |
2000 | } |
2001 | ||
2002 | /* | |
2003 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2004 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2005 | * Also schedule RCU core processing. |
64db4cff | 2006 | * |
9b2e4f18 | 2007 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2008 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2009 | * false, there is no point in invoking rcu_check_callbacks(). | |
2010 | */ | |
2011 | void rcu_check_callbacks(int cpu, int user) | |
2012 | { | |
300df91c | 2013 | trace_rcu_utilization("Start scheduler-tick"); |
a858af28 | 2014 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2015 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2016 | |
2017 | /* | |
2018 | * Get here if this CPU took its interrupt from user | |
2019 | * mode or from the idle loop, and if this is not a | |
2020 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2021 | * a quiescent state, so note it. |
64db4cff PM |
2022 | * |
2023 | * No memory barrier is required here because both | |
d6714c22 PM |
2024 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2025 | * variables that other CPUs neither access nor modify, | |
2026 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2027 | */ |
2028 | ||
d6714c22 PM |
2029 | rcu_sched_qs(cpu); |
2030 | rcu_bh_qs(cpu); | |
64db4cff PM |
2031 | |
2032 | } else if (!in_softirq()) { | |
2033 | ||
2034 | /* | |
2035 | * Get here if this CPU did not take its interrupt from | |
2036 | * softirq, in other words, if it is not interrupting | |
2037 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2038 | * critical section, so note it. |
64db4cff PM |
2039 | */ |
2040 | ||
d6714c22 | 2041 | rcu_bh_qs(cpu); |
64db4cff | 2042 | } |
f41d911f | 2043 | rcu_preempt_check_callbacks(cpu); |
d21670ac | 2044 | if (rcu_pending(cpu)) |
a46e0899 | 2045 | invoke_rcu_core(); |
300df91c | 2046 | trace_rcu_utilization("End scheduler-tick"); |
64db4cff PM |
2047 | } |
2048 | ||
64db4cff PM |
2049 | /* |
2050 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2051 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2052 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2053 | * | |
ee47eb9f | 2054 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2055 | */ |
45f014c5 | 2056 | static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *)) |
64db4cff PM |
2057 | { |
2058 | unsigned long bit; | |
2059 | int cpu; | |
2060 | unsigned long flags; | |
2061 | unsigned long mask; | |
a0b6c9a7 | 2062 | struct rcu_node *rnp; |
64db4cff | 2063 | |
a0b6c9a7 | 2064 | rcu_for_each_leaf_node(rsp, rnp) { |
b4be093f | 2065 | cond_resched(); |
64db4cff | 2066 | mask = 0; |
1304afb2 | 2067 | raw_spin_lock_irqsave(&rnp->lock, flags); |
ee47eb9f | 2068 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 2069 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 2070 | return; |
64db4cff | 2071 | } |
a0b6c9a7 | 2072 | if (rnp->qsmask == 0) { |
1217ed1b | 2073 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
64db4cff PM |
2074 | continue; |
2075 | } | |
a0b6c9a7 | 2076 | cpu = rnp->grplo; |
64db4cff | 2077 | bit = 1; |
a0b6c9a7 | 2078 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
394f99a9 LJ |
2079 | if ((rnp->qsmask & bit) != 0 && |
2080 | f(per_cpu_ptr(rsp->rda, cpu))) | |
64db4cff PM |
2081 | mask |= bit; |
2082 | } | |
45f014c5 | 2083 | if (mask != 0) { |
64db4cff | 2084 | |
d3f6bad3 PM |
2085 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
2086 | rcu_report_qs_rnp(mask, rsp, rnp, flags); | |
64db4cff PM |
2087 | continue; |
2088 | } | |
1304afb2 | 2089 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2090 | } |
27f4d280 | 2091 | rnp = rcu_get_root(rsp); |
1217ed1b PM |
2092 | if (rnp->qsmask == 0) { |
2093 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2094 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ | |
2095 | } | |
64db4cff PM |
2096 | } |
2097 | ||
2098 | /* | |
2099 | * Force quiescent states on reluctant CPUs, and also detect which | |
2100 | * CPUs are in dyntick-idle mode. | |
2101 | */ | |
4cdfc175 | 2102 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2103 | { |
2104 | unsigned long flags; | |
394f2769 PM |
2105 | bool ret; |
2106 | struct rcu_node *rnp; | |
2107 | struct rcu_node *rnp_old = NULL; | |
2108 | ||
2109 | /* Funnel through hierarchy to reduce memory contention. */ | |
2110 | rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode; | |
2111 | for (; rnp != NULL; rnp = rnp->parent) { | |
2112 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2113 | !raw_spin_trylock(&rnp->fqslock); | |
2114 | if (rnp_old != NULL) | |
2115 | raw_spin_unlock(&rnp_old->fqslock); | |
2116 | if (ret) { | |
2117 | rsp->n_force_qs_lh++; | |
2118 | return; | |
2119 | } | |
2120 | rnp_old = rnp; | |
2121 | } | |
2122 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2123 | |
394f2769 PM |
2124 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2125 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
2126 | raw_spin_unlock(&rnp_old->fqslock); | |
2127 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
2128 | rsp->n_force_qs_lh++; | |
2129 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); | |
4cdfc175 | 2130 | return; /* Someone beat us to it. */ |
46a1e34e | 2131 | } |
4cdfc175 | 2132 | rsp->gp_flags |= RCU_GP_FLAG_FQS; |
394f2769 | 2133 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2134 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ |
64db4cff PM |
2135 | } |
2136 | ||
64db4cff | 2137 | /* |
e0f23060 PM |
2138 | * This does the RCU core processing work for the specified rcu_state |
2139 | * and rcu_data structures. This may be called only from the CPU to | |
2140 | * whom the rdp belongs. | |
64db4cff PM |
2141 | */ |
2142 | static void | |
1bca8cf1 | 2143 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2144 | { |
2145 | unsigned long flags; | |
1bca8cf1 | 2146 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
64db4cff | 2147 | |
2e597558 PM |
2148 | WARN_ON_ONCE(rdp->beenonline == 0); |
2149 | ||
dc35c893 | 2150 | /* Handle the end of a grace period that some other CPU ended. */ |
64db4cff PM |
2151 | rcu_process_gp_end(rsp, rdp); |
2152 | ||
2153 | /* Update RCU state based on any recent quiescent states. */ | |
2154 | rcu_check_quiescent_state(rsp, rdp); | |
2155 | ||
2156 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2157 | local_irq_save(flags); |
64db4cff | 2158 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2159 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
64db4cff | 2160 | rcu_start_gp(rsp, flags); /* releases above lock */ |
dc35c893 PM |
2161 | } else { |
2162 | local_irq_restore(flags); | |
64db4cff PM |
2163 | } |
2164 | ||
2165 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2166 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2167 | invoke_rcu_callbacks(rsp, rdp); |
09223371 SL |
2168 | } |
2169 | ||
64db4cff | 2170 | /* |
e0f23060 | 2171 | * Do RCU core processing for the current CPU. |
64db4cff | 2172 | */ |
09223371 | 2173 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2174 | { |
6ce75a23 PM |
2175 | struct rcu_state *rsp; |
2176 | ||
bfa00b4c PM |
2177 | if (cpu_is_offline(smp_processor_id())) |
2178 | return; | |
300df91c | 2179 | trace_rcu_utilization("Start RCU core"); |
6ce75a23 PM |
2180 | for_each_rcu_flavor(rsp) |
2181 | __rcu_process_callbacks(rsp); | |
300df91c | 2182 | trace_rcu_utilization("End RCU core"); |
64db4cff PM |
2183 | } |
2184 | ||
a26ac245 | 2185 | /* |
e0f23060 PM |
2186 | * Schedule RCU callback invocation. If the specified type of RCU |
2187 | * does not support RCU priority boosting, just do a direct call, | |
2188 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
2189 | * are running on the current CPU with interrupts disabled, the | |
2190 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
a26ac245 | 2191 | */ |
a46e0899 | 2192 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2193 | { |
b0d30417 PM |
2194 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
2195 | return; | |
a46e0899 PM |
2196 | if (likely(!rsp->boost)) { |
2197 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2198 | return; |
2199 | } | |
a46e0899 | 2200 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2201 | } |
2202 | ||
a46e0899 | 2203 | static void invoke_rcu_core(void) |
09223371 SL |
2204 | { |
2205 | raise_softirq(RCU_SOFTIRQ); | |
2206 | } | |
2207 | ||
29154c57 PM |
2208 | /* |
2209 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2210 | */ | |
2211 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2212 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2213 | { |
62fde6ed PM |
2214 | /* |
2215 | * If called from an extended quiescent state, invoke the RCU | |
2216 | * core in order to force a re-evaluation of RCU's idleness. | |
2217 | */ | |
a16b7a69 | 2218 | if (rcu_is_cpu_idle() && cpu_online(smp_processor_id())) |
62fde6ed PM |
2219 | invoke_rcu_core(); |
2220 | ||
a16b7a69 | 2221 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2222 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2223 | return; |
64db4cff | 2224 | |
37c72e56 PM |
2225 | /* |
2226 | * Force the grace period if too many callbacks or too long waiting. | |
2227 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2228 | * if some other CPU has recently done so. Also, don't bother | |
2229 | * invoking force_quiescent_state() if the newly enqueued callback | |
2230 | * is the only one waiting for a grace period to complete. | |
2231 | */ | |
2655d57e | 2232 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2233 | |
2234 | /* Are we ignoring a completed grace period? */ | |
2235 | rcu_process_gp_end(rsp, rdp); | |
2236 | check_for_new_grace_period(rsp, rdp); | |
2237 | ||
2238 | /* Start a new grace period if one not already started. */ | |
2239 | if (!rcu_gp_in_progress(rsp)) { | |
2240 | unsigned long nestflag; | |
2241 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
2242 | ||
2243 | raw_spin_lock_irqsave(&rnp_root->lock, nestflag); | |
2244 | rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */ | |
2245 | } else { | |
2246 | /* Give the grace period a kick. */ | |
2247 | rdp->blimit = LONG_MAX; | |
2248 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2249 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2250 | force_quiescent_state(rsp); |
b52573d2 PM |
2251 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2252 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2253 | } | |
4cdfc175 | 2254 | } |
29154c57 PM |
2255 | } |
2256 | ||
3fbfbf7a PM |
2257 | /* |
2258 | * Helper function for call_rcu() and friends. The cpu argument will | |
2259 | * normally be -1, indicating "currently running CPU". It may specify | |
2260 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2261 | * is expected to specify a CPU. | |
2262 | */ | |
64db4cff PM |
2263 | static void |
2264 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 2265 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
2266 | { |
2267 | unsigned long flags; | |
2268 | struct rcu_data *rdp; | |
2269 | ||
0bb7b59d | 2270 | WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */ |
551d55a9 | 2271 | debug_rcu_head_queue(head); |
64db4cff PM |
2272 | head->func = func; |
2273 | head->next = NULL; | |
2274 | ||
64db4cff PM |
2275 | /* |
2276 | * Opportunistically note grace-period endings and beginnings. | |
2277 | * Note that we might see a beginning right after we see an | |
2278 | * end, but never vice versa, since this CPU has to pass through | |
2279 | * a quiescent state betweentimes. | |
2280 | */ | |
2281 | local_irq_save(flags); | |
394f99a9 | 2282 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
2283 | |
2284 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
2285 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
2286 | int offline; | |
2287 | ||
2288 | if (cpu != -1) | |
2289 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
2290 | offline = !__call_rcu_nocb(rdp, head, lazy); | |
2291 | WARN_ON_ONCE(offline); | |
0d8ee37e | 2292 | /* _call_rcu() is illegal on offline CPU; leak the callback. */ |
0d8ee37e PM |
2293 | local_irq_restore(flags); |
2294 | return; | |
2295 | } | |
29154c57 | 2296 | ACCESS_ONCE(rdp->qlen)++; |
486e2593 PM |
2297 | if (lazy) |
2298 | rdp->qlen_lazy++; | |
c57afe80 PM |
2299 | else |
2300 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2301 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
2302 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
2303 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 2304 | |
d4c08f2a PM |
2305 | if (__is_kfree_rcu_offset((unsigned long)func)) |
2306 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 2307 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2308 | else |
486e2593 | 2309 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2310 | |
29154c57 PM |
2311 | /* Go handle any RCU core processing required. */ |
2312 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
2313 | local_irq_restore(flags); |
2314 | } | |
2315 | ||
2316 | /* | |
d6714c22 | 2317 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 2318 | */ |
d6714c22 | 2319 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 2320 | { |
3fbfbf7a | 2321 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 2322 | } |
d6714c22 | 2323 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
2324 | |
2325 | /* | |
486e2593 | 2326 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
2327 | */ |
2328 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
2329 | { | |
3fbfbf7a | 2330 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
2331 | } |
2332 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
2333 | ||
6d813391 PM |
2334 | /* |
2335 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
2336 | * any blocking grace-period wait automatically implies a grace period | |
2337 | * if there is only one CPU online at any point time during execution | |
2338 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
2339 | * occasionally incorrectly indicate that there are multiple CPUs online | |
2340 | * when there was in fact only one the whole time, as this just adds | |
2341 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
2342 | */ |
2343 | static inline int rcu_blocking_is_gp(void) | |
2344 | { | |
95f0c1de PM |
2345 | int ret; |
2346 | ||
6d813391 | 2347 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
2348 | preempt_disable(); |
2349 | ret = num_online_cpus() <= 1; | |
2350 | preempt_enable(); | |
2351 | return ret; | |
6d813391 PM |
2352 | } |
2353 | ||
6ebb237b PM |
2354 | /** |
2355 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
2356 | * | |
2357 | * Control will return to the caller some time after a full rcu-sched | |
2358 | * grace period has elapsed, in other words after all currently executing | |
2359 | * rcu-sched read-side critical sections have completed. These read-side | |
2360 | * critical sections are delimited by rcu_read_lock_sched() and | |
2361 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
2362 | * local_irq_disable(), and so on may be used in place of | |
2363 | * rcu_read_lock_sched(). | |
2364 | * | |
2365 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
2366 | * non-threaded hardware-interrupt handlers, in progress on entry will |
2367 | * have completed before this primitive returns. However, this does not | |
2368 | * guarantee that softirq handlers will have completed, since in some | |
2369 | * kernels, these handlers can run in process context, and can block. | |
2370 | * | |
2371 | * Note that this guarantee implies further memory-ordering guarantees. | |
2372 | * On systems with more than one CPU, when synchronize_sched() returns, | |
2373 | * each CPU is guaranteed to have executed a full memory barrier since the | |
2374 | * end of its last RCU-sched read-side critical section whose beginning | |
2375 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
2376 | * an RCU read-side critical section that extends beyond the return from | |
2377 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
2378 | * after the beginning of synchronize_sched() and before the beginning of | |
2379 | * that RCU read-side critical section. Note that these guarantees include | |
2380 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
2381 | * that are executing in the kernel. | |
2382 | * | |
2383 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
2384 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
2385 | * to have executed a full memory barrier during the execution of | |
2386 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
2387 | * again only if the system has more than one CPU). | |
6ebb237b PM |
2388 | * |
2389 | * This primitive provides the guarantees made by the (now removed) | |
2390 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
2391 | * guarantees that rcu_read_lock() sections will have completed. | |
2392 | * In "classic RCU", these two guarantees happen to be one and | |
2393 | * the same, but can differ in realtime RCU implementations. | |
2394 | */ | |
2395 | void synchronize_sched(void) | |
2396 | { | |
fe15d706 PM |
2397 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2398 | !lock_is_held(&rcu_lock_map) && | |
2399 | !lock_is_held(&rcu_sched_lock_map), | |
2400 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
2401 | if (rcu_blocking_is_gp()) |
2402 | return; | |
3705b88d AM |
2403 | if (rcu_expedited) |
2404 | synchronize_sched_expedited(); | |
2405 | else | |
2406 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
2407 | } |
2408 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
2409 | ||
2410 | /** | |
2411 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
2412 | * | |
2413 | * Control will return to the caller some time after a full rcu_bh grace | |
2414 | * period has elapsed, in other words after all currently executing rcu_bh | |
2415 | * read-side critical sections have completed. RCU read-side critical | |
2416 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
2417 | * and may be nested. | |
f0a0e6f2 PM |
2418 | * |
2419 | * See the description of synchronize_sched() for more detailed information | |
2420 | * on memory ordering guarantees. | |
6ebb237b PM |
2421 | */ |
2422 | void synchronize_rcu_bh(void) | |
2423 | { | |
fe15d706 PM |
2424 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2425 | !lock_is_held(&rcu_lock_map) && | |
2426 | !lock_is_held(&rcu_sched_lock_map), | |
2427 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
2428 | if (rcu_blocking_is_gp()) |
2429 | return; | |
3705b88d AM |
2430 | if (rcu_expedited) |
2431 | synchronize_rcu_bh_expedited(); | |
2432 | else | |
2433 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
2434 | } |
2435 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
2436 | ||
3d3b7db0 PM |
2437 | static int synchronize_sched_expedited_cpu_stop(void *data) |
2438 | { | |
2439 | /* | |
2440 | * There must be a full memory barrier on each affected CPU | |
2441 | * between the time that try_stop_cpus() is called and the | |
2442 | * time that it returns. | |
2443 | * | |
2444 | * In the current initial implementation of cpu_stop, the | |
2445 | * above condition is already met when the control reaches | |
2446 | * this point and the following smp_mb() is not strictly | |
2447 | * necessary. Do smp_mb() anyway for documentation and | |
2448 | * robustness against future implementation changes. | |
2449 | */ | |
2450 | smp_mb(); /* See above comment block. */ | |
2451 | return 0; | |
2452 | } | |
2453 | ||
236fefaf PM |
2454 | /** |
2455 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
2456 | * | |
2457 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
2458 | * approach to force the grace period to end quickly. This consumes | |
2459 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
2460 | * so is thus not recommended for any sort of common-case code. In fact, | |
2461 | * if you are using synchronize_sched_expedited() in a loop, please | |
2462 | * restructure your code to batch your updates, and then use a single | |
2463 | * synchronize_sched() instead. | |
3d3b7db0 | 2464 | * |
236fefaf PM |
2465 | * Note that it is illegal to call this function while holding any lock |
2466 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal | |
2467 | * to call this function from a CPU-hotplug notifier. Failing to observe | |
2468 | * these restriction will result in deadlock. | |
3d3b7db0 PM |
2469 | * |
2470 | * This implementation can be thought of as an application of ticket | |
2471 | * locking to RCU, with sync_sched_expedited_started and | |
2472 | * sync_sched_expedited_done taking on the roles of the halves | |
2473 | * of the ticket-lock word. Each task atomically increments | |
2474 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
2475 | * then attempts to stop all the CPUs. If this succeeds, then each | |
2476 | * CPU will have executed a context switch, resulting in an RCU-sched | |
2477 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
2478 | * update sync_sched_expedited_done to match our snapshot -- but | |
2479 | * only if someone else has not already advanced past our snapshot. | |
2480 | * | |
2481 | * On the other hand, if try_stop_cpus() fails, we check the value | |
2482 | * of sync_sched_expedited_done. If it has advanced past our | |
2483 | * initial snapshot, then someone else must have forced a grace period | |
2484 | * some time after we took our snapshot. In this case, our work is | |
2485 | * done for us, and we can simply return. Otherwise, we try again, | |
2486 | * but keep our initial snapshot for purposes of checking for someone | |
2487 | * doing our work for us. | |
2488 | * | |
2489 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
2490 | */ | |
2491 | void synchronize_sched_expedited(void) | |
2492 | { | |
1924bcb0 PM |
2493 | long firstsnap, s, snap; |
2494 | int trycount = 0; | |
40694d66 | 2495 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 2496 | |
1924bcb0 PM |
2497 | /* |
2498 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
2499 | * By allowing sync_sched_expedited_started to advance no more than | |
2500 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
2501 | * that more than 3.5 billion CPUs would be required to force a | |
2502 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
2503 | * course be required on a 64-bit system. | |
2504 | */ | |
40694d66 PM |
2505 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
2506 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 PM |
2507 | ULONG_MAX / 8)) { |
2508 | synchronize_sched(); | |
a30489c5 | 2509 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
2510 | return; |
2511 | } | |
3d3b7db0 | 2512 | |
1924bcb0 PM |
2513 | /* |
2514 | * Take a ticket. Note that atomic_inc_return() implies a | |
2515 | * full memory barrier. | |
2516 | */ | |
40694d66 | 2517 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 2518 | firstsnap = snap; |
3d3b7db0 | 2519 | get_online_cpus(); |
1cc85961 | 2520 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 PM |
2521 | |
2522 | /* | |
2523 | * Each pass through the following loop attempts to force a | |
2524 | * context switch on each CPU. | |
2525 | */ | |
2526 | while (try_stop_cpus(cpu_online_mask, | |
2527 | synchronize_sched_expedited_cpu_stop, | |
2528 | NULL) == -EAGAIN) { | |
2529 | put_online_cpus(); | |
a30489c5 | 2530 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 2531 | |
1924bcb0 | 2532 | /* Check to see if someone else did our work for us. */ |
40694d66 | 2533 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2534 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2535 | /* ensure test happens before caller kfree */ |
2536 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2537 | atomic_long_inc(&rsp->expedited_workdone1); | |
1924bcb0 PM |
2538 | return; |
2539 | } | |
3d3b7db0 PM |
2540 | |
2541 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 2542 | if (trycount++ < 10) { |
3d3b7db0 | 2543 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 2544 | } else { |
3705b88d | 2545 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 2546 | atomic_long_inc(&rsp->expedited_normal); |
3d3b7db0 PM |
2547 | return; |
2548 | } | |
2549 | ||
1924bcb0 | 2550 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 2551 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2552 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2553 | /* ensure test happens before caller kfree */ |
2554 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2555 | atomic_long_inc(&rsp->expedited_workdone2); | |
3d3b7db0 PM |
2556 | return; |
2557 | } | |
2558 | ||
2559 | /* | |
2560 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
2561 | * callers to piggyback on our grace period. We retry |
2562 | * after they started, so our grace period works for them, | |
2563 | * and they started after our first try, so their grace | |
2564 | * period works for us. | |
3d3b7db0 PM |
2565 | */ |
2566 | get_online_cpus(); | |
40694d66 | 2567 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
2568 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
2569 | } | |
a30489c5 | 2570 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 PM |
2571 | |
2572 | /* | |
2573 | * Everyone up to our most recent fetch is covered by our grace | |
2574 | * period. Update the counter, but only if our work is still | |
2575 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 2576 | * than we did already did their update. |
3d3b7db0 PM |
2577 | */ |
2578 | do { | |
a30489c5 | 2579 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 2580 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2581 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 PM |
2582 | /* ensure test happens before caller kfree */ |
2583 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2584 | atomic_long_inc(&rsp->expedited_done_lost); | |
3d3b7db0 PM |
2585 | break; |
2586 | } | |
40694d66 | 2587 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 2588 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
2589 | |
2590 | put_online_cpus(); | |
2591 | } | |
2592 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
2593 | ||
64db4cff PM |
2594 | /* |
2595 | * Check to see if there is any immediate RCU-related work to be done | |
2596 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
2597 | * The checks are in order of increasing expense: checks that can be | |
2598 | * carried out against CPU-local state are performed first. However, | |
2599 | * we must check for CPU stalls first, else we might not get a chance. | |
2600 | */ | |
2601 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
2602 | { | |
2f51f988 PM |
2603 | struct rcu_node *rnp = rdp->mynode; |
2604 | ||
64db4cff PM |
2605 | rdp->n_rcu_pending++; |
2606 | ||
2607 | /* Check for CPU stalls, if enabled. */ | |
2608 | check_cpu_stall(rsp, rdp); | |
2609 | ||
2610 | /* Is the RCU core waiting for a quiescent state from this CPU? */ | |
5c51dd73 PM |
2611 | if (rcu_scheduler_fully_active && |
2612 | rdp->qs_pending && !rdp->passed_quiesce) { | |
d21670ac | 2613 | rdp->n_rp_qs_pending++; |
e4cc1f22 | 2614 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
d21670ac | 2615 | rdp->n_rp_report_qs++; |
64db4cff | 2616 | return 1; |
7ba5c840 | 2617 | } |
64db4cff PM |
2618 | |
2619 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
2620 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
2621 | rdp->n_rp_cb_ready++; | |
64db4cff | 2622 | return 1; |
7ba5c840 | 2623 | } |
64db4cff PM |
2624 | |
2625 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
2626 | if (cpu_needs_another_gp(rsp, rdp)) { |
2627 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 2628 | return 1; |
7ba5c840 | 2629 | } |
64db4cff PM |
2630 | |
2631 | /* Has another RCU grace period completed? */ | |
2f51f988 | 2632 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 2633 | rdp->n_rp_gp_completed++; |
64db4cff | 2634 | return 1; |
7ba5c840 | 2635 | } |
64db4cff PM |
2636 | |
2637 | /* Has a new RCU grace period started? */ | |
2f51f988 | 2638 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
7ba5c840 | 2639 | rdp->n_rp_gp_started++; |
64db4cff | 2640 | return 1; |
7ba5c840 | 2641 | } |
64db4cff | 2642 | |
64db4cff | 2643 | /* nothing to do */ |
7ba5c840 | 2644 | rdp->n_rp_need_nothing++; |
64db4cff PM |
2645 | return 0; |
2646 | } | |
2647 | ||
2648 | /* | |
2649 | * Check to see if there is any immediate RCU-related work to be done | |
2650 | * by the current CPU, returning 1 if so. This function is part of the | |
2651 | * RCU implementation; it is -not- an exported member of the RCU API. | |
2652 | */ | |
a157229c | 2653 | static int rcu_pending(int cpu) |
64db4cff | 2654 | { |
6ce75a23 PM |
2655 | struct rcu_state *rsp; |
2656 | ||
2657 | for_each_rcu_flavor(rsp) | |
2658 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu))) | |
2659 | return 1; | |
2660 | return 0; | |
64db4cff PM |
2661 | } |
2662 | ||
2663 | /* | |
2664 | * Check to see if any future RCU-related work will need to be done | |
2665 | * by the current CPU, even if none need be done immediately, returning | |
8bd93a2c | 2666 | * 1 if so. |
64db4cff | 2667 | */ |
aea1b35e | 2668 | static int rcu_cpu_has_callbacks(int cpu) |
64db4cff | 2669 | { |
6ce75a23 PM |
2670 | struct rcu_state *rsp; |
2671 | ||
64db4cff | 2672 | /* RCU callbacks either ready or pending? */ |
6ce75a23 PM |
2673 | for_each_rcu_flavor(rsp) |
2674 | if (per_cpu_ptr(rsp->rda, cpu)->nxtlist) | |
2675 | return 1; | |
2676 | return 0; | |
64db4cff PM |
2677 | } |
2678 | ||
a83eff0a PM |
2679 | /* |
2680 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
2681 | * the compiler is expected to optimize this away. | |
2682 | */ | |
2683 | static void _rcu_barrier_trace(struct rcu_state *rsp, char *s, | |
2684 | int cpu, unsigned long done) | |
2685 | { | |
2686 | trace_rcu_barrier(rsp->name, s, cpu, | |
2687 | atomic_read(&rsp->barrier_cpu_count), done); | |
2688 | } | |
2689 | ||
b1420f1c PM |
2690 | /* |
2691 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
2692 | * up the task executing _rcu_barrier(). | |
2693 | */ | |
24ebbca8 | 2694 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 2695 | { |
24ebbca8 PM |
2696 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
2697 | struct rcu_state *rsp = rdp->rsp; | |
2698 | ||
a83eff0a PM |
2699 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
2700 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 2701 | complete(&rsp->barrier_completion); |
a83eff0a PM |
2702 | } else { |
2703 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
2704 | } | |
d0ec774c PM |
2705 | } |
2706 | ||
2707 | /* | |
2708 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
2709 | */ | |
2710 | static void rcu_barrier_func(void *type) | |
2711 | { | |
037b64ed | 2712 | struct rcu_state *rsp = type; |
06668efa | 2713 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
d0ec774c | 2714 | |
a83eff0a | 2715 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 2716 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 2717 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
2718 | } |
2719 | ||
d0ec774c PM |
2720 | /* |
2721 | * Orchestrate the specified type of RCU barrier, waiting for all | |
2722 | * RCU callbacks of the specified type to complete. | |
2723 | */ | |
037b64ed | 2724 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 2725 | { |
b1420f1c | 2726 | int cpu; |
b1420f1c | 2727 | struct rcu_data *rdp; |
cf3a9c48 PM |
2728 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
2729 | unsigned long snap_done; | |
b1420f1c | 2730 | |
a83eff0a | 2731 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 2732 | |
e74f4c45 | 2733 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 2734 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 2735 | |
cf3a9c48 PM |
2736 | /* |
2737 | * Ensure that all prior references, including to ->n_barrier_done, | |
2738 | * are ordered before the _rcu_barrier() machinery. | |
2739 | */ | |
2740 | smp_mb(); /* See above block comment. */ | |
2741 | ||
2742 | /* | |
2743 | * Recheck ->n_barrier_done to see if others did our work for us. | |
2744 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
2745 | * transition. The "if" expression below therefore rounds the old | |
2746 | * value up to the next even number and adds two before comparing. | |
2747 | */ | |
2748 | snap_done = ACCESS_ONCE(rsp->n_barrier_done); | |
a83eff0a | 2749 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
cf3a9c48 | 2750 | if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) { |
a83eff0a | 2751 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
2752 | smp_mb(); /* caller's subsequent code after above check. */ |
2753 | mutex_unlock(&rsp->barrier_mutex); | |
2754 | return; | |
2755 | } | |
2756 | ||
2757 | /* | |
2758 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
2759 | * ACCESS_ONCE() to prevent the compiler from speculating | |
2760 | * the increment to precede the early-exit check. | |
2761 | */ | |
2762 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2763 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); | |
a83eff0a | 2764 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 2765 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 2766 | |
d0ec774c | 2767 | /* |
b1420f1c PM |
2768 | * Initialize the count to one rather than to zero in order to |
2769 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
2770 | * (or preemption of this task). Exclude CPU-hotplug operations |
2771 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 2772 | */ |
7db74df8 | 2773 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 2774 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 2775 | get_online_cpus(); |
b1420f1c PM |
2776 | |
2777 | /* | |
1331e7a1 PM |
2778 | * Force each CPU with callbacks to register a new callback. |
2779 | * When that callback is invoked, we will know that all of the | |
2780 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 2781 | */ |
3fbfbf7a PM |
2782 | for_each_possible_cpu(cpu) { |
2783 | if (!cpu_online(cpu) && !is_nocb_cpu(cpu)) | |
2784 | continue; | |
b1420f1c | 2785 | rdp = per_cpu_ptr(rsp->rda, cpu); |
3fbfbf7a PM |
2786 | if (is_nocb_cpu(cpu)) { |
2787 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, | |
2788 | rsp->n_barrier_done); | |
2789 | atomic_inc(&rsp->barrier_cpu_count); | |
2790 | __call_rcu(&rdp->barrier_head, rcu_barrier_callback, | |
2791 | rsp, cpu, 0); | |
2792 | } else if (ACCESS_ONCE(rdp->qlen)) { | |
a83eff0a PM |
2793 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
2794 | rsp->n_barrier_done); | |
037b64ed | 2795 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 2796 | } else { |
a83eff0a PM |
2797 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
2798 | rsp->n_barrier_done); | |
b1420f1c PM |
2799 | } |
2800 | } | |
1331e7a1 | 2801 | put_online_cpus(); |
b1420f1c PM |
2802 | |
2803 | /* | |
2804 | * Now that we have an rcu_barrier_callback() callback on each | |
2805 | * CPU, and thus each counted, remove the initial count. | |
2806 | */ | |
24ebbca8 | 2807 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 2808 | complete(&rsp->barrier_completion); |
b1420f1c | 2809 | |
cf3a9c48 PM |
2810 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
2811 | smp_mb(); /* Keep increment after above mechanism. */ | |
2812 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2813 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); | |
a83eff0a | 2814 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
2815 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
2816 | ||
b1420f1c | 2817 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 2818 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
2819 | |
2820 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 2821 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 2822 | } |
d0ec774c PM |
2823 | |
2824 | /** | |
2825 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
2826 | */ | |
2827 | void rcu_barrier_bh(void) | |
2828 | { | |
037b64ed | 2829 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
2830 | } |
2831 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
2832 | ||
2833 | /** | |
2834 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
2835 | */ | |
2836 | void rcu_barrier_sched(void) | |
2837 | { | |
037b64ed | 2838 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
2839 | } |
2840 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
2841 | ||
64db4cff | 2842 | /* |
27569620 | 2843 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 2844 | */ |
27569620 PM |
2845 | static void __init |
2846 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
2847 | { |
2848 | unsigned long flags; | |
394f99a9 | 2849 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
2850 | struct rcu_node *rnp = rcu_get_root(rsp); |
2851 | ||
2852 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 2853 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 2854 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
3f5d3ea6 | 2855 | init_callback_list(rdp); |
486e2593 | 2856 | rdp->qlen_lazy = 0; |
1d1fb395 | 2857 | ACCESS_ONCE(rdp->qlen) = 0; |
27569620 | 2858 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 2859 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 2860 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
c5d900bf FW |
2861 | #ifdef CONFIG_RCU_USER_QS |
2862 | WARN_ON_ONCE(rdp->dynticks->in_user); | |
2863 | #endif | |
27569620 | 2864 | rdp->cpu = cpu; |
d4c08f2a | 2865 | rdp->rsp = rsp; |
3fbfbf7a | 2866 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 2867 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
2868 | } |
2869 | ||
2870 | /* | |
2871 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
2872 | * offline event can be happening at a given time. Note also that we | |
2873 | * can accept some slop in the rsp->completed access due to the fact | |
2874 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 2875 | */ |
e4fa4c97 | 2876 | static void __cpuinit |
6cc68793 | 2877 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible) |
64db4cff PM |
2878 | { |
2879 | unsigned long flags; | |
64db4cff | 2880 | unsigned long mask; |
394f99a9 | 2881 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
2882 | struct rcu_node *rnp = rcu_get_root(rsp); |
2883 | ||
a4fbe35a PM |
2884 | /* Exclude new grace periods. */ |
2885 | mutex_lock(&rsp->onoff_mutex); | |
2886 | ||
64db4cff | 2887 | /* Set up local state, ensuring consistent view of global state. */ |
1304afb2 | 2888 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 2889 | rdp->beenonline = 1; /* We have now been online. */ |
6cc68793 | 2890 | rdp->preemptible = preemptible; |
37c72e56 PM |
2891 | rdp->qlen_last_fqs_check = 0; |
2892 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 2893 | rdp->blimit = blimit; |
0d8ee37e | 2894 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ |
29e37d81 | 2895 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
c92b131b PM |
2896 | atomic_set(&rdp->dynticks->dynticks, |
2897 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
7cb92499 | 2898 | rcu_prepare_for_idle_init(cpu); |
1304afb2 | 2899 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 2900 | |
64db4cff PM |
2901 | /* Add CPU to rcu_node bitmasks. */ |
2902 | rnp = rdp->mynode; | |
2903 | mask = rdp->grpmask; | |
2904 | do { | |
2905 | /* Exclude any attempts to start a new GP on small systems. */ | |
1304afb2 | 2906 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
2907 | rnp->qsmaskinit |= mask; |
2908 | mask = rnp->grpmask; | |
d09b62df | 2909 | if (rnp == rdp->mynode) { |
06ae115a PM |
2910 | /* |
2911 | * If there is a grace period in progress, we will | |
2912 | * set up to wait for it next time we run the | |
2913 | * RCU core code. | |
2914 | */ | |
2915 | rdp->gpnum = rnp->completed; | |
d09b62df | 2916 | rdp->completed = rnp->completed; |
06ae115a PM |
2917 | rdp->passed_quiesce = 0; |
2918 | rdp->qs_pending = 0; | |
d4c08f2a | 2919 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl"); |
d09b62df | 2920 | } |
1304afb2 | 2921 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
2922 | rnp = rnp->parent; |
2923 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); | |
a4fbe35a | 2924 | local_irq_restore(flags); |
64db4cff | 2925 | |
a4fbe35a | 2926 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
2927 | } |
2928 | ||
d72bce0e | 2929 | static void __cpuinit rcu_prepare_cpu(int cpu) |
64db4cff | 2930 | { |
6ce75a23 PM |
2931 | struct rcu_state *rsp; |
2932 | ||
2933 | for_each_rcu_flavor(rsp) | |
2934 | rcu_init_percpu_data(cpu, rsp, | |
2935 | strcmp(rsp->name, "rcu_preempt") == 0); | |
64db4cff PM |
2936 | } |
2937 | ||
2938 | /* | |
f41d911f | 2939 | * Handle CPU online/offline notification events. |
64db4cff | 2940 | */ |
9f680ab4 PM |
2941 | static int __cpuinit rcu_cpu_notify(struct notifier_block *self, |
2942 | unsigned long action, void *hcpu) | |
64db4cff PM |
2943 | { |
2944 | long cpu = (long)hcpu; | |
27f4d280 | 2945 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); |
a26ac245 | 2946 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 2947 | struct rcu_state *rsp; |
3fbfbf7a | 2948 | int ret = NOTIFY_OK; |
64db4cff | 2949 | |
300df91c | 2950 | trace_rcu_utilization("Start CPU hotplug"); |
64db4cff PM |
2951 | switch (action) { |
2952 | case CPU_UP_PREPARE: | |
2953 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
2954 | rcu_prepare_cpu(cpu); |
2955 | rcu_prepare_kthreads(cpu); | |
a26ac245 PM |
2956 | break; |
2957 | case CPU_ONLINE: | |
0f962a5e | 2958 | case CPU_DOWN_FAILED: |
5d01bbd1 | 2959 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
2960 | break; |
2961 | case CPU_DOWN_PREPARE: | |
3fbfbf7a PM |
2962 | if (nocb_cpu_expendable(cpu)) |
2963 | rcu_boost_kthread_setaffinity(rnp, cpu); | |
2964 | else | |
2965 | ret = NOTIFY_BAD; | |
64db4cff | 2966 | break; |
d0ec774c PM |
2967 | case CPU_DYING: |
2968 | case CPU_DYING_FROZEN: | |
2969 | /* | |
2d999e03 PM |
2970 | * The whole machine is "stopped" except this CPU, so we can |
2971 | * touch any data without introducing corruption. We send the | |
2972 | * dying CPU's callbacks to an arbitrarily chosen online CPU. | |
d0ec774c | 2973 | */ |
6ce75a23 PM |
2974 | for_each_rcu_flavor(rsp) |
2975 | rcu_cleanup_dying_cpu(rsp); | |
7cb92499 | 2976 | rcu_cleanup_after_idle(cpu); |
d0ec774c | 2977 | break; |
64db4cff PM |
2978 | case CPU_DEAD: |
2979 | case CPU_DEAD_FROZEN: | |
2980 | case CPU_UP_CANCELED: | |
2981 | case CPU_UP_CANCELED_FROZEN: | |
6ce75a23 PM |
2982 | for_each_rcu_flavor(rsp) |
2983 | rcu_cleanup_dead_cpu(cpu, rsp); | |
64db4cff PM |
2984 | break; |
2985 | default: | |
2986 | break; | |
2987 | } | |
300df91c | 2988 | trace_rcu_utilization("End CPU hotplug"); |
3fbfbf7a | 2989 | return ret; |
64db4cff PM |
2990 | } |
2991 | ||
b3dbec76 PM |
2992 | /* |
2993 | * Spawn the kthread that handles this RCU flavor's grace periods. | |
2994 | */ | |
2995 | static int __init rcu_spawn_gp_kthread(void) | |
2996 | { | |
2997 | unsigned long flags; | |
2998 | struct rcu_node *rnp; | |
2999 | struct rcu_state *rsp; | |
3000 | struct task_struct *t; | |
3001 | ||
3002 | for_each_rcu_flavor(rsp) { | |
3003 | t = kthread_run(rcu_gp_kthread, rsp, rsp->name); | |
3004 | BUG_ON(IS_ERR(t)); | |
3005 | rnp = rcu_get_root(rsp); | |
3006 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3007 | rsp->gp_kthread = t; | |
3008 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3fbfbf7a | 3009 | rcu_spawn_nocb_kthreads(rsp); |
b3dbec76 PM |
3010 | } |
3011 | return 0; | |
3012 | } | |
3013 | early_initcall(rcu_spawn_gp_kthread); | |
3014 | ||
bbad9379 PM |
3015 | /* |
3016 | * This function is invoked towards the end of the scheduler's initialization | |
3017 | * process. Before this is called, the idle task might contain | |
3018 | * RCU read-side critical sections (during which time, this idle | |
3019 | * task is booting the system). After this function is called, the | |
3020 | * idle tasks are prohibited from containing RCU read-side critical | |
3021 | * sections. This function also enables RCU lockdep checking. | |
3022 | */ | |
3023 | void rcu_scheduler_starting(void) | |
3024 | { | |
3025 | WARN_ON(num_online_cpus() != 1); | |
3026 | WARN_ON(nr_context_switches() > 0); | |
3027 | rcu_scheduler_active = 1; | |
3028 | } | |
3029 | ||
64db4cff PM |
3030 | /* |
3031 | * Compute the per-level fanout, either using the exact fanout specified | |
3032 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
3033 | */ | |
3034 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
3035 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3036 | { | |
3037 | int i; | |
3038 | ||
f885b7f2 | 3039 | for (i = rcu_num_lvls - 1; i > 0; i--) |
64db4cff | 3040 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
f885b7f2 | 3041 | rsp->levelspread[0] = rcu_fanout_leaf; |
64db4cff PM |
3042 | } |
3043 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3044 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3045 | { | |
3046 | int ccur; | |
3047 | int cprv; | |
3048 | int i; | |
3049 | ||
4dbd6bb3 | 3050 | cprv = nr_cpu_ids; |
f885b7f2 | 3051 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3052 | ccur = rsp->levelcnt[i]; |
3053 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
3054 | cprv = ccur; | |
3055 | } | |
3056 | } | |
3057 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3058 | ||
3059 | /* | |
3060 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3061 | */ | |
394f99a9 LJ |
3062 | static void __init rcu_init_one(struct rcu_state *rsp, |
3063 | struct rcu_data __percpu *rda) | |
64db4cff | 3064 | { |
394f2769 PM |
3065 | static char *buf[] = { "rcu_node_0", |
3066 | "rcu_node_1", | |
3067 | "rcu_node_2", | |
3068 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ | |
3069 | static char *fqs[] = { "rcu_node_fqs_0", | |
3070 | "rcu_node_fqs_1", | |
3071 | "rcu_node_fqs_2", | |
3072 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ | |
64db4cff PM |
3073 | int cpustride = 1; |
3074 | int i; | |
3075 | int j; | |
3076 | struct rcu_node *rnp; | |
3077 | ||
b6407e86 PM |
3078 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
3079 | ||
64db4cff PM |
3080 | /* Initialize the level-tracking arrays. */ |
3081 | ||
f885b7f2 PM |
3082 | for (i = 0; i < rcu_num_lvls; i++) |
3083 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
3084 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
3085 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
3086 | rcu_init_levelspread(rsp); | |
3087 | ||
3088 | /* Initialize the elements themselves, starting from the leaves. */ | |
3089 | ||
f885b7f2 | 3090 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3091 | cpustride *= rsp->levelspread[i]; |
3092 | rnp = rsp->level[i]; | |
3093 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 3094 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
3095 | lockdep_set_class_and_name(&rnp->lock, |
3096 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
3097 | raw_spin_lock_init(&rnp->fqslock); |
3098 | lockdep_set_class_and_name(&rnp->fqslock, | |
3099 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
3100 | rnp->gpnum = rsp->gpnum; |
3101 | rnp->completed = rsp->completed; | |
64db4cff PM |
3102 | rnp->qsmask = 0; |
3103 | rnp->qsmaskinit = 0; | |
3104 | rnp->grplo = j * cpustride; | |
3105 | rnp->grphi = (j + 1) * cpustride - 1; | |
3106 | if (rnp->grphi >= NR_CPUS) | |
3107 | rnp->grphi = NR_CPUS - 1; | |
3108 | if (i == 0) { | |
3109 | rnp->grpnum = 0; | |
3110 | rnp->grpmask = 0; | |
3111 | rnp->parent = NULL; | |
3112 | } else { | |
3113 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
3114 | rnp->grpmask = 1UL << rnp->grpnum; | |
3115 | rnp->parent = rsp->level[i - 1] + | |
3116 | j / rsp->levelspread[i - 1]; | |
3117 | } | |
3118 | rnp->level = i; | |
12f5f524 | 3119 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
64db4cff PM |
3120 | } |
3121 | } | |
0c34029a | 3122 | |
394f99a9 | 3123 | rsp->rda = rda; |
b3dbec76 | 3124 | init_waitqueue_head(&rsp->gp_wq); |
f885b7f2 | 3125 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 3126 | for_each_possible_cpu(i) { |
4a90a068 | 3127 | while (i > rnp->grphi) |
0c34029a | 3128 | rnp++; |
394f99a9 | 3129 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
3130 | rcu_boot_init_percpu_data(i, rsp); |
3131 | } | |
6ce75a23 | 3132 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
3133 | } |
3134 | ||
f885b7f2 PM |
3135 | /* |
3136 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
3137 | * replace the definitions in rcutree.h because those are needed to size | |
3138 | * the ->node array in the rcu_state structure. | |
3139 | */ | |
3140 | static void __init rcu_init_geometry(void) | |
3141 | { | |
3142 | int i; | |
3143 | int j; | |
cca6f393 | 3144 | int n = nr_cpu_ids; |
f885b7f2 PM |
3145 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
3146 | ||
3147 | /* If the compile-time values are accurate, just leave. */ | |
b17c7035 PM |
3148 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
3149 | nr_cpu_ids == NR_CPUS) | |
f885b7f2 PM |
3150 | return; |
3151 | ||
3152 | /* | |
3153 | * Compute number of nodes that can be handled an rcu_node tree | |
3154 | * with the given number of levels. Setting rcu_capacity[0] makes | |
3155 | * some of the arithmetic easier. | |
3156 | */ | |
3157 | rcu_capacity[0] = 1; | |
3158 | rcu_capacity[1] = rcu_fanout_leaf; | |
3159 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
3160 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
3161 | ||
3162 | /* | |
3163 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
3164 | * to increase the leaf-level fanout, not decrease it. Of course, | |
3165 | * the leaf-level fanout cannot exceed the number of bits in | |
3166 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
3167 | * the configured number of CPUs. Complain and fall back to the | |
3168 | * compile-time values if these limits are exceeded. | |
3169 | */ | |
3170 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
3171 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
3172 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
3173 | WARN_ON(1); | |
3174 | return; | |
3175 | } | |
3176 | ||
3177 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
3178 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
3179 | if (n <= rcu_capacity[i]) { | |
3180 | for (j = 0; j <= i; j++) | |
3181 | num_rcu_lvl[j] = | |
3182 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
3183 | rcu_num_lvls = i; | |
3184 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
3185 | num_rcu_lvl[j] = 0; | |
3186 | break; | |
3187 | } | |
3188 | ||
3189 | /* Calculate the total number of rcu_node structures. */ | |
3190 | rcu_num_nodes = 0; | |
3191 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
3192 | rcu_num_nodes += num_rcu_lvl[i]; | |
3193 | rcu_num_nodes -= n; | |
3194 | } | |
3195 | ||
9f680ab4 | 3196 | void __init rcu_init(void) |
64db4cff | 3197 | { |
017c4261 | 3198 | int cpu; |
9f680ab4 | 3199 | |
f41d911f | 3200 | rcu_bootup_announce(); |
f885b7f2 | 3201 | rcu_init_geometry(); |
394f99a9 LJ |
3202 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
3203 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); | |
f41d911f | 3204 | __rcu_init_preempt(); |
3fbfbf7a | 3205 | rcu_init_nocb(); |
09223371 | 3206 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
3207 | |
3208 | /* | |
3209 | * We don't need protection against CPU-hotplug here because | |
3210 | * this is called early in boot, before either interrupts | |
3211 | * or the scheduler are operational. | |
3212 | */ | |
3213 | cpu_notifier(rcu_cpu_notify, 0); | |
017c4261 PM |
3214 | for_each_online_cpu(cpu) |
3215 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
c68de209 | 3216 | check_cpu_stall_init(); |
64db4cff PM |
3217 | } |
3218 | ||
1eba8f84 | 3219 | #include "rcutree_plugin.h" |