]>
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 | |
87de1cfd PM |
15 | * along with this program; if not, you can access it online at |
16 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
64db4cff PM |
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> | |
4102adab | 44 | #include <linux/module.h> |
64db4cff PM |
45 | #include <linux/percpu.h> |
46 | #include <linux/notifier.h> | |
47 | #include <linux/cpu.h> | |
48 | #include <linux/mutex.h> | |
49 | #include <linux/time.h> | |
bbad9379 | 50 | #include <linux/kernel_stat.h> |
a26ac245 PM |
51 | #include <linux/wait.h> |
52 | #include <linux/kthread.h> | |
268bb0ce | 53 | #include <linux/prefetch.h> |
3d3b7db0 PM |
54 | #include <linux/delay.h> |
55 | #include <linux/stop_machine.h> | |
661a85dc | 56 | #include <linux/random.h> |
af658dca | 57 | #include <linux/trace_events.h> |
d1d74d14 | 58 | #include <linux/suspend.h> |
64db4cff | 59 | |
4102adab | 60 | #include "tree.h" |
29c00b4a | 61 | #include "rcu.h" |
9f77da9f | 62 | |
4102adab PM |
63 | MODULE_ALIAS("rcutree"); |
64 | #ifdef MODULE_PARAM_PREFIX | |
65 | #undef MODULE_PARAM_PREFIX | |
66 | #endif | |
67 | #define MODULE_PARAM_PREFIX "rcutree." | |
68 | ||
64db4cff PM |
69 | /* Data structures. */ |
70 | ||
f885b7f2 | 71 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
394f2769 | 72 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; |
385b73c0 | 73 | static struct lock_class_key rcu_exp_class[RCU_NUM_LVLS]; |
88b91c7c | 74 | |
f7f7bac9 SRRH |
75 | /* |
76 | * In order to export the rcu_state name to the tracing tools, it | |
77 | * needs to be added in the __tracepoint_string section. | |
78 | * This requires defining a separate variable tp_<sname>_varname | |
79 | * that points to the string being used, and this will allow | |
80 | * the tracing userspace tools to be able to decipher the string | |
81 | * address to the matching string. | |
82 | */ | |
a8a29b3b AB |
83 | #ifdef CONFIG_TRACING |
84 | # define DEFINE_RCU_TPS(sname) \ | |
f7f7bac9 | 85 | static char sname##_varname[] = #sname; \ |
a8a29b3b AB |
86 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; |
87 | # define RCU_STATE_NAME(sname) sname##_varname | |
88 | #else | |
89 | # define DEFINE_RCU_TPS(sname) | |
90 | # define RCU_STATE_NAME(sname) __stringify(sname) | |
91 | #endif | |
92 | ||
93 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ | |
94 | DEFINE_RCU_TPS(sname) \ | |
c92fb057 | 95 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \ |
a41bfeb2 | 96 | struct rcu_state sname##_state = { \ |
6c90cc7b | 97 | .level = { &sname##_state.node[0] }, \ |
2723249a | 98 | .rda = &sname##_data, \ |
037b64ed | 99 | .call = cr, \ |
77f81fe0 | 100 | .gp_state = RCU_GP_IDLE, \ |
42c3533e PM |
101 | .gpnum = 0UL - 300UL, \ |
102 | .completed = 0UL - 300UL, \ | |
7b2e6011 | 103 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
6c90cc7b PM |
104 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
105 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 106 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
a8a29b3b | 107 | .name = RCU_STATE_NAME(sname), \ |
a4889858 | 108 | .abbr = sabbr, \ |
2723249a | 109 | } |
64db4cff | 110 | |
a41bfeb2 SRRH |
111 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); |
112 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); | |
b1f77b05 | 113 | |
b28a7c01 | 114 | static struct rcu_state *const rcu_state_p; |
2927a689 | 115 | static struct rcu_data __percpu *const rcu_data_p; |
6ce75a23 | 116 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 117 | |
a3dc2948 PM |
118 | /* Dump rcu_node combining tree at boot to verify correct setup. */ |
119 | static bool dump_tree; | |
120 | module_param(dump_tree, bool, 0444); | |
7fa27001 PM |
121 | /* Control rcu_node-tree auto-balancing at boot time. */ |
122 | static bool rcu_fanout_exact; | |
123 | module_param(rcu_fanout_exact, bool, 0444); | |
47d631af PM |
124 | /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */ |
125 | static int rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
7e5c2dfb | 126 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 | 127 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
cb007102 AG |
128 | /* Number of rcu_nodes at specified level. */ |
129 | static int num_rcu_lvl[] = NUM_RCU_LVL_INIT; | |
f885b7f2 PM |
130 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ |
131 | ||
b0d30417 PM |
132 | /* |
133 | * The rcu_scheduler_active variable transitions from zero to one just | |
134 | * before the first task is spawned. So when this variable is zero, RCU | |
135 | * can assume that there is but one task, allowing RCU to (for example) | |
b44f6656 | 136 | * optimize synchronize_sched() to a simple barrier(). When this variable |
b0d30417 PM |
137 | * is one, RCU must actually do all the hard work required to detect real |
138 | * grace periods. This variable is also used to suppress boot-time false | |
139 | * positives from lockdep-RCU error checking. | |
140 | */ | |
bbad9379 PM |
141 | int rcu_scheduler_active __read_mostly; |
142 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
143 | ||
b0d30417 PM |
144 | /* |
145 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
146 | * during the early_initcall() processing, which is after the scheduler | |
147 | * is capable of creating new tasks. So RCU processing (for example, | |
148 | * creating tasks for RCU priority boosting) must be delayed until after | |
149 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
150 | * currently delay invocation of any RCU callbacks until after this point. | |
151 | * | |
152 | * It might later prove better for people registering RCU callbacks during | |
153 | * early boot to take responsibility for these callbacks, but one step at | |
154 | * a time. | |
155 | */ | |
156 | static int rcu_scheduler_fully_active __read_mostly; | |
157 | ||
0aa04b05 PM |
158 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); |
159 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); | |
5d01bbd1 | 160 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
161 | static void invoke_rcu_core(void); |
162 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
6587a23b PM |
163 | static void rcu_report_exp_rdp(struct rcu_state *rsp, |
164 | struct rcu_data *rdp, bool wake); | |
a26ac245 | 165 | |
a94844b2 | 166 | /* rcuc/rcub kthread realtime priority */ |
26730f55 | 167 | #ifdef CONFIG_RCU_KTHREAD_PRIO |
a94844b2 | 168 | static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO; |
26730f55 PM |
169 | #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */ |
170 | static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0; | |
171 | #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */ | |
a94844b2 PM |
172 | module_param(kthread_prio, int, 0644); |
173 | ||
8d7dc928 | 174 | /* Delay in jiffies for grace-period initialization delays, debug only. */ |
0f41c0dd PM |
175 | |
176 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT | |
177 | static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY; | |
178 | module_param(gp_preinit_delay, int, 0644); | |
179 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ | |
180 | static const int gp_preinit_delay; | |
181 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ | |
182 | ||
8d7dc928 PM |
183 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT |
184 | static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY; | |
37745d28 | 185 | module_param(gp_init_delay, int, 0644); |
8d7dc928 PM |
186 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ |
187 | static const int gp_init_delay; | |
188 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ | |
eab128e8 | 189 | |
0f41c0dd PM |
190 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP |
191 | static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY; | |
192 | module_param(gp_cleanup_delay, int, 0644); | |
193 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ | |
194 | static const int gp_cleanup_delay; | |
195 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ | |
196 | ||
eab128e8 PM |
197 | /* |
198 | * Number of grace periods between delays, normalized by the duration of | |
199 | * the delay. The longer the the delay, the more the grace periods between | |
200 | * each delay. The reason for this normalization is that it means that, | |
201 | * for non-zero delays, the overall slowdown of grace periods is constant | |
202 | * regardless of the duration of the delay. This arrangement balances | |
203 | * the need for long delays to increase some race probabilities with the | |
204 | * need for fast grace periods to increase other race probabilities. | |
205 | */ | |
206 | #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */ | |
37745d28 | 207 | |
4a298656 PM |
208 | /* |
209 | * Track the rcutorture test sequence number and the update version | |
210 | * number within a given test. The rcutorture_testseq is incremented | |
211 | * on every rcutorture module load and unload, so has an odd value | |
212 | * when a test is running. The rcutorture_vernum is set to zero | |
213 | * when rcutorture starts and is incremented on each rcutorture update. | |
214 | * These variables enable correlating rcutorture output with the | |
215 | * RCU tracing information. | |
216 | */ | |
217 | unsigned long rcutorture_testseq; | |
218 | unsigned long rcutorture_vernum; | |
219 | ||
0aa04b05 PM |
220 | /* |
221 | * Compute the mask of online CPUs for the specified rcu_node structure. | |
222 | * This will not be stable unless the rcu_node structure's ->lock is | |
223 | * held, but the bit corresponding to the current CPU will be stable | |
224 | * in most contexts. | |
225 | */ | |
226 | unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) | |
227 | { | |
7d0ae808 | 228 | return READ_ONCE(rnp->qsmaskinitnext); |
0aa04b05 PM |
229 | } |
230 | ||
fc2219d4 | 231 | /* |
7d0ae808 | 232 | * Return true if an RCU grace period is in progress. The READ_ONCE()s |
fc2219d4 PM |
233 | * permit this function to be invoked without holding the root rcu_node |
234 | * structure's ->lock, but of course results can be subject to change. | |
235 | */ | |
236 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
237 | { | |
7d0ae808 | 238 | return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum); |
fc2219d4 PM |
239 | } |
240 | ||
b1f77b05 | 241 | /* |
d6714c22 | 242 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 243 | * how many quiescent states passed, just if there was at least |
d6714c22 | 244 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 245 | * The caller must have disabled preemption. |
b1f77b05 | 246 | */ |
284a8c93 | 247 | void rcu_sched_qs(void) |
b1f77b05 | 248 | { |
338b0f76 PM |
249 | unsigned long flags; |
250 | ||
5b74c458 | 251 | if (__this_cpu_read(rcu_sched_data.cpu_no_qs.s)) { |
284a8c93 PM |
252 | trace_rcu_grace_period(TPS("rcu_sched"), |
253 | __this_cpu_read(rcu_sched_data.gpnum), | |
254 | TPS("cpuqs")); | |
5b74c458 | 255 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false); |
338b0f76 PM |
256 | if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) |
257 | return; | |
258 | local_irq_save(flags); | |
6587a23b PM |
259 | if (__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) { |
260 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false); | |
261 | rcu_report_exp_rdp(&rcu_sched_state, | |
262 | this_cpu_ptr(&rcu_sched_data), | |
263 | true); | |
264 | } | |
338b0f76 | 265 | local_irq_restore(flags); |
284a8c93 | 266 | } |
b1f77b05 IM |
267 | } |
268 | ||
284a8c93 | 269 | void rcu_bh_qs(void) |
b1f77b05 | 270 | { |
5b74c458 | 271 | if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) { |
284a8c93 PM |
272 | trace_rcu_grace_period(TPS("rcu_bh"), |
273 | __this_cpu_read(rcu_bh_data.gpnum), | |
274 | TPS("cpuqs")); | |
5b74c458 | 275 | __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false); |
284a8c93 | 276 | } |
b1f77b05 | 277 | } |
64db4cff | 278 | |
4a81e832 PM |
279 | static DEFINE_PER_CPU(int, rcu_sched_qs_mask); |
280 | ||
281 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { | |
282 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, | |
283 | .dynticks = ATOMIC_INIT(1), | |
284 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
285 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
286 | .dynticks_idle = ATOMIC_INIT(1), | |
287 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
288 | }; | |
289 | ||
5cd37193 PM |
290 | DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr); |
291 | EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr); | |
292 | ||
4a81e832 PM |
293 | /* |
294 | * Let the RCU core know that this CPU has gone through the scheduler, | |
295 | * which is a quiescent state. This is called when the need for a | |
296 | * quiescent state is urgent, so we burn an atomic operation and full | |
297 | * memory barriers to let the RCU core know about it, regardless of what | |
298 | * this CPU might (or might not) do in the near future. | |
299 | * | |
300 | * We inform the RCU core by emulating a zero-duration dyntick-idle | |
301 | * period, which we in turn do by incrementing the ->dynticks counter | |
302 | * by two. | |
303 | */ | |
304 | static void rcu_momentary_dyntick_idle(void) | |
305 | { | |
306 | unsigned long flags; | |
307 | struct rcu_data *rdp; | |
308 | struct rcu_dynticks *rdtp; | |
309 | int resched_mask; | |
310 | struct rcu_state *rsp; | |
311 | ||
312 | local_irq_save(flags); | |
313 | ||
314 | /* | |
315 | * Yes, we can lose flag-setting operations. This is OK, because | |
316 | * the flag will be set again after some delay. | |
317 | */ | |
318 | resched_mask = raw_cpu_read(rcu_sched_qs_mask); | |
319 | raw_cpu_write(rcu_sched_qs_mask, 0); | |
320 | ||
321 | /* Find the flavor that needs a quiescent state. */ | |
322 | for_each_rcu_flavor(rsp) { | |
323 | rdp = raw_cpu_ptr(rsp->rda); | |
324 | if (!(resched_mask & rsp->flavor_mask)) | |
325 | continue; | |
326 | smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */ | |
7d0ae808 PM |
327 | if (READ_ONCE(rdp->mynode->completed) != |
328 | READ_ONCE(rdp->cond_resched_completed)) | |
4a81e832 PM |
329 | continue; |
330 | ||
331 | /* | |
332 | * Pretend to be momentarily idle for the quiescent state. | |
333 | * This allows the grace-period kthread to record the | |
334 | * quiescent state, with no need for this CPU to do anything | |
335 | * further. | |
336 | */ | |
337 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
338 | smp_mb__before_atomic(); /* Earlier stuff before QS. */ | |
339 | atomic_add(2, &rdtp->dynticks); /* QS. */ | |
340 | smp_mb__after_atomic(); /* Later stuff after QS. */ | |
341 | break; | |
342 | } | |
343 | local_irq_restore(flags); | |
344 | } | |
345 | ||
25502a6c PM |
346 | /* |
347 | * Note a context switch. This is a quiescent state for RCU-sched, | |
348 | * and requires special handling for preemptible RCU. | |
e4cc1f22 | 349 | * The caller must have disabled preemption. |
25502a6c | 350 | */ |
38200cf2 | 351 | void rcu_note_context_switch(void) |
25502a6c | 352 | { |
bb73c52b | 353 | barrier(); /* Avoid RCU read-side critical sections leaking down. */ |
f7f7bac9 | 354 | trace_rcu_utilization(TPS("Start context switch")); |
284a8c93 | 355 | rcu_sched_qs(); |
38200cf2 | 356 | rcu_preempt_note_context_switch(); |
4a81e832 PM |
357 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) |
358 | rcu_momentary_dyntick_idle(); | |
f7f7bac9 | 359 | trace_rcu_utilization(TPS("End context switch")); |
bb73c52b | 360 | barrier(); /* Avoid RCU read-side critical sections leaking up. */ |
25502a6c | 361 | } |
29ce8310 | 362 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 363 | |
5cd37193 | 364 | /* |
1925d196 | 365 | * Register a quiescent state for all RCU flavors. If there is an |
5cd37193 PM |
366 | * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight |
367 | * dyntick-idle quiescent state visible to other CPUs (but only for those | |
1925d196 | 368 | * RCU flavors in desperate need of a quiescent state, which will normally |
5cd37193 PM |
369 | * be none of them). Either way, do a lightweight quiescent state for |
370 | * all RCU flavors. | |
bb73c52b BF |
371 | * |
372 | * The barrier() calls are redundant in the common case when this is | |
373 | * called externally, but just in case this is called from within this | |
374 | * file. | |
375 | * | |
5cd37193 PM |
376 | */ |
377 | void rcu_all_qs(void) | |
378 | { | |
bb73c52b | 379 | barrier(); /* Avoid RCU read-side critical sections leaking down. */ |
5cd37193 PM |
380 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) |
381 | rcu_momentary_dyntick_idle(); | |
382 | this_cpu_inc(rcu_qs_ctr); | |
bb73c52b | 383 | barrier(); /* Avoid RCU read-side critical sections leaking up. */ |
5cd37193 PM |
384 | } |
385 | EXPORT_SYMBOL_GPL(rcu_all_qs); | |
386 | ||
878d7439 ED |
387 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
388 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
389 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 390 | |
878d7439 ED |
391 | module_param(blimit, long, 0444); |
392 | module_param(qhimark, long, 0444); | |
393 | module_param(qlowmark, long, 0444); | |
3d76c082 | 394 | |
026ad283 PM |
395 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
396 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
397 | |
398 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
399 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
400 | ||
4a81e832 PM |
401 | /* |
402 | * How long the grace period must be before we start recruiting | |
403 | * quiescent-state help from rcu_note_context_switch(). | |
404 | */ | |
405 | static ulong jiffies_till_sched_qs = HZ / 20; | |
406 | module_param(jiffies_till_sched_qs, ulong, 0644); | |
407 | ||
48a7639c | 408 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
910ee45d | 409 | struct rcu_data *rdp); |
217af2a2 PM |
410 | static void force_qs_rnp(struct rcu_state *rsp, |
411 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
412 | unsigned long *maxj), | |
413 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 414 | static void force_quiescent_state(struct rcu_state *rsp); |
e3950ecd | 415 | static int rcu_pending(void); |
64db4cff PM |
416 | |
417 | /* | |
917963d0 | 418 | * Return the number of RCU batches started thus far for debug & stats. |
64db4cff | 419 | */ |
917963d0 PM |
420 | unsigned long rcu_batches_started(void) |
421 | { | |
422 | return rcu_state_p->gpnum; | |
423 | } | |
424 | EXPORT_SYMBOL_GPL(rcu_batches_started); | |
425 | ||
426 | /* | |
427 | * Return the number of RCU-sched batches started thus far for debug & stats. | |
64db4cff | 428 | */ |
917963d0 PM |
429 | unsigned long rcu_batches_started_sched(void) |
430 | { | |
431 | return rcu_sched_state.gpnum; | |
432 | } | |
433 | EXPORT_SYMBOL_GPL(rcu_batches_started_sched); | |
434 | ||
435 | /* | |
436 | * Return the number of RCU BH batches started thus far for debug & stats. | |
437 | */ | |
438 | unsigned long rcu_batches_started_bh(void) | |
439 | { | |
440 | return rcu_bh_state.gpnum; | |
441 | } | |
442 | EXPORT_SYMBOL_GPL(rcu_batches_started_bh); | |
443 | ||
444 | /* | |
445 | * Return the number of RCU batches completed thus far for debug & stats. | |
446 | */ | |
447 | unsigned long rcu_batches_completed(void) | |
448 | { | |
449 | return rcu_state_p->completed; | |
450 | } | |
451 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
452 | ||
453 | /* | |
454 | * Return the number of RCU-sched batches completed thus far for debug & stats. | |
64db4cff | 455 | */ |
9733e4f0 | 456 | unsigned long rcu_batches_completed_sched(void) |
64db4cff | 457 | { |
d6714c22 | 458 | return rcu_sched_state.completed; |
64db4cff | 459 | } |
d6714c22 | 460 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
461 | |
462 | /* | |
917963d0 | 463 | * Return the number of RCU BH batches completed thus far for debug & stats. |
64db4cff | 464 | */ |
9733e4f0 | 465 | unsigned long rcu_batches_completed_bh(void) |
64db4cff PM |
466 | { |
467 | return rcu_bh_state.completed; | |
468 | } | |
469 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
470 | ||
a381d757 ACB |
471 | /* |
472 | * Force a quiescent state. | |
473 | */ | |
474 | void rcu_force_quiescent_state(void) | |
475 | { | |
e534165b | 476 | force_quiescent_state(rcu_state_p); |
a381d757 ACB |
477 | } |
478 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
479 | ||
bf66f18e PM |
480 | /* |
481 | * Force a quiescent state for RCU BH. | |
482 | */ | |
483 | void rcu_bh_force_quiescent_state(void) | |
484 | { | |
4cdfc175 | 485 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
486 | } |
487 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
488 | ||
e7580f33 PM |
489 | /* |
490 | * Force a quiescent state for RCU-sched. | |
491 | */ | |
492 | void rcu_sched_force_quiescent_state(void) | |
493 | { | |
494 | force_quiescent_state(&rcu_sched_state); | |
495 | } | |
496 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
497 | ||
afea227f PM |
498 | /* |
499 | * Show the state of the grace-period kthreads. | |
500 | */ | |
501 | void show_rcu_gp_kthreads(void) | |
502 | { | |
503 | struct rcu_state *rsp; | |
504 | ||
505 | for_each_rcu_flavor(rsp) { | |
506 | pr_info("%s: wait state: %d ->state: %#lx\n", | |
507 | rsp->name, rsp->gp_state, rsp->gp_kthread->state); | |
508 | /* sched_show_task(rsp->gp_kthread); */ | |
509 | } | |
510 | } | |
511 | EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); | |
512 | ||
4a298656 PM |
513 | /* |
514 | * Record the number of times rcutorture tests have been initiated and | |
515 | * terminated. This information allows the debugfs tracing stats to be | |
516 | * correlated to the rcutorture messages, even when the rcutorture module | |
517 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
518 | * store this state in rcutorture itself. | |
519 | */ | |
520 | void rcutorture_record_test_transition(void) | |
521 | { | |
522 | rcutorture_testseq++; | |
523 | rcutorture_vernum = 0; | |
524 | } | |
525 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
526 | ||
ad0dc7f9 PM |
527 | /* |
528 | * Send along grace-period-related data for rcutorture diagnostics. | |
529 | */ | |
530 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
531 | unsigned long *gpnum, unsigned long *completed) | |
532 | { | |
533 | struct rcu_state *rsp = NULL; | |
534 | ||
535 | switch (test_type) { | |
536 | case RCU_FLAVOR: | |
e534165b | 537 | rsp = rcu_state_p; |
ad0dc7f9 PM |
538 | break; |
539 | case RCU_BH_FLAVOR: | |
540 | rsp = &rcu_bh_state; | |
541 | break; | |
542 | case RCU_SCHED_FLAVOR: | |
543 | rsp = &rcu_sched_state; | |
544 | break; | |
545 | default: | |
546 | break; | |
547 | } | |
548 | if (rsp != NULL) { | |
7d0ae808 PM |
549 | *flags = READ_ONCE(rsp->gp_flags); |
550 | *gpnum = READ_ONCE(rsp->gpnum); | |
551 | *completed = READ_ONCE(rsp->completed); | |
ad0dc7f9 PM |
552 | return; |
553 | } | |
554 | *flags = 0; | |
555 | *gpnum = 0; | |
556 | *completed = 0; | |
557 | } | |
558 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
559 | ||
4a298656 PM |
560 | /* |
561 | * Record the number of writer passes through the current rcutorture test. | |
562 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
563 | * messages. | |
564 | */ | |
565 | void rcutorture_record_progress(unsigned long vernum) | |
566 | { | |
567 | rcutorture_vernum++; | |
568 | } | |
569 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
570 | ||
64db4cff PM |
571 | /* |
572 | * Does the CPU have callbacks ready to be invoked? | |
573 | */ | |
574 | static int | |
575 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
576 | { | |
3fbfbf7a PM |
577 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
578 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
579 | } |
580 | ||
365187fb PM |
581 | /* |
582 | * Return the root node of the specified rcu_state structure. | |
583 | */ | |
584 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
585 | { | |
586 | return &rsp->node[0]; | |
587 | } | |
588 | ||
589 | /* | |
590 | * Is there any need for future grace periods? | |
591 | * Interrupts must be disabled. If the caller does not hold the root | |
592 | * rnp_node structure's ->lock, the results are advisory only. | |
593 | */ | |
594 | static int rcu_future_needs_gp(struct rcu_state *rsp) | |
595 | { | |
596 | struct rcu_node *rnp = rcu_get_root(rsp); | |
7d0ae808 | 597 | int idx = (READ_ONCE(rnp->completed) + 1) & 0x1; |
365187fb PM |
598 | int *fp = &rnp->need_future_gp[idx]; |
599 | ||
7d0ae808 | 600 | return READ_ONCE(*fp); |
365187fb PM |
601 | } |
602 | ||
64db4cff | 603 | /* |
dc35c893 PM |
604 | * Does the current CPU require a not-yet-started grace period? |
605 | * The caller must have disabled interrupts to prevent races with | |
606 | * normal callback registry. | |
64db4cff PM |
607 | */ |
608 | static int | |
609 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
610 | { | |
dc35c893 | 611 | int i; |
3fbfbf7a | 612 | |
dc35c893 PM |
613 | if (rcu_gp_in_progress(rsp)) |
614 | return 0; /* No, a grace period is already in progress. */ | |
365187fb | 615 | if (rcu_future_needs_gp(rsp)) |
34ed6246 | 616 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
617 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
618 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
619 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
620 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
621 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
622 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
7d0ae808 | 623 | ULONG_CMP_LT(READ_ONCE(rsp->completed), |
dc35c893 PM |
624 | rdp->nxtcompleted[i])) |
625 | return 1; /* Yes, CBs for future grace period. */ | |
626 | return 0; /* No grace period needed. */ | |
64db4cff PM |
627 | } |
628 | ||
9b2e4f18 | 629 | /* |
adf5091e | 630 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
631 | * |
632 | * If the new value of the ->dynticks_nesting counter now is zero, | |
633 | * we really have entered idle, and must do the appropriate accounting. | |
634 | * The caller must have disabled interrupts. | |
635 | */ | |
28ced795 | 636 | static void rcu_eqs_enter_common(long long oldval, bool user) |
9b2e4f18 | 637 | { |
96d3fd0d PM |
638 | struct rcu_state *rsp; |
639 | struct rcu_data *rdp; | |
28ced795 | 640 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
96d3fd0d | 641 | |
f7f7bac9 | 642 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
1ce46ee5 PM |
643 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
644 | !user && !is_idle_task(current)) { | |
289828e6 PM |
645 | struct task_struct *idle __maybe_unused = |
646 | idle_task(smp_processor_id()); | |
0989cb46 | 647 | |
f7f7bac9 | 648 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 649 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
650 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
651 | current->pid, current->comm, | |
652 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 653 | } |
96d3fd0d PM |
654 | for_each_rcu_flavor(rsp) { |
655 | rdp = this_cpu_ptr(rsp->rda); | |
656 | do_nocb_deferred_wakeup(rdp); | |
657 | } | |
198bbf81 | 658 | rcu_prepare_for_idle(); |
9b2e4f18 | 659 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 660 | smp_mb__before_atomic(); /* See above. */ |
9b2e4f18 | 661 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 662 | smp_mb__after_atomic(); /* Force ordering with next sojourn. */ |
1ce46ee5 PM |
663 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
664 | atomic_read(&rdtp->dynticks) & 0x1); | |
176f8f7a | 665 | rcu_dynticks_task_enter(); |
c44e2cdd PM |
666 | |
667 | /* | |
adf5091e | 668 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
669 | * in an RCU read-side critical section. |
670 | */ | |
f78f5b90 PM |
671 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map), |
672 | "Illegal idle entry in RCU read-side critical section."); | |
673 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map), | |
674 | "Illegal idle entry in RCU-bh read-side critical section."); | |
675 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map), | |
676 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 677 | } |
64db4cff | 678 | |
adf5091e FW |
679 | /* |
680 | * Enter an RCU extended quiescent state, which can be either the | |
681 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 682 | */ |
adf5091e | 683 | static void rcu_eqs_enter(bool user) |
64db4cff | 684 | { |
4145fa7f | 685 | long long oldval; |
64db4cff PM |
686 | struct rcu_dynticks *rdtp; |
687 | ||
c9d4b0af | 688 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 689 | oldval = rdtp->dynticks_nesting; |
1ce46ee5 PM |
690 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
691 | (oldval & DYNTICK_TASK_NEST_MASK) == 0); | |
3a592405 | 692 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { |
29e37d81 | 693 | rdtp->dynticks_nesting = 0; |
28ced795 | 694 | rcu_eqs_enter_common(oldval, user); |
3a592405 | 695 | } else { |
29e37d81 | 696 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; |
3a592405 | 697 | } |
64db4cff | 698 | } |
adf5091e FW |
699 | |
700 | /** | |
701 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
702 | * | |
703 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
704 | * read-side critical sections can occur. (Though RCU read-side | |
705 | * critical sections can occur in irq handlers in idle, a possibility | |
706 | * handled by irq_enter() and irq_exit().) | |
707 | * | |
708 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
709 | * the possibility of usermode upcalls having messed up our count | |
710 | * of interrupt nesting level during the prior busy period. | |
711 | */ | |
712 | void rcu_idle_enter(void) | |
713 | { | |
c5d900bf FW |
714 | unsigned long flags; |
715 | ||
716 | local_irq_save(flags); | |
cb349ca9 | 717 | rcu_eqs_enter(false); |
28ced795 | 718 | rcu_sysidle_enter(0); |
c5d900bf | 719 | local_irq_restore(flags); |
adf5091e | 720 | } |
8a2ecf47 | 721 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 722 | |
d1ec4c34 | 723 | #ifdef CONFIG_NO_HZ_FULL |
adf5091e FW |
724 | /** |
725 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
726 | * | |
727 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
728 | * is permitted between this call and rcu_user_exit(). This way the | |
729 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
730 | * when the CPU runs in userspace. | |
731 | */ | |
732 | void rcu_user_enter(void) | |
733 | { | |
91d1aa43 | 734 | rcu_eqs_enter(1); |
adf5091e | 735 | } |
d1ec4c34 | 736 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 737 | |
9b2e4f18 PM |
738 | /** |
739 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
740 | * | |
741 | * Exit from an interrupt handler, which might possibly result in entering | |
742 | * idle mode, in other words, leaving the mode in which read-side critical | |
743 | * sections can occur. | |
64db4cff | 744 | * |
9b2e4f18 PM |
745 | * This code assumes that the idle loop never does anything that might |
746 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
747 | * architecture violates this assumption, RCU will give you what you | |
748 | * deserve, good and hard. But very infrequently and irreproducibly. | |
749 | * | |
750 | * Use things like work queues to work around this limitation. | |
751 | * | |
752 | * You have been warned. | |
64db4cff | 753 | */ |
9b2e4f18 | 754 | void rcu_irq_exit(void) |
64db4cff PM |
755 | { |
756 | unsigned long flags; | |
4145fa7f | 757 | long long oldval; |
64db4cff PM |
758 | struct rcu_dynticks *rdtp; |
759 | ||
760 | local_irq_save(flags); | |
c9d4b0af | 761 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 762 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 | 763 | rdtp->dynticks_nesting--; |
1ce46ee5 PM |
764 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
765 | rdtp->dynticks_nesting < 0); | |
b6fc6020 | 766 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 767 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 768 | else |
28ced795 CL |
769 | rcu_eqs_enter_common(oldval, true); |
770 | rcu_sysidle_enter(1); | |
9b2e4f18 PM |
771 | local_irq_restore(flags); |
772 | } | |
773 | ||
774 | /* | |
adf5091e | 775 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
776 | * |
777 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
778 | * we really have exited idle, and must do the appropriate accounting. | |
779 | * The caller must have disabled interrupts. | |
780 | */ | |
28ced795 | 781 | static void rcu_eqs_exit_common(long long oldval, int user) |
9b2e4f18 | 782 | { |
28ced795 CL |
783 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
784 | ||
176f8f7a | 785 | rcu_dynticks_task_exit(); |
4e857c58 | 786 | smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */ |
23b5c8fa PM |
787 | atomic_inc(&rdtp->dynticks); |
788 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
4e857c58 | 789 | smp_mb__after_atomic(); /* See above. */ |
1ce46ee5 PM |
790 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
791 | !(atomic_read(&rdtp->dynticks) & 0x1)); | |
8fa7845d | 792 | rcu_cleanup_after_idle(); |
f7f7bac9 | 793 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
1ce46ee5 PM |
794 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
795 | !user && !is_idle_task(current)) { | |
289828e6 PM |
796 | struct task_struct *idle __maybe_unused = |
797 | idle_task(smp_processor_id()); | |
0989cb46 | 798 | |
f7f7bac9 | 799 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 800 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 801 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
802 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
803 | current->pid, current->comm, | |
804 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
805 | } |
806 | } | |
807 | ||
adf5091e FW |
808 | /* |
809 | * Exit an RCU extended quiescent state, which can be either the | |
810 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 811 | */ |
adf5091e | 812 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 813 | { |
9b2e4f18 PM |
814 | struct rcu_dynticks *rdtp; |
815 | long long oldval; | |
816 | ||
c9d4b0af | 817 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 | 818 | oldval = rdtp->dynticks_nesting; |
1ce46ee5 | 819 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0); |
3a592405 | 820 | if (oldval & DYNTICK_TASK_NEST_MASK) { |
29e37d81 | 821 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; |
3a592405 | 822 | } else { |
29e37d81 | 823 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
28ced795 | 824 | rcu_eqs_exit_common(oldval, user); |
3a592405 | 825 | } |
9b2e4f18 | 826 | } |
adf5091e FW |
827 | |
828 | /** | |
829 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
830 | * | |
831 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
832 | * read-side critical sections can occur. | |
833 | * | |
834 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
835 | * allow for the possibility of usermode upcalls messing up our count | |
836 | * of interrupt nesting level during the busy period that is just | |
837 | * now starting. | |
838 | */ | |
839 | void rcu_idle_exit(void) | |
840 | { | |
c5d900bf FW |
841 | unsigned long flags; |
842 | ||
843 | local_irq_save(flags); | |
cb349ca9 | 844 | rcu_eqs_exit(false); |
28ced795 | 845 | rcu_sysidle_exit(0); |
c5d900bf | 846 | local_irq_restore(flags); |
adf5091e | 847 | } |
8a2ecf47 | 848 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 849 | |
d1ec4c34 | 850 | #ifdef CONFIG_NO_HZ_FULL |
adf5091e FW |
851 | /** |
852 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
853 | * | |
854 | * Exit RCU idle mode while entering the kernel because it can | |
855 | * run a RCU read side critical section anytime. | |
856 | */ | |
857 | void rcu_user_exit(void) | |
858 | { | |
91d1aa43 | 859 | rcu_eqs_exit(1); |
adf5091e | 860 | } |
d1ec4c34 | 861 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 862 | |
9b2e4f18 PM |
863 | /** |
864 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
865 | * | |
866 | * Enter an interrupt handler, which might possibly result in exiting | |
867 | * idle mode, in other words, entering the mode in which read-side critical | |
868 | * sections can occur. | |
869 | * | |
870 | * Note that the Linux kernel is fully capable of entering an interrupt | |
871 | * handler that it never exits, for example when doing upcalls to | |
872 | * user mode! This code assumes that the idle loop never does upcalls to | |
873 | * user mode. If your architecture does do upcalls from the idle loop (or | |
874 | * does anything else that results in unbalanced calls to the irq_enter() | |
875 | * and irq_exit() functions), RCU will give you what you deserve, good | |
876 | * and hard. But very infrequently and irreproducibly. | |
877 | * | |
878 | * Use things like work queues to work around this limitation. | |
879 | * | |
880 | * You have been warned. | |
881 | */ | |
882 | void rcu_irq_enter(void) | |
883 | { | |
884 | unsigned long flags; | |
885 | struct rcu_dynticks *rdtp; | |
886 | long long oldval; | |
887 | ||
888 | local_irq_save(flags); | |
c9d4b0af | 889 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 PM |
890 | oldval = rdtp->dynticks_nesting; |
891 | rdtp->dynticks_nesting++; | |
1ce46ee5 PM |
892 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
893 | rdtp->dynticks_nesting == 0); | |
b6fc6020 | 894 | if (oldval) |
f7f7bac9 | 895 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 896 | else |
28ced795 CL |
897 | rcu_eqs_exit_common(oldval, true); |
898 | rcu_sysidle_exit(1); | |
64db4cff | 899 | local_irq_restore(flags); |
64db4cff PM |
900 | } |
901 | ||
902 | /** | |
903 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
904 | * | |
734d1680 PM |
905 | * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and |
906 | * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know | |
907 | * that the CPU is active. This implementation permits nested NMIs, as | |
908 | * long as the nesting level does not overflow an int. (You will probably | |
909 | * run out of stack space first.) | |
64db4cff PM |
910 | */ |
911 | void rcu_nmi_enter(void) | |
912 | { | |
c9d4b0af | 913 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
734d1680 | 914 | int incby = 2; |
64db4cff | 915 | |
734d1680 PM |
916 | /* Complain about underflow. */ |
917 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0); | |
918 | ||
919 | /* | |
920 | * If idle from RCU viewpoint, atomically increment ->dynticks | |
921 | * to mark non-idle and increment ->dynticks_nmi_nesting by one. | |
922 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means | |
923 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed | |
924 | * to be in the outermost NMI handler that interrupted an RCU-idle | |
925 | * period (observation due to Andy Lutomirski). | |
926 | */ | |
927 | if (!(atomic_read(&rdtp->dynticks) & 0x1)) { | |
928 | smp_mb__before_atomic(); /* Force delay from prior write. */ | |
929 | atomic_inc(&rdtp->dynticks); | |
930 | /* atomic_inc() before later RCU read-side crit sects */ | |
931 | smp_mb__after_atomic(); /* See above. */ | |
932 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
933 | incby = 1; | |
934 | } | |
935 | rdtp->dynticks_nmi_nesting += incby; | |
936 | barrier(); | |
64db4cff PM |
937 | } |
938 | ||
939 | /** | |
940 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
941 | * | |
734d1680 PM |
942 | * If we are returning from the outermost NMI handler that interrupted an |
943 | * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting | |
944 | * to let the RCU grace-period handling know that the CPU is back to | |
945 | * being RCU-idle. | |
64db4cff PM |
946 | */ |
947 | void rcu_nmi_exit(void) | |
948 | { | |
c9d4b0af | 949 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 950 | |
734d1680 PM |
951 | /* |
952 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. | |
953 | * (We are exiting an NMI handler, so RCU better be paying attention | |
954 | * to us!) | |
955 | */ | |
956 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0); | |
957 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
958 | ||
959 | /* | |
960 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so | |
961 | * leave it in non-RCU-idle state. | |
962 | */ | |
963 | if (rdtp->dynticks_nmi_nesting != 1) { | |
964 | rdtp->dynticks_nmi_nesting -= 2; | |
64db4cff | 965 | return; |
734d1680 PM |
966 | } |
967 | ||
968 | /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ | |
969 | rdtp->dynticks_nmi_nesting = 0; | |
23b5c8fa | 970 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 971 | smp_mb__before_atomic(); /* See above. */ |
23b5c8fa | 972 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 973 | smp_mb__after_atomic(); /* Force delay to next write. */ |
23b5c8fa | 974 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
64db4cff PM |
975 | } |
976 | ||
977 | /** | |
5c173eb8 PM |
978 | * __rcu_is_watching - are RCU read-side critical sections safe? |
979 | * | |
980 | * Return true if RCU is watching the running CPU, which means that | |
981 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
982 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
983 | * least disabled preemption. | |
984 | */ | |
9418fb20 | 985 | bool notrace __rcu_is_watching(void) |
5c173eb8 PM |
986 | { |
987 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
988 | } | |
989 | ||
990 | /** | |
991 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
64db4cff | 992 | * |
9b2e4f18 | 993 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 994 | * or NMI handler, return true. |
64db4cff | 995 | */ |
9418fb20 | 996 | bool notrace rcu_is_watching(void) |
64db4cff | 997 | { |
f534ed1f | 998 | bool ret; |
34240697 | 999 | |
46f00d18 | 1000 | preempt_disable_notrace(); |
5c173eb8 | 1001 | ret = __rcu_is_watching(); |
46f00d18 | 1002 | preempt_enable_notrace(); |
34240697 | 1003 | return ret; |
64db4cff | 1004 | } |
5c173eb8 | 1005 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 1006 | |
62fde6ed | 1007 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
1008 | |
1009 | /* | |
1010 | * Is the current CPU online? Disable preemption to avoid false positives | |
1011 | * that could otherwise happen due to the current CPU number being sampled, | |
1012 | * this task being preempted, its old CPU being taken offline, resuming | |
1013 | * on some other CPU, then determining that its old CPU is now offline. | |
1014 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
1015 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
1016 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
1017 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
1018 | * offline to continue to use RCU for one jiffy after marking itself | |
1019 | * offline in the cpu_online_mask. This leniency is necessary given the | |
1020 | * non-atomic nature of the online and offline processing, for example, | |
1021 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
1022 | * notifiers. | |
1023 | * | |
1024 | * This is also why RCU internally marks CPUs online during the | |
1025 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
1026 | * |
1027 | * Disable checking if in an NMI handler because we cannot safely report | |
1028 | * errors from NMI handlers anyway. | |
1029 | */ | |
1030 | bool rcu_lockdep_current_cpu_online(void) | |
1031 | { | |
2036d94a PM |
1032 | struct rcu_data *rdp; |
1033 | struct rcu_node *rnp; | |
c0d6d01b PM |
1034 | bool ret; |
1035 | ||
1036 | if (in_nmi()) | |
f6f7ee9a | 1037 | return true; |
c0d6d01b | 1038 | preempt_disable(); |
c9d4b0af | 1039 | rdp = this_cpu_ptr(&rcu_sched_data); |
2036d94a | 1040 | rnp = rdp->mynode; |
0aa04b05 | 1041 | ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) || |
c0d6d01b PM |
1042 | !rcu_scheduler_fully_active; |
1043 | preempt_enable(); | |
1044 | return ret; | |
1045 | } | |
1046 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
1047 | ||
62fde6ed | 1048 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 1049 | |
64db4cff | 1050 | /** |
9b2e4f18 | 1051 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 1052 | * |
9b2e4f18 PM |
1053 | * If the current CPU is idle or running at a first-level (not nested) |
1054 | * interrupt from idle, return true. The caller must have at least | |
1055 | * disabled preemption. | |
64db4cff | 1056 | */ |
62e3cb14 | 1057 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 1058 | { |
c9d4b0af | 1059 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
64db4cff PM |
1060 | } |
1061 | ||
64db4cff PM |
1062 | /* |
1063 | * Snapshot the specified CPU's dynticks counter so that we can later | |
1064 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 1065 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 1066 | */ |
217af2a2 PM |
1067 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
1068 | bool *isidle, unsigned long *maxj) | |
64db4cff | 1069 | { |
23b5c8fa | 1070 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 1071 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
7941dbde ACB |
1072 | if ((rdp->dynticks_snap & 0x1) == 0) { |
1073 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); | |
1074 | return 1; | |
1075 | } else { | |
7d0ae808 | 1076 | if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, |
e3663b10 | 1077 | rdp->mynode->gpnum)) |
7d0ae808 | 1078 | WRITE_ONCE(rdp->gpwrap, true); |
7941dbde ACB |
1079 | return 0; |
1080 | } | |
64db4cff PM |
1081 | } |
1082 | ||
1083 | /* | |
1084 | * Return true if the specified CPU has passed through a quiescent | |
1085 | * state by virtue of being in or having passed through an dynticks | |
1086 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 1087 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 1088 | */ |
217af2a2 PM |
1089 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
1090 | bool *isidle, unsigned long *maxj) | |
64db4cff | 1091 | { |
7eb4f455 | 1092 | unsigned int curr; |
4a81e832 | 1093 | int *rcrmp; |
7eb4f455 | 1094 | unsigned int snap; |
64db4cff | 1095 | |
7eb4f455 PM |
1096 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
1097 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
1098 | |
1099 | /* | |
1100 | * If the CPU passed through or entered a dynticks idle phase with | |
1101 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
1102 | * already acknowledged the request to pass through a quiescent | |
1103 | * state. Either way, that CPU cannot possibly be in an RCU | |
1104 | * read-side critical section that started before the beginning | |
1105 | * of the current RCU grace period. | |
1106 | */ | |
7eb4f455 | 1107 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 1108 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
1109 | rdp->dynticks_fqs++; |
1110 | return 1; | |
1111 | } | |
1112 | ||
a82dcc76 PM |
1113 | /* |
1114 | * Check for the CPU being offline, but only if the grace period | |
1115 | * is old enough. We don't need to worry about the CPU changing | |
1116 | * state: If we see it offline even once, it has been through a | |
1117 | * quiescent state. | |
1118 | * | |
1119 | * The reason for insisting that the grace period be at least | |
1120 | * one jiffy old is that CPUs that are not quite online and that | |
1121 | * have just gone offline can still execute RCU read-side critical | |
1122 | * sections. | |
1123 | */ | |
1124 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
1125 | return 0; /* Grace period is not old enough. */ | |
1126 | barrier(); | |
1127 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 1128 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
1129 | rdp->offline_fqs++; |
1130 | return 1; | |
1131 | } | |
65d798f0 PM |
1132 | |
1133 | /* | |
4a81e832 PM |
1134 | * A CPU running for an extended time within the kernel can |
1135 | * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode, | |
1136 | * even context-switching back and forth between a pair of | |
1137 | * in-kernel CPU-bound tasks cannot advance grace periods. | |
1138 | * So if the grace period is old enough, make the CPU pay attention. | |
1139 | * Note that the unsynchronized assignments to the per-CPU | |
1140 | * rcu_sched_qs_mask variable are safe. Yes, setting of | |
1141 | * bits can be lost, but they will be set again on the next | |
1142 | * force-quiescent-state pass. So lost bit sets do not result | |
1143 | * in incorrect behavior, merely in a grace period lasting | |
1144 | * a few jiffies longer than it might otherwise. Because | |
1145 | * there are at most four threads involved, and because the | |
1146 | * updates are only once every few jiffies, the probability of | |
1147 | * lossage (and thus of slight grace-period extension) is | |
1148 | * quite low. | |
1149 | * | |
1150 | * Note that if the jiffies_till_sched_qs boot/sysfs parameter | |
1151 | * is set too high, we override with half of the RCU CPU stall | |
1152 | * warning delay. | |
6193c76a | 1153 | */ |
4a81e832 PM |
1154 | rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu); |
1155 | if (ULONG_CMP_GE(jiffies, | |
1156 | rdp->rsp->gp_start + jiffies_till_sched_qs) || | |
cb1e78cf | 1157 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { |
7d0ae808 PM |
1158 | if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) { |
1159 | WRITE_ONCE(rdp->cond_resched_completed, | |
1160 | READ_ONCE(rdp->mynode->completed)); | |
4a81e832 | 1161 | smp_mb(); /* ->cond_resched_completed before *rcrmp. */ |
7d0ae808 PM |
1162 | WRITE_ONCE(*rcrmp, |
1163 | READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask); | |
4a81e832 PM |
1164 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ |
1165 | rdp->rsp->jiffies_resched += 5; /* Enable beating. */ | |
1166 | } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { | |
1167 | /* Time to beat on that CPU again! */ | |
1168 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1169 | rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */ | |
1170 | } | |
6193c76a PM |
1171 | } |
1172 | ||
a82dcc76 | 1173 | return 0; |
64db4cff PM |
1174 | } |
1175 | ||
64db4cff PM |
1176 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
1177 | { | |
cb1e78cf | 1178 | unsigned long j = jiffies; |
6193c76a | 1179 | unsigned long j1; |
26cdfedf PM |
1180 | |
1181 | rsp->gp_start = j; | |
1182 | smp_wmb(); /* Record start time before stall time. */ | |
6193c76a | 1183 | j1 = rcu_jiffies_till_stall_check(); |
7d0ae808 | 1184 | WRITE_ONCE(rsp->jiffies_stall, j + j1); |
6193c76a | 1185 | rsp->jiffies_resched = j + j1 / 2; |
7d0ae808 | 1186 | rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs); |
64db4cff PM |
1187 | } |
1188 | ||
fb81a44b PM |
1189 | /* |
1190 | * Complain about starvation of grace-period kthread. | |
1191 | */ | |
1192 | static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp) | |
1193 | { | |
1194 | unsigned long gpa; | |
1195 | unsigned long j; | |
1196 | ||
1197 | j = jiffies; | |
7d0ae808 | 1198 | gpa = READ_ONCE(rsp->gp_activity); |
fb81a44b | 1199 | if (j - gpa > 2 * HZ) |
319362c9 | 1200 | pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x s%d ->state=%#lx\n", |
81e701e4 | 1201 | rsp->name, j - gpa, |
319362c9 PM |
1202 | rsp->gpnum, rsp->completed, |
1203 | rsp->gp_flags, rsp->gp_state, | |
1204 | rsp->gp_kthread ? rsp->gp_kthread->state : 0); | |
64db4cff PM |
1205 | } |
1206 | ||
b637a328 | 1207 | /* |
bc1dce51 | 1208 | * Dump stacks of all tasks running on stalled CPUs. |
b637a328 PM |
1209 | */ |
1210 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
1211 | { | |
1212 | int cpu; | |
1213 | unsigned long flags; | |
1214 | struct rcu_node *rnp; | |
1215 | ||
1216 | rcu_for_each_leaf_node(rsp, rnp) { | |
1217 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1218 | if (rnp->qsmask != 0) { | |
1219 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1220 | if (rnp->qsmask & (1UL << cpu)) | |
1221 | dump_cpu_task(rnp->grplo + cpu); | |
1222 | } | |
1223 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1224 | } | |
1225 | } | |
1226 | ||
6ccd2ecd | 1227 | static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum) |
64db4cff PM |
1228 | { |
1229 | int cpu; | |
1230 | long delta; | |
1231 | unsigned long flags; | |
6ccd2ecd PM |
1232 | unsigned long gpa; |
1233 | unsigned long j; | |
285fe294 | 1234 | int ndetected = 0; |
64db4cff | 1235 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 1236 | long totqlen = 0; |
64db4cff PM |
1237 | |
1238 | /* Only let one CPU complain about others per time interval. */ | |
1239 | ||
1304afb2 | 1240 | raw_spin_lock_irqsave(&rnp->lock, flags); |
7d0ae808 | 1241 | delta = jiffies - READ_ONCE(rsp->jiffies_stall); |
fc2219d4 | 1242 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 1243 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1244 | return; |
1245 | } | |
7d0ae808 PM |
1246 | WRITE_ONCE(rsp->jiffies_stall, |
1247 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3); | |
1304afb2 | 1248 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1249 | |
8cdd32a9 PM |
1250 | /* |
1251 | * OK, time to rat on our buddy... | |
1252 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1253 | * RCU CPU stall warnings. | |
1254 | */ | |
d7f3e207 | 1255 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 1256 | rsp->name); |
a858af28 | 1257 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 1258 | rcu_for_each_leaf_node(rsp, rnp) { |
3acd9eb3 | 1259 | raw_spin_lock_irqsave(&rnp->lock, flags); |
9bc8b558 | 1260 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
1261 | if (rnp->qsmask != 0) { |
1262 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1263 | if (rnp->qsmask & (1UL << cpu)) { | |
1264 | print_cpu_stall_info(rsp, | |
1265 | rnp->grplo + cpu); | |
1266 | ndetected++; | |
1267 | } | |
1268 | } | |
3acd9eb3 | 1269 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1270 | } |
a858af28 | 1271 | |
a858af28 | 1272 | print_cpu_stall_info_end(); |
53bb857c PM |
1273 | for_each_possible_cpu(cpu) |
1274 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e | 1275 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", |
eee05882 | 1276 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
83ebe63e | 1277 | (long)rsp->gpnum, (long)rsp->completed, totqlen); |
6ccd2ecd | 1278 | if (ndetected) { |
b637a328 | 1279 | rcu_dump_cpu_stacks(rsp); |
6ccd2ecd | 1280 | } else { |
7d0ae808 PM |
1281 | if (READ_ONCE(rsp->gpnum) != gpnum || |
1282 | READ_ONCE(rsp->completed) == gpnum) { | |
6ccd2ecd PM |
1283 | pr_err("INFO: Stall ended before state dump start\n"); |
1284 | } else { | |
1285 | j = jiffies; | |
7d0ae808 | 1286 | gpa = READ_ONCE(rsp->gp_activity); |
237a0f21 | 1287 | pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n", |
6ccd2ecd | 1288 | rsp->name, j - gpa, j, gpa, |
237a0f21 PM |
1289 | jiffies_till_next_fqs, |
1290 | rcu_get_root(rsp)->qsmask); | |
6ccd2ecd PM |
1291 | /* In this case, the current CPU might be at fault. */ |
1292 | sched_show_task(current); | |
1293 | } | |
1294 | } | |
c1dc0b9c | 1295 | |
4cdfc175 | 1296 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
1297 | rcu_print_detail_task_stall(rsp); |
1298 | ||
fb81a44b PM |
1299 | rcu_check_gp_kthread_starvation(rsp); |
1300 | ||
4cdfc175 | 1301 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
1302 | } |
1303 | ||
1304 | static void print_cpu_stall(struct rcu_state *rsp) | |
1305 | { | |
53bb857c | 1306 | int cpu; |
64db4cff PM |
1307 | unsigned long flags; |
1308 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 1309 | long totqlen = 0; |
64db4cff | 1310 | |
8cdd32a9 PM |
1311 | /* |
1312 | * OK, time to rat on ourselves... | |
1313 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1314 | * RCU CPU stall warnings. | |
1315 | */ | |
d7f3e207 | 1316 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
1317 | print_cpu_stall_info_begin(); |
1318 | print_cpu_stall_info(rsp, smp_processor_id()); | |
1319 | print_cpu_stall_info_end(); | |
53bb857c PM |
1320 | for_each_possible_cpu(cpu) |
1321 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e PM |
1322 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", |
1323 | jiffies - rsp->gp_start, | |
1324 | (long)rsp->gpnum, (long)rsp->completed, totqlen); | |
fb81a44b PM |
1325 | |
1326 | rcu_check_gp_kthread_starvation(rsp); | |
1327 | ||
bc1dce51 | 1328 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 1329 | |
1304afb2 | 1330 | raw_spin_lock_irqsave(&rnp->lock, flags); |
7d0ae808 PM |
1331 | if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall))) |
1332 | WRITE_ONCE(rsp->jiffies_stall, | |
1333 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3); | |
1304afb2 | 1334 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 1335 | |
b021fe3e PZ |
1336 | /* |
1337 | * Attempt to revive the RCU machinery by forcing a context switch. | |
1338 | * | |
1339 | * A context switch would normally allow the RCU state machine to make | |
1340 | * progress and it could be we're stuck in kernel space without context | |
1341 | * switches for an entirely unreasonable amount of time. | |
1342 | */ | |
1343 | resched_cpu(smp_processor_id()); | |
64db4cff PM |
1344 | } |
1345 | ||
1346 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
1347 | { | |
26cdfedf PM |
1348 | unsigned long completed; |
1349 | unsigned long gpnum; | |
1350 | unsigned long gps; | |
bad6e139 PM |
1351 | unsigned long j; |
1352 | unsigned long js; | |
64db4cff PM |
1353 | struct rcu_node *rnp; |
1354 | ||
26cdfedf | 1355 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
c68de209 | 1356 | return; |
cb1e78cf | 1357 | j = jiffies; |
26cdfedf PM |
1358 | |
1359 | /* | |
1360 | * Lots of memory barriers to reject false positives. | |
1361 | * | |
1362 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
1363 | * then rsp->gp_start, and finally rsp->completed. These values | |
1364 | * are updated in the opposite order with memory barriers (or | |
1365 | * equivalent) during grace-period initialization and cleanup. | |
1366 | * Now, a false positive can occur if we get an new value of | |
1367 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
1368 | * the memory barriers, the only way that this can happen is if one | |
1369 | * grace period ends and another starts between these two fetches. | |
1370 | * Detect this by comparing rsp->completed with the previous fetch | |
1371 | * from rsp->gpnum. | |
1372 | * | |
1373 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
1374 | * and rsp->gp_start suffice to forestall false positives. | |
1375 | */ | |
7d0ae808 | 1376 | gpnum = READ_ONCE(rsp->gpnum); |
26cdfedf | 1377 | smp_rmb(); /* Pick up ->gpnum first... */ |
7d0ae808 | 1378 | js = READ_ONCE(rsp->jiffies_stall); |
26cdfedf | 1379 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
7d0ae808 | 1380 | gps = READ_ONCE(rsp->gp_start); |
26cdfedf | 1381 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ |
7d0ae808 | 1382 | completed = READ_ONCE(rsp->completed); |
26cdfedf PM |
1383 | if (ULONG_CMP_GE(completed, gpnum) || |
1384 | ULONG_CMP_LT(j, js) || | |
1385 | ULONG_CMP_GE(gps, js)) | |
1386 | return; /* No stall or GP completed since entering function. */ | |
64db4cff | 1387 | rnp = rdp->mynode; |
c96ea7cf | 1388 | if (rcu_gp_in_progress(rsp) && |
7d0ae808 | 1389 | (READ_ONCE(rnp->qsmask) & rdp->grpmask)) { |
64db4cff PM |
1390 | |
1391 | /* We haven't checked in, so go dump stack. */ | |
1392 | print_cpu_stall(rsp); | |
1393 | ||
bad6e139 PM |
1394 | } else if (rcu_gp_in_progress(rsp) && |
1395 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 1396 | |
bad6e139 | 1397 | /* They had a few time units to dump stack, so complain. */ |
6ccd2ecd | 1398 | print_other_cpu_stall(rsp, gpnum); |
64db4cff PM |
1399 | } |
1400 | } | |
1401 | ||
53d84e00 PM |
1402 | /** |
1403 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1404 | * | |
1405 | * Set the stall-warning timeout way off into the future, thus preventing | |
1406 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1407 | * RCU grace periods. | |
1408 | * | |
1409 | * The caller must disable hard irqs. | |
1410 | */ | |
1411 | void rcu_cpu_stall_reset(void) | |
1412 | { | |
6ce75a23 PM |
1413 | struct rcu_state *rsp; |
1414 | ||
1415 | for_each_rcu_flavor(rsp) | |
7d0ae808 | 1416 | WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2); |
53d84e00 PM |
1417 | } |
1418 | ||
3f5d3ea6 | 1419 | /* |
d3f3f3f2 PM |
1420 | * Initialize the specified rcu_data structure's default callback list |
1421 | * to empty. The default callback list is the one that is not used by | |
1422 | * no-callbacks CPUs. | |
3f5d3ea6 | 1423 | */ |
d3f3f3f2 | 1424 | static void init_default_callback_list(struct rcu_data *rdp) |
3f5d3ea6 PM |
1425 | { |
1426 | int i; | |
1427 | ||
1428 | rdp->nxtlist = NULL; | |
1429 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1430 | rdp->nxttail[i] = &rdp->nxtlist; | |
1431 | } | |
1432 | ||
d3f3f3f2 PM |
1433 | /* |
1434 | * Initialize the specified rcu_data structure's callback list to empty. | |
1435 | */ | |
1436 | static void init_callback_list(struct rcu_data *rdp) | |
1437 | { | |
1438 | if (init_nocb_callback_list(rdp)) | |
1439 | return; | |
1440 | init_default_callback_list(rdp); | |
1441 | } | |
1442 | ||
dc35c893 PM |
1443 | /* |
1444 | * Determine the value that ->completed will have at the end of the | |
1445 | * next subsequent grace period. This is used to tag callbacks so that | |
1446 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1447 | * been dyntick-idle for an extended period with callbacks under the | |
1448 | * influence of RCU_FAST_NO_HZ. | |
1449 | * | |
1450 | * The caller must hold rnp->lock with interrupts disabled. | |
1451 | */ | |
1452 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1453 | struct rcu_node *rnp) | |
1454 | { | |
1455 | /* | |
1456 | * If RCU is idle, we just wait for the next grace period. | |
1457 | * But we can only be sure that RCU is idle if we are looking | |
1458 | * at the root rcu_node structure -- otherwise, a new grace | |
1459 | * period might have started, but just not yet gotten around | |
1460 | * to initializing the current non-root rcu_node structure. | |
1461 | */ | |
1462 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1463 | return rnp->completed + 1; | |
1464 | ||
1465 | /* | |
1466 | * Otherwise, wait for a possible partial grace period and | |
1467 | * then the subsequent full grace period. | |
1468 | */ | |
1469 | return rnp->completed + 2; | |
1470 | } | |
1471 | ||
0446be48 PM |
1472 | /* |
1473 | * Trace-event helper function for rcu_start_future_gp() and | |
1474 | * rcu_nocb_wait_gp(). | |
1475 | */ | |
1476 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1477 | unsigned long c, const char *s) |
0446be48 PM |
1478 | { |
1479 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1480 | rnp->completed, c, rnp->level, | |
1481 | rnp->grplo, rnp->grphi, s); | |
1482 | } | |
1483 | ||
1484 | /* | |
1485 | * Start some future grace period, as needed to handle newly arrived | |
1486 | * callbacks. The required future grace periods are recorded in each | |
48a7639c PM |
1487 | * rcu_node structure's ->need_future_gp field. Returns true if there |
1488 | * is reason to awaken the grace-period kthread. | |
0446be48 PM |
1489 | * |
1490 | * The caller must hold the specified rcu_node structure's ->lock. | |
1491 | */ | |
48a7639c PM |
1492 | static bool __maybe_unused |
1493 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
1494 | unsigned long *c_out) | |
0446be48 PM |
1495 | { |
1496 | unsigned long c; | |
1497 | int i; | |
48a7639c | 1498 | bool ret = false; |
0446be48 PM |
1499 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); |
1500 | ||
1501 | /* | |
1502 | * Pick up grace-period number for new callbacks. If this | |
1503 | * grace period is already marked as needed, return to the caller. | |
1504 | */ | |
1505 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1506 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1507 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1508 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
48a7639c | 1509 | goto out; |
0446be48 PM |
1510 | } |
1511 | ||
1512 | /* | |
1513 | * If either this rcu_node structure or the root rcu_node structure | |
1514 | * believe that a grace period is in progress, then we must wait | |
1515 | * for the one following, which is in "c". Because our request | |
1516 | * will be noticed at the end of the current grace period, we don't | |
48bd8e9b PK |
1517 | * need to explicitly start one. We only do the lockless check |
1518 | * of rnp_root's fields if the current rcu_node structure thinks | |
1519 | * there is no grace period in flight, and because we hold rnp->lock, | |
1520 | * the only possible change is when rnp_root's two fields are | |
1521 | * equal, in which case rnp_root->gpnum might be concurrently | |
1522 | * incremented. But that is OK, as it will just result in our | |
1523 | * doing some extra useless work. | |
0446be48 PM |
1524 | */ |
1525 | if (rnp->gpnum != rnp->completed || | |
7d0ae808 | 1526 | READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) { |
0446be48 | 1527 | rnp->need_future_gp[c & 0x1]++; |
f7f7bac9 | 1528 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
48a7639c | 1529 | goto out; |
0446be48 PM |
1530 | } |
1531 | ||
1532 | /* | |
1533 | * There might be no grace period in progress. If we don't already | |
1534 | * hold it, acquire the root rcu_node structure's lock in order to | |
1535 | * start one (if needed). | |
1536 | */ | |
6303b9c8 | 1537 | if (rnp != rnp_root) { |
0446be48 | 1538 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 PM |
1539 | smp_mb__after_unlock_lock(); |
1540 | } | |
0446be48 PM |
1541 | |
1542 | /* | |
1543 | * Get a new grace-period number. If there really is no grace | |
1544 | * period in progress, it will be smaller than the one we obtained | |
1545 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1546 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1547 | */ | |
1548 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1549 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1550 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1551 | rdp->nxtcompleted[i] = c; | |
1552 | ||
1553 | /* | |
1554 | * If the needed for the required grace period is already | |
1555 | * recorded, trace and leave. | |
1556 | */ | |
1557 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1558 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1559 | goto unlock_out; |
1560 | } | |
1561 | ||
1562 | /* Record the need for the future grace period. */ | |
1563 | rnp_root->need_future_gp[c & 0x1]++; | |
1564 | ||
1565 | /* If a grace period is not already in progress, start one. */ | |
1566 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1567 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1568 | } else { |
f7f7bac9 | 1569 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
48a7639c | 1570 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1571 | } |
1572 | unlock_out: | |
1573 | if (rnp != rnp_root) | |
1574 | raw_spin_unlock(&rnp_root->lock); | |
48a7639c PM |
1575 | out: |
1576 | if (c_out != NULL) | |
1577 | *c_out = c; | |
1578 | return ret; | |
0446be48 PM |
1579 | } |
1580 | ||
1581 | /* | |
1582 | * Clean up any old requests for the just-ended grace period. Also return | |
1583 | * whether any additional grace periods have been requested. Also invoke | |
1584 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1585 | * waiting for this grace period to complete. | |
1586 | */ | |
1587 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1588 | { | |
1589 | int c = rnp->completed; | |
1590 | int needmore; | |
1591 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1592 | ||
1593 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1594 | rnp->need_future_gp[c & 0x1] = 0; | |
1595 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1596 | trace_rcu_future_gp(rnp, rdp, c, |
1597 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1598 | return needmore; |
1599 | } | |
1600 | ||
48a7639c PM |
1601 | /* |
1602 | * Awaken the grace-period kthread for the specified flavor of RCU. | |
1603 | * Don't do a self-awaken, and don't bother awakening when there is | |
1604 | * nothing for the grace-period kthread to do (as in several CPUs | |
1605 | * raced to awaken, and we lost), and finally don't try to awaken | |
1606 | * a kthread that has not yet been created. | |
1607 | */ | |
1608 | static void rcu_gp_kthread_wake(struct rcu_state *rsp) | |
1609 | { | |
1610 | if (current == rsp->gp_kthread || | |
7d0ae808 | 1611 | !READ_ONCE(rsp->gp_flags) || |
48a7639c PM |
1612 | !rsp->gp_kthread) |
1613 | return; | |
1614 | wake_up(&rsp->gp_wq); | |
1615 | } | |
1616 | ||
dc35c893 PM |
1617 | /* |
1618 | * If there is room, assign a ->completed number to any callbacks on | |
1619 | * this CPU that have not already been assigned. Also accelerate any | |
1620 | * callbacks that were previously assigned a ->completed number that has | |
1621 | * since proven to be too conservative, which can happen if callbacks get | |
1622 | * assigned a ->completed number while RCU is idle, but with reference to | |
1623 | * a non-root rcu_node structure. This function is idempotent, so it does | |
48a7639c PM |
1624 | * not hurt to call it repeatedly. Returns an flag saying that we should |
1625 | * awaken the RCU grace-period kthread. | |
dc35c893 PM |
1626 | * |
1627 | * The caller must hold rnp->lock with interrupts disabled. | |
1628 | */ | |
48a7639c | 1629 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1630 | struct rcu_data *rdp) |
1631 | { | |
1632 | unsigned long c; | |
1633 | int i; | |
48a7639c | 1634 | bool ret; |
dc35c893 PM |
1635 | |
1636 | /* If the CPU has no callbacks, nothing to do. */ | |
1637 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1638 | return false; |
dc35c893 PM |
1639 | |
1640 | /* | |
1641 | * Starting from the sublist containing the callbacks most | |
1642 | * recently assigned a ->completed number and working down, find the | |
1643 | * first sublist that is not assignable to an upcoming grace period. | |
1644 | * Such a sublist has something in it (first two tests) and has | |
1645 | * a ->completed number assigned that will complete sooner than | |
1646 | * the ->completed number for newly arrived callbacks (last test). | |
1647 | * | |
1648 | * The key point is that any later sublist can be assigned the | |
1649 | * same ->completed number as the newly arrived callbacks, which | |
1650 | * means that the callbacks in any of these later sublist can be | |
1651 | * grouped into a single sublist, whether or not they have already | |
1652 | * been assigned a ->completed number. | |
1653 | */ | |
1654 | c = rcu_cbs_completed(rsp, rnp); | |
1655 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1656 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1657 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1658 | break; | |
1659 | ||
1660 | /* | |
1661 | * If there are no sublist for unassigned callbacks, leave. | |
1662 | * At the same time, advance "i" one sublist, so that "i" will | |
1663 | * index into the sublist where all the remaining callbacks should | |
1664 | * be grouped into. | |
1665 | */ | |
1666 | if (++i >= RCU_NEXT_TAIL) | |
48a7639c | 1667 | return false; |
dc35c893 PM |
1668 | |
1669 | /* | |
1670 | * Assign all subsequent callbacks' ->completed number to the next | |
1671 | * full grace period and group them all in the sublist initially | |
1672 | * indexed by "i". | |
1673 | */ | |
1674 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1675 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1676 | rdp->nxtcompleted[i] = c; | |
1677 | } | |
910ee45d | 1678 | /* Record any needed additional grace periods. */ |
48a7639c | 1679 | ret = rcu_start_future_gp(rnp, rdp, NULL); |
6d4b418c PM |
1680 | |
1681 | /* Trace depending on how much we were able to accelerate. */ | |
1682 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1683 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1684 | else |
f7f7bac9 | 1685 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
48a7639c | 1686 | return ret; |
dc35c893 PM |
1687 | } |
1688 | ||
1689 | /* | |
1690 | * Move any callbacks whose grace period has completed to the | |
1691 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1692 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1693 | * sublist. This function is idempotent, so it does not hurt to | |
1694 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1695 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1696 | * |
1697 | * The caller must hold rnp->lock with interrupts disabled. | |
1698 | */ | |
48a7639c | 1699 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1700 | struct rcu_data *rdp) |
1701 | { | |
1702 | int i, j; | |
1703 | ||
1704 | /* If the CPU has no callbacks, nothing to do. */ | |
1705 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1706 | return false; |
dc35c893 PM |
1707 | |
1708 | /* | |
1709 | * Find all callbacks whose ->completed numbers indicate that they | |
1710 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1711 | */ | |
1712 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1713 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1714 | break; | |
1715 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1716 | } | |
1717 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1718 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1719 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1720 | ||
1721 | /* Copy down callbacks to fill in empty sublists. */ | |
1722 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1723 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1724 | break; | |
1725 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1726 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1727 | } | |
1728 | ||
1729 | /* Classify any remaining callbacks. */ | |
48a7639c | 1730 | return rcu_accelerate_cbs(rsp, rnp, rdp); |
dc35c893 PM |
1731 | } |
1732 | ||
d09b62df | 1733 | /* |
ba9fbe95 PM |
1734 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1735 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1736 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1737 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1738 | */ |
48a7639c PM |
1739 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, |
1740 | struct rcu_data *rdp) | |
d09b62df | 1741 | { |
48a7639c PM |
1742 | bool ret; |
1743 | ||
ba9fbe95 | 1744 | /* Handle the ends of any preceding grace periods first. */ |
e3663b10 | 1745 | if (rdp->completed == rnp->completed && |
7d0ae808 | 1746 | !unlikely(READ_ONCE(rdp->gpwrap))) { |
d09b62df | 1747 | |
ba9fbe95 | 1748 | /* No grace period end, so just accelerate recent callbacks. */ |
48a7639c | 1749 | ret = rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1750 | |
dc35c893 PM |
1751 | } else { |
1752 | ||
1753 | /* Advance callbacks. */ | |
48a7639c | 1754 | ret = rcu_advance_cbs(rsp, rnp, rdp); |
d09b62df PM |
1755 | |
1756 | /* Remember that we saw this grace-period completion. */ | |
1757 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1758 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1759 | } |
398ebe60 | 1760 | |
7d0ae808 | 1761 | if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) { |
6eaef633 PM |
1762 | /* |
1763 | * If the current grace period is waiting for this CPU, | |
1764 | * set up to detect a quiescent state, otherwise don't | |
1765 | * go looking for one. | |
1766 | */ | |
1767 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1768 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
5b74c458 | 1769 | rdp->cpu_no_qs.b.norm = true; |
5cd37193 | 1770 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); |
97c668b8 | 1771 | rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask); |
6eaef633 | 1772 | zero_cpu_stall_ticks(rdp); |
7d0ae808 | 1773 | WRITE_ONCE(rdp->gpwrap, false); |
6eaef633 | 1774 | } |
48a7639c | 1775 | return ret; |
6eaef633 PM |
1776 | } |
1777 | ||
d34ea322 | 1778 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1779 | { |
1780 | unsigned long flags; | |
48a7639c | 1781 | bool needwake; |
6eaef633 PM |
1782 | struct rcu_node *rnp; |
1783 | ||
1784 | local_irq_save(flags); | |
1785 | rnp = rdp->mynode; | |
7d0ae808 PM |
1786 | if ((rdp->gpnum == READ_ONCE(rnp->gpnum) && |
1787 | rdp->completed == READ_ONCE(rnp->completed) && | |
1788 | !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */ | |
6eaef633 PM |
1789 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
1790 | local_irq_restore(flags); | |
1791 | return; | |
1792 | } | |
6303b9c8 | 1793 | smp_mb__after_unlock_lock(); |
48a7639c | 1794 | needwake = __note_gp_changes(rsp, rnp, rdp); |
6eaef633 | 1795 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
48a7639c PM |
1796 | if (needwake) |
1797 | rcu_gp_kthread_wake(rsp); | |
6eaef633 PM |
1798 | } |
1799 | ||
0f41c0dd PM |
1800 | static void rcu_gp_slow(struct rcu_state *rsp, int delay) |
1801 | { | |
1802 | if (delay > 0 && | |
1803 | !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay))) | |
1804 | schedule_timeout_uninterruptible(delay); | |
1805 | } | |
1806 | ||
b3dbec76 | 1807 | /* |
f7be8209 | 1808 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1809 | */ |
7fdefc10 | 1810 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 | 1811 | { |
0aa04b05 | 1812 | unsigned long oldmask; |
b3dbec76 | 1813 | struct rcu_data *rdp; |
7fdefc10 | 1814 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1815 | |
7d0ae808 | 1816 | WRITE_ONCE(rsp->gp_activity, jiffies); |
7fdefc10 | 1817 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1818 | smp_mb__after_unlock_lock(); |
7d0ae808 | 1819 | if (!READ_ONCE(rsp->gp_flags)) { |
f7be8209 PM |
1820 | /* Spurious wakeup, tell caller to go back to sleep. */ |
1821 | raw_spin_unlock_irq(&rnp->lock); | |
1822 | return 0; | |
1823 | } | |
7d0ae808 | 1824 | WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */ |
b3dbec76 | 1825 | |
f7be8209 PM |
1826 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1827 | /* | |
1828 | * Grace period already in progress, don't start another. | |
1829 | * Not supposed to be able to happen. | |
1830 | */ | |
7fdefc10 PM |
1831 | raw_spin_unlock_irq(&rnp->lock); |
1832 | return 0; | |
1833 | } | |
1834 | ||
7fdefc10 | 1835 | /* Advance to a new grace period and initialize state. */ |
26cdfedf | 1836 | record_gp_stall_check_time(rsp); |
765a3f4f PM |
1837 | /* Record GP times before starting GP, hence smp_store_release(). */ |
1838 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); | |
f7f7bac9 | 1839 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1840 | raw_spin_unlock_irq(&rnp->lock); |
1841 | ||
0aa04b05 PM |
1842 | /* |
1843 | * Apply per-leaf buffered online and offline operations to the | |
1844 | * rcu_node tree. Note that this new grace period need not wait | |
1845 | * for subsequent online CPUs, and that quiescent-state forcing | |
1846 | * will handle subsequent offline CPUs. | |
1847 | */ | |
1848 | rcu_for_each_leaf_node(rsp, rnp) { | |
0f41c0dd | 1849 | rcu_gp_slow(rsp, gp_preinit_delay); |
0aa04b05 PM |
1850 | raw_spin_lock_irq(&rnp->lock); |
1851 | smp_mb__after_unlock_lock(); | |
1852 | if (rnp->qsmaskinit == rnp->qsmaskinitnext && | |
1853 | !rnp->wait_blkd_tasks) { | |
1854 | /* Nothing to do on this leaf rcu_node structure. */ | |
1855 | raw_spin_unlock_irq(&rnp->lock); | |
1856 | continue; | |
1857 | } | |
1858 | ||
1859 | /* Record old state, apply changes to ->qsmaskinit field. */ | |
1860 | oldmask = rnp->qsmaskinit; | |
1861 | rnp->qsmaskinit = rnp->qsmaskinitnext; | |
1862 | ||
1863 | /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ | |
1864 | if (!oldmask != !rnp->qsmaskinit) { | |
1865 | if (!oldmask) /* First online CPU for this rcu_node. */ | |
1866 | rcu_init_new_rnp(rnp); | |
1867 | else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */ | |
1868 | rnp->wait_blkd_tasks = true; | |
1869 | else /* Last offline CPU and can propagate. */ | |
1870 | rcu_cleanup_dead_rnp(rnp); | |
1871 | } | |
1872 | ||
1873 | /* | |
1874 | * If all waited-on tasks from prior grace period are | |
1875 | * done, and if all this rcu_node structure's CPUs are | |
1876 | * still offline, propagate up the rcu_node tree and | |
1877 | * clear ->wait_blkd_tasks. Otherwise, if one of this | |
1878 | * rcu_node structure's CPUs has since come back online, | |
1879 | * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp() | |
1880 | * checks for this, so just call it unconditionally). | |
1881 | */ | |
1882 | if (rnp->wait_blkd_tasks && | |
1883 | (!rcu_preempt_has_tasks(rnp) || | |
1884 | rnp->qsmaskinit)) { | |
1885 | rnp->wait_blkd_tasks = false; | |
1886 | rcu_cleanup_dead_rnp(rnp); | |
1887 | } | |
1888 | ||
1889 | raw_spin_unlock_irq(&rnp->lock); | |
1890 | } | |
7fdefc10 PM |
1891 | |
1892 | /* | |
1893 | * Set the quiescent-state-needed bits in all the rcu_node | |
1894 | * structures for all currently online CPUs in breadth-first order, | |
1895 | * starting from the root rcu_node structure, relying on the layout | |
1896 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1897 | * will access only the leaves of the hierarchy, thus seeing that no | |
1898 | * grace period is in progress, at least until the corresponding | |
1899 | * leaf node has been initialized. In addition, we have excluded | |
1900 | * CPU-hotplug operations. | |
1901 | * | |
1902 | * The grace period cannot complete until the initialization | |
1903 | * process finishes, because this kthread handles both. | |
1904 | */ | |
1905 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
0f41c0dd | 1906 | rcu_gp_slow(rsp, gp_init_delay); |
b3dbec76 | 1907 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1908 | smp_mb__after_unlock_lock(); |
b3dbec76 | 1909 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1910 | rcu_preempt_check_blocked_tasks(rnp); |
1911 | rnp->qsmask = rnp->qsmaskinit; | |
7d0ae808 | 1912 | WRITE_ONCE(rnp->gpnum, rsp->gpnum); |
3f47da0f | 1913 | if (WARN_ON_ONCE(rnp->completed != rsp->completed)) |
7d0ae808 | 1914 | WRITE_ONCE(rnp->completed, rsp->completed); |
7fdefc10 | 1915 | if (rnp == rdp->mynode) |
48a7639c | 1916 | (void)__note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1917 | rcu_preempt_boost_start_gp(rnp); |
1918 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1919 | rnp->level, rnp->grplo, | |
1920 | rnp->grphi, rnp->qsmask); | |
1921 | raw_spin_unlock_irq(&rnp->lock); | |
bde6c3aa | 1922 | cond_resched_rcu_qs(); |
7d0ae808 | 1923 | WRITE_ONCE(rsp->gp_activity, jiffies); |
7fdefc10 | 1924 | } |
b3dbec76 | 1925 | |
7fdefc10 PM |
1926 | return 1; |
1927 | } | |
b3dbec76 | 1928 | |
b9a425cf PM |
1929 | /* |
1930 | * Helper function for wait_event_interruptible_timeout() wakeup | |
1931 | * at force-quiescent-state time. | |
1932 | */ | |
1933 | static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp) | |
1934 | { | |
1935 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1936 | ||
1937 | /* Someone like call_rcu() requested a force-quiescent-state scan. */ | |
1938 | *gfp = READ_ONCE(rsp->gp_flags); | |
1939 | if (*gfp & RCU_GP_FLAG_FQS) | |
1940 | return true; | |
1941 | ||
1942 | /* The current grace period has completed. */ | |
1943 | if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp)) | |
1944 | return true; | |
1945 | ||
1946 | return false; | |
1947 | } | |
1948 | ||
4cdfc175 PM |
1949 | /* |
1950 | * Do one round of quiescent-state forcing. | |
1951 | */ | |
77f81fe0 | 1952 | static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time) |
4cdfc175 | 1953 | { |
217af2a2 PM |
1954 | bool isidle = false; |
1955 | unsigned long maxj; | |
4cdfc175 PM |
1956 | struct rcu_node *rnp = rcu_get_root(rsp); |
1957 | ||
7d0ae808 | 1958 | WRITE_ONCE(rsp->gp_activity, jiffies); |
4cdfc175 | 1959 | rsp->n_force_qs++; |
77f81fe0 | 1960 | if (first_time) { |
4cdfc175 | 1961 | /* Collect dyntick-idle snapshots. */ |
0edd1b17 | 1962 | if (is_sysidle_rcu_state(rsp)) { |
e02b2edf | 1963 | isidle = true; |
0edd1b17 PM |
1964 | maxj = jiffies - ULONG_MAX / 4; |
1965 | } | |
217af2a2 PM |
1966 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1967 | &isidle, &maxj); | |
0edd1b17 | 1968 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1969 | } else { |
1970 | /* Handle dyntick-idle and offline CPUs. */ | |
675da67f | 1971 | isidle = true; |
217af2a2 | 1972 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1973 | } |
1974 | /* Clear flag to prevent immediate re-entry. */ | |
7d0ae808 | 1975 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { |
4cdfc175 | 1976 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1977 | smp_mb__after_unlock_lock(); |
7d0ae808 PM |
1978 | WRITE_ONCE(rsp->gp_flags, |
1979 | READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS); | |
4cdfc175 PM |
1980 | raw_spin_unlock_irq(&rnp->lock); |
1981 | } | |
4cdfc175 PM |
1982 | } |
1983 | ||
7fdefc10 PM |
1984 | /* |
1985 | * Clean up after the old grace period. | |
1986 | */ | |
4cdfc175 | 1987 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1988 | { |
1989 | unsigned long gp_duration; | |
48a7639c | 1990 | bool needgp = false; |
dae6e64d | 1991 | int nocb = 0; |
7fdefc10 PM |
1992 | struct rcu_data *rdp; |
1993 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1994 | |
7d0ae808 | 1995 | WRITE_ONCE(rsp->gp_activity, jiffies); |
7fdefc10 | 1996 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 1997 | smp_mb__after_unlock_lock(); |
7fdefc10 PM |
1998 | gp_duration = jiffies - rsp->gp_start; |
1999 | if (gp_duration > rsp->gp_max) | |
2000 | rsp->gp_max = gp_duration; | |
b3dbec76 | 2001 | |
7fdefc10 PM |
2002 | /* |
2003 | * We know the grace period is complete, but to everyone else | |
2004 | * it appears to still be ongoing. But it is also the case | |
2005 | * that to everyone else it looks like there is nothing that | |
2006 | * they can do to advance the grace period. It is therefore | |
2007 | * safe for us to drop the lock in order to mark the grace | |
2008 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 2009 | */ |
5d4b8659 | 2010 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 2011 | |
5d4b8659 PM |
2012 | /* |
2013 | * Propagate new ->completed value to rcu_node structures so | |
2014 | * that other CPUs don't have to wait until the start of the next | |
2015 | * grace period to process their callbacks. This also avoids | |
2016 | * some nasty RCU grace-period initialization races by forcing | |
2017 | * the end of the current grace period to be completely recorded in | |
2018 | * all of the rcu_node structures before the beginning of the next | |
2019 | * grace period is recorded in any of the rcu_node structures. | |
2020 | */ | |
2021 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
755609a9 | 2022 | raw_spin_lock_irq(&rnp->lock); |
6303b9c8 | 2023 | smp_mb__after_unlock_lock(); |
5c60d25f PM |
2024 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
2025 | WARN_ON_ONCE(rnp->qsmask); | |
7d0ae808 | 2026 | WRITE_ONCE(rnp->completed, rsp->gpnum); |
b11cc576 PM |
2027 | rdp = this_cpu_ptr(rsp->rda); |
2028 | if (rnp == rdp->mynode) | |
48a7639c | 2029 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; |
78e4bc34 | 2030 | /* smp_mb() provided by prior unlock-lock pair. */ |
0446be48 | 2031 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 | 2032 | raw_spin_unlock_irq(&rnp->lock); |
bde6c3aa | 2033 | cond_resched_rcu_qs(); |
7d0ae808 | 2034 | WRITE_ONCE(rsp->gp_activity, jiffies); |
0f41c0dd | 2035 | rcu_gp_slow(rsp, gp_cleanup_delay); |
7fdefc10 | 2036 | } |
5d4b8659 PM |
2037 | rnp = rcu_get_root(rsp); |
2038 | raw_spin_lock_irq(&rnp->lock); | |
765a3f4f | 2039 | smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */ |
dae6e64d | 2040 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 | 2041 | |
765a3f4f | 2042 | /* Declare grace period done. */ |
7d0ae808 | 2043 | WRITE_ONCE(rsp->completed, rsp->gpnum); |
f7f7bac9 | 2044 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
77f81fe0 | 2045 | rsp->gp_state = RCU_GP_IDLE; |
5d4b8659 | 2046 | rdp = this_cpu_ptr(rsp->rda); |
48a7639c PM |
2047 | /* Advance CBs to reduce false positives below. */ |
2048 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; | |
2049 | if (needgp || cpu_needs_another_gp(rsp, rdp)) { | |
7d0ae808 | 2050 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); |
bb311ecc | 2051 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2052 | READ_ONCE(rsp->gpnum), |
bb311ecc PM |
2053 | TPS("newreq")); |
2054 | } | |
7fdefc10 | 2055 | raw_spin_unlock_irq(&rnp->lock); |
7fdefc10 PM |
2056 | } |
2057 | ||
2058 | /* | |
2059 | * Body of kthread that handles grace periods. | |
2060 | */ | |
2061 | static int __noreturn rcu_gp_kthread(void *arg) | |
2062 | { | |
77f81fe0 | 2063 | bool first_gp_fqs; |
88d6df61 | 2064 | int gf; |
d40011f6 | 2065 | unsigned long j; |
4cdfc175 | 2066 | int ret; |
7fdefc10 PM |
2067 | struct rcu_state *rsp = arg; |
2068 | struct rcu_node *rnp = rcu_get_root(rsp); | |
2069 | ||
5871968d | 2070 | rcu_bind_gp_kthread(); |
7fdefc10 PM |
2071 | for (;;) { |
2072 | ||
2073 | /* Handle grace-period start. */ | |
2074 | for (;;) { | |
63c4db78 | 2075 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2076 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2077 | TPS("reqwait")); |
afea227f | 2078 | rsp->gp_state = RCU_GP_WAIT_GPS; |
4cdfc175 | 2079 | wait_event_interruptible(rsp->gp_wq, |
7d0ae808 | 2080 | READ_ONCE(rsp->gp_flags) & |
4cdfc175 | 2081 | RCU_GP_FLAG_INIT); |
319362c9 | 2082 | rsp->gp_state = RCU_GP_DONE_GPS; |
78e4bc34 | 2083 | /* Locking provides needed memory barrier. */ |
f7be8209 | 2084 | if (rcu_gp_init(rsp)) |
7fdefc10 | 2085 | break; |
bde6c3aa | 2086 | cond_resched_rcu_qs(); |
7d0ae808 | 2087 | WRITE_ONCE(rsp->gp_activity, jiffies); |
73a860cd | 2088 | WARN_ON(signal_pending(current)); |
63c4db78 | 2089 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2090 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2091 | TPS("reqwaitsig")); |
7fdefc10 | 2092 | } |
cabc49c1 | 2093 | |
4cdfc175 | 2094 | /* Handle quiescent-state forcing. */ |
77f81fe0 | 2095 | first_gp_fqs = true; |
d40011f6 PM |
2096 | j = jiffies_till_first_fqs; |
2097 | if (j > HZ) { | |
2098 | j = HZ; | |
2099 | jiffies_till_first_fqs = HZ; | |
2100 | } | |
88d6df61 | 2101 | ret = 0; |
cabc49c1 | 2102 | for (;;) { |
88d6df61 PM |
2103 | if (!ret) |
2104 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 | 2105 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2106 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2107 | TPS("fqswait")); |
afea227f | 2108 | rsp->gp_state = RCU_GP_WAIT_FQS; |
4cdfc175 | 2109 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
b9a425cf | 2110 | rcu_gp_fqs_check_wake(rsp, &gf), j); |
32bb1c79 | 2111 | rsp->gp_state = RCU_GP_DOING_FQS; |
78e4bc34 | 2112 | /* Locking provides needed memory barriers. */ |
4cdfc175 | 2113 | /* If grace period done, leave loop. */ |
7d0ae808 | 2114 | if (!READ_ONCE(rnp->qsmask) && |
4cdfc175 | 2115 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 2116 | break; |
4cdfc175 | 2117 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
2118 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
2119 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 | 2120 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2121 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2122 | TPS("fqsstart")); |
77f81fe0 PM |
2123 | rcu_gp_fqs(rsp, first_gp_fqs); |
2124 | first_gp_fqs = false; | |
63c4db78 | 2125 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2126 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2127 | TPS("fqsend")); |
bde6c3aa | 2128 | cond_resched_rcu_qs(); |
7d0ae808 | 2129 | WRITE_ONCE(rsp->gp_activity, jiffies); |
4cdfc175 PM |
2130 | } else { |
2131 | /* Deal with stray signal. */ | |
bde6c3aa | 2132 | cond_resched_rcu_qs(); |
7d0ae808 | 2133 | WRITE_ONCE(rsp->gp_activity, jiffies); |
73a860cd | 2134 | WARN_ON(signal_pending(current)); |
63c4db78 | 2135 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2136 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2137 | TPS("fqswaitsig")); |
4cdfc175 | 2138 | } |
d40011f6 PM |
2139 | j = jiffies_till_next_fqs; |
2140 | if (j > HZ) { | |
2141 | j = HZ; | |
2142 | jiffies_till_next_fqs = HZ; | |
2143 | } else if (j < 1) { | |
2144 | j = 1; | |
2145 | jiffies_till_next_fqs = 1; | |
2146 | } | |
cabc49c1 | 2147 | } |
4cdfc175 PM |
2148 | |
2149 | /* Handle grace-period end. */ | |
319362c9 | 2150 | rsp->gp_state = RCU_GP_CLEANUP; |
4cdfc175 | 2151 | rcu_gp_cleanup(rsp); |
319362c9 | 2152 | rsp->gp_state = RCU_GP_CLEANED; |
b3dbec76 | 2153 | } |
b3dbec76 PM |
2154 | } |
2155 | ||
64db4cff PM |
2156 | /* |
2157 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
2158 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 2159 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
2160 | * |
2161 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
2162 | * invoke this function. This can happen when the dying CPU reports its | |
2163 | * quiescent state. | |
48a7639c PM |
2164 | * |
2165 | * Returns true if the grace-period kthread must be awakened. | |
64db4cff | 2166 | */ |
48a7639c | 2167 | static bool |
910ee45d PM |
2168 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
2169 | struct rcu_data *rdp) | |
64db4cff | 2170 | { |
b8462084 | 2171 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 2172 | /* |
b3dbec76 | 2173 | * Either we have not yet spawned the grace-period |
62da1921 PM |
2174 | * task, this CPU does not need another grace period, |
2175 | * or a grace period is already in progress. | |
b3dbec76 | 2176 | * Either way, don't start a new grace period. |
afe24b12 | 2177 | */ |
48a7639c | 2178 | return false; |
afe24b12 | 2179 | } |
7d0ae808 PM |
2180 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); |
2181 | trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum), | |
bb311ecc | 2182 | TPS("newreq")); |
62da1921 | 2183 | |
016a8d5b SR |
2184 | /* |
2185 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c | 2186 | * could cause possible deadlocks with the rq->lock. Defer |
48a7639c | 2187 | * the wakeup to our caller. |
016a8d5b | 2188 | */ |
48a7639c | 2189 | return true; |
64db4cff PM |
2190 | } |
2191 | ||
910ee45d PM |
2192 | /* |
2193 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
2194 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
2195 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
2196 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
2197 | * that is encountered beforehand. | |
48a7639c PM |
2198 | * |
2199 | * Returns true if the grace-period kthread needs to be awakened. | |
910ee45d | 2200 | */ |
48a7639c | 2201 | static bool rcu_start_gp(struct rcu_state *rsp) |
910ee45d PM |
2202 | { |
2203 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
2204 | struct rcu_node *rnp = rcu_get_root(rsp); | |
48a7639c | 2205 | bool ret = false; |
910ee45d PM |
2206 | |
2207 | /* | |
2208 | * If there is no grace period in progress right now, any | |
2209 | * callbacks we have up to this point will be satisfied by the | |
2210 | * next grace period. Also, advancing the callbacks reduces the | |
2211 | * probability of false positives from cpu_needs_another_gp() | |
2212 | * resulting in pointless grace periods. So, advance callbacks | |
2213 | * then start the grace period! | |
2214 | */ | |
48a7639c PM |
2215 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; |
2216 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; | |
2217 | return ret; | |
910ee45d PM |
2218 | } |
2219 | ||
f41d911f | 2220 | /* |
d3f6bad3 PM |
2221 | * Report a full set of quiescent states to the specified rcu_state |
2222 | * data structure. This involves cleaning up after the prior grace | |
2223 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
2224 | * if one is needed. Note that the caller must hold rnp->lock, which |
2225 | * is released before return. | |
f41d911f | 2226 | */ |
d3f6bad3 | 2227 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 2228 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 2229 | { |
fc2219d4 | 2230 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cd73ca21 | 2231 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); |
cabc49c1 | 2232 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
2aa792e6 | 2233 | rcu_gp_kthread_wake(rsp); |
f41d911f PM |
2234 | } |
2235 | ||
64db4cff | 2236 | /* |
d3f6bad3 PM |
2237 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
2238 | * Allows quiescent states for a group of CPUs to be reported at one go | |
2239 | * to the specified rcu_node structure, though all the CPUs in the group | |
654e9533 PM |
2240 | * must be represented by the same rcu_node structure (which need not be a |
2241 | * leaf rcu_node structure, though it often will be). The gps parameter | |
2242 | * is the grace-period snapshot, which means that the quiescent states | |
2243 | * are valid only if rnp->gpnum is equal to gps. That structure's lock | |
2244 | * must be held upon entry, and it is released before return. | |
64db4cff PM |
2245 | */ |
2246 | static void | |
d3f6bad3 | 2247 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
654e9533 | 2248 | struct rcu_node *rnp, unsigned long gps, unsigned long flags) |
64db4cff PM |
2249 | __releases(rnp->lock) |
2250 | { | |
654e9533 | 2251 | unsigned long oldmask = 0; |
28ecd580 PM |
2252 | struct rcu_node *rnp_c; |
2253 | ||
64db4cff PM |
2254 | /* Walk up the rcu_node hierarchy. */ |
2255 | for (;;) { | |
654e9533 | 2256 | if (!(rnp->qsmask & mask) || rnp->gpnum != gps) { |
64db4cff | 2257 | |
654e9533 PM |
2258 | /* |
2259 | * Our bit has already been cleared, or the | |
2260 | * relevant grace period is already over, so done. | |
2261 | */ | |
1304afb2 | 2262 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2263 | return; |
2264 | } | |
654e9533 | 2265 | WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ |
64db4cff | 2266 | rnp->qsmask &= ~mask; |
d4c08f2a PM |
2267 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
2268 | mask, rnp->qsmask, rnp->level, | |
2269 | rnp->grplo, rnp->grphi, | |
2270 | !!rnp->gp_tasks); | |
27f4d280 | 2271 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
2272 | |
2273 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 2274 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2275 | return; |
2276 | } | |
2277 | mask = rnp->grpmask; | |
2278 | if (rnp->parent == NULL) { | |
2279 | ||
2280 | /* No more levels. Exit loop holding root lock. */ | |
2281 | ||
2282 | break; | |
2283 | } | |
1304afb2 | 2284 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 2285 | rnp_c = rnp; |
64db4cff | 2286 | rnp = rnp->parent; |
1304afb2 | 2287 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2288 | smp_mb__after_unlock_lock(); |
654e9533 | 2289 | oldmask = rnp_c->qsmask; |
64db4cff PM |
2290 | } |
2291 | ||
2292 | /* | |
2293 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 2294 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 2295 | * to clean up and start the next grace period if one is needed. |
64db4cff | 2296 | */ |
d3f6bad3 | 2297 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
2298 | } |
2299 | ||
cc99a310 PM |
2300 | /* |
2301 | * Record a quiescent state for all tasks that were previously queued | |
2302 | * on the specified rcu_node structure and that were blocking the current | |
2303 | * RCU grace period. The caller must hold the specified rnp->lock with | |
2304 | * irqs disabled, and this lock is released upon return, but irqs remain | |
2305 | * disabled. | |
2306 | */ | |
0aa04b05 | 2307 | static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp, |
cc99a310 PM |
2308 | struct rcu_node *rnp, unsigned long flags) |
2309 | __releases(rnp->lock) | |
2310 | { | |
654e9533 | 2311 | unsigned long gps; |
cc99a310 PM |
2312 | unsigned long mask; |
2313 | struct rcu_node *rnp_p; | |
2314 | ||
a77da14c PM |
2315 | if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p || |
2316 | rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { | |
cc99a310 PM |
2317 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
2318 | return; /* Still need more quiescent states! */ | |
2319 | } | |
2320 | ||
2321 | rnp_p = rnp->parent; | |
2322 | if (rnp_p == NULL) { | |
2323 | /* | |
a77da14c PM |
2324 | * Only one rcu_node structure in the tree, so don't |
2325 | * try to report up to its nonexistent parent! | |
cc99a310 PM |
2326 | */ |
2327 | rcu_report_qs_rsp(rsp, flags); | |
2328 | return; | |
2329 | } | |
2330 | ||
654e9533 PM |
2331 | /* Report up the rest of the hierarchy, tracking current ->gpnum. */ |
2332 | gps = rnp->gpnum; | |
cc99a310 PM |
2333 | mask = rnp->grpmask; |
2334 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2335 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
2336 | smp_mb__after_unlock_lock(); | |
654e9533 | 2337 | rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags); |
cc99a310 PM |
2338 | } |
2339 | ||
64db4cff | 2340 | /* |
d3f6bad3 PM |
2341 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
2342 | * structure. This must be either called from the specified CPU, or | |
2343 | * called when the specified CPU is known to be offline (and when it is | |
2344 | * also known that no other CPU is concurrently trying to help the offline | |
2345 | * CPU). The lastcomp argument is used to make sure we are still in the | |
2346 | * grace period of interest. We don't want to end the current grace period | |
2347 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
2348 | */ |
2349 | static void | |
d7d6a11e | 2350 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2351 | { |
2352 | unsigned long flags; | |
2353 | unsigned long mask; | |
48a7639c | 2354 | bool needwake; |
64db4cff PM |
2355 | struct rcu_node *rnp; |
2356 | ||
2357 | rnp = rdp->mynode; | |
1304afb2 | 2358 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2359 | smp_mb__after_unlock_lock(); |
5b74c458 | 2360 | if ((rdp->cpu_no_qs.b.norm && |
5cd37193 PM |
2361 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) || |
2362 | rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum || | |
2363 | rdp->gpwrap) { | |
64db4cff PM |
2364 | |
2365 | /* | |
e4cc1f22 PM |
2366 | * The grace period in which this quiescent state was |
2367 | * recorded has ended, so don't report it upwards. | |
2368 | * We will instead need a new quiescent state that lies | |
2369 | * within the current grace period. | |
64db4cff | 2370 | */ |
5b74c458 | 2371 | rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */ |
5cd37193 | 2372 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); |
1304afb2 | 2373 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2374 | return; |
2375 | } | |
2376 | mask = rdp->grpmask; | |
2377 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 2378 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2379 | } else { |
97c668b8 | 2380 | rdp->core_needs_qs = 0; |
64db4cff PM |
2381 | |
2382 | /* | |
2383 | * This GP can't end until cpu checks in, so all of our | |
2384 | * callbacks can be processed during the next GP. | |
2385 | */ | |
48a7639c | 2386 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 2387 | |
654e9533 PM |
2388 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); |
2389 | /* ^^^ Released rnp->lock */ | |
48a7639c PM |
2390 | if (needwake) |
2391 | rcu_gp_kthread_wake(rsp); | |
64db4cff PM |
2392 | } |
2393 | } | |
2394 | ||
2395 | /* | |
2396 | * Check to see if there is a new grace period of which this CPU | |
2397 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2398 | * Otherwise, see if this CPU has just passed through its first | |
2399 | * quiescent state for this grace period, and record that fact if so. | |
2400 | */ | |
2401 | static void | |
2402 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
2403 | { | |
05eb552b PM |
2404 | /* Check for grace-period ends and beginnings. */ |
2405 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
2406 | |
2407 | /* | |
2408 | * Does this CPU still need to do its part for current grace period? | |
2409 | * If no, return and let the other CPUs do their part as well. | |
2410 | */ | |
97c668b8 | 2411 | if (!rdp->core_needs_qs) |
64db4cff PM |
2412 | return; |
2413 | ||
2414 | /* | |
2415 | * Was there a quiescent state since the beginning of the grace | |
2416 | * period? If no, then exit and wait for the next call. | |
2417 | */ | |
5b74c458 | 2418 | if (rdp->cpu_no_qs.b.norm && |
5cd37193 | 2419 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) |
64db4cff PM |
2420 | return; |
2421 | ||
d3f6bad3 PM |
2422 | /* |
2423 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2424 | * judge of that). | |
2425 | */ | |
d7d6a11e | 2426 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
2427 | } |
2428 | ||
e74f4c45 | 2429 | /* |
b1420f1c PM |
2430 | * Send the specified CPU's RCU callbacks to the orphanage. The |
2431 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 2432 | * ->orphan_lock. |
e74f4c45 | 2433 | */ |
b1420f1c PM |
2434 | static void |
2435 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
2436 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 2437 | { |
3fbfbf7a | 2438 | /* No-CBs CPUs do not have orphanable callbacks. */ |
ea46351c | 2439 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
2440 | return; |
2441 | ||
b1420f1c PM |
2442 | /* |
2443 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
2444 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
2445 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 2446 | */ |
a50c3af9 | 2447 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2448 | rsp->qlen_lazy += rdp->qlen_lazy; |
2449 | rsp->qlen += rdp->qlen; | |
2450 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 2451 | rdp->qlen_lazy = 0; |
7d0ae808 | 2452 | WRITE_ONCE(rdp->qlen, 0); |
a50c3af9 PM |
2453 | } |
2454 | ||
2455 | /* | |
b1420f1c PM |
2456 | * Next, move those callbacks still needing a grace period to |
2457 | * the orphanage, where some other CPU will pick them up. | |
2458 | * Some of the callbacks might have gone partway through a grace | |
2459 | * period, but that is too bad. They get to start over because we | |
2460 | * cannot assume that grace periods are synchronized across CPUs. | |
2461 | * We don't bother updating the ->nxttail[] array yet, instead | |
2462 | * we just reset the whole thing later on. | |
a50c3af9 | 2463 | */ |
b1420f1c PM |
2464 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
2465 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2466 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2467 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
2468 | } |
2469 | ||
2470 | /* | |
b1420f1c PM |
2471 | * Then move the ready-to-invoke callbacks to the orphanage, |
2472 | * where some other CPU will pick them up. These will not be | |
2473 | * required to pass though another grace period: They are done. | |
a50c3af9 | 2474 | */ |
e5601400 | 2475 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2476 | *rsp->orphan_donetail = rdp->nxtlist; |
2477 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 2478 | } |
e74f4c45 | 2479 | |
b33078b6 PM |
2480 | /* |
2481 | * Finally, initialize the rcu_data structure's list to empty and | |
2482 | * disallow further callbacks on this CPU. | |
2483 | */ | |
3f5d3ea6 | 2484 | init_callback_list(rdp); |
b33078b6 | 2485 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
b1420f1c PM |
2486 | } |
2487 | ||
2488 | /* | |
2489 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 2490 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c | 2491 | */ |
96d3fd0d | 2492 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) |
b1420f1c PM |
2493 | { |
2494 | int i; | |
fa07a58f | 2495 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
b1420f1c | 2496 | |
3fbfbf7a | 2497 | /* No-CBs CPUs are handled specially. */ |
ea46351c PM |
2498 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || |
2499 | rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) | |
3fbfbf7a PM |
2500 | return; |
2501 | ||
b1420f1c PM |
2502 | /* Do the accounting first. */ |
2503 | rdp->qlen_lazy += rsp->qlen_lazy; | |
2504 | rdp->qlen += rsp->qlen; | |
2505 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
2506 | if (rsp->qlen_lazy != rsp->qlen) |
2507 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2508 | rsp->qlen_lazy = 0; |
2509 | rsp->qlen = 0; | |
2510 | ||
2511 | /* | |
2512 | * We do not need a memory barrier here because the only way we | |
2513 | * can get here if there is an rcu_barrier() in flight is if | |
2514 | * we are the task doing the rcu_barrier(). | |
2515 | */ | |
2516 | ||
2517 | /* First adopt the ready-to-invoke callbacks. */ | |
2518 | if (rsp->orphan_donelist != NULL) { | |
2519 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2520 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
2521 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
2522 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2523 | rdp->nxttail[i] = rsp->orphan_donetail; | |
2524 | rsp->orphan_donelist = NULL; | |
2525 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2526 | } | |
2527 | ||
2528 | /* And then adopt the callbacks that still need a grace period. */ | |
2529 | if (rsp->orphan_nxtlist != NULL) { | |
2530 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
2531 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
2532 | rsp->orphan_nxtlist = NULL; | |
2533 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2534 | } | |
2535 | } | |
2536 | ||
2537 | /* | |
2538 | * Trace the fact that this CPU is going offline. | |
2539 | */ | |
2540 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
2541 | { | |
2542 | RCU_TRACE(unsigned long mask); | |
2543 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
2544 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
2545 | ||
ea46351c PM |
2546 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2547 | return; | |
2548 | ||
b1420f1c | 2549 | RCU_TRACE(mask = rdp->grpmask); |
e5601400 PM |
2550 | trace_rcu_grace_period(rsp->name, |
2551 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 2552 | TPS("cpuofl")); |
64db4cff PM |
2553 | } |
2554 | ||
8af3a5e7 PM |
2555 | /* |
2556 | * All CPUs for the specified rcu_node structure have gone offline, | |
2557 | * and all tasks that were preempted within an RCU read-side critical | |
2558 | * section while running on one of those CPUs have since exited their RCU | |
2559 | * read-side critical section. Some other CPU is reporting this fact with | |
2560 | * the specified rcu_node structure's ->lock held and interrupts disabled. | |
2561 | * This function therefore goes up the tree of rcu_node structures, | |
2562 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that | |
2563 | * the leaf rcu_node structure's ->qsmaskinit field has already been | |
2564 | * updated | |
2565 | * | |
2566 | * This function does check that the specified rcu_node structure has | |
2567 | * all CPUs offline and no blocked tasks, so it is OK to invoke it | |
2568 | * prematurely. That said, invoking it after the fact will cost you | |
2569 | * a needless lock acquisition. So once it has done its work, don't | |
2570 | * invoke it again. | |
2571 | */ | |
2572 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) | |
2573 | { | |
2574 | long mask; | |
2575 | struct rcu_node *rnp = rnp_leaf; | |
2576 | ||
ea46351c PM |
2577 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || |
2578 | rnp->qsmaskinit || rcu_preempt_has_tasks(rnp)) | |
8af3a5e7 PM |
2579 | return; |
2580 | for (;;) { | |
2581 | mask = rnp->grpmask; | |
2582 | rnp = rnp->parent; | |
2583 | if (!rnp) | |
2584 | break; | |
2585 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
2586 | smp_mb__after_unlock_lock(); /* GP memory ordering. */ | |
2587 | rnp->qsmaskinit &= ~mask; | |
0aa04b05 | 2588 | rnp->qsmask &= ~mask; |
8af3a5e7 PM |
2589 | if (rnp->qsmaskinit) { |
2590 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2591 | return; | |
2592 | } | |
2593 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2594 | } | |
2595 | } | |
2596 | ||
88428cc5 PM |
2597 | /* |
2598 | * The CPU is exiting the idle loop into the arch_cpu_idle_dead() | |
2599 | * function. We now remove it from the rcu_node tree's ->qsmaskinit | |
2600 | * bit masks. | |
2601 | */ | |
2602 | static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp) | |
2603 | { | |
2604 | unsigned long flags; | |
2605 | unsigned long mask; | |
2606 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
2607 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ | |
2608 | ||
ea46351c PM |
2609 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2610 | return; | |
2611 | ||
88428cc5 PM |
2612 | /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ |
2613 | mask = rdp->grpmask; | |
2614 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
2615 | smp_mb__after_unlock_lock(); /* Enforce GP memory-order guarantee. */ | |
2616 | rnp->qsmaskinitnext &= ~mask; | |
2617 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2618 | } | |
2619 | ||
64db4cff | 2620 | /* |
e5601400 | 2621 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
2622 | * this fact from process context. Do the remainder of the cleanup, |
2623 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
2624 | * adopting them. There can only be one CPU hotplug operation at a time, |
2625 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 2626 | */ |
e5601400 | 2627 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 2628 | { |
2036d94a | 2629 | unsigned long flags; |
e5601400 | 2630 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 2631 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2632 | |
ea46351c PM |
2633 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2634 | return; | |
2635 | ||
2036d94a | 2636 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2637 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 2638 | |
b1420f1c | 2639 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
78043c46 | 2640 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
b1420f1c | 2641 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); |
96d3fd0d | 2642 | rcu_adopt_orphan_cbs(rsp, flags); |
a8f4cbad | 2643 | raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags); |
b1420f1c | 2644 | |
cf01537e PM |
2645 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
2646 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2647 | cpu, rdp->qlen, rdp->nxtlist); | |
64db4cff PM |
2648 | } |
2649 | ||
64db4cff PM |
2650 | /* |
2651 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2652 | * period. Thottle as specified by rdp->blimit. | |
2653 | */ | |
37c72e56 | 2654 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2655 | { |
2656 | unsigned long flags; | |
2657 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
2658 | long bl, count, count_lazy; |
2659 | int i; | |
64db4cff | 2660 | |
dc35c893 | 2661 | /* If no callbacks are ready, just return. */ |
29c00b4a | 2662 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 2663 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
7d0ae808 | 2664 | trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist), |
4968c300 PM |
2665 | need_resched(), is_idle_task(current), |
2666 | rcu_is_callbacks_kthread()); | |
64db4cff | 2667 | return; |
29c00b4a | 2668 | } |
64db4cff PM |
2669 | |
2670 | /* | |
2671 | * Extract the list of ready callbacks, disabling to prevent | |
2672 | * races with call_rcu() from interrupt handlers. | |
2673 | */ | |
2674 | local_irq_save(flags); | |
8146c4e2 | 2675 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2676 | bl = rdp->blimit; |
486e2593 | 2677 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2678 | list = rdp->nxtlist; |
2679 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2680 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2681 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2682 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2683 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2684 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2685 | local_irq_restore(flags); |
2686 | ||
2687 | /* Invoke callbacks. */ | |
486e2593 | 2688 | count = count_lazy = 0; |
64db4cff PM |
2689 | while (list) { |
2690 | next = list->next; | |
2691 | prefetch(next); | |
551d55a9 | 2692 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2693 | if (__rcu_reclaim(rsp->name, list)) |
2694 | count_lazy++; | |
64db4cff | 2695 | list = next; |
dff1672d PM |
2696 | /* Stop only if limit reached and CPU has something to do. */ |
2697 | if (++count >= bl && | |
2698 | (need_resched() || | |
2699 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2700 | break; |
2701 | } | |
2702 | ||
2703 | local_irq_save(flags); | |
4968c300 PM |
2704 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2705 | is_idle_task(current), | |
2706 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2707 | |
2708 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2709 | if (list != NULL) { |
2710 | *tail = rdp->nxtlist; | |
2711 | rdp->nxtlist = list; | |
b41772ab PM |
2712 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2713 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2714 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2715 | else |
2716 | break; | |
2717 | } | |
b1420f1c PM |
2718 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2719 | rdp->qlen_lazy -= count_lazy; | |
7d0ae808 | 2720 | WRITE_ONCE(rdp->qlen, rdp->qlen - count); |
b1420f1c | 2721 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2722 | |
2723 | /* Reinstate batch limit if we have worked down the excess. */ | |
2724 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2725 | rdp->blimit = blimit; | |
2726 | ||
37c72e56 PM |
2727 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2728 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2729 | rdp->qlen_last_fqs_check = 0; | |
2730 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2731 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2732 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2733 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2734 | |
64db4cff PM |
2735 | local_irq_restore(flags); |
2736 | ||
e0f23060 | 2737 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2738 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2739 | invoke_rcu_core(); |
64db4cff PM |
2740 | } |
2741 | ||
2742 | /* | |
2743 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2744 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2745 | * Also schedule RCU core processing. |
64db4cff | 2746 | * |
9b2e4f18 | 2747 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2748 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2749 | * false, there is no point in invoking rcu_check_callbacks(). | |
2750 | */ | |
c3377c2d | 2751 | void rcu_check_callbacks(int user) |
64db4cff | 2752 | { |
f7f7bac9 | 2753 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2754 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2755 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2756 | |
2757 | /* | |
2758 | * Get here if this CPU took its interrupt from user | |
2759 | * mode or from the idle loop, and if this is not a | |
2760 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2761 | * a quiescent state, so note it. |
64db4cff PM |
2762 | * |
2763 | * No memory barrier is required here because both | |
d6714c22 PM |
2764 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2765 | * variables that other CPUs neither access nor modify, | |
2766 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2767 | */ |
2768 | ||
284a8c93 PM |
2769 | rcu_sched_qs(); |
2770 | rcu_bh_qs(); | |
64db4cff PM |
2771 | |
2772 | } else if (!in_softirq()) { | |
2773 | ||
2774 | /* | |
2775 | * Get here if this CPU did not take its interrupt from | |
2776 | * softirq, in other words, if it is not interrupting | |
2777 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2778 | * critical section, so note it. |
64db4cff PM |
2779 | */ |
2780 | ||
284a8c93 | 2781 | rcu_bh_qs(); |
64db4cff | 2782 | } |
86aea0e6 | 2783 | rcu_preempt_check_callbacks(); |
e3950ecd | 2784 | if (rcu_pending()) |
a46e0899 | 2785 | invoke_rcu_core(); |
8315f422 PM |
2786 | if (user) |
2787 | rcu_note_voluntary_context_switch(current); | |
f7f7bac9 | 2788 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2789 | } |
2790 | ||
64db4cff PM |
2791 | /* |
2792 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2793 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2794 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2795 | * | |
ee47eb9f | 2796 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2797 | */ |
217af2a2 PM |
2798 | static void force_qs_rnp(struct rcu_state *rsp, |
2799 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2800 | unsigned long *maxj), | |
2801 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2802 | { |
2803 | unsigned long bit; | |
2804 | int cpu; | |
2805 | unsigned long flags; | |
2806 | unsigned long mask; | |
a0b6c9a7 | 2807 | struct rcu_node *rnp; |
64db4cff | 2808 | |
a0b6c9a7 | 2809 | rcu_for_each_leaf_node(rsp, rnp) { |
bde6c3aa | 2810 | cond_resched_rcu_qs(); |
64db4cff | 2811 | mask = 0; |
1304afb2 | 2812 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 2813 | smp_mb__after_unlock_lock(); |
a0b6c9a7 | 2814 | if (rnp->qsmask == 0) { |
a77da14c PM |
2815 | if (rcu_state_p == &rcu_sched_state || |
2816 | rsp != rcu_state_p || | |
2817 | rcu_preempt_blocked_readers_cgp(rnp)) { | |
2818 | /* | |
2819 | * No point in scanning bits because they | |
2820 | * are all zero. But we might need to | |
2821 | * priority-boost blocked readers. | |
2822 | */ | |
2823 | rcu_initiate_boost(rnp, flags); | |
2824 | /* rcu_initiate_boost() releases rnp->lock */ | |
2825 | continue; | |
2826 | } | |
2827 | if (rnp->parent && | |
2828 | (rnp->parent->qsmask & rnp->grpmask)) { | |
2829 | /* | |
2830 | * Race between grace-period | |
2831 | * initialization and task exiting RCU | |
2832 | * read-side critical section: Report. | |
2833 | */ | |
2834 | rcu_report_unblock_qs_rnp(rsp, rnp, flags); | |
2835 | /* rcu_report_unblock_qs_rnp() rlses ->lock */ | |
2836 | continue; | |
2837 | } | |
64db4cff | 2838 | } |
a0b6c9a7 | 2839 | cpu = rnp->grplo; |
64db4cff | 2840 | bit = 1; |
a0b6c9a7 | 2841 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 | 2842 | if ((rnp->qsmask & bit) != 0) { |
0edd1b17 PM |
2843 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) |
2844 | mask |= bit; | |
2845 | } | |
64db4cff | 2846 | } |
45f014c5 | 2847 | if (mask != 0) { |
654e9533 PM |
2848 | /* Idle/offline CPUs, report (releases rnp->lock. */ |
2849 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); | |
0aa04b05 PM |
2850 | } else { |
2851 | /* Nothing to do here, so just drop the lock. */ | |
2852 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff | 2853 | } |
64db4cff | 2854 | } |
64db4cff PM |
2855 | } |
2856 | ||
2857 | /* | |
2858 | * Force quiescent states on reluctant CPUs, and also detect which | |
2859 | * CPUs are in dyntick-idle mode. | |
2860 | */ | |
4cdfc175 | 2861 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2862 | { |
2863 | unsigned long flags; | |
394f2769 PM |
2864 | bool ret; |
2865 | struct rcu_node *rnp; | |
2866 | struct rcu_node *rnp_old = NULL; | |
2867 | ||
2868 | /* Funnel through hierarchy to reduce memory contention. */ | |
d860d403 | 2869 | rnp = __this_cpu_read(rsp->rda->mynode); |
394f2769 | 2870 | for (; rnp != NULL; rnp = rnp->parent) { |
7d0ae808 | 2871 | ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || |
394f2769 PM |
2872 | !raw_spin_trylock(&rnp->fqslock); |
2873 | if (rnp_old != NULL) | |
2874 | raw_spin_unlock(&rnp_old->fqslock); | |
2875 | if (ret) { | |
a792563b | 2876 | rsp->n_force_qs_lh++; |
394f2769 PM |
2877 | return; |
2878 | } | |
2879 | rnp_old = rnp; | |
2880 | } | |
2881 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2882 | |
394f2769 PM |
2883 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2884 | raw_spin_lock_irqsave(&rnp_old->lock, flags); | |
6303b9c8 | 2885 | smp_mb__after_unlock_lock(); |
394f2769 | 2886 | raw_spin_unlock(&rnp_old->fqslock); |
7d0ae808 | 2887 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { |
a792563b | 2888 | rsp->n_force_qs_lh++; |
394f2769 | 2889 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2890 | return; /* Someone beat us to it. */ |
46a1e34e | 2891 | } |
7d0ae808 | 2892 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); |
394f2769 | 2893 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
2aa792e6 | 2894 | rcu_gp_kthread_wake(rsp); |
64db4cff PM |
2895 | } |
2896 | ||
64db4cff | 2897 | /* |
e0f23060 PM |
2898 | * This does the RCU core processing work for the specified rcu_state |
2899 | * and rcu_data structures. This may be called only from the CPU to | |
2900 | * whom the rdp belongs. | |
64db4cff PM |
2901 | */ |
2902 | static void | |
1bca8cf1 | 2903 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2904 | { |
2905 | unsigned long flags; | |
48a7639c | 2906 | bool needwake; |
fa07a58f | 2907 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
64db4cff | 2908 | |
2e597558 PM |
2909 | WARN_ON_ONCE(rdp->beenonline == 0); |
2910 | ||
64db4cff PM |
2911 | /* Update RCU state based on any recent quiescent states. */ |
2912 | rcu_check_quiescent_state(rsp, rdp); | |
2913 | ||
2914 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2915 | local_irq_save(flags); |
64db4cff | 2916 | if (cpu_needs_another_gp(rsp, rdp)) { |
dc35c893 | 2917 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
48a7639c | 2918 | needwake = rcu_start_gp(rsp); |
b8462084 | 2919 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
48a7639c PM |
2920 | if (needwake) |
2921 | rcu_gp_kthread_wake(rsp); | |
dc35c893 PM |
2922 | } else { |
2923 | local_irq_restore(flags); | |
64db4cff PM |
2924 | } |
2925 | ||
2926 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2927 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2928 | invoke_rcu_callbacks(rsp, rdp); |
96d3fd0d PM |
2929 | |
2930 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2931 | do_nocb_deferred_wakeup(rdp); | |
09223371 SL |
2932 | } |
2933 | ||
64db4cff | 2934 | /* |
e0f23060 | 2935 | * Do RCU core processing for the current CPU. |
64db4cff | 2936 | */ |
09223371 | 2937 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2938 | { |
6ce75a23 PM |
2939 | struct rcu_state *rsp; |
2940 | ||
bfa00b4c PM |
2941 | if (cpu_is_offline(smp_processor_id())) |
2942 | return; | |
f7f7bac9 | 2943 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2944 | for_each_rcu_flavor(rsp) |
2945 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2946 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2947 | } |
2948 | ||
a26ac245 | 2949 | /* |
e0f23060 PM |
2950 | * Schedule RCU callback invocation. If the specified type of RCU |
2951 | * does not support RCU priority boosting, just do a direct call, | |
2952 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
924df8a0 | 2953 | * are running on the current CPU with softirqs disabled, the |
e0f23060 | 2954 | * rcu_cpu_kthread_task cannot disappear out from under us. |
a26ac245 | 2955 | */ |
a46e0899 | 2956 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2957 | { |
7d0ae808 | 2958 | if (unlikely(!READ_ONCE(rcu_scheduler_fully_active))) |
b0d30417 | 2959 | return; |
a46e0899 PM |
2960 | if (likely(!rsp->boost)) { |
2961 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2962 | return; |
2963 | } | |
a46e0899 | 2964 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2965 | } |
2966 | ||
a46e0899 | 2967 | static void invoke_rcu_core(void) |
09223371 | 2968 | { |
b0f74036 PM |
2969 | if (cpu_online(smp_processor_id())) |
2970 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2971 | } |
2972 | ||
29154c57 PM |
2973 | /* |
2974 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2975 | */ | |
2976 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2977 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2978 | { |
48a7639c PM |
2979 | bool needwake; |
2980 | ||
62fde6ed PM |
2981 | /* |
2982 | * If called from an extended quiescent state, invoke the RCU | |
2983 | * core in order to force a re-evaluation of RCU's idleness. | |
2984 | */ | |
9910affa | 2985 | if (!rcu_is_watching()) |
62fde6ed PM |
2986 | invoke_rcu_core(); |
2987 | ||
a16b7a69 | 2988 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2989 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2990 | return; |
64db4cff | 2991 | |
37c72e56 PM |
2992 | /* |
2993 | * Force the grace period if too many callbacks or too long waiting. | |
2994 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2995 | * if some other CPU has recently done so. Also, don't bother | |
2996 | * invoking force_quiescent_state() if the newly enqueued callback | |
2997 | * is the only one waiting for a grace period to complete. | |
2998 | */ | |
2655d57e | 2999 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
3000 | |
3001 | /* Are we ignoring a completed grace period? */ | |
470716fc | 3002 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
3003 | |
3004 | /* Start a new grace period if one not already started. */ | |
3005 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
3006 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
3007 | ||
b8462084 | 3008 | raw_spin_lock(&rnp_root->lock); |
6303b9c8 | 3009 | smp_mb__after_unlock_lock(); |
48a7639c | 3010 | needwake = rcu_start_gp(rsp); |
b8462084 | 3011 | raw_spin_unlock(&rnp_root->lock); |
48a7639c PM |
3012 | if (needwake) |
3013 | rcu_gp_kthread_wake(rsp); | |
b52573d2 PM |
3014 | } else { |
3015 | /* Give the grace period a kick. */ | |
3016 | rdp->blimit = LONG_MAX; | |
3017 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
3018 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 3019 | force_quiescent_state(rsp); |
b52573d2 PM |
3020 | rdp->n_force_qs_snap = rsp->n_force_qs; |
3021 | rdp->qlen_last_fqs_check = rdp->qlen; | |
3022 | } | |
4cdfc175 | 3023 | } |
29154c57 PM |
3024 | } |
3025 | ||
ae150184 PM |
3026 | /* |
3027 | * RCU callback function to leak a callback. | |
3028 | */ | |
3029 | static void rcu_leak_callback(struct rcu_head *rhp) | |
3030 | { | |
3031 | } | |
3032 | ||
3fbfbf7a PM |
3033 | /* |
3034 | * Helper function for call_rcu() and friends. The cpu argument will | |
3035 | * normally be -1, indicating "currently running CPU". It may specify | |
3036 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
3037 | * is expected to specify a CPU. | |
3038 | */ | |
64db4cff | 3039 | static void |
b6a4ae76 | 3040 | __call_rcu(struct rcu_head *head, rcu_callback_t func, |
3fbfbf7a | 3041 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
3042 | { |
3043 | unsigned long flags; | |
3044 | struct rcu_data *rdp; | |
3045 | ||
1146edcb | 3046 | WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */ |
ae150184 PM |
3047 | if (debug_rcu_head_queue(head)) { |
3048 | /* Probable double call_rcu(), so leak the callback. */ | |
7d0ae808 | 3049 | WRITE_ONCE(head->func, rcu_leak_callback); |
ae150184 PM |
3050 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); |
3051 | return; | |
3052 | } | |
64db4cff PM |
3053 | head->func = func; |
3054 | head->next = NULL; | |
3055 | ||
64db4cff PM |
3056 | /* |
3057 | * Opportunistically note grace-period endings and beginnings. | |
3058 | * Note that we might see a beginning right after we see an | |
3059 | * end, but never vice versa, since this CPU has to pass through | |
3060 | * a quiescent state betweentimes. | |
3061 | */ | |
3062 | local_irq_save(flags); | |
394f99a9 | 3063 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
3064 | |
3065 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
3066 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
3067 | int offline; | |
3068 | ||
3069 | if (cpu != -1) | |
3070 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
143da9c2 PM |
3071 | if (likely(rdp->mynode)) { |
3072 | /* Post-boot, so this should be for a no-CBs CPU. */ | |
3073 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); | |
3074 | WARN_ON_ONCE(offline); | |
3075 | /* Offline CPU, _call_rcu() illegal, leak callback. */ | |
3076 | local_irq_restore(flags); | |
3077 | return; | |
3078 | } | |
3079 | /* | |
3080 | * Very early boot, before rcu_init(). Initialize if needed | |
3081 | * and then drop through to queue the callback. | |
3082 | */ | |
3083 | BUG_ON(cpu != -1); | |
34404ca8 | 3084 | WARN_ON_ONCE(!rcu_is_watching()); |
143da9c2 PM |
3085 | if (!likely(rdp->nxtlist)) |
3086 | init_default_callback_list(rdp); | |
0d8ee37e | 3087 | } |
7d0ae808 | 3088 | WRITE_ONCE(rdp->qlen, rdp->qlen + 1); |
486e2593 PM |
3089 | if (lazy) |
3090 | rdp->qlen_lazy++; | |
c57afe80 PM |
3091 | else |
3092 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
3093 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
3094 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
3095 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 3096 | |
d4c08f2a PM |
3097 | if (__is_kfree_rcu_offset((unsigned long)func)) |
3098 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 3099 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 3100 | else |
486e2593 | 3101 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 3102 | |
29154c57 PM |
3103 | /* Go handle any RCU core processing required. */ |
3104 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
3105 | local_irq_restore(flags); |
3106 | } | |
3107 | ||
3108 | /* | |
d6714c22 | 3109 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 3110 | */ |
b6a4ae76 | 3111 | void call_rcu_sched(struct rcu_head *head, rcu_callback_t func) |
64db4cff | 3112 | { |
3fbfbf7a | 3113 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 3114 | } |
d6714c22 | 3115 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
3116 | |
3117 | /* | |
486e2593 | 3118 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff | 3119 | */ |
b6a4ae76 | 3120 | void call_rcu_bh(struct rcu_head *head, rcu_callback_t func) |
64db4cff | 3121 | { |
3fbfbf7a | 3122 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
3123 | } |
3124 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
3125 | ||
495aa969 ACB |
3126 | /* |
3127 | * Queue an RCU callback for lazy invocation after a grace period. | |
3128 | * This will likely be later named something like "call_rcu_lazy()", | |
3129 | * but this change will require some way of tagging the lazy RCU | |
3130 | * callbacks in the list of pending callbacks. Until then, this | |
3131 | * function may only be called from __kfree_rcu(). | |
3132 | */ | |
3133 | void kfree_call_rcu(struct rcu_head *head, | |
b6a4ae76 | 3134 | rcu_callback_t func) |
495aa969 | 3135 | { |
e534165b | 3136 | __call_rcu(head, func, rcu_state_p, -1, 1); |
495aa969 ACB |
3137 | } |
3138 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
3139 | ||
6d813391 PM |
3140 | /* |
3141 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
3142 | * any blocking grace-period wait automatically implies a grace period | |
3143 | * if there is only one CPU online at any point time during execution | |
3144 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
3145 | * occasionally incorrectly indicate that there are multiple CPUs online | |
3146 | * when there was in fact only one the whole time, as this just adds | |
3147 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
3148 | */ |
3149 | static inline int rcu_blocking_is_gp(void) | |
3150 | { | |
95f0c1de PM |
3151 | int ret; |
3152 | ||
6d813391 | 3153 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
3154 | preempt_disable(); |
3155 | ret = num_online_cpus() <= 1; | |
3156 | preempt_enable(); | |
3157 | return ret; | |
6d813391 PM |
3158 | } |
3159 | ||
6ebb237b PM |
3160 | /** |
3161 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
3162 | * | |
3163 | * Control will return to the caller some time after a full rcu-sched | |
3164 | * grace period has elapsed, in other words after all currently executing | |
3165 | * rcu-sched read-side critical sections have completed. These read-side | |
3166 | * critical sections are delimited by rcu_read_lock_sched() and | |
3167 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
3168 | * local_irq_disable(), and so on may be used in place of | |
3169 | * rcu_read_lock_sched(). | |
3170 | * | |
3171 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
3172 | * non-threaded hardware-interrupt handlers, in progress on entry will |
3173 | * have completed before this primitive returns. However, this does not | |
3174 | * guarantee that softirq handlers will have completed, since in some | |
3175 | * kernels, these handlers can run in process context, and can block. | |
3176 | * | |
3177 | * Note that this guarantee implies further memory-ordering guarantees. | |
3178 | * On systems with more than one CPU, when synchronize_sched() returns, | |
3179 | * each CPU is guaranteed to have executed a full memory barrier since the | |
3180 | * end of its last RCU-sched read-side critical section whose beginning | |
3181 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
3182 | * an RCU read-side critical section that extends beyond the return from | |
3183 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
3184 | * after the beginning of synchronize_sched() and before the beginning of | |
3185 | * that RCU read-side critical section. Note that these guarantees include | |
3186 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
3187 | * that are executing in the kernel. | |
3188 | * | |
3189 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
3190 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
3191 | * to have executed a full memory barrier during the execution of | |
3192 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
3193 | * again only if the system has more than one CPU). | |
6ebb237b PM |
3194 | * |
3195 | * This primitive provides the guarantees made by the (now removed) | |
3196 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
3197 | * guarantees that rcu_read_lock() sections will have completed. | |
3198 | * In "classic RCU", these two guarantees happen to be one and | |
3199 | * the same, but can differ in realtime RCU implementations. | |
3200 | */ | |
3201 | void synchronize_sched(void) | |
3202 | { | |
f78f5b90 PM |
3203 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
3204 | lock_is_held(&rcu_lock_map) || | |
3205 | lock_is_held(&rcu_sched_lock_map), | |
3206 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
3207 | if (rcu_blocking_is_gp()) |
3208 | return; | |
5afff48b | 3209 | if (rcu_gp_is_expedited()) |
3705b88d AM |
3210 | synchronize_sched_expedited(); |
3211 | else | |
3212 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
3213 | } |
3214 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
3215 | ||
3216 | /** | |
3217 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
3218 | * | |
3219 | * Control will return to the caller some time after a full rcu_bh grace | |
3220 | * period has elapsed, in other words after all currently executing rcu_bh | |
3221 | * read-side critical sections have completed. RCU read-side critical | |
3222 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
3223 | * and may be nested. | |
f0a0e6f2 PM |
3224 | * |
3225 | * See the description of synchronize_sched() for more detailed information | |
3226 | * on memory ordering guarantees. | |
6ebb237b PM |
3227 | */ |
3228 | void synchronize_rcu_bh(void) | |
3229 | { | |
f78f5b90 PM |
3230 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
3231 | lock_is_held(&rcu_lock_map) || | |
3232 | lock_is_held(&rcu_sched_lock_map), | |
3233 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
3234 | if (rcu_blocking_is_gp()) |
3235 | return; | |
5afff48b | 3236 | if (rcu_gp_is_expedited()) |
3705b88d AM |
3237 | synchronize_rcu_bh_expedited(); |
3238 | else | |
3239 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
3240 | } |
3241 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
3242 | ||
765a3f4f PM |
3243 | /** |
3244 | * get_state_synchronize_rcu - Snapshot current RCU state | |
3245 | * | |
3246 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
3247 | * to determine whether or not a full grace period has elapsed in the | |
3248 | * meantime. | |
3249 | */ | |
3250 | unsigned long get_state_synchronize_rcu(void) | |
3251 | { | |
3252 | /* | |
3253 | * Any prior manipulation of RCU-protected data must happen | |
3254 | * before the load from ->gpnum. | |
3255 | */ | |
3256 | smp_mb(); /* ^^^ */ | |
3257 | ||
3258 | /* | |
3259 | * Make sure this load happens before the purportedly | |
3260 | * time-consuming work between get_state_synchronize_rcu() | |
3261 | * and cond_synchronize_rcu(). | |
3262 | */ | |
e534165b | 3263 | return smp_load_acquire(&rcu_state_p->gpnum); |
765a3f4f PM |
3264 | } |
3265 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
3266 | ||
3267 | /** | |
3268 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
3269 | * | |
3270 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
3271 | * | |
3272 | * If a full RCU grace period has elapsed since the earlier call to | |
3273 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
3274 | * synchronize_rcu() to wait for a full grace period. | |
3275 | * | |
3276 | * Yes, this function does not take counter wrap into account. But | |
3277 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3278 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3279 | * so waiting for one additional grace period should be just fine. | |
3280 | */ | |
3281 | void cond_synchronize_rcu(unsigned long oldstate) | |
3282 | { | |
3283 | unsigned long newstate; | |
3284 | ||
3285 | /* | |
3286 | * Ensure that this load happens before any RCU-destructive | |
3287 | * actions the caller might carry out after we return. | |
3288 | */ | |
e534165b | 3289 | newstate = smp_load_acquire(&rcu_state_p->completed); |
765a3f4f PM |
3290 | if (ULONG_CMP_GE(oldstate, newstate)) |
3291 | synchronize_rcu(); | |
3292 | } | |
3293 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
3294 | ||
24560056 PM |
3295 | /** |
3296 | * get_state_synchronize_sched - Snapshot current RCU-sched state | |
3297 | * | |
3298 | * Returns a cookie that is used by a later call to cond_synchronize_sched() | |
3299 | * to determine whether or not a full grace period has elapsed in the | |
3300 | * meantime. | |
3301 | */ | |
3302 | unsigned long get_state_synchronize_sched(void) | |
3303 | { | |
3304 | /* | |
3305 | * Any prior manipulation of RCU-protected data must happen | |
3306 | * before the load from ->gpnum. | |
3307 | */ | |
3308 | smp_mb(); /* ^^^ */ | |
3309 | ||
3310 | /* | |
3311 | * Make sure this load happens before the purportedly | |
3312 | * time-consuming work between get_state_synchronize_sched() | |
3313 | * and cond_synchronize_sched(). | |
3314 | */ | |
3315 | return smp_load_acquire(&rcu_sched_state.gpnum); | |
3316 | } | |
3317 | EXPORT_SYMBOL_GPL(get_state_synchronize_sched); | |
3318 | ||
3319 | /** | |
3320 | * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period | |
3321 | * | |
3322 | * @oldstate: return value from earlier call to get_state_synchronize_sched() | |
3323 | * | |
3324 | * If a full RCU-sched grace period has elapsed since the earlier call to | |
3325 | * get_state_synchronize_sched(), just return. Otherwise, invoke | |
3326 | * synchronize_sched() to wait for a full grace period. | |
3327 | * | |
3328 | * Yes, this function does not take counter wrap into account. But | |
3329 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3330 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3331 | * so waiting for one additional grace period should be just fine. | |
3332 | */ | |
3333 | void cond_synchronize_sched(unsigned long oldstate) | |
3334 | { | |
3335 | unsigned long newstate; | |
3336 | ||
3337 | /* | |
3338 | * Ensure that this load happens before any RCU-destructive | |
3339 | * actions the caller might carry out after we return. | |
3340 | */ | |
3341 | newstate = smp_load_acquire(&rcu_sched_state.completed); | |
3342 | if (ULONG_CMP_GE(oldstate, newstate)) | |
3343 | synchronize_sched(); | |
3344 | } | |
3345 | EXPORT_SYMBOL_GPL(cond_synchronize_sched); | |
3346 | ||
28f00767 PM |
3347 | /* Adjust sequence number for start of update-side operation. */ |
3348 | static void rcu_seq_start(unsigned long *sp) | |
3349 | { | |
3350 | WRITE_ONCE(*sp, *sp + 1); | |
3351 | smp_mb(); /* Ensure update-side operation after counter increment. */ | |
3352 | WARN_ON_ONCE(!(*sp & 0x1)); | |
3353 | } | |
3354 | ||
3355 | /* Adjust sequence number for end of update-side operation. */ | |
3356 | static void rcu_seq_end(unsigned long *sp) | |
3357 | { | |
3358 | smp_mb(); /* Ensure update-side operation before counter increment. */ | |
3359 | WRITE_ONCE(*sp, *sp + 1); | |
3360 | WARN_ON_ONCE(*sp & 0x1); | |
3361 | } | |
3362 | ||
3363 | /* Take a snapshot of the update side's sequence number. */ | |
3364 | static unsigned long rcu_seq_snap(unsigned long *sp) | |
3365 | { | |
3366 | unsigned long s; | |
3367 | ||
3368 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
3369 | s = (READ_ONCE(*sp) + 3) & ~0x1; | |
3370 | smp_mb(); /* Above access must not bleed into critical section. */ | |
3371 | return s; | |
3372 | } | |
3373 | ||
3374 | /* | |
3375 | * Given a snapshot from rcu_seq_snap(), determine whether or not a | |
3376 | * full update-side operation has occurred. | |
3377 | */ | |
3378 | static bool rcu_seq_done(unsigned long *sp, unsigned long s) | |
3379 | { | |
3380 | return ULONG_CMP_GE(READ_ONCE(*sp), s); | |
3381 | } | |
3382 | ||
3383 | /* Wrapper functions for expedited grace periods. */ | |
3384 | static void rcu_exp_gp_seq_start(struct rcu_state *rsp) | |
3385 | { | |
3386 | rcu_seq_start(&rsp->expedited_sequence); | |
3387 | } | |
3388 | static void rcu_exp_gp_seq_end(struct rcu_state *rsp) | |
3389 | { | |
3390 | rcu_seq_end(&rsp->expedited_sequence); | |
704dd435 | 3391 | smp_mb(); /* Ensure that consecutive grace periods serialize. */ |
28f00767 PM |
3392 | } |
3393 | static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp) | |
3394 | { | |
3395 | return rcu_seq_snap(&rsp->expedited_sequence); | |
3396 | } | |
3397 | static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s) | |
3398 | { | |
3399 | return rcu_seq_done(&rsp->expedited_sequence, s); | |
3400 | } | |
3401 | ||
b9585e94 PM |
3402 | /* |
3403 | * Reset the ->expmaskinit values in the rcu_node tree to reflect any | |
3404 | * recent CPU-online activity. Note that these masks are not cleared | |
3405 | * when CPUs go offline, so they reflect the union of all CPUs that have | |
3406 | * ever been online. This means that this function normally takes its | |
3407 | * no-work-to-do fastpath. | |
3408 | */ | |
3409 | static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp) | |
3410 | { | |
3411 | bool done; | |
3412 | unsigned long flags; | |
3413 | unsigned long mask; | |
3414 | unsigned long oldmask; | |
3415 | int ncpus = READ_ONCE(rsp->ncpus); | |
3416 | struct rcu_node *rnp; | |
3417 | struct rcu_node *rnp_up; | |
3418 | ||
3419 | /* If no new CPUs onlined since last time, nothing to do. */ | |
3420 | if (likely(ncpus == rsp->ncpus_snap)) | |
3421 | return; | |
3422 | rsp->ncpus_snap = ncpus; | |
3423 | ||
3424 | /* | |
3425 | * Each pass through the following loop propagates newly onlined | |
3426 | * CPUs for the current rcu_node structure up the rcu_node tree. | |
3427 | */ | |
3428 | rcu_for_each_leaf_node(rsp, rnp) { | |
3429 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3430 | smp_mb__after_unlock_lock(); | |
3431 | if (rnp->expmaskinit == rnp->expmaskinitnext) { | |
3432 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3433 | continue; /* No new CPUs, nothing to do. */ | |
3434 | } | |
3435 | ||
3436 | /* Update this node's mask, track old value for propagation. */ | |
3437 | oldmask = rnp->expmaskinit; | |
3438 | rnp->expmaskinit = rnp->expmaskinitnext; | |
3439 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3440 | ||
3441 | /* If was already nonzero, nothing to propagate. */ | |
3442 | if (oldmask) | |
3443 | continue; | |
3444 | ||
3445 | /* Propagate the new CPU up the tree. */ | |
3446 | mask = rnp->grpmask; | |
3447 | rnp_up = rnp->parent; | |
3448 | done = false; | |
3449 | while (rnp_up) { | |
3450 | raw_spin_lock_irqsave(&rnp_up->lock, flags); | |
3451 | smp_mb__after_unlock_lock(); | |
3452 | if (rnp_up->expmaskinit) | |
3453 | done = true; | |
3454 | rnp_up->expmaskinit |= mask; | |
3455 | raw_spin_unlock_irqrestore(&rnp_up->lock, flags); | |
3456 | if (done) | |
3457 | break; | |
3458 | mask = rnp_up->grpmask; | |
3459 | rnp_up = rnp_up->parent; | |
3460 | } | |
3461 | } | |
3462 | } | |
3463 | ||
3464 | /* | |
3465 | * Reset the ->expmask values in the rcu_node tree in preparation for | |
3466 | * a new expedited grace period. | |
3467 | */ | |
3468 | static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp) | |
3469 | { | |
3470 | unsigned long flags; | |
3471 | struct rcu_node *rnp; | |
3472 | ||
3473 | sync_exp_reset_tree_hotplug(rsp); | |
3474 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
3475 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3476 | smp_mb__after_unlock_lock(); | |
3477 | WARN_ON_ONCE(rnp->expmask); | |
3478 | rnp->expmask = rnp->expmaskinit; | |
3479 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3480 | } | |
3481 | } | |
3482 | ||
7922cd0e | 3483 | /* |
8203d6d0 | 3484 | * Return non-zero if there is no RCU expedited grace period in progress |
7922cd0e PM |
3485 | * for the specified rcu_node structure, in other words, if all CPUs and |
3486 | * tasks covered by the specified rcu_node structure have done their bit | |
3487 | * for the current expedited grace period. Works only for preemptible | |
3488 | * RCU -- other RCU implementation use other means. | |
3489 | * | |
3490 | * Caller must hold the root rcu_node's exp_funnel_mutex. | |
3491 | */ | |
3492 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) | |
3493 | { | |
8203d6d0 | 3494 | return rnp->exp_tasks == NULL && |
7922cd0e PM |
3495 | READ_ONCE(rnp->expmask) == 0; |
3496 | } | |
3497 | ||
3498 | /* | |
3499 | * Report the exit from RCU read-side critical section for the last task | |
3500 | * that queued itself during or before the current expedited preemptible-RCU | |
3501 | * grace period. This event is reported either to the rcu_node structure on | |
3502 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
3503 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
3504 | * iteratively!) | |
3505 | * | |
8203d6d0 PM |
3506 | * Caller must hold the root rcu_node's exp_funnel_mutex and the |
3507 | * specified rcu_node structure's ->lock. | |
7922cd0e | 3508 | */ |
8203d6d0 PM |
3509 | static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
3510 | bool wake, unsigned long flags) | |
3511 | __releases(rnp->lock) | |
7922cd0e | 3512 | { |
7922cd0e PM |
3513 | unsigned long mask; |
3514 | ||
7922cd0e PM |
3515 | for (;;) { |
3516 | if (!sync_rcu_preempt_exp_done(rnp)) { | |
8203d6d0 PM |
3517 | if (!rnp->expmask) |
3518 | rcu_initiate_boost(rnp, flags); | |
3519 | else | |
3520 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
7922cd0e PM |
3521 | break; |
3522 | } | |
3523 | if (rnp->parent == NULL) { | |
3524 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3525 | if (wake) { | |
3526 | smp_mb(); /* EGP done before wake_up(). */ | |
3527 | wake_up(&rsp->expedited_wq); | |
3528 | } | |
3529 | break; | |
3530 | } | |
3531 | mask = rnp->grpmask; | |
3532 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ | |
3533 | rnp = rnp->parent; | |
3534 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ | |
3535 | smp_mb__after_unlock_lock(); | |
8203d6d0 | 3536 | WARN_ON_ONCE(!(rnp->expmask & mask)); |
7922cd0e PM |
3537 | rnp->expmask &= ~mask; |
3538 | } | |
3539 | } | |
3540 | ||
8203d6d0 PM |
3541 | /* |
3542 | * Report expedited quiescent state for specified node. This is a | |
3543 | * lock-acquisition wrapper function for __rcu_report_exp_rnp(). | |
3544 | * | |
3545 | * Caller must hold the root rcu_node's exp_funnel_mutex. | |
3546 | */ | |
3547 | static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp, | |
3548 | struct rcu_node *rnp, bool wake) | |
3549 | { | |
3550 | unsigned long flags; | |
3551 | ||
3552 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3553 | smp_mb__after_unlock_lock(); | |
3554 | __rcu_report_exp_rnp(rsp, rnp, wake, flags); | |
3555 | } | |
3556 | ||
3557 | /* | |
3558 | * Report expedited quiescent state for multiple CPUs, all covered by the | |
3559 | * specified leaf rcu_node structure. Caller must hold the root | |
3560 | * rcu_node's exp_funnel_mutex. | |
3561 | */ | |
3562 | static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp, | |
3563 | unsigned long mask, bool wake) | |
3564 | { | |
3565 | unsigned long flags; | |
3566 | ||
3567 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3568 | smp_mb__after_unlock_lock(); | |
338b0f76 PM |
3569 | if (!(rnp->expmask & mask)) { |
3570 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3571 | return; | |
3572 | } | |
8203d6d0 PM |
3573 | rnp->expmask &= ~mask; |
3574 | __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */ | |
3575 | } | |
3576 | ||
3577 | /* | |
3578 | * Report expedited quiescent state for specified rcu_data (CPU). | |
3579 | * Caller must hold the root rcu_node's exp_funnel_mutex. | |
3580 | */ | |
6587a23b PM |
3581 | static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp, |
3582 | bool wake) | |
8203d6d0 PM |
3583 | { |
3584 | rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake); | |
3585 | } | |
3586 | ||
29fd9309 PM |
3587 | /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */ |
3588 | static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp, | |
2cd6ffaf | 3589 | struct rcu_data *rdp, |
29fd9309 | 3590 | atomic_long_t *stat, unsigned long s) |
3d3b7db0 | 3591 | { |
28f00767 | 3592 | if (rcu_exp_gp_seq_done(rsp, s)) { |
385b73c0 PM |
3593 | if (rnp) |
3594 | mutex_unlock(&rnp->exp_funnel_mutex); | |
2cd6ffaf PM |
3595 | else if (rdp) |
3596 | mutex_unlock(&rdp->exp_funnel_mutex); | |
385b73c0 PM |
3597 | /* Ensure test happens before caller kfree(). */ |
3598 | smp_mb__before_atomic(); /* ^^^ */ | |
3599 | atomic_long_inc(stat); | |
385b73c0 PM |
3600 | return true; |
3601 | } | |
3602 | return false; | |
3603 | } | |
3604 | ||
b09e5f86 PM |
3605 | /* |
3606 | * Funnel-lock acquisition for expedited grace periods. Returns a | |
3607 | * pointer to the root rcu_node structure, or NULL if some other | |
3608 | * task did the expedited grace period for us. | |
3609 | */ | |
3610 | static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s) | |
3611 | { | |
2cd6ffaf | 3612 | struct rcu_data *rdp; |
b09e5f86 PM |
3613 | struct rcu_node *rnp0; |
3614 | struct rcu_node *rnp1 = NULL; | |
3615 | ||
3d3b7db0 | 3616 | /* |
cdacbe1f PM |
3617 | * First try directly acquiring the root lock in order to reduce |
3618 | * latency in the common case where expedited grace periods are | |
3619 | * rare. We check mutex_is_locked() to avoid pathological levels of | |
3620 | * memory contention on ->exp_funnel_mutex in the heavy-load case. | |
3d3b7db0 | 3621 | */ |
cdacbe1f PM |
3622 | rnp0 = rcu_get_root(rsp); |
3623 | if (!mutex_is_locked(&rnp0->exp_funnel_mutex)) { | |
3624 | if (mutex_trylock(&rnp0->exp_funnel_mutex)) { | |
3625 | if (sync_exp_work_done(rsp, rnp0, NULL, | |
3626 | &rsp->expedited_workdone0, s)) | |
3627 | return NULL; | |
3628 | return rnp0; | |
3629 | } | |
3630 | } | |
3631 | ||
b09e5f86 PM |
3632 | /* |
3633 | * Each pass through the following loop works its way | |
3634 | * up the rcu_node tree, returning if others have done the | |
3635 | * work or otherwise falls through holding the root rnp's | |
3636 | * ->exp_funnel_mutex. The mapping from CPU to rcu_node structure | |
3637 | * can be inexact, as it is just promoting locality and is not | |
3638 | * strictly needed for correctness. | |
3639 | */ | |
2cd6ffaf PM |
3640 | rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id()); |
3641 | if (sync_exp_work_done(rsp, NULL, NULL, &rsp->expedited_workdone1, s)) | |
3642 | return NULL; | |
3643 | mutex_lock(&rdp->exp_funnel_mutex); | |
3644 | rnp0 = rdp->mynode; | |
b09e5f86 | 3645 | for (; rnp0 != NULL; rnp0 = rnp0->parent) { |
2cd6ffaf PM |
3646 | if (sync_exp_work_done(rsp, rnp1, rdp, |
3647 | &rsp->expedited_workdone2, s)) | |
b09e5f86 PM |
3648 | return NULL; |
3649 | mutex_lock(&rnp0->exp_funnel_mutex); | |
3650 | if (rnp1) | |
3651 | mutex_unlock(&rnp1->exp_funnel_mutex); | |
2cd6ffaf PM |
3652 | else |
3653 | mutex_unlock(&rdp->exp_funnel_mutex); | |
b09e5f86 PM |
3654 | rnp1 = rnp0; |
3655 | } | |
2cd6ffaf PM |
3656 | if (sync_exp_work_done(rsp, rnp1, rdp, |
3657 | &rsp->expedited_workdone3, s)) | |
b09e5f86 PM |
3658 | return NULL; |
3659 | return rnp1; | |
3660 | } | |
3661 | ||
cf3620a6 | 3662 | /* Invoked on each online non-idle CPU for expedited quiescent state. */ |
338b0f76 | 3663 | static void sync_sched_exp_handler(void *data) |
b09e5f86 | 3664 | { |
338b0f76 PM |
3665 | struct rcu_data *rdp; |
3666 | struct rcu_node *rnp; | |
3667 | struct rcu_state *rsp = data; | |
b09e5f86 | 3668 | |
338b0f76 PM |
3669 | rdp = this_cpu_ptr(rsp->rda); |
3670 | rnp = rdp->mynode; | |
3671 | if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) || | |
3672 | __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) | |
3673 | return; | |
6587a23b PM |
3674 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true); |
3675 | resched_cpu(smp_processor_id()); | |
3d3b7db0 PM |
3676 | } |
3677 | ||
338b0f76 PM |
3678 | /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */ |
3679 | static void sync_sched_exp_online_cleanup(int cpu) | |
3680 | { | |
3681 | struct rcu_data *rdp; | |
3682 | int ret; | |
3683 | struct rcu_node *rnp; | |
3684 | struct rcu_state *rsp = &rcu_sched_state; | |
3685 | ||
3686 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
3687 | rnp = rdp->mynode; | |
3688 | if (!(READ_ONCE(rnp->expmask) & rdp->grpmask)) | |
3689 | return; | |
3690 | ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0); | |
3691 | WARN_ON_ONCE(ret); | |
3692 | } | |
3693 | ||
bce5fa12 PM |
3694 | /* |
3695 | * Select the nodes that the upcoming expedited grace period needs | |
3696 | * to wait for. | |
3697 | */ | |
dcdb8807 PM |
3698 | static void sync_rcu_exp_select_cpus(struct rcu_state *rsp, |
3699 | smp_call_func_t func) | |
bce5fa12 PM |
3700 | { |
3701 | int cpu; | |
3702 | unsigned long flags; | |
3703 | unsigned long mask; | |
3704 | unsigned long mask_ofl_test; | |
3705 | unsigned long mask_ofl_ipi; | |
6587a23b | 3706 | int ret; |
bce5fa12 PM |
3707 | struct rcu_node *rnp; |
3708 | ||
3709 | sync_exp_reset_tree(rsp); | |
3710 | rcu_for_each_leaf_node(rsp, rnp) { | |
3711 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3712 | smp_mb__after_unlock_lock(); | |
3713 | ||
3714 | /* Each pass checks a CPU for identity, offline, and idle. */ | |
3715 | mask_ofl_test = 0; | |
3716 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) { | |
3717 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
3718 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); | |
3719 | ||
3720 | if (raw_smp_processor_id() == cpu || | |
bce5fa12 PM |
3721 | !(atomic_add_return(0, &rdtp->dynticks) & 0x1)) |
3722 | mask_ofl_test |= rdp->grpmask; | |
3723 | } | |
3724 | mask_ofl_ipi = rnp->expmask & ~mask_ofl_test; | |
3725 | ||
3726 | /* | |
3727 | * Need to wait for any blocked tasks as well. Note that | |
3728 | * additional blocking tasks will also block the expedited | |
3729 | * GP until such time as the ->expmask bits are cleared. | |
3730 | */ | |
3731 | if (rcu_preempt_has_tasks(rnp)) | |
3732 | rnp->exp_tasks = rnp->blkd_tasks.next; | |
3733 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3734 | ||
3735 | /* IPI the remaining CPUs for expedited quiescent state. */ | |
3736 | mask = 1; | |
3737 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) { | |
3738 | if (!(mask_ofl_ipi & mask)) | |
3739 | continue; | |
338b0f76 | 3740 | retry_ipi: |
dcdb8807 | 3741 | ret = smp_call_function_single(cpu, func, rsp, 0); |
338b0f76 | 3742 | if (!ret) { |
6587a23b | 3743 | mask_ofl_ipi &= ~mask; |
338b0f76 PM |
3744 | } else { |
3745 | /* Failed, raced with offline. */ | |
3746 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
3747 | if (cpu_online(cpu) && | |
3748 | (rnp->expmask & mask)) { | |
3749 | raw_spin_unlock_irqrestore(&rnp->lock, | |
3750 | flags); | |
3751 | schedule_timeout_uninterruptible(1); | |
3752 | if (cpu_online(cpu) && | |
3753 | (rnp->expmask & mask)) | |
3754 | goto retry_ipi; | |
3755 | raw_spin_lock_irqsave(&rnp->lock, | |
3756 | flags); | |
3757 | } | |
3758 | if (!(rnp->expmask & mask)) | |
3759 | mask_ofl_ipi &= ~mask; | |
3760 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3761 | } | |
bce5fa12 PM |
3762 | } |
3763 | /* Report quiescent states for those that went offline. */ | |
3764 | mask_ofl_test |= mask_ofl_ipi; | |
3765 | if (mask_ofl_test) | |
3766 | rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false); | |
3767 | } | |
3d3b7db0 PM |
3768 | } |
3769 | ||
cf3620a6 PM |
3770 | static void synchronize_sched_expedited_wait(struct rcu_state *rsp) |
3771 | { | |
3772 | int cpu; | |
3773 | unsigned long jiffies_stall; | |
3774 | unsigned long jiffies_start; | |
bce5fa12 PM |
3775 | unsigned long mask; |
3776 | struct rcu_node *rnp; | |
3777 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
cf3620a6 PM |
3778 | int ret; |
3779 | ||
3780 | jiffies_stall = rcu_jiffies_till_stall_check(); | |
3781 | jiffies_start = jiffies; | |
3782 | ||
3783 | for (;;) { | |
3784 | ret = wait_event_interruptible_timeout( | |
3785 | rsp->expedited_wq, | |
bce5fa12 | 3786 | sync_rcu_preempt_exp_done(rnp_root), |
cf3620a6 PM |
3787 | jiffies_stall); |
3788 | if (ret > 0) | |
3789 | return; | |
3790 | if (ret < 0) { | |
3791 | /* Hit a signal, disable CPU stall warnings. */ | |
3792 | wait_event(rsp->expedited_wq, | |
bce5fa12 | 3793 | sync_rcu_preempt_exp_done(rnp_root)); |
cf3620a6 PM |
3794 | return; |
3795 | } | |
c5865638 | 3796 | pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {", |
cf3620a6 | 3797 | rsp->name); |
bce5fa12 | 3798 | rcu_for_each_leaf_node(rsp, rnp) { |
74611ecb | 3799 | (void)rcu_print_task_exp_stall(rnp); |
bce5fa12 PM |
3800 | mask = 1; |
3801 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) { | |
74611ecb PM |
3802 | struct rcu_data *rdp; |
3803 | ||
bce5fa12 PM |
3804 | if (!(rnp->expmask & mask)) |
3805 | continue; | |
74611ecb PM |
3806 | rdp = per_cpu_ptr(rsp->rda, cpu); |
3807 | pr_cont(" %d-%c%c%c", cpu, | |
3808 | "O."[cpu_online(cpu)], | |
3809 | "o."[!!(rdp->grpmask & rnp->expmaskinit)], | |
3810 | "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]); | |
bce5fa12 PM |
3811 | } |
3812 | mask <<= 1; | |
cf3620a6 PM |
3813 | } |
3814 | pr_cont(" } %lu jiffies s: %lu\n", | |
3815 | jiffies - jiffies_start, rsp->expedited_sequence); | |
bce5fa12 PM |
3816 | rcu_for_each_leaf_node(rsp, rnp) { |
3817 | mask = 1; | |
3818 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) { | |
3819 | if (!(rnp->expmask & mask)) | |
3820 | continue; | |
3821 | dump_cpu_task(cpu); | |
3822 | } | |
cf3620a6 PM |
3823 | } |
3824 | jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3; | |
3825 | } | |
3826 | } | |
3827 | ||
236fefaf PM |
3828 | /** |
3829 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
3830 | * | |
3831 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
3832 | * approach to force the grace period to end quickly. This consumes | |
3833 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
3834 | * so is thus not recommended for any sort of common-case code. In fact, | |
3835 | * if you are using synchronize_sched_expedited() in a loop, please | |
3836 | * restructure your code to batch your updates, and then use a single | |
3837 | * synchronize_sched() instead. | |
3d3b7db0 | 3838 | * |
d6ada2cf PM |
3839 | * This implementation can be thought of as an application of sequence |
3840 | * locking to expedited grace periods, but using the sequence counter to | |
3841 | * determine when someone else has already done the work instead of for | |
385b73c0 | 3842 | * retrying readers. |
3d3b7db0 PM |
3843 | */ |
3844 | void synchronize_sched_expedited(void) | |
3845 | { | |
7fd0ddc5 | 3846 | unsigned long s; |
b09e5f86 | 3847 | struct rcu_node *rnp; |
40694d66 | 3848 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 3849 | |
d6ada2cf | 3850 | /* Take a snapshot of the sequence number. */ |
28f00767 | 3851 | s = rcu_exp_gp_seq_snap(rsp); |
3d3b7db0 | 3852 | |
b09e5f86 | 3853 | rnp = exp_funnel_lock(rsp, s); |
807226e2 | 3854 | if (rnp == NULL) |
b09e5f86 | 3855 | return; /* Someone else did our work for us. */ |
e0775cef | 3856 | |
28f00767 | 3857 | rcu_exp_gp_seq_start(rsp); |
338b0f76 | 3858 | sync_rcu_exp_select_cpus(rsp, sync_sched_exp_handler); |
bce5fa12 | 3859 | synchronize_sched_expedited_wait(rsp); |
e0775cef | 3860 | |
28f00767 | 3861 | rcu_exp_gp_seq_end(rsp); |
b09e5f86 | 3862 | mutex_unlock(&rnp->exp_funnel_mutex); |
3d3b7db0 PM |
3863 | } |
3864 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
3865 | ||
64db4cff PM |
3866 | /* |
3867 | * Check to see if there is any immediate RCU-related work to be done | |
3868 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
3869 | * The checks are in order of increasing expense: checks that can be | |
3870 | * carried out against CPU-local state are performed first. However, | |
3871 | * we must check for CPU stalls first, else we might not get a chance. | |
3872 | */ | |
3873 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
3874 | { | |
2f51f988 PM |
3875 | struct rcu_node *rnp = rdp->mynode; |
3876 | ||
64db4cff PM |
3877 | rdp->n_rcu_pending++; |
3878 | ||
3879 | /* Check for CPU stalls, if enabled. */ | |
3880 | check_cpu_stall(rsp, rdp); | |
3881 | ||
a096932f PM |
3882 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ |
3883 | if (rcu_nohz_full_cpu(rsp)) | |
3884 | return 0; | |
3885 | ||
64db4cff | 3886 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
5c51dd73 | 3887 | if (rcu_scheduler_fully_active && |
5b74c458 | 3888 | rdp->core_needs_qs && rdp->cpu_no_qs.b.norm && |
5cd37193 | 3889 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) { |
97c668b8 PM |
3890 | rdp->n_rp_core_needs_qs++; |
3891 | } else if (rdp->core_needs_qs && | |
5b74c458 | 3892 | (!rdp->cpu_no_qs.b.norm || |
5cd37193 | 3893 | rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) { |
d21670ac | 3894 | rdp->n_rp_report_qs++; |
64db4cff | 3895 | return 1; |
7ba5c840 | 3896 | } |
64db4cff PM |
3897 | |
3898 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
3899 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
3900 | rdp->n_rp_cb_ready++; | |
64db4cff | 3901 | return 1; |
7ba5c840 | 3902 | } |
64db4cff PM |
3903 | |
3904 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
3905 | if (cpu_needs_another_gp(rsp, rdp)) { |
3906 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 3907 | return 1; |
7ba5c840 | 3908 | } |
64db4cff PM |
3909 | |
3910 | /* Has another RCU grace period completed? */ | |
7d0ae808 | 3911 | if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 3912 | rdp->n_rp_gp_completed++; |
64db4cff | 3913 | return 1; |
7ba5c840 | 3914 | } |
64db4cff PM |
3915 | |
3916 | /* Has a new RCU grace period started? */ | |
7d0ae808 PM |
3917 | if (READ_ONCE(rnp->gpnum) != rdp->gpnum || |
3918 | unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */ | |
7ba5c840 | 3919 | rdp->n_rp_gp_started++; |
64db4cff | 3920 | return 1; |
7ba5c840 | 3921 | } |
64db4cff | 3922 | |
96d3fd0d PM |
3923 | /* Does this CPU need a deferred NOCB wakeup? */ |
3924 | if (rcu_nocb_need_deferred_wakeup(rdp)) { | |
3925 | rdp->n_rp_nocb_defer_wakeup++; | |
3926 | return 1; | |
3927 | } | |
3928 | ||
64db4cff | 3929 | /* nothing to do */ |
7ba5c840 | 3930 | rdp->n_rp_need_nothing++; |
64db4cff PM |
3931 | return 0; |
3932 | } | |
3933 | ||
3934 | /* | |
3935 | * Check to see if there is any immediate RCU-related work to be done | |
3936 | * by the current CPU, returning 1 if so. This function is part of the | |
3937 | * RCU implementation; it is -not- an exported member of the RCU API. | |
3938 | */ | |
e3950ecd | 3939 | static int rcu_pending(void) |
64db4cff | 3940 | { |
6ce75a23 PM |
3941 | struct rcu_state *rsp; |
3942 | ||
3943 | for_each_rcu_flavor(rsp) | |
e3950ecd | 3944 | if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda))) |
6ce75a23 PM |
3945 | return 1; |
3946 | return 0; | |
64db4cff PM |
3947 | } |
3948 | ||
3949 | /* | |
c0f4dfd4 PM |
3950 | * Return true if the specified CPU has any callback. If all_lazy is |
3951 | * non-NULL, store an indication of whether all callbacks are lazy. | |
3952 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 3953 | */ |
82072c4f | 3954 | static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy) |
64db4cff | 3955 | { |
c0f4dfd4 PM |
3956 | bool al = true; |
3957 | bool hc = false; | |
3958 | struct rcu_data *rdp; | |
6ce75a23 PM |
3959 | struct rcu_state *rsp; |
3960 | ||
c0f4dfd4 | 3961 | for_each_rcu_flavor(rsp) { |
aa6da514 | 3962 | rdp = this_cpu_ptr(rsp->rda); |
69c8d28c PM |
3963 | if (!rdp->nxtlist) |
3964 | continue; | |
3965 | hc = true; | |
3966 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
c0f4dfd4 | 3967 | al = false; |
69c8d28c PM |
3968 | break; |
3969 | } | |
c0f4dfd4 PM |
3970 | } |
3971 | if (all_lazy) | |
3972 | *all_lazy = al; | |
3973 | return hc; | |
64db4cff PM |
3974 | } |
3975 | ||
a83eff0a PM |
3976 | /* |
3977 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
3978 | * the compiler is expected to optimize this away. | |
3979 | */ | |
e66c33d5 | 3980 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
3981 | int cpu, unsigned long done) |
3982 | { | |
3983 | trace_rcu_barrier(rsp->name, s, cpu, | |
3984 | atomic_read(&rsp->barrier_cpu_count), done); | |
3985 | } | |
3986 | ||
b1420f1c PM |
3987 | /* |
3988 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
3989 | * up the task executing _rcu_barrier(). | |
3990 | */ | |
24ebbca8 | 3991 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3992 | { |
24ebbca8 PM |
3993 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
3994 | struct rcu_state *rsp = rdp->rsp; | |
3995 | ||
a83eff0a | 3996 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
4f525a52 | 3997 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence); |
7db74df8 | 3998 | complete(&rsp->barrier_completion); |
a83eff0a | 3999 | } else { |
4f525a52 | 4000 | _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence); |
a83eff0a | 4001 | } |
d0ec774c PM |
4002 | } |
4003 | ||
4004 | /* | |
4005 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
4006 | */ | |
4007 | static void rcu_barrier_func(void *type) | |
4008 | { | |
037b64ed | 4009 | struct rcu_state *rsp = type; |
fa07a58f | 4010 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
d0ec774c | 4011 | |
4f525a52 | 4012 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence); |
24ebbca8 | 4013 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 4014 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
4015 | } |
4016 | ||
d0ec774c PM |
4017 | /* |
4018 | * Orchestrate the specified type of RCU barrier, waiting for all | |
4019 | * RCU callbacks of the specified type to complete. | |
4020 | */ | |
037b64ed | 4021 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 4022 | { |
b1420f1c | 4023 | int cpu; |
b1420f1c | 4024 | struct rcu_data *rdp; |
4f525a52 | 4025 | unsigned long s = rcu_seq_snap(&rsp->barrier_sequence); |
b1420f1c | 4026 | |
4f525a52 | 4027 | _rcu_barrier_trace(rsp, "Begin", -1, s); |
b1420f1c | 4028 | |
e74f4c45 | 4029 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 4030 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 4031 | |
4f525a52 PM |
4032 | /* Did someone else do our work for us? */ |
4033 | if (rcu_seq_done(&rsp->barrier_sequence, s)) { | |
4034 | _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence); | |
cf3a9c48 PM |
4035 | smp_mb(); /* caller's subsequent code after above check. */ |
4036 | mutex_unlock(&rsp->barrier_mutex); | |
4037 | return; | |
4038 | } | |
4039 | ||
4f525a52 PM |
4040 | /* Mark the start of the barrier operation. */ |
4041 | rcu_seq_start(&rsp->barrier_sequence); | |
4042 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence); | |
b1420f1c | 4043 | |
d0ec774c | 4044 | /* |
b1420f1c PM |
4045 | * Initialize the count to one rather than to zero in order to |
4046 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
4047 | * (or preemption of this task). Exclude CPU-hotplug operations |
4048 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 4049 | */ |
7db74df8 | 4050 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 4051 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 4052 | get_online_cpus(); |
b1420f1c PM |
4053 | |
4054 | /* | |
1331e7a1 PM |
4055 | * Force each CPU with callbacks to register a new callback. |
4056 | * When that callback is invoked, we will know that all of the | |
4057 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 4058 | */ |
3fbfbf7a | 4059 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 4060 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 4061 | continue; |
b1420f1c | 4062 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 4063 | if (rcu_is_nocb_cpu(cpu)) { |
d7e29933 PM |
4064 | if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) { |
4065 | _rcu_barrier_trace(rsp, "OfflineNoCB", cpu, | |
4f525a52 | 4066 | rsp->barrier_sequence); |
d7e29933 PM |
4067 | } else { |
4068 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, | |
4f525a52 | 4069 | rsp->barrier_sequence); |
41050a00 | 4070 | smp_mb__before_atomic(); |
d7e29933 PM |
4071 | atomic_inc(&rsp->barrier_cpu_count); |
4072 | __call_rcu(&rdp->barrier_head, | |
4073 | rcu_barrier_callback, rsp, cpu, 0); | |
4074 | } | |
7d0ae808 | 4075 | } else if (READ_ONCE(rdp->qlen)) { |
a83eff0a | 4076 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
4f525a52 | 4077 | rsp->barrier_sequence); |
037b64ed | 4078 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 4079 | } else { |
a83eff0a | 4080 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
4f525a52 | 4081 | rsp->barrier_sequence); |
b1420f1c PM |
4082 | } |
4083 | } | |
1331e7a1 | 4084 | put_online_cpus(); |
b1420f1c PM |
4085 | |
4086 | /* | |
4087 | * Now that we have an rcu_barrier_callback() callback on each | |
4088 | * CPU, and thus each counted, remove the initial count. | |
4089 | */ | |
24ebbca8 | 4090 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 4091 | complete(&rsp->barrier_completion); |
b1420f1c PM |
4092 | |
4093 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ | |
7db74df8 | 4094 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c | 4095 | |
4f525a52 PM |
4096 | /* Mark the end of the barrier operation. */ |
4097 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence); | |
4098 | rcu_seq_end(&rsp->barrier_sequence); | |
4099 | ||
b1420f1c | 4100 | /* Other rcu_barrier() invocations can now safely proceed. */ |
7be7f0be | 4101 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 4102 | } |
d0ec774c PM |
4103 | |
4104 | /** | |
4105 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
4106 | */ | |
4107 | void rcu_barrier_bh(void) | |
4108 | { | |
037b64ed | 4109 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
4110 | } |
4111 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
4112 | ||
4113 | /** | |
4114 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
4115 | */ | |
4116 | void rcu_barrier_sched(void) | |
4117 | { | |
037b64ed | 4118 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
4119 | } |
4120 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
4121 | ||
0aa04b05 PM |
4122 | /* |
4123 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the | |
4124 | * first CPU in a given leaf rcu_node structure coming online. The caller | |
4125 | * must hold the corresponding leaf rcu_node ->lock with interrrupts | |
4126 | * disabled. | |
4127 | */ | |
4128 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) | |
4129 | { | |
4130 | long mask; | |
4131 | struct rcu_node *rnp = rnp_leaf; | |
4132 | ||
4133 | for (;;) { | |
4134 | mask = rnp->grpmask; | |
4135 | rnp = rnp->parent; | |
4136 | if (rnp == NULL) | |
4137 | return; | |
4138 | raw_spin_lock(&rnp->lock); /* Interrupts already disabled. */ | |
4139 | rnp->qsmaskinit |= mask; | |
4140 | raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */ | |
4141 | } | |
4142 | } | |
4143 | ||
64db4cff | 4144 | /* |
27569620 | 4145 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 4146 | */ |
27569620 PM |
4147 | static void __init |
4148 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
4149 | { |
4150 | unsigned long flags; | |
394f99a9 | 4151 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
4152 | struct rcu_node *rnp = rcu_get_root(rsp); |
4153 | ||
4154 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 4155 | raw_spin_lock_irqsave(&rnp->lock, flags); |
27569620 | 4156 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
27569620 | 4157 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 4158 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 4159 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 4160 | rdp->cpu = cpu; |
d4c08f2a | 4161 | rdp->rsp = rsp; |
2cd6ffaf | 4162 | mutex_init(&rdp->exp_funnel_mutex); |
3fbfbf7a | 4163 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 4164 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
4165 | } |
4166 | ||
4167 | /* | |
4168 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
4169 | * offline event can be happening at a given time. Note also that we | |
4170 | * can accept some slop in the rsp->completed access due to the fact | |
4171 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 4172 | */ |
49fb4c62 | 4173 | static void |
9b67122a | 4174 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp) |
64db4cff PM |
4175 | { |
4176 | unsigned long flags; | |
64db4cff | 4177 | unsigned long mask; |
394f99a9 | 4178 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
4179 | struct rcu_node *rnp = rcu_get_root(rsp); |
4180 | ||
4181 | /* Set up local state, ensuring consistent view of global state. */ | |
1304afb2 | 4182 | raw_spin_lock_irqsave(&rnp->lock, flags); |
37c72e56 PM |
4183 | rdp->qlen_last_fqs_check = 0; |
4184 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 4185 | rdp->blimit = blimit; |
39c8d313 PM |
4186 | if (!rdp->nxtlist) |
4187 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ | |
29e37d81 | 4188 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 4189 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
4190 | atomic_set(&rdp->dynticks->dynticks, |
4191 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 4192 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 4193 | |
0aa04b05 PM |
4194 | /* |
4195 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed | |
4196 | * propagation up the rcu_node tree will happen at the beginning | |
4197 | * of the next grace period. | |
4198 | */ | |
64db4cff PM |
4199 | rnp = rdp->mynode; |
4200 | mask = rdp->grpmask; | |
0aa04b05 PM |
4201 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
4202 | smp_mb__after_unlock_lock(); | |
4203 | rnp->qsmaskinitnext |= mask; | |
b9585e94 PM |
4204 | rnp->expmaskinitnext |= mask; |
4205 | if (!rdp->beenonline) | |
4206 | WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1); | |
4207 | rdp->beenonline = true; /* We have now been online. */ | |
0aa04b05 PM |
4208 | rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */ |
4209 | rdp->completed = rnp->completed; | |
5b74c458 | 4210 | rdp->cpu_no_qs.b.norm = true; |
a738eec6 | 4211 | rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu); |
97c668b8 | 4212 | rdp->core_needs_qs = false; |
0aa04b05 PM |
4213 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); |
4214 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff PM |
4215 | } |
4216 | ||
49fb4c62 | 4217 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 4218 | { |
6ce75a23 PM |
4219 | struct rcu_state *rsp; |
4220 | ||
4221 | for_each_rcu_flavor(rsp) | |
9b67122a | 4222 | rcu_init_percpu_data(cpu, rsp); |
64db4cff PM |
4223 | } |
4224 | ||
4225 | /* | |
f41d911f | 4226 | * Handle CPU online/offline notification events. |
64db4cff | 4227 | */ |
88428cc5 PM |
4228 | int rcu_cpu_notify(struct notifier_block *self, |
4229 | unsigned long action, void *hcpu) | |
64db4cff PM |
4230 | { |
4231 | long cpu = (long)hcpu; | |
e534165b | 4232 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
a26ac245 | 4233 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 4234 | struct rcu_state *rsp; |
64db4cff PM |
4235 | |
4236 | switch (action) { | |
4237 | case CPU_UP_PREPARE: | |
4238 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
4239 | rcu_prepare_cpu(cpu); |
4240 | rcu_prepare_kthreads(cpu); | |
35ce7f29 | 4241 | rcu_spawn_all_nocb_kthreads(cpu); |
a26ac245 PM |
4242 | break; |
4243 | case CPU_ONLINE: | |
0f962a5e | 4244 | case CPU_DOWN_FAILED: |
338b0f76 | 4245 | sync_sched_exp_online_cleanup(cpu); |
5d01bbd1 | 4246 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
4247 | break; |
4248 | case CPU_DOWN_PREPARE: | |
34ed6246 | 4249 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 4250 | break; |
d0ec774c PM |
4251 | case CPU_DYING: |
4252 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
4253 | for_each_rcu_flavor(rsp) |
4254 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 4255 | break; |
88428cc5 | 4256 | case CPU_DYING_IDLE: |
6587a23b | 4257 | /* QS for any half-done expedited RCU-sched GP. */ |
338b0f76 PM |
4258 | preempt_disable(); |
4259 | rcu_report_exp_rdp(&rcu_sched_state, | |
4260 | this_cpu_ptr(rcu_sched_state.rda), true); | |
4261 | preempt_enable(); | |
6587a23b | 4262 | |
88428cc5 PM |
4263 | for_each_rcu_flavor(rsp) { |
4264 | rcu_cleanup_dying_idle_cpu(cpu, rsp); | |
4265 | } | |
4266 | break; | |
64db4cff PM |
4267 | case CPU_DEAD: |
4268 | case CPU_DEAD_FROZEN: | |
4269 | case CPU_UP_CANCELED: | |
4270 | case CPU_UP_CANCELED_FROZEN: | |
776d6807 | 4271 | for_each_rcu_flavor(rsp) { |
6ce75a23 | 4272 | rcu_cleanup_dead_cpu(cpu, rsp); |
776d6807 PM |
4273 | do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu)); |
4274 | } | |
64db4cff PM |
4275 | break; |
4276 | default: | |
4277 | break; | |
4278 | } | |
34ed6246 | 4279 | return NOTIFY_OK; |
64db4cff PM |
4280 | } |
4281 | ||
d1d74d14 BP |
4282 | static int rcu_pm_notify(struct notifier_block *self, |
4283 | unsigned long action, void *hcpu) | |
4284 | { | |
4285 | switch (action) { | |
4286 | case PM_HIBERNATION_PREPARE: | |
4287 | case PM_SUSPEND_PREPARE: | |
4288 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
5afff48b | 4289 | rcu_expedite_gp(); |
d1d74d14 BP |
4290 | break; |
4291 | case PM_POST_HIBERNATION: | |
4292 | case PM_POST_SUSPEND: | |
5afff48b PM |
4293 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ |
4294 | rcu_unexpedite_gp(); | |
d1d74d14 BP |
4295 | break; |
4296 | default: | |
4297 | break; | |
4298 | } | |
4299 | return NOTIFY_OK; | |
4300 | } | |
4301 | ||
b3dbec76 | 4302 | /* |
9386c0b7 | 4303 | * Spawn the kthreads that handle each RCU flavor's grace periods. |
b3dbec76 PM |
4304 | */ |
4305 | static int __init rcu_spawn_gp_kthread(void) | |
4306 | { | |
4307 | unsigned long flags; | |
a94844b2 | 4308 | int kthread_prio_in = kthread_prio; |
b3dbec76 PM |
4309 | struct rcu_node *rnp; |
4310 | struct rcu_state *rsp; | |
a94844b2 | 4311 | struct sched_param sp; |
b3dbec76 PM |
4312 | struct task_struct *t; |
4313 | ||
a94844b2 PM |
4314 | /* Force priority into range. */ |
4315 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) | |
4316 | kthread_prio = 1; | |
4317 | else if (kthread_prio < 0) | |
4318 | kthread_prio = 0; | |
4319 | else if (kthread_prio > 99) | |
4320 | kthread_prio = 99; | |
4321 | if (kthread_prio != kthread_prio_in) | |
4322 | pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n", | |
4323 | kthread_prio, kthread_prio_in); | |
4324 | ||
9386c0b7 | 4325 | rcu_scheduler_fully_active = 1; |
b3dbec76 | 4326 | for_each_rcu_flavor(rsp) { |
a94844b2 | 4327 | t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
4328 | BUG_ON(IS_ERR(t)); |
4329 | rnp = rcu_get_root(rsp); | |
4330 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
4331 | rsp->gp_kthread = t; | |
a94844b2 PM |
4332 | if (kthread_prio) { |
4333 | sp.sched_priority = kthread_prio; | |
4334 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
4335 | } | |
4336 | wake_up_process(t); | |
b3dbec76 PM |
4337 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
4338 | } | |
35ce7f29 | 4339 | rcu_spawn_nocb_kthreads(); |
9386c0b7 | 4340 | rcu_spawn_boost_kthreads(); |
b3dbec76 PM |
4341 | return 0; |
4342 | } | |
4343 | early_initcall(rcu_spawn_gp_kthread); | |
4344 | ||
bbad9379 PM |
4345 | /* |
4346 | * This function is invoked towards the end of the scheduler's initialization | |
4347 | * process. Before this is called, the idle task might contain | |
4348 | * RCU read-side critical sections (during which time, this idle | |
4349 | * task is booting the system). After this function is called, the | |
4350 | * idle tasks are prohibited from containing RCU read-side critical | |
4351 | * sections. This function also enables RCU lockdep checking. | |
4352 | */ | |
4353 | void rcu_scheduler_starting(void) | |
4354 | { | |
4355 | WARN_ON(num_online_cpus() != 1); | |
4356 | WARN_ON(nr_context_switches() > 0); | |
4357 | rcu_scheduler_active = 1; | |
4358 | } | |
4359 | ||
64db4cff PM |
4360 | /* |
4361 | * Compute the per-level fanout, either using the exact fanout specified | |
7fa27001 | 4362 | * or balancing the tree, depending on the rcu_fanout_exact boot parameter. |
64db4cff | 4363 | */ |
199977bf | 4364 | static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt) |
64db4cff | 4365 | { |
64db4cff PM |
4366 | int i; |
4367 | ||
7fa27001 | 4368 | if (rcu_fanout_exact) { |
199977bf | 4369 | levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; |
66292405 | 4370 | for (i = rcu_num_lvls - 2; i >= 0; i--) |
199977bf | 4371 | levelspread[i] = RCU_FANOUT; |
66292405 PM |
4372 | } else { |
4373 | int ccur; | |
4374 | int cprv; | |
4375 | ||
4376 | cprv = nr_cpu_ids; | |
4377 | for (i = rcu_num_lvls - 1; i >= 0; i--) { | |
199977bf AG |
4378 | ccur = levelcnt[i]; |
4379 | levelspread[i] = (cprv + ccur - 1) / ccur; | |
66292405 PM |
4380 | cprv = ccur; |
4381 | } | |
64db4cff PM |
4382 | } |
4383 | } | |
64db4cff PM |
4384 | |
4385 | /* | |
4386 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
4387 | */ | |
394f99a9 LJ |
4388 | static void __init rcu_init_one(struct rcu_state *rsp, |
4389 | struct rcu_data __percpu *rda) | |
64db4cff | 4390 | { |
cb007102 AG |
4391 | static const char * const buf[] = RCU_NODE_NAME_INIT; |
4392 | static const char * const fqs[] = RCU_FQS_NAME_INIT; | |
385b73c0 | 4393 | static const char * const exp[] = RCU_EXP_NAME_INIT; |
4a81e832 | 4394 | static u8 fl_mask = 0x1; |
199977bf AG |
4395 | |
4396 | int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */ | |
4397 | int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */ | |
64db4cff PM |
4398 | int cpustride = 1; |
4399 | int i; | |
4400 | int j; | |
4401 | struct rcu_node *rnp; | |
4402 | ||
05b84aec | 4403 | BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
b6407e86 | 4404 | |
3eaaaf6c PM |
4405 | /* Silence gcc 4.8 false positive about array index out of range. */ |
4406 | if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS) | |
4407 | panic("rcu_init_one: rcu_num_lvls out of range"); | |
4930521a | 4408 | |
64db4cff PM |
4409 | /* Initialize the level-tracking arrays. */ |
4410 | ||
f885b7f2 | 4411 | for (i = 0; i < rcu_num_lvls; i++) |
199977bf | 4412 | levelcnt[i] = num_rcu_lvl[i]; |
f885b7f2 | 4413 | for (i = 1; i < rcu_num_lvls; i++) |
199977bf AG |
4414 | rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1]; |
4415 | rcu_init_levelspread(levelspread, levelcnt); | |
4a81e832 PM |
4416 | rsp->flavor_mask = fl_mask; |
4417 | fl_mask <<= 1; | |
64db4cff PM |
4418 | |
4419 | /* Initialize the elements themselves, starting from the leaves. */ | |
4420 | ||
f885b7f2 | 4421 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
199977bf | 4422 | cpustride *= levelspread[i]; |
64db4cff | 4423 | rnp = rsp->level[i]; |
199977bf | 4424 | for (j = 0; j < levelcnt[i]; j++, rnp++) { |
1304afb2 | 4425 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
4426 | lockdep_set_class_and_name(&rnp->lock, |
4427 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
4428 | raw_spin_lock_init(&rnp->fqslock); |
4429 | lockdep_set_class_and_name(&rnp->fqslock, | |
4430 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
4431 | rnp->gpnum = rsp->gpnum; |
4432 | rnp->completed = rsp->completed; | |
64db4cff PM |
4433 | rnp->qsmask = 0; |
4434 | rnp->qsmaskinit = 0; | |
4435 | rnp->grplo = j * cpustride; | |
4436 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
4437 | if (rnp->grphi >= nr_cpu_ids) |
4438 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
4439 | if (i == 0) { |
4440 | rnp->grpnum = 0; | |
4441 | rnp->grpmask = 0; | |
4442 | rnp->parent = NULL; | |
4443 | } else { | |
199977bf | 4444 | rnp->grpnum = j % levelspread[i - 1]; |
64db4cff PM |
4445 | rnp->grpmask = 1UL << rnp->grpnum; |
4446 | rnp->parent = rsp->level[i - 1] + | |
199977bf | 4447 | j / levelspread[i - 1]; |
64db4cff PM |
4448 | } |
4449 | rnp->level = i; | |
12f5f524 | 4450 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 4451 | rcu_init_one_nocb(rnp); |
385b73c0 | 4452 | mutex_init(&rnp->exp_funnel_mutex); |
83c2c735 PM |
4453 | lockdep_set_class_and_name(&rnp->exp_funnel_mutex, |
4454 | &rcu_exp_class[i], exp[i]); | |
64db4cff PM |
4455 | } |
4456 | } | |
0c34029a | 4457 | |
b3dbec76 | 4458 | init_waitqueue_head(&rsp->gp_wq); |
f4ecea30 | 4459 | init_waitqueue_head(&rsp->expedited_wq); |
f885b7f2 | 4460 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 4461 | for_each_possible_cpu(i) { |
4a90a068 | 4462 | while (i > rnp->grphi) |
0c34029a | 4463 | rnp++; |
394f99a9 | 4464 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
4465 | rcu_boot_init_percpu_data(i, rsp); |
4466 | } | |
6ce75a23 | 4467 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
4468 | } |
4469 | ||
f885b7f2 PM |
4470 | /* |
4471 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 4472 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
4473 | * the ->node array in the rcu_state structure. |
4474 | */ | |
4475 | static void __init rcu_init_geometry(void) | |
4476 | { | |
026ad283 | 4477 | ulong d; |
f885b7f2 | 4478 | int i; |
05b84aec | 4479 | int rcu_capacity[RCU_NUM_LVLS]; |
f885b7f2 | 4480 | |
026ad283 PM |
4481 | /* |
4482 | * Initialize any unspecified boot parameters. | |
4483 | * The default values of jiffies_till_first_fqs and | |
4484 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
4485 | * value, which is a function of HZ, then adding one for each | |
4486 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
4487 | */ | |
4488 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
4489 | if (jiffies_till_first_fqs == ULONG_MAX) | |
4490 | jiffies_till_first_fqs = d; | |
4491 | if (jiffies_till_next_fqs == ULONG_MAX) | |
4492 | jiffies_till_next_fqs = d; | |
4493 | ||
f885b7f2 | 4494 | /* If the compile-time values are accurate, just leave. */ |
47d631af | 4495 | if (rcu_fanout_leaf == RCU_FANOUT_LEAF && |
b17c7035 | 4496 | nr_cpu_ids == NR_CPUS) |
f885b7f2 | 4497 | return; |
39479098 PM |
4498 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", |
4499 | rcu_fanout_leaf, nr_cpu_ids); | |
f885b7f2 | 4500 | |
f885b7f2 | 4501 | /* |
ee968ac6 PM |
4502 | * The boot-time rcu_fanout_leaf parameter must be at least two |
4503 | * and cannot exceed the number of bits in the rcu_node masks. | |
4504 | * Complain and fall back to the compile-time values if this | |
4505 | * limit is exceeded. | |
f885b7f2 | 4506 | */ |
ee968ac6 | 4507 | if (rcu_fanout_leaf < 2 || |
75cf15a4 | 4508 | rcu_fanout_leaf > sizeof(unsigned long) * 8) { |
13bd6494 | 4509 | rcu_fanout_leaf = RCU_FANOUT_LEAF; |
f885b7f2 PM |
4510 | WARN_ON(1); |
4511 | return; | |
4512 | } | |
4513 | ||
f885b7f2 PM |
4514 | /* |
4515 | * Compute number of nodes that can be handled an rcu_node tree | |
9618138b | 4516 | * with the given number of levels. |
f885b7f2 | 4517 | */ |
9618138b | 4518 | rcu_capacity[0] = rcu_fanout_leaf; |
05b84aec | 4519 | for (i = 1; i < RCU_NUM_LVLS; i++) |
05c5df31 | 4520 | rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT; |
f885b7f2 PM |
4521 | |
4522 | /* | |
75cf15a4 | 4523 | * The tree must be able to accommodate the configured number of CPUs. |
ee968ac6 | 4524 | * If this limit is exceeded, fall back to the compile-time values. |
f885b7f2 | 4525 | */ |
ee968ac6 PM |
4526 | if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) { |
4527 | rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
4528 | WARN_ON(1); | |
4529 | return; | |
4530 | } | |
f885b7f2 | 4531 | |
679f9858 | 4532 | /* Calculate the number of levels in the tree. */ |
9618138b | 4533 | for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) { |
679f9858 | 4534 | } |
9618138b | 4535 | rcu_num_lvls = i + 1; |
679f9858 | 4536 | |
f885b7f2 | 4537 | /* Calculate the number of rcu_nodes at each level of the tree. */ |
679f9858 | 4538 | for (i = 0; i < rcu_num_lvls; i++) { |
9618138b | 4539 | int cap = rcu_capacity[(rcu_num_lvls - 1) - i]; |
679f9858 AG |
4540 | num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap); |
4541 | } | |
f885b7f2 PM |
4542 | |
4543 | /* Calculate the total number of rcu_node structures. */ | |
4544 | rcu_num_nodes = 0; | |
679f9858 | 4545 | for (i = 0; i < rcu_num_lvls; i++) |
f885b7f2 | 4546 | rcu_num_nodes += num_rcu_lvl[i]; |
f885b7f2 PM |
4547 | } |
4548 | ||
a3dc2948 PM |
4549 | /* |
4550 | * Dump out the structure of the rcu_node combining tree associated | |
4551 | * with the rcu_state structure referenced by rsp. | |
4552 | */ | |
4553 | static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp) | |
4554 | { | |
4555 | int level = 0; | |
4556 | struct rcu_node *rnp; | |
4557 | ||
4558 | pr_info("rcu_node tree layout dump\n"); | |
4559 | pr_info(" "); | |
4560 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
4561 | if (rnp->level != level) { | |
4562 | pr_cont("\n"); | |
4563 | pr_info(" "); | |
4564 | level = rnp->level; | |
4565 | } | |
4566 | pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum); | |
4567 | } | |
4568 | pr_cont("\n"); | |
4569 | } | |
4570 | ||
9f680ab4 | 4571 | void __init rcu_init(void) |
64db4cff | 4572 | { |
017c4261 | 4573 | int cpu; |
9f680ab4 | 4574 | |
47627678 PM |
4575 | rcu_early_boot_tests(); |
4576 | ||
f41d911f | 4577 | rcu_bootup_announce(); |
f885b7f2 | 4578 | rcu_init_geometry(); |
394f99a9 | 4579 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); |
69c8d28c | 4580 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
a3dc2948 PM |
4581 | if (dump_tree) |
4582 | rcu_dump_rcu_node_tree(&rcu_sched_state); | |
f41d911f | 4583 | __rcu_init_preempt(); |
b5b39360 | 4584 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
4585 | |
4586 | /* | |
4587 | * We don't need protection against CPU-hotplug here because | |
4588 | * this is called early in boot, before either interrupts | |
4589 | * or the scheduler are operational. | |
4590 | */ | |
4591 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 4592 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
4593 | for_each_online_cpu(cpu) |
4594 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
4595 | } |
4596 | ||
4102adab | 4597 | #include "tree_plugin.h" |