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 | |
a4889858 | 67 | #define RCU_STATE_INITIALIZER(sname, sabbr, 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, \ |
a4889858 | 79 | .abbr = sabbr, \ |
64db4cff PM |
80 | } |
81 | ||
037b64ed | 82 | struct rcu_state rcu_sched_state = |
a4889858 | 83 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); |
d6714c22 | 84 | DEFINE_PER_CPU(struct rcu_data, rcu_sched_data); |
64db4cff | 85 | |
a4889858 | 86 | struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); |
6258c4fb | 87 | DEFINE_PER_CPU(struct rcu_data, rcu_bh_data); |
b1f77b05 | 88 | |
27f4d280 | 89 | static struct rcu_state *rcu_state; |
6ce75a23 | 90 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 91 | |
f885b7f2 PM |
92 | /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */ |
93 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF; | |
7e5c2dfb | 94 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 PM |
95 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
96 | static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */ | |
97 | NUM_RCU_LVL_0, | |
98 | NUM_RCU_LVL_1, | |
99 | NUM_RCU_LVL_2, | |
100 | NUM_RCU_LVL_3, | |
101 | NUM_RCU_LVL_4, | |
102 | }; | |
103 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ | |
104 | ||
b0d30417 PM |
105 | /* |
106 | * The rcu_scheduler_active variable transitions from zero to one just | |
107 | * before the first task is spawned. So when this variable is zero, RCU | |
108 | * can assume that there is but one task, allowing RCU to (for example) | |
b44f6656 | 109 | * optimize synchronize_sched() to a simple barrier(). When this variable |
b0d30417 PM |
110 | * is one, RCU must actually do all the hard work required to detect real |
111 | * grace periods. This variable is also used to suppress boot-time false | |
112 | * positives from lockdep-RCU error checking. | |
113 | */ | |
bbad9379 PM |
114 | int rcu_scheduler_active __read_mostly; |
115 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
116 | ||
b0d30417 PM |
117 | /* |
118 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
119 | * during the early_initcall() processing, which is after the scheduler | |
120 | * is capable of creating new tasks. So RCU processing (for example, | |
121 | * creating tasks for RCU priority boosting) must be delayed until after | |
122 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
123 | * currently delay invocation of any RCU callbacks until after this point. | |
124 | * | |
125 | * It might later prove better for people registering RCU callbacks during | |
126 | * early boot to take responsibility for these callbacks, but one step at | |
127 | * a time. | |
128 | */ | |
129 | static int rcu_scheduler_fully_active __read_mostly; | |
130 | ||
a46e0899 PM |
131 | #ifdef CONFIG_RCU_BOOST |
132 | ||
a26ac245 PM |
133 | /* |
134 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
135 | * handle all flavors of RCU. | |
136 | */ | |
137 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
d71df90e | 138 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); |
5ece5bab | 139 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); |
d71df90e | 140 | DEFINE_PER_CPU(char, rcu_cpu_has_work); |
a26ac245 | 141 | |
a46e0899 PM |
142 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
143 | ||
5d01bbd1 | 144 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
145 | static void invoke_rcu_core(void); |
146 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 147 | |
4a298656 PM |
148 | /* |
149 | * Track the rcutorture test sequence number and the update version | |
150 | * number within a given test. The rcutorture_testseq is incremented | |
151 | * on every rcutorture module load and unload, so has an odd value | |
152 | * when a test is running. The rcutorture_vernum is set to zero | |
153 | * when rcutorture starts and is incremented on each rcutorture update. | |
154 | * These variables enable correlating rcutorture output with the | |
155 | * RCU tracing information. | |
156 | */ | |
157 | unsigned long rcutorture_testseq; | |
158 | unsigned long rcutorture_vernum; | |
159 | ||
fc2219d4 PM |
160 | /* |
161 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
162 | * permit this function to be invoked without holding the root rcu_node | |
163 | * structure's ->lock, but of course results can be subject to change. | |
164 | */ | |
165 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
166 | { | |
167 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
168 | } | |
169 | ||
b1f77b05 | 170 | /* |
d6714c22 | 171 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 172 | * how many quiescent states passed, just if there was at least |
d6714c22 | 173 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 174 | * The caller must have disabled preemption. |
b1f77b05 | 175 | */ |
d6714c22 | 176 | void rcu_sched_qs(int cpu) |
b1f77b05 | 177 | { |
25502a6c | 178 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu); |
f41d911f | 179 | |
e4cc1f22 | 180 | if (rdp->passed_quiesce == 0) |
d4c08f2a | 181 | trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs"); |
e4cc1f22 | 182 | rdp->passed_quiesce = 1; |
b1f77b05 IM |
183 | } |
184 | ||
d6714c22 | 185 | void rcu_bh_qs(int cpu) |
b1f77b05 | 186 | { |
25502a6c | 187 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu); |
f41d911f | 188 | |
e4cc1f22 | 189 | if (rdp->passed_quiesce == 0) |
d4c08f2a | 190 | trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs"); |
e4cc1f22 | 191 | rdp->passed_quiesce = 1; |
b1f77b05 | 192 | } |
64db4cff | 193 | |
25502a6c PM |
194 | /* |
195 | * Note a context switch. This is a quiescent state for RCU-sched, | |
196 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 197 | * The caller must have disabled preemption. |
25502a6c PM |
198 | */ |
199 | void rcu_note_context_switch(int cpu) | |
200 | { | |
300df91c | 201 | trace_rcu_utilization("Start context switch"); |
25502a6c | 202 | rcu_sched_qs(cpu); |
cba6d0d6 | 203 | rcu_preempt_note_context_switch(cpu); |
300df91c | 204 | trace_rcu_utilization("End context switch"); |
25502a6c | 205 | } |
29ce8310 | 206 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 207 | |
90a4d2c0 | 208 | DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { |
29e37d81 | 209 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, |
23b5c8fa | 210 | .dynticks = ATOMIC_INIT(1), |
90a4d2c0 | 211 | }; |
64db4cff | 212 | |
878d7439 ED |
213 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
214 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
215 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 216 | |
878d7439 ED |
217 | module_param(blimit, long, 0444); |
218 | module_param(qhimark, long, 0444); | |
219 | module_param(qlowmark, long, 0444); | |
3d76c082 | 220 | |
d40011f6 PM |
221 | static ulong jiffies_till_first_fqs = RCU_JIFFIES_TILL_FORCE_QS; |
222 | static ulong jiffies_till_next_fqs = RCU_JIFFIES_TILL_FORCE_QS; | |
223 | ||
224 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
225 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
226 | ||
4cdfc175 PM |
227 | static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *)); |
228 | static void force_quiescent_state(struct rcu_state *rsp); | |
a157229c | 229 | static int rcu_pending(int cpu); |
64db4cff PM |
230 | |
231 | /* | |
d6714c22 | 232 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
64db4cff | 233 | */ |
d6714c22 | 234 | long rcu_batches_completed_sched(void) |
64db4cff | 235 | { |
d6714c22 | 236 | return rcu_sched_state.completed; |
64db4cff | 237 | } |
d6714c22 | 238 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
239 | |
240 | /* | |
241 | * Return the number of RCU BH batches processed thus far for debug & stats. | |
242 | */ | |
243 | long rcu_batches_completed_bh(void) | |
244 | { | |
245 | return rcu_bh_state.completed; | |
246 | } | |
247 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
248 | ||
bf66f18e PM |
249 | /* |
250 | * Force a quiescent state for RCU BH. | |
251 | */ | |
252 | void rcu_bh_force_quiescent_state(void) | |
253 | { | |
4cdfc175 | 254 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
255 | } |
256 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
257 | ||
4a298656 PM |
258 | /* |
259 | * Record the number of times rcutorture tests have been initiated and | |
260 | * terminated. This information allows the debugfs tracing stats to be | |
261 | * correlated to the rcutorture messages, even when the rcutorture module | |
262 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
263 | * store this state in rcutorture itself. | |
264 | */ | |
265 | void rcutorture_record_test_transition(void) | |
266 | { | |
267 | rcutorture_testseq++; | |
268 | rcutorture_vernum = 0; | |
269 | } | |
270 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
271 | ||
272 | /* | |
273 | * Record the number of writer passes through the current rcutorture test. | |
274 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
275 | * messages. | |
276 | */ | |
277 | void rcutorture_record_progress(unsigned long vernum) | |
278 | { | |
279 | rcutorture_vernum++; | |
280 | } | |
281 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
282 | ||
bf66f18e PM |
283 | /* |
284 | * Force a quiescent state for RCU-sched. | |
285 | */ | |
286 | void rcu_sched_force_quiescent_state(void) | |
287 | { | |
4cdfc175 | 288 | force_quiescent_state(&rcu_sched_state); |
bf66f18e PM |
289 | } |
290 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
291 | ||
64db4cff PM |
292 | /* |
293 | * Does the CPU have callbacks ready to be invoked? | |
294 | */ | |
295 | static int | |
296 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
297 | { | |
3fbfbf7a PM |
298 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
299 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
300 | } |
301 | ||
302 | /* | |
dc35c893 PM |
303 | * Does the current CPU require a not-yet-started grace period? |
304 | * The caller must have disabled interrupts to prevent races with | |
305 | * normal callback registry. | |
64db4cff PM |
306 | */ |
307 | static int | |
308 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
309 | { | |
dc35c893 | 310 | int i; |
3fbfbf7a | 311 | |
dc35c893 PM |
312 | if (rcu_gp_in_progress(rsp)) |
313 | return 0; /* No, a grace period is already in progress. */ | |
dae6e64d | 314 | if (rcu_nocb_needs_gp(rsp)) |
34ed6246 | 315 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
316 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
317 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
318 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
319 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
320 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
321 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
322 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), | |
323 | rdp->nxtcompleted[i])) | |
324 | return 1; /* Yes, CBs for future grace period. */ | |
325 | return 0; /* No grace period needed. */ | |
64db4cff PM |
326 | } |
327 | ||
328 | /* | |
329 | * Return the root node of the specified rcu_state structure. | |
330 | */ | |
331 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
332 | { | |
333 | return &rsp->node[0]; | |
334 | } | |
335 | ||
9b2e4f18 | 336 | /* |
adf5091e | 337 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
338 | * |
339 | * If the new value of the ->dynticks_nesting counter now is zero, | |
340 | * we really have entered idle, and must do the appropriate accounting. | |
341 | * The caller must have disabled interrupts. | |
342 | */ | |
adf5091e FW |
343 | static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval, |
344 | bool user) | |
9b2e4f18 | 345 | { |
1bdc2b7d | 346 | trace_rcu_dyntick("Start", oldval, rdtp->dynticks_nesting); |
cb349ca9 | 347 | if (!user && !is_idle_task(current)) { |
0989cb46 PM |
348 | struct task_struct *idle = idle_task(smp_processor_id()); |
349 | ||
facc4e15 | 350 | trace_rcu_dyntick("Error on entry: not idle task", oldval, 0); |
bf1304e9 | 351 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
352 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
353 | current->pid, current->comm, | |
354 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 355 | } |
aea1b35e | 356 | rcu_prepare_for_idle(smp_processor_id()); |
9b2e4f18 PM |
357 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
358 | smp_mb__before_atomic_inc(); /* See above. */ | |
359 | atomic_inc(&rdtp->dynticks); | |
360 | smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */ | |
361 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
c44e2cdd PM |
362 | |
363 | /* | |
adf5091e | 364 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
365 | * in an RCU read-side critical section. |
366 | */ | |
367 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
368 | "Illegal idle entry in RCU read-side critical section."); | |
369 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
370 | "Illegal idle entry in RCU-bh read-side critical section."); | |
371 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
372 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 373 | } |
64db4cff | 374 | |
adf5091e FW |
375 | /* |
376 | * Enter an RCU extended quiescent state, which can be either the | |
377 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 378 | */ |
adf5091e | 379 | static void rcu_eqs_enter(bool user) |
64db4cff | 380 | { |
4145fa7f | 381 | long long oldval; |
64db4cff PM |
382 | struct rcu_dynticks *rdtp; |
383 | ||
64db4cff | 384 | rdtp = &__get_cpu_var(rcu_dynticks); |
4145fa7f | 385 | oldval = rdtp->dynticks_nesting; |
29e37d81 PM |
386 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
387 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) | |
388 | rdtp->dynticks_nesting = 0; | |
389 | else | |
390 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; | |
adf5091e | 391 | rcu_eqs_enter_common(rdtp, oldval, user); |
64db4cff | 392 | } |
adf5091e FW |
393 | |
394 | /** | |
395 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
396 | * | |
397 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
398 | * read-side critical sections can occur. (Though RCU read-side | |
399 | * critical sections can occur in irq handlers in idle, a possibility | |
400 | * handled by irq_enter() and irq_exit().) | |
401 | * | |
402 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
403 | * the possibility of usermode upcalls having messed up our count | |
404 | * of interrupt nesting level during the prior busy period. | |
405 | */ | |
406 | void rcu_idle_enter(void) | |
407 | { | |
c5d900bf FW |
408 | unsigned long flags; |
409 | ||
410 | local_irq_save(flags); | |
cb349ca9 | 411 | rcu_eqs_enter(false); |
c5d900bf | 412 | local_irq_restore(flags); |
adf5091e | 413 | } |
8a2ecf47 | 414 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 415 | |
2b1d5024 | 416 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
417 | /** |
418 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
419 | * | |
420 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
421 | * is permitted between this call and rcu_user_exit(). This way the | |
422 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
423 | * when the CPU runs in userspace. | |
424 | */ | |
425 | void rcu_user_enter(void) | |
426 | { | |
91d1aa43 | 427 | rcu_eqs_enter(1); |
adf5091e FW |
428 | } |
429 | ||
19dd1591 FW |
430 | /** |
431 | * rcu_user_enter_after_irq - inform RCU that we are going to resume userspace | |
432 | * after the current irq returns. | |
433 | * | |
434 | * This is similar to rcu_user_enter() but in the context of a non-nesting | |
435 | * irq. After this call, RCU enters into idle mode when the interrupt | |
436 | * returns. | |
437 | */ | |
438 | void rcu_user_enter_after_irq(void) | |
439 | { | |
440 | unsigned long flags; | |
441 | struct rcu_dynticks *rdtp; | |
442 | ||
443 | local_irq_save(flags); | |
444 | rdtp = &__get_cpu_var(rcu_dynticks); | |
445 | /* Ensure this irq is interrupting a non-idle RCU state. */ | |
446 | WARN_ON_ONCE(!(rdtp->dynticks_nesting & DYNTICK_TASK_MASK)); | |
447 | rdtp->dynticks_nesting = 1; | |
448 | local_irq_restore(flags); | |
449 | } | |
2b1d5024 | 450 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 451 | |
9b2e4f18 PM |
452 | /** |
453 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
454 | * | |
455 | * Exit from an interrupt handler, which might possibly result in entering | |
456 | * idle mode, in other words, leaving the mode in which read-side critical | |
457 | * sections can occur. | |
64db4cff | 458 | * |
9b2e4f18 PM |
459 | * This code assumes that the idle loop never does anything that might |
460 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
461 | * architecture violates this assumption, RCU will give you what you | |
462 | * deserve, good and hard. But very infrequently and irreproducibly. | |
463 | * | |
464 | * Use things like work queues to work around this limitation. | |
465 | * | |
466 | * You have been warned. | |
64db4cff | 467 | */ |
9b2e4f18 | 468 | void rcu_irq_exit(void) |
64db4cff PM |
469 | { |
470 | unsigned long flags; | |
4145fa7f | 471 | long long oldval; |
64db4cff PM |
472 | struct rcu_dynticks *rdtp; |
473 | ||
474 | local_irq_save(flags); | |
475 | rdtp = &__get_cpu_var(rcu_dynticks); | |
4145fa7f | 476 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
477 | rdtp->dynticks_nesting--; |
478 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 FW |
479 | if (rdtp->dynticks_nesting) |
480 | trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting); | |
481 | else | |
cb349ca9 | 482 | rcu_eqs_enter_common(rdtp, oldval, true); |
9b2e4f18 PM |
483 | local_irq_restore(flags); |
484 | } | |
485 | ||
486 | /* | |
adf5091e | 487 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
488 | * |
489 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
490 | * we really have exited idle, and must do the appropriate accounting. | |
491 | * The caller must have disabled interrupts. | |
492 | */ | |
adf5091e FW |
493 | static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval, |
494 | int user) | |
9b2e4f18 | 495 | { |
23b5c8fa PM |
496 | smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */ |
497 | atomic_inc(&rdtp->dynticks); | |
498 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
499 | smp_mb__after_atomic_inc(); /* See above. */ | |
500 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
7cb92499 | 501 | rcu_cleanup_after_idle(smp_processor_id()); |
4145fa7f | 502 | trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting); |
cb349ca9 | 503 | if (!user && !is_idle_task(current)) { |
0989cb46 PM |
504 | struct task_struct *idle = idle_task(smp_processor_id()); |
505 | ||
4145fa7f PM |
506 | trace_rcu_dyntick("Error on exit: not idle task", |
507 | oldval, rdtp->dynticks_nesting); | |
bf1304e9 | 508 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
509 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
510 | current->pid, current->comm, | |
511 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
512 | } |
513 | } | |
514 | ||
adf5091e FW |
515 | /* |
516 | * Exit an RCU extended quiescent state, which can be either the | |
517 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 518 | */ |
adf5091e | 519 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 520 | { |
9b2e4f18 PM |
521 | struct rcu_dynticks *rdtp; |
522 | long long oldval; | |
523 | ||
9b2e4f18 PM |
524 | rdtp = &__get_cpu_var(rcu_dynticks); |
525 | oldval = rdtp->dynticks_nesting; | |
29e37d81 PM |
526 | WARN_ON_ONCE(oldval < 0); |
527 | if (oldval & DYNTICK_TASK_NEST_MASK) | |
528 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; | |
529 | else | |
530 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; | |
adf5091e | 531 | rcu_eqs_exit_common(rdtp, oldval, user); |
9b2e4f18 | 532 | } |
adf5091e FW |
533 | |
534 | /** | |
535 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
536 | * | |
537 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
538 | * read-side critical sections can occur. | |
539 | * | |
540 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
541 | * allow for the possibility of usermode upcalls messing up our count | |
542 | * of interrupt nesting level during the busy period that is just | |
543 | * now starting. | |
544 | */ | |
545 | void rcu_idle_exit(void) | |
546 | { | |
c5d900bf FW |
547 | unsigned long flags; |
548 | ||
549 | local_irq_save(flags); | |
cb349ca9 | 550 | rcu_eqs_exit(false); |
c5d900bf | 551 | local_irq_restore(flags); |
adf5091e | 552 | } |
8a2ecf47 | 553 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 554 | |
2b1d5024 | 555 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
556 | /** |
557 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
558 | * | |
559 | * Exit RCU idle mode while entering the kernel because it can | |
560 | * run a RCU read side critical section anytime. | |
561 | */ | |
562 | void rcu_user_exit(void) | |
563 | { | |
91d1aa43 | 564 | rcu_eqs_exit(1); |
adf5091e FW |
565 | } |
566 | ||
19dd1591 FW |
567 | /** |
568 | * rcu_user_exit_after_irq - inform RCU that we won't resume to userspace | |
569 | * idle mode after the current non-nesting irq returns. | |
570 | * | |
571 | * This is similar to rcu_user_exit() but in the context of an irq. | |
572 | * This is called when the irq has interrupted a userspace RCU idle mode | |
573 | * context. When the current non-nesting interrupt returns after this call, | |
574 | * the CPU won't restore the RCU idle mode. | |
575 | */ | |
576 | void rcu_user_exit_after_irq(void) | |
577 | { | |
578 | unsigned long flags; | |
579 | struct rcu_dynticks *rdtp; | |
580 | ||
581 | local_irq_save(flags); | |
582 | rdtp = &__get_cpu_var(rcu_dynticks); | |
583 | /* Ensure we are interrupting an RCU idle mode. */ | |
584 | WARN_ON_ONCE(rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK); | |
585 | rdtp->dynticks_nesting += DYNTICK_TASK_EXIT_IDLE; | |
586 | local_irq_restore(flags); | |
587 | } | |
2b1d5024 | 588 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 589 | |
9b2e4f18 PM |
590 | /** |
591 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
592 | * | |
593 | * Enter an interrupt handler, which might possibly result in exiting | |
594 | * idle mode, in other words, entering the mode in which read-side critical | |
595 | * sections can occur. | |
596 | * | |
597 | * Note that the Linux kernel is fully capable of entering an interrupt | |
598 | * handler that it never exits, for example when doing upcalls to | |
599 | * user mode! This code assumes that the idle loop never does upcalls to | |
600 | * user mode. If your architecture does do upcalls from the idle loop (or | |
601 | * does anything else that results in unbalanced calls to the irq_enter() | |
602 | * and irq_exit() functions), RCU will give you what you deserve, good | |
603 | * and hard. But very infrequently and irreproducibly. | |
604 | * | |
605 | * Use things like work queues to work around this limitation. | |
606 | * | |
607 | * You have been warned. | |
608 | */ | |
609 | void rcu_irq_enter(void) | |
610 | { | |
611 | unsigned long flags; | |
612 | struct rcu_dynticks *rdtp; | |
613 | long long oldval; | |
614 | ||
615 | local_irq_save(flags); | |
616 | rdtp = &__get_cpu_var(rcu_dynticks); | |
617 | oldval = rdtp->dynticks_nesting; | |
618 | rdtp->dynticks_nesting++; | |
619 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 FW |
620 | if (oldval) |
621 | trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting); | |
622 | else | |
cb349ca9 | 623 | rcu_eqs_exit_common(rdtp, oldval, true); |
64db4cff | 624 | local_irq_restore(flags); |
64db4cff PM |
625 | } |
626 | ||
627 | /** | |
628 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
629 | * | |
630 | * If the CPU was idle with dynamic ticks active, and there is no | |
631 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
632 | * RCU grace-period handling know that the CPU is active. | |
633 | */ | |
634 | void rcu_nmi_enter(void) | |
635 | { | |
636 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); | |
637 | ||
23b5c8fa PM |
638 | if (rdtp->dynticks_nmi_nesting == 0 && |
639 | (atomic_read(&rdtp->dynticks) & 0x1)) | |
64db4cff | 640 | return; |
23b5c8fa PM |
641 | rdtp->dynticks_nmi_nesting++; |
642 | smp_mb__before_atomic_inc(); /* Force delay from prior write. */ | |
643 | atomic_inc(&rdtp->dynticks); | |
644 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
645 | smp_mb__after_atomic_inc(); /* See above. */ | |
646 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
64db4cff PM |
647 | } |
648 | ||
649 | /** | |
650 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
651 | * | |
652 | * If the CPU was idle with dynamic ticks active, and there is no | |
653 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
654 | * RCU grace-period handling know that the CPU is no longer active. | |
655 | */ | |
656 | void rcu_nmi_exit(void) | |
657 | { | |
658 | struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks); | |
659 | ||
23b5c8fa PM |
660 | if (rdtp->dynticks_nmi_nesting == 0 || |
661 | --rdtp->dynticks_nmi_nesting != 0) | |
64db4cff | 662 | return; |
23b5c8fa PM |
663 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
664 | smp_mb__before_atomic_inc(); /* See above. */ | |
665 | atomic_inc(&rdtp->dynticks); | |
666 | smp_mb__after_atomic_inc(); /* Force delay to next write. */ | |
667 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
64db4cff PM |
668 | } |
669 | ||
670 | /** | |
9b2e4f18 | 671 | * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle |
64db4cff | 672 | * |
9b2e4f18 | 673 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 674 | * or NMI handler, return true. |
64db4cff | 675 | */ |
9b2e4f18 | 676 | int rcu_is_cpu_idle(void) |
64db4cff | 677 | { |
34240697 PM |
678 | int ret; |
679 | ||
680 | preempt_disable(); | |
681 | ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0; | |
682 | preempt_enable(); | |
683 | return ret; | |
64db4cff | 684 | } |
e6b80a3b | 685 | EXPORT_SYMBOL(rcu_is_cpu_idle); |
64db4cff | 686 | |
62fde6ed | 687 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
688 | |
689 | /* | |
690 | * Is the current CPU online? Disable preemption to avoid false positives | |
691 | * that could otherwise happen due to the current CPU number being sampled, | |
692 | * this task being preempted, its old CPU being taken offline, resuming | |
693 | * on some other CPU, then determining that its old CPU is now offline. | |
694 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
695 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
696 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
697 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
698 | * offline to continue to use RCU for one jiffy after marking itself | |
699 | * offline in the cpu_online_mask. This leniency is necessary given the | |
700 | * non-atomic nature of the online and offline processing, for example, | |
701 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
702 | * notifiers. | |
703 | * | |
704 | * This is also why RCU internally marks CPUs online during the | |
705 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
706 | * |
707 | * Disable checking if in an NMI handler because we cannot safely report | |
708 | * errors from NMI handlers anyway. | |
709 | */ | |
710 | bool rcu_lockdep_current_cpu_online(void) | |
711 | { | |
2036d94a PM |
712 | struct rcu_data *rdp; |
713 | struct rcu_node *rnp; | |
c0d6d01b PM |
714 | bool ret; |
715 | ||
716 | if (in_nmi()) | |
717 | return 1; | |
718 | preempt_disable(); | |
2036d94a PM |
719 | rdp = &__get_cpu_var(rcu_sched_data); |
720 | rnp = rdp->mynode; | |
721 | ret = (rdp->grpmask & rnp->qsmaskinit) || | |
c0d6d01b PM |
722 | !rcu_scheduler_fully_active; |
723 | preempt_enable(); | |
724 | return ret; | |
725 | } | |
726 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
727 | ||
62fde6ed | 728 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 729 | |
64db4cff | 730 | /** |
9b2e4f18 | 731 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 732 | * |
9b2e4f18 PM |
733 | * If the current CPU is idle or running at a first-level (not nested) |
734 | * interrupt from idle, return true. The caller must have at least | |
735 | * disabled preemption. | |
64db4cff | 736 | */ |
62e3cb14 | 737 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 738 | { |
9b2e4f18 | 739 | return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1; |
64db4cff PM |
740 | } |
741 | ||
64db4cff PM |
742 | /* |
743 | * Snapshot the specified CPU's dynticks counter so that we can later | |
744 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 745 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff PM |
746 | */ |
747 | static int dyntick_save_progress_counter(struct rcu_data *rdp) | |
748 | { | |
23b5c8fa | 749 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
f0e7c19d | 750 | return (rdp->dynticks_snap & 0x1) == 0; |
64db4cff PM |
751 | } |
752 | ||
753 | /* | |
754 | * Return true if the specified CPU has passed through a quiescent | |
755 | * state by virtue of being in or having passed through an dynticks | |
756 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 757 | * for this same CPU, or by virtue of having been offline. |
64db4cff PM |
758 | */ |
759 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp) | |
760 | { | |
7eb4f455 PM |
761 | unsigned int curr; |
762 | unsigned int snap; | |
64db4cff | 763 | |
7eb4f455 PM |
764 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
765 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
766 | |
767 | /* | |
768 | * If the CPU passed through or entered a dynticks idle phase with | |
769 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
770 | * already acknowledged the request to pass through a quiescent | |
771 | * state. Either way, that CPU cannot possibly be in an RCU | |
772 | * read-side critical section that started before the beginning | |
773 | * of the current RCU grace period. | |
774 | */ | |
7eb4f455 | 775 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
d4c08f2a | 776 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti"); |
64db4cff PM |
777 | rdp->dynticks_fqs++; |
778 | return 1; | |
779 | } | |
780 | ||
a82dcc76 PM |
781 | /* |
782 | * Check for the CPU being offline, but only if the grace period | |
783 | * is old enough. We don't need to worry about the CPU changing | |
784 | * state: If we see it offline even once, it has been through a | |
785 | * quiescent state. | |
786 | * | |
787 | * The reason for insisting that the grace period be at least | |
788 | * one jiffy old is that CPUs that are not quite online and that | |
789 | * have just gone offline can still execute RCU read-side critical | |
790 | * sections. | |
791 | */ | |
792 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
793 | return 0; /* Grace period is not old enough. */ | |
794 | barrier(); | |
795 | if (cpu_is_offline(rdp->cpu)) { | |
796 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl"); | |
797 | rdp->offline_fqs++; | |
798 | return 1; | |
799 | } | |
800 | return 0; | |
64db4cff PM |
801 | } |
802 | ||
64db4cff PM |
803 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
804 | { | |
805 | rsp->gp_start = jiffies; | |
6bfc09e2 | 806 | rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check(); |
64db4cff PM |
807 | } |
808 | ||
b637a328 PM |
809 | /* |
810 | * Dump stacks of all tasks running on stalled CPUs. This is a fallback | |
811 | * for architectures that do not implement trigger_all_cpu_backtrace(). | |
812 | * The NMI-triggered stack traces are more accurate because they are | |
813 | * printed by the target CPU. | |
814 | */ | |
815 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
816 | { | |
817 | int cpu; | |
818 | unsigned long flags; | |
819 | struct rcu_node *rnp; | |
820 | ||
821 | rcu_for_each_leaf_node(rsp, rnp) { | |
822 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
823 | if (rnp->qsmask != 0) { | |
824 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
825 | if (rnp->qsmask & (1UL << cpu)) | |
826 | dump_cpu_task(rnp->grplo + cpu); | |
827 | } | |
828 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
829 | } | |
830 | } | |
831 | ||
64db4cff PM |
832 | static void print_other_cpu_stall(struct rcu_state *rsp) |
833 | { | |
834 | int cpu; | |
835 | long delta; | |
836 | unsigned long flags; | |
285fe294 | 837 | int ndetected = 0; |
64db4cff | 838 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 839 | long totqlen = 0; |
64db4cff PM |
840 | |
841 | /* Only let one CPU complain about others per time interval. */ | |
842 | ||
1304afb2 | 843 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 844 | delta = jiffies - rsp->jiffies_stall; |
fc2219d4 | 845 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 846 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
847 | return; |
848 | } | |
6bfc09e2 | 849 | rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 850 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 851 | |
8cdd32a9 PM |
852 | /* |
853 | * OK, time to rat on our buddy... | |
854 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
855 | * RCU CPU stall warnings. | |
856 | */ | |
a858af28 | 857 | printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 858 | rsp->name); |
a858af28 | 859 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 860 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 861 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 862 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
863 | if (rnp->qsmask != 0) { |
864 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
865 | if (rnp->qsmask & (1UL << cpu)) { | |
866 | print_cpu_stall_info(rsp, | |
867 | rnp->grplo + cpu); | |
868 | ndetected++; | |
869 | } | |
870 | } | |
3acd9eb3 | 871 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 872 | } |
a858af28 PM |
873 | |
874 | /* | |
875 | * Now rat on any tasks that got kicked up to the root rcu_node | |
876 | * due to CPU offlining. | |
877 | */ | |
878 | rnp = rcu_get_root(rsp); | |
879 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
285fe294 | 880 | ndetected += rcu_print_task_stall(rnp); |
a858af28 PM |
881 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
882 | ||
883 | print_cpu_stall_info_end(); | |
53bb857c PM |
884 | for_each_possible_cpu(cpu) |
885 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
886 | pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n", | |
eee05882 | 887 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
53bb857c | 888 | rsp->gpnum, rsp->completed, totqlen); |
9bc8b558 PM |
889 | if (ndetected == 0) |
890 | printk(KERN_ERR "INFO: Stall ended before state dump start\n"); | |
891 | else if (!trigger_all_cpu_backtrace()) | |
b637a328 | 892 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 893 | |
4cdfc175 | 894 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
895 | |
896 | rcu_print_detail_task_stall(rsp); | |
897 | ||
4cdfc175 | 898 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
899 | } |
900 | ||
901 | static void print_cpu_stall(struct rcu_state *rsp) | |
902 | { | |
53bb857c | 903 | int cpu; |
64db4cff PM |
904 | unsigned long flags; |
905 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 906 | long totqlen = 0; |
64db4cff | 907 | |
8cdd32a9 PM |
908 | /* |
909 | * OK, time to rat on ourselves... | |
910 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
911 | * RCU CPU stall warnings. | |
912 | */ | |
a858af28 PM |
913 | printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name); |
914 | print_cpu_stall_info_begin(); | |
915 | print_cpu_stall_info(rsp, smp_processor_id()); | |
916 | print_cpu_stall_info_end(); | |
53bb857c PM |
917 | for_each_possible_cpu(cpu) |
918 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
919 | pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n", | |
920 | jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen); | |
4627e240 PM |
921 | if (!trigger_all_cpu_backtrace()) |
922 | dump_stack(); | |
c1dc0b9c | 923 | |
1304afb2 | 924 | raw_spin_lock_irqsave(&rnp->lock, flags); |
20133cfc | 925 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall)) |
13cfcca0 | 926 | rsp->jiffies_stall = jiffies + |
6bfc09e2 | 927 | 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 928 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 929 | |
64db4cff PM |
930 | set_need_resched(); /* kick ourselves to get things going. */ |
931 | } | |
932 | ||
933 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
934 | { | |
bad6e139 PM |
935 | unsigned long j; |
936 | unsigned long js; | |
64db4cff PM |
937 | struct rcu_node *rnp; |
938 | ||
742734ee | 939 | if (rcu_cpu_stall_suppress) |
c68de209 | 940 | return; |
bad6e139 PM |
941 | j = ACCESS_ONCE(jiffies); |
942 | js = ACCESS_ONCE(rsp->jiffies_stall); | |
64db4cff | 943 | rnp = rdp->mynode; |
c96ea7cf PM |
944 | if (rcu_gp_in_progress(rsp) && |
945 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) { | |
64db4cff PM |
946 | |
947 | /* We haven't checked in, so go dump stack. */ | |
948 | print_cpu_stall(rsp); | |
949 | ||
bad6e139 PM |
950 | } else if (rcu_gp_in_progress(rsp) && |
951 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 952 | |
bad6e139 | 953 | /* They had a few time units to dump stack, so complain. */ |
64db4cff PM |
954 | print_other_cpu_stall(rsp); |
955 | } | |
956 | } | |
957 | ||
53d84e00 PM |
958 | /** |
959 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
960 | * | |
961 | * Set the stall-warning timeout way off into the future, thus preventing | |
962 | * any RCU CPU stall-warning messages from appearing in the current set of | |
963 | * RCU grace periods. | |
964 | * | |
965 | * The caller must disable hard irqs. | |
966 | */ | |
967 | void rcu_cpu_stall_reset(void) | |
968 | { | |
6ce75a23 PM |
969 | struct rcu_state *rsp; |
970 | ||
971 | for_each_rcu_flavor(rsp) | |
972 | rsp->jiffies_stall = jiffies + ULONG_MAX / 2; | |
53d84e00 PM |
973 | } |
974 | ||
64db4cff PM |
975 | /* |
976 | * Update CPU-local rcu_data state to record the newly noticed grace period. | |
977 | * This is used both when we started the grace period and when we notice | |
9160306e PM |
978 | * that someone else started the grace period. The caller must hold the |
979 | * ->lock of the leaf rcu_node structure corresponding to the current CPU, | |
980 | * and must have irqs disabled. | |
64db4cff | 981 | */ |
9160306e PM |
982 | static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) |
983 | { | |
984 | if (rdp->gpnum != rnp->gpnum) { | |
121dfc4b PM |
985 | /* |
986 | * If the current grace period is waiting for this CPU, | |
987 | * set up to detect a quiescent state, otherwise don't | |
988 | * go looking for one. | |
989 | */ | |
9160306e | 990 | rdp->gpnum = rnp->gpnum; |
d4c08f2a | 991 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart"); |
d7d6a11e PM |
992 | rdp->passed_quiesce = 0; |
993 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); | |
a858af28 | 994 | zero_cpu_stall_ticks(rdp); |
9160306e PM |
995 | } |
996 | } | |
997 | ||
64db4cff PM |
998 | static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp) |
999 | { | |
9160306e PM |
1000 | unsigned long flags; |
1001 | struct rcu_node *rnp; | |
1002 | ||
1003 | local_irq_save(flags); | |
1004 | rnp = rdp->mynode; | |
1005 | if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */ | |
1304afb2 | 1006 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
9160306e PM |
1007 | local_irq_restore(flags); |
1008 | return; | |
1009 | } | |
1010 | __note_new_gpnum(rsp, rnp, rdp); | |
1304afb2 | 1011 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1012 | } |
1013 | ||
1014 | /* | |
1015 | * Did someone else start a new RCU grace period start since we last | |
1016 | * checked? Update local state appropriately if so. Must be called | |
1017 | * on the CPU corresponding to rdp. | |
1018 | */ | |
1019 | static int | |
1020 | check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp) | |
1021 | { | |
1022 | unsigned long flags; | |
1023 | int ret = 0; | |
1024 | ||
1025 | local_irq_save(flags); | |
1026 | if (rdp->gpnum != rsp->gpnum) { | |
1027 | note_new_gpnum(rsp, rdp); | |
1028 | ret = 1; | |
1029 | } | |
1030 | local_irq_restore(flags); | |
1031 | return ret; | |
1032 | } | |
1033 | ||
3f5d3ea6 PM |
1034 | /* |
1035 | * Initialize the specified rcu_data structure's callback list to empty. | |
1036 | */ | |
1037 | static void init_callback_list(struct rcu_data *rdp) | |
1038 | { | |
1039 | int i; | |
1040 | ||
34ed6246 PM |
1041 | if (init_nocb_callback_list(rdp)) |
1042 | return; | |
3f5d3ea6 PM |
1043 | rdp->nxtlist = NULL; |
1044 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1045 | rdp->nxttail[i] = &rdp->nxtlist; | |
1046 | } | |
1047 | ||
dc35c893 PM |
1048 | /* |
1049 | * Determine the value that ->completed will have at the end of the | |
1050 | * next subsequent grace period. This is used to tag callbacks so that | |
1051 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1052 | * been dyntick-idle for an extended period with callbacks under the | |
1053 | * influence of RCU_FAST_NO_HZ. | |
1054 | * | |
1055 | * The caller must hold rnp->lock with interrupts disabled. | |
1056 | */ | |
1057 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1058 | struct rcu_node *rnp) | |
1059 | { | |
1060 | /* | |
1061 | * If RCU is idle, we just wait for the next grace period. | |
1062 | * But we can only be sure that RCU is idle if we are looking | |
1063 | * at the root rcu_node structure -- otherwise, a new grace | |
1064 | * period might have started, but just not yet gotten around | |
1065 | * to initializing the current non-root rcu_node structure. | |
1066 | */ | |
1067 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1068 | return rnp->completed + 1; | |
1069 | ||
1070 | /* | |
1071 | * Otherwise, wait for a possible partial grace period and | |
1072 | * then the subsequent full grace period. | |
1073 | */ | |
1074 | return rnp->completed + 2; | |
1075 | } | |
1076 | ||
1077 | /* | |
1078 | * If there is room, assign a ->completed number to any callbacks on | |
1079 | * this CPU that have not already been assigned. Also accelerate any | |
1080 | * callbacks that were previously assigned a ->completed number that has | |
1081 | * since proven to be too conservative, which can happen if callbacks get | |
1082 | * assigned a ->completed number while RCU is idle, but with reference to | |
1083 | * a non-root rcu_node structure. This function is idempotent, so it does | |
1084 | * not hurt to call it repeatedly. | |
1085 | * | |
1086 | * The caller must hold rnp->lock with interrupts disabled. | |
1087 | */ | |
1088 | static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1089 | struct rcu_data *rdp) | |
1090 | { | |
1091 | unsigned long c; | |
1092 | int i; | |
1093 | ||
1094 | /* If the CPU has no callbacks, nothing to do. */ | |
1095 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1096 | return; | |
1097 | ||
1098 | /* | |
1099 | * Starting from the sublist containing the callbacks most | |
1100 | * recently assigned a ->completed number and working down, find the | |
1101 | * first sublist that is not assignable to an upcoming grace period. | |
1102 | * Such a sublist has something in it (first two tests) and has | |
1103 | * a ->completed number assigned that will complete sooner than | |
1104 | * the ->completed number for newly arrived callbacks (last test). | |
1105 | * | |
1106 | * The key point is that any later sublist can be assigned the | |
1107 | * same ->completed number as the newly arrived callbacks, which | |
1108 | * means that the callbacks in any of these later sublist can be | |
1109 | * grouped into a single sublist, whether or not they have already | |
1110 | * been assigned a ->completed number. | |
1111 | */ | |
1112 | c = rcu_cbs_completed(rsp, rnp); | |
1113 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1114 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1115 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1116 | break; | |
1117 | ||
1118 | /* | |
1119 | * If there are no sublist for unassigned callbacks, leave. | |
1120 | * At the same time, advance "i" one sublist, so that "i" will | |
1121 | * index into the sublist where all the remaining callbacks should | |
1122 | * be grouped into. | |
1123 | */ | |
1124 | if (++i >= RCU_NEXT_TAIL) | |
1125 | return; | |
1126 | ||
1127 | /* | |
1128 | * Assign all subsequent callbacks' ->completed number to the next | |
1129 | * full grace period and group them all in the sublist initially | |
1130 | * indexed by "i". | |
1131 | */ | |
1132 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1133 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1134 | rdp->nxtcompleted[i] = c; | |
1135 | } | |
6d4b418c PM |
1136 | |
1137 | /* Trace depending on how much we were able to accelerate. */ | |
1138 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
1139 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccWaitCB"); | |
1140 | else | |
1141 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "AccReadyCB"); | |
dc35c893 PM |
1142 | } |
1143 | ||
1144 | /* | |
1145 | * Move any callbacks whose grace period has completed to the | |
1146 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1147 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1148 | * sublist. This function is idempotent, so it does not hurt to | |
1149 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
1150 | * | |
1151 | * The caller must hold rnp->lock with interrupts disabled. | |
1152 | */ | |
1153 | static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1154 | struct rcu_data *rdp) | |
1155 | { | |
1156 | int i, j; | |
1157 | ||
1158 | /* If the CPU has no callbacks, nothing to do. */ | |
1159 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1160 | return; | |
1161 | ||
1162 | /* | |
1163 | * Find all callbacks whose ->completed numbers indicate that they | |
1164 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1165 | */ | |
1166 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1167 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1168 | break; | |
1169 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1170 | } | |
1171 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1172 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1173 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1174 | ||
1175 | /* Copy down callbacks to fill in empty sublists. */ | |
1176 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1177 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1178 | break; | |
1179 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1180 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1181 | } | |
1182 | ||
1183 | /* Classify any remaining callbacks. */ | |
1184 | rcu_accelerate_cbs(rsp, rnp, rdp); | |
1185 | } | |
1186 | ||
d09b62df PM |
1187 | /* |
1188 | * Advance this CPU's callbacks, but only if the current grace period | |
1189 | * has ended. This may be called only from the CPU to whom the rdp | |
1190 | * belongs. In addition, the corresponding leaf rcu_node structure's | |
1191 | * ->lock must be held by the caller, with irqs disabled. | |
1192 | */ | |
1193 | static void | |
1194 | __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) | |
1195 | { | |
1196 | /* Did another grace period end? */ | |
dc35c893 | 1197 | if (rdp->completed == rnp->completed) { |
d09b62df | 1198 | |
dc35c893 PM |
1199 | /* No, so just accelerate recent callbacks. */ |
1200 | rcu_accelerate_cbs(rsp, rnp, rdp); | |
d09b62df | 1201 | |
dc35c893 PM |
1202 | } else { |
1203 | ||
1204 | /* Advance callbacks. */ | |
1205 | rcu_advance_cbs(rsp, rnp, rdp); | |
d09b62df PM |
1206 | |
1207 | /* Remember that we saw this grace-period completion. */ | |
1208 | rdp->completed = rnp->completed; | |
d4c08f2a | 1209 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend"); |
20377f32 | 1210 | |
5ff8e6f0 FW |
1211 | /* |
1212 | * If we were in an extended quiescent state, we may have | |
121dfc4b | 1213 | * missed some grace periods that others CPUs handled on |
5ff8e6f0 | 1214 | * our behalf. Catch up with this state to avoid noting |
121dfc4b PM |
1215 | * spurious new grace periods. If another grace period |
1216 | * has started, then rnp->gpnum will have advanced, so | |
d7d6a11e PM |
1217 | * we will detect this later on. Of course, any quiescent |
1218 | * states we found for the old GP are now invalid. | |
5ff8e6f0 | 1219 | */ |
d7d6a11e | 1220 | if (ULONG_CMP_LT(rdp->gpnum, rdp->completed)) { |
5ff8e6f0 | 1221 | rdp->gpnum = rdp->completed; |
d7d6a11e PM |
1222 | rdp->passed_quiesce = 0; |
1223 | } | |
5ff8e6f0 | 1224 | |
20377f32 | 1225 | /* |
121dfc4b PM |
1226 | * If RCU does not need a quiescent state from this CPU, |
1227 | * then make sure that this CPU doesn't go looking for one. | |
20377f32 | 1228 | */ |
121dfc4b | 1229 | if ((rnp->qsmask & rdp->grpmask) == 0) |
20377f32 | 1230 | rdp->qs_pending = 0; |
d09b62df PM |
1231 | } |
1232 | } | |
1233 | ||
1234 | /* | |
1235 | * Advance this CPU's callbacks, but only if the current grace period | |
1236 | * has ended. This may be called only from the CPU to whom the rdp | |
1237 | * belongs. | |
1238 | */ | |
1239 | static void | |
1240 | rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp) | |
1241 | { | |
1242 | unsigned long flags; | |
1243 | struct rcu_node *rnp; | |
1244 | ||
1245 | local_irq_save(flags); | |
1246 | rnp = rdp->mynode; | |
1247 | if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */ | |
1304afb2 | 1248 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
d09b62df PM |
1249 | local_irq_restore(flags); |
1250 | return; | |
1251 | } | |
1252 | __rcu_process_gp_end(rsp, rnp, rdp); | |
1304afb2 | 1253 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
d09b62df PM |
1254 | } |
1255 | ||
1256 | /* | |
1257 | * Do per-CPU grace-period initialization for running CPU. The caller | |
1258 | * must hold the lock of the leaf rcu_node structure corresponding to | |
1259 | * this CPU. | |
1260 | */ | |
1261 | static void | |
1262 | rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) | |
1263 | { | |
1264 | /* Prior grace period ended, so advance callbacks for current CPU. */ | |
1265 | __rcu_process_gp_end(rsp, rnp, rdp); | |
1266 | ||
9160306e PM |
1267 | /* Set state so that this CPU will detect the next quiescent state. */ |
1268 | __note_new_gpnum(rsp, rnp, rdp); | |
d09b62df PM |
1269 | } |
1270 | ||
b3dbec76 | 1271 | /* |
7fdefc10 | 1272 | * Initialize a new grace period. |
b3dbec76 | 1273 | */ |
7fdefc10 | 1274 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 PM |
1275 | { |
1276 | struct rcu_data *rdp; | |
7fdefc10 | 1277 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1278 | |
7fdefc10 | 1279 | raw_spin_lock_irq(&rnp->lock); |
4cdfc175 | 1280 | rsp->gp_flags = 0; /* Clear all flags: New grace period. */ |
b3dbec76 | 1281 | |
7fdefc10 PM |
1282 | if (rcu_gp_in_progress(rsp)) { |
1283 | /* Grace period already in progress, don't start another. */ | |
1284 | raw_spin_unlock_irq(&rnp->lock); | |
1285 | return 0; | |
1286 | } | |
1287 | ||
7fdefc10 PM |
1288 | /* Advance to a new grace period and initialize state. */ |
1289 | rsp->gpnum++; | |
1290 | trace_rcu_grace_period(rsp->name, rsp->gpnum, "start"); | |
7fdefc10 PM |
1291 | record_gp_stall_check_time(rsp); |
1292 | raw_spin_unlock_irq(&rnp->lock); | |
1293 | ||
1294 | /* Exclude any concurrent CPU-hotplug operations. */ | |
a4fbe35a | 1295 | mutex_lock(&rsp->onoff_mutex); |
7fdefc10 PM |
1296 | |
1297 | /* | |
1298 | * Set the quiescent-state-needed bits in all the rcu_node | |
1299 | * structures for all currently online CPUs in breadth-first order, | |
1300 | * starting from the root rcu_node structure, relying on the layout | |
1301 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1302 | * will access only the leaves of the hierarchy, thus seeing that no | |
1303 | * grace period is in progress, at least until the corresponding | |
1304 | * leaf node has been initialized. In addition, we have excluded | |
1305 | * CPU-hotplug operations. | |
1306 | * | |
1307 | * The grace period cannot complete until the initialization | |
1308 | * process finishes, because this kthread handles both. | |
1309 | */ | |
1310 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
b3dbec76 | 1311 | raw_spin_lock_irq(&rnp->lock); |
b3dbec76 | 1312 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1313 | rcu_preempt_check_blocked_tasks(rnp); |
1314 | rnp->qsmask = rnp->qsmaskinit; | |
1315 | rnp->gpnum = rsp->gpnum; | |
25d30cf4 | 1316 | WARN_ON_ONCE(rnp->completed != rsp->completed); |
7fdefc10 PM |
1317 | rnp->completed = rsp->completed; |
1318 | if (rnp == rdp->mynode) | |
1319 | rcu_start_gp_per_cpu(rsp, rnp, rdp); | |
1320 | rcu_preempt_boost_start_gp(rnp); | |
1321 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1322 | rnp->level, rnp->grplo, | |
1323 | rnp->grphi, rnp->qsmask); | |
1324 | raw_spin_unlock_irq(&rnp->lock); | |
661a85dc PM |
1325 | #ifdef CONFIG_PROVE_RCU_DELAY |
1326 | if ((random32() % (rcu_num_nodes * 8)) == 0) | |
1327 | schedule_timeout_uninterruptible(2); | |
1328 | #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */ | |
7fdefc10 PM |
1329 | cond_resched(); |
1330 | } | |
b3dbec76 | 1331 | |
a4fbe35a | 1332 | mutex_unlock(&rsp->onoff_mutex); |
7fdefc10 PM |
1333 | return 1; |
1334 | } | |
b3dbec76 | 1335 | |
4cdfc175 PM |
1336 | /* |
1337 | * Do one round of quiescent-state forcing. | |
1338 | */ | |
1339 | int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) | |
1340 | { | |
1341 | int fqs_state = fqs_state_in; | |
1342 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1343 | ||
1344 | rsp->n_force_qs++; | |
1345 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1346 | /* Collect dyntick-idle snapshots. */ | |
1347 | force_qs_rnp(rsp, dyntick_save_progress_counter); | |
1348 | fqs_state = RCU_FORCE_QS; | |
1349 | } else { | |
1350 | /* Handle dyntick-idle and offline CPUs. */ | |
1351 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs); | |
1352 | } | |
1353 | /* Clear flag to prevent immediate re-entry. */ | |
1354 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1355 | raw_spin_lock_irq(&rnp->lock); | |
1356 | rsp->gp_flags &= ~RCU_GP_FLAG_FQS; | |
1357 | raw_spin_unlock_irq(&rnp->lock); | |
1358 | } | |
1359 | return fqs_state; | |
1360 | } | |
1361 | ||
7fdefc10 PM |
1362 | /* |
1363 | * Clean up after the old grace period. | |
1364 | */ | |
4cdfc175 | 1365 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1366 | { |
1367 | unsigned long gp_duration; | |
dae6e64d | 1368 | int nocb = 0; |
7fdefc10 PM |
1369 | struct rcu_data *rdp; |
1370 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1371 | |
7fdefc10 PM |
1372 | raw_spin_lock_irq(&rnp->lock); |
1373 | gp_duration = jiffies - rsp->gp_start; | |
1374 | if (gp_duration > rsp->gp_max) | |
1375 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1376 | |
7fdefc10 PM |
1377 | /* |
1378 | * We know the grace period is complete, but to everyone else | |
1379 | * it appears to still be ongoing. But it is also the case | |
1380 | * that to everyone else it looks like there is nothing that | |
1381 | * they can do to advance the grace period. It is therefore | |
1382 | * safe for us to drop the lock in order to mark the grace | |
1383 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1384 | */ |
5d4b8659 | 1385 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1386 | |
5d4b8659 PM |
1387 | /* |
1388 | * Propagate new ->completed value to rcu_node structures so | |
1389 | * that other CPUs don't have to wait until the start of the next | |
1390 | * grace period to process their callbacks. This also avoids | |
1391 | * some nasty RCU grace-period initialization races by forcing | |
1392 | * the end of the current grace period to be completely recorded in | |
1393 | * all of the rcu_node structures before the beginning of the next | |
1394 | * grace period is recorded in any of the rcu_node structures. | |
1395 | */ | |
1396 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1397 | raw_spin_lock_irq(&rnp->lock); |
5d4b8659 | 1398 | rnp->completed = rsp->gpnum; |
b11cc576 PM |
1399 | rdp = this_cpu_ptr(rsp->rda); |
1400 | if (rnp == rdp->mynode) | |
1401 | __rcu_process_gp_end(rsp, rnp, rdp); | |
dae6e64d | 1402 | nocb += rcu_nocb_gp_cleanup(rsp, rnp); |
5d4b8659 PM |
1403 | raw_spin_unlock_irq(&rnp->lock); |
1404 | cond_resched(); | |
7fdefc10 | 1405 | } |
5d4b8659 PM |
1406 | rnp = rcu_get_root(rsp); |
1407 | raw_spin_lock_irq(&rnp->lock); | |
dae6e64d | 1408 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 PM |
1409 | |
1410 | rsp->completed = rsp->gpnum; /* Declare grace period done. */ | |
1411 | trace_rcu_grace_period(rsp->name, rsp->completed, "end"); | |
1412 | rsp->fqs_state = RCU_GP_IDLE; | |
5d4b8659 | 1413 | rdp = this_cpu_ptr(rsp->rda); |
b11cc576 | 1414 | rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */ |
7fdefc10 PM |
1415 | if (cpu_needs_another_gp(rsp, rdp)) |
1416 | rsp->gp_flags = 1; | |
1417 | raw_spin_unlock_irq(&rnp->lock); | |
7fdefc10 PM |
1418 | } |
1419 | ||
1420 | /* | |
1421 | * Body of kthread that handles grace periods. | |
1422 | */ | |
1423 | static int __noreturn rcu_gp_kthread(void *arg) | |
1424 | { | |
4cdfc175 | 1425 | int fqs_state; |
d40011f6 | 1426 | unsigned long j; |
4cdfc175 | 1427 | int ret; |
7fdefc10 PM |
1428 | struct rcu_state *rsp = arg; |
1429 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1430 | ||
1431 | for (;;) { | |
1432 | ||
1433 | /* Handle grace-period start. */ | |
1434 | for (;;) { | |
4cdfc175 PM |
1435 | wait_event_interruptible(rsp->gp_wq, |
1436 | rsp->gp_flags & | |
1437 | RCU_GP_FLAG_INIT); | |
1438 | if ((rsp->gp_flags & RCU_GP_FLAG_INIT) && | |
1439 | rcu_gp_init(rsp)) | |
7fdefc10 PM |
1440 | break; |
1441 | cond_resched(); | |
1442 | flush_signals(current); | |
1443 | } | |
cabc49c1 | 1444 | |
4cdfc175 PM |
1445 | /* Handle quiescent-state forcing. */ |
1446 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
1447 | j = jiffies_till_first_fqs; |
1448 | if (j > HZ) { | |
1449 | j = HZ; | |
1450 | jiffies_till_first_fqs = HZ; | |
1451 | } | |
cabc49c1 | 1452 | for (;;) { |
d40011f6 | 1453 | rsp->jiffies_force_qs = jiffies + j; |
4cdfc175 PM |
1454 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
1455 | (rsp->gp_flags & RCU_GP_FLAG_FQS) || | |
1456 | (!ACCESS_ONCE(rnp->qsmask) && | |
1457 | !rcu_preempt_blocked_readers_cgp(rnp)), | |
d40011f6 | 1458 | j); |
4cdfc175 | 1459 | /* If grace period done, leave loop. */ |
cabc49c1 | 1460 | if (!ACCESS_ONCE(rnp->qsmask) && |
4cdfc175 | 1461 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 1462 | break; |
4cdfc175 PM |
1463 | /* If time for quiescent-state forcing, do it. */ |
1464 | if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) { | |
1465 | fqs_state = rcu_gp_fqs(rsp, fqs_state); | |
1466 | cond_resched(); | |
1467 | } else { | |
1468 | /* Deal with stray signal. */ | |
1469 | cond_resched(); | |
1470 | flush_signals(current); | |
1471 | } | |
d40011f6 PM |
1472 | j = jiffies_till_next_fqs; |
1473 | if (j > HZ) { | |
1474 | j = HZ; | |
1475 | jiffies_till_next_fqs = HZ; | |
1476 | } else if (j < 1) { | |
1477 | j = 1; | |
1478 | jiffies_till_next_fqs = 1; | |
1479 | } | |
cabc49c1 | 1480 | } |
4cdfc175 PM |
1481 | |
1482 | /* Handle grace-period end. */ | |
1483 | rcu_gp_cleanup(rsp); | |
b3dbec76 | 1484 | } |
b3dbec76 PM |
1485 | } |
1486 | ||
64db4cff PM |
1487 | /* |
1488 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
1489 | * in preparation for detecting the next grace period. The caller must hold | |
1490 | * the root node's ->lock, which is released before return. Hard irqs must | |
1491 | * be disabled. | |
e5601400 PM |
1492 | * |
1493 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
1494 | * invoke this function. This can happen when the dying CPU reports its | |
1495 | * quiescent state. | |
64db4cff PM |
1496 | */ |
1497 | static void | |
1498 | rcu_start_gp(struct rcu_state *rsp, unsigned long flags) | |
1499 | __releases(rcu_get_root(rsp)->lock) | |
1500 | { | |
394f99a9 | 1501 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); |
64db4cff | 1502 | struct rcu_node *rnp = rcu_get_root(rsp); |
64db4cff | 1503 | |
b11cc576 PM |
1504 | /* |
1505 | * If there is no grace period in progress right now, any | |
1506 | * callbacks we have up to this point will be satisfied by the | |
1507 | * next grace period. Also, advancing the callbacks reduces the | |
1508 | * probability of false positives from cpu_needs_another_gp() | |
1509 | * resulting in pointless grace periods. So, advance callbacks! | |
1510 | */ | |
1511 | rcu_advance_cbs(rsp, rnp, rdp); | |
1512 | ||
b3dbec76 | 1513 | if (!rsp->gp_kthread || |
afe24b12 PM |
1514 | !cpu_needs_another_gp(rsp, rdp)) { |
1515 | /* | |
b3dbec76 | 1516 | * Either we have not yet spawned the grace-period |
62da1921 PM |
1517 | * task, this CPU does not need another grace period, |
1518 | * or a grace period is already in progress. | |
b3dbec76 | 1519 | * Either way, don't start a new grace period. |
afe24b12 PM |
1520 | */ |
1521 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1522 | return; | |
1523 | } | |
b32e9eb6 | 1524 | |
4cdfc175 | 1525 | rsp->gp_flags = RCU_GP_FLAG_INIT; |
62da1921 PM |
1526 | raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */ |
1527 | ||
1528 | /* Ensure that CPU is aware of completion of last grace period. */ | |
1529 | rcu_process_gp_end(rsp, rdp); | |
1530 | local_irq_restore(flags); | |
1531 | ||
1532 | /* Wake up rcu_gp_kthread() to start the grace period. */ | |
b3dbec76 | 1533 | wake_up(&rsp->gp_wq); |
64db4cff PM |
1534 | } |
1535 | ||
f41d911f | 1536 | /* |
d3f6bad3 PM |
1537 | * Report a full set of quiescent states to the specified rcu_state |
1538 | * data structure. This involves cleaning up after the prior grace | |
1539 | * period and letting rcu_start_gp() start up the next grace period | |
1540 | * if one is needed. Note that the caller must hold rnp->lock, as | |
1541 | * required by rcu_start_gp(), which will release it. | |
f41d911f | 1542 | */ |
d3f6bad3 | 1543 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 1544 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 1545 | { |
fc2219d4 | 1546 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cabc49c1 PM |
1547 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
1548 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ | |
f41d911f PM |
1549 | } |
1550 | ||
64db4cff | 1551 | /* |
d3f6bad3 PM |
1552 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
1553 | * Allows quiescent states for a group of CPUs to be reported at one go | |
1554 | * to the specified rcu_node structure, though all the CPUs in the group | |
1555 | * must be represented by the same rcu_node structure (which need not be | |
1556 | * a leaf rcu_node structure, though it often will be). That structure's | |
1557 | * lock must be held upon entry, and it is released before return. | |
64db4cff PM |
1558 | */ |
1559 | static void | |
d3f6bad3 PM |
1560 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
1561 | struct rcu_node *rnp, unsigned long flags) | |
64db4cff PM |
1562 | __releases(rnp->lock) |
1563 | { | |
28ecd580 PM |
1564 | struct rcu_node *rnp_c; |
1565 | ||
64db4cff PM |
1566 | /* Walk up the rcu_node hierarchy. */ |
1567 | for (;;) { | |
1568 | if (!(rnp->qsmask & mask)) { | |
1569 | ||
1570 | /* Our bit has already been cleared, so done. */ | |
1304afb2 | 1571 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1572 | return; |
1573 | } | |
1574 | rnp->qsmask &= ~mask; | |
d4c08f2a PM |
1575 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
1576 | mask, rnp->qsmask, rnp->level, | |
1577 | rnp->grplo, rnp->grphi, | |
1578 | !!rnp->gp_tasks); | |
27f4d280 | 1579 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
1580 | |
1581 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 1582 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1583 | return; |
1584 | } | |
1585 | mask = rnp->grpmask; | |
1586 | if (rnp->parent == NULL) { | |
1587 | ||
1588 | /* No more levels. Exit loop holding root lock. */ | |
1589 | ||
1590 | break; | |
1591 | } | |
1304afb2 | 1592 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 1593 | rnp_c = rnp; |
64db4cff | 1594 | rnp = rnp->parent; |
1304afb2 | 1595 | raw_spin_lock_irqsave(&rnp->lock, flags); |
28ecd580 | 1596 | WARN_ON_ONCE(rnp_c->qsmask); |
64db4cff PM |
1597 | } |
1598 | ||
1599 | /* | |
1600 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 1601 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 1602 | * to clean up and start the next grace period if one is needed. |
64db4cff | 1603 | */ |
d3f6bad3 | 1604 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
1605 | } |
1606 | ||
1607 | /* | |
d3f6bad3 PM |
1608 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
1609 | * structure. This must be either called from the specified CPU, or | |
1610 | * called when the specified CPU is known to be offline (and when it is | |
1611 | * also known that no other CPU is concurrently trying to help the offline | |
1612 | * CPU). The lastcomp argument is used to make sure we are still in the | |
1613 | * grace period of interest. We don't want to end the current grace period | |
1614 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
1615 | */ |
1616 | static void | |
d7d6a11e | 1617 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1618 | { |
1619 | unsigned long flags; | |
1620 | unsigned long mask; | |
1621 | struct rcu_node *rnp; | |
1622 | ||
1623 | rnp = rdp->mynode; | |
1304afb2 | 1624 | raw_spin_lock_irqsave(&rnp->lock, flags); |
d7d6a11e PM |
1625 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum || |
1626 | rnp->completed == rnp->gpnum) { | |
64db4cff PM |
1627 | |
1628 | /* | |
e4cc1f22 PM |
1629 | * The grace period in which this quiescent state was |
1630 | * recorded has ended, so don't report it upwards. | |
1631 | * We will instead need a new quiescent state that lies | |
1632 | * within the current grace period. | |
64db4cff | 1633 | */ |
e4cc1f22 | 1634 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
1304afb2 | 1635 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1636 | return; |
1637 | } | |
1638 | mask = rdp->grpmask; | |
1639 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 1640 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1641 | } else { |
1642 | rdp->qs_pending = 0; | |
1643 | ||
1644 | /* | |
1645 | * This GP can't end until cpu checks in, so all of our | |
1646 | * callbacks can be processed during the next GP. | |
1647 | */ | |
dc35c893 | 1648 | rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 1649 | |
d3f6bad3 | 1650 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
64db4cff PM |
1651 | } |
1652 | } | |
1653 | ||
1654 | /* | |
1655 | * Check to see if there is a new grace period of which this CPU | |
1656 | * is not yet aware, and if so, set up local rcu_data state for it. | |
1657 | * Otherwise, see if this CPU has just passed through its first | |
1658 | * quiescent state for this grace period, and record that fact if so. | |
1659 | */ | |
1660 | static void | |
1661 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
1662 | { | |
1663 | /* If there is now a new grace period, record and return. */ | |
1664 | if (check_for_new_grace_period(rsp, rdp)) | |
1665 | return; | |
1666 | ||
1667 | /* | |
1668 | * Does this CPU still need to do its part for current grace period? | |
1669 | * If no, return and let the other CPUs do their part as well. | |
1670 | */ | |
1671 | if (!rdp->qs_pending) | |
1672 | return; | |
1673 | ||
1674 | /* | |
1675 | * Was there a quiescent state since the beginning of the grace | |
1676 | * period? If no, then exit and wait for the next call. | |
1677 | */ | |
e4cc1f22 | 1678 | if (!rdp->passed_quiesce) |
64db4cff PM |
1679 | return; |
1680 | ||
d3f6bad3 PM |
1681 | /* |
1682 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
1683 | * judge of that). | |
1684 | */ | |
d7d6a11e | 1685 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
1686 | } |
1687 | ||
1688 | #ifdef CONFIG_HOTPLUG_CPU | |
1689 | ||
e74f4c45 | 1690 | /* |
b1420f1c PM |
1691 | * Send the specified CPU's RCU callbacks to the orphanage. The |
1692 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 1693 | * ->orphan_lock. |
e74f4c45 | 1694 | */ |
b1420f1c PM |
1695 | static void |
1696 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
1697 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 1698 | { |
3fbfbf7a PM |
1699 | /* No-CBs CPUs do not have orphanable callbacks. */ |
1700 | if (is_nocb_cpu(rdp->cpu)) | |
1701 | return; | |
1702 | ||
b1420f1c PM |
1703 | /* |
1704 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
1705 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
1706 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 1707 | */ |
a50c3af9 | 1708 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1709 | rsp->qlen_lazy += rdp->qlen_lazy; |
1710 | rsp->qlen += rdp->qlen; | |
1711 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 1712 | rdp->qlen_lazy = 0; |
1d1fb395 | 1713 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
1714 | } |
1715 | ||
1716 | /* | |
b1420f1c PM |
1717 | * Next, move those callbacks still needing a grace period to |
1718 | * the orphanage, where some other CPU will pick them up. | |
1719 | * Some of the callbacks might have gone partway through a grace | |
1720 | * period, but that is too bad. They get to start over because we | |
1721 | * cannot assume that grace periods are synchronized across CPUs. | |
1722 | * We don't bother updating the ->nxttail[] array yet, instead | |
1723 | * we just reset the whole thing later on. | |
a50c3af9 | 1724 | */ |
b1420f1c PM |
1725 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
1726 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1727 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
1728 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
1729 | } |
1730 | ||
1731 | /* | |
b1420f1c PM |
1732 | * Then move the ready-to-invoke callbacks to the orphanage, |
1733 | * where some other CPU will pick them up. These will not be | |
1734 | * required to pass though another grace period: They are done. | |
a50c3af9 | 1735 | */ |
e5601400 | 1736 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1737 | *rsp->orphan_donetail = rdp->nxtlist; |
1738 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 1739 | } |
e74f4c45 | 1740 | |
b1420f1c | 1741 | /* Finally, initialize the rcu_data structure's list to empty. */ |
3f5d3ea6 | 1742 | init_callback_list(rdp); |
b1420f1c PM |
1743 | } |
1744 | ||
1745 | /* | |
1746 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 1747 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c PM |
1748 | */ |
1749 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp) | |
1750 | { | |
1751 | int i; | |
1752 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); | |
1753 | ||
3fbfbf7a PM |
1754 | /* No-CBs CPUs are handled specially. */ |
1755 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp)) | |
1756 | return; | |
1757 | ||
b1420f1c PM |
1758 | /* Do the accounting first. */ |
1759 | rdp->qlen_lazy += rsp->qlen_lazy; | |
1760 | rdp->qlen += rsp->qlen; | |
1761 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
1762 | if (rsp->qlen_lazy != rsp->qlen) |
1763 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
1764 | rsp->qlen_lazy = 0; |
1765 | rsp->qlen = 0; | |
1766 | ||
1767 | /* | |
1768 | * We do not need a memory barrier here because the only way we | |
1769 | * can get here if there is an rcu_barrier() in flight is if | |
1770 | * we are the task doing the rcu_barrier(). | |
1771 | */ | |
1772 | ||
1773 | /* First adopt the ready-to-invoke callbacks. */ | |
1774 | if (rsp->orphan_donelist != NULL) { | |
1775 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1776 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
1777 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
1778 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
1779 | rdp->nxttail[i] = rsp->orphan_donetail; | |
1780 | rsp->orphan_donelist = NULL; | |
1781 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
1782 | } | |
1783 | ||
1784 | /* And then adopt the callbacks that still need a grace period. */ | |
1785 | if (rsp->orphan_nxtlist != NULL) { | |
1786 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
1787 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
1788 | rsp->orphan_nxtlist = NULL; | |
1789 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
1790 | } | |
1791 | } | |
1792 | ||
1793 | /* | |
1794 | * Trace the fact that this CPU is going offline. | |
1795 | */ | |
1796 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
1797 | { | |
1798 | RCU_TRACE(unsigned long mask); | |
1799 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
1800 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
1801 | ||
1802 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
1803 | trace_rcu_grace_period(rsp->name, |
1804 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
1805 | "cpuofl"); | |
64db4cff PM |
1806 | } |
1807 | ||
1808 | /* | |
e5601400 | 1809 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
1810 | * this fact from process context. Do the remainder of the cleanup, |
1811 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
1812 | * adopting them. There can only be one CPU hotplug operation at a time, |
1813 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 1814 | */ |
e5601400 | 1815 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 1816 | { |
2036d94a PM |
1817 | unsigned long flags; |
1818 | unsigned long mask; | |
1819 | int need_report = 0; | |
e5601400 | 1820 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 1821 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 1822 | |
2036d94a | 1823 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 1824 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 1825 | |
b1420f1c | 1826 | /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */ |
2036d94a PM |
1827 | |
1828 | /* Exclude any attempts to start a new grace period. */ | |
a4fbe35a | 1829 | mutex_lock(&rsp->onoff_mutex); |
7b2e6011 | 1830 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
2036d94a | 1831 | |
b1420f1c PM |
1832 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
1833 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
1834 | rcu_adopt_orphan_cbs(rsp); | |
1835 | ||
2036d94a PM |
1836 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
1837 | mask = rdp->grpmask; /* rnp->grplo is constant. */ | |
1838 | do { | |
1839 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
1840 | rnp->qsmaskinit &= ~mask; | |
1841 | if (rnp->qsmaskinit != 0) { | |
1842 | if (rnp != rdp->mynode) | |
1843 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1844 | break; | |
1845 | } | |
1846 | if (rnp == rdp->mynode) | |
1847 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); | |
1848 | else | |
1849 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1850 | mask = rnp->grpmask; | |
1851 | rnp = rnp->parent; | |
1852 | } while (rnp != NULL); | |
1853 | ||
1854 | /* | |
1855 | * We still hold the leaf rcu_node structure lock here, and | |
1856 | * irqs are still disabled. The reason for this subterfuge is | |
7b2e6011 | 1857 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock |
2036d94a PM |
1858 | * held leads to deadlock. |
1859 | */ | |
7b2e6011 | 1860 | raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */ |
2036d94a PM |
1861 | rnp = rdp->mynode; |
1862 | if (need_report & RCU_OFL_TASKS_NORM_GP) | |
1863 | rcu_report_unblock_qs_rnp(rnp, flags); | |
1864 | else | |
1865 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1866 | if (need_report & RCU_OFL_TASKS_EXP_GP) | |
1867 | rcu_report_exp_rnp(rsp, rnp, true); | |
cf01537e PM |
1868 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
1869 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
1870 | cpu, rdp->qlen, rdp->nxtlist); | |
0d8ee37e PM |
1871 | init_callback_list(rdp); |
1872 | /* Disallow further callbacks on this CPU. */ | |
1873 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; | |
a4fbe35a | 1874 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
1875 | } |
1876 | ||
1877 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1878 | ||
e5601400 | 1879 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
1880 | { |
1881 | } | |
1882 | ||
e5601400 | 1883 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
1884 | { |
1885 | } | |
1886 | ||
1887 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
1888 | ||
1889 | /* | |
1890 | * Invoke any RCU callbacks that have made it to the end of their grace | |
1891 | * period. Thottle as specified by rdp->blimit. | |
1892 | */ | |
37c72e56 | 1893 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1894 | { |
1895 | unsigned long flags; | |
1896 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
1897 | long bl, count, count_lazy; |
1898 | int i; | |
64db4cff | 1899 | |
dc35c893 | 1900 | /* If no callbacks are ready, just return. */ |
29c00b4a | 1901 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 1902 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
1903 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
1904 | need_resched(), is_idle_task(current), | |
1905 | rcu_is_callbacks_kthread()); | |
64db4cff | 1906 | return; |
29c00b4a | 1907 | } |
64db4cff PM |
1908 | |
1909 | /* | |
1910 | * Extract the list of ready callbacks, disabling to prevent | |
1911 | * races with call_rcu() from interrupt handlers. | |
1912 | */ | |
1913 | local_irq_save(flags); | |
8146c4e2 | 1914 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 1915 | bl = rdp->blimit; |
486e2593 | 1916 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
1917 | list = rdp->nxtlist; |
1918 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
1919 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
1920 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
1921 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
1922 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
1923 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
1924 | local_irq_restore(flags); |
1925 | ||
1926 | /* Invoke callbacks. */ | |
486e2593 | 1927 | count = count_lazy = 0; |
64db4cff PM |
1928 | while (list) { |
1929 | next = list->next; | |
1930 | prefetch(next); | |
551d55a9 | 1931 | debug_rcu_head_unqueue(list); |
486e2593 PM |
1932 | if (__rcu_reclaim(rsp->name, list)) |
1933 | count_lazy++; | |
64db4cff | 1934 | list = next; |
dff1672d PM |
1935 | /* Stop only if limit reached and CPU has something to do. */ |
1936 | if (++count >= bl && | |
1937 | (need_resched() || | |
1938 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
1939 | break; |
1940 | } | |
1941 | ||
1942 | local_irq_save(flags); | |
4968c300 PM |
1943 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
1944 | is_idle_task(current), | |
1945 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
1946 | |
1947 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
1948 | if (list != NULL) { |
1949 | *tail = rdp->nxtlist; | |
1950 | rdp->nxtlist = list; | |
b41772ab PM |
1951 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
1952 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
1953 | rdp->nxttail[i] = tail; | |
64db4cff PM |
1954 | else |
1955 | break; | |
1956 | } | |
b1420f1c PM |
1957 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
1958 | rdp->qlen_lazy -= count_lazy; | |
1d1fb395 | 1959 | ACCESS_ONCE(rdp->qlen) -= count; |
b1420f1c | 1960 | rdp->n_cbs_invoked += count; |
64db4cff PM |
1961 | |
1962 | /* Reinstate batch limit if we have worked down the excess. */ | |
1963 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
1964 | rdp->blimit = blimit; | |
1965 | ||
37c72e56 PM |
1966 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
1967 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
1968 | rdp->qlen_last_fqs_check = 0; | |
1969 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
1970 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
1971 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 1972 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 1973 | |
64db4cff PM |
1974 | local_irq_restore(flags); |
1975 | ||
e0f23060 | 1976 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 1977 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 1978 | invoke_rcu_core(); |
64db4cff PM |
1979 | } |
1980 | ||
1981 | /* | |
1982 | * Check to see if this CPU is in a non-context-switch quiescent state | |
1983 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 1984 | * Also schedule RCU core processing. |
64db4cff | 1985 | * |
9b2e4f18 | 1986 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
1987 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
1988 | * false, there is no point in invoking rcu_check_callbacks(). | |
1989 | */ | |
1990 | void rcu_check_callbacks(int cpu, int user) | |
1991 | { | |
300df91c | 1992 | trace_rcu_utilization("Start scheduler-tick"); |
a858af28 | 1993 | increment_cpu_stall_ticks(); |
9b2e4f18 | 1994 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
1995 | |
1996 | /* | |
1997 | * Get here if this CPU took its interrupt from user | |
1998 | * mode or from the idle loop, and if this is not a | |
1999 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2000 | * a quiescent state, so note it. |
64db4cff PM |
2001 | * |
2002 | * No memory barrier is required here because both | |
d6714c22 PM |
2003 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2004 | * variables that other CPUs neither access nor modify, | |
2005 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2006 | */ |
2007 | ||
d6714c22 PM |
2008 | rcu_sched_qs(cpu); |
2009 | rcu_bh_qs(cpu); | |
64db4cff PM |
2010 | |
2011 | } else if (!in_softirq()) { | |
2012 | ||
2013 | /* | |
2014 | * Get here if this CPU did not take its interrupt from | |
2015 | * softirq, in other words, if it is not interrupting | |
2016 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2017 | * critical section, so note it. |
64db4cff PM |
2018 | */ |
2019 | ||
d6714c22 | 2020 | rcu_bh_qs(cpu); |
64db4cff | 2021 | } |
f41d911f | 2022 | rcu_preempt_check_callbacks(cpu); |
d21670ac | 2023 | if (rcu_pending(cpu)) |
a46e0899 | 2024 | invoke_rcu_core(); |
300df91c | 2025 | trace_rcu_utilization("End scheduler-tick"); |
64db4cff PM |
2026 | } |
2027 | ||
64db4cff PM |
2028 | /* |
2029 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2030 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2031 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2032 | * | |
ee47eb9f | 2033 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2034 | */ |
45f014c5 | 2035 | static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *)) |
64db4cff PM |
2036 | { |
2037 | unsigned long bit; | |
2038 | int cpu; | |
2039 | unsigned long flags; | |
2040 | unsigned long mask; | |
a0b6c9a7 | 2041 | struct rcu_node *rnp; |
64db4cff | 2042 | |
a0b6c9a7 | 2043 | rcu_for_each_leaf_node(rsp, rnp) { |
b4be093f | 2044 | cond_resched(); |
64db4cff | 2045 | mask = 0; |
1304afb2 | 2046 | raw_spin_lock_irqsave(&rnp->lock, flags); |
ee47eb9f | 2047 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 2048 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 2049 | return; |
64db4cff | 2050 | } |
a0b6c9a7 | 2051 | if (rnp->qsmask == 0) { |
1217ed1b | 2052 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
64db4cff PM |
2053 | continue; |
2054 | } | |
a0b6c9a7 | 2055 | cpu = rnp->grplo; |
64db4cff | 2056 | bit = 1; |
a0b6c9a7 | 2057 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
394f99a9 LJ |
2058 | if ((rnp->qsmask & bit) != 0 && |
2059 | f(per_cpu_ptr(rsp->rda, cpu))) | |
64db4cff PM |
2060 | mask |= bit; |
2061 | } | |
45f014c5 | 2062 | if (mask != 0) { |
64db4cff | 2063 | |
d3f6bad3 PM |
2064 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
2065 | rcu_report_qs_rnp(mask, rsp, rnp, flags); | |
64db4cff PM |
2066 | continue; |
2067 | } | |
1304afb2 | 2068 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2069 | } |
27f4d280 | 2070 | rnp = rcu_get_root(rsp); |
1217ed1b PM |
2071 | if (rnp->qsmask == 0) { |
2072 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2073 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ | |
2074 | } | |
64db4cff PM |
2075 | } |
2076 | ||
2077 | /* | |
2078 | * Force quiescent states on reluctant CPUs, and also detect which | |
2079 | * CPUs are in dyntick-idle mode. | |
2080 | */ | |
4cdfc175 | 2081 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2082 | { |
2083 | unsigned long flags; | |
394f2769 PM |
2084 | bool ret; |
2085 | struct rcu_node *rnp; | |
2086 | struct rcu_node *rnp_old = NULL; | |
2087 | ||
2088 | /* Funnel through hierarchy to reduce memory contention. */ | |
2089 | rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode; | |
2090 | for (; rnp != NULL; rnp = rnp->parent) { | |
2091 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2092 | !raw_spin_trylock(&rnp->fqslock); | |
2093 | if (rnp_old != NULL) | |
2094 | raw_spin_unlock(&rnp_old->fqslock); | |
2095 | if (ret) { | |
2096 | rsp->n_force_qs_lh++; | |
2097 | return; | |
2098 | } | |
2099 | rnp_old = rnp; | |
2100 | } | |
2101 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2102 | |
394f2769 PM |
2103 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2104 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
2105 | raw_spin_unlock(&rnp_old->fqslock); | |
2106 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
2107 | rsp->n_force_qs_lh++; | |
2108 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); | |
4cdfc175 | 2109 | return; /* Someone beat us to it. */ |
46a1e34e | 2110 | } |
4cdfc175 | 2111 | rsp->gp_flags |= RCU_GP_FLAG_FQS; |
394f2769 | 2112 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2113 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ |
64db4cff PM |
2114 | } |
2115 | ||
64db4cff | 2116 | /* |
e0f23060 PM |
2117 | * This does the RCU core processing work for the specified rcu_state |
2118 | * and rcu_data structures. This may be called only from the CPU to | |
2119 | * whom the rdp belongs. | |
64db4cff PM |
2120 | */ |
2121 | static void | |
1bca8cf1 | 2122 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2123 | { |
2124 | unsigned long flags; | |
1bca8cf1 | 2125 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
64db4cff | 2126 | |
2e597558 PM |
2127 | WARN_ON_ONCE(rdp->beenonline == 0); |
2128 | ||
dc35c893 | 2129 | /* Handle the end of a grace period that some other CPU ended. */ |
64db4cff PM |
2130 | rcu_process_gp_end(rsp, rdp); |
2131 | ||
2132 | /* Update RCU state based on any recent quiescent states. */ | |
2133 | rcu_check_quiescent_state(rsp, rdp); | |
2134 | ||
2135 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2136 | local_irq_save(flags); |
64db4cff | 2137 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2138 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
64db4cff | 2139 | rcu_start_gp(rsp, flags); /* releases above lock */ |
dc35c893 PM |
2140 | } else { |
2141 | local_irq_restore(flags); | |
64db4cff PM |
2142 | } |
2143 | ||
2144 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2145 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2146 | invoke_rcu_callbacks(rsp, rdp); |
09223371 SL |
2147 | } |
2148 | ||
64db4cff | 2149 | /* |
e0f23060 | 2150 | * Do RCU core processing for the current CPU. |
64db4cff | 2151 | */ |
09223371 | 2152 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2153 | { |
6ce75a23 PM |
2154 | struct rcu_state *rsp; |
2155 | ||
bfa00b4c PM |
2156 | if (cpu_is_offline(smp_processor_id())) |
2157 | return; | |
300df91c | 2158 | trace_rcu_utilization("Start RCU core"); |
6ce75a23 PM |
2159 | for_each_rcu_flavor(rsp) |
2160 | __rcu_process_callbacks(rsp); | |
300df91c | 2161 | trace_rcu_utilization("End RCU core"); |
64db4cff PM |
2162 | } |
2163 | ||
a26ac245 | 2164 | /* |
e0f23060 PM |
2165 | * Schedule RCU callback invocation. If the specified type of RCU |
2166 | * does not support RCU priority boosting, just do a direct call, | |
2167 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
2168 | * are running on the current CPU with interrupts disabled, the | |
2169 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
a26ac245 | 2170 | */ |
a46e0899 | 2171 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2172 | { |
b0d30417 PM |
2173 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
2174 | return; | |
a46e0899 PM |
2175 | if (likely(!rsp->boost)) { |
2176 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2177 | return; |
2178 | } | |
a46e0899 | 2179 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2180 | } |
2181 | ||
a46e0899 | 2182 | static void invoke_rcu_core(void) |
09223371 SL |
2183 | { |
2184 | raise_softirq(RCU_SOFTIRQ); | |
2185 | } | |
2186 | ||
29154c57 PM |
2187 | /* |
2188 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2189 | */ | |
2190 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2191 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2192 | { |
62fde6ed PM |
2193 | /* |
2194 | * If called from an extended quiescent state, invoke the RCU | |
2195 | * core in order to force a re-evaluation of RCU's idleness. | |
2196 | */ | |
a16b7a69 | 2197 | if (rcu_is_cpu_idle() && cpu_online(smp_processor_id())) |
62fde6ed PM |
2198 | invoke_rcu_core(); |
2199 | ||
a16b7a69 | 2200 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2201 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2202 | return; |
64db4cff | 2203 | |
37c72e56 PM |
2204 | /* |
2205 | * Force the grace period if too many callbacks or too long waiting. | |
2206 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2207 | * if some other CPU has recently done so. Also, don't bother | |
2208 | * invoking force_quiescent_state() if the newly enqueued callback | |
2209 | * is the only one waiting for a grace period to complete. | |
2210 | */ | |
2655d57e | 2211 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2212 | |
2213 | /* Are we ignoring a completed grace period? */ | |
2214 | rcu_process_gp_end(rsp, rdp); | |
2215 | check_for_new_grace_period(rsp, rdp); | |
2216 | ||
2217 | /* Start a new grace period if one not already started. */ | |
2218 | if (!rcu_gp_in_progress(rsp)) { | |
2219 | unsigned long nestflag; | |
2220 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
2221 | ||
2222 | raw_spin_lock_irqsave(&rnp_root->lock, nestflag); | |
2223 | rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */ | |
2224 | } else { | |
2225 | /* Give the grace period a kick. */ | |
2226 | rdp->blimit = LONG_MAX; | |
2227 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2228 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2229 | force_quiescent_state(rsp); |
b52573d2 PM |
2230 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2231 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2232 | } | |
4cdfc175 | 2233 | } |
29154c57 PM |
2234 | } |
2235 | ||
3fbfbf7a PM |
2236 | /* |
2237 | * Helper function for call_rcu() and friends. The cpu argument will | |
2238 | * normally be -1, indicating "currently running CPU". It may specify | |
2239 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2240 | * is expected to specify a CPU. | |
2241 | */ | |
64db4cff PM |
2242 | static void |
2243 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 2244 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
2245 | { |
2246 | unsigned long flags; | |
2247 | struct rcu_data *rdp; | |
2248 | ||
0bb7b59d | 2249 | WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */ |
551d55a9 | 2250 | debug_rcu_head_queue(head); |
64db4cff PM |
2251 | head->func = func; |
2252 | head->next = NULL; | |
2253 | ||
64db4cff PM |
2254 | /* |
2255 | * Opportunistically note grace-period endings and beginnings. | |
2256 | * Note that we might see a beginning right after we see an | |
2257 | * end, but never vice versa, since this CPU has to pass through | |
2258 | * a quiescent state betweentimes. | |
2259 | */ | |
2260 | local_irq_save(flags); | |
394f99a9 | 2261 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
2262 | |
2263 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
2264 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
2265 | int offline; | |
2266 | ||
2267 | if (cpu != -1) | |
2268 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
2269 | offline = !__call_rcu_nocb(rdp, head, lazy); | |
2270 | WARN_ON_ONCE(offline); | |
0d8ee37e | 2271 | /* _call_rcu() is illegal on offline CPU; leak the callback. */ |
0d8ee37e PM |
2272 | local_irq_restore(flags); |
2273 | return; | |
2274 | } | |
29154c57 | 2275 | ACCESS_ONCE(rdp->qlen)++; |
486e2593 PM |
2276 | if (lazy) |
2277 | rdp->qlen_lazy++; | |
c57afe80 PM |
2278 | else |
2279 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2280 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
2281 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
2282 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 2283 | |
d4c08f2a PM |
2284 | if (__is_kfree_rcu_offset((unsigned long)func)) |
2285 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 2286 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2287 | else |
486e2593 | 2288 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2289 | |
29154c57 PM |
2290 | /* Go handle any RCU core processing required. */ |
2291 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
2292 | local_irq_restore(flags); |
2293 | } | |
2294 | ||
2295 | /* | |
d6714c22 | 2296 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 2297 | */ |
d6714c22 | 2298 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 2299 | { |
3fbfbf7a | 2300 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 2301 | } |
d6714c22 | 2302 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
2303 | |
2304 | /* | |
486e2593 | 2305 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
2306 | */ |
2307 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
2308 | { | |
3fbfbf7a | 2309 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
2310 | } |
2311 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
2312 | ||
6d813391 PM |
2313 | /* |
2314 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
2315 | * any blocking grace-period wait automatically implies a grace period | |
2316 | * if there is only one CPU online at any point time during execution | |
2317 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
2318 | * occasionally incorrectly indicate that there are multiple CPUs online | |
2319 | * when there was in fact only one the whole time, as this just adds | |
2320 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
2321 | */ |
2322 | static inline int rcu_blocking_is_gp(void) | |
2323 | { | |
95f0c1de PM |
2324 | int ret; |
2325 | ||
6d813391 | 2326 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
2327 | preempt_disable(); |
2328 | ret = num_online_cpus() <= 1; | |
2329 | preempt_enable(); | |
2330 | return ret; | |
6d813391 PM |
2331 | } |
2332 | ||
6ebb237b PM |
2333 | /** |
2334 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
2335 | * | |
2336 | * Control will return to the caller some time after a full rcu-sched | |
2337 | * grace period has elapsed, in other words after all currently executing | |
2338 | * rcu-sched read-side critical sections have completed. These read-side | |
2339 | * critical sections are delimited by rcu_read_lock_sched() and | |
2340 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
2341 | * local_irq_disable(), and so on may be used in place of | |
2342 | * rcu_read_lock_sched(). | |
2343 | * | |
2344 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
2345 | * non-threaded hardware-interrupt handlers, in progress on entry will |
2346 | * have completed before this primitive returns. However, this does not | |
2347 | * guarantee that softirq handlers will have completed, since in some | |
2348 | * kernels, these handlers can run in process context, and can block. | |
2349 | * | |
2350 | * Note that this guarantee implies further memory-ordering guarantees. | |
2351 | * On systems with more than one CPU, when synchronize_sched() returns, | |
2352 | * each CPU is guaranteed to have executed a full memory barrier since the | |
2353 | * end of its last RCU-sched read-side critical section whose beginning | |
2354 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
2355 | * an RCU read-side critical section that extends beyond the return from | |
2356 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
2357 | * after the beginning of synchronize_sched() and before the beginning of | |
2358 | * that RCU read-side critical section. Note that these guarantees include | |
2359 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
2360 | * that are executing in the kernel. | |
2361 | * | |
2362 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
2363 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
2364 | * to have executed a full memory barrier during the execution of | |
2365 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
2366 | * again only if the system has more than one CPU). | |
6ebb237b PM |
2367 | * |
2368 | * This primitive provides the guarantees made by the (now removed) | |
2369 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
2370 | * guarantees that rcu_read_lock() sections will have completed. | |
2371 | * In "classic RCU", these two guarantees happen to be one and | |
2372 | * the same, but can differ in realtime RCU implementations. | |
2373 | */ | |
2374 | void synchronize_sched(void) | |
2375 | { | |
fe15d706 PM |
2376 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2377 | !lock_is_held(&rcu_lock_map) && | |
2378 | !lock_is_held(&rcu_sched_lock_map), | |
2379 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
2380 | if (rcu_blocking_is_gp()) |
2381 | return; | |
3705b88d AM |
2382 | if (rcu_expedited) |
2383 | synchronize_sched_expedited(); | |
2384 | else | |
2385 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
2386 | } |
2387 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
2388 | ||
2389 | /** | |
2390 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
2391 | * | |
2392 | * Control will return to the caller some time after a full rcu_bh grace | |
2393 | * period has elapsed, in other words after all currently executing rcu_bh | |
2394 | * read-side critical sections have completed. RCU read-side critical | |
2395 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
2396 | * and may be nested. | |
f0a0e6f2 PM |
2397 | * |
2398 | * See the description of synchronize_sched() for more detailed information | |
2399 | * on memory ordering guarantees. | |
6ebb237b PM |
2400 | */ |
2401 | void synchronize_rcu_bh(void) | |
2402 | { | |
fe15d706 PM |
2403 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2404 | !lock_is_held(&rcu_lock_map) && | |
2405 | !lock_is_held(&rcu_sched_lock_map), | |
2406 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
2407 | if (rcu_blocking_is_gp()) |
2408 | return; | |
3705b88d AM |
2409 | if (rcu_expedited) |
2410 | synchronize_rcu_bh_expedited(); | |
2411 | else | |
2412 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
2413 | } |
2414 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
2415 | ||
3d3b7db0 PM |
2416 | static int synchronize_sched_expedited_cpu_stop(void *data) |
2417 | { | |
2418 | /* | |
2419 | * There must be a full memory barrier on each affected CPU | |
2420 | * between the time that try_stop_cpus() is called and the | |
2421 | * time that it returns. | |
2422 | * | |
2423 | * In the current initial implementation of cpu_stop, the | |
2424 | * above condition is already met when the control reaches | |
2425 | * this point and the following smp_mb() is not strictly | |
2426 | * necessary. Do smp_mb() anyway for documentation and | |
2427 | * robustness against future implementation changes. | |
2428 | */ | |
2429 | smp_mb(); /* See above comment block. */ | |
2430 | return 0; | |
2431 | } | |
2432 | ||
236fefaf PM |
2433 | /** |
2434 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
2435 | * | |
2436 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
2437 | * approach to force the grace period to end quickly. This consumes | |
2438 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
2439 | * so is thus not recommended for any sort of common-case code. In fact, | |
2440 | * if you are using synchronize_sched_expedited() in a loop, please | |
2441 | * restructure your code to batch your updates, and then use a single | |
2442 | * synchronize_sched() instead. | |
3d3b7db0 | 2443 | * |
236fefaf PM |
2444 | * Note that it is illegal to call this function while holding any lock |
2445 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal | |
2446 | * to call this function from a CPU-hotplug notifier. Failing to observe | |
2447 | * these restriction will result in deadlock. | |
3d3b7db0 PM |
2448 | * |
2449 | * This implementation can be thought of as an application of ticket | |
2450 | * locking to RCU, with sync_sched_expedited_started and | |
2451 | * sync_sched_expedited_done taking on the roles of the halves | |
2452 | * of the ticket-lock word. Each task atomically increments | |
2453 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
2454 | * then attempts to stop all the CPUs. If this succeeds, then each | |
2455 | * CPU will have executed a context switch, resulting in an RCU-sched | |
2456 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
2457 | * update sync_sched_expedited_done to match our snapshot -- but | |
2458 | * only if someone else has not already advanced past our snapshot. | |
2459 | * | |
2460 | * On the other hand, if try_stop_cpus() fails, we check the value | |
2461 | * of sync_sched_expedited_done. If it has advanced past our | |
2462 | * initial snapshot, then someone else must have forced a grace period | |
2463 | * some time after we took our snapshot. In this case, our work is | |
2464 | * done for us, and we can simply return. Otherwise, we try again, | |
2465 | * but keep our initial snapshot for purposes of checking for someone | |
2466 | * doing our work for us. | |
2467 | * | |
2468 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
2469 | */ | |
2470 | void synchronize_sched_expedited(void) | |
2471 | { | |
1924bcb0 PM |
2472 | long firstsnap, s, snap; |
2473 | int trycount = 0; | |
40694d66 | 2474 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 2475 | |
1924bcb0 PM |
2476 | /* |
2477 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
2478 | * By allowing sync_sched_expedited_started to advance no more than | |
2479 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
2480 | * that more than 3.5 billion CPUs would be required to force a | |
2481 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
2482 | * course be required on a 64-bit system. | |
2483 | */ | |
40694d66 PM |
2484 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
2485 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 PM |
2486 | ULONG_MAX / 8)) { |
2487 | synchronize_sched(); | |
a30489c5 | 2488 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
2489 | return; |
2490 | } | |
3d3b7db0 | 2491 | |
1924bcb0 PM |
2492 | /* |
2493 | * Take a ticket. Note that atomic_inc_return() implies a | |
2494 | * full memory barrier. | |
2495 | */ | |
40694d66 | 2496 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 2497 | firstsnap = snap; |
3d3b7db0 | 2498 | get_online_cpus(); |
1cc85961 | 2499 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 PM |
2500 | |
2501 | /* | |
2502 | * Each pass through the following loop attempts to force a | |
2503 | * context switch on each CPU. | |
2504 | */ | |
2505 | while (try_stop_cpus(cpu_online_mask, | |
2506 | synchronize_sched_expedited_cpu_stop, | |
2507 | NULL) == -EAGAIN) { | |
2508 | put_online_cpus(); | |
a30489c5 | 2509 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 2510 | |
1924bcb0 | 2511 | /* Check to see if someone else did our work for us. */ |
40694d66 | 2512 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2513 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2514 | /* ensure test happens before caller kfree */ |
2515 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2516 | atomic_long_inc(&rsp->expedited_workdone1); | |
1924bcb0 PM |
2517 | return; |
2518 | } | |
3d3b7db0 PM |
2519 | |
2520 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 2521 | if (trycount++ < 10) { |
3d3b7db0 | 2522 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 2523 | } else { |
3705b88d | 2524 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 2525 | atomic_long_inc(&rsp->expedited_normal); |
3d3b7db0 PM |
2526 | return; |
2527 | } | |
2528 | ||
1924bcb0 | 2529 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 2530 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2531 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2532 | /* ensure test happens before caller kfree */ |
2533 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2534 | atomic_long_inc(&rsp->expedited_workdone2); | |
3d3b7db0 PM |
2535 | return; |
2536 | } | |
2537 | ||
2538 | /* | |
2539 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
2540 | * callers to piggyback on our grace period. We retry |
2541 | * after they started, so our grace period works for them, | |
2542 | * and they started after our first try, so their grace | |
2543 | * period works for us. | |
3d3b7db0 PM |
2544 | */ |
2545 | get_online_cpus(); | |
40694d66 | 2546 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
2547 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
2548 | } | |
a30489c5 | 2549 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 PM |
2550 | |
2551 | /* | |
2552 | * Everyone up to our most recent fetch is covered by our grace | |
2553 | * period. Update the counter, but only if our work is still | |
2554 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 2555 | * than we did already did their update. |
3d3b7db0 PM |
2556 | */ |
2557 | do { | |
a30489c5 | 2558 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 2559 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2560 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 PM |
2561 | /* ensure test happens before caller kfree */ |
2562 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2563 | atomic_long_inc(&rsp->expedited_done_lost); | |
3d3b7db0 PM |
2564 | break; |
2565 | } | |
40694d66 | 2566 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 2567 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
2568 | |
2569 | put_online_cpus(); | |
2570 | } | |
2571 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
2572 | ||
64db4cff PM |
2573 | /* |
2574 | * Check to see if there is any immediate RCU-related work to be done | |
2575 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
2576 | * The checks are in order of increasing expense: checks that can be | |
2577 | * carried out against CPU-local state are performed first. However, | |
2578 | * we must check for CPU stalls first, else we might not get a chance. | |
2579 | */ | |
2580 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
2581 | { | |
2f51f988 PM |
2582 | struct rcu_node *rnp = rdp->mynode; |
2583 | ||
64db4cff PM |
2584 | rdp->n_rcu_pending++; |
2585 | ||
2586 | /* Check for CPU stalls, if enabled. */ | |
2587 | check_cpu_stall(rsp, rdp); | |
2588 | ||
2589 | /* Is the RCU core waiting for a quiescent state from this CPU? */ | |
5c51dd73 PM |
2590 | if (rcu_scheduler_fully_active && |
2591 | rdp->qs_pending && !rdp->passed_quiesce) { | |
d21670ac | 2592 | rdp->n_rp_qs_pending++; |
e4cc1f22 | 2593 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
d21670ac | 2594 | rdp->n_rp_report_qs++; |
64db4cff | 2595 | return 1; |
7ba5c840 | 2596 | } |
64db4cff PM |
2597 | |
2598 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
2599 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
2600 | rdp->n_rp_cb_ready++; | |
64db4cff | 2601 | return 1; |
7ba5c840 | 2602 | } |
64db4cff PM |
2603 | |
2604 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
2605 | if (cpu_needs_another_gp(rsp, rdp)) { |
2606 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 2607 | return 1; |
7ba5c840 | 2608 | } |
64db4cff PM |
2609 | |
2610 | /* Has another RCU grace period completed? */ | |
2f51f988 | 2611 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 2612 | rdp->n_rp_gp_completed++; |
64db4cff | 2613 | return 1; |
7ba5c840 | 2614 | } |
64db4cff PM |
2615 | |
2616 | /* Has a new RCU grace period started? */ | |
2f51f988 | 2617 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
7ba5c840 | 2618 | rdp->n_rp_gp_started++; |
64db4cff | 2619 | return 1; |
7ba5c840 | 2620 | } |
64db4cff | 2621 | |
64db4cff | 2622 | /* nothing to do */ |
7ba5c840 | 2623 | rdp->n_rp_need_nothing++; |
64db4cff PM |
2624 | return 0; |
2625 | } | |
2626 | ||
2627 | /* | |
2628 | * Check to see if there is any immediate RCU-related work to be done | |
2629 | * by the current CPU, returning 1 if so. This function is part of the | |
2630 | * RCU implementation; it is -not- an exported member of the RCU API. | |
2631 | */ | |
a157229c | 2632 | static int rcu_pending(int cpu) |
64db4cff | 2633 | { |
6ce75a23 PM |
2634 | struct rcu_state *rsp; |
2635 | ||
2636 | for_each_rcu_flavor(rsp) | |
2637 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu))) | |
2638 | return 1; | |
2639 | return 0; | |
64db4cff PM |
2640 | } |
2641 | ||
2642 | /* | |
2643 | * Check to see if any future RCU-related work will need to be done | |
2644 | * by the current CPU, even if none need be done immediately, returning | |
8bd93a2c | 2645 | * 1 if so. |
64db4cff | 2646 | */ |
aea1b35e | 2647 | static int rcu_cpu_has_callbacks(int cpu) |
64db4cff | 2648 | { |
6ce75a23 PM |
2649 | struct rcu_state *rsp; |
2650 | ||
64db4cff | 2651 | /* RCU callbacks either ready or pending? */ |
6ce75a23 PM |
2652 | for_each_rcu_flavor(rsp) |
2653 | if (per_cpu_ptr(rsp->rda, cpu)->nxtlist) | |
2654 | return 1; | |
2655 | return 0; | |
64db4cff PM |
2656 | } |
2657 | ||
a83eff0a PM |
2658 | /* |
2659 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
2660 | * the compiler is expected to optimize this away. | |
2661 | */ | |
2662 | static void _rcu_barrier_trace(struct rcu_state *rsp, char *s, | |
2663 | int cpu, unsigned long done) | |
2664 | { | |
2665 | trace_rcu_barrier(rsp->name, s, cpu, | |
2666 | atomic_read(&rsp->barrier_cpu_count), done); | |
2667 | } | |
2668 | ||
b1420f1c PM |
2669 | /* |
2670 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
2671 | * up the task executing _rcu_barrier(). | |
2672 | */ | |
24ebbca8 | 2673 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 2674 | { |
24ebbca8 PM |
2675 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
2676 | struct rcu_state *rsp = rdp->rsp; | |
2677 | ||
a83eff0a PM |
2678 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
2679 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 2680 | complete(&rsp->barrier_completion); |
a83eff0a PM |
2681 | } else { |
2682 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
2683 | } | |
d0ec774c PM |
2684 | } |
2685 | ||
2686 | /* | |
2687 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
2688 | */ | |
2689 | static void rcu_barrier_func(void *type) | |
2690 | { | |
037b64ed | 2691 | struct rcu_state *rsp = type; |
06668efa | 2692 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
d0ec774c | 2693 | |
a83eff0a | 2694 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 2695 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 2696 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
2697 | } |
2698 | ||
d0ec774c PM |
2699 | /* |
2700 | * Orchestrate the specified type of RCU barrier, waiting for all | |
2701 | * RCU callbacks of the specified type to complete. | |
2702 | */ | |
037b64ed | 2703 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 2704 | { |
b1420f1c | 2705 | int cpu; |
b1420f1c | 2706 | struct rcu_data *rdp; |
cf3a9c48 PM |
2707 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
2708 | unsigned long snap_done; | |
b1420f1c | 2709 | |
a83eff0a | 2710 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 2711 | |
e74f4c45 | 2712 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 2713 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 2714 | |
cf3a9c48 PM |
2715 | /* |
2716 | * Ensure that all prior references, including to ->n_barrier_done, | |
2717 | * are ordered before the _rcu_barrier() machinery. | |
2718 | */ | |
2719 | smp_mb(); /* See above block comment. */ | |
2720 | ||
2721 | /* | |
2722 | * Recheck ->n_barrier_done to see if others did our work for us. | |
2723 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
2724 | * transition. The "if" expression below therefore rounds the old | |
2725 | * value up to the next even number and adds two before comparing. | |
2726 | */ | |
2727 | snap_done = ACCESS_ONCE(rsp->n_barrier_done); | |
a83eff0a | 2728 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
cf3a9c48 | 2729 | if (ULONG_CMP_GE(snap_done, ((snap + 1) & ~0x1) + 2)) { |
a83eff0a | 2730 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
2731 | smp_mb(); /* caller's subsequent code after above check. */ |
2732 | mutex_unlock(&rsp->barrier_mutex); | |
2733 | return; | |
2734 | } | |
2735 | ||
2736 | /* | |
2737 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
2738 | * ACCESS_ONCE() to prevent the compiler from speculating | |
2739 | * the increment to precede the early-exit check. | |
2740 | */ | |
2741 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2742 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); | |
a83eff0a | 2743 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 2744 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 2745 | |
d0ec774c | 2746 | /* |
b1420f1c PM |
2747 | * Initialize the count to one rather than to zero in order to |
2748 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
2749 | * (or preemption of this task). Exclude CPU-hotplug operations |
2750 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 2751 | */ |
7db74df8 | 2752 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 2753 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 2754 | get_online_cpus(); |
b1420f1c PM |
2755 | |
2756 | /* | |
1331e7a1 PM |
2757 | * Force each CPU with callbacks to register a new callback. |
2758 | * When that callback is invoked, we will know that all of the | |
2759 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 2760 | */ |
3fbfbf7a PM |
2761 | for_each_possible_cpu(cpu) { |
2762 | if (!cpu_online(cpu) && !is_nocb_cpu(cpu)) | |
2763 | continue; | |
b1420f1c | 2764 | rdp = per_cpu_ptr(rsp->rda, cpu); |
3fbfbf7a PM |
2765 | if (is_nocb_cpu(cpu)) { |
2766 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, | |
2767 | rsp->n_barrier_done); | |
2768 | atomic_inc(&rsp->barrier_cpu_count); | |
2769 | __call_rcu(&rdp->barrier_head, rcu_barrier_callback, | |
2770 | rsp, cpu, 0); | |
2771 | } else if (ACCESS_ONCE(rdp->qlen)) { | |
a83eff0a PM |
2772 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
2773 | rsp->n_barrier_done); | |
037b64ed | 2774 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 2775 | } else { |
a83eff0a PM |
2776 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
2777 | rsp->n_barrier_done); | |
b1420f1c PM |
2778 | } |
2779 | } | |
1331e7a1 | 2780 | put_online_cpus(); |
b1420f1c PM |
2781 | |
2782 | /* | |
2783 | * Now that we have an rcu_barrier_callback() callback on each | |
2784 | * CPU, and thus each counted, remove the initial count. | |
2785 | */ | |
24ebbca8 | 2786 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 2787 | complete(&rsp->barrier_completion); |
b1420f1c | 2788 | |
cf3a9c48 PM |
2789 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
2790 | smp_mb(); /* Keep increment after above mechanism. */ | |
2791 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2792 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); | |
a83eff0a | 2793 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
2794 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
2795 | ||
b1420f1c | 2796 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 2797 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
2798 | |
2799 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 2800 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 2801 | } |
d0ec774c PM |
2802 | |
2803 | /** | |
2804 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
2805 | */ | |
2806 | void rcu_barrier_bh(void) | |
2807 | { | |
037b64ed | 2808 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
2809 | } |
2810 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
2811 | ||
2812 | /** | |
2813 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
2814 | */ | |
2815 | void rcu_barrier_sched(void) | |
2816 | { | |
037b64ed | 2817 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
2818 | } |
2819 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
2820 | ||
64db4cff | 2821 | /* |
27569620 | 2822 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 2823 | */ |
27569620 PM |
2824 | static void __init |
2825 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
2826 | { |
2827 | unsigned long flags; | |
394f99a9 | 2828 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
2829 | struct rcu_node *rnp = rcu_get_root(rsp); |
2830 | ||
2831 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 2832 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 2833 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
3f5d3ea6 | 2834 | init_callback_list(rdp); |
486e2593 | 2835 | rdp->qlen_lazy = 0; |
1d1fb395 | 2836 | ACCESS_ONCE(rdp->qlen) = 0; |
27569620 | 2837 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 2838 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 2839 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 2840 | rdp->cpu = cpu; |
d4c08f2a | 2841 | rdp->rsp = rsp; |
3fbfbf7a | 2842 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 2843 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
2844 | } |
2845 | ||
2846 | /* | |
2847 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
2848 | * offline event can be happening at a given time. Note also that we | |
2849 | * can accept some slop in the rsp->completed access due to the fact | |
2850 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 2851 | */ |
e4fa4c97 | 2852 | static void __cpuinit |
6cc68793 | 2853 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible) |
64db4cff PM |
2854 | { |
2855 | unsigned long flags; | |
64db4cff | 2856 | unsigned long mask; |
394f99a9 | 2857 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
2858 | struct rcu_node *rnp = rcu_get_root(rsp); |
2859 | ||
a4fbe35a PM |
2860 | /* Exclude new grace periods. */ |
2861 | mutex_lock(&rsp->onoff_mutex); | |
2862 | ||
64db4cff | 2863 | /* Set up local state, ensuring consistent view of global state. */ |
1304afb2 | 2864 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 2865 | rdp->beenonline = 1; /* We have now been online. */ |
6cc68793 | 2866 | rdp->preemptible = preemptible; |
37c72e56 PM |
2867 | rdp->qlen_last_fqs_check = 0; |
2868 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 2869 | rdp->blimit = blimit; |
0d8ee37e | 2870 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ |
29e37d81 | 2871 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
c92b131b PM |
2872 | atomic_set(&rdp->dynticks->dynticks, |
2873 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
7cb92499 | 2874 | rcu_prepare_for_idle_init(cpu); |
1304afb2 | 2875 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 2876 | |
64db4cff PM |
2877 | /* Add CPU to rcu_node bitmasks. */ |
2878 | rnp = rdp->mynode; | |
2879 | mask = rdp->grpmask; | |
2880 | do { | |
2881 | /* Exclude any attempts to start a new GP on small systems. */ | |
1304afb2 | 2882 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
2883 | rnp->qsmaskinit |= mask; |
2884 | mask = rnp->grpmask; | |
d09b62df | 2885 | if (rnp == rdp->mynode) { |
06ae115a PM |
2886 | /* |
2887 | * If there is a grace period in progress, we will | |
2888 | * set up to wait for it next time we run the | |
2889 | * RCU core code. | |
2890 | */ | |
2891 | rdp->gpnum = rnp->completed; | |
d09b62df | 2892 | rdp->completed = rnp->completed; |
06ae115a PM |
2893 | rdp->passed_quiesce = 0; |
2894 | rdp->qs_pending = 0; | |
d4c08f2a | 2895 | trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl"); |
d09b62df | 2896 | } |
1304afb2 | 2897 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
2898 | rnp = rnp->parent; |
2899 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); | |
a4fbe35a | 2900 | local_irq_restore(flags); |
64db4cff | 2901 | |
a4fbe35a | 2902 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
2903 | } |
2904 | ||
d72bce0e | 2905 | static void __cpuinit rcu_prepare_cpu(int cpu) |
64db4cff | 2906 | { |
6ce75a23 PM |
2907 | struct rcu_state *rsp; |
2908 | ||
2909 | for_each_rcu_flavor(rsp) | |
2910 | rcu_init_percpu_data(cpu, rsp, | |
2911 | strcmp(rsp->name, "rcu_preempt") == 0); | |
64db4cff PM |
2912 | } |
2913 | ||
2914 | /* | |
f41d911f | 2915 | * Handle CPU online/offline notification events. |
64db4cff | 2916 | */ |
9f680ab4 PM |
2917 | static int __cpuinit rcu_cpu_notify(struct notifier_block *self, |
2918 | unsigned long action, void *hcpu) | |
64db4cff PM |
2919 | { |
2920 | long cpu = (long)hcpu; | |
27f4d280 | 2921 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); |
a26ac245 | 2922 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 2923 | struct rcu_state *rsp; |
64db4cff | 2924 | |
300df91c | 2925 | trace_rcu_utilization("Start CPU hotplug"); |
64db4cff PM |
2926 | switch (action) { |
2927 | case CPU_UP_PREPARE: | |
2928 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
2929 | rcu_prepare_cpu(cpu); |
2930 | rcu_prepare_kthreads(cpu); | |
a26ac245 PM |
2931 | break; |
2932 | case CPU_ONLINE: | |
0f962a5e | 2933 | case CPU_DOWN_FAILED: |
5d01bbd1 | 2934 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
2935 | break; |
2936 | case CPU_DOWN_PREPARE: | |
34ed6246 | 2937 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 2938 | break; |
d0ec774c PM |
2939 | case CPU_DYING: |
2940 | case CPU_DYING_FROZEN: | |
2941 | /* | |
2d999e03 PM |
2942 | * The whole machine is "stopped" except this CPU, so we can |
2943 | * touch any data without introducing corruption. We send the | |
2944 | * dying CPU's callbacks to an arbitrarily chosen online CPU. | |
d0ec774c | 2945 | */ |
6ce75a23 PM |
2946 | for_each_rcu_flavor(rsp) |
2947 | rcu_cleanup_dying_cpu(rsp); | |
7cb92499 | 2948 | rcu_cleanup_after_idle(cpu); |
d0ec774c | 2949 | break; |
64db4cff PM |
2950 | case CPU_DEAD: |
2951 | case CPU_DEAD_FROZEN: | |
2952 | case CPU_UP_CANCELED: | |
2953 | case CPU_UP_CANCELED_FROZEN: | |
6ce75a23 PM |
2954 | for_each_rcu_flavor(rsp) |
2955 | rcu_cleanup_dead_cpu(cpu, rsp); | |
64db4cff PM |
2956 | break; |
2957 | default: | |
2958 | break; | |
2959 | } | |
300df91c | 2960 | trace_rcu_utilization("End CPU hotplug"); |
34ed6246 | 2961 | return NOTIFY_OK; |
64db4cff PM |
2962 | } |
2963 | ||
b3dbec76 PM |
2964 | /* |
2965 | * Spawn the kthread that handles this RCU flavor's grace periods. | |
2966 | */ | |
2967 | static int __init rcu_spawn_gp_kthread(void) | |
2968 | { | |
2969 | unsigned long flags; | |
2970 | struct rcu_node *rnp; | |
2971 | struct rcu_state *rsp; | |
2972 | struct task_struct *t; | |
2973 | ||
2974 | for_each_rcu_flavor(rsp) { | |
2975 | t = kthread_run(rcu_gp_kthread, rsp, rsp->name); | |
2976 | BUG_ON(IS_ERR(t)); | |
2977 | rnp = rcu_get_root(rsp); | |
2978 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2979 | rsp->gp_kthread = t; | |
2980 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3fbfbf7a | 2981 | rcu_spawn_nocb_kthreads(rsp); |
b3dbec76 PM |
2982 | } |
2983 | return 0; | |
2984 | } | |
2985 | early_initcall(rcu_spawn_gp_kthread); | |
2986 | ||
bbad9379 PM |
2987 | /* |
2988 | * This function is invoked towards the end of the scheduler's initialization | |
2989 | * process. Before this is called, the idle task might contain | |
2990 | * RCU read-side critical sections (during which time, this idle | |
2991 | * task is booting the system). After this function is called, the | |
2992 | * idle tasks are prohibited from containing RCU read-side critical | |
2993 | * sections. This function also enables RCU lockdep checking. | |
2994 | */ | |
2995 | void rcu_scheduler_starting(void) | |
2996 | { | |
2997 | WARN_ON(num_online_cpus() != 1); | |
2998 | WARN_ON(nr_context_switches() > 0); | |
2999 | rcu_scheduler_active = 1; | |
3000 | } | |
3001 | ||
64db4cff PM |
3002 | /* |
3003 | * Compute the per-level fanout, either using the exact fanout specified | |
3004 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
3005 | */ | |
3006 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
3007 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3008 | { | |
3009 | int i; | |
3010 | ||
f885b7f2 | 3011 | for (i = rcu_num_lvls - 1; i > 0; i--) |
64db4cff | 3012 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
f885b7f2 | 3013 | rsp->levelspread[0] = rcu_fanout_leaf; |
64db4cff PM |
3014 | } |
3015 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3016 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3017 | { | |
3018 | int ccur; | |
3019 | int cprv; | |
3020 | int i; | |
3021 | ||
4dbd6bb3 | 3022 | cprv = nr_cpu_ids; |
f885b7f2 | 3023 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3024 | ccur = rsp->levelcnt[i]; |
3025 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
3026 | cprv = ccur; | |
3027 | } | |
3028 | } | |
3029 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3030 | ||
3031 | /* | |
3032 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3033 | */ | |
394f99a9 LJ |
3034 | static void __init rcu_init_one(struct rcu_state *rsp, |
3035 | struct rcu_data __percpu *rda) | |
64db4cff | 3036 | { |
394f2769 PM |
3037 | static char *buf[] = { "rcu_node_0", |
3038 | "rcu_node_1", | |
3039 | "rcu_node_2", | |
3040 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ | |
3041 | static char *fqs[] = { "rcu_node_fqs_0", | |
3042 | "rcu_node_fqs_1", | |
3043 | "rcu_node_fqs_2", | |
3044 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ | |
64db4cff PM |
3045 | int cpustride = 1; |
3046 | int i; | |
3047 | int j; | |
3048 | struct rcu_node *rnp; | |
3049 | ||
b6407e86 PM |
3050 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
3051 | ||
4930521a PM |
3052 | /* Silence gcc 4.8 warning about array index out of range. */ |
3053 | if (rcu_num_lvls > RCU_NUM_LVLS) | |
3054 | panic("rcu_init_one: rcu_num_lvls overflow"); | |
3055 | ||
64db4cff PM |
3056 | /* Initialize the level-tracking arrays. */ |
3057 | ||
f885b7f2 PM |
3058 | for (i = 0; i < rcu_num_lvls; i++) |
3059 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
3060 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
3061 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
3062 | rcu_init_levelspread(rsp); | |
3063 | ||
3064 | /* Initialize the elements themselves, starting from the leaves. */ | |
3065 | ||
f885b7f2 | 3066 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3067 | cpustride *= rsp->levelspread[i]; |
3068 | rnp = rsp->level[i]; | |
3069 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 3070 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
3071 | lockdep_set_class_and_name(&rnp->lock, |
3072 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
3073 | raw_spin_lock_init(&rnp->fqslock); |
3074 | lockdep_set_class_and_name(&rnp->fqslock, | |
3075 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
3076 | rnp->gpnum = rsp->gpnum; |
3077 | rnp->completed = rsp->completed; | |
64db4cff PM |
3078 | rnp->qsmask = 0; |
3079 | rnp->qsmaskinit = 0; | |
3080 | rnp->grplo = j * cpustride; | |
3081 | rnp->grphi = (j + 1) * cpustride - 1; | |
3082 | if (rnp->grphi >= NR_CPUS) | |
3083 | rnp->grphi = NR_CPUS - 1; | |
3084 | if (i == 0) { | |
3085 | rnp->grpnum = 0; | |
3086 | rnp->grpmask = 0; | |
3087 | rnp->parent = NULL; | |
3088 | } else { | |
3089 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
3090 | rnp->grpmask = 1UL << rnp->grpnum; | |
3091 | rnp->parent = rsp->level[i - 1] + | |
3092 | j / rsp->levelspread[i - 1]; | |
3093 | } | |
3094 | rnp->level = i; | |
12f5f524 | 3095 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 3096 | rcu_init_one_nocb(rnp); |
64db4cff PM |
3097 | } |
3098 | } | |
0c34029a | 3099 | |
394f99a9 | 3100 | rsp->rda = rda; |
b3dbec76 | 3101 | init_waitqueue_head(&rsp->gp_wq); |
f885b7f2 | 3102 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 3103 | for_each_possible_cpu(i) { |
4a90a068 | 3104 | while (i > rnp->grphi) |
0c34029a | 3105 | rnp++; |
394f99a9 | 3106 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
3107 | rcu_boot_init_percpu_data(i, rsp); |
3108 | } | |
6ce75a23 | 3109 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
3110 | } |
3111 | ||
f885b7f2 PM |
3112 | /* |
3113 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
3114 | * replace the definitions in rcutree.h because those are needed to size | |
3115 | * the ->node array in the rcu_state structure. | |
3116 | */ | |
3117 | static void __init rcu_init_geometry(void) | |
3118 | { | |
3119 | int i; | |
3120 | int j; | |
cca6f393 | 3121 | int n = nr_cpu_ids; |
f885b7f2 PM |
3122 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
3123 | ||
3124 | /* If the compile-time values are accurate, just leave. */ | |
b17c7035 PM |
3125 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
3126 | nr_cpu_ids == NR_CPUS) | |
f885b7f2 PM |
3127 | return; |
3128 | ||
3129 | /* | |
3130 | * Compute number of nodes that can be handled an rcu_node tree | |
3131 | * with the given number of levels. Setting rcu_capacity[0] makes | |
3132 | * some of the arithmetic easier. | |
3133 | */ | |
3134 | rcu_capacity[0] = 1; | |
3135 | rcu_capacity[1] = rcu_fanout_leaf; | |
3136 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
3137 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
3138 | ||
3139 | /* | |
3140 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
3141 | * to increase the leaf-level fanout, not decrease it. Of course, | |
3142 | * the leaf-level fanout cannot exceed the number of bits in | |
3143 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
3144 | * the configured number of CPUs. Complain and fall back to the | |
3145 | * compile-time values if these limits are exceeded. | |
3146 | */ | |
3147 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
3148 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
3149 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
3150 | WARN_ON(1); | |
3151 | return; | |
3152 | } | |
3153 | ||
3154 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
3155 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
3156 | if (n <= rcu_capacity[i]) { | |
3157 | for (j = 0; j <= i; j++) | |
3158 | num_rcu_lvl[j] = | |
3159 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
3160 | rcu_num_lvls = i; | |
3161 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
3162 | num_rcu_lvl[j] = 0; | |
3163 | break; | |
3164 | } | |
3165 | ||
3166 | /* Calculate the total number of rcu_node structures. */ | |
3167 | rcu_num_nodes = 0; | |
3168 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
3169 | rcu_num_nodes += num_rcu_lvl[i]; | |
3170 | rcu_num_nodes -= n; | |
3171 | } | |
3172 | ||
9f680ab4 | 3173 | void __init rcu_init(void) |
64db4cff | 3174 | { |
017c4261 | 3175 | int cpu; |
9f680ab4 | 3176 | |
f41d911f | 3177 | rcu_bootup_announce(); |
f885b7f2 | 3178 | rcu_init_geometry(); |
394f99a9 LJ |
3179 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
3180 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); | |
f41d911f | 3181 | __rcu_init_preempt(); |
09223371 | 3182 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
3183 | |
3184 | /* | |
3185 | * We don't need protection against CPU-hotplug here because | |
3186 | * this is called early in boot, before either interrupts | |
3187 | * or the scheduler are operational. | |
3188 | */ | |
3189 | cpu_notifier(rcu_cpu_notify, 0); | |
017c4261 PM |
3190 | for_each_online_cpu(cpu) |
3191 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
3192 | } |
3193 | ||
1eba8f84 | 3194 | #include "rcutree_plugin.h" |