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 | ||
f7f7bac9 SRRH |
71 | /* |
72 | * In order to export the rcu_state name to the tracing tools, it | |
73 | * needs to be added in the __tracepoint_string section. | |
74 | * This requires defining a separate variable tp_<sname>_varname | |
75 | * that points to the string being used, and this will allow | |
76 | * the tracing userspace tools to be able to decipher the string | |
77 | * address to the matching string. | |
78 | */ | |
a8a29b3b AB |
79 | #ifdef CONFIG_TRACING |
80 | # define DEFINE_RCU_TPS(sname) \ | |
f7f7bac9 | 81 | static char sname##_varname[] = #sname; \ |
a8a29b3b AB |
82 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; |
83 | # define RCU_STATE_NAME(sname) sname##_varname | |
84 | #else | |
85 | # define DEFINE_RCU_TPS(sname) | |
86 | # define RCU_STATE_NAME(sname) __stringify(sname) | |
87 | #endif | |
88 | ||
89 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ | |
90 | DEFINE_RCU_TPS(sname) \ | |
c92fb057 | 91 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \ |
a41bfeb2 | 92 | struct rcu_state sname##_state = { \ |
6c90cc7b | 93 | .level = { &sname##_state.node[0] }, \ |
2723249a | 94 | .rda = &sname##_data, \ |
037b64ed | 95 | .call = cr, \ |
77f81fe0 | 96 | .gp_state = RCU_GP_IDLE, \ |
42c3533e PM |
97 | .gpnum = 0UL - 300UL, \ |
98 | .completed = 0UL - 300UL, \ | |
7b2e6011 | 99 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
6c90cc7b PM |
100 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
101 | .orphan_donetail = &sname##_state.orphan_donelist, \ | |
7be7f0be | 102 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
a8a29b3b | 103 | .name = RCU_STATE_NAME(sname), \ |
a4889858 | 104 | .abbr = sabbr, \ |
2723249a | 105 | } |
64db4cff | 106 | |
a41bfeb2 SRRH |
107 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); |
108 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); | |
b1f77b05 | 109 | |
b28a7c01 | 110 | static struct rcu_state *const rcu_state_p; |
2927a689 | 111 | static struct rcu_data __percpu *const rcu_data_p; |
6ce75a23 | 112 | LIST_HEAD(rcu_struct_flavors); |
27f4d280 | 113 | |
a3dc2948 PM |
114 | /* Dump rcu_node combining tree at boot to verify correct setup. */ |
115 | static bool dump_tree; | |
116 | module_param(dump_tree, bool, 0444); | |
7fa27001 PM |
117 | /* Control rcu_node-tree auto-balancing at boot time. */ |
118 | static bool rcu_fanout_exact; | |
119 | module_param(rcu_fanout_exact, bool, 0444); | |
47d631af PM |
120 | /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */ |
121 | static int rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
7e5c2dfb | 122 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 | 123 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
cb007102 AG |
124 | /* Number of rcu_nodes at specified level. */ |
125 | static int num_rcu_lvl[] = NUM_RCU_LVL_INIT; | |
f885b7f2 PM |
126 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ |
127 | ||
b0d30417 PM |
128 | /* |
129 | * The rcu_scheduler_active variable transitions from zero to one just | |
130 | * before the first task is spawned. So when this variable is zero, RCU | |
131 | * can assume that there is but one task, allowing RCU to (for example) | |
b44f6656 | 132 | * optimize synchronize_sched() to a simple barrier(). When this variable |
b0d30417 PM |
133 | * is one, RCU must actually do all the hard work required to detect real |
134 | * grace periods. This variable is also used to suppress boot-time false | |
135 | * positives from lockdep-RCU error checking. | |
136 | */ | |
bbad9379 PM |
137 | int rcu_scheduler_active __read_mostly; |
138 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
139 | ||
b0d30417 PM |
140 | /* |
141 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
142 | * during the early_initcall() processing, which is after the scheduler | |
143 | * is capable of creating new tasks. So RCU processing (for example, | |
144 | * creating tasks for RCU priority boosting) must be delayed until after | |
145 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
146 | * currently delay invocation of any RCU callbacks until after this point. | |
147 | * | |
148 | * It might later prove better for people registering RCU callbacks during | |
149 | * early boot to take responsibility for these callbacks, but one step at | |
150 | * a time. | |
151 | */ | |
152 | static int rcu_scheduler_fully_active __read_mostly; | |
153 | ||
0aa04b05 PM |
154 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); |
155 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); | |
5d01bbd1 | 156 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 PM |
157 | static void invoke_rcu_core(void); |
158 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); | |
6587a23b PM |
159 | static void rcu_report_exp_rdp(struct rcu_state *rsp, |
160 | struct rcu_data *rdp, bool wake); | |
a26ac245 | 161 | |
a94844b2 | 162 | /* rcuc/rcub kthread realtime priority */ |
26730f55 | 163 | #ifdef CONFIG_RCU_KTHREAD_PRIO |
a94844b2 | 164 | static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO; |
26730f55 PM |
165 | #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */ |
166 | static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0; | |
167 | #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */ | |
a94844b2 PM |
168 | module_param(kthread_prio, int, 0644); |
169 | ||
8d7dc928 | 170 | /* Delay in jiffies for grace-period initialization delays, debug only. */ |
0f41c0dd PM |
171 | |
172 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT | |
173 | static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY; | |
174 | module_param(gp_preinit_delay, int, 0644); | |
175 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ | |
176 | static const int gp_preinit_delay; | |
177 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ | |
178 | ||
8d7dc928 PM |
179 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT |
180 | static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY; | |
37745d28 | 181 | module_param(gp_init_delay, int, 0644); |
8d7dc928 PM |
182 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ |
183 | static const int gp_init_delay; | |
184 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ | |
eab128e8 | 185 | |
0f41c0dd PM |
186 | #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP |
187 | static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY; | |
188 | module_param(gp_cleanup_delay, int, 0644); | |
189 | #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ | |
190 | static const int gp_cleanup_delay; | |
191 | #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ | |
192 | ||
eab128e8 PM |
193 | /* |
194 | * Number of grace periods between delays, normalized by the duration of | |
195 | * the delay. The longer the the delay, the more the grace periods between | |
196 | * each delay. The reason for this normalization is that it means that, | |
197 | * for non-zero delays, the overall slowdown of grace periods is constant | |
198 | * regardless of the duration of the delay. This arrangement balances | |
199 | * the need for long delays to increase some race probabilities with the | |
200 | * need for fast grace periods to increase other race probabilities. | |
201 | */ | |
202 | #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */ | |
37745d28 | 203 | |
4a298656 PM |
204 | /* |
205 | * Track the rcutorture test sequence number and the update version | |
206 | * number within a given test. The rcutorture_testseq is incremented | |
207 | * on every rcutorture module load and unload, so has an odd value | |
208 | * when a test is running. The rcutorture_vernum is set to zero | |
209 | * when rcutorture starts and is incremented on each rcutorture update. | |
210 | * These variables enable correlating rcutorture output with the | |
211 | * RCU tracing information. | |
212 | */ | |
213 | unsigned long rcutorture_testseq; | |
214 | unsigned long rcutorture_vernum; | |
215 | ||
0aa04b05 PM |
216 | /* |
217 | * Compute the mask of online CPUs for the specified rcu_node structure. | |
218 | * This will not be stable unless the rcu_node structure's ->lock is | |
219 | * held, but the bit corresponding to the current CPU will be stable | |
220 | * in most contexts. | |
221 | */ | |
222 | unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) | |
223 | { | |
7d0ae808 | 224 | return READ_ONCE(rnp->qsmaskinitnext); |
0aa04b05 PM |
225 | } |
226 | ||
fc2219d4 | 227 | /* |
7d0ae808 | 228 | * Return true if an RCU grace period is in progress. The READ_ONCE()s |
fc2219d4 PM |
229 | * permit this function to be invoked without holding the root rcu_node |
230 | * structure's ->lock, but of course results can be subject to change. | |
231 | */ | |
232 | static int rcu_gp_in_progress(struct rcu_state *rsp) | |
233 | { | |
7d0ae808 | 234 | return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum); |
fc2219d4 PM |
235 | } |
236 | ||
b1f77b05 | 237 | /* |
d6714c22 | 238 | * Note a quiescent state. Because we do not need to know |
b1f77b05 | 239 | * how many quiescent states passed, just if there was at least |
d6714c22 | 240 | * one since the start of the grace period, this just sets a flag. |
e4cc1f22 | 241 | * The caller must have disabled preemption. |
b1f77b05 | 242 | */ |
284a8c93 | 243 | void rcu_sched_qs(void) |
b1f77b05 | 244 | { |
fecbf6f0 PM |
245 | if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s)) |
246 | return; | |
247 | trace_rcu_grace_period(TPS("rcu_sched"), | |
248 | __this_cpu_read(rcu_sched_data.gpnum), | |
249 | TPS("cpuqs")); | |
250 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false); | |
251 | if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) | |
252 | return; | |
46a5d164 PM |
253 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false); |
254 | rcu_report_exp_rdp(&rcu_sched_state, | |
255 | this_cpu_ptr(&rcu_sched_data), true); | |
b1f77b05 IM |
256 | } |
257 | ||
284a8c93 | 258 | void rcu_bh_qs(void) |
b1f77b05 | 259 | { |
5b74c458 | 260 | if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) { |
284a8c93 PM |
261 | trace_rcu_grace_period(TPS("rcu_bh"), |
262 | __this_cpu_read(rcu_bh_data.gpnum), | |
263 | TPS("cpuqs")); | |
5b74c458 | 264 | __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false); |
284a8c93 | 265 | } |
b1f77b05 | 266 | } |
64db4cff | 267 | |
4a81e832 PM |
268 | static DEFINE_PER_CPU(int, rcu_sched_qs_mask); |
269 | ||
270 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { | |
271 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, | |
272 | .dynticks = ATOMIC_INIT(1), | |
273 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
274 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, | |
275 | .dynticks_idle = ATOMIC_INIT(1), | |
276 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
277 | }; | |
278 | ||
5cd37193 PM |
279 | DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr); |
280 | EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr); | |
281 | ||
4a81e832 PM |
282 | /* |
283 | * Let the RCU core know that this CPU has gone through the scheduler, | |
284 | * which is a quiescent state. This is called when the need for a | |
285 | * quiescent state is urgent, so we burn an atomic operation and full | |
286 | * memory barriers to let the RCU core know about it, regardless of what | |
287 | * this CPU might (or might not) do in the near future. | |
288 | * | |
289 | * We inform the RCU core by emulating a zero-duration dyntick-idle | |
290 | * period, which we in turn do by incrementing the ->dynticks counter | |
291 | * by two. | |
46a5d164 PM |
292 | * |
293 | * The caller must have disabled interrupts. | |
4a81e832 PM |
294 | */ |
295 | static void rcu_momentary_dyntick_idle(void) | |
296 | { | |
4a81e832 PM |
297 | struct rcu_data *rdp; |
298 | struct rcu_dynticks *rdtp; | |
299 | int resched_mask; | |
300 | struct rcu_state *rsp; | |
301 | ||
4a81e832 PM |
302 | /* |
303 | * Yes, we can lose flag-setting operations. This is OK, because | |
304 | * the flag will be set again after some delay. | |
305 | */ | |
306 | resched_mask = raw_cpu_read(rcu_sched_qs_mask); | |
307 | raw_cpu_write(rcu_sched_qs_mask, 0); | |
308 | ||
309 | /* Find the flavor that needs a quiescent state. */ | |
310 | for_each_rcu_flavor(rsp) { | |
311 | rdp = raw_cpu_ptr(rsp->rda); | |
312 | if (!(resched_mask & rsp->flavor_mask)) | |
313 | continue; | |
314 | smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */ | |
7d0ae808 PM |
315 | if (READ_ONCE(rdp->mynode->completed) != |
316 | READ_ONCE(rdp->cond_resched_completed)) | |
4a81e832 PM |
317 | continue; |
318 | ||
319 | /* | |
320 | * Pretend to be momentarily idle for the quiescent state. | |
321 | * This allows the grace-period kthread to record the | |
322 | * quiescent state, with no need for this CPU to do anything | |
323 | * further. | |
324 | */ | |
325 | rdtp = this_cpu_ptr(&rcu_dynticks); | |
326 | smp_mb__before_atomic(); /* Earlier stuff before QS. */ | |
327 | atomic_add(2, &rdtp->dynticks); /* QS. */ | |
328 | smp_mb__after_atomic(); /* Later stuff after QS. */ | |
329 | break; | |
330 | } | |
4a81e832 PM |
331 | } |
332 | ||
25502a6c PM |
333 | /* |
334 | * Note a context switch. This is a quiescent state for RCU-sched, | |
335 | * and requires special handling for preemptible RCU. | |
46a5d164 | 336 | * The caller must have disabled interrupts. |
25502a6c | 337 | */ |
38200cf2 | 338 | void rcu_note_context_switch(void) |
25502a6c | 339 | { |
bb73c52b | 340 | barrier(); /* Avoid RCU read-side critical sections leaking down. */ |
f7f7bac9 | 341 | trace_rcu_utilization(TPS("Start context switch")); |
284a8c93 | 342 | rcu_sched_qs(); |
38200cf2 | 343 | rcu_preempt_note_context_switch(); |
4a81e832 PM |
344 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) |
345 | rcu_momentary_dyntick_idle(); | |
f7f7bac9 | 346 | trace_rcu_utilization(TPS("End context switch")); |
bb73c52b | 347 | barrier(); /* Avoid RCU read-side critical sections leaking up. */ |
25502a6c | 348 | } |
29ce8310 | 349 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
25502a6c | 350 | |
5cd37193 | 351 | /* |
1925d196 | 352 | * Register a quiescent state for all RCU flavors. If there is an |
5cd37193 PM |
353 | * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight |
354 | * dyntick-idle quiescent state visible to other CPUs (but only for those | |
1925d196 | 355 | * RCU flavors in desperate need of a quiescent state, which will normally |
5cd37193 PM |
356 | * be none of them). Either way, do a lightweight quiescent state for |
357 | * all RCU flavors. | |
bb73c52b BF |
358 | * |
359 | * The barrier() calls are redundant in the common case when this is | |
360 | * called externally, but just in case this is called from within this | |
361 | * file. | |
362 | * | |
5cd37193 PM |
363 | */ |
364 | void rcu_all_qs(void) | |
365 | { | |
46a5d164 PM |
366 | unsigned long flags; |
367 | ||
bb73c52b | 368 | barrier(); /* Avoid RCU read-side critical sections leaking down. */ |
46a5d164 PM |
369 | if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) { |
370 | local_irq_save(flags); | |
5cd37193 | 371 | rcu_momentary_dyntick_idle(); |
46a5d164 PM |
372 | local_irq_restore(flags); |
373 | } | |
5cd37193 | 374 | this_cpu_inc(rcu_qs_ctr); |
bb73c52b | 375 | barrier(); /* Avoid RCU read-side critical sections leaking up. */ |
5cd37193 PM |
376 | } |
377 | EXPORT_SYMBOL_GPL(rcu_all_qs); | |
378 | ||
878d7439 ED |
379 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
380 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ | |
381 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ | |
64db4cff | 382 | |
878d7439 ED |
383 | module_param(blimit, long, 0444); |
384 | module_param(qhimark, long, 0444); | |
385 | module_param(qlowmark, long, 0444); | |
3d76c082 | 386 | |
026ad283 PM |
387 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
388 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
d40011f6 PM |
389 | |
390 | module_param(jiffies_till_first_fqs, ulong, 0644); | |
391 | module_param(jiffies_till_next_fqs, ulong, 0644); | |
392 | ||
4a81e832 PM |
393 | /* |
394 | * How long the grace period must be before we start recruiting | |
395 | * quiescent-state help from rcu_note_context_switch(). | |
396 | */ | |
397 | static ulong jiffies_till_sched_qs = HZ / 20; | |
398 | module_param(jiffies_till_sched_qs, ulong, 0644); | |
399 | ||
48a7639c | 400 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
910ee45d | 401 | struct rcu_data *rdp); |
217af2a2 PM |
402 | static void force_qs_rnp(struct rcu_state *rsp, |
403 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
404 | unsigned long *maxj), | |
405 | bool *isidle, unsigned long *maxj); | |
4cdfc175 | 406 | static void force_quiescent_state(struct rcu_state *rsp); |
e3950ecd | 407 | static int rcu_pending(void); |
64db4cff PM |
408 | |
409 | /* | |
917963d0 | 410 | * Return the number of RCU batches started thus far for debug & stats. |
64db4cff | 411 | */ |
917963d0 PM |
412 | unsigned long rcu_batches_started(void) |
413 | { | |
414 | return rcu_state_p->gpnum; | |
415 | } | |
416 | EXPORT_SYMBOL_GPL(rcu_batches_started); | |
417 | ||
418 | /* | |
419 | * Return the number of RCU-sched batches started thus far for debug & stats. | |
64db4cff | 420 | */ |
917963d0 PM |
421 | unsigned long rcu_batches_started_sched(void) |
422 | { | |
423 | return rcu_sched_state.gpnum; | |
424 | } | |
425 | EXPORT_SYMBOL_GPL(rcu_batches_started_sched); | |
426 | ||
427 | /* | |
428 | * Return the number of RCU BH batches started thus far for debug & stats. | |
429 | */ | |
430 | unsigned long rcu_batches_started_bh(void) | |
431 | { | |
432 | return rcu_bh_state.gpnum; | |
433 | } | |
434 | EXPORT_SYMBOL_GPL(rcu_batches_started_bh); | |
435 | ||
436 | /* | |
437 | * Return the number of RCU batches completed thus far for debug & stats. | |
438 | */ | |
439 | unsigned long rcu_batches_completed(void) | |
440 | { | |
441 | return rcu_state_p->completed; | |
442 | } | |
443 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
444 | ||
445 | /* | |
446 | * Return the number of RCU-sched batches completed thus far for debug & stats. | |
64db4cff | 447 | */ |
9733e4f0 | 448 | unsigned long rcu_batches_completed_sched(void) |
64db4cff | 449 | { |
d6714c22 | 450 | return rcu_sched_state.completed; |
64db4cff | 451 | } |
d6714c22 | 452 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
64db4cff PM |
453 | |
454 | /* | |
917963d0 | 455 | * Return the number of RCU BH batches completed thus far for debug & stats. |
64db4cff | 456 | */ |
9733e4f0 | 457 | unsigned long rcu_batches_completed_bh(void) |
64db4cff PM |
458 | { |
459 | return rcu_bh_state.completed; | |
460 | } | |
461 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
462 | ||
a381d757 ACB |
463 | /* |
464 | * Force a quiescent state. | |
465 | */ | |
466 | void rcu_force_quiescent_state(void) | |
467 | { | |
e534165b | 468 | force_quiescent_state(rcu_state_p); |
a381d757 ACB |
469 | } |
470 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
471 | ||
bf66f18e PM |
472 | /* |
473 | * Force a quiescent state for RCU BH. | |
474 | */ | |
475 | void rcu_bh_force_quiescent_state(void) | |
476 | { | |
4cdfc175 | 477 | force_quiescent_state(&rcu_bh_state); |
bf66f18e PM |
478 | } |
479 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); | |
480 | ||
e7580f33 PM |
481 | /* |
482 | * Force a quiescent state for RCU-sched. | |
483 | */ | |
484 | void rcu_sched_force_quiescent_state(void) | |
485 | { | |
486 | force_quiescent_state(&rcu_sched_state); | |
487 | } | |
488 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); | |
489 | ||
afea227f PM |
490 | /* |
491 | * Show the state of the grace-period kthreads. | |
492 | */ | |
493 | void show_rcu_gp_kthreads(void) | |
494 | { | |
495 | struct rcu_state *rsp; | |
496 | ||
497 | for_each_rcu_flavor(rsp) { | |
498 | pr_info("%s: wait state: %d ->state: %#lx\n", | |
499 | rsp->name, rsp->gp_state, rsp->gp_kthread->state); | |
500 | /* sched_show_task(rsp->gp_kthread); */ | |
501 | } | |
502 | } | |
503 | EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); | |
504 | ||
4a298656 PM |
505 | /* |
506 | * Record the number of times rcutorture tests have been initiated and | |
507 | * terminated. This information allows the debugfs tracing stats to be | |
508 | * correlated to the rcutorture messages, even when the rcutorture module | |
509 | * is being repeatedly loaded and unloaded. In other words, we cannot | |
510 | * store this state in rcutorture itself. | |
511 | */ | |
512 | void rcutorture_record_test_transition(void) | |
513 | { | |
514 | rcutorture_testseq++; | |
515 | rcutorture_vernum = 0; | |
516 | } | |
517 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); | |
518 | ||
ad0dc7f9 PM |
519 | /* |
520 | * Send along grace-period-related data for rcutorture diagnostics. | |
521 | */ | |
522 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
523 | unsigned long *gpnum, unsigned long *completed) | |
524 | { | |
525 | struct rcu_state *rsp = NULL; | |
526 | ||
527 | switch (test_type) { | |
528 | case RCU_FLAVOR: | |
e534165b | 529 | rsp = rcu_state_p; |
ad0dc7f9 PM |
530 | break; |
531 | case RCU_BH_FLAVOR: | |
532 | rsp = &rcu_bh_state; | |
533 | break; | |
534 | case RCU_SCHED_FLAVOR: | |
535 | rsp = &rcu_sched_state; | |
536 | break; | |
537 | default: | |
538 | break; | |
539 | } | |
540 | if (rsp != NULL) { | |
7d0ae808 PM |
541 | *flags = READ_ONCE(rsp->gp_flags); |
542 | *gpnum = READ_ONCE(rsp->gpnum); | |
543 | *completed = READ_ONCE(rsp->completed); | |
ad0dc7f9 PM |
544 | return; |
545 | } | |
546 | *flags = 0; | |
547 | *gpnum = 0; | |
548 | *completed = 0; | |
549 | } | |
550 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
551 | ||
4a298656 PM |
552 | /* |
553 | * Record the number of writer passes through the current rcutorture test. | |
554 | * This is also used to correlate debugfs tracing stats with the rcutorture | |
555 | * messages. | |
556 | */ | |
557 | void rcutorture_record_progress(unsigned long vernum) | |
558 | { | |
559 | rcutorture_vernum++; | |
560 | } | |
561 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); | |
562 | ||
64db4cff PM |
563 | /* |
564 | * Does the CPU have callbacks ready to be invoked? | |
565 | */ | |
566 | static int | |
567 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) | |
568 | { | |
3fbfbf7a PM |
569 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
570 | rdp->nxttail[RCU_DONE_TAIL] != NULL; | |
64db4cff PM |
571 | } |
572 | ||
365187fb PM |
573 | /* |
574 | * Return the root node of the specified rcu_state structure. | |
575 | */ | |
576 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) | |
577 | { | |
578 | return &rsp->node[0]; | |
579 | } | |
580 | ||
581 | /* | |
582 | * Is there any need for future grace periods? | |
583 | * Interrupts must be disabled. If the caller does not hold the root | |
584 | * rnp_node structure's ->lock, the results are advisory only. | |
585 | */ | |
586 | static int rcu_future_needs_gp(struct rcu_state *rsp) | |
587 | { | |
588 | struct rcu_node *rnp = rcu_get_root(rsp); | |
7d0ae808 | 589 | int idx = (READ_ONCE(rnp->completed) + 1) & 0x1; |
365187fb PM |
590 | int *fp = &rnp->need_future_gp[idx]; |
591 | ||
7d0ae808 | 592 | return READ_ONCE(*fp); |
365187fb PM |
593 | } |
594 | ||
64db4cff | 595 | /* |
dc35c893 PM |
596 | * Does the current CPU require a not-yet-started grace period? |
597 | * The caller must have disabled interrupts to prevent races with | |
598 | * normal callback registry. | |
64db4cff PM |
599 | */ |
600 | static int | |
601 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) | |
602 | { | |
dc35c893 | 603 | int i; |
3fbfbf7a | 604 | |
dc35c893 PM |
605 | if (rcu_gp_in_progress(rsp)) |
606 | return 0; /* No, a grace period is already in progress. */ | |
365187fb | 607 | if (rcu_future_needs_gp(rsp)) |
34ed6246 | 608 | return 1; /* Yes, a no-CBs CPU needs one. */ |
dc35c893 PM |
609 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
610 | return 0; /* No, this is a no-CBs (or offline) CPU. */ | |
611 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) | |
612 | return 1; /* Yes, this CPU has newly registered callbacks. */ | |
613 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) | |
614 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && | |
7d0ae808 | 615 | ULONG_CMP_LT(READ_ONCE(rsp->completed), |
dc35c893 PM |
616 | rdp->nxtcompleted[i])) |
617 | return 1; /* Yes, CBs for future grace period. */ | |
618 | return 0; /* No grace period needed. */ | |
64db4cff PM |
619 | } |
620 | ||
9b2e4f18 | 621 | /* |
adf5091e | 622 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
9b2e4f18 PM |
623 | * |
624 | * If the new value of the ->dynticks_nesting counter now is zero, | |
625 | * we really have entered idle, and must do the appropriate accounting. | |
626 | * The caller must have disabled interrupts. | |
627 | */ | |
28ced795 | 628 | static void rcu_eqs_enter_common(long long oldval, bool user) |
9b2e4f18 | 629 | { |
96d3fd0d PM |
630 | struct rcu_state *rsp; |
631 | struct rcu_data *rdp; | |
28ced795 | 632 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
96d3fd0d | 633 | |
f7f7bac9 | 634 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
1ce46ee5 PM |
635 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
636 | !user && !is_idle_task(current)) { | |
289828e6 PM |
637 | struct task_struct *idle __maybe_unused = |
638 | idle_task(smp_processor_id()); | |
0989cb46 | 639 | |
f7f7bac9 | 640 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
bf1304e9 | 641 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
642 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
643 | current->pid, current->comm, | |
644 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 | 645 | } |
96d3fd0d PM |
646 | for_each_rcu_flavor(rsp) { |
647 | rdp = this_cpu_ptr(rsp->rda); | |
648 | do_nocb_deferred_wakeup(rdp); | |
649 | } | |
198bbf81 | 650 | rcu_prepare_for_idle(); |
9b2e4f18 | 651 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 652 | smp_mb__before_atomic(); /* See above. */ |
9b2e4f18 | 653 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 654 | smp_mb__after_atomic(); /* Force ordering with next sojourn. */ |
1ce46ee5 PM |
655 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
656 | atomic_read(&rdtp->dynticks) & 0x1); | |
176f8f7a | 657 | rcu_dynticks_task_enter(); |
c44e2cdd PM |
658 | |
659 | /* | |
adf5091e | 660 | * It is illegal to enter an extended quiescent state while |
c44e2cdd PM |
661 | * in an RCU read-side critical section. |
662 | */ | |
f78f5b90 PM |
663 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map), |
664 | "Illegal idle entry in RCU read-side critical section."); | |
665 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map), | |
666 | "Illegal idle entry in RCU-bh read-side critical section."); | |
667 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map), | |
668 | "Illegal idle entry in RCU-sched read-side critical section."); | |
9b2e4f18 | 669 | } |
64db4cff | 670 | |
adf5091e FW |
671 | /* |
672 | * Enter an RCU extended quiescent state, which can be either the | |
673 | * idle loop or adaptive-tickless usermode execution. | |
64db4cff | 674 | */ |
adf5091e | 675 | static void rcu_eqs_enter(bool user) |
64db4cff | 676 | { |
4145fa7f | 677 | long long oldval; |
64db4cff PM |
678 | struct rcu_dynticks *rdtp; |
679 | ||
c9d4b0af | 680 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 681 | oldval = rdtp->dynticks_nesting; |
1ce46ee5 PM |
682 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
683 | (oldval & DYNTICK_TASK_NEST_MASK) == 0); | |
3a592405 | 684 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { |
29e37d81 | 685 | rdtp->dynticks_nesting = 0; |
28ced795 | 686 | rcu_eqs_enter_common(oldval, user); |
3a592405 | 687 | } else { |
29e37d81 | 688 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; |
3a592405 | 689 | } |
64db4cff | 690 | } |
adf5091e FW |
691 | |
692 | /** | |
693 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
694 | * | |
695 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
696 | * read-side critical sections can occur. (Though RCU read-side | |
697 | * critical sections can occur in irq handlers in idle, a possibility | |
698 | * handled by irq_enter() and irq_exit().) | |
699 | * | |
700 | * We crowbar the ->dynticks_nesting field to zero to allow for | |
701 | * the possibility of usermode upcalls having messed up our count | |
702 | * of interrupt nesting level during the prior busy period. | |
703 | */ | |
704 | void rcu_idle_enter(void) | |
705 | { | |
c5d900bf FW |
706 | unsigned long flags; |
707 | ||
708 | local_irq_save(flags); | |
cb349ca9 | 709 | rcu_eqs_enter(false); |
28ced795 | 710 | rcu_sysidle_enter(0); |
c5d900bf | 711 | local_irq_restore(flags); |
adf5091e | 712 | } |
8a2ecf47 | 713 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 714 | |
d1ec4c34 | 715 | #ifdef CONFIG_NO_HZ_FULL |
adf5091e FW |
716 | /** |
717 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
718 | * | |
719 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
720 | * is permitted between this call and rcu_user_exit(). This way the | |
721 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
722 | * when the CPU runs in userspace. | |
723 | */ | |
724 | void rcu_user_enter(void) | |
725 | { | |
91d1aa43 | 726 | rcu_eqs_enter(1); |
adf5091e | 727 | } |
d1ec4c34 | 728 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 729 | |
9b2e4f18 PM |
730 | /** |
731 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
732 | * | |
733 | * Exit from an interrupt handler, which might possibly result in entering | |
734 | * idle mode, in other words, leaving the mode in which read-side critical | |
735 | * sections can occur. | |
64db4cff | 736 | * |
9b2e4f18 PM |
737 | * This code assumes that the idle loop never does anything that might |
738 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
739 | * architecture violates this assumption, RCU will give you what you | |
740 | * deserve, good and hard. But very infrequently and irreproducibly. | |
741 | * | |
742 | * Use things like work queues to work around this limitation. | |
743 | * | |
744 | * You have been warned. | |
64db4cff | 745 | */ |
9b2e4f18 | 746 | void rcu_irq_exit(void) |
64db4cff PM |
747 | { |
748 | unsigned long flags; | |
4145fa7f | 749 | long long oldval; |
64db4cff PM |
750 | struct rcu_dynticks *rdtp; |
751 | ||
752 | local_irq_save(flags); | |
c9d4b0af | 753 | rdtp = this_cpu_ptr(&rcu_dynticks); |
4145fa7f | 754 | oldval = rdtp->dynticks_nesting; |
9b2e4f18 | 755 | rdtp->dynticks_nesting--; |
1ce46ee5 PM |
756 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
757 | rdtp->dynticks_nesting < 0); | |
b6fc6020 | 758 | if (rdtp->dynticks_nesting) |
f7f7bac9 | 759 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 760 | else |
28ced795 CL |
761 | rcu_eqs_enter_common(oldval, true); |
762 | rcu_sysidle_enter(1); | |
9b2e4f18 PM |
763 | local_irq_restore(flags); |
764 | } | |
765 | ||
766 | /* | |
adf5091e | 767 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
9b2e4f18 PM |
768 | * |
769 | * If the new value of the ->dynticks_nesting counter was previously zero, | |
770 | * we really have exited idle, and must do the appropriate accounting. | |
771 | * The caller must have disabled interrupts. | |
772 | */ | |
28ced795 | 773 | static void rcu_eqs_exit_common(long long oldval, int user) |
9b2e4f18 | 774 | { |
28ced795 CL |
775 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
776 | ||
176f8f7a | 777 | rcu_dynticks_task_exit(); |
4e857c58 | 778 | smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */ |
23b5c8fa PM |
779 | atomic_inc(&rdtp->dynticks); |
780 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ | |
4e857c58 | 781 | smp_mb__after_atomic(); /* See above. */ |
1ce46ee5 PM |
782 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
783 | !(atomic_read(&rdtp->dynticks) & 0x1)); | |
8fa7845d | 784 | rcu_cleanup_after_idle(); |
f7f7bac9 | 785 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
1ce46ee5 PM |
786 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
787 | !user && !is_idle_task(current)) { | |
289828e6 PM |
788 | struct task_struct *idle __maybe_unused = |
789 | idle_task(smp_processor_id()); | |
0989cb46 | 790 | |
f7f7bac9 | 791 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
4145fa7f | 792 | oldval, rdtp->dynticks_nesting); |
bf1304e9 | 793 | ftrace_dump(DUMP_ORIG); |
0989cb46 PM |
794 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
795 | current->pid, current->comm, | |
796 | idle->pid, idle->comm); /* must be idle task! */ | |
9b2e4f18 PM |
797 | } |
798 | } | |
799 | ||
adf5091e FW |
800 | /* |
801 | * Exit an RCU extended quiescent state, which can be either the | |
802 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 803 | */ |
adf5091e | 804 | static void rcu_eqs_exit(bool user) |
9b2e4f18 | 805 | { |
9b2e4f18 PM |
806 | struct rcu_dynticks *rdtp; |
807 | long long oldval; | |
808 | ||
c9d4b0af | 809 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 | 810 | oldval = rdtp->dynticks_nesting; |
1ce46ee5 | 811 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0); |
3a592405 | 812 | if (oldval & DYNTICK_TASK_NEST_MASK) { |
29e37d81 | 813 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; |
3a592405 | 814 | } else { |
29e37d81 | 815 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
28ced795 | 816 | rcu_eqs_exit_common(oldval, user); |
3a592405 | 817 | } |
9b2e4f18 | 818 | } |
adf5091e FW |
819 | |
820 | /** | |
821 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
822 | * | |
823 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
824 | * read-side critical sections can occur. | |
825 | * | |
826 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to | |
827 | * allow for the possibility of usermode upcalls messing up our count | |
828 | * of interrupt nesting level during the busy period that is just | |
829 | * now starting. | |
830 | */ | |
831 | void rcu_idle_exit(void) | |
832 | { | |
c5d900bf FW |
833 | unsigned long flags; |
834 | ||
835 | local_irq_save(flags); | |
cb349ca9 | 836 | rcu_eqs_exit(false); |
28ced795 | 837 | rcu_sysidle_exit(0); |
c5d900bf | 838 | local_irq_restore(flags); |
adf5091e | 839 | } |
8a2ecf47 | 840 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 841 | |
d1ec4c34 | 842 | #ifdef CONFIG_NO_HZ_FULL |
adf5091e FW |
843 | /** |
844 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
845 | * | |
846 | * Exit RCU idle mode while entering the kernel because it can | |
847 | * run a RCU read side critical section anytime. | |
848 | */ | |
849 | void rcu_user_exit(void) | |
850 | { | |
91d1aa43 | 851 | rcu_eqs_exit(1); |
adf5091e | 852 | } |
d1ec4c34 | 853 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 854 | |
9b2e4f18 PM |
855 | /** |
856 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle | |
857 | * | |
858 | * Enter an interrupt handler, which might possibly result in exiting | |
859 | * idle mode, in other words, entering the mode in which read-side critical | |
860 | * sections can occur. | |
861 | * | |
862 | * Note that the Linux kernel is fully capable of entering an interrupt | |
863 | * handler that it never exits, for example when doing upcalls to | |
864 | * user mode! This code assumes that the idle loop never does upcalls to | |
865 | * user mode. If your architecture does do upcalls from the idle loop (or | |
866 | * does anything else that results in unbalanced calls to the irq_enter() | |
867 | * and irq_exit() functions), RCU will give you what you deserve, good | |
868 | * and hard. But very infrequently and irreproducibly. | |
869 | * | |
870 | * Use things like work queues to work around this limitation. | |
871 | * | |
872 | * You have been warned. | |
873 | */ | |
874 | void rcu_irq_enter(void) | |
875 | { | |
876 | unsigned long flags; | |
877 | struct rcu_dynticks *rdtp; | |
878 | long long oldval; | |
879 | ||
880 | local_irq_save(flags); | |
c9d4b0af | 881 | rdtp = this_cpu_ptr(&rcu_dynticks); |
9b2e4f18 PM |
882 | oldval = rdtp->dynticks_nesting; |
883 | rdtp->dynticks_nesting++; | |
1ce46ee5 PM |
884 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
885 | rdtp->dynticks_nesting == 0); | |
b6fc6020 | 886 | if (oldval) |
f7f7bac9 | 887 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
b6fc6020 | 888 | else |
28ced795 CL |
889 | rcu_eqs_exit_common(oldval, true); |
890 | rcu_sysidle_exit(1); | |
64db4cff | 891 | local_irq_restore(flags); |
64db4cff PM |
892 | } |
893 | ||
894 | /** | |
895 | * rcu_nmi_enter - inform RCU of entry to NMI context | |
896 | * | |
734d1680 PM |
897 | * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and |
898 | * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know | |
899 | * that the CPU is active. This implementation permits nested NMIs, as | |
900 | * long as the nesting level does not overflow an int. (You will probably | |
901 | * run out of stack space first.) | |
64db4cff PM |
902 | */ |
903 | void rcu_nmi_enter(void) | |
904 | { | |
c9d4b0af | 905 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
734d1680 | 906 | int incby = 2; |
64db4cff | 907 | |
734d1680 PM |
908 | /* Complain about underflow. */ |
909 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0); | |
910 | ||
911 | /* | |
912 | * If idle from RCU viewpoint, atomically increment ->dynticks | |
913 | * to mark non-idle and increment ->dynticks_nmi_nesting by one. | |
914 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means | |
915 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed | |
916 | * to be in the outermost NMI handler that interrupted an RCU-idle | |
917 | * period (observation due to Andy Lutomirski). | |
918 | */ | |
919 | if (!(atomic_read(&rdtp->dynticks) & 0x1)) { | |
920 | smp_mb__before_atomic(); /* Force delay from prior write. */ | |
921 | atomic_inc(&rdtp->dynticks); | |
922 | /* atomic_inc() before later RCU read-side crit sects */ | |
923 | smp_mb__after_atomic(); /* See above. */ | |
924 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
925 | incby = 1; | |
926 | } | |
927 | rdtp->dynticks_nmi_nesting += incby; | |
928 | barrier(); | |
64db4cff PM |
929 | } |
930 | ||
931 | /** | |
932 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
933 | * | |
734d1680 PM |
934 | * If we are returning from the outermost NMI handler that interrupted an |
935 | * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting | |
936 | * to let the RCU grace-period handling know that the CPU is back to | |
937 | * being RCU-idle. | |
64db4cff PM |
938 | */ |
939 | void rcu_nmi_exit(void) | |
940 | { | |
c9d4b0af | 941 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
64db4cff | 942 | |
734d1680 PM |
943 | /* |
944 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. | |
945 | * (We are exiting an NMI handler, so RCU better be paying attention | |
946 | * to us!) | |
947 | */ | |
948 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0); | |
949 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); | |
950 | ||
951 | /* | |
952 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so | |
953 | * leave it in non-RCU-idle state. | |
954 | */ | |
955 | if (rdtp->dynticks_nmi_nesting != 1) { | |
956 | rdtp->dynticks_nmi_nesting -= 2; | |
64db4cff | 957 | return; |
734d1680 PM |
958 | } |
959 | ||
960 | /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ | |
961 | rdtp->dynticks_nmi_nesting = 0; | |
23b5c8fa | 962 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
4e857c58 | 963 | smp_mb__before_atomic(); /* See above. */ |
23b5c8fa | 964 | atomic_inc(&rdtp->dynticks); |
4e857c58 | 965 | smp_mb__after_atomic(); /* Force delay to next write. */ |
23b5c8fa | 966 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
64db4cff PM |
967 | } |
968 | ||
969 | /** | |
5c173eb8 PM |
970 | * __rcu_is_watching - are RCU read-side critical sections safe? |
971 | * | |
972 | * Return true if RCU is watching the running CPU, which means that | |
973 | * this CPU can safely enter RCU read-side critical sections. Unlike | |
974 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at | |
975 | * least disabled preemption. | |
976 | */ | |
9418fb20 | 977 | bool notrace __rcu_is_watching(void) |
5c173eb8 PM |
978 | { |
979 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; | |
980 | } | |
981 | ||
982 | /** | |
983 | * rcu_is_watching - see if RCU thinks that the current CPU is idle | |
64db4cff | 984 | * |
9b2e4f18 | 985 | * If the current CPU is in its idle loop and is neither in an interrupt |
34240697 | 986 | * or NMI handler, return true. |
64db4cff | 987 | */ |
9418fb20 | 988 | bool notrace rcu_is_watching(void) |
64db4cff | 989 | { |
f534ed1f | 990 | bool ret; |
34240697 | 991 | |
46f00d18 | 992 | preempt_disable_notrace(); |
5c173eb8 | 993 | ret = __rcu_is_watching(); |
46f00d18 | 994 | preempt_enable_notrace(); |
34240697 | 995 | return ret; |
64db4cff | 996 | } |
5c173eb8 | 997 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 998 | |
62fde6ed | 999 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
1000 | |
1001 | /* | |
1002 | * Is the current CPU online? Disable preemption to avoid false positives | |
1003 | * that could otherwise happen due to the current CPU number being sampled, | |
1004 | * this task being preempted, its old CPU being taken offline, resuming | |
1005 | * on some other CPU, then determining that its old CPU is now offline. | |
1006 | * It is OK to use RCU on an offline processor during initial boot, hence | |
2036d94a PM |
1007 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
1008 | * for a CPU coming online to use RCU for one jiffy prior to marking itself | |
1009 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going | |
1010 | * offline to continue to use RCU for one jiffy after marking itself | |
1011 | * offline in the cpu_online_mask. This leniency is necessary given the | |
1012 | * non-atomic nature of the online and offline processing, for example, | |
1013 | * the fact that a CPU enters the scheduler after completing the CPU_DYING | |
1014 | * notifiers. | |
1015 | * | |
1016 | * This is also why RCU internally marks CPUs online during the | |
1017 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. | |
c0d6d01b PM |
1018 | * |
1019 | * Disable checking if in an NMI handler because we cannot safely report | |
1020 | * errors from NMI handlers anyway. | |
1021 | */ | |
1022 | bool rcu_lockdep_current_cpu_online(void) | |
1023 | { | |
2036d94a PM |
1024 | struct rcu_data *rdp; |
1025 | struct rcu_node *rnp; | |
c0d6d01b PM |
1026 | bool ret; |
1027 | ||
1028 | if (in_nmi()) | |
f6f7ee9a | 1029 | return true; |
c0d6d01b | 1030 | preempt_disable(); |
c9d4b0af | 1031 | rdp = this_cpu_ptr(&rcu_sched_data); |
2036d94a | 1032 | rnp = rdp->mynode; |
0aa04b05 | 1033 | ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) || |
c0d6d01b PM |
1034 | !rcu_scheduler_fully_active; |
1035 | preempt_enable(); | |
1036 | return ret; | |
1037 | } | |
1038 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
1039 | ||
62fde6ed | 1040 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 1041 | |
64db4cff | 1042 | /** |
9b2e4f18 | 1043 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
64db4cff | 1044 | * |
9b2e4f18 PM |
1045 | * If the current CPU is idle or running at a first-level (not nested) |
1046 | * interrupt from idle, return true. The caller must have at least | |
1047 | * disabled preemption. | |
64db4cff | 1048 | */ |
62e3cb14 | 1049 | static int rcu_is_cpu_rrupt_from_idle(void) |
64db4cff | 1050 | { |
c9d4b0af | 1051 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
64db4cff PM |
1052 | } |
1053 | ||
64db4cff PM |
1054 | /* |
1055 | * Snapshot the specified CPU's dynticks counter so that we can later | |
1056 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 1057 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 1058 | */ |
217af2a2 PM |
1059 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
1060 | bool *isidle, unsigned long *maxj) | |
64db4cff | 1061 | { |
23b5c8fa | 1062 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
0edd1b17 | 1063 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
7941dbde ACB |
1064 | if ((rdp->dynticks_snap & 0x1) == 0) { |
1065 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); | |
1066 | return 1; | |
1067 | } else { | |
7d0ae808 | 1068 | if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, |
e3663b10 | 1069 | rdp->mynode->gpnum)) |
7d0ae808 | 1070 | WRITE_ONCE(rdp->gpwrap, true); |
7941dbde ACB |
1071 | return 0; |
1072 | } | |
64db4cff PM |
1073 | } |
1074 | ||
1075 | /* | |
1076 | * Return true if the specified CPU has passed through a quiescent | |
1077 | * state by virtue of being in or having passed through an dynticks | |
1078 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 1079 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 1080 | */ |
217af2a2 PM |
1081 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
1082 | bool *isidle, unsigned long *maxj) | |
64db4cff | 1083 | { |
7eb4f455 | 1084 | unsigned int curr; |
4a81e832 | 1085 | int *rcrmp; |
7eb4f455 | 1086 | unsigned int snap; |
64db4cff | 1087 | |
7eb4f455 PM |
1088 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
1089 | snap = (unsigned int)rdp->dynticks_snap; | |
64db4cff PM |
1090 | |
1091 | /* | |
1092 | * If the CPU passed through or entered a dynticks idle phase with | |
1093 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
1094 | * already acknowledged the request to pass through a quiescent | |
1095 | * state. Either way, that CPU cannot possibly be in an RCU | |
1096 | * read-side critical section that started before the beginning | |
1097 | * of the current RCU grace period. | |
1098 | */ | |
7eb4f455 | 1099 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
f7f7bac9 | 1100 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
64db4cff PM |
1101 | rdp->dynticks_fqs++; |
1102 | return 1; | |
1103 | } | |
1104 | ||
a82dcc76 PM |
1105 | /* |
1106 | * Check for the CPU being offline, but only if the grace period | |
1107 | * is old enough. We don't need to worry about the CPU changing | |
1108 | * state: If we see it offline even once, it has been through a | |
1109 | * quiescent state. | |
1110 | * | |
1111 | * The reason for insisting that the grace period be at least | |
1112 | * one jiffy old is that CPUs that are not quite online and that | |
1113 | * have just gone offline can still execute RCU read-side critical | |
1114 | * sections. | |
1115 | */ | |
1116 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) | |
1117 | return 0; /* Grace period is not old enough. */ | |
1118 | barrier(); | |
1119 | if (cpu_is_offline(rdp->cpu)) { | |
f7f7bac9 | 1120 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
a82dcc76 PM |
1121 | rdp->offline_fqs++; |
1122 | return 1; | |
1123 | } | |
65d798f0 PM |
1124 | |
1125 | /* | |
4a81e832 PM |
1126 | * A CPU running for an extended time within the kernel can |
1127 | * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode, | |
1128 | * even context-switching back and forth between a pair of | |
1129 | * in-kernel CPU-bound tasks cannot advance grace periods. | |
1130 | * So if the grace period is old enough, make the CPU pay attention. | |
1131 | * Note that the unsynchronized assignments to the per-CPU | |
1132 | * rcu_sched_qs_mask variable are safe. Yes, setting of | |
1133 | * bits can be lost, but they will be set again on the next | |
1134 | * force-quiescent-state pass. So lost bit sets do not result | |
1135 | * in incorrect behavior, merely in a grace period lasting | |
1136 | * a few jiffies longer than it might otherwise. Because | |
1137 | * there are at most four threads involved, and because the | |
1138 | * updates are only once every few jiffies, the probability of | |
1139 | * lossage (and thus of slight grace-period extension) is | |
1140 | * quite low. | |
1141 | * | |
1142 | * Note that if the jiffies_till_sched_qs boot/sysfs parameter | |
1143 | * is set too high, we override with half of the RCU CPU stall | |
1144 | * warning delay. | |
6193c76a | 1145 | */ |
4a81e832 PM |
1146 | rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu); |
1147 | if (ULONG_CMP_GE(jiffies, | |
1148 | rdp->rsp->gp_start + jiffies_till_sched_qs) || | |
cb1e78cf | 1149 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { |
7d0ae808 PM |
1150 | if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) { |
1151 | WRITE_ONCE(rdp->cond_resched_completed, | |
1152 | READ_ONCE(rdp->mynode->completed)); | |
4a81e832 | 1153 | smp_mb(); /* ->cond_resched_completed before *rcrmp. */ |
7d0ae808 PM |
1154 | WRITE_ONCE(*rcrmp, |
1155 | READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask); | |
4a81e832 PM |
1156 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ |
1157 | rdp->rsp->jiffies_resched += 5; /* Enable beating. */ | |
1158 | } else if (ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { | |
1159 | /* Time to beat on that CPU again! */ | |
1160 | resched_cpu(rdp->cpu); /* Force CPU into scheduler. */ | |
1161 | rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */ | |
1162 | } | |
6193c76a PM |
1163 | } |
1164 | ||
a82dcc76 | 1165 | return 0; |
64db4cff PM |
1166 | } |
1167 | ||
64db4cff PM |
1168 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
1169 | { | |
cb1e78cf | 1170 | unsigned long j = jiffies; |
6193c76a | 1171 | unsigned long j1; |
26cdfedf PM |
1172 | |
1173 | rsp->gp_start = j; | |
1174 | smp_wmb(); /* Record start time before stall time. */ | |
6193c76a | 1175 | j1 = rcu_jiffies_till_stall_check(); |
7d0ae808 | 1176 | WRITE_ONCE(rsp->jiffies_stall, j + j1); |
6193c76a | 1177 | rsp->jiffies_resched = j + j1 / 2; |
7d0ae808 | 1178 | rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs); |
64db4cff PM |
1179 | } |
1180 | ||
fb81a44b PM |
1181 | /* |
1182 | * Complain about starvation of grace-period kthread. | |
1183 | */ | |
1184 | static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp) | |
1185 | { | |
1186 | unsigned long gpa; | |
1187 | unsigned long j; | |
1188 | ||
1189 | j = jiffies; | |
7d0ae808 | 1190 | gpa = READ_ONCE(rsp->gp_activity); |
fb81a44b | 1191 | if (j - gpa > 2 * HZ) |
319362c9 | 1192 | pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x s%d ->state=%#lx\n", |
81e701e4 | 1193 | rsp->name, j - gpa, |
319362c9 PM |
1194 | rsp->gpnum, rsp->completed, |
1195 | rsp->gp_flags, rsp->gp_state, | |
1196 | rsp->gp_kthread ? rsp->gp_kthread->state : 0); | |
64db4cff PM |
1197 | } |
1198 | ||
b637a328 | 1199 | /* |
bc1dce51 | 1200 | * Dump stacks of all tasks running on stalled CPUs. |
b637a328 PM |
1201 | */ |
1202 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) | |
1203 | { | |
1204 | int cpu; | |
1205 | unsigned long flags; | |
1206 | struct rcu_node *rnp; | |
1207 | ||
1208 | rcu_for_each_leaf_node(rsp, rnp) { | |
6cf10081 | 1209 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
b637a328 PM |
1210 | if (rnp->qsmask != 0) { |
1211 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1212 | if (rnp->qsmask & (1UL << cpu)) | |
1213 | dump_cpu_task(rnp->grplo + cpu); | |
1214 | } | |
1215 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1216 | } | |
1217 | } | |
1218 | ||
6ccd2ecd | 1219 | static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum) |
64db4cff PM |
1220 | { |
1221 | int cpu; | |
1222 | long delta; | |
1223 | unsigned long flags; | |
6ccd2ecd PM |
1224 | unsigned long gpa; |
1225 | unsigned long j; | |
285fe294 | 1226 | int ndetected = 0; |
64db4cff | 1227 | struct rcu_node *rnp = rcu_get_root(rsp); |
53bb857c | 1228 | long totqlen = 0; |
64db4cff PM |
1229 | |
1230 | /* Only let one CPU complain about others per time interval. */ | |
1231 | ||
6cf10081 | 1232 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
7d0ae808 | 1233 | delta = jiffies - READ_ONCE(rsp->jiffies_stall); |
fc2219d4 | 1234 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
1304afb2 | 1235 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
1236 | return; |
1237 | } | |
7d0ae808 PM |
1238 | WRITE_ONCE(rsp->jiffies_stall, |
1239 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3); | |
1304afb2 | 1240 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1241 | |
8cdd32a9 PM |
1242 | /* |
1243 | * OK, time to rat on our buddy... | |
1244 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1245 | * RCU CPU stall warnings. | |
1246 | */ | |
d7f3e207 | 1247 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
4300aa64 | 1248 | rsp->name); |
a858af28 | 1249 | print_cpu_stall_info_begin(); |
a0b6c9a7 | 1250 | rcu_for_each_leaf_node(rsp, rnp) { |
6cf10081 | 1251 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
9bc8b558 | 1252 | ndetected += rcu_print_task_stall(rnp); |
c8020a67 PM |
1253 | if (rnp->qsmask != 0) { |
1254 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) | |
1255 | if (rnp->qsmask & (1UL << cpu)) { | |
1256 | print_cpu_stall_info(rsp, | |
1257 | rnp->grplo + cpu); | |
1258 | ndetected++; | |
1259 | } | |
1260 | } | |
3acd9eb3 | 1261 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 1262 | } |
a858af28 | 1263 | |
a858af28 | 1264 | print_cpu_stall_info_end(); |
53bb857c PM |
1265 | for_each_possible_cpu(cpu) |
1266 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e | 1267 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", |
eee05882 | 1268 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
83ebe63e | 1269 | (long)rsp->gpnum, (long)rsp->completed, totqlen); |
6ccd2ecd | 1270 | if (ndetected) { |
b637a328 | 1271 | rcu_dump_cpu_stacks(rsp); |
6ccd2ecd | 1272 | } else { |
7d0ae808 PM |
1273 | if (READ_ONCE(rsp->gpnum) != gpnum || |
1274 | READ_ONCE(rsp->completed) == gpnum) { | |
6ccd2ecd PM |
1275 | pr_err("INFO: Stall ended before state dump start\n"); |
1276 | } else { | |
1277 | j = jiffies; | |
7d0ae808 | 1278 | gpa = READ_ONCE(rsp->gp_activity); |
237a0f21 | 1279 | pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n", |
6ccd2ecd | 1280 | rsp->name, j - gpa, j, gpa, |
237a0f21 PM |
1281 | jiffies_till_next_fqs, |
1282 | rcu_get_root(rsp)->qsmask); | |
6ccd2ecd PM |
1283 | /* In this case, the current CPU might be at fault. */ |
1284 | sched_show_task(current); | |
1285 | } | |
1286 | } | |
c1dc0b9c | 1287 | |
4cdfc175 | 1288 | /* Complain about tasks blocking the grace period. */ |
1ed509a2 PM |
1289 | rcu_print_detail_task_stall(rsp); |
1290 | ||
fb81a44b PM |
1291 | rcu_check_gp_kthread_starvation(rsp); |
1292 | ||
4cdfc175 | 1293 | force_quiescent_state(rsp); /* Kick them all. */ |
64db4cff PM |
1294 | } |
1295 | ||
1296 | static void print_cpu_stall(struct rcu_state *rsp) | |
1297 | { | |
53bb857c | 1298 | int cpu; |
64db4cff PM |
1299 | unsigned long flags; |
1300 | struct rcu_node *rnp = rcu_get_root(rsp); | |
53bb857c | 1301 | long totqlen = 0; |
64db4cff | 1302 | |
8cdd32a9 PM |
1303 | /* |
1304 | * OK, time to rat on ourselves... | |
1305 | * See Documentation/RCU/stallwarn.txt for info on how to debug | |
1306 | * RCU CPU stall warnings. | |
1307 | */ | |
d7f3e207 | 1308 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
a858af28 PM |
1309 | print_cpu_stall_info_begin(); |
1310 | print_cpu_stall_info(rsp, smp_processor_id()); | |
1311 | print_cpu_stall_info_end(); | |
53bb857c PM |
1312 | for_each_possible_cpu(cpu) |
1313 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; | |
83ebe63e PM |
1314 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", |
1315 | jiffies - rsp->gp_start, | |
1316 | (long)rsp->gpnum, (long)rsp->completed, totqlen); | |
fb81a44b PM |
1317 | |
1318 | rcu_check_gp_kthread_starvation(rsp); | |
1319 | ||
bc1dce51 | 1320 | rcu_dump_cpu_stacks(rsp); |
c1dc0b9c | 1321 | |
6cf10081 | 1322 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
7d0ae808 PM |
1323 | if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall))) |
1324 | WRITE_ONCE(rsp->jiffies_stall, | |
1325 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3); | |
1304afb2 | 1326 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c1dc0b9c | 1327 | |
b021fe3e PZ |
1328 | /* |
1329 | * Attempt to revive the RCU machinery by forcing a context switch. | |
1330 | * | |
1331 | * A context switch would normally allow the RCU state machine to make | |
1332 | * progress and it could be we're stuck in kernel space without context | |
1333 | * switches for an entirely unreasonable amount of time. | |
1334 | */ | |
1335 | resched_cpu(smp_processor_id()); | |
64db4cff PM |
1336 | } |
1337 | ||
1338 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) | |
1339 | { | |
26cdfedf PM |
1340 | unsigned long completed; |
1341 | unsigned long gpnum; | |
1342 | unsigned long gps; | |
bad6e139 PM |
1343 | unsigned long j; |
1344 | unsigned long js; | |
64db4cff PM |
1345 | struct rcu_node *rnp; |
1346 | ||
26cdfedf | 1347 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
c68de209 | 1348 | return; |
cb1e78cf | 1349 | j = jiffies; |
26cdfedf PM |
1350 | |
1351 | /* | |
1352 | * Lots of memory barriers to reject false positives. | |
1353 | * | |
1354 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, | |
1355 | * then rsp->gp_start, and finally rsp->completed. These values | |
1356 | * are updated in the opposite order with memory barriers (or | |
1357 | * equivalent) during grace-period initialization and cleanup. | |
1358 | * Now, a false positive can occur if we get an new value of | |
1359 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given | |
1360 | * the memory barriers, the only way that this can happen is if one | |
1361 | * grace period ends and another starts between these two fetches. | |
1362 | * Detect this by comparing rsp->completed with the previous fetch | |
1363 | * from rsp->gpnum. | |
1364 | * | |
1365 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, | |
1366 | * and rsp->gp_start suffice to forestall false positives. | |
1367 | */ | |
7d0ae808 | 1368 | gpnum = READ_ONCE(rsp->gpnum); |
26cdfedf | 1369 | smp_rmb(); /* Pick up ->gpnum first... */ |
7d0ae808 | 1370 | js = READ_ONCE(rsp->jiffies_stall); |
26cdfedf | 1371 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
7d0ae808 | 1372 | gps = READ_ONCE(rsp->gp_start); |
26cdfedf | 1373 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ |
7d0ae808 | 1374 | completed = READ_ONCE(rsp->completed); |
26cdfedf PM |
1375 | if (ULONG_CMP_GE(completed, gpnum) || |
1376 | ULONG_CMP_LT(j, js) || | |
1377 | ULONG_CMP_GE(gps, js)) | |
1378 | return; /* No stall or GP completed since entering function. */ | |
64db4cff | 1379 | rnp = rdp->mynode; |
c96ea7cf | 1380 | if (rcu_gp_in_progress(rsp) && |
7d0ae808 | 1381 | (READ_ONCE(rnp->qsmask) & rdp->grpmask)) { |
64db4cff PM |
1382 | |
1383 | /* We haven't checked in, so go dump stack. */ | |
1384 | print_cpu_stall(rsp); | |
1385 | ||
bad6e139 PM |
1386 | } else if (rcu_gp_in_progress(rsp) && |
1387 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { | |
64db4cff | 1388 | |
bad6e139 | 1389 | /* They had a few time units to dump stack, so complain. */ |
6ccd2ecd | 1390 | print_other_cpu_stall(rsp, gpnum); |
64db4cff PM |
1391 | } |
1392 | } | |
1393 | ||
53d84e00 PM |
1394 | /** |
1395 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period | |
1396 | * | |
1397 | * Set the stall-warning timeout way off into the future, thus preventing | |
1398 | * any RCU CPU stall-warning messages from appearing in the current set of | |
1399 | * RCU grace periods. | |
1400 | * | |
1401 | * The caller must disable hard irqs. | |
1402 | */ | |
1403 | void rcu_cpu_stall_reset(void) | |
1404 | { | |
6ce75a23 PM |
1405 | struct rcu_state *rsp; |
1406 | ||
1407 | for_each_rcu_flavor(rsp) | |
7d0ae808 | 1408 | WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2); |
53d84e00 PM |
1409 | } |
1410 | ||
3f5d3ea6 | 1411 | /* |
d3f3f3f2 PM |
1412 | * Initialize the specified rcu_data structure's default callback list |
1413 | * to empty. The default callback list is the one that is not used by | |
1414 | * no-callbacks CPUs. | |
3f5d3ea6 | 1415 | */ |
d3f3f3f2 | 1416 | static void init_default_callback_list(struct rcu_data *rdp) |
3f5d3ea6 PM |
1417 | { |
1418 | int i; | |
1419 | ||
1420 | rdp->nxtlist = NULL; | |
1421 | for (i = 0; i < RCU_NEXT_SIZE; i++) | |
1422 | rdp->nxttail[i] = &rdp->nxtlist; | |
1423 | } | |
1424 | ||
d3f3f3f2 PM |
1425 | /* |
1426 | * Initialize the specified rcu_data structure's callback list to empty. | |
1427 | */ | |
1428 | static void init_callback_list(struct rcu_data *rdp) | |
1429 | { | |
1430 | if (init_nocb_callback_list(rdp)) | |
1431 | return; | |
1432 | init_default_callback_list(rdp); | |
1433 | } | |
1434 | ||
dc35c893 PM |
1435 | /* |
1436 | * Determine the value that ->completed will have at the end of the | |
1437 | * next subsequent grace period. This is used to tag callbacks so that | |
1438 | * a CPU can invoke callbacks in a timely fashion even if that CPU has | |
1439 | * been dyntick-idle for an extended period with callbacks under the | |
1440 | * influence of RCU_FAST_NO_HZ. | |
1441 | * | |
1442 | * The caller must hold rnp->lock with interrupts disabled. | |
1443 | */ | |
1444 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, | |
1445 | struct rcu_node *rnp) | |
1446 | { | |
1447 | /* | |
1448 | * If RCU is idle, we just wait for the next grace period. | |
1449 | * But we can only be sure that RCU is idle if we are looking | |
1450 | * at the root rcu_node structure -- otherwise, a new grace | |
1451 | * period might have started, but just not yet gotten around | |
1452 | * to initializing the current non-root rcu_node structure. | |
1453 | */ | |
1454 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) | |
1455 | return rnp->completed + 1; | |
1456 | ||
1457 | /* | |
1458 | * Otherwise, wait for a possible partial grace period and | |
1459 | * then the subsequent full grace period. | |
1460 | */ | |
1461 | return rnp->completed + 2; | |
1462 | } | |
1463 | ||
0446be48 PM |
1464 | /* |
1465 | * Trace-event helper function for rcu_start_future_gp() and | |
1466 | * rcu_nocb_wait_gp(). | |
1467 | */ | |
1468 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
e66c33d5 | 1469 | unsigned long c, const char *s) |
0446be48 PM |
1470 | { |
1471 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, | |
1472 | rnp->completed, c, rnp->level, | |
1473 | rnp->grplo, rnp->grphi, s); | |
1474 | } | |
1475 | ||
1476 | /* | |
1477 | * Start some future grace period, as needed to handle newly arrived | |
1478 | * callbacks. The required future grace periods are recorded in each | |
48a7639c PM |
1479 | * rcu_node structure's ->need_future_gp field. Returns true if there |
1480 | * is reason to awaken the grace-period kthread. | |
0446be48 PM |
1481 | * |
1482 | * The caller must hold the specified rcu_node structure's ->lock. | |
1483 | */ | |
48a7639c PM |
1484 | static bool __maybe_unused |
1485 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
1486 | unsigned long *c_out) | |
0446be48 PM |
1487 | { |
1488 | unsigned long c; | |
1489 | int i; | |
48a7639c | 1490 | bool ret = false; |
0446be48 PM |
1491 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); |
1492 | ||
1493 | /* | |
1494 | * Pick up grace-period number for new callbacks. If this | |
1495 | * grace period is already marked as needed, return to the caller. | |
1496 | */ | |
1497 | c = rcu_cbs_completed(rdp->rsp, rnp); | |
f7f7bac9 | 1498 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
0446be48 | 1499 | if (rnp->need_future_gp[c & 0x1]) { |
f7f7bac9 | 1500 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
48a7639c | 1501 | goto out; |
0446be48 PM |
1502 | } |
1503 | ||
1504 | /* | |
1505 | * If either this rcu_node structure or the root rcu_node structure | |
1506 | * believe that a grace period is in progress, then we must wait | |
1507 | * for the one following, which is in "c". Because our request | |
1508 | * will be noticed at the end of the current grace period, we don't | |
48bd8e9b PK |
1509 | * need to explicitly start one. We only do the lockless check |
1510 | * of rnp_root's fields if the current rcu_node structure thinks | |
1511 | * there is no grace period in flight, and because we hold rnp->lock, | |
1512 | * the only possible change is when rnp_root's two fields are | |
1513 | * equal, in which case rnp_root->gpnum might be concurrently | |
1514 | * incremented. But that is OK, as it will just result in our | |
1515 | * doing some extra useless work. | |
0446be48 PM |
1516 | */ |
1517 | if (rnp->gpnum != rnp->completed || | |
7d0ae808 | 1518 | READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) { |
0446be48 | 1519 | rnp->need_future_gp[c & 0x1]++; |
f7f7bac9 | 1520 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
48a7639c | 1521 | goto out; |
0446be48 PM |
1522 | } |
1523 | ||
1524 | /* | |
1525 | * There might be no grace period in progress. If we don't already | |
1526 | * hold it, acquire the root rcu_node structure's lock in order to | |
1527 | * start one (if needed). | |
1528 | */ | |
2a67e741 PZ |
1529 | if (rnp != rnp_root) |
1530 | raw_spin_lock_rcu_node(rnp_root); | |
0446be48 PM |
1531 | |
1532 | /* | |
1533 | * Get a new grace-period number. If there really is no grace | |
1534 | * period in progress, it will be smaller than the one we obtained | |
1535 | * earlier. Adjust callbacks as needed. Note that even no-CBs | |
1536 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. | |
1537 | */ | |
1538 | c = rcu_cbs_completed(rdp->rsp, rnp_root); | |
1539 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) | |
1540 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) | |
1541 | rdp->nxtcompleted[i] = c; | |
1542 | ||
1543 | /* | |
1544 | * If the needed for the required grace period is already | |
1545 | * recorded, trace and leave. | |
1546 | */ | |
1547 | if (rnp_root->need_future_gp[c & 0x1]) { | |
f7f7bac9 | 1548 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
0446be48 PM |
1549 | goto unlock_out; |
1550 | } | |
1551 | ||
1552 | /* Record the need for the future grace period. */ | |
1553 | rnp_root->need_future_gp[c & 0x1]++; | |
1554 | ||
1555 | /* If a grace period is not already in progress, start one. */ | |
1556 | if (rnp_root->gpnum != rnp_root->completed) { | |
f7f7bac9 | 1557 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
0446be48 | 1558 | } else { |
f7f7bac9 | 1559 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
48a7639c | 1560 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
0446be48 PM |
1561 | } |
1562 | unlock_out: | |
1563 | if (rnp != rnp_root) | |
1564 | raw_spin_unlock(&rnp_root->lock); | |
48a7639c PM |
1565 | out: |
1566 | if (c_out != NULL) | |
1567 | *c_out = c; | |
1568 | return ret; | |
0446be48 PM |
1569 | } |
1570 | ||
1571 | /* | |
1572 | * Clean up any old requests for the just-ended grace period. Also return | |
1573 | * whether any additional grace periods have been requested. Also invoke | |
1574 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads | |
1575 | * waiting for this grace period to complete. | |
1576 | */ | |
1577 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) | |
1578 | { | |
1579 | int c = rnp->completed; | |
1580 | int needmore; | |
1581 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
1582 | ||
1583 | rcu_nocb_gp_cleanup(rsp, rnp); | |
1584 | rnp->need_future_gp[c & 0x1] = 0; | |
1585 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; | |
f7f7bac9 SRRH |
1586 | trace_rcu_future_gp(rnp, rdp, c, |
1587 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); | |
0446be48 PM |
1588 | return needmore; |
1589 | } | |
1590 | ||
48a7639c PM |
1591 | /* |
1592 | * Awaken the grace-period kthread for the specified flavor of RCU. | |
1593 | * Don't do a self-awaken, and don't bother awakening when there is | |
1594 | * nothing for the grace-period kthread to do (as in several CPUs | |
1595 | * raced to awaken, and we lost), and finally don't try to awaken | |
1596 | * a kthread that has not yet been created. | |
1597 | */ | |
1598 | static void rcu_gp_kthread_wake(struct rcu_state *rsp) | |
1599 | { | |
1600 | if (current == rsp->gp_kthread || | |
7d0ae808 | 1601 | !READ_ONCE(rsp->gp_flags) || |
48a7639c PM |
1602 | !rsp->gp_kthread) |
1603 | return; | |
1604 | wake_up(&rsp->gp_wq); | |
1605 | } | |
1606 | ||
dc35c893 PM |
1607 | /* |
1608 | * If there is room, assign a ->completed number to any callbacks on | |
1609 | * this CPU that have not already been assigned. Also accelerate any | |
1610 | * callbacks that were previously assigned a ->completed number that has | |
1611 | * since proven to be too conservative, which can happen if callbacks get | |
1612 | * assigned a ->completed number while RCU is idle, but with reference to | |
1613 | * a non-root rcu_node structure. This function is idempotent, so it does | |
48a7639c PM |
1614 | * not hurt to call it repeatedly. Returns an flag saying that we should |
1615 | * awaken the RCU grace-period kthread. | |
dc35c893 PM |
1616 | * |
1617 | * The caller must hold rnp->lock with interrupts disabled. | |
1618 | */ | |
48a7639c | 1619 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1620 | struct rcu_data *rdp) |
1621 | { | |
1622 | unsigned long c; | |
1623 | int i; | |
48a7639c | 1624 | bool ret; |
dc35c893 PM |
1625 | |
1626 | /* If the CPU has no callbacks, nothing to do. */ | |
1627 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1628 | return false; |
dc35c893 PM |
1629 | |
1630 | /* | |
1631 | * Starting from the sublist containing the callbacks most | |
1632 | * recently assigned a ->completed number and working down, find the | |
1633 | * first sublist that is not assignable to an upcoming grace period. | |
1634 | * Such a sublist has something in it (first two tests) and has | |
1635 | * a ->completed number assigned that will complete sooner than | |
1636 | * the ->completed number for newly arrived callbacks (last test). | |
1637 | * | |
1638 | * The key point is that any later sublist can be assigned the | |
1639 | * same ->completed number as the newly arrived callbacks, which | |
1640 | * means that the callbacks in any of these later sublist can be | |
1641 | * grouped into a single sublist, whether or not they have already | |
1642 | * been assigned a ->completed number. | |
1643 | */ | |
1644 | c = rcu_cbs_completed(rsp, rnp); | |
1645 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) | |
1646 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && | |
1647 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) | |
1648 | break; | |
1649 | ||
1650 | /* | |
1651 | * If there are no sublist for unassigned callbacks, leave. | |
1652 | * At the same time, advance "i" one sublist, so that "i" will | |
1653 | * index into the sublist where all the remaining callbacks should | |
1654 | * be grouped into. | |
1655 | */ | |
1656 | if (++i >= RCU_NEXT_TAIL) | |
48a7639c | 1657 | return false; |
dc35c893 PM |
1658 | |
1659 | /* | |
1660 | * Assign all subsequent callbacks' ->completed number to the next | |
1661 | * full grace period and group them all in the sublist initially | |
1662 | * indexed by "i". | |
1663 | */ | |
1664 | for (; i <= RCU_NEXT_TAIL; i++) { | |
1665 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; | |
1666 | rdp->nxtcompleted[i] = c; | |
1667 | } | |
910ee45d | 1668 | /* Record any needed additional grace periods. */ |
48a7639c | 1669 | ret = rcu_start_future_gp(rnp, rdp, NULL); |
6d4b418c PM |
1670 | |
1671 | /* Trace depending on how much we were able to accelerate. */ | |
1672 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) | |
f7f7bac9 | 1673 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
6d4b418c | 1674 | else |
f7f7bac9 | 1675 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
48a7639c | 1676 | return ret; |
dc35c893 PM |
1677 | } |
1678 | ||
1679 | /* | |
1680 | * Move any callbacks whose grace period has completed to the | |
1681 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
1682 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL | |
1683 | * sublist. This function is idempotent, so it does not hurt to | |
1684 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1685 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1686 | * |
1687 | * The caller must hold rnp->lock with interrupts disabled. | |
1688 | */ | |
48a7639c | 1689 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
dc35c893 PM |
1690 | struct rcu_data *rdp) |
1691 | { | |
1692 | int i, j; | |
1693 | ||
1694 | /* If the CPU has no callbacks, nothing to do. */ | |
1695 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) | |
48a7639c | 1696 | return false; |
dc35c893 PM |
1697 | |
1698 | /* | |
1699 | * Find all callbacks whose ->completed numbers indicate that they | |
1700 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. | |
1701 | */ | |
1702 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { | |
1703 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) | |
1704 | break; | |
1705 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; | |
1706 | } | |
1707 | /* Clean up any sublist tail pointers that were misordered above. */ | |
1708 | for (j = RCU_WAIT_TAIL; j < i; j++) | |
1709 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; | |
1710 | ||
1711 | /* Copy down callbacks to fill in empty sublists. */ | |
1712 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { | |
1713 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) | |
1714 | break; | |
1715 | rdp->nxttail[j] = rdp->nxttail[i]; | |
1716 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; | |
1717 | } | |
1718 | ||
1719 | /* Classify any remaining callbacks. */ | |
48a7639c | 1720 | return rcu_accelerate_cbs(rsp, rnp, rdp); |
dc35c893 PM |
1721 | } |
1722 | ||
d09b62df | 1723 | /* |
ba9fbe95 PM |
1724 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1725 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1726 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1727 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1728 | */ |
48a7639c PM |
1729 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, |
1730 | struct rcu_data *rdp) | |
d09b62df | 1731 | { |
48a7639c PM |
1732 | bool ret; |
1733 | ||
ba9fbe95 | 1734 | /* Handle the ends of any preceding grace periods first. */ |
e3663b10 | 1735 | if (rdp->completed == rnp->completed && |
7d0ae808 | 1736 | !unlikely(READ_ONCE(rdp->gpwrap))) { |
d09b62df | 1737 | |
ba9fbe95 | 1738 | /* No grace period end, so just accelerate recent callbacks. */ |
48a7639c | 1739 | ret = rcu_accelerate_cbs(rsp, rnp, rdp); |
d09b62df | 1740 | |
dc35c893 PM |
1741 | } else { |
1742 | ||
1743 | /* Advance callbacks. */ | |
48a7639c | 1744 | ret = rcu_advance_cbs(rsp, rnp, rdp); |
d09b62df PM |
1745 | |
1746 | /* Remember that we saw this grace-period completion. */ | |
1747 | rdp->completed = rnp->completed; | |
f7f7bac9 | 1748 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
d09b62df | 1749 | } |
398ebe60 | 1750 | |
7d0ae808 | 1751 | if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) { |
6eaef633 PM |
1752 | /* |
1753 | * If the current grace period is waiting for this CPU, | |
1754 | * set up to detect a quiescent state, otherwise don't | |
1755 | * go looking for one. | |
1756 | */ | |
1757 | rdp->gpnum = rnp->gpnum; | |
f7f7bac9 | 1758 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
5b74c458 | 1759 | rdp->cpu_no_qs.b.norm = true; |
5cd37193 | 1760 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); |
97c668b8 | 1761 | rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask); |
6eaef633 | 1762 | zero_cpu_stall_ticks(rdp); |
7d0ae808 | 1763 | WRITE_ONCE(rdp->gpwrap, false); |
6eaef633 | 1764 | } |
48a7639c | 1765 | return ret; |
6eaef633 PM |
1766 | } |
1767 | ||
d34ea322 | 1768 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
6eaef633 PM |
1769 | { |
1770 | unsigned long flags; | |
48a7639c | 1771 | bool needwake; |
6eaef633 PM |
1772 | struct rcu_node *rnp; |
1773 | ||
1774 | local_irq_save(flags); | |
1775 | rnp = rdp->mynode; | |
7d0ae808 PM |
1776 | if ((rdp->gpnum == READ_ONCE(rnp->gpnum) && |
1777 | rdp->completed == READ_ONCE(rnp->completed) && | |
1778 | !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */ | |
2a67e741 | 1779 | !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */ |
6eaef633 PM |
1780 | local_irq_restore(flags); |
1781 | return; | |
1782 | } | |
48a7639c | 1783 | needwake = __note_gp_changes(rsp, rnp, rdp); |
6eaef633 | 1784 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
48a7639c PM |
1785 | if (needwake) |
1786 | rcu_gp_kthread_wake(rsp); | |
6eaef633 PM |
1787 | } |
1788 | ||
0f41c0dd PM |
1789 | static void rcu_gp_slow(struct rcu_state *rsp, int delay) |
1790 | { | |
1791 | if (delay > 0 && | |
1792 | !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay))) | |
1793 | schedule_timeout_uninterruptible(delay); | |
1794 | } | |
1795 | ||
b3dbec76 | 1796 | /* |
f7be8209 | 1797 | * Initialize a new grace period. Return 0 if no grace period required. |
b3dbec76 | 1798 | */ |
7fdefc10 | 1799 | static int rcu_gp_init(struct rcu_state *rsp) |
b3dbec76 | 1800 | { |
0aa04b05 | 1801 | unsigned long oldmask; |
b3dbec76 | 1802 | struct rcu_data *rdp; |
7fdefc10 | 1803 | struct rcu_node *rnp = rcu_get_root(rsp); |
b3dbec76 | 1804 | |
7d0ae808 | 1805 | WRITE_ONCE(rsp->gp_activity, jiffies); |
2a67e741 | 1806 | raw_spin_lock_irq_rcu_node(rnp); |
7d0ae808 | 1807 | if (!READ_ONCE(rsp->gp_flags)) { |
f7be8209 PM |
1808 | /* Spurious wakeup, tell caller to go back to sleep. */ |
1809 | raw_spin_unlock_irq(&rnp->lock); | |
1810 | return 0; | |
1811 | } | |
7d0ae808 | 1812 | WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */ |
b3dbec76 | 1813 | |
f7be8209 PM |
1814 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
1815 | /* | |
1816 | * Grace period already in progress, don't start another. | |
1817 | * Not supposed to be able to happen. | |
1818 | */ | |
7fdefc10 PM |
1819 | raw_spin_unlock_irq(&rnp->lock); |
1820 | return 0; | |
1821 | } | |
1822 | ||
7fdefc10 | 1823 | /* Advance to a new grace period and initialize state. */ |
26cdfedf | 1824 | record_gp_stall_check_time(rsp); |
765a3f4f PM |
1825 | /* Record GP times before starting GP, hence smp_store_release(). */ |
1826 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); | |
f7f7bac9 | 1827 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
7fdefc10 PM |
1828 | raw_spin_unlock_irq(&rnp->lock); |
1829 | ||
0aa04b05 PM |
1830 | /* |
1831 | * Apply per-leaf buffered online and offline operations to the | |
1832 | * rcu_node tree. Note that this new grace period need not wait | |
1833 | * for subsequent online CPUs, and that quiescent-state forcing | |
1834 | * will handle subsequent offline CPUs. | |
1835 | */ | |
1836 | rcu_for_each_leaf_node(rsp, rnp) { | |
0f41c0dd | 1837 | rcu_gp_slow(rsp, gp_preinit_delay); |
2a67e741 | 1838 | raw_spin_lock_irq_rcu_node(rnp); |
0aa04b05 PM |
1839 | if (rnp->qsmaskinit == rnp->qsmaskinitnext && |
1840 | !rnp->wait_blkd_tasks) { | |
1841 | /* Nothing to do on this leaf rcu_node structure. */ | |
1842 | raw_spin_unlock_irq(&rnp->lock); | |
1843 | continue; | |
1844 | } | |
1845 | ||
1846 | /* Record old state, apply changes to ->qsmaskinit field. */ | |
1847 | oldmask = rnp->qsmaskinit; | |
1848 | rnp->qsmaskinit = rnp->qsmaskinitnext; | |
1849 | ||
1850 | /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ | |
1851 | if (!oldmask != !rnp->qsmaskinit) { | |
1852 | if (!oldmask) /* First online CPU for this rcu_node. */ | |
1853 | rcu_init_new_rnp(rnp); | |
1854 | else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */ | |
1855 | rnp->wait_blkd_tasks = true; | |
1856 | else /* Last offline CPU and can propagate. */ | |
1857 | rcu_cleanup_dead_rnp(rnp); | |
1858 | } | |
1859 | ||
1860 | /* | |
1861 | * If all waited-on tasks from prior grace period are | |
1862 | * done, and if all this rcu_node structure's CPUs are | |
1863 | * still offline, propagate up the rcu_node tree and | |
1864 | * clear ->wait_blkd_tasks. Otherwise, if one of this | |
1865 | * rcu_node structure's CPUs has since come back online, | |
1866 | * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp() | |
1867 | * checks for this, so just call it unconditionally). | |
1868 | */ | |
1869 | if (rnp->wait_blkd_tasks && | |
1870 | (!rcu_preempt_has_tasks(rnp) || | |
1871 | rnp->qsmaskinit)) { | |
1872 | rnp->wait_blkd_tasks = false; | |
1873 | rcu_cleanup_dead_rnp(rnp); | |
1874 | } | |
1875 | ||
1876 | raw_spin_unlock_irq(&rnp->lock); | |
1877 | } | |
7fdefc10 PM |
1878 | |
1879 | /* | |
1880 | * Set the quiescent-state-needed bits in all the rcu_node | |
1881 | * structures for all currently online CPUs in breadth-first order, | |
1882 | * starting from the root rcu_node structure, relying on the layout | |
1883 | * of the tree within the rsp->node[] array. Note that other CPUs | |
1884 | * will access only the leaves of the hierarchy, thus seeing that no | |
1885 | * grace period is in progress, at least until the corresponding | |
1886 | * leaf node has been initialized. In addition, we have excluded | |
1887 | * CPU-hotplug operations. | |
1888 | * | |
1889 | * The grace period cannot complete until the initialization | |
1890 | * process finishes, because this kthread handles both. | |
1891 | */ | |
1892 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
0f41c0dd | 1893 | rcu_gp_slow(rsp, gp_init_delay); |
2a67e741 | 1894 | raw_spin_lock_irq_rcu_node(rnp); |
b3dbec76 | 1895 | rdp = this_cpu_ptr(rsp->rda); |
7fdefc10 PM |
1896 | rcu_preempt_check_blocked_tasks(rnp); |
1897 | rnp->qsmask = rnp->qsmaskinit; | |
7d0ae808 | 1898 | WRITE_ONCE(rnp->gpnum, rsp->gpnum); |
3f47da0f | 1899 | if (WARN_ON_ONCE(rnp->completed != rsp->completed)) |
7d0ae808 | 1900 | WRITE_ONCE(rnp->completed, rsp->completed); |
7fdefc10 | 1901 | if (rnp == rdp->mynode) |
48a7639c | 1902 | (void)__note_gp_changes(rsp, rnp, rdp); |
7fdefc10 PM |
1903 | rcu_preempt_boost_start_gp(rnp); |
1904 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, | |
1905 | rnp->level, rnp->grplo, | |
1906 | rnp->grphi, rnp->qsmask); | |
1907 | raw_spin_unlock_irq(&rnp->lock); | |
bde6c3aa | 1908 | cond_resched_rcu_qs(); |
7d0ae808 | 1909 | WRITE_ONCE(rsp->gp_activity, jiffies); |
7fdefc10 | 1910 | } |
b3dbec76 | 1911 | |
7fdefc10 PM |
1912 | return 1; |
1913 | } | |
b3dbec76 | 1914 | |
b9a425cf PM |
1915 | /* |
1916 | * Helper function for wait_event_interruptible_timeout() wakeup | |
1917 | * at force-quiescent-state time. | |
1918 | */ | |
1919 | static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp) | |
1920 | { | |
1921 | struct rcu_node *rnp = rcu_get_root(rsp); | |
1922 | ||
1923 | /* Someone like call_rcu() requested a force-quiescent-state scan. */ | |
1924 | *gfp = READ_ONCE(rsp->gp_flags); | |
1925 | if (*gfp & RCU_GP_FLAG_FQS) | |
1926 | return true; | |
1927 | ||
1928 | /* The current grace period has completed. */ | |
1929 | if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp)) | |
1930 | return true; | |
1931 | ||
1932 | return false; | |
1933 | } | |
1934 | ||
4cdfc175 PM |
1935 | /* |
1936 | * Do one round of quiescent-state forcing. | |
1937 | */ | |
77f81fe0 | 1938 | static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time) |
4cdfc175 | 1939 | { |
217af2a2 PM |
1940 | bool isidle = false; |
1941 | unsigned long maxj; | |
4cdfc175 PM |
1942 | struct rcu_node *rnp = rcu_get_root(rsp); |
1943 | ||
7d0ae808 | 1944 | WRITE_ONCE(rsp->gp_activity, jiffies); |
4cdfc175 | 1945 | rsp->n_force_qs++; |
77f81fe0 | 1946 | if (first_time) { |
4cdfc175 | 1947 | /* Collect dyntick-idle snapshots. */ |
0edd1b17 | 1948 | if (is_sysidle_rcu_state(rsp)) { |
e02b2edf | 1949 | isidle = true; |
0edd1b17 PM |
1950 | maxj = jiffies - ULONG_MAX / 4; |
1951 | } | |
217af2a2 PM |
1952 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
1953 | &isidle, &maxj); | |
0edd1b17 | 1954 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
4cdfc175 PM |
1955 | } else { |
1956 | /* Handle dyntick-idle and offline CPUs. */ | |
675da67f | 1957 | isidle = true; |
217af2a2 | 1958 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
4cdfc175 PM |
1959 | } |
1960 | /* Clear flag to prevent immediate re-entry. */ | |
7d0ae808 | 1961 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { |
2a67e741 | 1962 | raw_spin_lock_irq_rcu_node(rnp); |
7d0ae808 PM |
1963 | WRITE_ONCE(rsp->gp_flags, |
1964 | READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS); | |
4cdfc175 PM |
1965 | raw_spin_unlock_irq(&rnp->lock); |
1966 | } | |
4cdfc175 PM |
1967 | } |
1968 | ||
7fdefc10 PM |
1969 | /* |
1970 | * Clean up after the old grace period. | |
1971 | */ | |
4cdfc175 | 1972 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
7fdefc10 PM |
1973 | { |
1974 | unsigned long gp_duration; | |
48a7639c | 1975 | bool needgp = false; |
dae6e64d | 1976 | int nocb = 0; |
7fdefc10 PM |
1977 | struct rcu_data *rdp; |
1978 | struct rcu_node *rnp = rcu_get_root(rsp); | |
b3dbec76 | 1979 | |
7d0ae808 | 1980 | WRITE_ONCE(rsp->gp_activity, jiffies); |
2a67e741 | 1981 | raw_spin_lock_irq_rcu_node(rnp); |
7fdefc10 PM |
1982 | gp_duration = jiffies - rsp->gp_start; |
1983 | if (gp_duration > rsp->gp_max) | |
1984 | rsp->gp_max = gp_duration; | |
b3dbec76 | 1985 | |
7fdefc10 PM |
1986 | /* |
1987 | * We know the grace period is complete, but to everyone else | |
1988 | * it appears to still be ongoing. But it is also the case | |
1989 | * that to everyone else it looks like there is nothing that | |
1990 | * they can do to advance the grace period. It is therefore | |
1991 | * safe for us to drop the lock in order to mark the grace | |
1992 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1993 | */ |
5d4b8659 | 1994 | raw_spin_unlock_irq(&rnp->lock); |
b3dbec76 | 1995 | |
5d4b8659 PM |
1996 | /* |
1997 | * Propagate new ->completed value to rcu_node structures so | |
1998 | * that other CPUs don't have to wait until the start of the next | |
1999 | * grace period to process their callbacks. This also avoids | |
2000 | * some nasty RCU grace-period initialization races by forcing | |
2001 | * the end of the current grace period to be completely recorded in | |
2002 | * all of the rcu_node structures before the beginning of the next | |
2003 | * grace period is recorded in any of the rcu_node structures. | |
2004 | */ | |
2005 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
2a67e741 | 2006 | raw_spin_lock_irq_rcu_node(rnp); |
5c60d25f PM |
2007 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
2008 | WARN_ON_ONCE(rnp->qsmask); | |
7d0ae808 | 2009 | WRITE_ONCE(rnp->completed, rsp->gpnum); |
b11cc576 PM |
2010 | rdp = this_cpu_ptr(rsp->rda); |
2011 | if (rnp == rdp->mynode) | |
48a7639c | 2012 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; |
78e4bc34 | 2013 | /* smp_mb() provided by prior unlock-lock pair. */ |
0446be48 | 2014 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
5d4b8659 | 2015 | raw_spin_unlock_irq(&rnp->lock); |
bde6c3aa | 2016 | cond_resched_rcu_qs(); |
7d0ae808 | 2017 | WRITE_ONCE(rsp->gp_activity, jiffies); |
0f41c0dd | 2018 | rcu_gp_slow(rsp, gp_cleanup_delay); |
7fdefc10 | 2019 | } |
5d4b8659 | 2020 | rnp = rcu_get_root(rsp); |
2a67e741 | 2021 | raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */ |
dae6e64d | 2022 | rcu_nocb_gp_set(rnp, nocb); |
7fdefc10 | 2023 | |
765a3f4f | 2024 | /* Declare grace period done. */ |
7d0ae808 | 2025 | WRITE_ONCE(rsp->completed, rsp->gpnum); |
f7f7bac9 | 2026 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
77f81fe0 | 2027 | rsp->gp_state = RCU_GP_IDLE; |
5d4b8659 | 2028 | rdp = this_cpu_ptr(rsp->rda); |
48a7639c PM |
2029 | /* Advance CBs to reduce false positives below. */ |
2030 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; | |
2031 | if (needgp || cpu_needs_another_gp(rsp, rdp)) { | |
7d0ae808 | 2032 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); |
bb311ecc | 2033 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2034 | READ_ONCE(rsp->gpnum), |
bb311ecc PM |
2035 | TPS("newreq")); |
2036 | } | |
7fdefc10 | 2037 | raw_spin_unlock_irq(&rnp->lock); |
7fdefc10 PM |
2038 | } |
2039 | ||
2040 | /* | |
2041 | * Body of kthread that handles grace periods. | |
2042 | */ | |
2043 | static int __noreturn rcu_gp_kthread(void *arg) | |
2044 | { | |
77f81fe0 | 2045 | bool first_gp_fqs; |
88d6df61 | 2046 | int gf; |
d40011f6 | 2047 | unsigned long j; |
4cdfc175 | 2048 | int ret; |
7fdefc10 PM |
2049 | struct rcu_state *rsp = arg; |
2050 | struct rcu_node *rnp = rcu_get_root(rsp); | |
2051 | ||
5871968d | 2052 | rcu_bind_gp_kthread(); |
7fdefc10 PM |
2053 | for (;;) { |
2054 | ||
2055 | /* Handle grace-period start. */ | |
2056 | for (;;) { | |
63c4db78 | 2057 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2058 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2059 | TPS("reqwait")); |
afea227f | 2060 | rsp->gp_state = RCU_GP_WAIT_GPS; |
4cdfc175 | 2061 | wait_event_interruptible(rsp->gp_wq, |
7d0ae808 | 2062 | READ_ONCE(rsp->gp_flags) & |
4cdfc175 | 2063 | RCU_GP_FLAG_INIT); |
319362c9 | 2064 | rsp->gp_state = RCU_GP_DONE_GPS; |
78e4bc34 | 2065 | /* Locking provides needed memory barrier. */ |
f7be8209 | 2066 | if (rcu_gp_init(rsp)) |
7fdefc10 | 2067 | break; |
bde6c3aa | 2068 | cond_resched_rcu_qs(); |
7d0ae808 | 2069 | WRITE_ONCE(rsp->gp_activity, jiffies); |
73a860cd | 2070 | WARN_ON(signal_pending(current)); |
63c4db78 | 2071 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2072 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2073 | TPS("reqwaitsig")); |
7fdefc10 | 2074 | } |
cabc49c1 | 2075 | |
4cdfc175 | 2076 | /* Handle quiescent-state forcing. */ |
77f81fe0 | 2077 | first_gp_fqs = true; |
d40011f6 PM |
2078 | j = jiffies_till_first_fqs; |
2079 | if (j > HZ) { | |
2080 | j = HZ; | |
2081 | jiffies_till_first_fqs = HZ; | |
2082 | } | |
88d6df61 | 2083 | ret = 0; |
cabc49c1 | 2084 | for (;;) { |
88d6df61 PM |
2085 | if (!ret) |
2086 | rsp->jiffies_force_qs = jiffies + j; | |
63c4db78 | 2087 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2088 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2089 | TPS("fqswait")); |
afea227f | 2090 | rsp->gp_state = RCU_GP_WAIT_FQS; |
4cdfc175 | 2091 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
b9a425cf | 2092 | rcu_gp_fqs_check_wake(rsp, &gf), j); |
32bb1c79 | 2093 | rsp->gp_state = RCU_GP_DOING_FQS; |
78e4bc34 | 2094 | /* Locking provides needed memory barriers. */ |
4cdfc175 | 2095 | /* If grace period done, leave loop. */ |
7d0ae808 | 2096 | if (!READ_ONCE(rnp->qsmask) && |
4cdfc175 | 2097 | !rcu_preempt_blocked_readers_cgp(rnp)) |
cabc49c1 | 2098 | break; |
4cdfc175 | 2099 | /* If time for quiescent-state forcing, do it. */ |
88d6df61 PM |
2100 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
2101 | (gf & RCU_GP_FLAG_FQS)) { | |
63c4db78 | 2102 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2103 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2104 | TPS("fqsstart")); |
77f81fe0 PM |
2105 | rcu_gp_fqs(rsp, first_gp_fqs); |
2106 | first_gp_fqs = false; | |
63c4db78 | 2107 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2108 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2109 | TPS("fqsend")); |
bde6c3aa | 2110 | cond_resched_rcu_qs(); |
7d0ae808 | 2111 | WRITE_ONCE(rsp->gp_activity, jiffies); |
4cdfc175 PM |
2112 | } else { |
2113 | /* Deal with stray signal. */ | |
bde6c3aa | 2114 | cond_resched_rcu_qs(); |
7d0ae808 | 2115 | WRITE_ONCE(rsp->gp_activity, jiffies); |
73a860cd | 2116 | WARN_ON(signal_pending(current)); |
63c4db78 | 2117 | trace_rcu_grace_period(rsp->name, |
7d0ae808 | 2118 | READ_ONCE(rsp->gpnum), |
63c4db78 | 2119 | TPS("fqswaitsig")); |
4cdfc175 | 2120 | } |
d40011f6 PM |
2121 | j = jiffies_till_next_fqs; |
2122 | if (j > HZ) { | |
2123 | j = HZ; | |
2124 | jiffies_till_next_fqs = HZ; | |
2125 | } else if (j < 1) { | |
2126 | j = 1; | |
2127 | jiffies_till_next_fqs = 1; | |
2128 | } | |
cabc49c1 | 2129 | } |
4cdfc175 PM |
2130 | |
2131 | /* Handle grace-period end. */ | |
319362c9 | 2132 | rsp->gp_state = RCU_GP_CLEANUP; |
4cdfc175 | 2133 | rcu_gp_cleanup(rsp); |
319362c9 | 2134 | rsp->gp_state = RCU_GP_CLEANED; |
b3dbec76 | 2135 | } |
b3dbec76 PM |
2136 | } |
2137 | ||
64db4cff PM |
2138 | /* |
2139 | * Start a new RCU grace period if warranted, re-initializing the hierarchy | |
2140 | * in preparation for detecting the next grace period. The caller must hold | |
b8462084 | 2141 | * the root node's ->lock and hard irqs must be disabled. |
e5601400 PM |
2142 | * |
2143 | * Note that it is legal for a dying CPU (which is marked as offline) to | |
2144 | * invoke this function. This can happen when the dying CPU reports its | |
2145 | * quiescent state. | |
48a7639c PM |
2146 | * |
2147 | * Returns true if the grace-period kthread must be awakened. | |
64db4cff | 2148 | */ |
48a7639c | 2149 | static bool |
910ee45d PM |
2150 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
2151 | struct rcu_data *rdp) | |
64db4cff | 2152 | { |
b8462084 | 2153 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
afe24b12 | 2154 | /* |
b3dbec76 | 2155 | * Either we have not yet spawned the grace-period |
62da1921 PM |
2156 | * task, this CPU does not need another grace period, |
2157 | * or a grace period is already in progress. | |
b3dbec76 | 2158 | * Either way, don't start a new grace period. |
afe24b12 | 2159 | */ |
48a7639c | 2160 | return false; |
afe24b12 | 2161 | } |
7d0ae808 PM |
2162 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); |
2163 | trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum), | |
bb311ecc | 2164 | TPS("newreq")); |
62da1921 | 2165 | |
016a8d5b SR |
2166 | /* |
2167 | * We can't do wakeups while holding the rnp->lock, as that | |
1eafd31c | 2168 | * could cause possible deadlocks with the rq->lock. Defer |
48a7639c | 2169 | * the wakeup to our caller. |
016a8d5b | 2170 | */ |
48a7639c | 2171 | return true; |
64db4cff PM |
2172 | } |
2173 | ||
910ee45d PM |
2174 | /* |
2175 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's | |
2176 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it | |
2177 | * is invoked indirectly from rcu_advance_cbs(), which would result in | |
2178 | * endless recursion -- or would do so if it wasn't for the self-deadlock | |
2179 | * that is encountered beforehand. | |
48a7639c PM |
2180 | * |
2181 | * Returns true if the grace-period kthread needs to be awakened. | |
910ee45d | 2182 | */ |
48a7639c | 2183 | static bool rcu_start_gp(struct rcu_state *rsp) |
910ee45d PM |
2184 | { |
2185 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); | |
2186 | struct rcu_node *rnp = rcu_get_root(rsp); | |
48a7639c | 2187 | bool ret = false; |
910ee45d PM |
2188 | |
2189 | /* | |
2190 | * If there is no grace period in progress right now, any | |
2191 | * callbacks we have up to this point will be satisfied by the | |
2192 | * next grace period. Also, advancing the callbacks reduces the | |
2193 | * probability of false positives from cpu_needs_another_gp() | |
2194 | * resulting in pointless grace periods. So, advance callbacks | |
2195 | * then start the grace period! | |
2196 | */ | |
48a7639c PM |
2197 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; |
2198 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; | |
2199 | return ret; | |
910ee45d PM |
2200 | } |
2201 | ||
f41d911f | 2202 | /* |
d3f6bad3 PM |
2203 | * Report a full set of quiescent states to the specified rcu_state |
2204 | * data structure. This involves cleaning up after the prior grace | |
2205 | * period and letting rcu_start_gp() start up the next grace period | |
b8462084 PM |
2206 | * if one is needed. Note that the caller must hold rnp->lock, which |
2207 | * is released before return. | |
f41d911f | 2208 | */ |
d3f6bad3 | 2209 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
fc2219d4 | 2210 | __releases(rcu_get_root(rsp)->lock) |
f41d911f | 2211 | { |
fc2219d4 | 2212 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
cd73ca21 | 2213 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); |
cabc49c1 | 2214 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
2aa792e6 | 2215 | rcu_gp_kthread_wake(rsp); |
f41d911f PM |
2216 | } |
2217 | ||
64db4cff | 2218 | /* |
d3f6bad3 PM |
2219 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
2220 | * Allows quiescent states for a group of CPUs to be reported at one go | |
2221 | * to the specified rcu_node structure, though all the CPUs in the group | |
654e9533 PM |
2222 | * must be represented by the same rcu_node structure (which need not be a |
2223 | * leaf rcu_node structure, though it often will be). The gps parameter | |
2224 | * is the grace-period snapshot, which means that the quiescent states | |
2225 | * are valid only if rnp->gpnum is equal to gps. That structure's lock | |
2226 | * must be held upon entry, and it is released before return. | |
64db4cff PM |
2227 | */ |
2228 | static void | |
d3f6bad3 | 2229 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
654e9533 | 2230 | struct rcu_node *rnp, unsigned long gps, unsigned long flags) |
64db4cff PM |
2231 | __releases(rnp->lock) |
2232 | { | |
654e9533 | 2233 | unsigned long oldmask = 0; |
28ecd580 PM |
2234 | struct rcu_node *rnp_c; |
2235 | ||
64db4cff PM |
2236 | /* Walk up the rcu_node hierarchy. */ |
2237 | for (;;) { | |
654e9533 | 2238 | if (!(rnp->qsmask & mask) || rnp->gpnum != gps) { |
64db4cff | 2239 | |
654e9533 PM |
2240 | /* |
2241 | * Our bit has already been cleared, or the | |
2242 | * relevant grace period is already over, so done. | |
2243 | */ | |
1304afb2 | 2244 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2245 | return; |
2246 | } | |
654e9533 | 2247 | WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ |
64db4cff | 2248 | rnp->qsmask &= ~mask; |
d4c08f2a PM |
2249 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
2250 | mask, rnp->qsmask, rnp->level, | |
2251 | rnp->grplo, rnp->grphi, | |
2252 | !!rnp->gp_tasks); | |
27f4d280 | 2253 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
2254 | |
2255 | /* Other bits still set at this level, so done. */ | |
1304afb2 | 2256 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2257 | return; |
2258 | } | |
2259 | mask = rnp->grpmask; | |
2260 | if (rnp->parent == NULL) { | |
2261 | ||
2262 | /* No more levels. Exit loop holding root lock. */ | |
2263 | ||
2264 | break; | |
2265 | } | |
1304afb2 | 2266 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
28ecd580 | 2267 | rnp_c = rnp; |
64db4cff | 2268 | rnp = rnp->parent; |
2a67e741 | 2269 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
654e9533 | 2270 | oldmask = rnp_c->qsmask; |
64db4cff PM |
2271 | } |
2272 | ||
2273 | /* | |
2274 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 2275 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 2276 | * to clean up and start the next grace period if one is needed. |
64db4cff | 2277 | */ |
d3f6bad3 | 2278 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
64db4cff PM |
2279 | } |
2280 | ||
cc99a310 PM |
2281 | /* |
2282 | * Record a quiescent state for all tasks that were previously queued | |
2283 | * on the specified rcu_node structure and that were blocking the current | |
2284 | * RCU grace period. The caller must hold the specified rnp->lock with | |
2285 | * irqs disabled, and this lock is released upon return, but irqs remain | |
2286 | * disabled. | |
2287 | */ | |
0aa04b05 | 2288 | static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp, |
cc99a310 PM |
2289 | struct rcu_node *rnp, unsigned long flags) |
2290 | __releases(rnp->lock) | |
2291 | { | |
654e9533 | 2292 | unsigned long gps; |
cc99a310 PM |
2293 | unsigned long mask; |
2294 | struct rcu_node *rnp_p; | |
2295 | ||
a77da14c PM |
2296 | if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p || |
2297 | rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { | |
cc99a310 PM |
2298 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
2299 | return; /* Still need more quiescent states! */ | |
2300 | } | |
2301 | ||
2302 | rnp_p = rnp->parent; | |
2303 | if (rnp_p == NULL) { | |
2304 | /* | |
a77da14c PM |
2305 | * Only one rcu_node structure in the tree, so don't |
2306 | * try to report up to its nonexistent parent! | |
cc99a310 PM |
2307 | */ |
2308 | rcu_report_qs_rsp(rsp, flags); | |
2309 | return; | |
2310 | } | |
2311 | ||
654e9533 PM |
2312 | /* Report up the rest of the hierarchy, tracking current ->gpnum. */ |
2313 | gps = rnp->gpnum; | |
cc99a310 PM |
2314 | mask = rnp->grpmask; |
2315 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2a67e741 | 2316 | raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */ |
654e9533 | 2317 | rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags); |
cc99a310 PM |
2318 | } |
2319 | ||
64db4cff | 2320 | /* |
d3f6bad3 PM |
2321 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
2322 | * structure. This must be either called from the specified CPU, or | |
2323 | * called when the specified CPU is known to be offline (and when it is | |
2324 | * also known that no other CPU is concurrently trying to help the offline | |
2325 | * CPU). The lastcomp argument is used to make sure we are still in the | |
2326 | * grace period of interest. We don't want to end the current grace period | |
2327 | * based on quiescent states detected in an earlier grace period! | |
64db4cff PM |
2328 | */ |
2329 | static void | |
d7d6a11e | 2330 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2331 | { |
2332 | unsigned long flags; | |
2333 | unsigned long mask; | |
48a7639c | 2334 | bool needwake; |
64db4cff PM |
2335 | struct rcu_node *rnp; |
2336 | ||
2337 | rnp = rdp->mynode; | |
2a67e741 | 2338 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
5b74c458 | 2339 | if ((rdp->cpu_no_qs.b.norm && |
5cd37193 PM |
2340 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) || |
2341 | rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum || | |
2342 | rdp->gpwrap) { | |
64db4cff PM |
2343 | |
2344 | /* | |
e4cc1f22 PM |
2345 | * The grace period in which this quiescent state was |
2346 | * recorded has ended, so don't report it upwards. | |
2347 | * We will instead need a new quiescent state that lies | |
2348 | * within the current grace period. | |
64db4cff | 2349 | */ |
5b74c458 | 2350 | rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */ |
5cd37193 | 2351 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); |
1304afb2 | 2352 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff PM |
2353 | return; |
2354 | } | |
2355 | mask = rdp->grpmask; | |
2356 | if ((rnp->qsmask & mask) == 0) { | |
1304afb2 | 2357 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
64db4cff | 2358 | } else { |
97c668b8 | 2359 | rdp->core_needs_qs = 0; |
64db4cff PM |
2360 | |
2361 | /* | |
2362 | * This GP can't end until cpu checks in, so all of our | |
2363 | * callbacks can be processed during the next GP. | |
2364 | */ | |
48a7639c | 2365 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
64db4cff | 2366 | |
654e9533 PM |
2367 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); |
2368 | /* ^^^ Released rnp->lock */ | |
48a7639c PM |
2369 | if (needwake) |
2370 | rcu_gp_kthread_wake(rsp); | |
64db4cff PM |
2371 | } |
2372 | } | |
2373 | ||
2374 | /* | |
2375 | * Check to see if there is a new grace period of which this CPU | |
2376 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2377 | * Otherwise, see if this CPU has just passed through its first | |
2378 | * quiescent state for this grace period, and record that fact if so. | |
2379 | */ | |
2380 | static void | |
2381 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) | |
2382 | { | |
05eb552b PM |
2383 | /* Check for grace-period ends and beginnings. */ |
2384 | note_gp_changes(rsp, rdp); | |
64db4cff PM |
2385 | |
2386 | /* | |
2387 | * Does this CPU still need to do its part for current grace period? | |
2388 | * If no, return and let the other CPUs do their part as well. | |
2389 | */ | |
97c668b8 | 2390 | if (!rdp->core_needs_qs) |
64db4cff PM |
2391 | return; |
2392 | ||
2393 | /* | |
2394 | * Was there a quiescent state since the beginning of the grace | |
2395 | * period? If no, then exit and wait for the next call. | |
2396 | */ | |
5b74c458 | 2397 | if (rdp->cpu_no_qs.b.norm && |
5cd37193 | 2398 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) |
64db4cff PM |
2399 | return; |
2400 | ||
d3f6bad3 PM |
2401 | /* |
2402 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2403 | * judge of that). | |
2404 | */ | |
d7d6a11e | 2405 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
64db4cff PM |
2406 | } |
2407 | ||
e74f4c45 | 2408 | /* |
b1420f1c PM |
2409 | * Send the specified CPU's RCU callbacks to the orphanage. The |
2410 | * specified CPU must be offline, and the caller must hold the | |
7b2e6011 | 2411 | * ->orphan_lock. |
e74f4c45 | 2412 | */ |
b1420f1c PM |
2413 | static void |
2414 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, | |
2415 | struct rcu_node *rnp, struct rcu_data *rdp) | |
e74f4c45 | 2416 | { |
3fbfbf7a | 2417 | /* No-CBs CPUs do not have orphanable callbacks. */ |
ea46351c | 2418 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a PM |
2419 | return; |
2420 | ||
b1420f1c PM |
2421 | /* |
2422 | * Orphan the callbacks. First adjust the counts. This is safe | |
abfd6e58 PM |
2423 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
2424 | * cannot be running now. Thus no memory barrier is required. | |
b1420f1c | 2425 | */ |
a50c3af9 | 2426 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2427 | rsp->qlen_lazy += rdp->qlen_lazy; |
2428 | rsp->qlen += rdp->qlen; | |
2429 | rdp->n_cbs_orphaned += rdp->qlen; | |
a50c3af9 | 2430 | rdp->qlen_lazy = 0; |
7d0ae808 | 2431 | WRITE_ONCE(rdp->qlen, 0); |
a50c3af9 PM |
2432 | } |
2433 | ||
2434 | /* | |
b1420f1c PM |
2435 | * Next, move those callbacks still needing a grace period to |
2436 | * the orphanage, where some other CPU will pick them up. | |
2437 | * Some of the callbacks might have gone partway through a grace | |
2438 | * period, but that is too bad. They get to start over because we | |
2439 | * cannot assume that grace periods are synchronized across CPUs. | |
2440 | * We don't bother updating the ->nxttail[] array yet, instead | |
2441 | * we just reset the whole thing later on. | |
a50c3af9 | 2442 | */ |
b1420f1c PM |
2443 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
2444 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2445 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2446 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
a50c3af9 PM |
2447 | } |
2448 | ||
2449 | /* | |
b1420f1c PM |
2450 | * Then move the ready-to-invoke callbacks to the orphanage, |
2451 | * where some other CPU will pick them up. These will not be | |
2452 | * required to pass though another grace period: They are done. | |
a50c3af9 | 2453 | */ |
e5601400 | 2454 | if (rdp->nxtlist != NULL) { |
b1420f1c PM |
2455 | *rsp->orphan_donetail = rdp->nxtlist; |
2456 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; | |
e5601400 | 2457 | } |
e74f4c45 | 2458 | |
b33078b6 PM |
2459 | /* |
2460 | * Finally, initialize the rcu_data structure's list to empty and | |
2461 | * disallow further callbacks on this CPU. | |
2462 | */ | |
3f5d3ea6 | 2463 | init_callback_list(rdp); |
b33078b6 | 2464 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
b1420f1c PM |
2465 | } |
2466 | ||
2467 | /* | |
2468 | * Adopt the RCU callbacks from the specified rcu_state structure's | |
7b2e6011 | 2469 | * orphanage. The caller must hold the ->orphan_lock. |
b1420f1c | 2470 | */ |
96d3fd0d | 2471 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) |
b1420f1c PM |
2472 | { |
2473 | int i; | |
fa07a58f | 2474 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
b1420f1c | 2475 | |
3fbfbf7a | 2476 | /* No-CBs CPUs are handled specially. */ |
ea46351c PM |
2477 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || |
2478 | rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) | |
3fbfbf7a PM |
2479 | return; |
2480 | ||
b1420f1c PM |
2481 | /* Do the accounting first. */ |
2482 | rdp->qlen_lazy += rsp->qlen_lazy; | |
2483 | rdp->qlen += rsp->qlen; | |
2484 | rdp->n_cbs_adopted += rsp->qlen; | |
8f5af6f1 PM |
2485 | if (rsp->qlen_lazy != rsp->qlen) |
2486 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
2487 | rsp->qlen_lazy = 0; |
2488 | rsp->qlen = 0; | |
2489 | ||
2490 | /* | |
2491 | * We do not need a memory barrier here because the only way we | |
2492 | * can get here if there is an rcu_barrier() in flight is if | |
2493 | * we are the task doing the rcu_barrier(). | |
2494 | */ | |
2495 | ||
2496 | /* First adopt the ready-to-invoke callbacks. */ | |
2497 | if (rsp->orphan_donelist != NULL) { | |
2498 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; | |
2499 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; | |
2500 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) | |
2501 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2502 | rdp->nxttail[i] = rsp->orphan_donetail; | |
2503 | rsp->orphan_donelist = NULL; | |
2504 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2505 | } | |
2506 | ||
2507 | /* And then adopt the callbacks that still need a grace period. */ | |
2508 | if (rsp->orphan_nxtlist != NULL) { | |
2509 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; | |
2510 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; | |
2511 | rsp->orphan_nxtlist = NULL; | |
2512 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2513 | } | |
2514 | } | |
2515 | ||
2516 | /* | |
2517 | * Trace the fact that this CPU is going offline. | |
2518 | */ | |
2519 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) | |
2520 | { | |
2521 | RCU_TRACE(unsigned long mask); | |
2522 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); | |
2523 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); | |
2524 | ||
ea46351c PM |
2525 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2526 | return; | |
2527 | ||
b1420f1c | 2528 | RCU_TRACE(mask = rdp->grpmask); |
e5601400 PM |
2529 | trace_rcu_grace_period(rsp->name, |
2530 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), | |
f7f7bac9 | 2531 | TPS("cpuofl")); |
64db4cff PM |
2532 | } |
2533 | ||
8af3a5e7 PM |
2534 | /* |
2535 | * All CPUs for the specified rcu_node structure have gone offline, | |
2536 | * and all tasks that were preempted within an RCU read-side critical | |
2537 | * section while running on one of those CPUs have since exited their RCU | |
2538 | * read-side critical section. Some other CPU is reporting this fact with | |
2539 | * the specified rcu_node structure's ->lock held and interrupts disabled. | |
2540 | * This function therefore goes up the tree of rcu_node structures, | |
2541 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that | |
2542 | * the leaf rcu_node structure's ->qsmaskinit field has already been | |
2543 | * updated | |
2544 | * | |
2545 | * This function does check that the specified rcu_node structure has | |
2546 | * all CPUs offline and no blocked tasks, so it is OK to invoke it | |
2547 | * prematurely. That said, invoking it after the fact will cost you | |
2548 | * a needless lock acquisition. So once it has done its work, don't | |
2549 | * invoke it again. | |
2550 | */ | |
2551 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) | |
2552 | { | |
2553 | long mask; | |
2554 | struct rcu_node *rnp = rnp_leaf; | |
2555 | ||
ea46351c PM |
2556 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || |
2557 | rnp->qsmaskinit || rcu_preempt_has_tasks(rnp)) | |
8af3a5e7 PM |
2558 | return; |
2559 | for (;;) { | |
2560 | mask = rnp->grpmask; | |
2561 | rnp = rnp->parent; | |
2562 | if (!rnp) | |
2563 | break; | |
2a67e741 | 2564 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
8af3a5e7 | 2565 | rnp->qsmaskinit &= ~mask; |
0aa04b05 | 2566 | rnp->qsmask &= ~mask; |
8af3a5e7 PM |
2567 | if (rnp->qsmaskinit) { |
2568 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2569 | return; | |
2570 | } | |
2571 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
2572 | } | |
2573 | } | |
2574 | ||
88428cc5 PM |
2575 | /* |
2576 | * The CPU is exiting the idle loop into the arch_cpu_idle_dead() | |
2577 | * function. We now remove it from the rcu_node tree's ->qsmaskinit | |
2578 | * bit masks. | |
2579 | */ | |
2580 | static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp) | |
2581 | { | |
2582 | unsigned long flags; | |
2583 | unsigned long mask; | |
2584 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
2585 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ | |
2586 | ||
ea46351c PM |
2587 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2588 | return; | |
2589 | ||
88428cc5 PM |
2590 | /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ |
2591 | mask = rdp->grpmask; | |
2a67e741 | 2592 | raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */ |
88428cc5 PM |
2593 | rnp->qsmaskinitnext &= ~mask; |
2594 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
2595 | } | |
2596 | ||
64db4cff | 2597 | /* |
e5601400 | 2598 | * The CPU has been completely removed, and some other CPU is reporting |
b1420f1c PM |
2599 | * this fact from process context. Do the remainder of the cleanup, |
2600 | * including orphaning the outgoing CPU's RCU callbacks, and also | |
1331e7a1 PM |
2601 | * adopting them. There can only be one CPU hotplug operation at a time, |
2602 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. | |
64db4cff | 2603 | */ |
e5601400 | 2604 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
64db4cff | 2605 | { |
2036d94a | 2606 | unsigned long flags; |
e5601400 | 2607 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
b1420f1c | 2608 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2609 | |
ea46351c PM |
2610 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
2611 | return; | |
2612 | ||
2036d94a | 2613 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2614 | rcu_boost_kthread_setaffinity(rnp, -1); |
2036d94a | 2615 | |
b1420f1c | 2616 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
78043c46 | 2617 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
b1420f1c | 2618 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); |
96d3fd0d | 2619 | rcu_adopt_orphan_cbs(rsp, flags); |
a8f4cbad | 2620 | raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags); |
b1420f1c | 2621 | |
cf01537e PM |
2622 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
2623 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", | |
2624 | cpu, rdp->qlen, rdp->nxtlist); | |
64db4cff PM |
2625 | } |
2626 | ||
64db4cff PM |
2627 | /* |
2628 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2629 | * period. Thottle as specified by rdp->blimit. | |
2630 | */ | |
37c72e56 | 2631 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
64db4cff PM |
2632 | { |
2633 | unsigned long flags; | |
2634 | struct rcu_head *next, *list, **tail; | |
878d7439 ED |
2635 | long bl, count, count_lazy; |
2636 | int i; | |
64db4cff | 2637 | |
dc35c893 | 2638 | /* If no callbacks are ready, just return. */ |
29c00b4a | 2639 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
486e2593 | 2640 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
7d0ae808 | 2641 | trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist), |
4968c300 PM |
2642 | need_resched(), is_idle_task(current), |
2643 | rcu_is_callbacks_kthread()); | |
64db4cff | 2644 | return; |
29c00b4a | 2645 | } |
64db4cff PM |
2646 | |
2647 | /* | |
2648 | * Extract the list of ready callbacks, disabling to prevent | |
2649 | * races with call_rcu() from interrupt handlers. | |
2650 | */ | |
2651 | local_irq_save(flags); | |
8146c4e2 | 2652 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
29c00b4a | 2653 | bl = rdp->blimit; |
486e2593 | 2654 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
64db4cff PM |
2655 | list = rdp->nxtlist; |
2656 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; | |
2657 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; | |
2658 | tail = rdp->nxttail[RCU_DONE_TAIL]; | |
b41772ab PM |
2659 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
2660 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) | |
2661 | rdp->nxttail[i] = &rdp->nxtlist; | |
64db4cff PM |
2662 | local_irq_restore(flags); |
2663 | ||
2664 | /* Invoke callbacks. */ | |
486e2593 | 2665 | count = count_lazy = 0; |
64db4cff PM |
2666 | while (list) { |
2667 | next = list->next; | |
2668 | prefetch(next); | |
551d55a9 | 2669 | debug_rcu_head_unqueue(list); |
486e2593 PM |
2670 | if (__rcu_reclaim(rsp->name, list)) |
2671 | count_lazy++; | |
64db4cff | 2672 | list = next; |
dff1672d PM |
2673 | /* Stop only if limit reached and CPU has something to do. */ |
2674 | if (++count >= bl && | |
2675 | (need_resched() || | |
2676 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff PM |
2677 | break; |
2678 | } | |
2679 | ||
2680 | local_irq_save(flags); | |
4968c300 PM |
2681 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
2682 | is_idle_task(current), | |
2683 | rcu_is_callbacks_kthread()); | |
64db4cff PM |
2684 | |
2685 | /* Update count, and requeue any remaining callbacks. */ | |
64db4cff PM |
2686 | if (list != NULL) { |
2687 | *tail = rdp->nxtlist; | |
2688 | rdp->nxtlist = list; | |
b41772ab PM |
2689 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
2690 | if (&rdp->nxtlist == rdp->nxttail[i]) | |
2691 | rdp->nxttail[i] = tail; | |
64db4cff PM |
2692 | else |
2693 | break; | |
2694 | } | |
b1420f1c PM |
2695 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
2696 | rdp->qlen_lazy -= count_lazy; | |
7d0ae808 | 2697 | WRITE_ONCE(rdp->qlen, rdp->qlen - count); |
b1420f1c | 2698 | rdp->n_cbs_invoked += count; |
64db4cff PM |
2699 | |
2700 | /* Reinstate batch limit if we have worked down the excess. */ | |
2701 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) | |
2702 | rdp->blimit = blimit; | |
2703 | ||
37c72e56 PM |
2704 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
2705 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { | |
2706 | rdp->qlen_last_fqs_check = 0; | |
2707 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
2708 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) | |
2709 | rdp->qlen_last_fqs_check = rdp->qlen; | |
cfca9279 | 2710 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
37c72e56 | 2711 | |
64db4cff PM |
2712 | local_irq_restore(flags); |
2713 | ||
e0f23060 | 2714 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
64db4cff | 2715 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2716 | invoke_rcu_core(); |
64db4cff PM |
2717 | } |
2718 | ||
2719 | /* | |
2720 | * Check to see if this CPU is in a non-context-switch quiescent state | |
2721 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). | |
e0f23060 | 2722 | * Also schedule RCU core processing. |
64db4cff | 2723 | * |
9b2e4f18 | 2724 | * This function must be called from hardirq context. It is normally |
64db4cff PM |
2725 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
2726 | * false, there is no point in invoking rcu_check_callbacks(). | |
2727 | */ | |
c3377c2d | 2728 | void rcu_check_callbacks(int user) |
64db4cff | 2729 | { |
f7f7bac9 | 2730 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a858af28 | 2731 | increment_cpu_stall_ticks(); |
9b2e4f18 | 2732 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
64db4cff PM |
2733 | |
2734 | /* | |
2735 | * Get here if this CPU took its interrupt from user | |
2736 | * mode or from the idle loop, and if this is not a | |
2737 | * nested interrupt. In this case, the CPU is in | |
d6714c22 | 2738 | * a quiescent state, so note it. |
64db4cff PM |
2739 | * |
2740 | * No memory barrier is required here because both | |
d6714c22 PM |
2741 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
2742 | * variables that other CPUs neither access nor modify, | |
2743 | * at least not while the corresponding CPU is online. | |
64db4cff PM |
2744 | */ |
2745 | ||
284a8c93 PM |
2746 | rcu_sched_qs(); |
2747 | rcu_bh_qs(); | |
64db4cff PM |
2748 | |
2749 | } else if (!in_softirq()) { | |
2750 | ||
2751 | /* | |
2752 | * Get here if this CPU did not take its interrupt from | |
2753 | * softirq, in other words, if it is not interrupting | |
2754 | * a rcu_bh read-side critical section. This is an _bh | |
d6714c22 | 2755 | * critical section, so note it. |
64db4cff PM |
2756 | */ |
2757 | ||
284a8c93 | 2758 | rcu_bh_qs(); |
64db4cff | 2759 | } |
86aea0e6 | 2760 | rcu_preempt_check_callbacks(); |
e3950ecd | 2761 | if (rcu_pending()) |
a46e0899 | 2762 | invoke_rcu_core(); |
8315f422 PM |
2763 | if (user) |
2764 | rcu_note_voluntary_context_switch(current); | |
f7f7bac9 | 2765 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2766 | } |
2767 | ||
64db4cff PM |
2768 | /* |
2769 | * Scan the leaf rcu_node structures, processing dyntick state for any that | |
2770 | * have not yet encountered a quiescent state, using the function specified. | |
27f4d280 PM |
2771 | * Also initiate boosting for any threads blocked on the root rcu_node. |
2772 | * | |
ee47eb9f | 2773 | * The caller must have suppressed start of new grace periods. |
64db4cff | 2774 | */ |
217af2a2 PM |
2775 | static void force_qs_rnp(struct rcu_state *rsp, |
2776 | int (*f)(struct rcu_data *rsp, bool *isidle, | |
2777 | unsigned long *maxj), | |
2778 | bool *isidle, unsigned long *maxj) | |
64db4cff PM |
2779 | { |
2780 | unsigned long bit; | |
2781 | int cpu; | |
2782 | unsigned long flags; | |
2783 | unsigned long mask; | |
a0b6c9a7 | 2784 | struct rcu_node *rnp; |
64db4cff | 2785 | |
a0b6c9a7 | 2786 | rcu_for_each_leaf_node(rsp, rnp) { |
bde6c3aa | 2787 | cond_resched_rcu_qs(); |
64db4cff | 2788 | mask = 0; |
2a67e741 | 2789 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
a0b6c9a7 | 2790 | if (rnp->qsmask == 0) { |
a77da14c PM |
2791 | if (rcu_state_p == &rcu_sched_state || |
2792 | rsp != rcu_state_p || | |
2793 | rcu_preempt_blocked_readers_cgp(rnp)) { | |
2794 | /* | |
2795 | * No point in scanning bits because they | |
2796 | * are all zero. But we might need to | |
2797 | * priority-boost blocked readers. | |
2798 | */ | |
2799 | rcu_initiate_boost(rnp, flags); | |
2800 | /* rcu_initiate_boost() releases rnp->lock */ | |
2801 | continue; | |
2802 | } | |
2803 | if (rnp->parent && | |
2804 | (rnp->parent->qsmask & rnp->grpmask)) { | |
2805 | /* | |
2806 | * Race between grace-period | |
2807 | * initialization and task exiting RCU | |
2808 | * read-side critical section: Report. | |
2809 | */ | |
2810 | rcu_report_unblock_qs_rnp(rsp, rnp, flags); | |
2811 | /* rcu_report_unblock_qs_rnp() rlses ->lock */ | |
2812 | continue; | |
2813 | } | |
64db4cff | 2814 | } |
a0b6c9a7 | 2815 | cpu = rnp->grplo; |
64db4cff | 2816 | bit = 1; |
a0b6c9a7 | 2817 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
0edd1b17 | 2818 | if ((rnp->qsmask & bit) != 0) { |
0edd1b17 PM |
2819 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) |
2820 | mask |= bit; | |
2821 | } | |
64db4cff | 2822 | } |
45f014c5 | 2823 | if (mask != 0) { |
654e9533 PM |
2824 | /* Idle/offline CPUs, report (releases rnp->lock. */ |
2825 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); | |
0aa04b05 PM |
2826 | } else { |
2827 | /* Nothing to do here, so just drop the lock. */ | |
2828 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff | 2829 | } |
64db4cff | 2830 | } |
64db4cff PM |
2831 | } |
2832 | ||
2833 | /* | |
2834 | * Force quiescent states on reluctant CPUs, and also detect which | |
2835 | * CPUs are in dyntick-idle mode. | |
2836 | */ | |
4cdfc175 | 2837 | static void force_quiescent_state(struct rcu_state *rsp) |
64db4cff PM |
2838 | { |
2839 | unsigned long flags; | |
394f2769 PM |
2840 | bool ret; |
2841 | struct rcu_node *rnp; | |
2842 | struct rcu_node *rnp_old = NULL; | |
2843 | ||
2844 | /* Funnel through hierarchy to reduce memory contention. */ | |
d860d403 | 2845 | rnp = __this_cpu_read(rsp->rda->mynode); |
394f2769 | 2846 | for (; rnp != NULL; rnp = rnp->parent) { |
7d0ae808 | 2847 | ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || |
394f2769 PM |
2848 | !raw_spin_trylock(&rnp->fqslock); |
2849 | if (rnp_old != NULL) | |
2850 | raw_spin_unlock(&rnp_old->fqslock); | |
2851 | if (ret) { | |
a792563b | 2852 | rsp->n_force_qs_lh++; |
394f2769 PM |
2853 | return; |
2854 | } | |
2855 | rnp_old = rnp; | |
2856 | } | |
2857 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ | |
64db4cff | 2858 | |
394f2769 | 2859 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2a67e741 | 2860 | raw_spin_lock_irqsave_rcu_node(rnp_old, flags); |
394f2769 | 2861 | raw_spin_unlock(&rnp_old->fqslock); |
7d0ae808 | 2862 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { |
a792563b | 2863 | rsp->n_force_qs_lh++; |
394f2769 | 2864 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
4cdfc175 | 2865 | return; /* Someone beat us to it. */ |
46a1e34e | 2866 | } |
7d0ae808 | 2867 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); |
394f2769 | 2868 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
2aa792e6 | 2869 | rcu_gp_kthread_wake(rsp); |
64db4cff PM |
2870 | } |
2871 | ||
64db4cff | 2872 | /* |
e0f23060 PM |
2873 | * This does the RCU core processing work for the specified rcu_state |
2874 | * and rcu_data structures. This may be called only from the CPU to | |
2875 | * whom the rdp belongs. | |
64db4cff PM |
2876 | */ |
2877 | static void | |
1bca8cf1 | 2878 | __rcu_process_callbacks(struct rcu_state *rsp) |
64db4cff PM |
2879 | { |
2880 | unsigned long flags; | |
48a7639c | 2881 | bool needwake; |
fa07a58f | 2882 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
64db4cff | 2883 | |
2e597558 PM |
2884 | WARN_ON_ONCE(rdp->beenonline == 0); |
2885 | ||
64db4cff PM |
2886 | /* Update RCU state based on any recent quiescent states. */ |
2887 | rcu_check_quiescent_state(rsp, rdp); | |
2888 | ||
2889 | /* Does this CPU require a not-yet-started grace period? */ | |
dc35c893 | 2890 | local_irq_save(flags); |
64db4cff | 2891 | if (cpu_needs_another_gp(rsp, rdp)) { |
6cf10081 | 2892 | raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */ |
48a7639c | 2893 | needwake = rcu_start_gp(rsp); |
b8462084 | 2894 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
48a7639c PM |
2895 | if (needwake) |
2896 | rcu_gp_kthread_wake(rsp); | |
dc35c893 PM |
2897 | } else { |
2898 | local_irq_restore(flags); | |
64db4cff PM |
2899 | } |
2900 | ||
2901 | /* If there are callbacks ready, invoke them. */ | |
09223371 | 2902 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
a46e0899 | 2903 | invoke_rcu_callbacks(rsp, rdp); |
96d3fd0d PM |
2904 | |
2905 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2906 | do_nocb_deferred_wakeup(rdp); | |
09223371 SL |
2907 | } |
2908 | ||
64db4cff | 2909 | /* |
e0f23060 | 2910 | * Do RCU core processing for the current CPU. |
64db4cff | 2911 | */ |
09223371 | 2912 | static void rcu_process_callbacks(struct softirq_action *unused) |
64db4cff | 2913 | { |
6ce75a23 PM |
2914 | struct rcu_state *rsp; |
2915 | ||
bfa00b4c PM |
2916 | if (cpu_is_offline(smp_processor_id())) |
2917 | return; | |
f7f7bac9 | 2918 | trace_rcu_utilization(TPS("Start RCU core")); |
6ce75a23 PM |
2919 | for_each_rcu_flavor(rsp) |
2920 | __rcu_process_callbacks(rsp); | |
f7f7bac9 | 2921 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2922 | } |
2923 | ||
a26ac245 | 2924 | /* |
e0f23060 PM |
2925 | * Schedule RCU callback invocation. If the specified type of RCU |
2926 | * does not support RCU priority boosting, just do a direct call, | |
2927 | * otherwise wake up the per-CPU kernel kthread. Note that because we | |
924df8a0 | 2928 | * are running on the current CPU with softirqs disabled, the |
e0f23060 | 2929 | * rcu_cpu_kthread_task cannot disappear out from under us. |
a26ac245 | 2930 | */ |
a46e0899 | 2931 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
a26ac245 | 2932 | { |
7d0ae808 | 2933 | if (unlikely(!READ_ONCE(rcu_scheduler_fully_active))) |
b0d30417 | 2934 | return; |
a46e0899 PM |
2935 | if (likely(!rsp->boost)) { |
2936 | rcu_do_batch(rsp, rdp); | |
a26ac245 PM |
2937 | return; |
2938 | } | |
a46e0899 | 2939 | invoke_rcu_callbacks_kthread(); |
a26ac245 PM |
2940 | } |
2941 | ||
a46e0899 | 2942 | static void invoke_rcu_core(void) |
09223371 | 2943 | { |
b0f74036 PM |
2944 | if (cpu_online(smp_processor_id())) |
2945 | raise_softirq(RCU_SOFTIRQ); | |
09223371 SL |
2946 | } |
2947 | ||
29154c57 PM |
2948 | /* |
2949 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2950 | */ | |
2951 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, | |
2952 | struct rcu_head *head, unsigned long flags) | |
64db4cff | 2953 | { |
48a7639c PM |
2954 | bool needwake; |
2955 | ||
62fde6ed PM |
2956 | /* |
2957 | * If called from an extended quiescent state, invoke the RCU | |
2958 | * core in order to force a re-evaluation of RCU's idleness. | |
2959 | */ | |
9910affa | 2960 | if (!rcu_is_watching()) |
62fde6ed PM |
2961 | invoke_rcu_core(); |
2962 | ||
a16b7a69 | 2963 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2964 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2965 | return; |
64db4cff | 2966 | |
37c72e56 PM |
2967 | /* |
2968 | * Force the grace period if too many callbacks or too long waiting. | |
2969 | * Enforce hysteresis, and don't invoke force_quiescent_state() | |
2970 | * if some other CPU has recently done so. Also, don't bother | |
2971 | * invoking force_quiescent_state() if the newly enqueued callback | |
2972 | * is the only one waiting for a grace period to complete. | |
2973 | */ | |
2655d57e | 2974 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
b52573d2 PM |
2975 | |
2976 | /* Are we ignoring a completed grace period? */ | |
470716fc | 2977 | note_gp_changes(rsp, rdp); |
b52573d2 PM |
2978 | |
2979 | /* Start a new grace period if one not already started. */ | |
2980 | if (!rcu_gp_in_progress(rsp)) { | |
b52573d2 PM |
2981 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
2982 | ||
2a67e741 | 2983 | raw_spin_lock_rcu_node(rnp_root); |
48a7639c | 2984 | needwake = rcu_start_gp(rsp); |
b8462084 | 2985 | raw_spin_unlock(&rnp_root->lock); |
48a7639c PM |
2986 | if (needwake) |
2987 | rcu_gp_kthread_wake(rsp); | |
b52573d2 PM |
2988 | } else { |
2989 | /* Give the grace period a kick. */ | |
2990 | rdp->blimit = LONG_MAX; | |
2991 | if (rsp->n_force_qs == rdp->n_force_qs_snap && | |
2992 | *rdp->nxttail[RCU_DONE_TAIL] != head) | |
4cdfc175 | 2993 | force_quiescent_state(rsp); |
b52573d2 PM |
2994 | rdp->n_force_qs_snap = rsp->n_force_qs; |
2995 | rdp->qlen_last_fqs_check = rdp->qlen; | |
2996 | } | |
4cdfc175 | 2997 | } |
29154c57 PM |
2998 | } |
2999 | ||
ae150184 PM |
3000 | /* |
3001 | * RCU callback function to leak a callback. | |
3002 | */ | |
3003 | static void rcu_leak_callback(struct rcu_head *rhp) | |
3004 | { | |
3005 | } | |
3006 | ||
3fbfbf7a PM |
3007 | /* |
3008 | * Helper function for call_rcu() and friends. The cpu argument will | |
3009 | * normally be -1, indicating "currently running CPU". It may specify | |
3010 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() | |
3011 | * is expected to specify a CPU. | |
3012 | */ | |
64db4cff | 3013 | static void |
b6a4ae76 | 3014 | __call_rcu(struct rcu_head *head, rcu_callback_t func, |
3fbfbf7a | 3015 | struct rcu_state *rsp, int cpu, bool lazy) |
64db4cff PM |
3016 | { |
3017 | unsigned long flags; | |
3018 | struct rcu_data *rdp; | |
3019 | ||
1146edcb | 3020 | WARN_ON_ONCE((unsigned long)head & 0x1); /* Misaligned rcu_head! */ |
ae150184 PM |
3021 | if (debug_rcu_head_queue(head)) { |
3022 | /* Probable double call_rcu(), so leak the callback. */ | |
7d0ae808 | 3023 | WRITE_ONCE(head->func, rcu_leak_callback); |
ae150184 PM |
3024 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); |
3025 | return; | |
3026 | } | |
64db4cff PM |
3027 | head->func = func; |
3028 | head->next = NULL; | |
3029 | ||
64db4cff PM |
3030 | /* |
3031 | * Opportunistically note grace-period endings and beginnings. | |
3032 | * Note that we might see a beginning right after we see an | |
3033 | * end, but never vice versa, since this CPU has to pass through | |
3034 | * a quiescent state betweentimes. | |
3035 | */ | |
3036 | local_irq_save(flags); | |
394f99a9 | 3037 | rdp = this_cpu_ptr(rsp->rda); |
64db4cff PM |
3038 | |
3039 | /* Add the callback to our list. */ | |
3fbfbf7a PM |
3040 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
3041 | int offline; | |
3042 | ||
3043 | if (cpu != -1) | |
3044 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
143da9c2 PM |
3045 | if (likely(rdp->mynode)) { |
3046 | /* Post-boot, so this should be for a no-CBs CPU. */ | |
3047 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); | |
3048 | WARN_ON_ONCE(offline); | |
3049 | /* Offline CPU, _call_rcu() illegal, leak callback. */ | |
3050 | local_irq_restore(flags); | |
3051 | return; | |
3052 | } | |
3053 | /* | |
3054 | * Very early boot, before rcu_init(). Initialize if needed | |
3055 | * and then drop through to queue the callback. | |
3056 | */ | |
3057 | BUG_ON(cpu != -1); | |
34404ca8 | 3058 | WARN_ON_ONCE(!rcu_is_watching()); |
143da9c2 PM |
3059 | if (!likely(rdp->nxtlist)) |
3060 | init_default_callback_list(rdp); | |
0d8ee37e | 3061 | } |
7d0ae808 | 3062 | WRITE_ONCE(rdp->qlen, rdp->qlen + 1); |
486e2593 PM |
3063 | if (lazy) |
3064 | rdp->qlen_lazy++; | |
c57afe80 PM |
3065 | else |
3066 | rcu_idle_count_callbacks_posted(); | |
b1420f1c PM |
3067 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
3068 | *rdp->nxttail[RCU_NEXT_TAIL] = head; | |
3069 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; | |
2655d57e | 3070 | |
d4c08f2a PM |
3071 | if (__is_kfree_rcu_offset((unsigned long)func)) |
3072 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, | |
486e2593 | 3073 | rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 3074 | else |
486e2593 | 3075 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
d4c08f2a | 3076 | |
29154c57 PM |
3077 | /* Go handle any RCU core processing required. */ |
3078 | __call_rcu_core(rsp, rdp, head, flags); | |
64db4cff PM |
3079 | local_irq_restore(flags); |
3080 | } | |
3081 | ||
3082 | /* | |
d6714c22 | 3083 | * Queue an RCU-sched callback for invocation after a grace period. |
64db4cff | 3084 | */ |
b6a4ae76 | 3085 | void call_rcu_sched(struct rcu_head *head, rcu_callback_t func) |
64db4cff | 3086 | { |
3fbfbf7a | 3087 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
64db4cff | 3088 | } |
d6714c22 | 3089 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
64db4cff PM |
3090 | |
3091 | /* | |
486e2593 | 3092 | * Queue an RCU callback for invocation after a quicker grace period. |
64db4cff | 3093 | */ |
b6a4ae76 | 3094 | void call_rcu_bh(struct rcu_head *head, rcu_callback_t func) |
64db4cff | 3095 | { |
3fbfbf7a | 3096 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
64db4cff PM |
3097 | } |
3098 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
3099 | ||
495aa969 ACB |
3100 | /* |
3101 | * Queue an RCU callback for lazy invocation after a grace period. | |
3102 | * This will likely be later named something like "call_rcu_lazy()", | |
3103 | * but this change will require some way of tagging the lazy RCU | |
3104 | * callbacks in the list of pending callbacks. Until then, this | |
3105 | * function may only be called from __kfree_rcu(). | |
3106 | */ | |
3107 | void kfree_call_rcu(struct rcu_head *head, | |
b6a4ae76 | 3108 | rcu_callback_t func) |
495aa969 | 3109 | { |
e534165b | 3110 | __call_rcu(head, func, rcu_state_p, -1, 1); |
495aa969 ACB |
3111 | } |
3112 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
3113 | ||
6d813391 PM |
3114 | /* |
3115 | * Because a context switch is a grace period for RCU-sched and RCU-bh, | |
3116 | * any blocking grace-period wait automatically implies a grace period | |
3117 | * if there is only one CPU online at any point time during execution | |
3118 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to | |
3119 | * occasionally incorrectly indicate that there are multiple CPUs online | |
3120 | * when there was in fact only one the whole time, as this just adds | |
3121 | * some overhead: RCU still operates correctly. | |
6d813391 PM |
3122 | */ |
3123 | static inline int rcu_blocking_is_gp(void) | |
3124 | { | |
95f0c1de PM |
3125 | int ret; |
3126 | ||
6d813391 | 3127 | might_sleep(); /* Check for RCU read-side critical section. */ |
95f0c1de PM |
3128 | preempt_disable(); |
3129 | ret = num_online_cpus() <= 1; | |
3130 | preempt_enable(); | |
3131 | return ret; | |
6d813391 PM |
3132 | } |
3133 | ||
6ebb237b PM |
3134 | /** |
3135 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. | |
3136 | * | |
3137 | * Control will return to the caller some time after a full rcu-sched | |
3138 | * grace period has elapsed, in other words after all currently executing | |
3139 | * rcu-sched read-side critical sections have completed. These read-side | |
3140 | * critical sections are delimited by rcu_read_lock_sched() and | |
3141 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), | |
3142 | * local_irq_disable(), and so on may be used in place of | |
3143 | * rcu_read_lock_sched(). | |
3144 | * | |
3145 | * This means that all preempt_disable code sequences, including NMI and | |
f0a0e6f2 PM |
3146 | * non-threaded hardware-interrupt handlers, in progress on entry will |
3147 | * have completed before this primitive returns. However, this does not | |
3148 | * guarantee that softirq handlers will have completed, since in some | |
3149 | * kernels, these handlers can run in process context, and can block. | |
3150 | * | |
3151 | * Note that this guarantee implies further memory-ordering guarantees. | |
3152 | * On systems with more than one CPU, when synchronize_sched() returns, | |
3153 | * each CPU is guaranteed to have executed a full memory barrier since the | |
3154 | * end of its last RCU-sched read-side critical section whose beginning | |
3155 | * preceded the call to synchronize_sched(). In addition, each CPU having | |
3156 | * an RCU read-side critical section that extends beyond the return from | |
3157 | * synchronize_sched() is guaranteed to have executed a full memory barrier | |
3158 | * after the beginning of synchronize_sched() and before the beginning of | |
3159 | * that RCU read-side critical section. Note that these guarantees include | |
3160 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
3161 | * that are executing in the kernel. | |
3162 | * | |
3163 | * Furthermore, if CPU A invoked synchronize_sched(), which returned | |
3164 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
3165 | * to have executed a full memory barrier during the execution of | |
3166 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but | |
3167 | * again only if the system has more than one CPU). | |
6ebb237b PM |
3168 | * |
3169 | * This primitive provides the guarantees made by the (now removed) | |
3170 | * synchronize_kernel() API. In contrast, synchronize_rcu() only | |
3171 | * guarantees that rcu_read_lock() sections will have completed. | |
3172 | * In "classic RCU", these two guarantees happen to be one and | |
3173 | * the same, but can differ in realtime RCU implementations. | |
3174 | */ | |
3175 | void synchronize_sched(void) | |
3176 | { | |
f78f5b90 PM |
3177 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
3178 | lock_is_held(&rcu_lock_map) || | |
3179 | lock_is_held(&rcu_sched_lock_map), | |
3180 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); | |
6ebb237b PM |
3181 | if (rcu_blocking_is_gp()) |
3182 | return; | |
5afff48b | 3183 | if (rcu_gp_is_expedited()) |
3705b88d AM |
3184 | synchronize_sched_expedited(); |
3185 | else | |
3186 | wait_rcu_gp(call_rcu_sched); | |
6ebb237b PM |
3187 | } |
3188 | EXPORT_SYMBOL_GPL(synchronize_sched); | |
3189 | ||
3190 | /** | |
3191 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. | |
3192 | * | |
3193 | * Control will return to the caller some time after a full rcu_bh grace | |
3194 | * period has elapsed, in other words after all currently executing rcu_bh | |
3195 | * read-side critical sections have completed. RCU read-side critical | |
3196 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), | |
3197 | * and may be nested. | |
f0a0e6f2 PM |
3198 | * |
3199 | * See the description of synchronize_sched() for more detailed information | |
3200 | * on memory ordering guarantees. | |
6ebb237b PM |
3201 | */ |
3202 | void synchronize_rcu_bh(void) | |
3203 | { | |
f78f5b90 PM |
3204 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
3205 | lock_is_held(&rcu_lock_map) || | |
3206 | lock_is_held(&rcu_sched_lock_map), | |
3207 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); | |
6ebb237b PM |
3208 | if (rcu_blocking_is_gp()) |
3209 | return; | |
5afff48b | 3210 | if (rcu_gp_is_expedited()) |
3705b88d AM |
3211 | synchronize_rcu_bh_expedited(); |
3212 | else | |
3213 | wait_rcu_gp(call_rcu_bh); | |
6ebb237b PM |
3214 | } |
3215 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); | |
3216 | ||
765a3f4f PM |
3217 | /** |
3218 | * get_state_synchronize_rcu - Snapshot current RCU state | |
3219 | * | |
3220 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
3221 | * to determine whether or not a full grace period has elapsed in the | |
3222 | * meantime. | |
3223 | */ | |
3224 | unsigned long get_state_synchronize_rcu(void) | |
3225 | { | |
3226 | /* | |
3227 | * Any prior manipulation of RCU-protected data must happen | |
3228 | * before the load from ->gpnum. | |
3229 | */ | |
3230 | smp_mb(); /* ^^^ */ | |
3231 | ||
3232 | /* | |
3233 | * Make sure this load happens before the purportedly | |
3234 | * time-consuming work between get_state_synchronize_rcu() | |
3235 | * and cond_synchronize_rcu(). | |
3236 | */ | |
e534165b | 3237 | return smp_load_acquire(&rcu_state_p->gpnum); |
765a3f4f PM |
3238 | } |
3239 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
3240 | ||
3241 | /** | |
3242 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
3243 | * | |
3244 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
3245 | * | |
3246 | * If a full RCU grace period has elapsed since the earlier call to | |
3247 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
3248 | * synchronize_rcu() to wait for a full grace period. | |
3249 | * | |
3250 | * Yes, this function does not take counter wrap into account. But | |
3251 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3252 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3253 | * so waiting for one additional grace period should be just fine. | |
3254 | */ | |
3255 | void cond_synchronize_rcu(unsigned long oldstate) | |
3256 | { | |
3257 | unsigned long newstate; | |
3258 | ||
3259 | /* | |
3260 | * Ensure that this load happens before any RCU-destructive | |
3261 | * actions the caller might carry out after we return. | |
3262 | */ | |
e534165b | 3263 | newstate = smp_load_acquire(&rcu_state_p->completed); |
765a3f4f PM |
3264 | if (ULONG_CMP_GE(oldstate, newstate)) |
3265 | synchronize_rcu(); | |
3266 | } | |
3267 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
3268 | ||
24560056 PM |
3269 | /** |
3270 | * get_state_synchronize_sched - Snapshot current RCU-sched state | |
3271 | * | |
3272 | * Returns a cookie that is used by a later call to cond_synchronize_sched() | |
3273 | * to determine whether or not a full grace period has elapsed in the | |
3274 | * meantime. | |
3275 | */ | |
3276 | unsigned long get_state_synchronize_sched(void) | |
3277 | { | |
3278 | /* | |
3279 | * Any prior manipulation of RCU-protected data must happen | |
3280 | * before the load from ->gpnum. | |
3281 | */ | |
3282 | smp_mb(); /* ^^^ */ | |
3283 | ||
3284 | /* | |
3285 | * Make sure this load happens before the purportedly | |
3286 | * time-consuming work between get_state_synchronize_sched() | |
3287 | * and cond_synchronize_sched(). | |
3288 | */ | |
3289 | return smp_load_acquire(&rcu_sched_state.gpnum); | |
3290 | } | |
3291 | EXPORT_SYMBOL_GPL(get_state_synchronize_sched); | |
3292 | ||
3293 | /** | |
3294 | * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period | |
3295 | * | |
3296 | * @oldstate: return value from earlier call to get_state_synchronize_sched() | |
3297 | * | |
3298 | * If a full RCU-sched grace period has elapsed since the earlier call to | |
3299 | * get_state_synchronize_sched(), just return. Otherwise, invoke | |
3300 | * synchronize_sched() to wait for a full grace period. | |
3301 | * | |
3302 | * Yes, this function does not take counter wrap into account. But | |
3303 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3304 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3305 | * so waiting for one additional grace period should be just fine. | |
3306 | */ | |
3307 | void cond_synchronize_sched(unsigned long oldstate) | |
3308 | { | |
3309 | unsigned long newstate; | |
3310 | ||
3311 | /* | |
3312 | * Ensure that this load happens before any RCU-destructive | |
3313 | * actions the caller might carry out after we return. | |
3314 | */ | |
3315 | newstate = smp_load_acquire(&rcu_sched_state.completed); | |
3316 | if (ULONG_CMP_GE(oldstate, newstate)) | |
3317 | synchronize_sched(); | |
3318 | } | |
3319 | EXPORT_SYMBOL_GPL(cond_synchronize_sched); | |
3320 | ||
28f00767 PM |
3321 | /* Adjust sequence number for start of update-side operation. */ |
3322 | static void rcu_seq_start(unsigned long *sp) | |
3323 | { | |
3324 | WRITE_ONCE(*sp, *sp + 1); | |
3325 | smp_mb(); /* Ensure update-side operation after counter increment. */ | |
3326 | WARN_ON_ONCE(!(*sp & 0x1)); | |
3327 | } | |
3328 | ||
3329 | /* Adjust sequence number for end of update-side operation. */ | |
3330 | static void rcu_seq_end(unsigned long *sp) | |
3331 | { | |
3332 | smp_mb(); /* Ensure update-side operation before counter increment. */ | |
3333 | WRITE_ONCE(*sp, *sp + 1); | |
3334 | WARN_ON_ONCE(*sp & 0x1); | |
3335 | } | |
3336 | ||
3337 | /* Take a snapshot of the update side's sequence number. */ | |
3338 | static unsigned long rcu_seq_snap(unsigned long *sp) | |
3339 | { | |
3340 | unsigned long s; | |
3341 | ||
3342 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
3343 | s = (READ_ONCE(*sp) + 3) & ~0x1; | |
3344 | smp_mb(); /* Above access must not bleed into critical section. */ | |
3345 | return s; | |
3346 | } | |
3347 | ||
3348 | /* | |
3349 | * Given a snapshot from rcu_seq_snap(), determine whether or not a | |
3350 | * full update-side operation has occurred. | |
3351 | */ | |
3352 | static bool rcu_seq_done(unsigned long *sp, unsigned long s) | |
3353 | { | |
3354 | return ULONG_CMP_GE(READ_ONCE(*sp), s); | |
3355 | } | |
3356 | ||
3357 | /* Wrapper functions for expedited grace periods. */ | |
3358 | static void rcu_exp_gp_seq_start(struct rcu_state *rsp) | |
3359 | { | |
3360 | rcu_seq_start(&rsp->expedited_sequence); | |
3361 | } | |
3362 | static void rcu_exp_gp_seq_end(struct rcu_state *rsp) | |
3363 | { | |
3364 | rcu_seq_end(&rsp->expedited_sequence); | |
704dd435 | 3365 | smp_mb(); /* Ensure that consecutive grace periods serialize. */ |
28f00767 PM |
3366 | } |
3367 | static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp) | |
3368 | { | |
3369 | return rcu_seq_snap(&rsp->expedited_sequence); | |
3370 | } | |
3371 | static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s) | |
3372 | { | |
3373 | return rcu_seq_done(&rsp->expedited_sequence, s); | |
3374 | } | |
3375 | ||
b9585e94 PM |
3376 | /* |
3377 | * Reset the ->expmaskinit values in the rcu_node tree to reflect any | |
3378 | * recent CPU-online activity. Note that these masks are not cleared | |
3379 | * when CPUs go offline, so they reflect the union of all CPUs that have | |
3380 | * ever been online. This means that this function normally takes its | |
3381 | * no-work-to-do fastpath. | |
3382 | */ | |
3383 | static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp) | |
3384 | { | |
3385 | bool done; | |
3386 | unsigned long flags; | |
3387 | unsigned long mask; | |
3388 | unsigned long oldmask; | |
3389 | int ncpus = READ_ONCE(rsp->ncpus); | |
3390 | struct rcu_node *rnp; | |
3391 | struct rcu_node *rnp_up; | |
3392 | ||
3393 | /* If no new CPUs onlined since last time, nothing to do. */ | |
3394 | if (likely(ncpus == rsp->ncpus_snap)) | |
3395 | return; | |
3396 | rsp->ncpus_snap = ncpus; | |
3397 | ||
3398 | /* | |
3399 | * Each pass through the following loop propagates newly onlined | |
3400 | * CPUs for the current rcu_node structure up the rcu_node tree. | |
3401 | */ | |
3402 | rcu_for_each_leaf_node(rsp, rnp) { | |
2a67e741 | 3403 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
b9585e94 PM |
3404 | if (rnp->expmaskinit == rnp->expmaskinitnext) { |
3405 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3406 | continue; /* No new CPUs, nothing to do. */ | |
3407 | } | |
3408 | ||
3409 | /* Update this node's mask, track old value for propagation. */ | |
3410 | oldmask = rnp->expmaskinit; | |
3411 | rnp->expmaskinit = rnp->expmaskinitnext; | |
3412 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3413 | ||
3414 | /* If was already nonzero, nothing to propagate. */ | |
3415 | if (oldmask) | |
3416 | continue; | |
3417 | ||
3418 | /* Propagate the new CPU up the tree. */ | |
3419 | mask = rnp->grpmask; | |
3420 | rnp_up = rnp->parent; | |
3421 | done = false; | |
3422 | while (rnp_up) { | |
2a67e741 | 3423 | raw_spin_lock_irqsave_rcu_node(rnp_up, flags); |
b9585e94 PM |
3424 | if (rnp_up->expmaskinit) |
3425 | done = true; | |
3426 | rnp_up->expmaskinit |= mask; | |
3427 | raw_spin_unlock_irqrestore(&rnp_up->lock, flags); | |
3428 | if (done) | |
3429 | break; | |
3430 | mask = rnp_up->grpmask; | |
3431 | rnp_up = rnp_up->parent; | |
3432 | } | |
3433 | } | |
3434 | } | |
3435 | ||
3436 | /* | |
3437 | * Reset the ->expmask values in the rcu_node tree in preparation for | |
3438 | * a new expedited grace period. | |
3439 | */ | |
3440 | static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp) | |
3441 | { | |
3442 | unsigned long flags; | |
3443 | struct rcu_node *rnp; | |
3444 | ||
3445 | sync_exp_reset_tree_hotplug(rsp); | |
3446 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
2a67e741 | 3447 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
b9585e94 PM |
3448 | WARN_ON_ONCE(rnp->expmask); |
3449 | rnp->expmask = rnp->expmaskinit; | |
3450 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3451 | } | |
3452 | } | |
3453 | ||
7922cd0e | 3454 | /* |
8203d6d0 | 3455 | * Return non-zero if there is no RCU expedited grace period in progress |
7922cd0e PM |
3456 | * for the specified rcu_node structure, in other words, if all CPUs and |
3457 | * tasks covered by the specified rcu_node structure have done their bit | |
3458 | * for the current expedited grace period. Works only for preemptible | |
3459 | * RCU -- other RCU implementation use other means. | |
3460 | * | |
3461 | * Caller must hold the root rcu_node's exp_funnel_mutex. | |
3462 | */ | |
3463 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) | |
3464 | { | |
8203d6d0 | 3465 | return rnp->exp_tasks == NULL && |
7922cd0e PM |
3466 | READ_ONCE(rnp->expmask) == 0; |
3467 | } | |
3468 | ||
3469 | /* | |
3470 | * Report the exit from RCU read-side critical section for the last task | |
3471 | * that queued itself during or before the current expedited preemptible-RCU | |
3472 | * grace period. This event is reported either to the rcu_node structure on | |
3473 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
3474 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
3475 | * iteratively!) | |
3476 | * | |
8203d6d0 PM |
3477 | * Caller must hold the root rcu_node's exp_funnel_mutex and the |
3478 | * specified rcu_node structure's ->lock. | |
7922cd0e | 3479 | */ |
8203d6d0 PM |
3480 | static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
3481 | bool wake, unsigned long flags) | |
3482 | __releases(rnp->lock) | |
7922cd0e | 3483 | { |
7922cd0e PM |
3484 | unsigned long mask; |
3485 | ||
7922cd0e PM |
3486 | for (;;) { |
3487 | if (!sync_rcu_preempt_exp_done(rnp)) { | |
8203d6d0 PM |
3488 | if (!rnp->expmask) |
3489 | rcu_initiate_boost(rnp, flags); | |
3490 | else | |
3491 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
7922cd0e PM |
3492 | break; |
3493 | } | |
3494 | if (rnp->parent == NULL) { | |
3495 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3496 | if (wake) { | |
3497 | smp_mb(); /* EGP done before wake_up(). */ | |
3498 | wake_up(&rsp->expedited_wq); | |
3499 | } | |
3500 | break; | |
3501 | } | |
3502 | mask = rnp->grpmask; | |
3503 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ | |
3504 | rnp = rnp->parent; | |
2a67e741 | 3505 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled */ |
8203d6d0 | 3506 | WARN_ON_ONCE(!(rnp->expmask & mask)); |
7922cd0e PM |
3507 | rnp->expmask &= ~mask; |
3508 | } | |
3509 | } | |
3510 | ||
8203d6d0 PM |
3511 | /* |
3512 | * Report expedited quiescent state for specified node. This is a | |
3513 | * lock-acquisition wrapper function for __rcu_report_exp_rnp(). | |
3514 | * | |
3515 | * Caller must hold the root rcu_node's exp_funnel_mutex. | |
3516 | */ | |
3517 | static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp, | |
3518 | struct rcu_node *rnp, bool wake) | |
3519 | { | |
3520 | unsigned long flags; | |
3521 | ||
2a67e741 | 3522 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
8203d6d0 PM |
3523 | __rcu_report_exp_rnp(rsp, rnp, wake, flags); |
3524 | } | |
3525 | ||
3526 | /* | |
3527 | * Report expedited quiescent state for multiple CPUs, all covered by the | |
3528 | * specified leaf rcu_node structure. Caller must hold the root | |
3529 | * rcu_node's exp_funnel_mutex. | |
3530 | */ | |
3531 | static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp, | |
3532 | unsigned long mask, bool wake) | |
3533 | { | |
3534 | unsigned long flags; | |
3535 | ||
2a67e741 | 3536 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
338b0f76 PM |
3537 | if (!(rnp->expmask & mask)) { |
3538 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3539 | return; | |
3540 | } | |
8203d6d0 PM |
3541 | rnp->expmask &= ~mask; |
3542 | __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */ | |
3543 | } | |
3544 | ||
3545 | /* | |
3546 | * Report expedited quiescent state for specified rcu_data (CPU). | |
3547 | * Caller must hold the root rcu_node's exp_funnel_mutex. | |
3548 | */ | |
6587a23b PM |
3549 | static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp, |
3550 | bool wake) | |
8203d6d0 PM |
3551 | { |
3552 | rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake); | |
3553 | } | |
3554 | ||
29fd9309 PM |
3555 | /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */ |
3556 | static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp, | |
2cd6ffaf | 3557 | struct rcu_data *rdp, |
29fd9309 | 3558 | atomic_long_t *stat, unsigned long s) |
3d3b7db0 | 3559 | { |
28f00767 | 3560 | if (rcu_exp_gp_seq_done(rsp, s)) { |
385b73c0 PM |
3561 | if (rnp) |
3562 | mutex_unlock(&rnp->exp_funnel_mutex); | |
2cd6ffaf PM |
3563 | else if (rdp) |
3564 | mutex_unlock(&rdp->exp_funnel_mutex); | |
385b73c0 PM |
3565 | /* Ensure test happens before caller kfree(). */ |
3566 | smp_mb__before_atomic(); /* ^^^ */ | |
3567 | atomic_long_inc(stat); | |
385b73c0 PM |
3568 | return true; |
3569 | } | |
3570 | return false; | |
3571 | } | |
3572 | ||
b09e5f86 PM |
3573 | /* |
3574 | * Funnel-lock acquisition for expedited grace periods. Returns a | |
3575 | * pointer to the root rcu_node structure, or NULL if some other | |
3576 | * task did the expedited grace period for us. | |
3577 | */ | |
3578 | static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s) | |
3579 | { | |
2cd6ffaf | 3580 | struct rcu_data *rdp; |
b09e5f86 PM |
3581 | struct rcu_node *rnp0; |
3582 | struct rcu_node *rnp1 = NULL; | |
3583 | ||
3d3b7db0 | 3584 | /* |
cdacbe1f PM |
3585 | * First try directly acquiring the root lock in order to reduce |
3586 | * latency in the common case where expedited grace periods are | |
3587 | * rare. We check mutex_is_locked() to avoid pathological levels of | |
3588 | * memory contention on ->exp_funnel_mutex in the heavy-load case. | |
3d3b7db0 | 3589 | */ |
cdacbe1f PM |
3590 | rnp0 = rcu_get_root(rsp); |
3591 | if (!mutex_is_locked(&rnp0->exp_funnel_mutex)) { | |
3592 | if (mutex_trylock(&rnp0->exp_funnel_mutex)) { | |
3593 | if (sync_exp_work_done(rsp, rnp0, NULL, | |
3594 | &rsp->expedited_workdone0, s)) | |
3595 | return NULL; | |
3596 | return rnp0; | |
3597 | } | |
3598 | } | |
3599 | ||
b09e5f86 PM |
3600 | /* |
3601 | * Each pass through the following loop works its way | |
3602 | * up the rcu_node tree, returning if others have done the | |
3603 | * work or otherwise falls through holding the root rnp's | |
3604 | * ->exp_funnel_mutex. The mapping from CPU to rcu_node structure | |
3605 | * can be inexact, as it is just promoting locality and is not | |
3606 | * strictly needed for correctness. | |
3607 | */ | |
2cd6ffaf PM |
3608 | rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id()); |
3609 | if (sync_exp_work_done(rsp, NULL, NULL, &rsp->expedited_workdone1, s)) | |
3610 | return NULL; | |
3611 | mutex_lock(&rdp->exp_funnel_mutex); | |
3612 | rnp0 = rdp->mynode; | |
b09e5f86 | 3613 | for (; rnp0 != NULL; rnp0 = rnp0->parent) { |
2cd6ffaf PM |
3614 | if (sync_exp_work_done(rsp, rnp1, rdp, |
3615 | &rsp->expedited_workdone2, s)) | |
b09e5f86 PM |
3616 | return NULL; |
3617 | mutex_lock(&rnp0->exp_funnel_mutex); | |
3618 | if (rnp1) | |
3619 | mutex_unlock(&rnp1->exp_funnel_mutex); | |
2cd6ffaf PM |
3620 | else |
3621 | mutex_unlock(&rdp->exp_funnel_mutex); | |
b09e5f86 PM |
3622 | rnp1 = rnp0; |
3623 | } | |
2cd6ffaf PM |
3624 | if (sync_exp_work_done(rsp, rnp1, rdp, |
3625 | &rsp->expedited_workdone3, s)) | |
b09e5f86 PM |
3626 | return NULL; |
3627 | return rnp1; | |
3628 | } | |
3629 | ||
cf3620a6 | 3630 | /* Invoked on each online non-idle CPU for expedited quiescent state. */ |
338b0f76 | 3631 | static void sync_sched_exp_handler(void *data) |
b09e5f86 | 3632 | { |
338b0f76 PM |
3633 | struct rcu_data *rdp; |
3634 | struct rcu_node *rnp; | |
3635 | struct rcu_state *rsp = data; | |
b09e5f86 | 3636 | |
338b0f76 PM |
3637 | rdp = this_cpu_ptr(rsp->rda); |
3638 | rnp = rdp->mynode; | |
3639 | if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) || | |
3640 | __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) | |
3641 | return; | |
6587a23b PM |
3642 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true); |
3643 | resched_cpu(smp_processor_id()); | |
3d3b7db0 PM |
3644 | } |
3645 | ||
338b0f76 PM |
3646 | /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */ |
3647 | static void sync_sched_exp_online_cleanup(int cpu) | |
3648 | { | |
3649 | struct rcu_data *rdp; | |
3650 | int ret; | |
3651 | struct rcu_node *rnp; | |
3652 | struct rcu_state *rsp = &rcu_sched_state; | |
3653 | ||
3654 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
3655 | rnp = rdp->mynode; | |
3656 | if (!(READ_ONCE(rnp->expmask) & rdp->grpmask)) | |
3657 | return; | |
3658 | ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0); | |
3659 | WARN_ON_ONCE(ret); | |
3660 | } | |
3661 | ||
bce5fa12 PM |
3662 | /* |
3663 | * Select the nodes that the upcoming expedited grace period needs | |
3664 | * to wait for. | |
3665 | */ | |
dcdb8807 PM |
3666 | static void sync_rcu_exp_select_cpus(struct rcu_state *rsp, |
3667 | smp_call_func_t func) | |
bce5fa12 PM |
3668 | { |
3669 | int cpu; | |
3670 | unsigned long flags; | |
3671 | unsigned long mask; | |
3672 | unsigned long mask_ofl_test; | |
3673 | unsigned long mask_ofl_ipi; | |
6587a23b | 3674 | int ret; |
bce5fa12 PM |
3675 | struct rcu_node *rnp; |
3676 | ||
3677 | sync_exp_reset_tree(rsp); | |
3678 | rcu_for_each_leaf_node(rsp, rnp) { | |
2a67e741 | 3679 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
bce5fa12 PM |
3680 | |
3681 | /* Each pass checks a CPU for identity, offline, and idle. */ | |
3682 | mask_ofl_test = 0; | |
3683 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) { | |
3684 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
3685 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); | |
3686 | ||
3687 | if (raw_smp_processor_id() == cpu || | |
bce5fa12 PM |
3688 | !(atomic_add_return(0, &rdtp->dynticks) & 0x1)) |
3689 | mask_ofl_test |= rdp->grpmask; | |
3690 | } | |
3691 | mask_ofl_ipi = rnp->expmask & ~mask_ofl_test; | |
3692 | ||
3693 | /* | |
3694 | * Need to wait for any blocked tasks as well. Note that | |
3695 | * additional blocking tasks will also block the expedited | |
3696 | * GP until such time as the ->expmask bits are cleared. | |
3697 | */ | |
3698 | if (rcu_preempt_has_tasks(rnp)) | |
3699 | rnp->exp_tasks = rnp->blkd_tasks.next; | |
3700 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3701 | ||
3702 | /* IPI the remaining CPUs for expedited quiescent state. */ | |
3703 | mask = 1; | |
3704 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) { | |
3705 | if (!(mask_ofl_ipi & mask)) | |
3706 | continue; | |
338b0f76 | 3707 | retry_ipi: |
dcdb8807 | 3708 | ret = smp_call_function_single(cpu, func, rsp, 0); |
338b0f76 | 3709 | if (!ret) { |
6587a23b | 3710 | mask_ofl_ipi &= ~mask; |
338b0f76 PM |
3711 | } else { |
3712 | /* Failed, raced with offline. */ | |
6cf10081 | 3713 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
338b0f76 PM |
3714 | if (cpu_online(cpu) && |
3715 | (rnp->expmask & mask)) { | |
3716 | raw_spin_unlock_irqrestore(&rnp->lock, | |
3717 | flags); | |
3718 | schedule_timeout_uninterruptible(1); | |
3719 | if (cpu_online(cpu) && | |
3720 | (rnp->expmask & mask)) | |
3721 | goto retry_ipi; | |
6cf10081 PM |
3722 | raw_spin_lock_irqsave_rcu_node(rnp, |
3723 | flags); | |
338b0f76 PM |
3724 | } |
3725 | if (!(rnp->expmask & mask)) | |
3726 | mask_ofl_ipi &= ~mask; | |
3727 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
3728 | } | |
bce5fa12 PM |
3729 | } |
3730 | /* Report quiescent states for those that went offline. */ | |
3731 | mask_ofl_test |= mask_ofl_ipi; | |
3732 | if (mask_ofl_test) | |
3733 | rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false); | |
3734 | } | |
3d3b7db0 PM |
3735 | } |
3736 | ||
cf3620a6 PM |
3737 | static void synchronize_sched_expedited_wait(struct rcu_state *rsp) |
3738 | { | |
3739 | int cpu; | |
3740 | unsigned long jiffies_stall; | |
3741 | unsigned long jiffies_start; | |
bce5fa12 PM |
3742 | unsigned long mask; |
3743 | struct rcu_node *rnp; | |
3744 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
cf3620a6 PM |
3745 | int ret; |
3746 | ||
3747 | jiffies_stall = rcu_jiffies_till_stall_check(); | |
3748 | jiffies_start = jiffies; | |
3749 | ||
3750 | for (;;) { | |
3751 | ret = wait_event_interruptible_timeout( | |
3752 | rsp->expedited_wq, | |
bce5fa12 | 3753 | sync_rcu_preempt_exp_done(rnp_root), |
cf3620a6 PM |
3754 | jiffies_stall); |
3755 | if (ret > 0) | |
3756 | return; | |
3757 | if (ret < 0) { | |
3758 | /* Hit a signal, disable CPU stall warnings. */ | |
3759 | wait_event(rsp->expedited_wq, | |
bce5fa12 | 3760 | sync_rcu_preempt_exp_done(rnp_root)); |
cf3620a6 PM |
3761 | return; |
3762 | } | |
c5865638 | 3763 | pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {", |
cf3620a6 | 3764 | rsp->name); |
bce5fa12 | 3765 | rcu_for_each_leaf_node(rsp, rnp) { |
74611ecb | 3766 | (void)rcu_print_task_exp_stall(rnp); |
bce5fa12 PM |
3767 | mask = 1; |
3768 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) { | |
74611ecb PM |
3769 | struct rcu_data *rdp; |
3770 | ||
bce5fa12 PM |
3771 | if (!(rnp->expmask & mask)) |
3772 | continue; | |
74611ecb PM |
3773 | rdp = per_cpu_ptr(rsp->rda, cpu); |
3774 | pr_cont(" %d-%c%c%c", cpu, | |
3775 | "O."[cpu_online(cpu)], | |
3776 | "o."[!!(rdp->grpmask & rnp->expmaskinit)], | |
3777 | "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]); | |
bce5fa12 PM |
3778 | } |
3779 | mask <<= 1; | |
cf3620a6 PM |
3780 | } |
3781 | pr_cont(" } %lu jiffies s: %lu\n", | |
3782 | jiffies - jiffies_start, rsp->expedited_sequence); | |
bce5fa12 PM |
3783 | rcu_for_each_leaf_node(rsp, rnp) { |
3784 | mask = 1; | |
3785 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask <<= 1) { | |
3786 | if (!(rnp->expmask & mask)) | |
3787 | continue; | |
3788 | dump_cpu_task(cpu); | |
3789 | } | |
cf3620a6 PM |
3790 | } |
3791 | jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3; | |
3792 | } | |
3793 | } | |
3794 | ||
236fefaf PM |
3795 | /** |
3796 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
3797 | * | |
3798 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
3799 | * approach to force the grace period to end quickly. This consumes | |
3800 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
3801 | * so is thus not recommended for any sort of common-case code. In fact, | |
3802 | * if you are using synchronize_sched_expedited() in a loop, please | |
3803 | * restructure your code to batch your updates, and then use a single | |
3804 | * synchronize_sched() instead. | |
3d3b7db0 | 3805 | * |
d6ada2cf PM |
3806 | * This implementation can be thought of as an application of sequence |
3807 | * locking to expedited grace periods, but using the sequence counter to | |
3808 | * determine when someone else has already done the work instead of for | |
385b73c0 | 3809 | * retrying readers. |
3d3b7db0 PM |
3810 | */ |
3811 | void synchronize_sched_expedited(void) | |
3812 | { | |
7fd0ddc5 | 3813 | unsigned long s; |
b09e5f86 | 3814 | struct rcu_node *rnp; |
40694d66 | 3815 | struct rcu_state *rsp = &rcu_sched_state; |
3d3b7db0 | 3816 | |
d6ada2cf | 3817 | /* Take a snapshot of the sequence number. */ |
28f00767 | 3818 | s = rcu_exp_gp_seq_snap(rsp); |
3d3b7db0 | 3819 | |
b09e5f86 | 3820 | rnp = exp_funnel_lock(rsp, s); |
807226e2 | 3821 | if (rnp == NULL) |
b09e5f86 | 3822 | return; /* Someone else did our work for us. */ |
e0775cef | 3823 | |
28f00767 | 3824 | rcu_exp_gp_seq_start(rsp); |
338b0f76 | 3825 | sync_rcu_exp_select_cpus(rsp, sync_sched_exp_handler); |
bce5fa12 | 3826 | synchronize_sched_expedited_wait(rsp); |
e0775cef | 3827 | |
28f00767 | 3828 | rcu_exp_gp_seq_end(rsp); |
b09e5f86 | 3829 | mutex_unlock(&rnp->exp_funnel_mutex); |
3d3b7db0 PM |
3830 | } |
3831 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
3832 | ||
64db4cff PM |
3833 | /* |
3834 | * Check to see if there is any immediate RCU-related work to be done | |
3835 | * by the current CPU, for the specified type of RCU, returning 1 if so. | |
3836 | * The checks are in order of increasing expense: checks that can be | |
3837 | * carried out against CPU-local state are performed first. However, | |
3838 | * we must check for CPU stalls first, else we might not get a chance. | |
3839 | */ | |
3840 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) | |
3841 | { | |
2f51f988 PM |
3842 | struct rcu_node *rnp = rdp->mynode; |
3843 | ||
64db4cff PM |
3844 | rdp->n_rcu_pending++; |
3845 | ||
3846 | /* Check for CPU stalls, if enabled. */ | |
3847 | check_cpu_stall(rsp, rdp); | |
3848 | ||
a096932f PM |
3849 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ |
3850 | if (rcu_nohz_full_cpu(rsp)) | |
3851 | return 0; | |
3852 | ||
64db4cff | 3853 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
5c51dd73 | 3854 | if (rcu_scheduler_fully_active && |
5b74c458 | 3855 | rdp->core_needs_qs && rdp->cpu_no_qs.b.norm && |
5cd37193 | 3856 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) { |
97c668b8 PM |
3857 | rdp->n_rp_core_needs_qs++; |
3858 | } else if (rdp->core_needs_qs && | |
5b74c458 | 3859 | (!rdp->cpu_no_qs.b.norm || |
5cd37193 | 3860 | rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) { |
d21670ac | 3861 | rdp->n_rp_report_qs++; |
64db4cff | 3862 | return 1; |
7ba5c840 | 3863 | } |
64db4cff PM |
3864 | |
3865 | /* Does this CPU have callbacks ready to invoke? */ | |
7ba5c840 PM |
3866 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
3867 | rdp->n_rp_cb_ready++; | |
64db4cff | 3868 | return 1; |
7ba5c840 | 3869 | } |
64db4cff PM |
3870 | |
3871 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
7ba5c840 PM |
3872 | if (cpu_needs_another_gp(rsp, rdp)) { |
3873 | rdp->n_rp_cpu_needs_gp++; | |
64db4cff | 3874 | return 1; |
7ba5c840 | 3875 | } |
64db4cff PM |
3876 | |
3877 | /* Has another RCU grace period completed? */ | |
7d0ae808 | 3878 | if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
7ba5c840 | 3879 | rdp->n_rp_gp_completed++; |
64db4cff | 3880 | return 1; |
7ba5c840 | 3881 | } |
64db4cff PM |
3882 | |
3883 | /* Has a new RCU grace period started? */ | |
7d0ae808 PM |
3884 | if (READ_ONCE(rnp->gpnum) != rdp->gpnum || |
3885 | unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */ | |
7ba5c840 | 3886 | rdp->n_rp_gp_started++; |
64db4cff | 3887 | return 1; |
7ba5c840 | 3888 | } |
64db4cff | 3889 | |
96d3fd0d PM |
3890 | /* Does this CPU need a deferred NOCB wakeup? */ |
3891 | if (rcu_nocb_need_deferred_wakeup(rdp)) { | |
3892 | rdp->n_rp_nocb_defer_wakeup++; | |
3893 | return 1; | |
3894 | } | |
3895 | ||
64db4cff | 3896 | /* nothing to do */ |
7ba5c840 | 3897 | rdp->n_rp_need_nothing++; |
64db4cff PM |
3898 | return 0; |
3899 | } | |
3900 | ||
3901 | /* | |
3902 | * Check to see if there is any immediate RCU-related work to be done | |
3903 | * by the current CPU, returning 1 if so. This function is part of the | |
3904 | * RCU implementation; it is -not- an exported member of the RCU API. | |
3905 | */ | |
e3950ecd | 3906 | static int rcu_pending(void) |
64db4cff | 3907 | { |
6ce75a23 PM |
3908 | struct rcu_state *rsp; |
3909 | ||
3910 | for_each_rcu_flavor(rsp) | |
e3950ecd | 3911 | if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda))) |
6ce75a23 PM |
3912 | return 1; |
3913 | return 0; | |
64db4cff PM |
3914 | } |
3915 | ||
3916 | /* | |
c0f4dfd4 PM |
3917 | * Return true if the specified CPU has any callback. If all_lazy is |
3918 | * non-NULL, store an indication of whether all callbacks are lazy. | |
3919 | * (If there are no callbacks, all of them are deemed to be lazy.) | |
64db4cff | 3920 | */ |
82072c4f | 3921 | static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy) |
64db4cff | 3922 | { |
c0f4dfd4 PM |
3923 | bool al = true; |
3924 | bool hc = false; | |
3925 | struct rcu_data *rdp; | |
6ce75a23 PM |
3926 | struct rcu_state *rsp; |
3927 | ||
c0f4dfd4 | 3928 | for_each_rcu_flavor(rsp) { |
aa6da514 | 3929 | rdp = this_cpu_ptr(rsp->rda); |
69c8d28c PM |
3930 | if (!rdp->nxtlist) |
3931 | continue; | |
3932 | hc = true; | |
3933 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { | |
c0f4dfd4 | 3934 | al = false; |
69c8d28c PM |
3935 | break; |
3936 | } | |
c0f4dfd4 PM |
3937 | } |
3938 | if (all_lazy) | |
3939 | *all_lazy = al; | |
3940 | return hc; | |
64db4cff PM |
3941 | } |
3942 | ||
a83eff0a PM |
3943 | /* |
3944 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, | |
3945 | * the compiler is expected to optimize this away. | |
3946 | */ | |
e66c33d5 | 3947 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
a83eff0a PM |
3948 | int cpu, unsigned long done) |
3949 | { | |
3950 | trace_rcu_barrier(rsp->name, s, cpu, | |
3951 | atomic_read(&rsp->barrier_cpu_count), done); | |
3952 | } | |
3953 | ||
b1420f1c PM |
3954 | /* |
3955 | * RCU callback function for _rcu_barrier(). If we are last, wake | |
3956 | * up the task executing _rcu_barrier(). | |
3957 | */ | |
24ebbca8 | 3958 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3959 | { |
24ebbca8 PM |
3960 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
3961 | struct rcu_state *rsp = rdp->rsp; | |
3962 | ||
a83eff0a | 3963 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
4f525a52 | 3964 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence); |
7db74df8 | 3965 | complete(&rsp->barrier_completion); |
a83eff0a | 3966 | } else { |
4f525a52 | 3967 | _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence); |
a83eff0a | 3968 | } |
d0ec774c PM |
3969 | } |
3970 | ||
3971 | /* | |
3972 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3973 | */ | |
3974 | static void rcu_barrier_func(void *type) | |
3975 | { | |
037b64ed | 3976 | struct rcu_state *rsp = type; |
fa07a58f | 3977 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); |
d0ec774c | 3978 | |
4f525a52 | 3979 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence); |
24ebbca8 | 3980 | atomic_inc(&rsp->barrier_cpu_count); |
06668efa | 3981 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
d0ec774c PM |
3982 | } |
3983 | ||
d0ec774c PM |
3984 | /* |
3985 | * Orchestrate the specified type of RCU barrier, waiting for all | |
3986 | * RCU callbacks of the specified type to complete. | |
3987 | */ | |
037b64ed | 3988 | static void _rcu_barrier(struct rcu_state *rsp) |
d0ec774c | 3989 | { |
b1420f1c | 3990 | int cpu; |
b1420f1c | 3991 | struct rcu_data *rdp; |
4f525a52 | 3992 | unsigned long s = rcu_seq_snap(&rsp->barrier_sequence); |
b1420f1c | 3993 | |
4f525a52 | 3994 | _rcu_barrier_trace(rsp, "Begin", -1, s); |
b1420f1c | 3995 | |
e74f4c45 | 3996 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
7be7f0be | 3997 | mutex_lock(&rsp->barrier_mutex); |
b1420f1c | 3998 | |
4f525a52 PM |
3999 | /* Did someone else do our work for us? */ |
4000 | if (rcu_seq_done(&rsp->barrier_sequence, s)) { | |
4001 | _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence); | |
cf3a9c48 PM |
4002 | smp_mb(); /* caller's subsequent code after above check. */ |
4003 | mutex_unlock(&rsp->barrier_mutex); | |
4004 | return; | |
4005 | } | |
4006 | ||
4f525a52 PM |
4007 | /* Mark the start of the barrier operation. */ |
4008 | rcu_seq_start(&rsp->barrier_sequence); | |
4009 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence); | |
b1420f1c | 4010 | |
d0ec774c | 4011 | /* |
b1420f1c PM |
4012 | * Initialize the count to one rather than to zero in order to |
4013 | * avoid a too-soon return to zero in case of a short grace period | |
1331e7a1 PM |
4014 | * (or preemption of this task). Exclude CPU-hotplug operations |
4015 | * to ensure that no offline CPU has callbacks queued. | |
d0ec774c | 4016 | */ |
7db74df8 | 4017 | init_completion(&rsp->barrier_completion); |
24ebbca8 | 4018 | atomic_set(&rsp->barrier_cpu_count, 1); |
1331e7a1 | 4019 | get_online_cpus(); |
b1420f1c PM |
4020 | |
4021 | /* | |
1331e7a1 PM |
4022 | * Force each CPU with callbacks to register a new callback. |
4023 | * When that callback is invoked, we will know that all of the | |
4024 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 4025 | */ |
3fbfbf7a | 4026 | for_each_possible_cpu(cpu) { |
d1e43fa5 | 4027 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
3fbfbf7a | 4028 | continue; |
b1420f1c | 4029 | rdp = per_cpu_ptr(rsp->rda, cpu); |
d1e43fa5 | 4030 | if (rcu_is_nocb_cpu(cpu)) { |
d7e29933 PM |
4031 | if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) { |
4032 | _rcu_barrier_trace(rsp, "OfflineNoCB", cpu, | |
4f525a52 | 4033 | rsp->barrier_sequence); |
d7e29933 PM |
4034 | } else { |
4035 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, | |
4f525a52 | 4036 | rsp->barrier_sequence); |
41050a00 | 4037 | smp_mb__before_atomic(); |
d7e29933 PM |
4038 | atomic_inc(&rsp->barrier_cpu_count); |
4039 | __call_rcu(&rdp->barrier_head, | |
4040 | rcu_barrier_callback, rsp, cpu, 0); | |
4041 | } | |
7d0ae808 | 4042 | } else if (READ_ONCE(rdp->qlen)) { |
a83eff0a | 4043 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
4f525a52 | 4044 | rsp->barrier_sequence); |
037b64ed | 4045 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
b1420f1c | 4046 | } else { |
a83eff0a | 4047 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
4f525a52 | 4048 | rsp->barrier_sequence); |
b1420f1c PM |
4049 | } |
4050 | } | |
1331e7a1 | 4051 | put_online_cpus(); |
b1420f1c PM |
4052 | |
4053 | /* | |
4054 | * Now that we have an rcu_barrier_callback() callback on each | |
4055 | * CPU, and thus each counted, remove the initial count. | |
4056 | */ | |
24ebbca8 | 4057 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
7db74df8 | 4058 | complete(&rsp->barrier_completion); |
b1420f1c PM |
4059 | |
4060 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ | |
7db74df8 | 4061 | wait_for_completion(&rsp->barrier_completion); |
b1420f1c | 4062 | |
4f525a52 PM |
4063 | /* Mark the end of the barrier operation. */ |
4064 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence); | |
4065 | rcu_seq_end(&rsp->barrier_sequence); | |
4066 | ||
b1420f1c | 4067 | /* Other rcu_barrier() invocations can now safely proceed. */ |
7be7f0be | 4068 | mutex_unlock(&rsp->barrier_mutex); |
d0ec774c | 4069 | } |
d0ec774c PM |
4070 | |
4071 | /** | |
4072 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. | |
4073 | */ | |
4074 | void rcu_barrier_bh(void) | |
4075 | { | |
037b64ed | 4076 | _rcu_barrier(&rcu_bh_state); |
d0ec774c PM |
4077 | } |
4078 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); | |
4079 | ||
4080 | /** | |
4081 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. | |
4082 | */ | |
4083 | void rcu_barrier_sched(void) | |
4084 | { | |
037b64ed | 4085 | _rcu_barrier(&rcu_sched_state); |
d0ec774c PM |
4086 | } |
4087 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); | |
4088 | ||
0aa04b05 PM |
4089 | /* |
4090 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the | |
4091 | * first CPU in a given leaf rcu_node structure coming online. The caller | |
4092 | * must hold the corresponding leaf rcu_node ->lock with interrrupts | |
4093 | * disabled. | |
4094 | */ | |
4095 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) | |
4096 | { | |
4097 | long mask; | |
4098 | struct rcu_node *rnp = rnp_leaf; | |
4099 | ||
4100 | for (;;) { | |
4101 | mask = rnp->grpmask; | |
4102 | rnp = rnp->parent; | |
4103 | if (rnp == NULL) | |
4104 | return; | |
6cf10081 | 4105 | raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */ |
0aa04b05 PM |
4106 | rnp->qsmaskinit |= mask; |
4107 | raw_spin_unlock(&rnp->lock); /* Interrupts remain disabled. */ | |
4108 | } | |
4109 | } | |
4110 | ||
64db4cff | 4111 | /* |
27569620 | 4112 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 4113 | */ |
27569620 PM |
4114 | static void __init |
4115 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) | |
64db4cff PM |
4116 | { |
4117 | unsigned long flags; | |
394f99a9 | 4118 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
27569620 PM |
4119 | struct rcu_node *rnp = rcu_get_root(rsp); |
4120 | ||
4121 | /* Set up local state, ensuring consistent view of global state. */ | |
6cf10081 | 4122 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
27569620 | 4123 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
27569620 | 4124 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
29e37d81 | 4125 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
9b2e4f18 | 4126 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
27569620 | 4127 | rdp->cpu = cpu; |
d4c08f2a | 4128 | rdp->rsp = rsp; |
2cd6ffaf | 4129 | mutex_init(&rdp->exp_funnel_mutex); |
3fbfbf7a | 4130 | rcu_boot_init_nocb_percpu_data(rdp); |
1304afb2 | 4131 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27569620 PM |
4132 | } |
4133 | ||
4134 | /* | |
4135 | * Initialize a CPU's per-CPU RCU data. Note that only one online or | |
4136 | * offline event can be happening at a given time. Note also that we | |
4137 | * can accept some slop in the rsp->completed access due to the fact | |
4138 | * that this CPU cannot possibly have any RCU callbacks in flight yet. | |
64db4cff | 4139 | */ |
49fb4c62 | 4140 | static void |
9b67122a | 4141 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp) |
64db4cff PM |
4142 | { |
4143 | unsigned long flags; | |
64db4cff | 4144 | unsigned long mask; |
394f99a9 | 4145 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
64db4cff PM |
4146 | struct rcu_node *rnp = rcu_get_root(rsp); |
4147 | ||
4148 | /* Set up local state, ensuring consistent view of global state. */ | |
6cf10081 | 4149 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
37c72e56 PM |
4150 | rdp->qlen_last_fqs_check = 0; |
4151 | rdp->n_force_qs_snap = rsp->n_force_qs; | |
64db4cff | 4152 | rdp->blimit = blimit; |
39c8d313 PM |
4153 | if (!rdp->nxtlist) |
4154 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ | |
29e37d81 | 4155 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
2333210b | 4156 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
c92b131b PM |
4157 | atomic_set(&rdp->dynticks->dynticks, |
4158 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); | |
1304afb2 | 4159 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
64db4cff | 4160 | |
0aa04b05 PM |
4161 | /* |
4162 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed | |
4163 | * propagation up the rcu_node tree will happen at the beginning | |
4164 | * of the next grace period. | |
4165 | */ | |
64db4cff PM |
4166 | rnp = rdp->mynode; |
4167 | mask = rdp->grpmask; | |
2a67e741 | 4168 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
0aa04b05 | 4169 | rnp->qsmaskinitnext |= mask; |
b9585e94 PM |
4170 | rnp->expmaskinitnext |= mask; |
4171 | if (!rdp->beenonline) | |
4172 | WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1); | |
4173 | rdp->beenonline = true; /* We have now been online. */ | |
0aa04b05 PM |
4174 | rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */ |
4175 | rdp->completed = rnp->completed; | |
5b74c458 | 4176 | rdp->cpu_no_qs.b.norm = true; |
a738eec6 | 4177 | rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu); |
97c668b8 | 4178 | rdp->core_needs_qs = false; |
0aa04b05 PM |
4179 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); |
4180 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
64db4cff PM |
4181 | } |
4182 | ||
49fb4c62 | 4183 | static void rcu_prepare_cpu(int cpu) |
64db4cff | 4184 | { |
6ce75a23 PM |
4185 | struct rcu_state *rsp; |
4186 | ||
4187 | for_each_rcu_flavor(rsp) | |
9b67122a | 4188 | rcu_init_percpu_data(cpu, rsp); |
64db4cff PM |
4189 | } |
4190 | ||
4191 | /* | |
f41d911f | 4192 | * Handle CPU online/offline notification events. |
64db4cff | 4193 | */ |
88428cc5 PM |
4194 | int rcu_cpu_notify(struct notifier_block *self, |
4195 | unsigned long action, void *hcpu) | |
64db4cff PM |
4196 | { |
4197 | long cpu = (long)hcpu; | |
e534165b | 4198 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
a26ac245 | 4199 | struct rcu_node *rnp = rdp->mynode; |
6ce75a23 | 4200 | struct rcu_state *rsp; |
64db4cff PM |
4201 | |
4202 | switch (action) { | |
4203 | case CPU_UP_PREPARE: | |
4204 | case CPU_UP_PREPARE_FROZEN: | |
d72bce0e PZ |
4205 | rcu_prepare_cpu(cpu); |
4206 | rcu_prepare_kthreads(cpu); | |
35ce7f29 | 4207 | rcu_spawn_all_nocb_kthreads(cpu); |
a26ac245 PM |
4208 | break; |
4209 | case CPU_ONLINE: | |
0f962a5e | 4210 | case CPU_DOWN_FAILED: |
338b0f76 | 4211 | sync_sched_exp_online_cleanup(cpu); |
5d01bbd1 | 4212 | rcu_boost_kthread_setaffinity(rnp, -1); |
0f962a5e PM |
4213 | break; |
4214 | case CPU_DOWN_PREPARE: | |
34ed6246 | 4215 | rcu_boost_kthread_setaffinity(rnp, cpu); |
64db4cff | 4216 | break; |
d0ec774c PM |
4217 | case CPU_DYING: |
4218 | case CPU_DYING_FROZEN: | |
6ce75a23 PM |
4219 | for_each_rcu_flavor(rsp) |
4220 | rcu_cleanup_dying_cpu(rsp); | |
d0ec774c | 4221 | break; |
88428cc5 | 4222 | case CPU_DYING_IDLE: |
6587a23b | 4223 | /* QS for any half-done expedited RCU-sched GP. */ |
338b0f76 PM |
4224 | preempt_disable(); |
4225 | rcu_report_exp_rdp(&rcu_sched_state, | |
4226 | this_cpu_ptr(rcu_sched_state.rda), true); | |
4227 | preempt_enable(); | |
6587a23b | 4228 | |
88428cc5 PM |
4229 | for_each_rcu_flavor(rsp) { |
4230 | rcu_cleanup_dying_idle_cpu(cpu, rsp); | |
4231 | } | |
4232 | break; | |
64db4cff PM |
4233 | case CPU_DEAD: |
4234 | case CPU_DEAD_FROZEN: | |
4235 | case CPU_UP_CANCELED: | |
4236 | case CPU_UP_CANCELED_FROZEN: | |
776d6807 | 4237 | for_each_rcu_flavor(rsp) { |
6ce75a23 | 4238 | rcu_cleanup_dead_cpu(cpu, rsp); |
776d6807 PM |
4239 | do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu)); |
4240 | } | |
64db4cff PM |
4241 | break; |
4242 | default: | |
4243 | break; | |
4244 | } | |
34ed6246 | 4245 | return NOTIFY_OK; |
64db4cff PM |
4246 | } |
4247 | ||
d1d74d14 BP |
4248 | static int rcu_pm_notify(struct notifier_block *self, |
4249 | unsigned long action, void *hcpu) | |
4250 | { | |
4251 | switch (action) { | |
4252 | case PM_HIBERNATION_PREPARE: | |
4253 | case PM_SUSPEND_PREPARE: | |
4254 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ | |
5afff48b | 4255 | rcu_expedite_gp(); |
d1d74d14 BP |
4256 | break; |
4257 | case PM_POST_HIBERNATION: | |
4258 | case PM_POST_SUSPEND: | |
5afff48b PM |
4259 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ |
4260 | rcu_unexpedite_gp(); | |
d1d74d14 BP |
4261 | break; |
4262 | default: | |
4263 | break; | |
4264 | } | |
4265 | return NOTIFY_OK; | |
4266 | } | |
4267 | ||
b3dbec76 | 4268 | /* |
9386c0b7 | 4269 | * Spawn the kthreads that handle each RCU flavor's grace periods. |
b3dbec76 PM |
4270 | */ |
4271 | static int __init rcu_spawn_gp_kthread(void) | |
4272 | { | |
4273 | unsigned long flags; | |
a94844b2 | 4274 | int kthread_prio_in = kthread_prio; |
b3dbec76 PM |
4275 | struct rcu_node *rnp; |
4276 | struct rcu_state *rsp; | |
a94844b2 | 4277 | struct sched_param sp; |
b3dbec76 PM |
4278 | struct task_struct *t; |
4279 | ||
a94844b2 PM |
4280 | /* Force priority into range. */ |
4281 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) | |
4282 | kthread_prio = 1; | |
4283 | else if (kthread_prio < 0) | |
4284 | kthread_prio = 0; | |
4285 | else if (kthread_prio > 99) | |
4286 | kthread_prio = 99; | |
4287 | if (kthread_prio != kthread_prio_in) | |
4288 | pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n", | |
4289 | kthread_prio, kthread_prio_in); | |
4290 | ||
9386c0b7 | 4291 | rcu_scheduler_fully_active = 1; |
b3dbec76 | 4292 | for_each_rcu_flavor(rsp) { |
a94844b2 | 4293 | t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name); |
b3dbec76 PM |
4294 | BUG_ON(IS_ERR(t)); |
4295 | rnp = rcu_get_root(rsp); | |
6cf10081 | 4296 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
b3dbec76 | 4297 | rsp->gp_kthread = t; |
a94844b2 PM |
4298 | if (kthread_prio) { |
4299 | sp.sched_priority = kthread_prio; | |
4300 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
4301 | } | |
4302 | wake_up_process(t); | |
b3dbec76 PM |
4303 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
4304 | } | |
35ce7f29 | 4305 | rcu_spawn_nocb_kthreads(); |
9386c0b7 | 4306 | rcu_spawn_boost_kthreads(); |
b3dbec76 PM |
4307 | return 0; |
4308 | } | |
4309 | early_initcall(rcu_spawn_gp_kthread); | |
4310 | ||
bbad9379 PM |
4311 | /* |
4312 | * This function is invoked towards the end of the scheduler's initialization | |
4313 | * process. Before this is called, the idle task might contain | |
4314 | * RCU read-side critical sections (during which time, this idle | |
4315 | * task is booting the system). After this function is called, the | |
4316 | * idle tasks are prohibited from containing RCU read-side critical | |
4317 | * sections. This function also enables RCU lockdep checking. | |
4318 | */ | |
4319 | void rcu_scheduler_starting(void) | |
4320 | { | |
4321 | WARN_ON(num_online_cpus() != 1); | |
4322 | WARN_ON(nr_context_switches() > 0); | |
4323 | rcu_scheduler_active = 1; | |
4324 | } | |
4325 | ||
64db4cff PM |
4326 | /* |
4327 | * Compute the per-level fanout, either using the exact fanout specified | |
7fa27001 | 4328 | * or balancing the tree, depending on the rcu_fanout_exact boot parameter. |
64db4cff | 4329 | */ |
199977bf | 4330 | static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt) |
64db4cff | 4331 | { |
64db4cff PM |
4332 | int i; |
4333 | ||
7fa27001 | 4334 | if (rcu_fanout_exact) { |
199977bf | 4335 | levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; |
66292405 | 4336 | for (i = rcu_num_lvls - 2; i >= 0; i--) |
199977bf | 4337 | levelspread[i] = RCU_FANOUT; |
66292405 PM |
4338 | } else { |
4339 | int ccur; | |
4340 | int cprv; | |
4341 | ||
4342 | cprv = nr_cpu_ids; | |
4343 | for (i = rcu_num_lvls - 1; i >= 0; i--) { | |
199977bf AG |
4344 | ccur = levelcnt[i]; |
4345 | levelspread[i] = (cprv + ccur - 1) / ccur; | |
66292405 PM |
4346 | cprv = ccur; |
4347 | } | |
64db4cff PM |
4348 | } |
4349 | } | |
64db4cff PM |
4350 | |
4351 | /* | |
4352 | * Helper function for rcu_init() that initializes one rcu_state structure. | |
4353 | */ | |
a87f203e | 4354 | static void __init rcu_init_one(struct rcu_state *rsp) |
64db4cff | 4355 | { |
cb007102 AG |
4356 | static const char * const buf[] = RCU_NODE_NAME_INIT; |
4357 | static const char * const fqs[] = RCU_FQS_NAME_INIT; | |
385b73c0 | 4358 | static const char * const exp[] = RCU_EXP_NAME_INIT; |
3dc5dbe9 PM |
4359 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
4360 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; | |
4361 | static struct lock_class_key rcu_exp_class[RCU_NUM_LVLS]; | |
4a81e832 | 4362 | static u8 fl_mask = 0x1; |
199977bf AG |
4363 | |
4364 | int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */ | |
4365 | int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */ | |
64db4cff PM |
4366 | int cpustride = 1; |
4367 | int i; | |
4368 | int j; | |
4369 | struct rcu_node *rnp; | |
4370 | ||
05b84aec | 4371 | BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
b6407e86 | 4372 | |
3eaaaf6c PM |
4373 | /* Silence gcc 4.8 false positive about array index out of range. */ |
4374 | if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS) | |
4375 | panic("rcu_init_one: rcu_num_lvls out of range"); | |
4930521a | 4376 | |
64db4cff PM |
4377 | /* Initialize the level-tracking arrays. */ |
4378 | ||
f885b7f2 | 4379 | for (i = 0; i < rcu_num_lvls; i++) |
199977bf | 4380 | levelcnt[i] = num_rcu_lvl[i]; |
f885b7f2 | 4381 | for (i = 1; i < rcu_num_lvls; i++) |
199977bf AG |
4382 | rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1]; |
4383 | rcu_init_levelspread(levelspread, levelcnt); | |
4a81e832 PM |
4384 | rsp->flavor_mask = fl_mask; |
4385 | fl_mask <<= 1; | |
64db4cff PM |
4386 | |
4387 | /* Initialize the elements themselves, starting from the leaves. */ | |
4388 | ||
f885b7f2 | 4389 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
199977bf | 4390 | cpustride *= levelspread[i]; |
64db4cff | 4391 | rnp = rsp->level[i]; |
199977bf | 4392 | for (j = 0; j < levelcnt[i]; j++, rnp++) { |
1304afb2 | 4393 | raw_spin_lock_init(&rnp->lock); |
b6407e86 PM |
4394 | lockdep_set_class_and_name(&rnp->lock, |
4395 | &rcu_node_class[i], buf[i]); | |
394f2769 PM |
4396 | raw_spin_lock_init(&rnp->fqslock); |
4397 | lockdep_set_class_and_name(&rnp->fqslock, | |
4398 | &rcu_fqs_class[i], fqs[i]); | |
25d30cf4 PM |
4399 | rnp->gpnum = rsp->gpnum; |
4400 | rnp->completed = rsp->completed; | |
64db4cff PM |
4401 | rnp->qsmask = 0; |
4402 | rnp->qsmaskinit = 0; | |
4403 | rnp->grplo = j * cpustride; | |
4404 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
4405 | if (rnp->grphi >= nr_cpu_ids) |
4406 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
4407 | if (i == 0) { |
4408 | rnp->grpnum = 0; | |
4409 | rnp->grpmask = 0; | |
4410 | rnp->parent = NULL; | |
4411 | } else { | |
199977bf | 4412 | rnp->grpnum = j % levelspread[i - 1]; |
64db4cff PM |
4413 | rnp->grpmask = 1UL << rnp->grpnum; |
4414 | rnp->parent = rsp->level[i - 1] + | |
199977bf | 4415 | j / levelspread[i - 1]; |
64db4cff PM |
4416 | } |
4417 | rnp->level = i; | |
12f5f524 | 4418 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 4419 | rcu_init_one_nocb(rnp); |
385b73c0 | 4420 | mutex_init(&rnp->exp_funnel_mutex); |
83c2c735 PM |
4421 | lockdep_set_class_and_name(&rnp->exp_funnel_mutex, |
4422 | &rcu_exp_class[i], exp[i]); | |
64db4cff PM |
4423 | } |
4424 | } | |
0c34029a | 4425 | |
b3dbec76 | 4426 | init_waitqueue_head(&rsp->gp_wq); |
f4ecea30 | 4427 | init_waitqueue_head(&rsp->expedited_wq); |
f885b7f2 | 4428 | rnp = rsp->level[rcu_num_lvls - 1]; |
0c34029a | 4429 | for_each_possible_cpu(i) { |
4a90a068 | 4430 | while (i > rnp->grphi) |
0c34029a | 4431 | rnp++; |
394f99a9 | 4432 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
0c34029a LJ |
4433 | rcu_boot_init_percpu_data(i, rsp); |
4434 | } | |
6ce75a23 | 4435 | list_add(&rsp->flavors, &rcu_struct_flavors); |
64db4cff PM |
4436 | } |
4437 | ||
f885b7f2 PM |
4438 | /* |
4439 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 4440 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
4441 | * the ->node array in the rcu_state structure. |
4442 | */ | |
4443 | static void __init rcu_init_geometry(void) | |
4444 | { | |
026ad283 | 4445 | ulong d; |
f885b7f2 | 4446 | int i; |
05b84aec | 4447 | int rcu_capacity[RCU_NUM_LVLS]; |
f885b7f2 | 4448 | |
026ad283 PM |
4449 | /* |
4450 | * Initialize any unspecified boot parameters. | |
4451 | * The default values of jiffies_till_first_fqs and | |
4452 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
4453 | * value, which is a function of HZ, then adding one for each | |
4454 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
4455 | */ | |
4456 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
4457 | if (jiffies_till_first_fqs == ULONG_MAX) | |
4458 | jiffies_till_first_fqs = d; | |
4459 | if (jiffies_till_next_fqs == ULONG_MAX) | |
4460 | jiffies_till_next_fqs = d; | |
4461 | ||
f885b7f2 | 4462 | /* If the compile-time values are accurate, just leave. */ |
47d631af | 4463 | if (rcu_fanout_leaf == RCU_FANOUT_LEAF && |
b17c7035 | 4464 | nr_cpu_ids == NR_CPUS) |
f885b7f2 | 4465 | return; |
39479098 PM |
4466 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", |
4467 | rcu_fanout_leaf, nr_cpu_ids); | |
f885b7f2 | 4468 | |
f885b7f2 | 4469 | /* |
ee968ac6 PM |
4470 | * The boot-time rcu_fanout_leaf parameter must be at least two |
4471 | * and cannot exceed the number of bits in the rcu_node masks. | |
4472 | * Complain and fall back to the compile-time values if this | |
4473 | * limit is exceeded. | |
f885b7f2 | 4474 | */ |
ee968ac6 | 4475 | if (rcu_fanout_leaf < 2 || |
75cf15a4 | 4476 | rcu_fanout_leaf > sizeof(unsigned long) * 8) { |
13bd6494 | 4477 | rcu_fanout_leaf = RCU_FANOUT_LEAF; |
f885b7f2 PM |
4478 | WARN_ON(1); |
4479 | return; | |
4480 | } | |
4481 | ||
f885b7f2 PM |
4482 | /* |
4483 | * Compute number of nodes that can be handled an rcu_node tree | |
9618138b | 4484 | * with the given number of levels. |
f885b7f2 | 4485 | */ |
9618138b | 4486 | rcu_capacity[0] = rcu_fanout_leaf; |
05b84aec | 4487 | for (i = 1; i < RCU_NUM_LVLS; i++) |
05c5df31 | 4488 | rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT; |
f885b7f2 PM |
4489 | |
4490 | /* | |
75cf15a4 | 4491 | * The tree must be able to accommodate the configured number of CPUs. |
ee968ac6 | 4492 | * If this limit is exceeded, fall back to the compile-time values. |
f885b7f2 | 4493 | */ |
ee968ac6 PM |
4494 | if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) { |
4495 | rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
4496 | WARN_ON(1); | |
4497 | return; | |
4498 | } | |
f885b7f2 | 4499 | |
679f9858 | 4500 | /* Calculate the number of levels in the tree. */ |
9618138b | 4501 | for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) { |
679f9858 | 4502 | } |
9618138b | 4503 | rcu_num_lvls = i + 1; |
679f9858 | 4504 | |
f885b7f2 | 4505 | /* Calculate the number of rcu_nodes at each level of the tree. */ |
679f9858 | 4506 | for (i = 0; i < rcu_num_lvls; i++) { |
9618138b | 4507 | int cap = rcu_capacity[(rcu_num_lvls - 1) - i]; |
679f9858 AG |
4508 | num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap); |
4509 | } | |
f885b7f2 PM |
4510 | |
4511 | /* Calculate the total number of rcu_node structures. */ | |
4512 | rcu_num_nodes = 0; | |
679f9858 | 4513 | for (i = 0; i < rcu_num_lvls; i++) |
f885b7f2 | 4514 | rcu_num_nodes += num_rcu_lvl[i]; |
f885b7f2 PM |
4515 | } |
4516 | ||
a3dc2948 PM |
4517 | /* |
4518 | * Dump out the structure of the rcu_node combining tree associated | |
4519 | * with the rcu_state structure referenced by rsp. | |
4520 | */ | |
4521 | static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp) | |
4522 | { | |
4523 | int level = 0; | |
4524 | struct rcu_node *rnp; | |
4525 | ||
4526 | pr_info("rcu_node tree layout dump\n"); | |
4527 | pr_info(" "); | |
4528 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
4529 | if (rnp->level != level) { | |
4530 | pr_cont("\n"); | |
4531 | pr_info(" "); | |
4532 | level = rnp->level; | |
4533 | } | |
4534 | pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum); | |
4535 | } | |
4536 | pr_cont("\n"); | |
4537 | } | |
4538 | ||
9f680ab4 | 4539 | void __init rcu_init(void) |
64db4cff | 4540 | { |
017c4261 | 4541 | int cpu; |
9f680ab4 | 4542 | |
47627678 PM |
4543 | rcu_early_boot_tests(); |
4544 | ||
f41d911f | 4545 | rcu_bootup_announce(); |
f885b7f2 | 4546 | rcu_init_geometry(); |
a87f203e PM |
4547 | rcu_init_one(&rcu_bh_state); |
4548 | rcu_init_one(&rcu_sched_state); | |
a3dc2948 PM |
4549 | if (dump_tree) |
4550 | rcu_dump_rcu_node_tree(&rcu_sched_state); | |
f41d911f | 4551 | __rcu_init_preempt(); |
b5b39360 | 4552 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
9f680ab4 PM |
4553 | |
4554 | /* | |
4555 | * We don't need protection against CPU-hotplug here because | |
4556 | * this is called early in boot, before either interrupts | |
4557 | * or the scheduler are operational. | |
4558 | */ | |
4559 | cpu_notifier(rcu_cpu_notify, 0); | |
d1d74d14 | 4560 | pm_notifier(rcu_pm_notify, 0); |
017c4261 PM |
4561 | for_each_online_cpu(cpu) |
4562 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
64db4cff PM |
4563 | } |
4564 | ||
4102adab | 4565 | #include "tree_plugin.h" |