Commit | Line | Data |
---|---|---|
64db4cff PM |
1 | /* |
2 | * Read-Copy Update mechanism for mutual exclusion | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public License | |
15 | * along with this program; if not, write to the Free Software | |
16 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. | |
17 | * | |
18 | * Copyright IBM Corporation, 2008 | |
19 | * | |
20 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> | |
21 | * Manfred Spraul <manfred@colorfullife.com> | |
22 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version | |
23 | * | |
24 | * Based on the original work by Paul McKenney <paulmck@us.ibm.com> | |
25 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. | |
26 | * | |
27 | * For detailed explanation of Read-Copy Update mechanism see - | |
a71fca58 | 28 | * Documentation/RCU |
64db4cff PM |
29 | */ |
30 | #include <linux/types.h> | |
31 | #include <linux/kernel.h> | |
32 | #include <linux/init.h> | |
33 | #include <linux/spinlock.h> | |
34 | #include <linux/smp.h> | |
35 | #include <linux/rcupdate.h> | |
36 | #include <linux/interrupt.h> | |
37 | #include <linux/sched.h> | |
c1dc0b9c | 38 | #include <linux/nmi.h> |
8826f3b0 | 39 | #include <linux/atomic.h> |
64db4cff | 40 | #include <linux/bitops.h> |
9984de1a | 41 | #include <linux/export.h> |
64db4cff PM |
42 | #include <linux/completion.h> |
43 | #include <linux/moduleparam.h> | |
44 | #include <linux/percpu.h> | |
45 | #include <linux/notifier.h> | |
46 | #include <linux/cpu.h> | |
47 | #include <linux/mutex.h> | |
48 | #include <linux/time.h> | |
bbad9379 | 49 | #include <linux/kernel_stat.h> |
a26ac245 PM |
50 | #include <linux/wait.h> |
51 | #include <linux/kthread.h> | |
268bb0ce | 52 | #include <linux/prefetch.h> |
3d3b7db0 PM |
53 | #include <linux/delay.h> |
54 | #include <linux/stop_machine.h> | |
661a85dc | 55 | #include <linux/random.h> |
f7f7bac9 | 56 | #include <linux/ftrace_event.h> |
d1d74d14 | 57 | #include <linux/suspend.h> |
64db4cff | 58 | |
9f77da9f | 59 | #include "rcutree.h" |
29c00b4a PM |
60 | #include <trace/events/rcu.h> |
61 | ||
62 | #include "rcu.h" | |
9f77da9f | 63 | |
64db4cff PM |
64 | /* Data structures. */ |
65 | ||
f885b7f2 | 66 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
394f2769 | 67 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; |
88b91c7c | 68 | |
f7f7bac9 SRRH |
69 | /* |
70 | * In order to export the rcu_state name to the tracing tools, it | |
71 | * needs to be added in the __tracepoint_string section. | |
72 | * This requires defining a separate variable tp_<sname>_varname | |
73 | * that points to the string being used, and this will allow | |
74 | * the tracing userspace tools to be able to decipher the string | |
75 | * address to the matching string. | |
76 | */ | |
a41bfeb2 | 77 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ |
f7f7bac9 SRRH |
78 | static char sname##_varname[] = #sname; \ |
79 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \ | |
a41bfeb2 | 80 | struct rcu_state sname##_state = { \ |
6c90cc7b | 81 | .level = { &sname##_state.node[0] }, \ |
037b64ed | 82 | .call = cr, \ |
af446b70 | 83 | .fqs_state = RCU_GP_IDLE, \ |
42c3533e PM |
84 | .gpnum = 0UL - 300UL, \ |
85 | .completed = 0UL - 300UL, \ | |
7b2e6011 | 86 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
6c90cc7b PM |
87 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
88 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 89 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
a4fbe35a | 90 | .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \ |
f7f7bac9 | 91 | .name = sname##_varname, \ |
a4889858 | 92 | .abbr = sabbr, \ |
a41bfeb2 SRRH |
93 | }; \ |
94 | DEFINE_PER_CPU(struct rcu_data, sname##_data) | |
64db4cff | 95 | |
a41bfeb2 SRRH |
96 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); |
97 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); | |
b1f77b05 | 98 | |
27f4d280 | 99 | static struct rcu_state *rcu_state; |
6ce75a23 | 100 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 101 | |
f885b7f2 PM |
102 | /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */ |
103 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF; | |
7e5c2dfb | 104 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 PM |
105 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
106 | static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */ | |
107 | NUM_RCU_LVL_0, | |
108 | NUM_RCU_LVL_1, | |
109 | NUM_RCU_LVL_2, | |
110 | NUM_RCU_LVL_3, | |
111 | NUM_RCU_LVL_4, | |
112 | }; | |
113 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ | |
114 | ||
b0d30417 PM |
115 | /* |
116 | * The rcu_scheduler_active variable transitions from zero to one just | |
117 | * before the first task is spawned. So when this variable is zero, RCU | |
118 | * can assume that there is but one task, allowing RCU to (for example) | |
b44f6656 | 119 | * optimize synchronize_sched() to a simple barrier(). When this variable |
b0d30417 PM |
120 | * is one, RCU must actually do all the hard work required to detect real |
121 | * grace periods. This variable is also used to suppress boot-time false | |
122 | * positives from lockdep-RCU error checking. | |
123 | */ | |
bbad9379 PM |
124 | int rcu_scheduler_active __read_mostly; |
125 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
126 | ||
b0d30417 PM |
127 | /* |
128 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
129 | * during the early_initcall() processing, which is after the scheduler | |
130 | * is capable of creating new tasks. So RCU processing (for example, | |
131 | * creating tasks for RCU priority boosting) must be delayed until after | |
132 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
133 | * currently delay invocation of any RCU callbacks until after this point. | |
134 | * | |
135 | * It might later prove better for people registering RCU callbacks during | |
136 | * early boot to take responsibility for these callbacks, but one step at | |
137 | * a time. | |
138 | */ | |
139 | static int rcu_scheduler_fully_active __read_mostly; | |
140 | ||
a46e0899 PM |
141 | #ifdef CONFIG_RCU_BOOST |
142 | ||
a26ac245 PM |
143 | /* |
144 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
145 | * handle all flavors of RCU. | |
146 | */ | |
147 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
d71df90e | 148 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); |
5ece5bab | 149 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); |
d71df90e | 150 | DEFINE_PER_CPU(char, rcu_cpu_has_work); |
a26ac245 | 151 | |
a46e0899 PM |
152 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
153 | ||
5d01bbd1 | 154 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
155 | static void invoke_rcu_core(void); |
156 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
a26ac245 | 157 | |
4a298656 PM |
158 | /* |
159 | * Track the rcutorture test sequence number and the update version | |
160 | * number within a given test. The rcutorture_testseq is incremented | |
161 | * on every rcutorture module load and unload, so has an odd value | |
162 | * when a test is running. The rcutorture_vernum is set to zero | |
163 | * when rcutorture starts and is incremented on each rcutorture update. | |
164 | * These variables enable correlating rcutorture output with the | |
165 | * RCU tracing information. | |
166 | */ | |
167 | unsigned long rcutorture_testseq; | |
168 | unsigned long rcutorture_vernum; | |
169 | ||
fc2219d4 PM |
170 | /* |
171 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s | |
172 | * permit this function to be invoked without holding the root rcu_node | |
173 | * structure's ->lock, but of course results can be subject to change. | |
174 | */ | |
175 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
176 | { | |
177 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); | |
178 | } | |
179 | ||
b1f77b05 | 180 | /* |
d6714c22 | 181 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 182 | * how many quiescent states passed, just if there was at least |
d6714c22 | 183 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 184 | * The caller must have disabled preemption. |
b1f77b05 | 185 | */ |
d6714c22 | 186 | void rcu_sched_qs(int cpu) |
b1f77b05 | 187 | { |
25502a6c | 188 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu); |
f41d911f | 189 | |
e4cc1f22 | 190 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 191 | trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 192 | rdp->passed_quiesce = 1; |
b1f77b05 IM |
193 | } |
194 | ||
d6714c22 | 195 | void rcu_bh_qs(int cpu) |
b1f77b05 | 196 | { |
25502a6c | 197 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu); |
f41d911f | 198 | |
e4cc1f22 | 199 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 200 | trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 201 | rdp->passed_quiesce = 1; |
b1f77b05 | 202 | } |
64db4cff | 203 | |
25502a6c PM |
204 | /* |
205 | * Note a context switch. This is a quiescent state for RCU-sched, | |
206 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 207 | * The caller must have disabled preemption. |
25502a6c PM |
208 | */ |
209 | void rcu_note_context_switch(int cpu) | |
210 | { | |
f7f7bac9 | 211 | trace_rcu_utilization(TPS("Start context switch")); |
25502a6c | 212 | rcu_sched_qs(cpu); |
cba6d0d6 | 213 | rcu_preempt_note_context_switch(cpu); |
f7f7bac9 | 214 | trace_rcu_utilization(TPS("End context switch")); |
25502a6c | 215 | } |
29ce8310 | 216 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 217 | |
01896f7e | 218 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { |
29e37d81 | 219 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, |
23b5c8fa | 220 | .dynticks = ATOMIC_INIT(1), |
2333210b PM |
221 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
222 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
223 | .dynticks_idle = ATOMIC_INIT(1), | |
224 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
90a4d2c0 | 225 | }; |
64db4cff | 226 | |
878d7439 ED |
227 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
228 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
229 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 230 | |
878d7439 ED |
231 | module_param(blimit, long, 0444); |
232 | module_param(qhimark, long, 0444); | |
233 | module_param(qlowmark, long, 0444); | |
3d76c082 | 234 | |
026ad283 PM |
235 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
236 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
237 | |
238 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
239 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
240 | ||
910ee45d PM |
241 | static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
242 | struct rcu_data *rdp); | |
217af2a2 PM |
243 | static void force_qs_rnp(struct rcu_state *rsp, |
244 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
245 | unsigned long *maxj), | |
246 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 247 | static void force_quiescent_state(struct rcu_state *rsp); |
a157229c | 248 | static int rcu_pending(int cpu); |
64db4cff PM |
249 | |
250 | /* | |
d6714c22 | 251 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
64db4cff | 252 | */ |
d6714c22 | 253 | long rcu_batches_completed_sched(void) |
64db4cff | 254 | { |
d6714c22 | 255 | return rcu_sched_state.completed; |
64db4cff | 256 | } |
d6714c22 | 257 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
258 | |
259 | /* | |
260 | * Return the number of RCU BH batches processed thus far for debug & stats. | |
261 | */ | |
262 | long rcu_batches_completed_bh(void) | |
263 | { | |
264 | return rcu_bh_state.completed; | |
265 | } | |
266 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
267 | ||
bf66f18e PM |
268 | /* |
269 | * Force a quiescent state for RCU BH. | |
270 | */ | |
271 | void rcu_bh_force_quiescent_state(void) | |
272 | { | |
4cdfc175 | 273 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
274 | } |
275 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
276 | ||
4a298656 PM |
277 | /* |
278 | * Record the number of times rcutorture tests have been initiated and | |
279 | * terminated. This information allows the debugfs tracing stats to be | |
280 | * correlated to the rcutorture messages, even when the rcutorture module | |
281 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
282 | * store this state in rcutorture itself. | |
283 | */ | |
284 | void rcutorture_record_test_transition(void) | |
285 | { | |
286 | rcutorture_testseq++; | |
287 | rcutorture_vernum = 0; | |
288 | } | |
289 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
290 | ||
291 | /* | |
292 | * Record the number of writer passes through the current rcutorture test. | |
293 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
294 | * messages. | |
295 | */ | |
296 | void rcutorture_record_progress(unsigned long vernum) | |
297 | { | |
298 | rcutorture_vernum++; | |
299 | } | |
300 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
301 | ||
bf66f18e PM |
302 | /* |
303 | * Force a quiescent state for RCU-sched. | |
304 | */ | |
305 | void rcu_sched_force_quiescent_state(void) | |
306 | { | |
4cdfc175 | 307 | force_quiescent_state(&rcu_sched_state); |
bf66f18e PM |
308 | } |
309 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
310 | ||
64db4cff PM |
311 | /* |
312 | * Does the CPU have callbacks ready to be invoked? | |
313 | */ | |
314 | static int | |
315 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
316 | { | |
3fbfbf7a PM |
317 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
318 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
319 | } |
320 | ||
321 | /* | |
dc35c893 PM |
322 | * Does the current CPU require a not-yet-started grace period? |
323 | * The caller must have disabled interrupts to prevent races with | |
324 | * normal callback registry. | |
64db4cff PM |
325 | */ |
326 | static int | |
327 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
328 | { | |
dc35c893 | 329 | int i; |
3fbfbf7a | 330 | |
dc35c893 PM |
331 | if (rcu_gp_in_progress(rsp)) |
332 | return 0; /* No, a grace period is already in progress. */ | |
dae6e64d | 333 | if (rcu_nocb_needs_gp(rsp)) |
34ed6246 | 334 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
335 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
336 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
337 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
338 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
339 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
340 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
341 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), | |
342 | rdp->nxtcompleted[i])) | |
343 | return 1; /* Yes, CBs for future grace period. */ | |
344 | return 0; /* No grace period needed. */ | |
64db4cff PM |
345 | } |
346 | ||
347 | /* | |
348 | * Return the root node of the specified rcu_state structure. | |
349 | */ | |
350 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
351 | { | |
352 | return &rsp->node[0]; | |
353 | } | |
354 | ||
9b2e4f18 | 355 | /* |
adf5091e | 356 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
357 | * |
358 | * If the new value of the ->dynticks_nesting counter now is zero, | |
359 | * we really have entered idle, and must do the appropriate accounting. | |
360 | * The caller must have disabled interrupts. | |
361 | */ | |
adf5091e FW |
362 | static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval, |
363 | bool user) | |
9b2e4f18 | 364 | { |
f7f7bac9 | 365 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 366 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
367 | struct task_struct *idle __maybe_unused = |
368 | idle_task(smp_processor_id()); | |
0989cb46 | 369 | |
f7f7bac9 | 370 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 371 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
372 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
373 | current->pid, current->comm, | |
374 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 375 | } |
aea1b35e | 376 | rcu_prepare_for_idle(smp_processor_id()); |
9b2e4f18 PM |
377 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
378 | smp_mb__before_atomic_inc(); /* See above. */ | |
379 | atomic_inc(&rdtp->dynticks); | |
380 | smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */ | |
381 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
c44e2cdd PM |
382 | |
383 | /* | |
adf5091e | 384 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
385 | * in an RCU read-side critical section. |
386 | */ | |
387 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), | |
388 | "Illegal idle entry in RCU read-side critical section."); | |
389 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), | |
390 | "Illegal idle entry in RCU-bh read-side critical section."); | |
391 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), | |
392 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 393 | } |
64db4cff | 394 | |
adf5091e FW |
395 | /* |
396 | * Enter an RCU extended quiescent state, which can be either the | |
397 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 398 | */ |
adf5091e | 399 | static void rcu_eqs_enter(bool user) |
64db4cff | 400 | { |
4145fa7f | 401 | long long oldval; |
64db4cff PM |
402 | struct rcu_dynticks *rdtp; |
403 | ||
c9d4b0af | 404 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 405 | oldval = rdtp->dynticks_nesting; |
29e37d81 PM |
406 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
407 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) | |
408 | rdtp->dynticks_nesting = 0; | |
409 | else | |
410 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; | |
adf5091e | 411 | rcu_eqs_enter_common(rdtp, oldval, user); |
64db4cff | 412 | } |
adf5091e FW |
413 | |
414 | /** | |
415 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
416 | * | |
417 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
418 | * read-side critical sections can occur. (Though RCU read-side | |
419 | * critical sections can occur in irq handlers in idle, a possibility | |
420 | * handled by irq_enter() and irq_exit().) | |
421 | * | |
422 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
423 | * the possibility of usermode upcalls having messed up our count | |
424 | * of interrupt nesting level during the prior busy period. | |
425 | */ | |
426 | void rcu_idle_enter(void) | |
427 | { | |
c5d900bf FW |
428 | unsigned long flags; |
429 | ||
430 | local_irq_save(flags); | |
cb349ca9 | 431 | rcu_eqs_enter(false); |
c9d4b0af | 432 | rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0); |
c5d900bf | 433 | local_irq_restore(flags); |
adf5091e | 434 | } |
8a2ecf47 | 435 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 436 | |
2b1d5024 | 437 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
438 | /** |
439 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
440 | * | |
441 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
442 | * is permitted between this call and rcu_user_exit(). This way the | |
443 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
444 | * when the CPU runs in userspace. | |
445 | */ | |
446 | void rcu_user_enter(void) | |
447 | { | |
91d1aa43 | 448 | rcu_eqs_enter(1); |
adf5091e | 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); | |
c9d4b0af | 475 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 476 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 PM |
477 | rdtp->dynticks_nesting--; |
478 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); | |
b6fc6020 | 479 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 480 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 481 | else |
cb349ca9 | 482 | rcu_eqs_enter_common(rdtp, oldval, true); |
eb348b89 | 483 | rcu_sysidle_enter(rdtp, 1); |
9b2e4f18 PM |
484 | local_irq_restore(flags); |
485 | } | |
486 | ||
487 | /* | |
adf5091e | 488 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
489 | * |
490 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
491 | * we really have exited idle, and must do the appropriate accounting. | |
492 | * The caller must have disabled interrupts. | |
493 | */ | |
adf5091e FW |
494 | static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval, |
495 | int user) | |
9b2e4f18 | 496 | { |
23b5c8fa PM |
497 | smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */ |
498 | atomic_inc(&rdtp->dynticks); | |
499 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
500 | smp_mb__after_atomic_inc(); /* See above. */ | |
501 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
7cb92499 | 502 | rcu_cleanup_after_idle(smp_processor_id()); |
f7f7bac9 | 503 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
cb349ca9 | 504 | if (!user && !is_idle_task(current)) { |
289828e6 PM |
505 | struct task_struct *idle __maybe_unused = |
506 | idle_task(smp_processor_id()); | |
0989cb46 | 507 | |
f7f7bac9 | 508 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 509 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 510 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
511 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
512 | current->pid, current->comm, | |
513 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
514 | } |
515 | } | |
516 | ||
adf5091e FW |
517 | /* |
518 | * Exit an RCU extended quiescent state, which can be either the | |
519 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 520 | */ |
adf5091e | 521 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 522 | { |
9b2e4f18 PM |
523 | struct rcu_dynticks *rdtp; |
524 | long long oldval; | |
525 | ||
c9d4b0af | 526 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 | 527 | oldval = rdtp->dynticks_nesting; |
29e37d81 PM |
528 | WARN_ON_ONCE(oldval < 0); |
529 | if (oldval & DYNTICK_TASK_NEST_MASK) | |
530 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; | |
531 | else | |
532 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; | |
adf5091e | 533 | rcu_eqs_exit_common(rdtp, oldval, user); |
9b2e4f18 | 534 | } |
adf5091e FW |
535 | |
536 | /** | |
537 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
538 | * | |
539 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
540 | * read-side critical sections can occur. | |
541 | * | |
542 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
543 | * allow for the possibility of usermode upcalls messing up our count | |
544 | * of interrupt nesting level during the busy period that is just | |
545 | * now starting. | |
546 | */ | |
547 | void rcu_idle_exit(void) | |
548 | { | |
c5d900bf FW |
549 | unsigned long flags; |
550 | ||
551 | local_irq_save(flags); | |
cb349ca9 | 552 | rcu_eqs_exit(false); |
c9d4b0af | 553 | rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0); |
c5d900bf | 554 | local_irq_restore(flags); |
adf5091e | 555 | } |
8a2ecf47 | 556 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 557 | |
2b1d5024 | 558 | #ifdef CONFIG_RCU_USER_QS |
adf5091e FW |
559 | /** |
560 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
561 | * | |
562 | * Exit RCU idle mode while entering the kernel because it can | |
563 | * run a RCU read side critical section anytime. | |
564 | */ | |
565 | void rcu_user_exit(void) | |
566 | { | |
91d1aa43 | 567 | rcu_eqs_exit(1); |
adf5091e | 568 | } |
2b1d5024 | 569 | #endif /* CONFIG_RCU_USER_QS */ |
19dd1591 | 570 | |
9b2e4f18 PM |
571 | /** |
572 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
573 | * | |
574 | * Enter an interrupt handler, which might possibly result in exiting | |
575 | * idle mode, in other words, entering the mode in which read-side critical | |
576 | * sections can occur. | |
577 | * | |
578 | * Note that the Linux kernel is fully capable of entering an interrupt | |
579 | * handler that it never exits, for example when doing upcalls to | |
580 | * user mode! This code assumes that the idle loop never does upcalls to | |
581 | * user mode. If your architecture does do upcalls from the idle loop (or | |
582 | * does anything else that results in unbalanced calls to the irq_enter() | |
583 | * and irq_exit() functions), RCU will give you what you deserve, good | |
584 | * and hard. But very infrequently and irreproducibly. | |
585 | * | |
586 | * Use things like work queues to work around this limitation. | |
587 | * | |
588 | * You have been warned. | |
589 | */ | |
590 | void rcu_irq_enter(void) | |
591 | { | |
592 | unsigned long flags; | |
593 | struct rcu_dynticks *rdtp; | |
594 | long long oldval; | |
595 | ||
596 | local_irq_save(flags); | |
c9d4b0af | 597 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 PM |
598 | oldval = rdtp->dynticks_nesting; |
599 | rdtp->dynticks_nesting++; | |
600 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); | |
b6fc6020 | 601 | if (oldval) |
f7f7bac9 | 602 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 603 | else |
cb349ca9 | 604 | rcu_eqs_exit_common(rdtp, oldval, true); |
eb348b89 | 605 | rcu_sysidle_exit(rdtp, 1); |
64db4cff | 606 | local_irq_restore(flags); |
64db4cff PM |
607 | } |
608 | ||
609 | /** | |
610 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
611 | * | |
612 | * If the CPU was idle with dynamic ticks active, and there is no | |
613 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
614 | * RCU grace-period handling know that the CPU is active. | |
615 | */ | |
616 | void rcu_nmi_enter(void) | |
617 | { | |
c9d4b0af | 618 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 619 | |
23b5c8fa PM |
620 | if (rdtp->dynticks_nmi_nesting == 0 && |
621 | (atomic_read(&rdtp->dynticks) & 0x1)) | |
64db4cff | 622 | return; |
23b5c8fa PM |
623 | rdtp->dynticks_nmi_nesting++; |
624 | smp_mb__before_atomic_inc(); /* Force delay from prior write. */ | |
625 | atomic_inc(&rdtp->dynticks); | |
626 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
627 | smp_mb__after_atomic_inc(); /* See above. */ | |
628 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
64db4cff PM |
629 | } |
630 | ||
631 | /** | |
632 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
633 | * | |
634 | * If the CPU was idle with dynamic ticks active, and there is no | |
635 | * irq handler running, this updates rdtp->dynticks_nmi to let the | |
636 | * RCU grace-period handling know that the CPU is no longer active. | |
637 | */ | |
638 | void rcu_nmi_exit(void) | |
639 | { | |
c9d4b0af | 640 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 641 | |
23b5c8fa PM |
642 | if (rdtp->dynticks_nmi_nesting == 0 || |
643 | --rdtp->dynticks_nmi_nesting != 0) | |
64db4cff | 644 | return; |
23b5c8fa PM |
645 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
646 | smp_mb__before_atomic_inc(); /* See above. */ | |
647 | atomic_inc(&rdtp->dynticks); | |
648 | smp_mb__after_atomic_inc(); /* Force delay to next write. */ | |
649 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); | |
64db4cff PM |
650 | } |
651 | ||
652 | /** | |
5c173eb8 PM |
653 | * __rcu_is_watching - are RCU read-side critical sections safe? |
654 | * | |
655 | * Return true if RCU is watching the running CPU, which means that | |
656 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
657 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
658 | * least disabled preemption. | |
659 | */ | |
660 | bool __rcu_is_watching(void) | |
661 | { | |
662 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
663 | } | |
664 | ||
665 | /** | |
666 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
64db4cff | 667 | * |
9b2e4f18 | 668 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 669 | * or NMI handler, return true. |
64db4cff | 670 | */ |
5c173eb8 | 671 | bool rcu_is_watching(void) |
64db4cff | 672 | { |
34240697 PM |
673 | int ret; |
674 | ||
675 | preempt_disable(); | |
5c173eb8 | 676 | ret = __rcu_is_watching(); |
34240697 PM |
677 | preempt_enable(); |
678 | return ret; | |
64db4cff | 679 | } |
5c173eb8 | 680 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 681 | |
62fde6ed | 682 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
683 | |
684 | /* | |
685 | * Is the current CPU online? Disable preemption to avoid false positives | |
686 | * that could otherwise happen due to the current CPU number being sampled, | |
687 | * this task being preempted, its old CPU being taken offline, resuming | |
688 | * on some other CPU, then determining that its old CPU is now offline. | |
689 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
690 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
691 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
692 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
693 | * offline to continue to use RCU for one jiffy after marking itself | |
694 | * offline in the cpu_online_mask. This leniency is necessary given the | |
695 | * non-atomic nature of the online and offline processing, for example, | |
696 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
697 | * notifiers. | |
698 | * | |
699 | * This is also why RCU internally marks CPUs online during the | |
700 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
701 | * |
702 | * Disable checking if in an NMI handler because we cannot safely report | |
703 | * errors from NMI handlers anyway. | |
704 | */ | |
705 | bool rcu_lockdep_current_cpu_online(void) | |
706 | { | |
2036d94a PM |
707 | struct rcu_data *rdp; |
708 | struct rcu_node *rnp; | |
c0d6d01b PM |
709 | bool ret; |
710 | ||
711 | if (in_nmi()) | |
712 | return 1; | |
713 | preempt_disable(); | |
c9d4b0af | 714 | rdp = this_cpu_ptr(&rcu_sched_data); |
2036d94a PM |
715 | rnp = rdp->mynode; |
716 | ret = (rdp->grpmask & rnp->qsmaskinit) || | |
c0d6d01b PM |
717 | !rcu_scheduler_fully_active; |
718 | preempt_enable(); | |
719 | return ret; | |
720 | } | |
721 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
722 | ||
62fde6ed | 723 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 724 | |
64db4cff | 725 | /** |
9b2e4f18 | 726 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 727 | * |
9b2e4f18 PM |
728 | * If the current CPU is idle or running at a first-level (not nested) |
729 | * interrupt from idle, return true. The caller must have at least | |
730 | * disabled preemption. | |
64db4cff | 731 | */ |
62e3cb14 | 732 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 733 | { |
c9d4b0af | 734 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
64db4cff PM |
735 | } |
736 | ||
64db4cff PM |
737 | /* |
738 | * Snapshot the specified CPU's dynticks counter so that we can later | |
739 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 740 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 741 | */ |
217af2a2 PM |
742 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
743 | bool *isidle, unsigned long *maxj) | |
64db4cff | 744 | { |
23b5c8fa | 745 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 746 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
f0e7c19d | 747 | return (rdp->dynticks_snap & 0x1) == 0; |
64db4cff PM |
748 | } |
749 | ||
750 | /* | |
751 | * Return true if the specified CPU has passed through a quiescent | |
752 | * state by virtue of being in or having passed through an dynticks | |
753 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 754 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 755 | */ |
217af2a2 PM |
756 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
757 | bool *isidle, unsigned long *maxj) | |
64db4cff | 758 | { |
7eb4f455 PM |
759 | unsigned int curr; |
760 | unsigned int snap; | |
64db4cff | 761 | |
7eb4f455 PM |
762 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
763 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
764 | |
765 | /* | |
766 | * If the CPU passed through or entered a dynticks idle phase with | |
767 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
768 | * already acknowledged the request to pass through a quiescent | |
769 | * state. Either way, that CPU cannot possibly be in an RCU | |
770 | * read-side critical section that started before the beginning | |
771 | * of the current RCU grace period. | |
772 | */ | |
7eb4f455 | 773 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 774 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
775 | rdp->dynticks_fqs++; |
776 | return 1; | |
777 | } | |
778 | ||
a82dcc76 PM |
779 | /* |
780 | * Check for the CPU being offline, but only if the grace period | |
781 | * is old enough. We don't need to worry about the CPU changing | |
782 | * state: If we see it offline even once, it has been through a | |
783 | * quiescent state. | |
784 | * | |
785 | * The reason for insisting that the grace period be at least | |
786 | * one jiffy old is that CPUs that are not quite online and that | |
787 | * have just gone offline can still execute RCU read-side critical | |
788 | * sections. | |
789 | */ | |
790 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
791 | return 0; /* Grace period is not old enough. */ | |
792 | barrier(); | |
793 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 794 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
795 | rdp->offline_fqs++; |
796 | return 1; | |
797 | } | |
65d798f0 PM |
798 | |
799 | /* | |
800 | * There is a possibility that a CPU in adaptive-ticks state | |
801 | * might run in the kernel with the scheduling-clock tick disabled | |
802 | * for an extended time period. Invoke rcu_kick_nohz_cpu() to | |
803 | * force the CPU to restart the scheduling-clock tick in this | |
804 | * CPU is in this state. | |
805 | */ | |
806 | rcu_kick_nohz_cpu(rdp->cpu); | |
807 | ||
a82dcc76 | 808 | return 0; |
64db4cff PM |
809 | } |
810 | ||
64db4cff PM |
811 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
812 | { | |
26cdfedf PM |
813 | unsigned long j = ACCESS_ONCE(jiffies); |
814 | ||
815 | rsp->gp_start = j; | |
816 | smp_wmb(); /* Record start time before stall time. */ | |
817 | rsp->jiffies_stall = j + rcu_jiffies_till_stall_check(); | |
64db4cff PM |
818 | } |
819 | ||
b637a328 PM |
820 | /* |
821 | * Dump stacks of all tasks running on stalled CPUs. This is a fallback | |
822 | * for architectures that do not implement trigger_all_cpu_backtrace(). | |
823 | * The NMI-triggered stack traces are more accurate because they are | |
824 | * printed by the target CPU. | |
825 | */ | |
826 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
827 | { | |
828 | int cpu; | |
829 | unsigned long flags; | |
830 | struct rcu_node *rnp; | |
831 | ||
832 | rcu_for_each_leaf_node(rsp, rnp) { | |
833 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
834 | if (rnp->qsmask != 0) { | |
835 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
836 | if (rnp->qsmask & (1UL << cpu)) | |
837 | dump_cpu_task(rnp->grplo + cpu); | |
838 | } | |
839 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
840 | } | |
841 | } | |
842 | ||
64db4cff PM |
843 | static void print_other_cpu_stall(struct rcu_state *rsp) |
844 | { | |
845 | int cpu; | |
846 | long delta; | |
847 | unsigned long flags; | |
285fe294 | 848 | int ndetected = 0; |
64db4cff | 849 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 850 | long totqlen = 0; |
64db4cff PM |
851 | |
852 | /* Only let one CPU complain about others per time interval. */ | |
853 | ||
1304afb2 | 854 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 855 | delta = jiffies - rsp->jiffies_stall; |
fc2219d4 | 856 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 857 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
858 | return; |
859 | } | |
6bfc09e2 | 860 | rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 861 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 862 | |
8cdd32a9 PM |
863 | /* |
864 | * OK, time to rat on our buddy... | |
865 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
866 | * RCU CPU stall warnings. | |
867 | */ | |
d7f3e207 | 868 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 869 | rsp->name); |
a858af28 | 870 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 871 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 872 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 873 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
874 | if (rnp->qsmask != 0) { |
875 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
876 | if (rnp->qsmask & (1UL << cpu)) { | |
877 | print_cpu_stall_info(rsp, | |
878 | rnp->grplo + cpu); | |
879 | ndetected++; | |
880 | } | |
881 | } | |
3acd9eb3 | 882 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 883 | } |
a858af28 PM |
884 | |
885 | /* | |
886 | * Now rat on any tasks that got kicked up to the root rcu_node | |
887 | * due to CPU offlining. | |
888 | */ | |
889 | rnp = rcu_get_root(rsp); | |
890 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
285fe294 | 891 | ndetected += rcu_print_task_stall(rnp); |
a858af28 PM |
892 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
893 | ||
894 | print_cpu_stall_info_end(); | |
53bb857c PM |
895 | for_each_possible_cpu(cpu) |
896 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
897 | pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n", | |
eee05882 | 898 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
53bb857c | 899 | rsp->gpnum, rsp->completed, totqlen); |
9bc8b558 | 900 | if (ndetected == 0) |
d7f3e207 | 901 | pr_err("INFO: Stall ended before state dump start\n"); |
9bc8b558 | 902 | else if (!trigger_all_cpu_backtrace()) |
b637a328 | 903 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 904 | |
4cdfc175 | 905 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
906 | |
907 | rcu_print_detail_task_stall(rsp); | |
908 | ||
4cdfc175 | 909 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
910 | } |
911 | ||
912 | static void print_cpu_stall(struct rcu_state *rsp) | |
913 | { | |
53bb857c | 914 | int cpu; |
64db4cff PM |
915 | unsigned long flags; |
916 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 917 | long totqlen = 0; |
64db4cff | 918 | |
8cdd32a9 PM |
919 | /* |
920 | * OK, time to rat on ourselves... | |
921 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
922 | * RCU CPU stall warnings. | |
923 | */ | |
d7f3e207 | 924 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
925 | print_cpu_stall_info_begin(); |
926 | print_cpu_stall_info(rsp, smp_processor_id()); | |
927 | print_cpu_stall_info_end(); | |
53bb857c PM |
928 | for_each_possible_cpu(cpu) |
929 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
930 | pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n", | |
931 | jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen); | |
4627e240 PM |
932 | if (!trigger_all_cpu_backtrace()) |
933 | dump_stack(); | |
c1dc0b9c | 934 | |
1304afb2 | 935 | raw_spin_lock_irqsave(&rnp->lock, flags); |
20133cfc | 936 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall)) |
13cfcca0 | 937 | rsp->jiffies_stall = jiffies + |
6bfc09e2 | 938 | 3 * rcu_jiffies_till_stall_check() + 3; |
1304afb2 | 939 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 940 | |
64db4cff PM |
941 | set_need_resched(); /* kick ourselves to get things going. */ |
942 | } | |
943 | ||
944 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
945 | { | |
26cdfedf PM |
946 | unsigned long completed; |
947 | unsigned long gpnum; | |
948 | unsigned long gps; | |
bad6e139 PM |
949 | unsigned long j; |
950 | unsigned long js; | |
64db4cff PM |
951 | struct rcu_node *rnp; |
952 | ||
26cdfedf | 953 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
c68de209 | 954 | return; |
bad6e139 | 955 | j = ACCESS_ONCE(jiffies); |
26cdfedf PM |
956 | |
957 | /* | |
958 | * Lots of memory barriers to reject false positives. | |
959 | * | |
960 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
961 | * then rsp->gp_start, and finally rsp->completed. These values | |
962 | * are updated in the opposite order with memory barriers (or | |
963 | * equivalent) during grace-period initialization and cleanup. | |
964 | * Now, a false positive can occur if we get an new value of | |
965 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
966 | * the memory barriers, the only way that this can happen is if one | |
967 | * grace period ends and another starts between these two fetches. | |
968 | * Detect this by comparing rsp->completed with the previous fetch | |
969 | * from rsp->gpnum. | |
970 | * | |
971 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
972 | * and rsp->gp_start suffice to forestall false positives. | |
973 | */ | |
974 | gpnum = ACCESS_ONCE(rsp->gpnum); | |
975 | smp_rmb(); /* Pick up ->gpnum first... */ | |
bad6e139 | 976 | js = ACCESS_ONCE(rsp->jiffies_stall); |
26cdfedf PM |
977 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
978 | gps = ACCESS_ONCE(rsp->gp_start); | |
979 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ | |
980 | completed = ACCESS_ONCE(rsp->completed); | |
981 | if (ULONG_CMP_GE(completed, gpnum) || | |
982 | ULONG_CMP_LT(j, js) || | |
983 | ULONG_CMP_GE(gps, js)) | |
984 | return; /* No stall or GP completed since entering function. */ | |
64db4cff | 985 | rnp = rdp->mynode; |
c96ea7cf | 986 | if (rcu_gp_in_progress(rsp) && |
26cdfedf | 987 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) { |
64db4cff PM |
988 | |
989 | /* We haven't checked in, so go dump stack. */ | |
990 | print_cpu_stall(rsp); | |
991 | ||
bad6e139 PM |
992 | } else if (rcu_gp_in_progress(rsp) && |
993 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 994 | |
bad6e139 | 995 | /* They had a few time units to dump stack, so complain. */ |
64db4cff PM |
996 | print_other_cpu_stall(rsp); |
997 | } | |
998 | } | |
999 | ||
53d84e00 PM |
1000 | /** |
1001 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1002 | * | |
1003 | * Set the stall-warning timeout way off into the future, thus preventing | |
1004 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1005 | * RCU grace periods. | |
1006 | * | |
1007 | * The caller must disable hard irqs. | |
1008 | */ | |
1009 | void rcu_cpu_stall_reset(void) | |
1010 | { | |
6ce75a23 PM |
1011 | struct rcu_state *rsp; |
1012 | ||
1013 | for_each_rcu_flavor(rsp) | |
1014 | rsp->jiffies_stall = jiffies + ULONG_MAX / 2; | |
53d84e00 PM |
1015 | } |
1016 | ||
3f5d3ea6 PM |
1017 | /* |
1018 | * Initialize the specified rcu_data structure's callback list to empty. | |
1019 | */ | |
1020 | static void init_callback_list(struct rcu_data *rdp) | |
1021 | { | |
1022 | int i; | |
1023 | ||
34ed6246 PM |
1024 | if (init_nocb_callback_list(rdp)) |
1025 | return; | |
3f5d3ea6 PM |
1026 | rdp->nxtlist = NULL; |
1027 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1028 | rdp->nxttail[i] = &rdp->nxtlist; | |
1029 | } | |
1030 | ||
dc35c893 PM |
1031 | /* |
1032 | * Determine the value that ->completed will have at the end of the | |
1033 | * next subsequent grace period. This is used to tag callbacks so that | |
1034 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1035 | * been dyntick-idle for an extended period with callbacks under the | |
1036 | * influence of RCU_FAST_NO_HZ. | |
1037 | * | |
1038 | * The caller must hold rnp->lock with interrupts disabled. | |
1039 | */ | |
1040 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1041 | struct rcu_node *rnp) | |
1042 | { | |
1043 | /* | |
1044 | * If RCU is idle, we just wait for the next grace period. | |
1045 | * But we can only be sure that RCU is idle if we are looking | |
1046 | * at the root rcu_node structure -- otherwise, a new grace | |
1047 | * period might have started, but just not yet gotten around | |
1048 | * to initializing the current non-root rcu_node structure. | |
1049 | */ | |
1050 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1051 | return rnp->completed + 1; | |
1052 | ||
1053 | /* | |
1054 | * Otherwise, wait for a possible partial grace period and | |
1055 | * then the subsequent full grace period. | |
1056 | */ | |
1057 | return rnp->completed + 2; | |
1058 | } | |
1059 | ||
0446be48 PM |
1060 | /* |
1061 | * Trace-event helper function for rcu_start_future_gp() and | |
1062 | * rcu_nocb_wait_gp(). | |
1063 | */ | |
1064 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1065 | unsigned long c, const char *s) |
0446be48 PM |
1066 | { |
1067 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1068 | rnp->completed, c, rnp->level, | |
1069 | rnp->grplo, rnp->grphi, s); | |
1070 | } | |
1071 | ||
1072 | /* | |
1073 | * Start some future grace period, as needed to handle newly arrived | |
1074 | * callbacks. The required future grace periods are recorded in each | |
1075 | * rcu_node structure's ->need_future_gp field. | |
1076 | * | |
1077 | * The caller must hold the specified rcu_node structure's ->lock. | |
1078 | */ | |
1079 | static unsigned long __maybe_unused | |
1080 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp) | |
1081 | { | |
1082 | unsigned long c; | |
1083 | int i; | |
1084 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); | |
1085 | ||
1086 | /* | |
1087 | * Pick up grace-period number for new callbacks. If this | |
1088 | * grace period is already marked as needed, return to the caller. | |
1089 | */ | |
1090 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1091 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1092 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1093 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
0446be48 PM |
1094 | return c; |
1095 | } | |
1096 | ||
1097 | /* | |
1098 | * If either this rcu_node structure or the root rcu_node structure | |
1099 | * believe that a grace period is in progress, then we must wait | |
1100 | * for the one following, which is in "c". Because our request | |
1101 | * will be noticed at the end of the current grace period, we don't | |
1102 | * need to explicitly start one. | |
1103 | */ | |
1104 | if (rnp->gpnum != rnp->completed || | |
1105 | ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) { | |
1106 | rnp->need_future_gp[c & 0x1]++; | |
f7f7bac9 | 1107 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
0446be48 PM |
1108 | return c; |
1109 | } | |
1110 | ||
1111 | /* | |
1112 | * There might be no grace period in progress. If we don't already | |
1113 | * hold it, acquire the root rcu_node structure's lock in order to | |
1114 | * start one (if needed). | |
1115 | */ | |
1116 | if (rnp != rnp_root) | |
1117 | raw_spin_lock(&rnp_root->lock); | |
1118 | ||
1119 | /* | |
1120 | * Get a new grace-period number. If there really is no grace | |
1121 | * period in progress, it will be smaller than the one we obtained | |
1122 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1123 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1124 | */ | |
1125 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1126 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1127 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1128 | rdp->nxtcompleted[i] = c; | |
1129 | ||
1130 | /* | |
1131 | * If the needed for the required grace period is already | |
1132 | * recorded, trace and leave. | |
1133 | */ | |
1134 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1135 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1136 | goto unlock_out; |
1137 | } | |
1138 | ||
1139 | /* Record the need for the future grace period. */ | |
1140 | rnp_root->need_future_gp[c & 0x1]++; | |
1141 | ||
1142 | /* If a grace period is not already in progress, start one. */ | |
1143 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1144 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1145 | } else { |
f7f7bac9 | 1146 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
910ee45d | 1147 | rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1148 | } |
1149 | unlock_out: | |
1150 | if (rnp != rnp_root) | |
1151 | raw_spin_unlock(&rnp_root->lock); | |
1152 | return c; | |
1153 | } | |
1154 | ||
1155 | /* | |
1156 | * Clean up any old requests for the just-ended grace period. Also return | |
1157 | * whether any additional grace periods have been requested. Also invoke | |
1158 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1159 | * waiting for this grace period to complete. | |
1160 | */ | |
1161 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1162 | { | |
1163 | int c = rnp->completed; | |
1164 | int needmore; | |
1165 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1166 | ||
1167 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1168 | rnp->need_future_gp[c & 0x1] = 0; | |
1169 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1170 | trace_rcu_future_gp(rnp, rdp, c, |
1171 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1172 | return needmore; |
1173 | } | |
1174 | ||
dc35c893 PM |
1175 | /* |
1176 | * If there is room, assign a ->completed number to any callbacks on | |
1177 | * this CPU that have not already been assigned. Also accelerate any | |
1178 | * callbacks that were previously assigned a ->completed number that has | |
1179 | * since proven to be too conservative, which can happen if callbacks get | |
1180 | * assigned a ->completed number while RCU is idle, but with reference to | |
1181 | * a non-root rcu_node structure. This function is idempotent, so it does | |
1182 | * not hurt to call it repeatedly. | |
1183 | * | |
1184 | * The caller must hold rnp->lock with interrupts disabled. | |
1185 | */ | |
1186 | static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1187 | struct rcu_data *rdp) | |
1188 | { | |
1189 | unsigned long c; | |
1190 | int i; | |
1191 | ||
1192 | /* If the CPU has no callbacks, nothing to do. */ | |
1193 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1194 | return; | |
1195 | ||
1196 | /* | |
1197 | * Starting from the sublist containing the callbacks most | |
1198 | * recently assigned a ->completed number and working down, find the | |
1199 | * first sublist that is not assignable to an upcoming grace period. | |
1200 | * Such a sublist has something in it (first two tests) and has | |
1201 | * a ->completed number assigned that will complete sooner than | |
1202 | * the ->completed number for newly arrived callbacks (last test). | |
1203 | * | |
1204 | * The key point is that any later sublist can be assigned the | |
1205 | * same ->completed number as the newly arrived callbacks, which | |
1206 | * means that the callbacks in any of these later sublist can be | |
1207 | * grouped into a single sublist, whether or not they have already | |
1208 | * been assigned a ->completed number. | |
1209 | */ | |
1210 | c = rcu_cbs_completed(rsp, rnp); | |
1211 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1212 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1213 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1214 | break; | |
1215 | ||
1216 | /* | |
1217 | * If there are no sublist for unassigned callbacks, leave. | |
1218 | * At the same time, advance "i" one sublist, so that "i" will | |
1219 | * index into the sublist where all the remaining callbacks should | |
1220 | * be grouped into. | |
1221 | */ | |
1222 | if (++i >= RCU_NEXT_TAIL) | |
1223 | return; | |
1224 | ||
1225 | /* | |
1226 | * Assign all subsequent callbacks' ->completed number to the next | |
1227 | * full grace period and group them all in the sublist initially | |
1228 | * indexed by "i". | |
1229 | */ | |
1230 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1231 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1232 | rdp->nxtcompleted[i] = c; | |
1233 | } | |
910ee45d PM |
1234 | /* Record any needed additional grace periods. */ |
1235 | rcu_start_future_gp(rnp, rdp); | |
6d4b418c PM |
1236 | |
1237 | /* Trace depending on how much we were able to accelerate. */ | |
1238 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1239 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1240 | else |
f7f7bac9 | 1241 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
dc35c893 PM |
1242 | } |
1243 | ||
1244 | /* | |
1245 | * Move any callbacks whose grace period has completed to the | |
1246 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1247 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1248 | * sublist. This function is idempotent, so it does not hurt to | |
1249 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
1250 | * | |
1251 | * The caller must hold rnp->lock with interrupts disabled. | |
1252 | */ | |
1253 | static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, | |
1254 | struct rcu_data *rdp) | |
1255 | { | |
1256 | int i, j; | |
1257 | ||
1258 | /* If the CPU has no callbacks, nothing to do. */ | |
1259 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
1260 | return; | |
1261 | ||
1262 | /* | |
1263 | * Find all callbacks whose ->completed numbers indicate that they | |
1264 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1265 | */ | |
1266 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1267 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1268 | break; | |
1269 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1270 | } | |
1271 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1272 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1273 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1274 | ||
1275 | /* Copy down callbacks to fill in empty sublists. */ | |
1276 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1277 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1278 | break; | |
1279 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1280 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1281 | } | |
1282 | ||
1283 | /* Classify any remaining callbacks. */ | |
1284 | rcu_accelerate_cbs(rsp, rnp, rdp); | |
1285 | } | |
1286 | ||
d09b62df | 1287 | /* |
ba9fbe95 PM |
1288 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1289 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1290 | * structure corresponding to the current CPU, and must have irqs disabled. | |
d09b62df | 1291 | */ |
ba9fbe95 | 1292 | static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp) |
d09b62df | 1293 | { |
ba9fbe95 | 1294 | /* Handle the ends of any preceding grace periods first. */ |
dc35c893 | 1295 | if (rdp->completed == rnp->completed) { |
d09b62df | 1296 | |
ba9fbe95 | 1297 | /* No grace period end, so just accelerate recent callbacks. */ |
dc35c893 | 1298 | rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1299 | |
dc35c893 PM |
1300 | } else { |
1301 | ||
1302 | /* Advance callbacks. */ | |
1303 | rcu_advance_cbs(rsp, rnp, rdp); | |
d09b62df PM |
1304 | |
1305 | /* Remember that we saw this grace-period completion. */ | |
1306 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1307 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1308 | } |
398ebe60 | 1309 | |
6eaef633 PM |
1310 | if (rdp->gpnum != rnp->gpnum) { |
1311 | /* | |
1312 | * If the current grace period is waiting for this CPU, | |
1313 | * set up to detect a quiescent state, otherwise don't | |
1314 | * go looking for one. | |
1315 | */ | |
1316 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1317 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
6eaef633 PM |
1318 | rdp->passed_quiesce = 0; |
1319 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); | |
1320 | zero_cpu_stall_ticks(rdp); | |
1321 | } | |
1322 | } | |
1323 | ||
d34ea322 | 1324 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1325 | { |
1326 | unsigned long flags; | |
1327 | struct rcu_node *rnp; | |
1328 | ||
1329 | local_irq_save(flags); | |
1330 | rnp = rdp->mynode; | |
d34ea322 PM |
1331 | if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) && |
1332 | rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */ | |
6eaef633 PM |
1333 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
1334 | local_irq_restore(flags); | |
1335 | return; | |
1336 | } | |
d34ea322 | 1337 | __note_gp_changes(rsp, rnp, rdp); |
6eaef633 PM |
1338 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1339 | } | |
1340 | ||
b3dbec76 | 1341 | /* |
f7be8209 | 1342 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1343 | */ |
7fdefc10 | 1344 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 PM |
1345 | { |
1346 | struct rcu_data *rdp; | |
7fdefc10 | 1347 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1348 | |
eb75767b | 1349 | rcu_bind_gp_kthread(); |
7fdefc10 | 1350 | raw_spin_lock_irq(&rnp->lock); |
f7be8209 PM |
1351 | if (rsp->gp_flags == 0) { |
1352 | /* Spurious wakeup, tell caller to go back to sleep. */ | |
1353 | raw_spin_unlock_irq(&rnp->lock); | |
1354 | return 0; | |
1355 | } | |
4cdfc175 | 1356 | rsp->gp_flags = 0; /* Clear all flags: New grace period. */ |
b3dbec76 | 1357 | |
f7be8209 PM |
1358 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1359 | /* | |
1360 | * Grace period already in progress, don't start another. | |
1361 | * Not supposed to be able to happen. | |
1362 | */ | |
7fdefc10 PM |
1363 | raw_spin_unlock_irq(&rnp->lock); |
1364 | return 0; | |
1365 | } | |
1366 | ||
7fdefc10 | 1367 | /* Advance to a new grace period and initialize state. */ |
26cdfedf PM |
1368 | record_gp_stall_check_time(rsp); |
1369 | smp_wmb(); /* Record GP times before starting GP. */ | |
7fdefc10 | 1370 | rsp->gpnum++; |
f7f7bac9 | 1371 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1372 | raw_spin_unlock_irq(&rnp->lock); |
1373 | ||
1374 | /* Exclude any concurrent CPU-hotplug operations. */ | |
a4fbe35a | 1375 | mutex_lock(&rsp->onoff_mutex); |
7fdefc10 PM |
1376 | |
1377 | /* | |
1378 | * Set the quiescent-state-needed bits in all the rcu_node | |
1379 | * structures for all currently online CPUs in breadth-first order, | |
1380 | * starting from the root rcu_node structure, relying on the layout | |
1381 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1382 | * will access only the leaves of the hierarchy, thus seeing that no | |
1383 | * grace period is in progress, at least until the corresponding | |
1384 | * leaf node has been initialized. In addition, we have excluded | |
1385 | * CPU-hotplug operations. | |
1386 | * | |
1387 | * The grace period cannot complete until the initialization | |
1388 | * process finishes, because this kthread handles both. | |
1389 | */ | |
1390 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
b3dbec76 | 1391 | raw_spin_lock_irq(&rnp->lock); |
b3dbec76 | 1392 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1393 | rcu_preempt_check_blocked_tasks(rnp); |
1394 | rnp->qsmask = rnp->qsmaskinit; | |
0446be48 | 1395 | ACCESS_ONCE(rnp->gpnum) = rsp->gpnum; |
25d30cf4 | 1396 | WARN_ON_ONCE(rnp->completed != rsp->completed); |
0446be48 | 1397 | ACCESS_ONCE(rnp->completed) = rsp->completed; |
7fdefc10 | 1398 | if (rnp == rdp->mynode) |
ce3d9c03 | 1399 | __note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1400 | rcu_preempt_boost_start_gp(rnp); |
1401 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1402 | rnp->level, rnp->grplo, | |
1403 | rnp->grphi, rnp->qsmask); | |
1404 | raw_spin_unlock_irq(&rnp->lock); | |
661a85dc | 1405 | #ifdef CONFIG_PROVE_RCU_DELAY |
971394f3 | 1406 | if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 && |
81e59494 | 1407 | system_state == SYSTEM_RUNNING) |
971394f3 | 1408 | udelay(200); |
661a85dc | 1409 | #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */ |
7fdefc10 PM |
1410 | cond_resched(); |
1411 | } | |
b3dbec76 | 1412 | |
a4fbe35a | 1413 | mutex_unlock(&rsp->onoff_mutex); |
7fdefc10 PM |
1414 | return 1; |
1415 | } | |
b3dbec76 | 1416 | |
4cdfc175 PM |
1417 | /* |
1418 | * Do one round of quiescent-state forcing. | |
1419 | */ | |
01896f7e | 1420 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) |
4cdfc175 PM |
1421 | { |
1422 | int fqs_state = fqs_state_in; | |
217af2a2 PM |
1423 | bool isidle = false; |
1424 | unsigned long maxj; | |
4cdfc175 PM |
1425 | struct rcu_node *rnp = rcu_get_root(rsp); |
1426 | ||
1427 | rsp->n_force_qs++; | |
1428 | if (fqs_state == RCU_SAVE_DYNTICK) { | |
1429 | /* Collect dyntick-idle snapshots. */ | |
0edd1b17 PM |
1430 | if (is_sysidle_rcu_state(rsp)) { |
1431 | isidle = 1; | |
1432 | maxj = jiffies - ULONG_MAX / 4; | |
1433 | } | |
217af2a2 PM |
1434 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1435 | &isidle, &maxj); | |
0edd1b17 | 1436 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1437 | fqs_state = RCU_FORCE_QS; |
1438 | } else { | |
1439 | /* Handle dyntick-idle and offline CPUs. */ | |
0edd1b17 | 1440 | isidle = 0; |
217af2a2 | 1441 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1442 | } |
1443 | /* Clear flag to prevent immediate re-entry. */ | |
1444 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
1445 | raw_spin_lock_irq(&rnp->lock); | |
1446 | rsp->gp_flags &= ~RCU_GP_FLAG_FQS; | |
1447 | raw_spin_unlock_irq(&rnp->lock); | |
1448 | } | |
1449 | return fqs_state; | |
1450 | } | |
1451 | ||
7fdefc10 PM |
1452 | /* |
1453 | * Clean up after the old grace period. | |
1454 | */ | |
4cdfc175 | 1455 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1456 | { |
1457 | unsigned long gp_duration; | |
dae6e64d | 1458 | int nocb = 0; |
7fdefc10 PM |
1459 | struct rcu_data *rdp; |
1460 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1461 | |
7fdefc10 PM |
1462 | raw_spin_lock_irq(&rnp->lock); |
1463 | gp_duration = jiffies - rsp->gp_start; | |
1464 | if (gp_duration > rsp->gp_max) | |
1465 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1466 | |
7fdefc10 PM |
1467 | /* |
1468 | * We know the grace period is complete, but to everyone else | |
1469 | * it appears to still be ongoing. But it is also the case | |
1470 | * that to everyone else it looks like there is nothing that | |
1471 | * they can do to advance the grace period. It is therefore | |
1472 | * safe for us to drop the lock in order to mark the grace | |
1473 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1474 | */ |
5d4b8659 | 1475 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1476 | |
5d4b8659 PM |
1477 | /* |
1478 | * Propagate new ->completed value to rcu_node structures so | |
1479 | * that other CPUs don't have to wait until the start of the next | |
1480 | * grace period to process their callbacks. This also avoids | |
1481 | * some nasty RCU grace-period initialization races by forcing | |
1482 | * the end of the current grace period to be completely recorded in | |
1483 | * all of the rcu_node structures before the beginning of the next | |
1484 | * grace period is recorded in any of the rcu_node structures. | |
1485 | */ | |
1486 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 1487 | raw_spin_lock_irq(&rnp->lock); |
0446be48 | 1488 | ACCESS_ONCE(rnp->completed) = rsp->gpnum; |
b11cc576 PM |
1489 | rdp = this_cpu_ptr(rsp->rda); |
1490 | if (rnp == rdp->mynode) | |
470716fc | 1491 | __note_gp_changes(rsp, rnp, rdp); |
0446be48 | 1492 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 PM |
1493 | raw_spin_unlock_irq(&rnp->lock); |
1494 | cond_resched(); | |
7fdefc10 | 1495 | } |
5d4b8659 PM |
1496 | rnp = rcu_get_root(rsp); |
1497 | raw_spin_lock_irq(&rnp->lock); | |
dae6e64d | 1498 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 PM |
1499 | |
1500 | rsp->completed = rsp->gpnum; /* Declare grace period done. */ | |
f7f7bac9 | 1501 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
7fdefc10 | 1502 | rsp->fqs_state = RCU_GP_IDLE; |
5d4b8659 | 1503 | rdp = this_cpu_ptr(rsp->rda); |
b11cc576 | 1504 | rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */ |
bb311ecc | 1505 | if (cpu_needs_another_gp(rsp, rdp)) { |
b3f2d025 | 1506 | rsp->gp_flags = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1507 | trace_rcu_grace_period(rsp->name, |
1508 | ACCESS_ONCE(rsp->gpnum), | |
1509 | TPS("newreq")); | |
1510 | } | |
7fdefc10 | 1511 | raw_spin_unlock_irq(&rnp->lock); |
7fdefc10 PM |
1512 | } |
1513 | ||
1514 | /* | |
1515 | * Body of kthread that handles grace periods. | |
1516 | */ | |
1517 | static int __noreturn rcu_gp_kthread(void *arg) | |
1518 | { | |
4cdfc175 | 1519 | int fqs_state; |
88d6df61 | 1520 | int gf; |
d40011f6 | 1521 | unsigned long j; |
4cdfc175 | 1522 | int ret; |
7fdefc10 PM |
1523 | struct rcu_state *rsp = arg; |
1524 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1525 | ||
1526 | for (;;) { | |
1527 | ||
1528 | /* Handle grace-period start. */ | |
1529 | for (;;) { | |
63c4db78 PM |
1530 | trace_rcu_grace_period(rsp->name, |
1531 | ACCESS_ONCE(rsp->gpnum), | |
1532 | TPS("reqwait")); | |
4cdfc175 | 1533 | wait_event_interruptible(rsp->gp_wq, |
591c6d17 | 1534 | ACCESS_ONCE(rsp->gp_flags) & |
4cdfc175 | 1535 | RCU_GP_FLAG_INIT); |
f7be8209 | 1536 | if (rcu_gp_init(rsp)) |
7fdefc10 PM |
1537 | break; |
1538 | cond_resched(); | |
1539 | flush_signals(current); | |
63c4db78 PM |
1540 | trace_rcu_grace_period(rsp->name, |
1541 | ACCESS_ONCE(rsp->gpnum), | |
1542 | TPS("reqwaitsig")); | |
7fdefc10 | 1543 | } |
cabc49c1 | 1544 | |
4cdfc175 PM |
1545 | /* Handle quiescent-state forcing. */ |
1546 | fqs_state = RCU_SAVE_DYNTICK; | |
d40011f6 PM |
1547 | j = jiffies_till_first_fqs; |
1548 | if (j > HZ) { | |
1549 | j = HZ; | |
1550 | jiffies_till_first_fqs = HZ; | |
1551 | } | |
88d6df61 | 1552 | ret = 0; |
cabc49c1 | 1553 | for (;;) { |
88d6df61 PM |
1554 | if (!ret) |
1555 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 PM |
1556 | trace_rcu_grace_period(rsp->name, |
1557 | ACCESS_ONCE(rsp->gpnum), | |
1558 | TPS("fqswait")); | |
4cdfc175 | 1559 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
88d6df61 PM |
1560 | ((gf = ACCESS_ONCE(rsp->gp_flags)) & |
1561 | RCU_GP_FLAG_FQS) || | |
4cdfc175 PM |
1562 | (!ACCESS_ONCE(rnp->qsmask) && |
1563 | !rcu_preempt_blocked_readers_cgp(rnp)), | |
d40011f6 | 1564 | j); |
4cdfc175 | 1565 | /* If grace period done, leave loop. */ |
cabc49c1 | 1566 | if (!ACCESS_ONCE(rnp->qsmask) && |
4cdfc175 | 1567 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 1568 | break; |
4cdfc175 | 1569 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
1570 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
1571 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 PM |
1572 | trace_rcu_grace_period(rsp->name, |
1573 | ACCESS_ONCE(rsp->gpnum), | |
1574 | TPS("fqsstart")); | |
4cdfc175 | 1575 | fqs_state = rcu_gp_fqs(rsp, fqs_state); |
63c4db78 PM |
1576 | trace_rcu_grace_period(rsp->name, |
1577 | ACCESS_ONCE(rsp->gpnum), | |
1578 | TPS("fqsend")); | |
4cdfc175 PM |
1579 | cond_resched(); |
1580 | } else { | |
1581 | /* Deal with stray signal. */ | |
1582 | cond_resched(); | |
1583 | flush_signals(current); | |
63c4db78 PM |
1584 | trace_rcu_grace_period(rsp->name, |
1585 | ACCESS_ONCE(rsp->gpnum), | |
1586 | TPS("fqswaitsig")); | |
4cdfc175 | 1587 | } |
d40011f6 PM |
1588 | j = jiffies_till_next_fqs; |
1589 | if (j > HZ) { | |
1590 | j = HZ; | |
1591 | jiffies_till_next_fqs = HZ; | |
1592 | } else if (j < 1) { | |
1593 | j = 1; | |
1594 | jiffies_till_next_fqs = 1; | |
1595 | } | |
cabc49c1 | 1596 | } |
4cdfc175 PM |
1597 | |
1598 | /* Handle grace-period end. */ | |
1599 | rcu_gp_cleanup(rsp); | |
b3dbec76 | 1600 | } |
b3dbec76 PM |
1601 | } |
1602 | ||
016a8d5b SR |
1603 | static void rsp_wakeup(struct irq_work *work) |
1604 | { | |
1605 | struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work); | |
1606 | ||
1607 | /* Wake up rcu_gp_kthread() to start the grace period. */ | |
1608 | wake_up(&rsp->gp_wq); | |
1609 | } | |
1610 | ||
64db4cff PM |
1611 | /* |
1612 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
1613 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 1614 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
1615 | * |
1616 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
1617 | * invoke this function. This can happen when the dying CPU reports its | |
1618 | * quiescent state. | |
64db4cff PM |
1619 | */ |
1620 | static void | |
910ee45d PM |
1621 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
1622 | struct rcu_data *rdp) | |
64db4cff | 1623 | { |
b8462084 | 1624 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 1625 | /* |
b3dbec76 | 1626 | * Either we have not yet spawned the grace-period |
62da1921 PM |
1627 | * task, this CPU does not need another grace period, |
1628 | * or a grace period is already in progress. | |
b3dbec76 | 1629 | * Either way, don't start a new grace period. |
afe24b12 | 1630 | */ |
afe24b12 PM |
1631 | return; |
1632 | } | |
4cdfc175 | 1633 | rsp->gp_flags = RCU_GP_FLAG_INIT; |
bb311ecc PM |
1634 | trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum), |
1635 | TPS("newreq")); | |
62da1921 | 1636 | |
016a8d5b SR |
1637 | /* |
1638 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c PM |
1639 | * could cause possible deadlocks with the rq->lock. Defer |
1640 | * the wakeup to interrupt context. And don't bother waking | |
1641 | * up the running kthread. | |
016a8d5b | 1642 | */ |
1eafd31c PM |
1643 | if (current != rsp->gp_kthread) |
1644 | irq_work_queue(&rsp->wakeup_work); | |
64db4cff PM |
1645 | } |
1646 | ||
910ee45d PM |
1647 | /* |
1648 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
1649 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
1650 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
1651 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
1652 | * that is encountered beforehand. | |
1653 | */ | |
1654 | static void | |
1655 | rcu_start_gp(struct rcu_state *rsp) | |
1656 | { | |
1657 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1658 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1659 | ||
1660 | /* | |
1661 | * If there is no grace period in progress right now, any | |
1662 | * callbacks we have up to this point will be satisfied by the | |
1663 | * next grace period. Also, advancing the callbacks reduces the | |
1664 | * probability of false positives from cpu_needs_another_gp() | |
1665 | * resulting in pointless grace periods. So, advance callbacks | |
1666 | * then start the grace period! | |
1667 | */ | |
1668 | rcu_advance_cbs(rsp, rnp, rdp); | |
1669 | rcu_start_gp_advanced(rsp, rnp, rdp); | |
1670 | } | |
1671 | ||
f41d911f | 1672 | /* |
d3f6bad3 PM |
1673 | * Report a full set of quiescent states to the specified rcu_state |
1674 | * data structure. This involves cleaning up after the prior grace | |
1675 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
1676 | * if one is needed. Note that the caller must hold rnp->lock, which |
1677 | * is released before return. | |
f41d911f | 1678 | */ |
d3f6bad3 | 1679 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 1680 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 1681 | { |
fc2219d4 | 1682 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cabc49c1 PM |
1683 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
1684 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ | |
f41d911f PM |
1685 | } |
1686 | ||
64db4cff | 1687 | /* |
d3f6bad3 PM |
1688 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
1689 | * Allows quiescent states for a group of CPUs to be reported at one go | |
1690 | * to the specified rcu_node structure, though all the CPUs in the group | |
1691 | * must be represented by the same rcu_node structure (which need not be | |
1692 | * a leaf rcu_node structure, though it often will be). That structure's | |
1693 | * lock must be held upon entry, and it is released before return. | |
64db4cff PM |
1694 | */ |
1695 | static void | |
d3f6bad3 PM |
1696 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
1697 | struct rcu_node *rnp, unsigned long flags) | |
64db4cff PM |
1698 | __releases(rnp->lock) |
1699 | { | |
28ecd580 PM |
1700 | struct rcu_node *rnp_c; |
1701 | ||
64db4cff PM |
1702 | /* Walk up the rcu_node hierarchy. */ |
1703 | for (;;) { | |
1704 | if (!(rnp->qsmask & mask)) { | |
1705 | ||
1706 | /* Our bit has already been cleared, so done. */ | |
1304afb2 | 1707 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1708 | return; |
1709 | } | |
1710 | rnp->qsmask &= ~mask; | |
d4c08f2a PM |
1711 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
1712 | mask, rnp->qsmask, rnp->level, | |
1713 | rnp->grplo, rnp->grphi, | |
1714 | !!rnp->gp_tasks); | |
27f4d280 | 1715 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
1716 | |
1717 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 1718 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1719 | return; |
1720 | } | |
1721 | mask = rnp->grpmask; | |
1722 | if (rnp->parent == NULL) { | |
1723 | ||
1724 | /* No more levels. Exit loop holding root lock. */ | |
1725 | ||
1726 | break; | |
1727 | } | |
1304afb2 | 1728 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 1729 | rnp_c = rnp; |
64db4cff | 1730 | rnp = rnp->parent; |
1304afb2 | 1731 | raw_spin_lock_irqsave(&rnp->lock, flags); |
28ecd580 | 1732 | WARN_ON_ONCE(rnp_c->qsmask); |
64db4cff PM |
1733 | } |
1734 | ||
1735 | /* | |
1736 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 1737 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 1738 | * to clean up and start the next grace period if one is needed. |
64db4cff | 1739 | */ |
d3f6bad3 | 1740 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
1741 | } |
1742 | ||
1743 | /* | |
d3f6bad3 PM |
1744 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
1745 | * structure. This must be either called from the specified CPU, or | |
1746 | * called when the specified CPU is known to be offline (and when it is | |
1747 | * also known that no other CPU is concurrently trying to help the offline | |
1748 | * CPU). The lastcomp argument is used to make sure we are still in the | |
1749 | * grace period of interest. We don't want to end the current grace period | |
1750 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
1751 | */ |
1752 | static void | |
d7d6a11e | 1753 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
1754 | { |
1755 | unsigned long flags; | |
1756 | unsigned long mask; | |
1757 | struct rcu_node *rnp; | |
1758 | ||
1759 | rnp = rdp->mynode; | |
1304afb2 | 1760 | raw_spin_lock_irqsave(&rnp->lock, flags); |
d7d6a11e PM |
1761 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum || |
1762 | rnp->completed == rnp->gpnum) { | |
64db4cff PM |
1763 | |
1764 | /* | |
e4cc1f22 PM |
1765 | * The grace period in which this quiescent state was |
1766 | * recorded has ended, so don't report it upwards. | |
1767 | * We will instead need a new quiescent state that lies | |
1768 | * within the current grace period. | |
64db4cff | 1769 | */ |
e4cc1f22 | 1770 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
1304afb2 | 1771 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1772 | return; |
1773 | } | |
1774 | mask = rdp->grpmask; | |
1775 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 1776 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1777 | } else { |
1778 | rdp->qs_pending = 0; | |
1779 | ||
1780 | /* | |
1781 | * This GP can't end until cpu checks in, so all of our | |
1782 | * callbacks can be processed during the next GP. | |
1783 | */ | |
dc35c893 | 1784 | rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 1785 | |
d3f6bad3 | 1786 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
64db4cff PM |
1787 | } |
1788 | } | |
1789 | ||
1790 | /* | |
1791 | * Check to see if there is a new grace period of which this CPU | |
1792 | * is not yet aware, and if so, set up local rcu_data state for it. | |
1793 | * Otherwise, see if this CPU has just passed through its first | |
1794 | * quiescent state for this grace period, and record that fact if so. | |
1795 | */ | |
1796 | static void | |
1797 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
1798 | { | |
05eb552b PM |
1799 | /* Check for grace-period ends and beginnings. */ |
1800 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
1801 | |
1802 | /* | |
1803 | * Does this CPU still need to do its part for current grace period? | |
1804 | * If no, return and let the other CPUs do their part as well. | |
1805 | */ | |
1806 | if (!rdp->qs_pending) | |
1807 | return; | |
1808 | ||
1809 | /* | |
1810 | * Was there a quiescent state since the beginning of the grace | |
1811 | * period? If no, then exit and wait for the next call. | |
1812 | */ | |
e4cc1f22 | 1813 | if (!rdp->passed_quiesce) |
64db4cff PM |
1814 | return; |
1815 | ||
d3f6bad3 PM |
1816 | /* |
1817 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
1818 | * judge of that). | |
1819 | */ | |
d7d6a11e | 1820 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
1821 | } |
1822 | ||
1823 | #ifdef CONFIG_HOTPLUG_CPU | |
1824 | ||
e74f4c45 | 1825 | /* |
b1420f1c PM |
1826 | * Send the specified CPU's RCU callbacks to the orphanage. The |
1827 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 1828 | * ->orphan_lock. |
e74f4c45 | 1829 | */ |
b1420f1c PM |
1830 | static void |
1831 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
1832 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 1833 | { |
3fbfbf7a | 1834 | /* No-CBs CPUs do not have orphanable callbacks. */ |
d1e43fa5 | 1835 | if (rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
1836 | return; |
1837 | ||
b1420f1c PM |
1838 | /* |
1839 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
1840 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
1841 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 1842 | */ |
a50c3af9 | 1843 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1844 | rsp->qlen_lazy += rdp->qlen_lazy; |
1845 | rsp->qlen += rdp->qlen; | |
1846 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 1847 | rdp->qlen_lazy = 0; |
1d1fb395 | 1848 | ACCESS_ONCE(rdp->qlen) = 0; |
a50c3af9 PM |
1849 | } |
1850 | ||
1851 | /* | |
b1420f1c PM |
1852 | * Next, move those callbacks still needing a grace period to |
1853 | * the orphanage, where some other CPU will pick them up. | |
1854 | * Some of the callbacks might have gone partway through a grace | |
1855 | * period, but that is too bad. They get to start over because we | |
1856 | * cannot assume that grace periods are synchronized across CPUs. | |
1857 | * We don't bother updating the ->nxttail[] array yet, instead | |
1858 | * we just reset the whole thing later on. | |
a50c3af9 | 1859 | */ |
b1420f1c PM |
1860 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
1861 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1862 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
1863 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
1864 | } |
1865 | ||
1866 | /* | |
b1420f1c PM |
1867 | * Then move the ready-to-invoke callbacks to the orphanage, |
1868 | * where some other CPU will pick them up. These will not be | |
1869 | * required to pass though another grace period: They are done. | |
a50c3af9 | 1870 | */ |
e5601400 | 1871 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
1872 | *rsp->orphan_donetail = rdp->nxtlist; |
1873 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 1874 | } |
e74f4c45 | 1875 | |
b1420f1c | 1876 | /* Finally, initialize the rcu_data structure's list to empty. */ |
3f5d3ea6 | 1877 | init_callback_list(rdp); |
b1420f1c PM |
1878 | } |
1879 | ||
1880 | /* | |
1881 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 1882 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c PM |
1883 | */ |
1884 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp) | |
1885 | { | |
1886 | int i; | |
1887 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); | |
1888 | ||
3fbfbf7a PM |
1889 | /* No-CBs CPUs are handled specially. */ |
1890 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp)) | |
1891 | return; | |
1892 | ||
b1420f1c PM |
1893 | /* Do the accounting first. */ |
1894 | rdp->qlen_lazy += rsp->qlen_lazy; | |
1895 | rdp->qlen += rsp->qlen; | |
1896 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
1897 | if (rsp->qlen_lazy != rsp->qlen) |
1898 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
1899 | rsp->qlen_lazy = 0; |
1900 | rsp->qlen = 0; | |
1901 | ||
1902 | /* | |
1903 | * We do not need a memory barrier here because the only way we | |
1904 | * can get here if there is an rcu_barrier() in flight is if | |
1905 | * we are the task doing the rcu_barrier(). | |
1906 | */ | |
1907 | ||
1908 | /* First adopt the ready-to-invoke callbacks. */ | |
1909 | if (rsp->orphan_donelist != NULL) { | |
1910 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
1911 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
1912 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
1913 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
1914 | rdp->nxttail[i] = rsp->orphan_donetail; | |
1915 | rsp->orphan_donelist = NULL; | |
1916 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
1917 | } | |
1918 | ||
1919 | /* And then adopt the callbacks that still need a grace period. */ | |
1920 | if (rsp->orphan_nxtlist != NULL) { | |
1921 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
1922 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
1923 | rsp->orphan_nxtlist = NULL; | |
1924 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
1925 | } | |
1926 | } | |
1927 | ||
1928 | /* | |
1929 | * Trace the fact that this CPU is going offline. | |
1930 | */ | |
1931 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
1932 | { | |
1933 | RCU_TRACE(unsigned long mask); | |
1934 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
1935 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
1936 | ||
1937 | RCU_TRACE(mask = rdp->grpmask); | |
e5601400 PM |
1938 | trace_rcu_grace_period(rsp->name, |
1939 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 1940 | TPS("cpuofl")); |
64db4cff PM |
1941 | } |
1942 | ||
1943 | /* | |
e5601400 | 1944 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
1945 | * this fact from process context. Do the remainder of the cleanup, |
1946 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
1947 | * adopting them. There can only be one CPU hotplug operation at a time, |
1948 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 1949 | */ |
e5601400 | 1950 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 1951 | { |
2036d94a PM |
1952 | unsigned long flags; |
1953 | unsigned long mask; | |
1954 | int need_report = 0; | |
e5601400 | 1955 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 1956 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 1957 | |
2036d94a | 1958 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 1959 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 1960 | |
b1420f1c | 1961 | /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */ |
2036d94a PM |
1962 | |
1963 | /* Exclude any attempts to start a new grace period. */ | |
a4fbe35a | 1964 | mutex_lock(&rsp->onoff_mutex); |
7b2e6011 | 1965 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
2036d94a | 1966 | |
b1420f1c PM |
1967 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
1968 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); | |
1969 | rcu_adopt_orphan_cbs(rsp); | |
1970 | ||
2036d94a PM |
1971 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
1972 | mask = rdp->grpmask; /* rnp->grplo is constant. */ | |
1973 | do { | |
1974 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
1975 | rnp->qsmaskinit &= ~mask; | |
1976 | if (rnp->qsmaskinit != 0) { | |
1977 | if (rnp != rdp->mynode) | |
1978 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1979 | break; | |
1980 | } | |
1981 | if (rnp == rdp->mynode) | |
1982 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); | |
1983 | else | |
1984 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
1985 | mask = rnp->grpmask; | |
1986 | rnp = rnp->parent; | |
1987 | } while (rnp != NULL); | |
1988 | ||
1989 | /* | |
1990 | * We still hold the leaf rcu_node structure lock here, and | |
1991 | * irqs are still disabled. The reason for this subterfuge is | |
7b2e6011 | 1992 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock |
2036d94a PM |
1993 | * held leads to deadlock. |
1994 | */ | |
7b2e6011 | 1995 | raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */ |
2036d94a PM |
1996 | rnp = rdp->mynode; |
1997 | if (need_report & RCU_OFL_TASKS_NORM_GP) | |
1998 | rcu_report_unblock_qs_rnp(rnp, flags); | |
1999 | else | |
2000 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2001 | if (need_report & RCU_OFL_TASKS_EXP_GP) | |
2002 | rcu_report_exp_rnp(rsp, rnp, true); | |
cf01537e PM |
2003 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
2004 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2005 | cpu, rdp->qlen, rdp->nxtlist); | |
0d8ee37e PM |
2006 | init_callback_list(rdp); |
2007 | /* Disallow further callbacks on this CPU. */ | |
2008 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; | |
a4fbe35a | 2009 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
2010 | } |
2011 | ||
2012 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ | |
2013 | ||
e5601400 | 2014 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
e74f4c45 PM |
2015 | { |
2016 | } | |
2017 | ||
e5601400 | 2018 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff PM |
2019 | { |
2020 | } | |
2021 | ||
2022 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ | |
2023 | ||
2024 | /* | |
2025 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2026 | * period. Thottle as specified by rdp->blimit. | |
2027 | */ | |
37c72e56 | 2028 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2029 | { |
2030 | unsigned long flags; | |
2031 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
2032 | long bl, count, count_lazy; |
2033 | int i; | |
64db4cff | 2034 | |
dc35c893 | 2035 | /* If no callbacks are ready, just return. */ |
29c00b4a | 2036 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 2037 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
4968c300 PM |
2038 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
2039 | need_resched(), is_idle_task(current), | |
2040 | rcu_is_callbacks_kthread()); | |
64db4cff | 2041 | return; |
29c00b4a | 2042 | } |
64db4cff PM |
2043 | |
2044 | /* | |
2045 | * Extract the list of ready callbacks, disabling to prevent | |
2046 | * races with call_rcu() from interrupt handlers. | |
2047 | */ | |
2048 | local_irq_save(flags); | |
8146c4e2 | 2049 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2050 | bl = rdp->blimit; |
486e2593 | 2051 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2052 | list = rdp->nxtlist; |
2053 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2054 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2055 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2056 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2057 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2058 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2059 | local_irq_restore(flags); |
2060 | ||
2061 | /* Invoke callbacks. */ | |
486e2593 | 2062 | count = count_lazy = 0; |
64db4cff PM |
2063 | while (list) { |
2064 | next = list->next; | |
2065 | prefetch(next); | |
551d55a9 | 2066 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2067 | if (__rcu_reclaim(rsp->name, list)) |
2068 | count_lazy++; | |
64db4cff | 2069 | list = next; |
dff1672d PM |
2070 | /* Stop only if limit reached and CPU has something to do. */ |
2071 | if (++count >= bl && | |
2072 | (need_resched() || | |
2073 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2074 | break; |
2075 | } | |
2076 | ||
2077 | local_irq_save(flags); | |
4968c300 PM |
2078 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2079 | is_idle_task(current), | |
2080 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2081 | |
2082 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2083 | if (list != NULL) { |
2084 | *tail = rdp->nxtlist; | |
2085 | rdp->nxtlist = list; | |
b41772ab PM |
2086 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2087 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2088 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2089 | else |
2090 | break; | |
2091 | } | |
b1420f1c PM |
2092 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2093 | rdp->qlen_lazy -= count_lazy; | |
1d1fb395 | 2094 | ACCESS_ONCE(rdp->qlen) -= count; |
b1420f1c | 2095 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2096 | |
2097 | /* Reinstate batch limit if we have worked down the excess. */ | |
2098 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2099 | rdp->blimit = blimit; | |
2100 | ||
37c72e56 PM |
2101 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2102 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2103 | rdp->qlen_last_fqs_check = 0; | |
2104 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2105 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2106 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2107 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2108 | |
64db4cff PM |
2109 | local_irq_restore(flags); |
2110 | ||
e0f23060 | 2111 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2112 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2113 | invoke_rcu_core(); |
64db4cff PM |
2114 | } |
2115 | ||
2116 | /* | |
2117 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2118 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2119 | * Also schedule RCU core processing. |
64db4cff | 2120 | * |
9b2e4f18 | 2121 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2122 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2123 | * false, there is no point in invoking rcu_check_callbacks(). | |
2124 | */ | |
2125 | void rcu_check_callbacks(int cpu, int user) | |
2126 | { | |
f7f7bac9 | 2127 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2128 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2129 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2130 | |
2131 | /* | |
2132 | * Get here if this CPU took its interrupt from user | |
2133 | * mode or from the idle loop, and if this is not a | |
2134 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2135 | * a quiescent state, so note it. |
64db4cff PM |
2136 | * |
2137 | * No memory barrier is required here because both | |
d6714c22 PM |
2138 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2139 | * variables that other CPUs neither access nor modify, | |
2140 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2141 | */ |
2142 | ||
d6714c22 PM |
2143 | rcu_sched_qs(cpu); |
2144 | rcu_bh_qs(cpu); | |
64db4cff PM |
2145 | |
2146 | } else if (!in_softirq()) { | |
2147 | ||
2148 | /* | |
2149 | * Get here if this CPU did not take its interrupt from | |
2150 | * softirq, in other words, if it is not interrupting | |
2151 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2152 | * critical section, so note it. |
64db4cff PM |
2153 | */ |
2154 | ||
d6714c22 | 2155 | rcu_bh_qs(cpu); |
64db4cff | 2156 | } |
f41d911f | 2157 | rcu_preempt_check_callbacks(cpu); |
d21670ac | 2158 | if (rcu_pending(cpu)) |
a46e0899 | 2159 | invoke_rcu_core(); |
f7f7bac9 | 2160 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2161 | } |
2162 | ||
64db4cff PM |
2163 | /* |
2164 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2165 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2166 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2167 | * | |
ee47eb9f | 2168 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2169 | */ |
217af2a2 PM |
2170 | static void force_qs_rnp(struct rcu_state *rsp, |
2171 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2172 | unsigned long *maxj), | |
2173 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2174 | { |
2175 | unsigned long bit; | |
2176 | int cpu; | |
2177 | unsigned long flags; | |
2178 | unsigned long mask; | |
a0b6c9a7 | 2179 | struct rcu_node *rnp; |
64db4cff | 2180 | |
a0b6c9a7 | 2181 | rcu_for_each_leaf_node(rsp, rnp) { |
b4be093f | 2182 | cond_resched(); |
64db4cff | 2183 | mask = 0; |
1304afb2 | 2184 | raw_spin_lock_irqsave(&rnp->lock, flags); |
ee47eb9f | 2185 | if (!rcu_gp_in_progress(rsp)) { |
1304afb2 | 2186 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
0f10dc82 | 2187 | return; |
64db4cff | 2188 | } |
a0b6c9a7 | 2189 | if (rnp->qsmask == 0) { |
1217ed1b | 2190 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
64db4cff PM |
2191 | continue; |
2192 | } | |
a0b6c9a7 | 2193 | cpu = rnp->grplo; |
64db4cff | 2194 | bit = 1; |
a0b6c9a7 | 2195 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 PM |
2196 | if ((rnp->qsmask & bit) != 0) { |
2197 | if ((rnp->qsmaskinit & bit) != 0) | |
2198 | *isidle = 0; | |
2199 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) | |
2200 | mask |= bit; | |
2201 | } | |
64db4cff | 2202 | } |
45f014c5 | 2203 | if (mask != 0) { |
64db4cff | 2204 | |
d3f6bad3 PM |
2205 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
2206 | rcu_report_qs_rnp(mask, rsp, rnp, flags); | |
64db4cff PM |
2207 | continue; |
2208 | } | |
1304afb2 | 2209 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2210 | } |
27f4d280 | 2211 | rnp = rcu_get_root(rsp); |
1217ed1b PM |
2212 | if (rnp->qsmask == 0) { |
2213 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2214 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ | |
2215 | } | |
64db4cff PM |
2216 | } |
2217 | ||
2218 | /* | |
2219 | * Force quiescent states on reluctant CPUs, and also detect which | |
2220 | * CPUs are in dyntick-idle mode. | |
2221 | */ | |
4cdfc175 | 2222 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2223 | { |
2224 | unsigned long flags; | |
394f2769 PM |
2225 | bool ret; |
2226 | struct rcu_node *rnp; | |
2227 | struct rcu_node *rnp_old = NULL; | |
2228 | ||
2229 | /* Funnel through hierarchy to reduce memory contention. */ | |
2230 | rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode; | |
2231 | for (; rnp != NULL; rnp = rnp->parent) { | |
2232 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || | |
2233 | !raw_spin_trylock(&rnp->fqslock); | |
2234 | if (rnp_old != NULL) | |
2235 | raw_spin_unlock(&rnp_old->fqslock); | |
2236 | if (ret) { | |
2237 | rsp->n_force_qs_lh++; | |
2238 | return; | |
2239 | } | |
2240 | rnp_old = rnp; | |
2241 | } | |
2242 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2243 | |
394f2769 PM |
2244 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2245 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
2246 | raw_spin_unlock(&rnp_old->fqslock); | |
2247 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { | |
2248 | rsp->n_force_qs_lh++; | |
2249 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); | |
4cdfc175 | 2250 | return; /* Someone beat us to it. */ |
46a1e34e | 2251 | } |
4cdfc175 | 2252 | rsp->gp_flags |= RCU_GP_FLAG_FQS; |
394f2769 | 2253 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2254 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ |
64db4cff PM |
2255 | } |
2256 | ||
64db4cff | 2257 | /* |
e0f23060 PM |
2258 | * This does the RCU core processing work for the specified rcu_state |
2259 | * and rcu_data structures. This may be called only from the CPU to | |
2260 | * whom the rdp belongs. | |
64db4cff PM |
2261 | */ |
2262 | static void | |
1bca8cf1 | 2263 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2264 | { |
2265 | unsigned long flags; | |
1bca8cf1 | 2266 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
64db4cff | 2267 | |
2e597558 PM |
2268 | WARN_ON_ONCE(rdp->beenonline == 0); |
2269 | ||
64db4cff PM |
2270 | /* Update RCU state based on any recent quiescent states. */ |
2271 | rcu_check_quiescent_state(rsp, rdp); | |
2272 | ||
2273 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2274 | local_irq_save(flags); |
64db4cff | 2275 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2276 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
b8462084 PM |
2277 | rcu_start_gp(rsp); |
2278 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); | |
dc35c893 PM |
2279 | } else { |
2280 | local_irq_restore(flags); | |
64db4cff PM |
2281 | } |
2282 | ||
2283 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2284 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2285 | invoke_rcu_callbacks(rsp, rdp); |
09223371 SL |
2286 | } |
2287 | ||
64db4cff | 2288 | /* |
e0f23060 | 2289 | * Do RCU core processing for the current CPU. |
64db4cff | 2290 | */ |
09223371 | 2291 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2292 | { |
6ce75a23 PM |
2293 | struct rcu_state *rsp; |
2294 | ||
bfa00b4c PM |
2295 | if (cpu_is_offline(smp_processor_id())) |
2296 | return; | |
f7f7bac9 | 2297 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2298 | for_each_rcu_flavor(rsp) |
2299 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2300 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2301 | } |
2302 | ||
a26ac245 | 2303 | /* |
e0f23060 PM |
2304 | * Schedule RCU callback invocation. If the specified type of RCU |
2305 | * does not support RCU priority boosting, just do a direct call, | |
2306 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
2307 | * are running on the current CPU with interrupts disabled, the | |
2308 | * rcu_cpu_kthread_task cannot disappear out from under us. | |
a26ac245 | 2309 | */ |
a46e0899 | 2310 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2311 | { |
b0d30417 PM |
2312 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
2313 | return; | |
a46e0899 PM |
2314 | if (likely(!rsp->boost)) { |
2315 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2316 | return; |
2317 | } | |
a46e0899 | 2318 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2319 | } |
2320 | ||
a46e0899 | 2321 | static void invoke_rcu_core(void) |
09223371 | 2322 | { |
b0f74036 PM |
2323 | if (cpu_online(smp_processor_id())) |
2324 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2325 | } |
2326 | ||
29154c57 PM |
2327 | /* |
2328 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2329 | */ | |
2330 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2331 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2332 | { |
62fde6ed PM |
2333 | /* |
2334 | * If called from an extended quiescent state, invoke the RCU | |
2335 | * core in order to force a re-evaluation of RCU's idleness. | |
2336 | */ | |
5c173eb8 | 2337 | if (!rcu_is_watching() && cpu_online(smp_processor_id())) |
62fde6ed PM |
2338 | invoke_rcu_core(); |
2339 | ||
a16b7a69 | 2340 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2341 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2342 | return; |
64db4cff | 2343 | |
37c72e56 PM |
2344 | /* |
2345 | * Force the grace period if too many callbacks or too long waiting. | |
2346 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2347 | * if some other CPU has recently done so. Also, don't bother | |
2348 | * invoking force_quiescent_state() if the newly enqueued callback | |
2349 | * is the only one waiting for a grace period to complete. | |
2350 | */ | |
2655d57e | 2351 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2352 | |
2353 | /* Are we ignoring a completed grace period? */ | |
470716fc | 2354 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
2355 | |
2356 | /* Start a new grace period if one not already started. */ | |
2357 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
2358 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
2359 | ||
b8462084 PM |
2360 | raw_spin_lock(&rnp_root->lock); |
2361 | rcu_start_gp(rsp); | |
2362 | raw_spin_unlock(&rnp_root->lock); | |
b52573d2 PM |
2363 | } else { |
2364 | /* Give the grace period a kick. */ | |
2365 | rdp->blimit = LONG_MAX; | |
2366 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2367 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2368 | force_quiescent_state(rsp); |
b52573d2 PM |
2369 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2370 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2371 | } | |
4cdfc175 | 2372 | } |
29154c57 PM |
2373 | } |
2374 | ||
ae150184 PM |
2375 | /* |
2376 | * RCU callback function to leak a callback. | |
2377 | */ | |
2378 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2379 | { | |
2380 | } | |
2381 | ||
3fbfbf7a PM |
2382 | /* |
2383 | * Helper function for call_rcu() and friends. The cpu argument will | |
2384 | * normally be -1, indicating "currently running CPU". It may specify | |
2385 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
2386 | * is expected to specify a CPU. | |
2387 | */ | |
64db4cff PM |
2388 | static void |
2389 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), | |
3fbfbf7a | 2390 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
2391 | { |
2392 | unsigned long flags; | |
2393 | struct rcu_data *rdp; | |
2394 | ||
0bb7b59d | 2395 | WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */ |
ae150184 PM |
2396 | if (debug_rcu_head_queue(head)) { |
2397 | /* Probable double call_rcu(), so leak the callback. */ | |
2398 | ACCESS_ONCE(head->func) = rcu_leak_callback; | |
2399 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); | |
2400 | return; | |
2401 | } | |
64db4cff PM |
2402 | head->func = func; |
2403 | head->next = NULL; | |
2404 | ||
64db4cff PM |
2405 | /* |
2406 | * Opportunistically note grace-period endings and beginnings. | |
2407 | * Note that we might see a beginning right after we see an | |
2408 | * end, but never vice versa, since this CPU has to pass through | |
2409 | * a quiescent state betweentimes. | |
2410 | */ | |
2411 | local_irq_save(flags); | |
394f99a9 | 2412 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
2413 | |
2414 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
2415 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
2416 | int offline; | |
2417 | ||
2418 | if (cpu != -1) | |
2419 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
2420 | offline = !__call_rcu_nocb(rdp, head, lazy); | |
2421 | WARN_ON_ONCE(offline); | |
0d8ee37e | 2422 | /* _call_rcu() is illegal on offline CPU; leak the callback. */ |
0d8ee37e PM |
2423 | local_irq_restore(flags); |
2424 | return; | |
2425 | } | |
29154c57 | 2426 | ACCESS_ONCE(rdp->qlen)++; |
486e2593 PM |
2427 | if (lazy) |
2428 | rdp->qlen_lazy++; | |
c57afe80 PM |
2429 | else |
2430 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2431 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
2432 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
2433 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 2434 | |
d4c08f2a PM |
2435 | if (__is_kfree_rcu_offset((unsigned long)func)) |
2436 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 2437 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2438 | else |
486e2593 | 2439 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 2440 | |
29154c57 PM |
2441 | /* Go handle any RCU core processing required. */ |
2442 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
2443 | local_irq_restore(flags); |
2444 | } | |
2445 | ||
2446 | /* | |
d6714c22 | 2447 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 2448 | */ |
d6714c22 | 2449 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
64db4cff | 2450 | { |
3fbfbf7a | 2451 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 2452 | } |
d6714c22 | 2453 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
2454 | |
2455 | /* | |
486e2593 | 2456 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff PM |
2457 | */ |
2458 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
2459 | { | |
3fbfbf7a | 2460 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
2461 | } |
2462 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
2463 | ||
6d813391 PM |
2464 | /* |
2465 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
2466 | * any blocking grace-period wait automatically implies a grace period | |
2467 | * if there is only one CPU online at any point time during execution | |
2468 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
2469 | * occasionally incorrectly indicate that there are multiple CPUs online | |
2470 | * when there was in fact only one the whole time, as this just adds | |
2471 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
2472 | */ |
2473 | static inline int rcu_blocking_is_gp(void) | |
2474 | { | |
95f0c1de PM |
2475 | int ret; |
2476 | ||
6d813391 | 2477 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
2478 | preempt_disable(); |
2479 | ret = num_online_cpus() <= 1; | |
2480 | preempt_enable(); | |
2481 | return ret; | |
6d813391 PM |
2482 | } |
2483 | ||
6ebb237b PM |
2484 | /** |
2485 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
2486 | * | |
2487 | * Control will return to the caller some time after a full rcu-sched | |
2488 | * grace period has elapsed, in other words after all currently executing | |
2489 | * rcu-sched read-side critical sections have completed. These read-side | |
2490 | * critical sections are delimited by rcu_read_lock_sched() and | |
2491 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
2492 | * local_irq_disable(), and so on may be used in place of | |
2493 | * rcu_read_lock_sched(). | |
2494 | * | |
2495 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
2496 | * non-threaded hardware-interrupt handlers, in progress on entry will |
2497 | * have completed before this primitive returns. However, this does not | |
2498 | * guarantee that softirq handlers will have completed, since in some | |
2499 | * kernels, these handlers can run in process context, and can block. | |
2500 | * | |
2501 | * Note that this guarantee implies further memory-ordering guarantees. | |
2502 | * On systems with more than one CPU, when synchronize_sched() returns, | |
2503 | * each CPU is guaranteed to have executed a full memory barrier since the | |
2504 | * end of its last RCU-sched read-side critical section whose beginning | |
2505 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
2506 | * an RCU read-side critical section that extends beyond the return from | |
2507 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
2508 | * after the beginning of synchronize_sched() and before the beginning of | |
2509 | * that RCU read-side critical section. Note that these guarantees include | |
2510 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
2511 | * that are executing in the kernel. | |
2512 | * | |
2513 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
2514 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
2515 | * to have executed a full memory barrier during the execution of | |
2516 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
2517 | * again only if the system has more than one CPU). | |
6ebb237b PM |
2518 | * |
2519 | * This primitive provides the guarantees made by the (now removed) | |
2520 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
2521 | * guarantees that rcu_read_lock() sections will have completed. | |
2522 | * In "classic RCU", these two guarantees happen to be one and | |
2523 | * the same, but can differ in realtime RCU implementations. | |
2524 | */ | |
2525 | void synchronize_sched(void) | |
2526 | { | |
fe15d706 PM |
2527 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2528 | !lock_is_held(&rcu_lock_map) && | |
2529 | !lock_is_held(&rcu_sched_lock_map), | |
2530 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
2531 | if (rcu_blocking_is_gp()) |
2532 | return; | |
3705b88d AM |
2533 | if (rcu_expedited) |
2534 | synchronize_sched_expedited(); | |
2535 | else | |
2536 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
2537 | } |
2538 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
2539 | ||
2540 | /** | |
2541 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
2542 | * | |
2543 | * Control will return to the caller some time after a full rcu_bh grace | |
2544 | * period has elapsed, in other words after all currently executing rcu_bh | |
2545 | * read-side critical sections have completed. RCU read-side critical | |
2546 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
2547 | * and may be nested. | |
f0a0e6f2 PM |
2548 | * |
2549 | * See the description of synchronize_sched() for more detailed information | |
2550 | * on memory ordering guarantees. | |
6ebb237b PM |
2551 | */ |
2552 | void synchronize_rcu_bh(void) | |
2553 | { | |
fe15d706 PM |
2554 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
2555 | !lock_is_held(&rcu_lock_map) && | |
2556 | !lock_is_held(&rcu_sched_lock_map), | |
2557 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
2558 | if (rcu_blocking_is_gp()) |
2559 | return; | |
3705b88d AM |
2560 | if (rcu_expedited) |
2561 | synchronize_rcu_bh_expedited(); | |
2562 | else | |
2563 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
2564 | } |
2565 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
2566 | ||
3d3b7db0 PM |
2567 | static int synchronize_sched_expedited_cpu_stop(void *data) |
2568 | { | |
2569 | /* | |
2570 | * There must be a full memory barrier on each affected CPU | |
2571 | * between the time that try_stop_cpus() is called and the | |
2572 | * time that it returns. | |
2573 | * | |
2574 | * In the current initial implementation of cpu_stop, the | |
2575 | * above condition is already met when the control reaches | |
2576 | * this point and the following smp_mb() is not strictly | |
2577 | * necessary. Do smp_mb() anyway for documentation and | |
2578 | * robustness against future implementation changes. | |
2579 | */ | |
2580 | smp_mb(); /* See above comment block. */ | |
2581 | return 0; | |
2582 | } | |
2583 | ||
236fefaf PM |
2584 | /** |
2585 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
2586 | * | |
2587 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
2588 | * approach to force the grace period to end quickly. This consumes | |
2589 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
2590 | * so is thus not recommended for any sort of common-case code. In fact, | |
2591 | * if you are using synchronize_sched_expedited() in a loop, please | |
2592 | * restructure your code to batch your updates, and then use a single | |
2593 | * synchronize_sched() instead. | |
3d3b7db0 | 2594 | * |
236fefaf PM |
2595 | * Note that it is illegal to call this function while holding any lock |
2596 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal | |
2597 | * to call this function from a CPU-hotplug notifier. Failing to observe | |
2598 | * these restriction will result in deadlock. | |
3d3b7db0 PM |
2599 | * |
2600 | * This implementation can be thought of as an application of ticket | |
2601 | * locking to RCU, with sync_sched_expedited_started and | |
2602 | * sync_sched_expedited_done taking on the roles of the halves | |
2603 | * of the ticket-lock word. Each task atomically increments | |
2604 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
2605 | * then attempts to stop all the CPUs. If this succeeds, then each | |
2606 | * CPU will have executed a context switch, resulting in an RCU-sched | |
2607 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
2608 | * update sync_sched_expedited_done to match our snapshot -- but | |
2609 | * only if someone else has not already advanced past our snapshot. | |
2610 | * | |
2611 | * On the other hand, if try_stop_cpus() fails, we check the value | |
2612 | * of sync_sched_expedited_done. If it has advanced past our | |
2613 | * initial snapshot, then someone else must have forced a grace period | |
2614 | * some time after we took our snapshot. In this case, our work is | |
2615 | * done for us, and we can simply return. Otherwise, we try again, | |
2616 | * but keep our initial snapshot for purposes of checking for someone | |
2617 | * doing our work for us. | |
2618 | * | |
2619 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
2620 | */ | |
2621 | void synchronize_sched_expedited(void) | |
2622 | { | |
1924bcb0 PM |
2623 | long firstsnap, s, snap; |
2624 | int trycount = 0; | |
40694d66 | 2625 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 2626 | |
1924bcb0 PM |
2627 | /* |
2628 | * If we are in danger of counter wrap, just do synchronize_sched(). | |
2629 | * By allowing sync_sched_expedited_started to advance no more than | |
2630 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring | |
2631 | * that more than 3.5 billion CPUs would be required to force a | |
2632 | * counter wrap on a 32-bit system. Quite a few more CPUs would of | |
2633 | * course be required on a 64-bit system. | |
2634 | */ | |
40694d66 PM |
2635 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
2636 | (ulong)atomic_long_read(&rsp->expedited_done) + | |
1924bcb0 PM |
2637 | ULONG_MAX / 8)) { |
2638 | synchronize_sched(); | |
a30489c5 | 2639 | atomic_long_inc(&rsp->expedited_wrap); |
1924bcb0 PM |
2640 | return; |
2641 | } | |
3d3b7db0 | 2642 | |
1924bcb0 PM |
2643 | /* |
2644 | * Take a ticket. Note that atomic_inc_return() implies a | |
2645 | * full memory barrier. | |
2646 | */ | |
40694d66 | 2647 | snap = atomic_long_inc_return(&rsp->expedited_start); |
1924bcb0 | 2648 | firstsnap = snap; |
3d3b7db0 | 2649 | get_online_cpus(); |
1cc85961 | 2650 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
3d3b7db0 PM |
2651 | |
2652 | /* | |
2653 | * Each pass through the following loop attempts to force a | |
2654 | * context switch on each CPU. | |
2655 | */ | |
2656 | while (try_stop_cpus(cpu_online_mask, | |
2657 | synchronize_sched_expedited_cpu_stop, | |
2658 | NULL) == -EAGAIN) { | |
2659 | put_online_cpus(); | |
a30489c5 | 2660 | atomic_long_inc(&rsp->expedited_tryfail); |
3d3b7db0 | 2661 | |
1924bcb0 | 2662 | /* Check to see if someone else did our work for us. */ |
40694d66 | 2663 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2664 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2665 | /* ensure test happens before caller kfree */ |
2666 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2667 | atomic_long_inc(&rsp->expedited_workdone1); | |
1924bcb0 PM |
2668 | return; |
2669 | } | |
3d3b7db0 PM |
2670 | |
2671 | /* No joy, try again later. Or just synchronize_sched(). */ | |
c701d5d9 | 2672 | if (trycount++ < 10) { |
3d3b7db0 | 2673 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 2674 | } else { |
3705b88d | 2675 | wait_rcu_gp(call_rcu_sched); |
a30489c5 | 2676 | atomic_long_inc(&rsp->expedited_normal); |
3d3b7db0 PM |
2677 | return; |
2678 | } | |
2679 | ||
1924bcb0 | 2680 | /* Recheck to see if someone else did our work for us. */ |
40694d66 | 2681 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2682 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
a30489c5 PM |
2683 | /* ensure test happens before caller kfree */ |
2684 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2685 | atomic_long_inc(&rsp->expedited_workdone2); | |
3d3b7db0 PM |
2686 | return; |
2687 | } | |
2688 | ||
2689 | /* | |
2690 | * Refetching sync_sched_expedited_started allows later | |
1924bcb0 PM |
2691 | * callers to piggyback on our grace period. We retry |
2692 | * after they started, so our grace period works for them, | |
2693 | * and they started after our first try, so their grace | |
2694 | * period works for us. | |
3d3b7db0 PM |
2695 | */ |
2696 | get_online_cpus(); | |
40694d66 | 2697 | snap = atomic_long_read(&rsp->expedited_start); |
3d3b7db0 PM |
2698 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
2699 | } | |
a30489c5 | 2700 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
3d3b7db0 PM |
2701 | |
2702 | /* | |
2703 | * Everyone up to our most recent fetch is covered by our grace | |
2704 | * period. Update the counter, but only if our work is still | |
2705 | * relevant -- which it won't be if someone who started later | |
1924bcb0 | 2706 | * than we did already did their update. |
3d3b7db0 PM |
2707 | */ |
2708 | do { | |
a30489c5 | 2709 | atomic_long_inc(&rsp->expedited_done_tries); |
40694d66 | 2710 | s = atomic_long_read(&rsp->expedited_done); |
1924bcb0 | 2711 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
a30489c5 PM |
2712 | /* ensure test happens before caller kfree */ |
2713 | smp_mb__before_atomic_inc(); /* ^^^ */ | |
2714 | atomic_long_inc(&rsp->expedited_done_lost); | |
3d3b7db0 PM |
2715 | break; |
2716 | } | |
40694d66 | 2717 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
a30489c5 | 2718 | atomic_long_inc(&rsp->expedited_done_exit); |
3d3b7db0 PM |
2719 | |
2720 | put_online_cpus(); | |
2721 | } | |
2722 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
2723 | ||
64db4cff PM |
2724 | /* |
2725 | * Check to see if there is any immediate RCU-related work to be done | |
2726 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
2727 | * The checks are in order of increasing expense: checks that can be | |
2728 | * carried out against CPU-local state are performed first. However, | |
2729 | * we must check for CPU stalls first, else we might not get a chance. | |
2730 | */ | |
2731 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
2732 | { | |
2f51f988 PM |
2733 | struct rcu_node *rnp = rdp->mynode; |
2734 | ||
64db4cff PM |
2735 | rdp->n_rcu_pending++; |
2736 | ||
2737 | /* Check for CPU stalls, if enabled. */ | |
2738 | check_cpu_stall(rsp, rdp); | |
2739 | ||
2740 | /* Is the RCU core waiting for a quiescent state from this CPU? */ | |
5c51dd73 PM |
2741 | if (rcu_scheduler_fully_active && |
2742 | rdp->qs_pending && !rdp->passed_quiesce) { | |
d21670ac | 2743 | rdp->n_rp_qs_pending++; |
e4cc1f22 | 2744 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
d21670ac | 2745 | rdp->n_rp_report_qs++; |
64db4cff | 2746 | return 1; |
7ba5c840 | 2747 | } |
64db4cff PM |
2748 | |
2749 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
2750 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
2751 | rdp->n_rp_cb_ready++; | |
64db4cff | 2752 | return 1; |
7ba5c840 | 2753 | } |
64db4cff PM |
2754 | |
2755 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
2756 | if (cpu_needs_another_gp(rsp, rdp)) { |
2757 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 2758 | return 1; |
7ba5c840 | 2759 | } |
64db4cff PM |
2760 | |
2761 | /* Has another RCU grace period completed? */ | |
2f51f988 | 2762 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 2763 | rdp->n_rp_gp_completed++; |
64db4cff | 2764 | return 1; |
7ba5c840 | 2765 | } |
64db4cff PM |
2766 | |
2767 | /* Has a new RCU grace period started? */ | |
2f51f988 | 2768 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
7ba5c840 | 2769 | rdp->n_rp_gp_started++; |
64db4cff | 2770 | return 1; |
7ba5c840 | 2771 | } |
64db4cff | 2772 | |
64db4cff | 2773 | /* nothing to do */ |
7ba5c840 | 2774 | rdp->n_rp_need_nothing++; |
64db4cff PM |
2775 | return 0; |
2776 | } | |
2777 | ||
2778 | /* | |
2779 | * Check to see if there is any immediate RCU-related work to be done | |
2780 | * by the current CPU, returning 1 if so. This function is part of the | |
2781 | * RCU implementation; it is -not- an exported member of the RCU API. | |
2782 | */ | |
a157229c | 2783 | static int rcu_pending(int cpu) |
64db4cff | 2784 | { |
6ce75a23 PM |
2785 | struct rcu_state *rsp; |
2786 | ||
2787 | for_each_rcu_flavor(rsp) | |
2788 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu))) | |
2789 | return 1; | |
2790 | return 0; | |
64db4cff PM |
2791 | } |
2792 | ||
2793 | /* | |
c0f4dfd4 PM |
2794 | * Return true if the specified CPU has any callback. If all_lazy is |
2795 | * non-NULL, store an indication of whether all callbacks are lazy. | |
2796 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 2797 | */ |
c0f4dfd4 | 2798 | static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy) |
64db4cff | 2799 | { |
c0f4dfd4 PM |
2800 | bool al = true; |
2801 | bool hc = false; | |
2802 | struct rcu_data *rdp; | |
6ce75a23 PM |
2803 | struct rcu_state *rsp; |
2804 | ||
c0f4dfd4 PM |
2805 | for_each_rcu_flavor(rsp) { |
2806 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
69c8d28c PM |
2807 | if (!rdp->nxtlist) |
2808 | continue; | |
2809 | hc = true; | |
2810 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
c0f4dfd4 | 2811 | al = false; |
69c8d28c PM |
2812 | break; |
2813 | } | |
c0f4dfd4 PM |
2814 | } |
2815 | if (all_lazy) | |
2816 | *all_lazy = al; | |
2817 | return hc; | |
64db4cff PM |
2818 | } |
2819 | ||
a83eff0a PM |
2820 | /* |
2821 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
2822 | * the compiler is expected to optimize this away. | |
2823 | */ | |
e66c33d5 | 2824 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
2825 | int cpu, unsigned long done) |
2826 | { | |
2827 | trace_rcu_barrier(rsp->name, s, cpu, | |
2828 | atomic_read(&rsp->barrier_cpu_count), done); | |
2829 | } | |
2830 | ||
b1420f1c PM |
2831 | /* |
2832 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
2833 | * up the task executing _rcu_barrier(). | |
2834 | */ | |
24ebbca8 | 2835 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 2836 | { |
24ebbca8 PM |
2837 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
2838 | struct rcu_state *rsp = rdp->rsp; | |
2839 | ||
a83eff0a PM |
2840 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
2841 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); | |
7db74df8 | 2842 | complete(&rsp->barrier_completion); |
a83eff0a PM |
2843 | } else { |
2844 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); | |
2845 | } | |
d0ec774c PM |
2846 | } |
2847 | ||
2848 | /* | |
2849 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
2850 | */ | |
2851 | static void rcu_barrier_func(void *type) | |
2852 | { | |
037b64ed | 2853 | struct rcu_state *rsp = type; |
06668efa | 2854 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
d0ec774c | 2855 | |
a83eff0a | 2856 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
24ebbca8 | 2857 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 2858 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
2859 | } |
2860 | ||
d0ec774c PM |
2861 | /* |
2862 | * Orchestrate the specified type of RCU barrier, waiting for all | |
2863 | * RCU callbacks of the specified type to complete. | |
2864 | */ | |
037b64ed | 2865 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 2866 | { |
b1420f1c | 2867 | int cpu; |
b1420f1c | 2868 | struct rcu_data *rdp; |
cf3a9c48 PM |
2869 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
2870 | unsigned long snap_done; | |
b1420f1c | 2871 | |
a83eff0a | 2872 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
b1420f1c | 2873 | |
e74f4c45 | 2874 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 2875 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 2876 | |
cf3a9c48 PM |
2877 | /* |
2878 | * Ensure that all prior references, including to ->n_barrier_done, | |
2879 | * are ordered before the _rcu_barrier() machinery. | |
2880 | */ | |
2881 | smp_mb(); /* See above block comment. */ | |
2882 | ||
2883 | /* | |
2884 | * Recheck ->n_barrier_done to see if others did our work for us. | |
2885 | * This means checking ->n_barrier_done for an even-to-odd-to-even | |
2886 | * transition. The "if" expression below therefore rounds the old | |
2887 | * value up to the next even number and adds two before comparing. | |
2888 | */ | |
458fb381 | 2889 | snap_done = rsp->n_barrier_done; |
a83eff0a | 2890 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
458fb381 PM |
2891 | |
2892 | /* | |
2893 | * If the value in snap is odd, we needed to wait for the current | |
2894 | * rcu_barrier() to complete, then wait for the next one, in other | |
2895 | * words, we need the value of snap_done to be three larger than | |
2896 | * the value of snap. On the other hand, if the value in snap is | |
2897 | * even, we only had to wait for the next rcu_barrier() to complete, | |
2898 | * in other words, we need the value of snap_done to be only two | |
2899 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes | |
2900 | * this for us (thank you, Linus!). | |
2901 | */ | |
2902 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) { | |
a83eff0a | 2903 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
cf3a9c48 PM |
2904 | smp_mb(); /* caller's subsequent code after above check. */ |
2905 | mutex_unlock(&rsp->barrier_mutex); | |
2906 | return; | |
2907 | } | |
2908 | ||
2909 | /* | |
2910 | * Increment ->n_barrier_done to avoid duplicate work. Use | |
2911 | * ACCESS_ONCE() to prevent the compiler from speculating | |
2912 | * the increment to precede the early-exit check. | |
2913 | */ | |
2914 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2915 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); | |
a83eff0a | 2916 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
cf3a9c48 | 2917 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
b1420f1c | 2918 | |
d0ec774c | 2919 | /* |
b1420f1c PM |
2920 | * Initialize the count to one rather than to zero in order to |
2921 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
2922 | * (or preemption of this task). Exclude CPU-hotplug operations |
2923 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 2924 | */ |
7db74df8 | 2925 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 2926 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 2927 | get_online_cpus(); |
b1420f1c PM |
2928 | |
2929 | /* | |
1331e7a1 PM |
2930 | * Force each CPU with callbacks to register a new callback. |
2931 | * When that callback is invoked, we will know that all of the | |
2932 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 2933 | */ |
3fbfbf7a | 2934 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 2935 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 2936 | continue; |
b1420f1c | 2937 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 2938 | if (rcu_is_nocb_cpu(cpu)) { |
3fbfbf7a PM |
2939 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, |
2940 | rsp->n_barrier_done); | |
2941 | atomic_inc(&rsp->barrier_cpu_count); | |
2942 | __call_rcu(&rdp->barrier_head, rcu_barrier_callback, | |
2943 | rsp, cpu, 0); | |
2944 | } else if (ACCESS_ONCE(rdp->qlen)) { | |
a83eff0a PM |
2945 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
2946 | rsp->n_barrier_done); | |
037b64ed | 2947 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 2948 | } else { |
a83eff0a PM |
2949 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
2950 | rsp->n_barrier_done); | |
b1420f1c PM |
2951 | } |
2952 | } | |
1331e7a1 | 2953 | put_online_cpus(); |
b1420f1c PM |
2954 | |
2955 | /* | |
2956 | * Now that we have an rcu_barrier_callback() callback on each | |
2957 | * CPU, and thus each counted, remove the initial count. | |
2958 | */ | |
24ebbca8 | 2959 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 2960 | complete(&rsp->barrier_completion); |
b1420f1c | 2961 | |
cf3a9c48 PM |
2962 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
2963 | smp_mb(); /* Keep increment after above mechanism. */ | |
2964 | ACCESS_ONCE(rsp->n_barrier_done)++; | |
2965 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); | |
a83eff0a | 2966 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
cf3a9c48 PM |
2967 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
2968 | ||
b1420f1c | 2969 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
7db74df8 | 2970 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c PM |
2971 | |
2972 | /* Other rcu_barrier() invocations can now safely proceed. */ | |
7be7f0be | 2973 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 2974 | } |
d0ec774c PM |
2975 | |
2976 | /** | |
2977 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
2978 | */ | |
2979 | void rcu_barrier_bh(void) | |
2980 | { | |
037b64ed | 2981 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
2982 | } |
2983 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
2984 | ||
2985 | /** | |
2986 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
2987 | */ | |
2988 | void rcu_barrier_sched(void) | |
2989 | { | |
037b64ed | 2990 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
2991 | } |
2992 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
2993 | ||
64db4cff | 2994 | /* |
27569620 | 2995 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 2996 | */ |
27569620 PM |
2997 | static void __init |
2998 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
2999 | { |
3000 | unsigned long flags; | |
394f99a9 | 3001 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
3002 | struct rcu_node *rnp = rcu_get_root(rsp); |
3003 | ||
3004 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 3005 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 3006 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
3f5d3ea6 | 3007 | init_callback_list(rdp); |
486e2593 | 3008 | rdp->qlen_lazy = 0; |
1d1fb395 | 3009 | ACCESS_ONCE(rdp->qlen) = 0; |
27569620 | 3010 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 3011 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 3012 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 3013 | rdp->cpu = cpu; |
d4c08f2a | 3014 | rdp->rsp = rsp; |
3fbfbf7a | 3015 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 3016 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
3017 | } |
3018 | ||
3019 | /* | |
3020 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
3021 | * offline event can be happening at a given time. Note also that we | |
3022 | * can accept some slop in the rsp->completed access due to the fact | |
3023 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 3024 | */ |
49fb4c62 | 3025 | static void |
6cc68793 | 3026 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible) |
64db4cff PM |
3027 | { |
3028 | unsigned long flags; | |
64db4cff | 3029 | unsigned long mask; |
394f99a9 | 3030 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
3031 | struct rcu_node *rnp = rcu_get_root(rsp); |
3032 | ||
a4fbe35a PM |
3033 | /* Exclude new grace periods. */ |
3034 | mutex_lock(&rsp->onoff_mutex); | |
3035 | ||
64db4cff | 3036 | /* Set up local state, ensuring consistent view of global state. */ |
1304afb2 | 3037 | raw_spin_lock_irqsave(&rnp->lock, flags); |
64db4cff | 3038 | rdp->beenonline = 1; /* We have now been online. */ |
6cc68793 | 3039 | rdp->preemptible = preemptible; |
37c72e56 PM |
3040 | rdp->qlen_last_fqs_check = 0; |
3041 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 3042 | rdp->blimit = blimit; |
0d8ee37e | 3043 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ |
29e37d81 | 3044 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 3045 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
3046 | atomic_set(&rdp->dynticks->dynticks, |
3047 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 3048 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 3049 | |
64db4cff PM |
3050 | /* Add CPU to rcu_node bitmasks. */ |
3051 | rnp = rdp->mynode; | |
3052 | mask = rdp->grpmask; | |
3053 | do { | |
3054 | /* Exclude any attempts to start a new GP on small systems. */ | |
1304afb2 | 3055 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3056 | rnp->qsmaskinit |= mask; |
3057 | mask = rnp->grpmask; | |
d09b62df | 3058 | if (rnp == rdp->mynode) { |
06ae115a PM |
3059 | /* |
3060 | * If there is a grace period in progress, we will | |
3061 | * set up to wait for it next time we run the | |
3062 | * RCU core code. | |
3063 | */ | |
3064 | rdp->gpnum = rnp->completed; | |
d09b62df | 3065 | rdp->completed = rnp->completed; |
06ae115a PM |
3066 | rdp->passed_quiesce = 0; |
3067 | rdp->qs_pending = 0; | |
f7f7bac9 | 3068 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); |
d09b62df | 3069 | } |
1304afb2 | 3070 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
64db4cff PM |
3071 | rnp = rnp->parent; |
3072 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); | |
a4fbe35a | 3073 | local_irq_restore(flags); |
64db4cff | 3074 | |
a4fbe35a | 3075 | mutex_unlock(&rsp->onoff_mutex); |
64db4cff PM |
3076 | } |
3077 | ||
49fb4c62 | 3078 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 3079 | { |
6ce75a23 PM |
3080 | struct rcu_state *rsp; |
3081 | ||
3082 | for_each_rcu_flavor(rsp) | |
3083 | rcu_init_percpu_data(cpu, rsp, | |
3084 | strcmp(rsp->name, "rcu_preempt") == 0); | |
64db4cff PM |
3085 | } |
3086 | ||
3087 | /* | |
f41d911f | 3088 | * Handle CPU online/offline notification events. |
64db4cff | 3089 | */ |
49fb4c62 | 3090 | static int rcu_cpu_notify(struct notifier_block *self, |
9f680ab4 | 3091 | unsigned long action, void *hcpu) |
64db4cff PM |
3092 | { |
3093 | long cpu = (long)hcpu; | |
27f4d280 | 3094 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); |
a26ac245 | 3095 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 3096 | struct rcu_state *rsp; |
64db4cff | 3097 | |
f7f7bac9 | 3098 | trace_rcu_utilization(TPS("Start CPU hotplug")); |
64db4cff PM |
3099 | switch (action) { |
3100 | case CPU_UP_PREPARE: | |
3101 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
3102 | rcu_prepare_cpu(cpu); |
3103 | rcu_prepare_kthreads(cpu); | |
a26ac245 PM |
3104 | break; |
3105 | case CPU_ONLINE: | |
0f962a5e | 3106 | case CPU_DOWN_FAILED: |
5d01bbd1 | 3107 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
3108 | break; |
3109 | case CPU_DOWN_PREPARE: | |
34ed6246 | 3110 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 3111 | break; |
d0ec774c PM |
3112 | case CPU_DYING: |
3113 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
3114 | for_each_rcu_flavor(rsp) |
3115 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 3116 | break; |
64db4cff PM |
3117 | case CPU_DEAD: |
3118 | case CPU_DEAD_FROZEN: | |
3119 | case CPU_UP_CANCELED: | |
3120 | case CPU_UP_CANCELED_FROZEN: | |
6ce75a23 PM |
3121 | for_each_rcu_flavor(rsp) |
3122 | rcu_cleanup_dead_cpu(cpu, rsp); | |
64db4cff PM |
3123 | break; |
3124 | default: | |
3125 | break; | |
3126 | } | |
f7f7bac9 | 3127 | trace_rcu_utilization(TPS("End CPU hotplug")); |
34ed6246 | 3128 | return NOTIFY_OK; |
64db4cff PM |
3129 | } |
3130 | ||
d1d74d14 BP |
3131 | static int rcu_pm_notify(struct notifier_block *self, |
3132 | unsigned long action, void *hcpu) | |
3133 | { | |
3134 | switch (action) { | |
3135 | case PM_HIBERNATION_PREPARE: | |
3136 | case PM_SUSPEND_PREPARE: | |
3137 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
3138 | rcu_expedited = 1; | |
3139 | break; | |
3140 | case PM_POST_HIBERNATION: | |
3141 | case PM_POST_SUSPEND: | |
3142 | rcu_expedited = 0; | |
3143 | break; | |
3144 | default: | |
3145 | break; | |
3146 | } | |
3147 | return NOTIFY_OK; | |
3148 | } | |
3149 | ||
b3dbec76 PM |
3150 | /* |
3151 | * Spawn the kthread that handles this RCU flavor's grace periods. | |
3152 | */ | |
3153 | static int __init rcu_spawn_gp_kthread(void) | |
3154 | { | |
3155 | unsigned long flags; | |
3156 | struct rcu_node *rnp; | |
3157 | struct rcu_state *rsp; | |
3158 | struct task_struct *t; | |
3159 | ||
3160 | for_each_rcu_flavor(rsp) { | |
f170168b | 3161 | t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
3162 | BUG_ON(IS_ERR(t)); |
3163 | rnp = rcu_get_root(rsp); | |
3164 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3165 | rsp->gp_kthread = t; | |
3166 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3fbfbf7a | 3167 | rcu_spawn_nocb_kthreads(rsp); |
b3dbec76 PM |
3168 | } |
3169 | return 0; | |
3170 | } | |
3171 | early_initcall(rcu_spawn_gp_kthread); | |
3172 | ||
bbad9379 PM |
3173 | /* |
3174 | * This function is invoked towards the end of the scheduler's initialization | |
3175 | * process. Before this is called, the idle task might contain | |
3176 | * RCU read-side critical sections (during which time, this idle | |
3177 | * task is booting the system). After this function is called, the | |
3178 | * idle tasks are prohibited from containing RCU read-side critical | |
3179 | * sections. This function also enables RCU lockdep checking. | |
3180 | */ | |
3181 | void rcu_scheduler_starting(void) | |
3182 | { | |
3183 | WARN_ON(num_online_cpus() != 1); | |
3184 | WARN_ON(nr_context_switches() > 0); | |
3185 | rcu_scheduler_active = 1; | |
3186 | } | |
3187 | ||
64db4cff PM |
3188 | /* |
3189 | * Compute the per-level fanout, either using the exact fanout specified | |
3190 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. | |
3191 | */ | |
3192 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
3193 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3194 | { | |
3195 | int i; | |
3196 | ||
f885b7f2 | 3197 | for (i = rcu_num_lvls - 1; i > 0; i--) |
64db4cff | 3198 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
f885b7f2 | 3199 | rsp->levelspread[0] = rcu_fanout_leaf; |
64db4cff PM |
3200 | } |
3201 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3202 | static void __init rcu_init_levelspread(struct rcu_state *rsp) | |
3203 | { | |
3204 | int ccur; | |
3205 | int cprv; | |
3206 | int i; | |
3207 | ||
4dbd6bb3 | 3208 | cprv = nr_cpu_ids; |
f885b7f2 | 3209 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3210 | ccur = rsp->levelcnt[i]; |
3211 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; | |
3212 | cprv = ccur; | |
3213 | } | |
3214 | } | |
3215 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ | |
3216 | ||
3217 | /* | |
3218 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
3219 | */ | |
394f99a9 LJ |
3220 | static void __init rcu_init_one(struct rcu_state *rsp, |
3221 | struct rcu_data __percpu *rda) | |
64db4cff | 3222 | { |
394f2769 PM |
3223 | static char *buf[] = { "rcu_node_0", |
3224 | "rcu_node_1", | |
3225 | "rcu_node_2", | |
3226 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ | |
3227 | static char *fqs[] = { "rcu_node_fqs_0", | |
3228 | "rcu_node_fqs_1", | |
3229 | "rcu_node_fqs_2", | |
3230 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ | |
64db4cff PM |
3231 | int cpustride = 1; |
3232 | int i; | |
3233 | int j; | |
3234 | struct rcu_node *rnp; | |
3235 | ||
b6407e86 PM |
3236 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
3237 | ||
4930521a PM |
3238 | /* Silence gcc 4.8 warning about array index out of range. */ |
3239 | if (rcu_num_lvls > RCU_NUM_LVLS) | |
3240 | panic("rcu_init_one: rcu_num_lvls overflow"); | |
3241 | ||
64db4cff PM |
3242 | /* Initialize the level-tracking arrays. */ |
3243 | ||
f885b7f2 PM |
3244 | for (i = 0; i < rcu_num_lvls; i++) |
3245 | rsp->levelcnt[i] = num_rcu_lvl[i]; | |
3246 | for (i = 1; i < rcu_num_lvls; i++) | |
64db4cff PM |
3247 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
3248 | rcu_init_levelspread(rsp); | |
3249 | ||
3250 | /* Initialize the elements themselves, starting from the leaves. */ | |
3251 | ||
f885b7f2 | 3252 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
64db4cff PM |
3253 | cpustride *= rsp->levelspread[i]; |
3254 | rnp = rsp->level[i]; | |
3255 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { | |
1304afb2 | 3256 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
3257 | lockdep_set_class_and_name(&rnp->lock, |
3258 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
3259 | raw_spin_lock_init(&rnp->fqslock); |
3260 | lockdep_set_class_and_name(&rnp->fqslock, | |
3261 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
3262 | rnp->gpnum = rsp->gpnum; |
3263 | rnp->completed = rsp->completed; | |
64db4cff PM |
3264 | rnp->qsmask = 0; |
3265 | rnp->qsmaskinit = 0; | |
3266 | rnp->grplo = j * cpustride; | |
3267 | rnp->grphi = (j + 1) * cpustride - 1; | |
3268 | if (rnp->grphi >= NR_CPUS) | |
3269 | rnp->grphi = NR_CPUS - 1; | |
3270 | if (i == 0) { | |
3271 | rnp->grpnum = 0; | |
3272 | rnp->grpmask = 0; | |
3273 | rnp->parent = NULL; | |
3274 | } else { | |
3275 | rnp->grpnum = j % rsp->levelspread[i - 1]; | |
3276 | rnp->grpmask = 1UL << rnp->grpnum; | |
3277 | rnp->parent = rsp->level[i - 1] + | |
3278 | j / rsp->levelspread[i - 1]; | |
3279 | } | |
3280 | rnp->level = i; | |
12f5f524 | 3281 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 3282 | rcu_init_one_nocb(rnp); |
64db4cff PM |
3283 | } |
3284 | } | |
0c34029a | 3285 | |
394f99a9 | 3286 | rsp->rda = rda; |
b3dbec76 | 3287 | init_waitqueue_head(&rsp->gp_wq); |
016a8d5b | 3288 | init_irq_work(&rsp->wakeup_work, rsp_wakeup); |
f885b7f2 | 3289 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 3290 | for_each_possible_cpu(i) { |
4a90a068 | 3291 | while (i > rnp->grphi) |
0c34029a | 3292 | rnp++; |
394f99a9 | 3293 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
3294 | rcu_boot_init_percpu_data(i, rsp); |
3295 | } | |
6ce75a23 | 3296 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
3297 | } |
3298 | ||
f885b7f2 PM |
3299 | /* |
3300 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
3301 | * replace the definitions in rcutree.h because those are needed to size | |
3302 | * the ->node array in the rcu_state structure. | |
3303 | */ | |
3304 | static void __init rcu_init_geometry(void) | |
3305 | { | |
026ad283 | 3306 | ulong d; |
f885b7f2 PM |
3307 | int i; |
3308 | int j; | |
cca6f393 | 3309 | int n = nr_cpu_ids; |
f885b7f2 PM |
3310 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
3311 | ||
026ad283 PM |
3312 | /* |
3313 | * Initialize any unspecified boot parameters. | |
3314 | * The default values of jiffies_till_first_fqs and | |
3315 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
3316 | * value, which is a function of HZ, then adding one for each | |
3317 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
3318 | */ | |
3319 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
3320 | if (jiffies_till_first_fqs == ULONG_MAX) | |
3321 | jiffies_till_first_fqs = d; | |
3322 | if (jiffies_till_next_fqs == ULONG_MAX) | |
3323 | jiffies_till_next_fqs = d; | |
3324 | ||
f885b7f2 | 3325 | /* If the compile-time values are accurate, just leave. */ |
b17c7035 PM |
3326 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
3327 | nr_cpu_ids == NR_CPUS) | |
f885b7f2 PM |
3328 | return; |
3329 | ||
3330 | /* | |
3331 | * Compute number of nodes that can be handled an rcu_node tree | |
3332 | * with the given number of levels. Setting rcu_capacity[0] makes | |
3333 | * some of the arithmetic easier. | |
3334 | */ | |
3335 | rcu_capacity[0] = 1; | |
3336 | rcu_capacity[1] = rcu_fanout_leaf; | |
3337 | for (i = 2; i <= MAX_RCU_LVLS; i++) | |
3338 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; | |
3339 | ||
3340 | /* | |
3341 | * The boot-time rcu_fanout_leaf parameter is only permitted | |
3342 | * to increase the leaf-level fanout, not decrease it. Of course, | |
3343 | * the leaf-level fanout cannot exceed the number of bits in | |
3344 | * the rcu_node masks. Finally, the tree must be able to accommodate | |
3345 | * the configured number of CPUs. Complain and fall back to the | |
3346 | * compile-time values if these limits are exceeded. | |
3347 | */ | |
3348 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || | |
3349 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || | |
3350 | n > rcu_capacity[MAX_RCU_LVLS]) { | |
3351 | WARN_ON(1); | |
3352 | return; | |
3353 | } | |
3354 | ||
3355 | /* Calculate the number of rcu_nodes at each level of the tree. */ | |
3356 | for (i = 1; i <= MAX_RCU_LVLS; i++) | |
3357 | if (n <= rcu_capacity[i]) { | |
3358 | for (j = 0; j <= i; j++) | |
3359 | num_rcu_lvl[j] = | |
3360 | DIV_ROUND_UP(n, rcu_capacity[i - j]); | |
3361 | rcu_num_lvls = i; | |
3362 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) | |
3363 | num_rcu_lvl[j] = 0; | |
3364 | break; | |
3365 | } | |
3366 | ||
3367 | /* Calculate the total number of rcu_node structures. */ | |
3368 | rcu_num_nodes = 0; | |
3369 | for (i = 0; i <= MAX_RCU_LVLS; i++) | |
3370 | rcu_num_nodes += num_rcu_lvl[i]; | |
3371 | rcu_num_nodes -= n; | |
3372 | } | |
3373 | ||
9f680ab4 | 3374 | void __init rcu_init(void) |
64db4cff | 3375 | { |
017c4261 | 3376 | int cpu; |
9f680ab4 | 3377 | |
f41d911f | 3378 | rcu_bootup_announce(); |
f885b7f2 | 3379 | rcu_init_geometry(); |
394f99a9 | 3380 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); |
69c8d28c | 3381 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
f41d911f | 3382 | __rcu_init_preempt(); |
b5b39360 | 3383 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
3384 | |
3385 | /* | |
3386 | * We don't need protection against CPU-hotplug here because | |
3387 | * this is called early in boot, before either interrupts | |
3388 | * or the scheduler are operational. | |
3389 | */ | |
3390 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 3391 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
3392 | for_each_online_cpu(cpu) |
3393 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
3394 | } |
3395 | ||
1eba8f84 | 3396 | #include "rcutree_plugin.h" |