]>
Commit | Line | Data |
---|---|---|
22e40925 | 1 | // SPDX-License-Identifier: GPL-2.0+ |
64db4cff | 2 | /* |
65bb0dc4 | 3 | * Read-Copy Update mechanism for mutual exclusion (tree-based version) |
64db4cff | 4 | * |
64db4cff PM |
5 | * Copyright IBM Corporation, 2008 |
6 | * | |
7 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> | |
8 | * Manfred Spraul <manfred@colorfullife.com> | |
65bb0dc4 | 9 | * Paul E. McKenney <paulmck@linux.ibm.com> |
64db4cff | 10 | * |
22e40925 | 11 | * Based on the original work by Paul McKenney <paulmck@linux.ibm.com> |
64db4cff PM |
12 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. |
13 | * | |
14 | * For detailed explanation of Read-Copy Update mechanism see - | |
a71fca58 | 15 | * Documentation/RCU |
64db4cff | 16 | */ |
a7538352 JP |
17 | |
18 | #define pr_fmt(fmt) "rcu: " fmt | |
19 | ||
64db4cff PM |
20 | #include <linux/types.h> |
21 | #include <linux/kernel.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/spinlock.h> | |
24 | #include <linux/smp.h> | |
f9411ebe | 25 | #include <linux/rcupdate_wait.h> |
64db4cff PM |
26 | #include <linux/interrupt.h> |
27 | #include <linux/sched.h> | |
b17b0153 | 28 | #include <linux/sched/debug.h> |
c1dc0b9c | 29 | #include <linux/nmi.h> |
8826f3b0 | 30 | #include <linux/atomic.h> |
64db4cff | 31 | #include <linux/bitops.h> |
9984de1a | 32 | #include <linux/export.h> |
64db4cff PM |
33 | #include <linux/completion.h> |
34 | #include <linux/moduleparam.h> | |
35 | #include <linux/percpu.h> | |
36 | #include <linux/notifier.h> | |
37 | #include <linux/cpu.h> | |
38 | #include <linux/mutex.h> | |
39 | #include <linux/time.h> | |
bbad9379 | 40 | #include <linux/kernel_stat.h> |
a26ac245 PM |
41 | #include <linux/wait.h> |
42 | #include <linux/kthread.h> | |
ae7e81c0 | 43 | #include <uapi/linux/sched/types.h> |
268bb0ce | 44 | #include <linux/prefetch.h> |
3d3b7db0 | 45 | #include <linux/delay.h> |
661a85dc | 46 | #include <linux/random.h> |
af658dca | 47 | #include <linux/trace_events.h> |
d1d74d14 | 48 | #include <linux/suspend.h> |
a278d471 | 49 | #include <linux/ftrace.h> |
d3052109 | 50 | #include <linux/tick.h> |
2ccaff10 | 51 | #include <linux/sysrq.h> |
c13324a5 | 52 | #include <linux/kprobes.h> |
48d07c04 SAS |
53 | #include <linux/gfp.h> |
54 | #include <linux/oom.h> | |
55 | #include <linux/smpboot.h> | |
56 | #include <linux/jiffies.h> | |
77a40f97 | 57 | #include <linux/slab.h> |
48d07c04 | 58 | #include <linux/sched/isolation.h> |
cfcdef5e | 59 | #include <linux/sched/clock.h> |
48d07c04 | 60 | #include "../time/tick-internal.h" |
64db4cff | 61 | |
4102adab | 62 | #include "tree.h" |
29c00b4a | 63 | #include "rcu.h" |
9f77da9f | 64 | |
4102adab PM |
65 | #ifdef MODULE_PARAM_PREFIX |
66 | #undef MODULE_PARAM_PREFIX | |
67 | #endif | |
68 | #define MODULE_PARAM_PREFIX "rcutree." | |
69 | ||
35315936 PM |
70 | #ifndef data_race |
71 | #define data_race(expr) \ | |
72 | ({ \ | |
73 | expr; \ | |
74 | }) | |
75 | #endif | |
76 | #ifndef ASSERT_EXCLUSIVE_WRITER | |
77 | #define ASSERT_EXCLUSIVE_WRITER(var) do { } while (0) | |
78 | #endif | |
79 | #ifndef ASSERT_EXCLUSIVE_ACCESS | |
80 | #define ASSERT_EXCLUSIVE_ACCESS(var) do { } while (0) | |
81 | #endif | |
82 | ||
64db4cff PM |
83 | /* Data structures. */ |
84 | ||
f7f7bac9 | 85 | /* |
dc5a4f29 PM |
86 | * Steal a bit from the bottom of ->dynticks for idle entry/exit |
87 | * control. Initially this is for TLB flushing. | |
f7f7bac9 | 88 | */ |
dc5a4f29 PM |
89 | #define RCU_DYNTICK_CTRL_MASK 0x1 |
90 | #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1) | |
a8a29b3b | 91 | |
4c5273bf PM |
92 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = { |
93 | .dynticks_nesting = 1, | |
94 | .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE, | |
dc5a4f29 | 95 | .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR), |
4c5273bf | 96 | }; |
c30fe541 | 97 | static struct rcu_state rcu_state = { |
358be2d3 | 98 | .level = { &rcu_state.node[0] }, |
358be2d3 PM |
99 | .gp_state = RCU_GP_IDLE, |
100 | .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT, | |
101 | .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex), | |
102 | .name = RCU_NAME, | |
103 | .abbr = RCU_ABBR, | |
104 | .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex), | |
105 | .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex), | |
894d45bb | 106 | .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock), |
358be2d3 | 107 | }; |
27f4d280 | 108 | |
a3dc2948 PM |
109 | /* Dump rcu_node combining tree at boot to verify correct setup. */ |
110 | static bool dump_tree; | |
111 | module_param(dump_tree, bool, 0444); | |
48d07c04 | 112 | /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */ |
da44cd6c | 113 | static bool use_softirq = true; |
48d07c04 | 114 | module_param(use_softirq, bool, 0444); |
7fa27001 PM |
115 | /* Control rcu_node-tree auto-balancing at boot time. */ |
116 | static bool rcu_fanout_exact; | |
117 | module_param(rcu_fanout_exact, bool, 0444); | |
47d631af PM |
118 | /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */ |
119 | static int rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
7e5c2dfb | 120 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 | 121 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
cb007102 | 122 | /* Number of rcu_nodes at specified level. */ |
e95d68d2 | 123 | int num_rcu_lvl[] = NUM_RCU_LVL_INIT; |
f885b7f2 PM |
124 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ |
125 | ||
b0d30417 | 126 | /* |
52d7e48b PM |
127 | * The rcu_scheduler_active variable is initialized to the value |
128 | * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the | |
129 | * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE, | |
130 | * RCU can assume that there is but one task, allowing RCU to (for example) | |
0d95092c | 131 | * optimize synchronize_rcu() to a simple barrier(). When this variable |
52d7e48b PM |
132 | * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required |
133 | * to detect real grace periods. This variable is also used to suppress | |
134 | * boot-time false positives from lockdep-RCU error checking. Finally, it | |
135 | * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU | |
136 | * is fully initialized, including all of its kthreads having been spawned. | |
b0d30417 | 137 | */ |
bbad9379 PM |
138 | int rcu_scheduler_active __read_mostly; |
139 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
140 | ||
b0d30417 PM |
141 | /* |
142 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
143 | * during the early_initcall() processing, which is after the scheduler | |
144 | * is capable of creating new tasks. So RCU processing (for example, | |
145 | * creating tasks for RCU priority boosting) must be delayed until after | |
146 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
147 | * currently delay invocation of any RCU callbacks until after this point. | |
148 | * | |
149 | * It might later prove better for people registering RCU callbacks during | |
150 | * early boot to take responsibility for these callbacks, but one step at | |
151 | * a time. | |
152 | */ | |
153 | static int rcu_scheduler_fully_active __read_mostly; | |
154 | ||
b50912d0 PM |
155 | static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp, |
156 | unsigned long gps, unsigned long flags); | |
0aa04b05 PM |
157 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); |
158 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); | |
5d01bbd1 | 159 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 | 160 | static void invoke_rcu_core(void); |
63d4c8c9 | 161 | static void rcu_report_exp_rdp(struct rcu_data *rdp); |
3549c2bc | 162 | static void sync_sched_exp_online_cleanup(int cpu); |
b2b00ddf | 163 | static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp); |
a26ac245 | 164 | |
a94844b2 | 165 | /* rcuc/rcub kthread realtime priority */ |
26730f55 | 166 | static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0; |
3ffe3d1a | 167 | module_param(kthread_prio, int, 0444); |
a94844b2 | 168 | |
8d7dc928 | 169 | /* Delay in jiffies for grace-period initialization delays, debug only. */ |
0f41c0dd | 170 | |
90040c9e PM |
171 | static int gp_preinit_delay; |
172 | module_param(gp_preinit_delay, int, 0444); | |
173 | static int gp_init_delay; | |
174 | module_param(gp_init_delay, int, 0444); | |
175 | static int gp_cleanup_delay; | |
176 | module_param(gp_cleanup_delay, int, 0444); | |
0f41c0dd | 177 | |
4cf439a2 | 178 | /* Retrieve RCU kthreads priority for rcutorture */ |
4babd855 JFG |
179 | int rcu_get_gp_kthreads_prio(void) |
180 | { | |
181 | return kthread_prio; | |
182 | } | |
183 | EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio); | |
184 | ||
eab128e8 PM |
185 | /* |
186 | * Number of grace periods between delays, normalized by the duration of | |
bfd090be | 187 | * the delay. The longer the delay, the more the grace periods between |
eab128e8 PM |
188 | * each delay. The reason for this normalization is that it means that, |
189 | * for non-zero delays, the overall slowdown of grace periods is constant | |
190 | * regardless of the duration of the delay. This arrangement balances | |
191 | * the need for long delays to increase some race probabilities with the | |
192 | * need for fast grace periods to increase other race probabilities. | |
193 | */ | |
194 | #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */ | |
37745d28 | 195 | |
0aa04b05 PM |
196 | /* |
197 | * Compute the mask of online CPUs for the specified rcu_node structure. | |
198 | * This will not be stable unless the rcu_node structure's ->lock is | |
199 | * held, but the bit corresponding to the current CPU will be stable | |
200 | * in most contexts. | |
201 | */ | |
c30fe541 | 202 | static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) |
0aa04b05 | 203 | { |
7d0ae808 | 204 | return READ_ONCE(rnp->qsmaskinitnext); |
0aa04b05 PM |
205 | } |
206 | ||
fc2219d4 | 207 | /* |
7d0ae808 | 208 | * Return true if an RCU grace period is in progress. The READ_ONCE()s |
fc2219d4 PM |
209 | * permit this function to be invoked without holding the root rcu_node |
210 | * structure's ->lock, but of course results can be subject to change. | |
211 | */ | |
de8e8730 | 212 | static int rcu_gp_in_progress(void) |
fc2219d4 | 213 | { |
de8e8730 | 214 | return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq)); |
b1f77b05 IM |
215 | } |
216 | ||
903ee83d PM |
217 | /* |
218 | * Return the number of callbacks queued on the specified CPU. | |
219 | * Handles both the nocbs and normal cases. | |
220 | */ | |
221 | static long rcu_get_n_cbs_cpu(int cpu) | |
222 | { | |
223 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
224 | ||
c035280f | 225 | if (rcu_segcblist_is_enabled(&rdp->cblist)) |
903ee83d | 226 | return rcu_segcblist_n_cbs(&rdp->cblist); |
c035280f | 227 | return 0; |
903ee83d PM |
228 | } |
229 | ||
d28139c4 | 230 | void rcu_softirq_qs(void) |
b1f77b05 | 231 | { |
45975c7d | 232 | rcu_qs(); |
d28139c4 | 233 | rcu_preempt_deferred_qs(current); |
b1f77b05 | 234 | } |
64db4cff | 235 | |
2625d469 PM |
236 | /* |
237 | * Record entry into an extended quiescent state. This is only to be | |
ac3caf82 PM |
238 | * called when not already in an extended quiescent state, that is, |
239 | * RCU is watching prior to the call to this function and is no longer | |
240 | * watching upon return. | |
2625d469 | 241 | */ |
ff5c4f5c | 242 | static noinstr void rcu_dynticks_eqs_enter(void) |
2625d469 | 243 | { |
dc5a4f29 | 244 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
b8c17e66 | 245 | int seq; |
2625d469 PM |
246 | |
247 | /* | |
b8c17e66 | 248 | * CPUs seeing atomic_add_return() must see prior RCU read-side |
2625d469 PM |
249 | * critical sections, and we also must force ordering with the |
250 | * next idle sojourn. | |
251 | */ | |
7d0c9c50 | 252 | rcu_dynticks_task_trace_enter(); // Before ->dynticks update! |
dc5a4f29 | 253 | seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks); |
ac3caf82 | 254 | // RCU is no longer watching. Better be in extended quiescent state! |
b8c17e66 PM |
255 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
256 | (seq & RCU_DYNTICK_CTRL_CTR)); | |
257 | /* Better not have special action (TLB flush) pending! */ | |
258 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
259 | (seq & RCU_DYNTICK_CTRL_MASK)); | |
2625d469 PM |
260 | } |
261 | ||
262 | /* | |
263 | * Record exit from an extended quiescent state. This is only to be | |
ac3caf82 PM |
264 | * called from an extended quiescent state, that is, RCU is not watching |
265 | * prior to the call to this function and is watching upon return. | |
2625d469 | 266 | */ |
ff5c4f5c | 267 | static noinstr void rcu_dynticks_eqs_exit(void) |
2625d469 | 268 | { |
dc5a4f29 | 269 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
b8c17e66 | 270 | int seq; |
2625d469 PM |
271 | |
272 | /* | |
b8c17e66 | 273 | * CPUs seeing atomic_add_return() must see prior idle sojourns, |
2625d469 PM |
274 | * and we also must force ordering with the next RCU read-side |
275 | * critical section. | |
276 | */ | |
dc5a4f29 | 277 | seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks); |
ac3caf82 | 278 | // RCU is now watching. Better not be in an extended quiescent state! |
7d0c9c50 | 279 | rcu_dynticks_task_trace_exit(); // After ->dynticks update! |
b8c17e66 PM |
280 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
281 | !(seq & RCU_DYNTICK_CTRL_CTR)); | |
282 | if (seq & RCU_DYNTICK_CTRL_MASK) { | |
dc5a4f29 | 283 | atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks); |
b8c17e66 | 284 | smp_mb__after_atomic(); /* _exit after clearing mask. */ |
b8c17e66 | 285 | } |
2625d469 PM |
286 | } |
287 | ||
288 | /* | |
289 | * Reset the current CPU's ->dynticks counter to indicate that the | |
290 | * newly onlined CPU is no longer in an extended quiescent state. | |
291 | * This will either leave the counter unchanged, or increment it | |
292 | * to the next non-quiescent value. | |
293 | * | |
294 | * The non-atomic test/increment sequence works because the upper bits | |
295 | * of the ->dynticks counter are manipulated only by the corresponding CPU, | |
296 | * or when the corresponding CPU is offline. | |
297 | */ | |
298 | static void rcu_dynticks_eqs_online(void) | |
299 | { | |
dc5a4f29 | 300 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
2625d469 | 301 | |
dc5a4f29 | 302 | if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR) |
2625d469 | 303 | return; |
dc5a4f29 | 304 | atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks); |
2625d469 PM |
305 | } |
306 | ||
02a5c550 PM |
307 | /* |
308 | * Is the current CPU in an extended quiescent state? | |
309 | * | |
310 | * No ordering, as we are sampling CPU-local information. | |
311 | */ | |
ff5c4f5c | 312 | static __always_inline bool rcu_dynticks_curr_cpu_in_eqs(void) |
02a5c550 | 313 | { |
dc5a4f29 | 314 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
02a5c550 | 315 | |
dc5a4f29 | 316 | return !(atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR); |
02a5c550 PM |
317 | } |
318 | ||
8b2f63ab PM |
319 | /* |
320 | * Snapshot the ->dynticks counter with full ordering so as to allow | |
321 | * stable comparison of this counter with past and future snapshots. | |
322 | */ | |
c30fe541 | 323 | static int rcu_dynticks_snap(struct rcu_data *rdp) |
8b2f63ab | 324 | { |
dc5a4f29 | 325 | int snap = atomic_add_return(0, &rdp->dynticks); |
8b2f63ab | 326 | |
b8c17e66 | 327 | return snap & ~RCU_DYNTICK_CTRL_MASK; |
8b2f63ab PM |
328 | } |
329 | ||
02a5c550 PM |
330 | /* |
331 | * Return true if the snapshot returned from rcu_dynticks_snap() | |
332 | * indicates that RCU is in an extended quiescent state. | |
333 | */ | |
334 | static bool rcu_dynticks_in_eqs(int snap) | |
335 | { | |
b8c17e66 | 336 | return !(snap & RCU_DYNTICK_CTRL_CTR); |
02a5c550 PM |
337 | } |
338 | ||
339 | /* | |
dc5a4f29 | 340 | * Return true if the CPU corresponding to the specified rcu_data |
02a5c550 PM |
341 | * structure has spent some time in an extended quiescent state since |
342 | * rcu_dynticks_snap() returned the specified snapshot. | |
343 | */ | |
dc5a4f29 | 344 | static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap) |
02a5c550 | 345 | { |
dc5a4f29 | 346 | return snap != rcu_dynticks_snap(rdp); |
02a5c550 PM |
347 | } |
348 | ||
7d0c9c50 PM |
349 | /* |
350 | * Return true if the referenced integer is zero while the specified | |
351 | * CPU remains within a single extended quiescent state. | |
352 | */ | |
353 | bool rcu_dynticks_zero_in_eqs(int cpu, int *vp) | |
354 | { | |
355 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
356 | int snap; | |
357 | ||
358 | // If not quiescent, force back to earlier extended quiescent state. | |
359 | snap = atomic_read(&rdp->dynticks) & ~(RCU_DYNTICK_CTRL_MASK | | |
360 | RCU_DYNTICK_CTRL_CTR); | |
361 | ||
362 | smp_rmb(); // Order ->dynticks and *vp reads. | |
363 | if (READ_ONCE(*vp)) | |
364 | return false; // Non-zero, so report failure; | |
365 | smp_rmb(); // Order *vp read and ->dynticks re-read. | |
366 | ||
367 | // If still in the same extended quiescent state, we are good! | |
368 | return snap == (atomic_read(&rdp->dynticks) & ~RCU_DYNTICK_CTRL_MASK); | |
369 | } | |
370 | ||
b8c17e66 PM |
371 | /* |
372 | * Set the special (bottom) bit of the specified CPU so that it | |
373 | * will take special action (such as flushing its TLB) on the | |
374 | * next exit from an extended quiescent state. Returns true if | |
375 | * the bit was successfully set, or false if the CPU was not in | |
376 | * an extended quiescent state. | |
377 | */ | |
378 | bool rcu_eqs_special_set(int cpu) | |
379 | { | |
380 | int old; | |
381 | int new; | |
faa059c3 | 382 | int new_old; |
dc5a4f29 | 383 | struct rcu_data *rdp = &per_cpu(rcu_data, cpu); |
b8c17e66 | 384 | |
faa059c3 | 385 | new_old = atomic_read(&rdp->dynticks); |
b8c17e66 | 386 | do { |
faa059c3 | 387 | old = new_old; |
b8c17e66 PM |
388 | if (old & RCU_DYNTICK_CTRL_CTR) |
389 | return false; | |
390 | new = old | RCU_DYNTICK_CTRL_MASK; | |
faa059c3 PM |
391 | new_old = atomic_cmpxchg(&rdp->dynticks, old, new); |
392 | } while (new_old != old); | |
b8c17e66 | 393 | return true; |
6563de9d | 394 | } |
5cd37193 | 395 | |
4a81e832 PM |
396 | /* |
397 | * Let the RCU core know that this CPU has gone through the scheduler, | |
398 | * which is a quiescent state. This is called when the need for a | |
399 | * quiescent state is urgent, so we burn an atomic operation and full | |
400 | * memory barriers to let the RCU core know about it, regardless of what | |
401 | * this CPU might (or might not) do in the near future. | |
402 | * | |
0f9be8ca | 403 | * We inform the RCU core by emulating a zero-duration dyntick-idle period. |
46a5d164 | 404 | * |
3b57a399 | 405 | * The caller must have disabled interrupts and must not be idle. |
4a81e832 | 406 | */ |
366237e7 | 407 | void rcu_momentary_dyntick_idle(void) |
4a81e832 | 408 | { |
3b57a399 PM |
409 | int special; |
410 | ||
2dba13f0 | 411 | raw_cpu_write(rcu_data.rcu_need_heavy_qs, false); |
dc5a4f29 PM |
412 | special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR, |
413 | &this_cpu_ptr(&rcu_data)->dynticks); | |
3b57a399 PM |
414 | /* It is illegal to call this from idle state. */ |
415 | WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR)); | |
3e310098 | 416 | rcu_preempt_deferred_qs(current); |
4a81e832 | 417 | } |
79ba7ff5 | 418 | EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle); |
4a81e832 | 419 | |
45975c7d | 420 | /** |
806f04e9 | 421 | * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle |
bb73c52b | 422 | * |
eddded80 | 423 | * If the current CPU is idle and running at a first-level (not nested) |
806f04e9 PZ |
424 | * interrupt, or directly, from idle, return true. |
425 | * | |
426 | * The caller must have at least disabled IRQs. | |
5cd37193 | 427 | */ |
45975c7d | 428 | static int rcu_is_cpu_rrupt_from_idle(void) |
5cd37193 | 429 | { |
806f04e9 PZ |
430 | long nesting; |
431 | ||
432 | /* | |
433 | * Usually called from the tick; but also used from smp_function_call() | |
434 | * for expedited grace periods. This latter can result in running from | |
435 | * the idle task, instead of an actual IPI. | |
436 | */ | |
437 | lockdep_assert_irqs_disabled(); | |
eddded80 JFG |
438 | |
439 | /* Check for counter underflows */ | |
440 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0, | |
441 | "RCU dynticks_nesting counter underflow!"); | |
442 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0, | |
443 | "RCU dynticks_nmi_nesting counter underflow/zero!"); | |
444 | ||
445 | /* Are we at first interrupt nesting level? */ | |
806f04e9 PZ |
446 | nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting); |
447 | if (nesting > 1) | |
eddded80 JFG |
448 | return false; |
449 | ||
806f04e9 PZ |
450 | /* |
451 | * If we're not in an interrupt, we must be in the idle task! | |
452 | */ | |
453 | WARN_ON_ONCE(!nesting && !is_idle_task(current)); | |
454 | ||
eddded80 JFG |
455 | /* Does CPU appear to be idle from an RCU standpoint? */ |
456 | return __this_cpu_read(rcu_data.dynticks_nesting) == 0; | |
5cd37193 | 457 | } |
5cd37193 | 458 | |
d5a9a8c3 PM |
459 | #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch ... */ |
460 | #define DEFAULT_MAX_RCU_BLIMIT 10000 /* ... even during callback flood. */ | |
17c7798b PM |
461 | static long blimit = DEFAULT_RCU_BLIMIT; |
462 | #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */ | |
463 | static long qhimark = DEFAULT_RCU_QHIMARK; | |
464 | #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */ | |
465 | static long qlowmark = DEFAULT_RCU_QLOMARK; | |
b2b00ddf PM |
466 | #define DEFAULT_RCU_QOVLD_MULT 2 |
467 | #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK) | |
468 | static long qovld = DEFAULT_RCU_QOVLD; /* If this many pending, hammer QS. */ | |
469 | static long qovld_calc = -1; /* No pre-initialization lock acquisitions! */ | |
64db4cff | 470 | |
878d7439 ED |
471 | module_param(blimit, long, 0444); |
472 | module_param(qhimark, long, 0444); | |
473 | module_param(qlowmark, long, 0444); | |
b2b00ddf | 474 | module_param(qovld, long, 0444); |
3d76c082 | 475 | |
026ad283 PM |
476 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
477 | static ulong jiffies_till_next_fqs = ULONG_MAX; | |
8c7c4829 | 478 | static bool rcu_kick_kthreads; |
cfcdef5e ED |
479 | static int rcu_divisor = 7; |
480 | module_param(rcu_divisor, int, 0644); | |
481 | ||
482 | /* Force an exit from rcu_do_batch() after 3 milliseconds. */ | |
483 | static long rcu_resched_ns = 3 * NSEC_PER_MSEC; | |
484 | module_param(rcu_resched_ns, long, 0644); | |
d40011f6 | 485 | |
c06aed0e PM |
486 | /* |
487 | * How long the grace period must be before we start recruiting | |
488 | * quiescent-state help from rcu_note_context_switch(). | |
489 | */ | |
490 | static ulong jiffies_till_sched_qs = ULONG_MAX; | |
491 | module_param(jiffies_till_sched_qs, ulong, 0444); | |
85f2b60c | 492 | static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */ |
c06aed0e PM |
493 | module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */ |
494 | ||
495 | /* | |
496 | * Make sure that we give the grace-period kthread time to detect any | |
497 | * idle CPUs before taking active measures to force quiescent states. | |
498 | * However, don't go below 100 milliseconds, adjusted upwards for really | |
499 | * large systems. | |
500 | */ | |
501 | static void adjust_jiffies_till_sched_qs(void) | |
502 | { | |
503 | unsigned long j; | |
504 | ||
505 | /* If jiffies_till_sched_qs was specified, respect the request. */ | |
506 | if (jiffies_till_sched_qs != ULONG_MAX) { | |
507 | WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs); | |
508 | return; | |
509 | } | |
85f2b60c | 510 | /* Otherwise, set to third fqs scan, but bound below on large system. */ |
c06aed0e PM |
511 | j = READ_ONCE(jiffies_till_first_fqs) + |
512 | 2 * READ_ONCE(jiffies_till_next_fqs); | |
513 | if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV) | |
514 | j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
515 | pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j); | |
516 | WRITE_ONCE(jiffies_to_sched_qs, j); | |
517 | } | |
518 | ||
67abb96c BP |
519 | static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp) |
520 | { | |
521 | ulong j; | |
522 | int ret = kstrtoul(val, 0, &j); | |
523 | ||
c06aed0e | 524 | if (!ret) { |
67abb96c | 525 | WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j); |
c06aed0e PM |
526 | adjust_jiffies_till_sched_qs(); |
527 | } | |
67abb96c BP |
528 | return ret; |
529 | } | |
530 | ||
531 | static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp) | |
532 | { | |
533 | ulong j; | |
534 | int ret = kstrtoul(val, 0, &j); | |
535 | ||
c06aed0e | 536 | if (!ret) { |
67abb96c | 537 | WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1)); |
c06aed0e PM |
538 | adjust_jiffies_till_sched_qs(); |
539 | } | |
67abb96c BP |
540 | return ret; |
541 | } | |
542 | ||
543 | static struct kernel_param_ops first_fqs_jiffies_ops = { | |
544 | .set = param_set_first_fqs_jiffies, | |
545 | .get = param_get_ulong, | |
546 | }; | |
547 | ||
548 | static struct kernel_param_ops next_fqs_jiffies_ops = { | |
549 | .set = param_set_next_fqs_jiffies, | |
550 | .get = param_get_ulong, | |
551 | }; | |
552 | ||
553 | module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644); | |
554 | module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644); | |
8c7c4829 | 555 | module_param(rcu_kick_kthreads, bool, 0644); |
d40011f6 | 556 | |
8ff0b907 | 557 | static void force_qs_rnp(int (*f)(struct rcu_data *rdp)); |
dd7dafd1 | 558 | static int rcu_pending(int user); |
64db4cff PM |
559 | |
560 | /* | |
17ef2fe9 | 561 | * Return the number of RCU GPs completed thus far for debug & stats. |
64db4cff | 562 | */ |
17ef2fe9 | 563 | unsigned long rcu_get_gp_seq(void) |
917963d0 | 564 | { |
16fc9c60 | 565 | return READ_ONCE(rcu_state.gp_seq); |
917963d0 | 566 | } |
17ef2fe9 | 567 | EXPORT_SYMBOL_GPL(rcu_get_gp_seq); |
917963d0 | 568 | |
291783b8 PM |
569 | /* |
570 | * Return the number of RCU expedited batches completed thus far for | |
571 | * debug & stats. Odd numbers mean that a batch is in progress, even | |
572 | * numbers mean idle. The value returned will thus be roughly double | |
573 | * the cumulative batches since boot. | |
574 | */ | |
575 | unsigned long rcu_exp_batches_completed(void) | |
576 | { | |
16fc9c60 | 577 | return rcu_state.expedited_sequence; |
291783b8 PM |
578 | } |
579 | EXPORT_SYMBOL_GPL(rcu_exp_batches_completed); | |
580 | ||
fd897573 PM |
581 | /* |
582 | * Return the root node of the rcu_state structure. | |
583 | */ | |
584 | static struct rcu_node *rcu_get_root(void) | |
585 | { | |
586 | return &rcu_state.node[0]; | |
587 | } | |
588 | ||
ad0dc7f9 PM |
589 | /* |
590 | * Send along grace-period-related data for rcutorture diagnostics. | |
591 | */ | |
592 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
aebc8264 | 593 | unsigned long *gp_seq) |
ad0dc7f9 | 594 | { |
ad0dc7f9 PM |
595 | switch (test_type) { |
596 | case RCU_FLAVOR: | |
f7dd7d44 PM |
597 | *flags = READ_ONCE(rcu_state.gp_flags); |
598 | *gp_seq = rcu_seq_current(&rcu_state.gp_seq); | |
ad0dc7f9 PM |
599 | break; |
600 | default: | |
601 | break; | |
602 | } | |
ad0dc7f9 PM |
603 | } |
604 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
605 | ||
9b2e4f18 | 606 | /* |
215bba9f PM |
607 | * Enter an RCU extended quiescent state, which can be either the |
608 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 609 | * |
215bba9f PM |
610 | * We crowbar the ->dynticks_nmi_nesting field to zero to allow for |
611 | * the possibility of usermode upcalls having messed up our count | |
612 | * of interrupt nesting level during the prior busy period. | |
9b2e4f18 | 613 | */ |
ff5c4f5c | 614 | static noinstr void rcu_eqs_enter(bool user) |
9b2e4f18 | 615 | { |
4c5273bf | 616 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
96d3fd0d | 617 | |
4c5273bf PM |
618 | WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE); |
619 | WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); | |
215bba9f | 620 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
4c5273bf PM |
621 | rdp->dynticks_nesting == 0); |
622 | if (rdp->dynticks_nesting != 1) { | |
ac3caf82 | 623 | // RCU will still be watching, so just do accounting and leave. |
4c5273bf | 624 | rdp->dynticks_nesting--; |
215bba9f | 625 | return; |
9b2e4f18 | 626 | } |
96d3fd0d | 627 | |
b04db8e1 | 628 | lockdep_assert_irqs_disabled(); |
ff5c4f5c | 629 | instrumentation_begin(); |
6cf539a8 | 630 | trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks)); |
e68bbb26 | 631 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current)); |
b97d23c5 PM |
632 | rdp = this_cpu_ptr(&rcu_data); |
633 | do_nocb_deferred_wakeup(rdp); | |
198bbf81 | 634 | rcu_prepare_for_idle(); |
3e310098 | 635 | rcu_preempt_deferred_qs(current); |
ff5c4f5c | 636 | instrumentation_end(); |
4c5273bf | 637 | WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */ |
ac3caf82 | 638 | // RCU is watching here ... |
844ccdd7 | 639 | rcu_dynticks_eqs_enter(); |
ac3caf82 | 640 | // ... but is no longer watching here. |
176f8f7a | 641 | rcu_dynticks_task_enter(); |
64db4cff | 642 | } |
adf5091e FW |
643 | |
644 | /** | |
645 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
646 | * | |
647 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
648 | * read-side critical sections can occur. (Though RCU read-side | |
649 | * critical sections can occur in irq handlers in idle, a possibility | |
650 | * handled by irq_enter() and irq_exit().) | |
651 | * | |
c0da313e PM |
652 | * If you add or remove a call to rcu_idle_enter(), be sure to test with |
653 | * CONFIG_RCU_EQS_DEBUG=y. | |
adf5091e FW |
654 | */ |
655 | void rcu_idle_enter(void) | |
656 | { | |
b04db8e1 | 657 | lockdep_assert_irqs_disabled(); |
cb349ca9 | 658 | rcu_eqs_enter(false); |
adf5091e | 659 | } |
64db4cff | 660 | |
d1ec4c34 | 661 | #ifdef CONFIG_NO_HZ_FULL |
adf5091e FW |
662 | /** |
663 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
664 | * | |
665 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
666 | * is permitted between this call and rcu_user_exit(). This way the | |
667 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
668 | * when the CPU runs in userspace. | |
c0da313e PM |
669 | * |
670 | * If you add or remove a call to rcu_user_enter(), be sure to test with | |
671 | * CONFIG_RCU_EQS_DEBUG=y. | |
adf5091e | 672 | */ |
ff5c4f5c | 673 | noinstr void rcu_user_enter(void) |
adf5091e | 674 | { |
b04db8e1 | 675 | lockdep_assert_irqs_disabled(); |
d4db30af | 676 | rcu_eqs_enter(true); |
adf5091e | 677 | } |
d1ec4c34 | 678 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 679 | |
9ea366f6 PM |
680 | /** |
681 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
682 | * | |
fd581a91 | 683 | * If we are returning from the outermost NMI handler that interrupted an |
dc5a4f29 | 684 | * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting |
fd581a91 PM |
685 | * to let the RCU grace-period handling know that the CPU is back to |
686 | * being RCU-idle. | |
687 | * | |
9ea366f6 | 688 | * If you add or remove a call to rcu_nmi_exit(), be sure to test |
fd581a91 PM |
689 | * with CONFIG_RCU_EQS_DEBUG=y. |
690 | */ | |
9ea366f6 | 691 | noinstr void rcu_nmi_exit(void) |
fd581a91 | 692 | { |
4c5273bf | 693 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
fd581a91 PM |
694 | |
695 | /* | |
696 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. | |
697 | * (We are exiting an NMI handler, so RCU better be paying attention | |
698 | * to us!) | |
699 | */ | |
4c5273bf | 700 | WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0); |
fd581a91 PM |
701 | WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs()); |
702 | ||
703 | /* | |
704 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so | |
705 | * leave it in non-RCU-idle state. | |
706 | */ | |
4c5273bf | 707 | if (rdp->dynticks_nmi_nesting != 1) { |
ff5c4f5c | 708 | instrumentation_begin(); |
6cf539a8 ME |
709 | trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2, |
710 | atomic_read(&rdp->dynticks)); | |
4c5273bf PM |
711 | WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */ |
712 | rdp->dynticks_nmi_nesting - 2); | |
ff5c4f5c | 713 | instrumentation_end(); |
fd581a91 PM |
714 | return; |
715 | } | |
716 | ||
ff5c4f5c | 717 | instrumentation_begin(); |
fd581a91 | 718 | /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ |
6cf539a8 | 719 | trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks)); |
4c5273bf | 720 | WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */ |
cf7614e1 | 721 | |
9ea366f6 | 722 | if (!in_nmi()) |
cf7614e1 | 723 | rcu_prepare_for_idle(); |
ff5c4f5c | 724 | instrumentation_end(); |
cf7614e1 | 725 | |
ac3caf82 | 726 | // RCU is watching here ... |
fd581a91 | 727 | rcu_dynticks_eqs_enter(); |
ac3caf82 | 728 | // ... but is no longer watching here. |
cf7614e1 | 729 | |
9ea366f6 | 730 | if (!in_nmi()) |
cf7614e1 BP |
731 | rcu_dynticks_task_enter(); |
732 | } | |
733 | ||
9b2e4f18 PM |
734 | /** |
735 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
736 | * | |
737 | * Exit from an interrupt handler, which might possibly result in entering | |
738 | * idle mode, in other words, leaving the mode in which read-side critical | |
7c9906ca | 739 | * sections can occur. The caller must have disabled interrupts. |
64db4cff | 740 | * |
9b2e4f18 PM |
741 | * This code assumes that the idle loop never does anything that might |
742 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
58721f5d PM |
743 | * architecture's idle loop violates this assumption, RCU will give you what |
744 | * you deserve, good and hard. But very infrequently and irreproducibly. | |
9b2e4f18 PM |
745 | * |
746 | * Use things like work queues to work around this limitation. | |
747 | * | |
748 | * You have been warned. | |
c0da313e PM |
749 | * |
750 | * If you add or remove a call to rcu_irq_exit(), be sure to test with | |
751 | * CONFIG_RCU_EQS_DEBUG=y. | |
64db4cff | 752 | */ |
ff5c4f5c | 753 | void noinstr rcu_irq_exit(void) |
64db4cff | 754 | { |
b04db8e1 | 755 | lockdep_assert_irqs_disabled(); |
9ea366f6 | 756 | rcu_nmi_exit(); |
7c9906ca PM |
757 | } |
758 | ||
8ae0ae67 TG |
759 | /** |
760 | * rcu_irq_exit_preempt - Inform RCU that current CPU is exiting irq | |
761 | * towards in kernel preemption | |
762 | * | |
763 | * Same as rcu_irq_exit() but has a sanity check that scheduling is safe | |
764 | * from RCU point of view. Invoked from return from interrupt before kernel | |
765 | * preemption. | |
766 | */ | |
767 | void rcu_irq_exit_preempt(void) | |
768 | { | |
769 | lockdep_assert_irqs_disabled(); | |
770 | rcu_nmi_exit(); | |
771 | ||
772 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0, | |
773 | "RCU dynticks_nesting counter underflow/zero!"); | |
774 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) != | |
775 | DYNTICK_IRQ_NONIDLE, | |
776 | "Bad RCU dynticks_nmi_nesting counter\n"); | |
777 | RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(), | |
778 | "RCU in extended quiescent state!"); | |
779 | } | |
780 | ||
07325d4a TG |
781 | #ifdef CONFIG_PROVE_RCU |
782 | /** | |
783 | * rcu_irq_exit_check_preempt - Validate that scheduling is possible | |
784 | */ | |
785 | void rcu_irq_exit_check_preempt(void) | |
786 | { | |
787 | lockdep_assert_irqs_disabled(); | |
788 | ||
789 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0, | |
790 | "RCU dynticks_nesting counter underflow/zero!"); | |
791 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) != | |
792 | DYNTICK_IRQ_NONIDLE, | |
793 | "Bad RCU dynticks_nmi_nesting counter\n"); | |
794 | RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(), | |
795 | "RCU in extended quiescent state!"); | |
796 | } | |
797 | #endif /* #ifdef CONFIG_PROVE_RCU */ | |
798 | ||
7c9906ca PM |
799 | /* |
800 | * Wrapper for rcu_irq_exit() where interrupts are enabled. | |
c0da313e PM |
801 | * |
802 | * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test | |
803 | * with CONFIG_RCU_EQS_DEBUG=y. | |
7c9906ca PM |
804 | */ |
805 | void rcu_irq_exit_irqson(void) | |
806 | { | |
807 | unsigned long flags; | |
808 | ||
809 | local_irq_save(flags); | |
810 | rcu_irq_exit(); | |
9b2e4f18 PM |
811 | local_irq_restore(flags); |
812 | } | |
813 | ||
adf5091e FW |
814 | /* |
815 | * Exit an RCU extended quiescent state, which can be either the | |
816 | * idle loop or adaptive-tickless usermode execution. | |
51a1fd30 PM |
817 | * |
818 | * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to | |
819 | * allow for the possibility of usermode upcalls messing up our count of | |
820 | * interrupt nesting level during the busy period that is just now starting. | |
9b2e4f18 | 821 | */ |
ff5c4f5c | 822 | static void noinstr rcu_eqs_exit(bool user) |
9b2e4f18 | 823 | { |
4c5273bf | 824 | struct rcu_data *rdp; |
84585aa8 | 825 | long oldval; |
9b2e4f18 | 826 | |
b04db8e1 | 827 | lockdep_assert_irqs_disabled(); |
4c5273bf PM |
828 | rdp = this_cpu_ptr(&rcu_data); |
829 | oldval = rdp->dynticks_nesting; | |
1ce46ee5 | 830 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0); |
51a1fd30 | 831 | if (oldval) { |
ac3caf82 | 832 | // RCU was already watching, so just do accounting and leave. |
4c5273bf | 833 | rdp->dynticks_nesting++; |
9dd238e2 | 834 | return; |
3a592405 | 835 | } |
9dd238e2 | 836 | rcu_dynticks_task_exit(); |
ac3caf82 | 837 | // RCU is not watching here ... |
9dd238e2 | 838 | rcu_dynticks_eqs_exit(); |
ac3caf82 | 839 | // ... but is watching here. |
ff5c4f5c | 840 | instrumentation_begin(); |
9dd238e2 | 841 | rcu_cleanup_after_idle(); |
6cf539a8 | 842 | trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks)); |
e68bbb26 | 843 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current)); |
4c5273bf PM |
844 | WRITE_ONCE(rdp->dynticks_nesting, 1); |
845 | WARN_ON_ONCE(rdp->dynticks_nmi_nesting); | |
846 | WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE); | |
ff5c4f5c | 847 | instrumentation_end(); |
9b2e4f18 | 848 | } |
adf5091e FW |
849 | |
850 | /** | |
851 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
852 | * | |
853 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
854 | * read-side critical sections can occur. | |
855 | * | |
c0da313e PM |
856 | * If you add or remove a call to rcu_idle_exit(), be sure to test with |
857 | * CONFIG_RCU_EQS_DEBUG=y. | |
adf5091e FW |
858 | */ |
859 | void rcu_idle_exit(void) | |
860 | { | |
c5d900bf FW |
861 | unsigned long flags; |
862 | ||
863 | local_irq_save(flags); | |
cb349ca9 | 864 | rcu_eqs_exit(false); |
c5d900bf | 865 | local_irq_restore(flags); |
adf5091e | 866 | } |
9b2e4f18 | 867 | |
d1ec4c34 | 868 | #ifdef CONFIG_NO_HZ_FULL |
adf5091e FW |
869 | /** |
870 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
871 | * | |
872 | * Exit RCU idle mode while entering the kernel because it can | |
873 | * run a RCU read side critical section anytime. | |
c0da313e PM |
874 | * |
875 | * If you add or remove a call to rcu_user_exit(), be sure to test with | |
876 | * CONFIG_RCU_EQS_DEBUG=y. | |
adf5091e | 877 | */ |
ff5c4f5c | 878 | void noinstr rcu_user_exit(void) |
adf5091e | 879 | { |
91d1aa43 | 880 | rcu_eqs_exit(1); |
adf5091e | 881 | } |
aaf2bc50 PM |
882 | |
883 | /** | |
884 | * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it. | |
885 | * | |
886 | * The scheduler tick is not normally enabled when CPUs enter the kernel | |
887 | * from nohz_full userspace execution. After all, nohz_full userspace | |
888 | * execution is an RCU quiescent state and the time executing in the kernel | |
889 | * is quite short. Except of course when it isn't. And it is not hard to | |
890 | * cause a large system to spend tens of seconds or even minutes looping | |
891 | * in the kernel, which can cause a number of problems, include RCU CPU | |
892 | * stall warnings. | |
893 | * | |
894 | * Therefore, if a nohz_full CPU fails to report a quiescent state | |
895 | * in a timely manner, the RCU grace-period kthread sets that CPU's | |
896 | * ->rcu_urgent_qs flag with the expectation that the next interrupt or | |
897 | * exception will invoke this function, which will turn on the scheduler | |
898 | * tick, which will enable RCU to detect that CPU's quiescent states, | |
899 | * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels. | |
900 | * The tick will be disabled once a quiescent state is reported for | |
901 | * this CPU. | |
902 | * | |
903 | * Of course, in carefully tuned systems, there might never be an | |
904 | * interrupt or exception. In that case, the RCU grace-period kthread | |
905 | * will eventually cause one to happen. However, in less carefully | |
906 | * controlled environments, this function allows RCU to get what it | |
907 | * needs without creating otherwise useless interruptions. | |
908 | */ | |
909 | void __rcu_irq_enter_check_tick(void) | |
910 | { | |
911 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); | |
912 | ||
913 | // Enabling the tick is unsafe in NMI handlers. | |
914 | if (WARN_ON_ONCE(in_nmi())) | |
915 | return; | |
916 | ||
917 | RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(), | |
918 | "Illegal rcu_irq_enter_check_tick() from extended quiescent state"); | |
919 | ||
920 | if (!tick_nohz_full_cpu(rdp->cpu) || | |
921 | !READ_ONCE(rdp->rcu_urgent_qs) || | |
922 | READ_ONCE(rdp->rcu_forced_tick)) { | |
923 | // RCU doesn't need nohz_full help from this CPU, or it is | |
924 | // already getting that help. | |
925 | return; | |
926 | } | |
927 | ||
928 | // We get here only when not in an extended quiescent state and | |
929 | // from interrupts (as opposed to NMIs). Therefore, (1) RCU is | |
930 | // already watching and (2) The fact that we are in an interrupt | |
931 | // handler and that the rcu_node lock is an irq-disabled lock | |
932 | // prevents self-deadlock. So we can safely recheck under the lock. | |
933 | // Note that the nohz_full state currently cannot change. | |
934 | raw_spin_lock_rcu_node(rdp->mynode); | |
935 | if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) { | |
936 | // A nohz_full CPU is in the kernel and RCU needs a | |
937 | // quiescent state. Turn on the tick! | |
938 | WRITE_ONCE(rdp->rcu_forced_tick, true); | |
939 | tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU); | |
940 | } | |
941 | raw_spin_unlock_rcu_node(rdp->mynode); | |
942 | } | |
d1ec4c34 | 943 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 944 | |
64db4cff | 945 | /** |
9ea366f6 | 946 | * rcu_nmi_enter - inform RCU of entry to NMI context |
cf7614e1 | 947 | * @irq: Is this call from rcu_irq_enter? |
64db4cff | 948 | * |
dc5a4f29 | 949 | * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and |
4c5273bf | 950 | * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know |
734d1680 PM |
951 | * that the CPU is active. This implementation permits nested NMIs, as |
952 | * long as the nesting level does not overflow an int. (You will probably | |
953 | * run out of stack space first.) | |
c0da313e | 954 | * |
9ea366f6 | 955 | * If you add or remove a call to rcu_nmi_enter(), be sure to test |
c0da313e | 956 | * with CONFIG_RCU_EQS_DEBUG=y. |
64db4cff | 957 | */ |
9ea366f6 | 958 | noinstr void rcu_nmi_enter(void) |
64db4cff | 959 | { |
84585aa8 | 960 | long incby = 2; |
5b14557b | 961 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
64db4cff | 962 | |
734d1680 | 963 | /* Complain about underflow. */ |
4c5273bf | 964 | WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0); |
734d1680 PM |
965 | |
966 | /* | |
967 | * If idle from RCU viewpoint, atomically increment ->dynticks | |
968 | * to mark non-idle and increment ->dynticks_nmi_nesting by one. | |
969 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means | |
970 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed | |
971 | * to be in the outermost NMI handler that interrupted an RCU-idle | |
972 | * period (observation due to Andy Lutomirski). | |
973 | */ | |
02a5c550 | 974 | if (rcu_dynticks_curr_cpu_in_eqs()) { |
cf7614e1 | 975 | |
9ea366f6 | 976 | if (!in_nmi()) |
cf7614e1 BP |
977 | rcu_dynticks_task_exit(); |
978 | ||
ac3caf82 | 979 | // RCU is not watching here ... |
2625d469 | 980 | rcu_dynticks_eqs_exit(); |
ac3caf82 | 981 | // ... but is watching here. |
cf7614e1 | 982 | |
9ea366f6 | 983 | if (!in_nmi()) |
cf7614e1 BP |
984 | rcu_cleanup_after_idle(); |
985 | ||
734d1680 | 986 | incby = 1; |
9ea366f6 | 987 | } else if (!in_nmi()) { |
ff5c4f5c | 988 | instrumentation_begin(); |
aaf2bc50 | 989 | rcu_irq_enter_check_tick(); |
ff5c4f5c | 990 | instrumentation_end(); |
734d1680 | 991 | } |
ff5c4f5c | 992 | instrumentation_begin(); |
bd2b879a | 993 | trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="), |
4c5273bf | 994 | rdp->dynticks_nmi_nesting, |
6cf539a8 | 995 | rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks)); |
ff5c4f5c | 996 | instrumentation_end(); |
4c5273bf PM |
997 | WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */ |
998 | rdp->dynticks_nmi_nesting + incby); | |
734d1680 | 999 | barrier(); |
64db4cff PM |
1000 | } |
1001 | ||
1002 | /** | |
9b2e4f18 | 1003 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle |
64db4cff | 1004 | * |
9b2e4f18 PM |
1005 | * Enter an interrupt handler, which might possibly result in exiting |
1006 | * idle mode, in other words, entering the mode in which read-side critical | |
7c9906ca | 1007 | * sections can occur. The caller must have disabled interrupts. |
c0da313e | 1008 | * |
9b2e4f18 | 1009 | * Note that the Linux kernel is fully capable of entering an interrupt |
58721f5d PM |
1010 | * handler that it never exits, for example when doing upcalls to user mode! |
1011 | * This code assumes that the idle loop never does upcalls to user mode. | |
1012 | * If your architecture's idle loop does do upcalls to user mode (or does | |
1013 | * anything else that results in unbalanced calls to the irq_enter() and | |
1014 | * irq_exit() functions), RCU will give you what you deserve, good and hard. | |
1015 | * But very infrequently and irreproducibly. | |
9b2e4f18 PM |
1016 | * |
1017 | * Use things like work queues to work around this limitation. | |
1018 | * | |
1019 | * You have been warned. | |
c0da313e PM |
1020 | * |
1021 | * If you add or remove a call to rcu_irq_enter(), be sure to test with | |
1022 | * CONFIG_RCU_EQS_DEBUG=y. | |
64db4cff | 1023 | */ |
ff5c4f5c | 1024 | noinstr void rcu_irq_enter(void) |
64db4cff | 1025 | { |
b04db8e1 | 1026 | lockdep_assert_irqs_disabled(); |
9ea366f6 | 1027 | rcu_nmi_enter(); |
7c9906ca | 1028 | } |
734d1680 | 1029 | |
7c9906ca PM |
1030 | /* |
1031 | * Wrapper for rcu_irq_enter() where interrupts are enabled. | |
c0da313e PM |
1032 | * |
1033 | * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test | |
1034 | * with CONFIG_RCU_EQS_DEBUG=y. | |
7c9906ca PM |
1035 | */ |
1036 | void rcu_irq_enter_irqson(void) | |
1037 | { | |
1038 | unsigned long flags; | |
734d1680 | 1039 | |
7c9906ca PM |
1040 | local_irq_save(flags); |
1041 | rcu_irq_enter(); | |
64db4cff | 1042 | local_irq_restore(flags); |
64db4cff PM |
1043 | } |
1044 | ||
66e4c33b | 1045 | /* |
516e5ae0 JFG |
1046 | * If any sort of urgency was applied to the current CPU (for example, |
1047 | * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order | |
1048 | * to get to a quiescent state, disable it. | |
66e4c33b | 1049 | */ |
516e5ae0 | 1050 | static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp) |
66e4c33b | 1051 | { |
5b14557b | 1052 | raw_lockdep_assert_held_rcu_node(rdp->mynode); |
516e5ae0 JFG |
1053 | WRITE_ONCE(rdp->rcu_urgent_qs, false); |
1054 | WRITE_ONCE(rdp->rcu_need_heavy_qs, false); | |
66e4c33b PM |
1055 | if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) { |
1056 | tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU); | |
2a2ae872 | 1057 | WRITE_ONCE(rdp->rcu_forced_tick, false); |
66e4c33b PM |
1058 | } |
1059 | } | |
1060 | ||
b1fcf9b8 TG |
1061 | noinstr bool __rcu_is_watching(void) |
1062 | { | |
1063 | return !rcu_dynticks_curr_cpu_in_eqs(); | |
1064 | } | |
1065 | ||
5c173eb8 | 1066 | /** |
2320bda2 | 1067 | * rcu_is_watching - see if RCU thinks that the current CPU is not idle |
64db4cff | 1068 | * |
791875d1 PM |
1069 | * Return true if RCU is watching the running CPU, which means that this |
1070 | * CPU can safely enter RCU read-side critical sections. In other words, | |
2320bda2 ZZ |
1071 | * if the current CPU is not in its idle loop or is in an interrupt or |
1072 | * NMI handler, return true. | |
64db4cff | 1073 | */ |
ff5c4f5c | 1074 | bool rcu_is_watching(void) |
64db4cff | 1075 | { |
f534ed1f | 1076 | bool ret; |
34240697 | 1077 | |
46f00d18 | 1078 | preempt_disable_notrace(); |
791875d1 | 1079 | ret = !rcu_dynticks_curr_cpu_in_eqs(); |
46f00d18 | 1080 | preempt_enable_notrace(); |
34240697 | 1081 | return ret; |
64db4cff | 1082 | } |
5c173eb8 | 1083 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 1084 | |
bcbfdd01 PM |
1085 | /* |
1086 | * If a holdout task is actually running, request an urgent quiescent | |
1087 | * state from its CPU. This is unsynchronized, so migrations can cause | |
1088 | * the request to go to the wrong CPU. Which is OK, all that will happen | |
1089 | * is that the CPU's next context switch will be a bit slower and next | |
1090 | * time around this task will generate another request. | |
1091 | */ | |
1092 | void rcu_request_urgent_qs_task(struct task_struct *t) | |
1093 | { | |
1094 | int cpu; | |
1095 | ||
1096 | barrier(); | |
1097 | cpu = task_cpu(t); | |
1098 | if (!task_curr(t)) | |
1099 | return; /* This task is not running on that CPU. */ | |
2dba13f0 | 1100 | smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true); |
bcbfdd01 PM |
1101 | } |
1102 | ||
62fde6ed | 1103 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
1104 | |
1105 | /* | |
5554788e | 1106 | * Is the current CPU online as far as RCU is concerned? |
2036d94a | 1107 | * |
5554788e PM |
1108 | * Disable preemption to avoid false positives that could otherwise |
1109 | * happen due to the current CPU number being sampled, this task being | |
1110 | * preempted, its old CPU being taken offline, resuming on some other CPU, | |
49918a54 | 1111 | * then determining that its old CPU is now offline. |
c0d6d01b | 1112 | * |
5554788e PM |
1113 | * Disable checking if in an NMI handler because we cannot safely |
1114 | * report errors from NMI handlers anyway. In addition, it is OK to use | |
1115 | * RCU on an offline processor during initial boot, hence the check for | |
1116 | * rcu_scheduler_fully_active. | |
c0d6d01b PM |
1117 | */ |
1118 | bool rcu_lockdep_current_cpu_online(void) | |
1119 | { | |
2036d94a PM |
1120 | struct rcu_data *rdp; |
1121 | struct rcu_node *rnp; | |
b97d23c5 | 1122 | bool ret = false; |
c0d6d01b | 1123 | |
5554788e | 1124 | if (in_nmi() || !rcu_scheduler_fully_active) |
f6f7ee9a | 1125 | return true; |
ff5c4f5c | 1126 | preempt_disable_notrace(); |
b97d23c5 PM |
1127 | rdp = this_cpu_ptr(&rcu_data); |
1128 | rnp = rdp->mynode; | |
1129 | if (rdp->grpmask & rcu_rnp_online_cpus(rnp)) | |
1130 | ret = true; | |
ff5c4f5c | 1131 | preempt_enable_notrace(); |
b97d23c5 | 1132 | return ret; |
c0d6d01b PM |
1133 | } |
1134 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
1135 | ||
62fde6ed | 1136 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 1137 | |
9b9500da PM |
1138 | /* |
1139 | * We are reporting a quiescent state on behalf of some other CPU, so | |
1140 | * it is our responsibility to check for and handle potential overflow | |
a66ae8ae | 1141 | * of the rcu_node ->gp_seq counter with respect to the rcu_data counters. |
9b9500da PM |
1142 | * After all, the CPU might be in deep idle state, and thus executing no |
1143 | * code whatsoever. | |
1144 | */ | |
1145 | static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp) | |
1146 | { | |
a32e01ee | 1147 | raw_lockdep_assert_held_rcu_node(rnp); |
a66ae8ae PM |
1148 | if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4, |
1149 | rnp->gp_seq)) | |
9b9500da | 1150 | WRITE_ONCE(rdp->gpwrap, true); |
8aa670cd PM |
1151 | if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq)) |
1152 | rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4; | |
9b9500da PM |
1153 | } |
1154 | ||
64db4cff PM |
1155 | /* |
1156 | * Snapshot the specified CPU's dynticks counter so that we can later | |
1157 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 1158 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 1159 | */ |
fe5ac724 | 1160 | static int dyntick_save_progress_counter(struct rcu_data *rdp) |
64db4cff | 1161 | { |
dc5a4f29 | 1162 | rdp->dynticks_snap = rcu_dynticks_snap(rdp); |
02a5c550 | 1163 | if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) { |
88d1bead | 1164 | trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti")); |
9b9500da | 1165 | rcu_gpnum_ovf(rdp->mynode, rdp); |
23a9bacd | 1166 | return 1; |
7941dbde | 1167 | } |
23a9bacd | 1168 | return 0; |
64db4cff PM |
1169 | } |
1170 | ||
1171 | /* | |
1172 | * Return true if the specified CPU has passed through a quiescent | |
1173 | * state by virtue of being in or having passed through an dynticks | |
1174 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 1175 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 1176 | */ |
fe5ac724 | 1177 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp) |
64db4cff | 1178 | { |
3a19b46a | 1179 | unsigned long jtsq; |
0f9be8ca | 1180 | bool *rnhqp; |
9226b10d | 1181 | bool *ruqp; |
9b9500da | 1182 | struct rcu_node *rnp = rdp->mynode; |
64db4cff PM |
1183 | |
1184 | /* | |
1185 | * If the CPU passed through or entered a dynticks idle phase with | |
1186 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
1187 | * already acknowledged the request to pass through a quiescent | |
1188 | * state. Either way, that CPU cannot possibly be in an RCU | |
1189 | * read-side critical section that started before the beginning | |
1190 | * of the current RCU grace period. | |
1191 | */ | |
dc5a4f29 | 1192 | if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) { |
88d1bead | 1193 | trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti")); |
9b9500da | 1194 | rcu_gpnum_ovf(rnp, rdp); |
3a19b46a PM |
1195 | return 1; |
1196 | } | |
1197 | ||
f2e2df59 PM |
1198 | /* If waiting too long on an offline CPU, complain. */ |
1199 | if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp)) && | |
88d1bead | 1200 | time_after(jiffies, rcu_state.gp_start + HZ)) { |
f2e2df59 PM |
1201 | bool onl; |
1202 | struct rcu_node *rnp1; | |
1203 | ||
1204 | WARN_ON(1); /* Offline CPUs are supposed to report QS! */ | |
1205 | pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n", | |
1206 | __func__, rnp->grplo, rnp->grphi, rnp->level, | |
1207 | (long)rnp->gp_seq, (long)rnp->completedqs); | |
1208 | for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent) | |
1209 | pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n", | |
1210 | __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask); | |
1211 | onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp)); | |
1212 | pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n", | |
1213 | __func__, rdp->cpu, ".o"[onl], | |
1214 | (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags, | |
1215 | (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags); | |
1216 | return 1; /* Break things loose after complaining. */ | |
1217 | } | |
1218 | ||
65d798f0 | 1219 | /* |
4a81e832 | 1220 | * A CPU running for an extended time within the kernel can |
c06aed0e PM |
1221 | * delay RCU grace periods: (1) At age jiffies_to_sched_qs, |
1222 | * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set | |
7e28c5af PM |
1223 | * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the |
1224 | * unsynchronized assignments to the per-CPU rcu_need_heavy_qs | |
1225 | * variable are safe because the assignments are repeated if this | |
1226 | * CPU failed to pass through a quiescent state. This code | |
c06aed0e | 1227 | * also checks .jiffies_resched in case jiffies_to_sched_qs |
7e28c5af | 1228 | * is set way high. |
6193c76a | 1229 | */ |
c06aed0e | 1230 | jtsq = READ_ONCE(jiffies_to_sched_qs); |
2dba13f0 PM |
1231 | ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu); |
1232 | rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu); | |
0f9be8ca | 1233 | if (!READ_ONCE(*rnhqp) && |
7e28c5af | 1234 | (time_after(jiffies, rcu_state.gp_start + jtsq * 2) || |
b2b00ddf PM |
1235 | time_after(jiffies, rcu_state.jiffies_resched) || |
1236 | rcu_state.cbovld)) { | |
0f9be8ca | 1237 | WRITE_ONCE(*rnhqp, true); |
9226b10d PM |
1238 | /* Store rcu_need_heavy_qs before rcu_urgent_qs. */ |
1239 | smp_store_release(ruqp, true); | |
7e28c5af PM |
1240 | } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) { |
1241 | WRITE_ONCE(*ruqp, true); | |
6193c76a PM |
1242 | } |
1243 | ||
28053bc7 | 1244 | /* |
c98cac60 | 1245 | * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq! |
d3052109 PM |
1246 | * The above code handles this, but only for straight cond_resched(). |
1247 | * And some in-kernel loops check need_resched() before calling | |
1248 | * cond_resched(), which defeats the above code for CPUs that are | |
1249 | * running in-kernel with scheduling-clock interrupts disabled. | |
1250 | * So hit them over the head with the resched_cpu() hammer! | |
28053bc7 | 1251 | */ |
d3052109 | 1252 | if (tick_nohz_full_cpu(rdp->cpu) && |
b2b00ddf PM |
1253 | (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) || |
1254 | rcu_state.cbovld)) { | |
05ef9e9e | 1255 | WRITE_ONCE(*ruqp, true); |
28053bc7 | 1256 | resched_cpu(rdp->cpu); |
d3052109 PM |
1257 | WRITE_ONCE(rdp->last_fqs_resched, jiffies); |
1258 | } | |
1259 | ||
1260 | /* | |
1261 | * If more than halfway to RCU CPU stall-warning time, invoke | |
1262 | * resched_cpu() more frequently to try to loosen things up a bit. | |
1263 | * Also check to see if the CPU is getting hammered with interrupts, | |
1264 | * but only once per grace period, just to keep the IPIs down to | |
1265 | * a dull roar. | |
1266 | */ | |
1267 | if (time_after(jiffies, rcu_state.jiffies_resched)) { | |
1268 | if (time_after(jiffies, | |
1269 | READ_ONCE(rdp->last_fqs_resched) + jtsq)) { | |
1270 | resched_cpu(rdp->cpu); | |
1271 | WRITE_ONCE(rdp->last_fqs_resched, jiffies); | |
1272 | } | |
9b9500da | 1273 | if (IS_ENABLED(CONFIG_IRQ_WORK) && |
8aa670cd | 1274 | !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq && |
9b9500da PM |
1275 | (rnp->ffmask & rdp->grpmask)) { |
1276 | init_irq_work(&rdp->rcu_iw, rcu_iw_handler); | |
49915ac3 | 1277 | atomic_set(&rdp->rcu_iw.flags, IRQ_WORK_HARD_IRQ); |
9b9500da | 1278 | rdp->rcu_iw_pending = true; |
8aa670cd | 1279 | rdp->rcu_iw_gp_seq = rnp->gp_seq; |
9b9500da PM |
1280 | irq_work_queue_on(&rdp->rcu_iw, rdp->cpu); |
1281 | } | |
1282 | } | |
4914950a | 1283 | |
a82dcc76 | 1284 | return 0; |
64db4cff PM |
1285 | } |
1286 | ||
41e80595 PM |
1287 | /* Trace-event wrapper function for trace_rcu_future_grace_period. */ |
1288 | static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
b73de91d | 1289 | unsigned long gp_seq_req, const char *s) |
0446be48 | 1290 | { |
0937d045 PM |
1291 | trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq), |
1292 | gp_seq_req, rnp->level, | |
1293 | rnp->grplo, rnp->grphi, s); | |
0446be48 PM |
1294 | } |
1295 | ||
1296 | /* | |
b73de91d | 1297 | * rcu_start_this_gp - Request the start of a particular grace period |
df2bf8f7 | 1298 | * @rnp_start: The leaf node of the CPU from which to start. |
b73de91d JF |
1299 | * @rdp: The rcu_data corresponding to the CPU from which to start. |
1300 | * @gp_seq_req: The gp_seq of the grace period to start. | |
1301 | * | |
41e80595 | 1302 | * Start the specified grace period, as needed to handle newly arrived |
0446be48 | 1303 | * callbacks. The required future grace periods are recorded in each |
7a1d0f23 | 1304 | * rcu_node structure's ->gp_seq_needed field. Returns true if there |
48a7639c | 1305 | * is reason to awaken the grace-period kthread. |
0446be48 | 1306 | * |
d5cd9685 PM |
1307 | * The caller must hold the specified rcu_node structure's ->lock, which |
1308 | * is why the caller is responsible for waking the grace-period kthread. | |
b73de91d JF |
1309 | * |
1310 | * Returns true if the GP thread needs to be awakened else false. | |
0446be48 | 1311 | */ |
df2bf8f7 | 1312 | static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp, |
b73de91d | 1313 | unsigned long gp_seq_req) |
0446be48 | 1314 | { |
48a7639c | 1315 | bool ret = false; |
df2bf8f7 | 1316 | struct rcu_node *rnp; |
0446be48 PM |
1317 | |
1318 | /* | |
360e0da6 PM |
1319 | * Use funnel locking to either acquire the root rcu_node |
1320 | * structure's lock or bail out if the need for this grace period | |
df2bf8f7 JFG |
1321 | * has already been recorded -- or if that grace period has in |
1322 | * fact already started. If there is already a grace period in | |
1323 | * progress in a non-leaf node, no recording is needed because the | |
1324 | * end of the grace period will scan the leaf rcu_node structures. | |
1325 | * Note that rnp_start->lock must not be released. | |
0446be48 | 1326 | */ |
df2bf8f7 JFG |
1327 | raw_lockdep_assert_held_rcu_node(rnp_start); |
1328 | trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf")); | |
1329 | for (rnp = rnp_start; 1; rnp = rnp->parent) { | |
1330 | if (rnp != rnp_start) | |
1331 | raw_spin_lock_rcu_node(rnp); | |
1332 | if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) || | |
1333 | rcu_seq_started(&rnp->gp_seq, gp_seq_req) || | |
1334 | (rnp != rnp_start && | |
1335 | rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) { | |
1336 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, | |
b73de91d | 1337 | TPS("Prestarted")); |
360e0da6 PM |
1338 | goto unlock_out; |
1339 | } | |
8ff37290 | 1340 | WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req); |
226ca5e7 | 1341 | if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) { |
a2165e41 | 1342 | /* |
226ca5e7 JFG |
1343 | * We just marked the leaf or internal node, and a |
1344 | * grace period is in progress, which means that | |
1345 | * rcu_gp_cleanup() will see the marking. Bail to | |
1346 | * reduce contention. | |
a2165e41 | 1347 | */ |
df2bf8f7 | 1348 | trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, |
b73de91d | 1349 | TPS("Startedleaf")); |
a2165e41 PM |
1350 | goto unlock_out; |
1351 | } | |
df2bf8f7 JFG |
1352 | if (rnp != rnp_start && rnp->parent != NULL) |
1353 | raw_spin_unlock_rcu_node(rnp); | |
1354 | if (!rnp->parent) | |
360e0da6 | 1355 | break; /* At root, and perhaps also leaf. */ |
0446be48 PM |
1356 | } |
1357 | ||
360e0da6 | 1358 | /* If GP already in progress, just leave, otherwise start one. */ |
de8e8730 | 1359 | if (rcu_gp_in_progress()) { |
df2bf8f7 | 1360 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot")); |
0446be48 PM |
1361 | goto unlock_out; |
1362 | } | |
df2bf8f7 | 1363 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot")); |
9cbc5b97 | 1364 | WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT); |
2906d215 | 1365 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); |
5648d659 | 1366 | if (!READ_ONCE(rcu_state.gp_kthread)) { |
df2bf8f7 | 1367 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread")); |
360e0da6 | 1368 | goto unlock_out; |
0446be48 | 1369 | } |
62ae1951 | 1370 | trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq")); |
360e0da6 | 1371 | ret = true; /* Caller must wake GP kthread. */ |
0446be48 | 1372 | unlock_out: |
ab5e869c | 1373 | /* Push furthest requested GP to leaf node and rcu_data structure. */ |
df2bf8f7 | 1374 | if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) { |
8ff37290 PM |
1375 | WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed); |
1376 | WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed); | |
ab5e869c | 1377 | } |
df2bf8f7 JFG |
1378 | if (rnp != rnp_start) |
1379 | raw_spin_unlock_rcu_node(rnp); | |
48a7639c | 1380 | return ret; |
0446be48 PM |
1381 | } |
1382 | ||
1383 | /* | |
1384 | * Clean up any old requests for the just-ended grace period. Also return | |
d1e4f01d | 1385 | * whether any additional grace periods have been requested. |
0446be48 | 1386 | */ |
3481f2ea | 1387 | static bool rcu_future_gp_cleanup(struct rcu_node *rnp) |
0446be48 | 1388 | { |
fb31340f | 1389 | bool needmore; |
da1df50d | 1390 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
0446be48 | 1391 | |
7a1d0f23 PM |
1392 | needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed); |
1393 | if (!needmore) | |
1394 | rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */ | |
b73de91d | 1395 | trace_rcu_this_gp(rnp, rdp, rnp->gp_seq, |
41e80595 | 1396 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); |
0446be48 PM |
1397 | return needmore; |
1398 | } | |
1399 | ||
48a7639c | 1400 | /* |
5648d659 PM |
1401 | * Awaken the grace-period kthread. Don't do a self-awaken (unless in an |
1402 | * interrupt or softirq handler, in which case we just might immediately | |
1403 | * sleep upon return, resulting in a grace-period hang), and don't bother | |
1404 | * awakening when there is nothing for the grace-period kthread to do | |
1405 | * (as in several CPUs raced to awaken, we lost), and finally don't try | |
1406 | * to awaken a kthread that has not yet been created. If all those checks | |
1407 | * are passed, track some debug information and awaken. | |
1d1f898d ZJ |
1408 | * |
1409 | * So why do the self-wakeup when in an interrupt or softirq handler | |
1410 | * in the grace-period kthread's context? Because the kthread might have | |
1411 | * been interrupted just as it was going to sleep, and just after the final | |
1412 | * pre-sleep check of the awaken condition. In this case, a wakeup really | |
1413 | * is required, and is therefore supplied. | |
48a7639c | 1414 | */ |
532c00c9 | 1415 | static void rcu_gp_kthread_wake(void) |
48a7639c | 1416 | { |
5648d659 PM |
1417 | struct task_struct *t = READ_ONCE(rcu_state.gp_kthread); |
1418 | ||
1419 | if ((current == t && !in_irq() && !in_serving_softirq()) || | |
1420 | !READ_ONCE(rcu_state.gp_flags) || !t) | |
48a7639c | 1421 | return; |
fd897573 PM |
1422 | WRITE_ONCE(rcu_state.gp_wake_time, jiffies); |
1423 | WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq)); | |
532c00c9 | 1424 | swake_up_one(&rcu_state.gp_wq); |
48a7639c PM |
1425 | } |
1426 | ||
dc35c893 | 1427 | /* |
29365e56 PM |
1428 | * If there is room, assign a ->gp_seq number to any callbacks on this |
1429 | * CPU that have not already been assigned. Also accelerate any callbacks | |
1430 | * that were previously assigned a ->gp_seq number that has since proven | |
1431 | * to be too conservative, which can happen if callbacks get assigned a | |
1432 | * ->gp_seq number while RCU is idle, but with reference to a non-root | |
1433 | * rcu_node structure. This function is idempotent, so it does not hurt | |
1434 | * to call it repeatedly. Returns an flag saying that we should awaken | |
1435 | * the RCU grace-period kthread. | |
dc35c893 PM |
1436 | * |
1437 | * The caller must hold rnp->lock with interrupts disabled. | |
1438 | */ | |
02f50142 | 1439 | static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp) |
dc35c893 | 1440 | { |
b73de91d | 1441 | unsigned long gp_seq_req; |
15fecf89 | 1442 | bool ret = false; |
dc35c893 | 1443 | |
d1b222c6 | 1444 | rcu_lockdep_assert_cblist_protected(rdp); |
a32e01ee | 1445 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1446 | |
15fecf89 PM |
1447 | /* If no pending (not yet ready to invoke) callbacks, nothing to do. */ |
1448 | if (!rcu_segcblist_pend_cbs(&rdp->cblist)) | |
48a7639c | 1449 | return false; |
dc35c893 PM |
1450 | |
1451 | /* | |
15fecf89 PM |
1452 | * Callbacks are often registered with incomplete grace-period |
1453 | * information. Something about the fact that getting exact | |
1454 | * information requires acquiring a global lock... RCU therefore | |
1455 | * makes a conservative estimate of the grace period number at which | |
1456 | * a given callback will become ready to invoke. The following | |
1457 | * code checks this estimate and improves it when possible, thus | |
1458 | * accelerating callback invocation to an earlier grace-period | |
1459 | * number. | |
dc35c893 | 1460 | */ |
9cbc5b97 | 1461 | gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq); |
b73de91d JF |
1462 | if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req)) |
1463 | ret = rcu_start_this_gp(rnp, rdp, gp_seq_req); | |
6d4b418c PM |
1464 | |
1465 | /* Trace depending on how much we were able to accelerate. */ | |
15fecf89 | 1466 | if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL)) |
9cbc5b97 | 1467 | trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccWaitCB")); |
6d4b418c | 1468 | else |
9cbc5b97 | 1469 | trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("AccReadyCB")); |
48a7639c | 1470 | return ret; |
dc35c893 PM |
1471 | } |
1472 | ||
e44e73ca PM |
1473 | /* |
1474 | * Similar to rcu_accelerate_cbs(), but does not require that the leaf | |
1475 | * rcu_node structure's ->lock be held. It consults the cached value | |
1476 | * of ->gp_seq_needed in the rcu_data structure, and if that indicates | |
1477 | * that a new grace-period request be made, invokes rcu_accelerate_cbs() | |
1478 | * while holding the leaf rcu_node structure's ->lock. | |
1479 | */ | |
c6e09b97 | 1480 | static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp, |
e44e73ca PM |
1481 | struct rcu_data *rdp) |
1482 | { | |
1483 | unsigned long c; | |
1484 | bool needwake; | |
1485 | ||
d1b222c6 | 1486 | rcu_lockdep_assert_cblist_protected(rdp); |
c6e09b97 | 1487 | c = rcu_seq_snap(&rcu_state.gp_seq); |
a5b89501 | 1488 | if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) { |
e44e73ca PM |
1489 | /* Old request still live, so mark recent callbacks. */ |
1490 | (void)rcu_segcblist_accelerate(&rdp->cblist, c); | |
1491 | return; | |
1492 | } | |
1493 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ | |
02f50142 | 1494 | needwake = rcu_accelerate_cbs(rnp, rdp); |
e44e73ca PM |
1495 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
1496 | if (needwake) | |
532c00c9 | 1497 | rcu_gp_kthread_wake(); |
e44e73ca PM |
1498 | } |
1499 | ||
dc35c893 PM |
1500 | /* |
1501 | * Move any callbacks whose grace period has completed to the | |
1502 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
29365e56 | 1503 | * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL |
dc35c893 PM |
1504 | * sublist. This function is idempotent, so it does not hurt to |
1505 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1506 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1507 | * |
1508 | * The caller must hold rnp->lock with interrupts disabled. | |
1509 | */ | |
834f56bf | 1510 | static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp) |
dc35c893 | 1511 | { |
d1b222c6 | 1512 | rcu_lockdep_assert_cblist_protected(rdp); |
a32e01ee | 1513 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1514 | |
15fecf89 PM |
1515 | /* If no pending (not yet ready to invoke) callbacks, nothing to do. */ |
1516 | if (!rcu_segcblist_pend_cbs(&rdp->cblist)) | |
48a7639c | 1517 | return false; |
dc35c893 PM |
1518 | |
1519 | /* | |
29365e56 | 1520 | * Find all callbacks whose ->gp_seq numbers indicate that they |
dc35c893 PM |
1521 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. |
1522 | */ | |
29365e56 | 1523 | rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq); |
dc35c893 PM |
1524 | |
1525 | /* Classify any remaining callbacks. */ | |
02f50142 | 1526 | return rcu_accelerate_cbs(rnp, rdp); |
dc35c893 PM |
1527 | } |
1528 | ||
7f36ef82 PM |
1529 | /* |
1530 | * Move and classify callbacks, but only if doing so won't require | |
1531 | * that the RCU grace-period kthread be awakened. | |
1532 | */ | |
1533 | static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp, | |
1534 | struct rcu_data *rdp) | |
1535 | { | |
d1b222c6 | 1536 | rcu_lockdep_assert_cblist_protected(rdp); |
6608c3a0 PM |
1537 | if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || |
1538 | !raw_spin_trylock_rcu_node(rnp)) | |
7f36ef82 PM |
1539 | return; |
1540 | WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp)); | |
6608c3a0 | 1541 | raw_spin_unlock_rcu_node(rnp); |
7f36ef82 PM |
1542 | } |
1543 | ||
d09b62df | 1544 | /* |
ba9fbe95 PM |
1545 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1546 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1547 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1548 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1549 | */ |
c7e48f7b | 1550 | static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp) |
d09b62df | 1551 | { |
5d6742b3 | 1552 | bool ret = false; |
b5ea0370 | 1553 | bool need_qs; |
5d6742b3 PM |
1554 | const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) && |
1555 | rcu_segcblist_is_offloaded(&rdp->cblist); | |
48a7639c | 1556 | |
a32e01ee | 1557 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1558 | |
67e14c1e PM |
1559 | if (rdp->gp_seq == rnp->gp_seq) |
1560 | return false; /* Nothing to do. */ | |
d09b62df | 1561 | |
67e14c1e PM |
1562 | /* Handle the ends of any preceding grace periods first. */ |
1563 | if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) || | |
1564 | unlikely(READ_ONCE(rdp->gpwrap))) { | |
5d6742b3 PM |
1565 | if (!offloaded) |
1566 | ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */ | |
b5ea0370 | 1567 | rdp->core_needs_qs = false; |
9cbc5b97 | 1568 | trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend")); |
67e14c1e | 1569 | } else { |
5d6742b3 PM |
1570 | if (!offloaded) |
1571 | ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */ | |
b5ea0370 PM |
1572 | if (rdp->core_needs_qs) |
1573 | rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask); | |
d09b62df | 1574 | } |
398ebe60 | 1575 | |
67e14c1e PM |
1576 | /* Now handle the beginnings of any new-to-this-CPU grace periods. */ |
1577 | if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) || | |
1578 | unlikely(READ_ONCE(rdp->gpwrap))) { | |
6eaef633 PM |
1579 | /* |
1580 | * If the current grace period is waiting for this CPU, | |
1581 | * set up to detect a quiescent state, otherwise don't | |
1582 | * go looking for one. | |
1583 | */ | |
9cbc5b97 | 1584 | trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart")); |
b5ea0370 PM |
1585 | need_qs = !!(rnp->qsmask & rdp->grpmask); |
1586 | rdp->cpu_no_qs.b.norm = need_qs; | |
1587 | rdp->core_needs_qs = need_qs; | |
6eaef633 PM |
1588 | zero_cpu_stall_ticks(rdp); |
1589 | } | |
67e14c1e | 1590 | rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */ |
13dc7d0c | 1591 | if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap) |
8ff37290 | 1592 | WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed); |
3d18469a PM |
1593 | WRITE_ONCE(rdp->gpwrap, false); |
1594 | rcu_gpnum_ovf(rnp, rdp); | |
48a7639c | 1595 | return ret; |
6eaef633 PM |
1596 | } |
1597 | ||
15cabdff | 1598 | static void note_gp_changes(struct rcu_data *rdp) |
6eaef633 PM |
1599 | { |
1600 | unsigned long flags; | |
48a7639c | 1601 | bool needwake; |
6eaef633 PM |
1602 | struct rcu_node *rnp; |
1603 | ||
1604 | local_irq_save(flags); | |
1605 | rnp = rdp->mynode; | |
67e14c1e | 1606 | if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) && |
7d0ae808 | 1607 | !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */ |
2a67e741 | 1608 | !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */ |
6eaef633 PM |
1609 | local_irq_restore(flags); |
1610 | return; | |
1611 | } | |
c7e48f7b | 1612 | needwake = __note_gp_changes(rnp, rdp); |
67c583a7 | 1613 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
48a7639c | 1614 | if (needwake) |
532c00c9 | 1615 | rcu_gp_kthread_wake(); |
6eaef633 PM |
1616 | } |
1617 | ||
22212332 | 1618 | static void rcu_gp_slow(int delay) |
0f41c0dd PM |
1619 | { |
1620 | if (delay > 0 && | |
22212332 | 1621 | !(rcu_seq_ctr(rcu_state.gp_seq) % |
dee4f422 | 1622 | (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay))) |
0f41c0dd PM |
1623 | schedule_timeout_uninterruptible(delay); |
1624 | } | |
1625 | ||
55b2dcf5 PM |
1626 | static unsigned long sleep_duration; |
1627 | ||
1628 | /* Allow rcutorture to stall the grace-period kthread. */ | |
1629 | void rcu_gp_set_torture_wait(int duration) | |
1630 | { | |
1631 | if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0) | |
1632 | WRITE_ONCE(sleep_duration, duration); | |
1633 | } | |
1634 | EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait); | |
1635 | ||
1636 | /* Actually implement the aforementioned wait. */ | |
1637 | static void rcu_gp_torture_wait(void) | |
1638 | { | |
1639 | unsigned long duration; | |
1640 | ||
1641 | if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST)) | |
1642 | return; | |
1643 | duration = xchg(&sleep_duration, 0UL); | |
1644 | if (duration > 0) { | |
1645 | pr_alert("%s: Waiting %lu jiffies\n", __func__, duration); | |
1646 | schedule_timeout_uninterruptible(duration); | |
1647 | pr_alert("%s: Wait complete\n", __func__); | |
1648 | } | |
1649 | } | |
1650 | ||
b3dbec76 | 1651 | /* |
45fed3e7 | 1652 | * Initialize a new grace period. Return false if no grace period required. |
b3dbec76 | 1653 | */ |
0854a05c | 1654 | static bool rcu_gp_init(void) |
b3dbec76 | 1655 | { |
ec2c2976 | 1656 | unsigned long flags; |
0aa04b05 | 1657 | unsigned long oldmask; |
ec2c2976 | 1658 | unsigned long mask; |
b3dbec76 | 1659 | struct rcu_data *rdp; |
336a4f6c | 1660 | struct rcu_node *rnp = rcu_get_root(); |
b3dbec76 | 1661 | |
9cbc5b97 | 1662 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
2a67e741 | 1663 | raw_spin_lock_irq_rcu_node(rnp); |
9cbc5b97 | 1664 | if (!READ_ONCE(rcu_state.gp_flags)) { |
f7be8209 | 1665 | /* Spurious wakeup, tell caller to go back to sleep. */ |
67c583a7 | 1666 | raw_spin_unlock_irq_rcu_node(rnp); |
45fed3e7 | 1667 | return false; |
f7be8209 | 1668 | } |
9cbc5b97 | 1669 | WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */ |
b3dbec76 | 1670 | |
de8e8730 | 1671 | if (WARN_ON_ONCE(rcu_gp_in_progress())) { |
f7be8209 PM |
1672 | /* |
1673 | * Grace period already in progress, don't start another. | |
1674 | * Not supposed to be able to happen. | |
1675 | */ | |
67c583a7 | 1676 | raw_spin_unlock_irq_rcu_node(rnp); |
45fed3e7 | 1677 | return false; |
7fdefc10 PM |
1678 | } |
1679 | ||
7fdefc10 | 1680 | /* Advance to a new grace period and initialize state. */ |
ad3832e9 | 1681 | record_gp_stall_check_time(); |
ff3bb6f4 | 1682 | /* Record GP times before starting GP, hence rcu_seq_start(). */ |
9cbc5b97 | 1683 | rcu_seq_start(&rcu_state.gp_seq); |
62ae1951 | 1684 | ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); |
9cbc5b97 | 1685 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); |
67c583a7 | 1686 | raw_spin_unlock_irq_rcu_node(rnp); |
7fdefc10 | 1687 | |
0aa04b05 PM |
1688 | /* |
1689 | * Apply per-leaf buffered online and offline operations to the | |
1690 | * rcu_node tree. Note that this new grace period need not wait | |
1691 | * for subsequent online CPUs, and that quiescent-state forcing | |
1692 | * will handle subsequent offline CPUs. | |
1693 | */ | |
9cbc5b97 | 1694 | rcu_state.gp_state = RCU_GP_ONOFF; |
aedf4ba9 | 1695 | rcu_for_each_leaf_node(rnp) { |
894d45bb | 1696 | raw_spin_lock(&rcu_state.ofl_lock); |
2a67e741 | 1697 | raw_spin_lock_irq_rcu_node(rnp); |
0aa04b05 PM |
1698 | if (rnp->qsmaskinit == rnp->qsmaskinitnext && |
1699 | !rnp->wait_blkd_tasks) { | |
1700 | /* Nothing to do on this leaf rcu_node structure. */ | |
67c583a7 | 1701 | raw_spin_unlock_irq_rcu_node(rnp); |
894d45bb | 1702 | raw_spin_unlock(&rcu_state.ofl_lock); |
0aa04b05 PM |
1703 | continue; |
1704 | } | |
1705 | ||
1706 | /* Record old state, apply changes to ->qsmaskinit field. */ | |
1707 | oldmask = rnp->qsmaskinit; | |
1708 | rnp->qsmaskinit = rnp->qsmaskinitnext; | |
1709 | ||
1710 | /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ | |
1711 | if (!oldmask != !rnp->qsmaskinit) { | |
962aff03 PM |
1712 | if (!oldmask) { /* First online CPU for rcu_node. */ |
1713 | if (!rnp->wait_blkd_tasks) /* Ever offline? */ | |
1714 | rcu_init_new_rnp(rnp); | |
1715 | } else if (rcu_preempt_has_tasks(rnp)) { | |
1716 | rnp->wait_blkd_tasks = true; /* blocked tasks */ | |
1717 | } else { /* Last offline CPU and can propagate. */ | |
0aa04b05 | 1718 | rcu_cleanup_dead_rnp(rnp); |
962aff03 | 1719 | } |
0aa04b05 PM |
1720 | } |
1721 | ||
1722 | /* | |
1723 | * If all waited-on tasks from prior grace period are | |
1724 | * done, and if all this rcu_node structure's CPUs are | |
1725 | * still offline, propagate up the rcu_node tree and | |
1726 | * clear ->wait_blkd_tasks. Otherwise, if one of this | |
1727 | * rcu_node structure's CPUs has since come back online, | |
962aff03 | 1728 | * simply clear ->wait_blkd_tasks. |
0aa04b05 PM |
1729 | */ |
1730 | if (rnp->wait_blkd_tasks && | |
962aff03 | 1731 | (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) { |
0aa04b05 | 1732 | rnp->wait_blkd_tasks = false; |
962aff03 PM |
1733 | if (!rnp->qsmaskinit) |
1734 | rcu_cleanup_dead_rnp(rnp); | |
0aa04b05 PM |
1735 | } |
1736 | ||
67c583a7 | 1737 | raw_spin_unlock_irq_rcu_node(rnp); |
894d45bb | 1738 | raw_spin_unlock(&rcu_state.ofl_lock); |
0aa04b05 | 1739 | } |
22212332 | 1740 | rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */ |
7fdefc10 PM |
1741 | |
1742 | /* | |
1743 | * Set the quiescent-state-needed bits in all the rcu_node | |
9cbc5b97 PM |
1744 | * structures for all currently online CPUs in breadth-first |
1745 | * order, starting from the root rcu_node structure, relying on the | |
1746 | * layout of the tree within the rcu_state.node[] array. Note that | |
1747 | * other CPUs will access only the leaves of the hierarchy, thus | |
1748 | * seeing that no grace period is in progress, at least until the | |
1749 | * corresponding leaf node has been initialized. | |
7fdefc10 PM |
1750 | * |
1751 | * The grace period cannot complete until the initialization | |
1752 | * process finishes, because this kthread handles both. | |
1753 | */ | |
9cbc5b97 | 1754 | rcu_state.gp_state = RCU_GP_INIT; |
aedf4ba9 | 1755 | rcu_for_each_node_breadth_first(rnp) { |
22212332 | 1756 | rcu_gp_slow(gp_init_delay); |
ec2c2976 | 1757 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
da1df50d | 1758 | rdp = this_cpu_ptr(&rcu_data); |
81ab59a3 | 1759 | rcu_preempt_check_blocked_tasks(rnp); |
7fdefc10 | 1760 | rnp->qsmask = rnp->qsmaskinit; |
9cbc5b97 | 1761 | WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq); |
7fdefc10 | 1762 | if (rnp == rdp->mynode) |
c7e48f7b | 1763 | (void)__note_gp_changes(rnp, rdp); |
7fdefc10 | 1764 | rcu_preempt_boost_start_gp(rnp); |
9cbc5b97 | 1765 | trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq, |
7fdefc10 PM |
1766 | rnp->level, rnp->grplo, |
1767 | rnp->grphi, rnp->qsmask); | |
ec2c2976 PM |
1768 | /* Quiescent states for tasks on any now-offline CPUs. */ |
1769 | mask = rnp->qsmask & ~rnp->qsmaskinitnext; | |
f2e2df59 | 1770 | rnp->rcu_gp_init_mask = mask; |
ec2c2976 | 1771 | if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp)) |
b50912d0 | 1772 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
ec2c2976 PM |
1773 | else |
1774 | raw_spin_unlock_irq_rcu_node(rnp); | |
cee43939 | 1775 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 1776 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
7fdefc10 | 1777 | } |
b3dbec76 | 1778 | |
45fed3e7 | 1779 | return true; |
7fdefc10 | 1780 | } |
b3dbec76 | 1781 | |
b9a425cf | 1782 | /* |
b3dae109 | 1783 | * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state |
d5374226 | 1784 | * time. |
b9a425cf | 1785 | */ |
0854a05c | 1786 | static bool rcu_gp_fqs_check_wake(int *gfp) |
b9a425cf | 1787 | { |
336a4f6c | 1788 | struct rcu_node *rnp = rcu_get_root(); |
b9a425cf | 1789 | |
1fca4d12 PM |
1790 | // If under overload conditions, force an immediate FQS scan. |
1791 | if (*gfp & RCU_GP_FLAG_OVLD) | |
1792 | return true; | |
1793 | ||
1794 | // Someone like call_rcu() requested a force-quiescent-state scan. | |
0854a05c | 1795 | *gfp = READ_ONCE(rcu_state.gp_flags); |
b9a425cf PM |
1796 | if (*gfp & RCU_GP_FLAG_FQS) |
1797 | return true; | |
1798 | ||
1fca4d12 | 1799 | // The current grace period has completed. |
b9a425cf PM |
1800 | if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp)) |
1801 | return true; | |
1802 | ||
1803 | return false; | |
1804 | } | |
1805 | ||
4cdfc175 PM |
1806 | /* |
1807 | * Do one round of quiescent-state forcing. | |
1808 | */ | |
0854a05c | 1809 | static void rcu_gp_fqs(bool first_time) |
4cdfc175 | 1810 | { |
336a4f6c | 1811 | struct rcu_node *rnp = rcu_get_root(); |
4cdfc175 | 1812 | |
9cbc5b97 PM |
1813 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
1814 | rcu_state.n_force_qs++; | |
77f81fe0 | 1815 | if (first_time) { |
4cdfc175 | 1816 | /* Collect dyntick-idle snapshots. */ |
e9ecb780 | 1817 | force_qs_rnp(dyntick_save_progress_counter); |
4cdfc175 PM |
1818 | } else { |
1819 | /* Handle dyntick-idle and offline CPUs. */ | |
e9ecb780 | 1820 | force_qs_rnp(rcu_implicit_dynticks_qs); |
4cdfc175 PM |
1821 | } |
1822 | /* Clear flag to prevent immediate re-entry. */ | |
9cbc5b97 | 1823 | if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) { |
2a67e741 | 1824 | raw_spin_lock_irq_rcu_node(rnp); |
9cbc5b97 PM |
1825 | WRITE_ONCE(rcu_state.gp_flags, |
1826 | READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS); | |
67c583a7 | 1827 | raw_spin_unlock_irq_rcu_node(rnp); |
4cdfc175 | 1828 | } |
4cdfc175 PM |
1829 | } |
1830 | ||
c3854a05 PM |
1831 | /* |
1832 | * Loop doing repeated quiescent-state forcing until the grace period ends. | |
1833 | */ | |
1834 | static void rcu_gp_fqs_loop(void) | |
1835 | { | |
1836 | bool first_gp_fqs; | |
1fca4d12 | 1837 | int gf = 0; |
c3854a05 PM |
1838 | unsigned long j; |
1839 | int ret; | |
1840 | struct rcu_node *rnp = rcu_get_root(); | |
1841 | ||
1842 | first_gp_fqs = true; | |
c06aed0e | 1843 | j = READ_ONCE(jiffies_till_first_fqs); |
1fca4d12 PM |
1844 | if (rcu_state.cbovld) |
1845 | gf = RCU_GP_FLAG_OVLD; | |
c3854a05 PM |
1846 | ret = 0; |
1847 | for (;;) { | |
1848 | if (!ret) { | |
1849 | rcu_state.jiffies_force_qs = jiffies + j; | |
1850 | WRITE_ONCE(rcu_state.jiffies_kick_kthreads, | |
9cf422a8 | 1851 | jiffies + (j ? 3 * j : 2)); |
c3854a05 | 1852 | } |
0f11ad32 | 1853 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
1854 | TPS("fqswait")); |
1855 | rcu_state.gp_state = RCU_GP_WAIT_FQS; | |
1856 | ret = swait_event_idle_timeout_exclusive( | |
1857 | rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j); | |
55b2dcf5 | 1858 | rcu_gp_torture_wait(); |
c3854a05 PM |
1859 | rcu_state.gp_state = RCU_GP_DOING_FQS; |
1860 | /* Locking provides needed memory barriers. */ | |
1861 | /* If grace period done, leave loop. */ | |
1862 | if (!READ_ONCE(rnp->qsmask) && | |
1863 | !rcu_preempt_blocked_readers_cgp(rnp)) | |
1864 | break; | |
1865 | /* If time for quiescent-state forcing, do it. */ | |
29ffebc5 | 1866 | if (!time_after(rcu_state.jiffies_force_qs, jiffies) || |
c3854a05 | 1867 | (gf & RCU_GP_FLAG_FQS)) { |
0f11ad32 | 1868 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
1869 | TPS("fqsstart")); |
1870 | rcu_gp_fqs(first_gp_fqs); | |
1fca4d12 PM |
1871 | gf = 0; |
1872 | if (first_gp_fqs) { | |
1873 | first_gp_fqs = false; | |
1874 | gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0; | |
1875 | } | |
0f11ad32 | 1876 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
1877 | TPS("fqsend")); |
1878 | cond_resched_tasks_rcu_qs(); | |
1879 | WRITE_ONCE(rcu_state.gp_activity, jiffies); | |
1880 | ret = 0; /* Force full wait till next FQS. */ | |
c06aed0e | 1881 | j = READ_ONCE(jiffies_till_next_fqs); |
c3854a05 PM |
1882 | } else { |
1883 | /* Deal with stray signal. */ | |
1884 | cond_resched_tasks_rcu_qs(); | |
1885 | WRITE_ONCE(rcu_state.gp_activity, jiffies); | |
1886 | WARN_ON(signal_pending(current)); | |
0f11ad32 | 1887 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
1888 | TPS("fqswaitsig")); |
1889 | ret = 1; /* Keep old FQS timing. */ | |
1890 | j = jiffies; | |
1891 | if (time_after(jiffies, rcu_state.jiffies_force_qs)) | |
1892 | j = 1; | |
1893 | else | |
1894 | j = rcu_state.jiffies_force_qs - j; | |
1fca4d12 | 1895 | gf = 0; |
c3854a05 PM |
1896 | } |
1897 | } | |
1898 | } | |
1899 | ||
7fdefc10 PM |
1900 | /* |
1901 | * Clean up after the old grace period. | |
1902 | */ | |
0854a05c | 1903 | static void rcu_gp_cleanup(void) |
7fdefc10 | 1904 | { |
b2b00ddf | 1905 | int cpu; |
48a7639c | 1906 | bool needgp = false; |
b2b00ddf | 1907 | unsigned long gp_duration; |
de30ad51 | 1908 | unsigned long new_gp_seq; |
5d6742b3 | 1909 | bool offloaded; |
7fdefc10 | 1910 | struct rcu_data *rdp; |
336a4f6c | 1911 | struct rcu_node *rnp = rcu_get_root(); |
abedf8e2 | 1912 | struct swait_queue_head *sq; |
b3dbec76 | 1913 | |
9cbc5b97 | 1914 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
2a67e741 | 1915 | raw_spin_lock_irq_rcu_node(rnp); |
c51d7b5e PM |
1916 | rcu_state.gp_end = jiffies; |
1917 | gp_duration = rcu_state.gp_end - rcu_state.gp_start; | |
9cbc5b97 PM |
1918 | if (gp_duration > rcu_state.gp_max) |
1919 | rcu_state.gp_max = gp_duration; | |
b3dbec76 | 1920 | |
7fdefc10 PM |
1921 | /* |
1922 | * We know the grace period is complete, but to everyone else | |
1923 | * it appears to still be ongoing. But it is also the case | |
1924 | * that to everyone else it looks like there is nothing that | |
1925 | * they can do to advance the grace period. It is therefore | |
1926 | * safe for us to drop the lock in order to mark the grace | |
1927 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 1928 | */ |
67c583a7 | 1929 | raw_spin_unlock_irq_rcu_node(rnp); |
b3dbec76 | 1930 | |
5d4b8659 | 1931 | /* |
ff3bb6f4 PM |
1932 | * Propagate new ->gp_seq value to rcu_node structures so that |
1933 | * other CPUs don't have to wait until the start of the next grace | |
1934 | * period to process their callbacks. This also avoids some nasty | |
1935 | * RCU grace-period initialization races by forcing the end of | |
1936 | * the current grace period to be completely recorded in all of | |
1937 | * the rcu_node structures before the beginning of the next grace | |
1938 | * period is recorded in any of the rcu_node structures. | |
5d4b8659 | 1939 | */ |
9cbc5b97 | 1940 | new_gp_seq = rcu_state.gp_seq; |
de30ad51 | 1941 | rcu_seq_end(&new_gp_seq); |
aedf4ba9 | 1942 | rcu_for_each_node_breadth_first(rnp) { |
2a67e741 | 1943 | raw_spin_lock_irq_rcu_node(rnp); |
4bc8d555 | 1944 | if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp))) |
81ab59a3 | 1945 | dump_blkd_tasks(rnp, 10); |
5c60d25f | 1946 | WARN_ON_ONCE(rnp->qsmask); |
de30ad51 | 1947 | WRITE_ONCE(rnp->gp_seq, new_gp_seq); |
da1df50d | 1948 | rdp = this_cpu_ptr(&rcu_data); |
b11cc576 | 1949 | if (rnp == rdp->mynode) |
c7e48f7b | 1950 | needgp = __note_gp_changes(rnp, rdp) || needgp; |
78e4bc34 | 1951 | /* smp_mb() provided by prior unlock-lock pair. */ |
3481f2ea | 1952 | needgp = rcu_future_gp_cleanup(rnp) || needgp; |
b2b00ddf PM |
1953 | // Reset overload indication for CPUs no longer overloaded |
1954 | if (rcu_is_leaf_node(rnp)) | |
1955 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) { | |
1956 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
1957 | check_cb_ovld_locked(rdp, rnp); | |
1958 | } | |
065bb78c | 1959 | sq = rcu_nocb_gp_get(rnp); |
67c583a7 | 1960 | raw_spin_unlock_irq_rcu_node(rnp); |
065bb78c | 1961 | rcu_nocb_gp_cleanup(sq); |
cee43939 | 1962 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 1963 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
22212332 | 1964 | rcu_gp_slow(gp_cleanup_delay); |
7fdefc10 | 1965 | } |
336a4f6c | 1966 | rnp = rcu_get_root(); |
9cbc5b97 | 1967 | raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */ |
7fdefc10 | 1968 | |
0a89e5a4 | 1969 | /* Declare grace period done, trace first to use old GP number. */ |
9cbc5b97 | 1970 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end")); |
0a89e5a4 | 1971 | rcu_seq_end(&rcu_state.gp_seq); |
62ae1951 | 1972 | ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); |
9cbc5b97 | 1973 | rcu_state.gp_state = RCU_GP_IDLE; |
fb31340f | 1974 | /* Check for GP requests since above loop. */ |
da1df50d | 1975 | rdp = this_cpu_ptr(&rcu_data); |
5b55072f | 1976 | if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) { |
abd13fdd | 1977 | trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed, |
41e80595 | 1978 | TPS("CleanupMore")); |
fb31340f PM |
1979 | needgp = true; |
1980 | } | |
48a7639c | 1981 | /* Advance CBs to reduce false positives below. */ |
5d6742b3 PM |
1982 | offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) && |
1983 | rcu_segcblist_is_offloaded(&rdp->cblist); | |
1984 | if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) { | |
9cbc5b97 | 1985 | WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT); |
2906d215 | 1986 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); |
9cbc5b97 | 1987 | trace_rcu_grace_period(rcu_state.name, |
0f11ad32 | 1988 | rcu_state.gp_seq, |
bb311ecc | 1989 | TPS("newreq")); |
18390aea | 1990 | } else { |
9cbc5b97 PM |
1991 | WRITE_ONCE(rcu_state.gp_flags, |
1992 | rcu_state.gp_flags & RCU_GP_FLAG_INIT); | |
bb311ecc | 1993 | } |
67c583a7 | 1994 | raw_spin_unlock_irq_rcu_node(rnp); |
7fdefc10 PM |
1995 | } |
1996 | ||
1997 | /* | |
1998 | * Body of kthread that handles grace periods. | |
1999 | */ | |
0854a05c | 2000 | static int __noreturn rcu_gp_kthread(void *unused) |
7fdefc10 | 2001 | { |
5871968d | 2002 | rcu_bind_gp_kthread(); |
7fdefc10 PM |
2003 | for (;;) { |
2004 | ||
2005 | /* Handle grace-period start. */ | |
2006 | for (;;) { | |
0f11ad32 | 2007 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
63c4db78 | 2008 | TPS("reqwait")); |
9cbc5b97 PM |
2009 | rcu_state.gp_state = RCU_GP_WAIT_GPS; |
2010 | swait_event_idle_exclusive(rcu_state.gp_wq, | |
2011 | READ_ONCE(rcu_state.gp_flags) & | |
2012 | RCU_GP_FLAG_INIT); | |
55b2dcf5 | 2013 | rcu_gp_torture_wait(); |
9cbc5b97 | 2014 | rcu_state.gp_state = RCU_GP_DONE_GPS; |
78e4bc34 | 2015 | /* Locking provides needed memory barrier. */ |
0854a05c | 2016 | if (rcu_gp_init()) |
7fdefc10 | 2017 | break; |
cee43939 | 2018 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 2019 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
73a860cd | 2020 | WARN_ON(signal_pending(current)); |
0f11ad32 | 2021 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
63c4db78 | 2022 | TPS("reqwaitsig")); |
7fdefc10 | 2023 | } |
cabc49c1 | 2024 | |
4cdfc175 | 2025 | /* Handle quiescent-state forcing. */ |
c3854a05 | 2026 | rcu_gp_fqs_loop(); |
4cdfc175 PM |
2027 | |
2028 | /* Handle grace-period end. */ | |
9cbc5b97 | 2029 | rcu_state.gp_state = RCU_GP_CLEANUP; |
0854a05c | 2030 | rcu_gp_cleanup(); |
9cbc5b97 | 2031 | rcu_state.gp_state = RCU_GP_CLEANED; |
b3dbec76 | 2032 | } |
b3dbec76 PM |
2033 | } |
2034 | ||
f41d911f | 2035 | /* |
49918a54 PM |
2036 | * Report a full set of quiescent states to the rcu_state data structure. |
2037 | * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if | |
2038 | * another grace period is required. Whether we wake the grace-period | |
2039 | * kthread or it awakens itself for the next round of quiescent-state | |
2040 | * forcing, that kthread will clean up after the just-completed grace | |
2041 | * period. Note that the caller must hold rnp->lock, which is released | |
2042 | * before return. | |
f41d911f | 2043 | */ |
aff4e9ed | 2044 | static void rcu_report_qs_rsp(unsigned long flags) |
336a4f6c | 2045 | __releases(rcu_get_root()->lock) |
f41d911f | 2046 | { |
336a4f6c | 2047 | raw_lockdep_assert_held_rcu_node(rcu_get_root()); |
de8e8730 | 2048 | WARN_ON_ONCE(!rcu_gp_in_progress()); |
9cbc5b97 PM |
2049 | WRITE_ONCE(rcu_state.gp_flags, |
2050 | READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS); | |
336a4f6c | 2051 | raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags); |
532c00c9 | 2052 | rcu_gp_kthread_wake(); |
f41d911f PM |
2053 | } |
2054 | ||
64db4cff | 2055 | /* |
d3f6bad3 PM |
2056 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
2057 | * Allows quiescent states for a group of CPUs to be reported at one go | |
2058 | * to the specified rcu_node structure, though all the CPUs in the group | |
654e9533 PM |
2059 | * must be represented by the same rcu_node structure (which need not be a |
2060 | * leaf rcu_node structure, though it often will be). The gps parameter | |
2061 | * is the grace-period snapshot, which means that the quiescent states | |
c9a24e2d | 2062 | * are valid only if rnp->gp_seq is equal to gps. That structure's lock |
654e9533 | 2063 | * must be held upon entry, and it is released before return. |
ec2c2976 PM |
2064 | * |
2065 | * As a special case, if mask is zero, the bit-already-cleared check is | |
2066 | * disabled. This allows propagating quiescent state due to resumed tasks | |
2067 | * during grace-period initialization. | |
64db4cff | 2068 | */ |
b50912d0 PM |
2069 | static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp, |
2070 | unsigned long gps, unsigned long flags) | |
64db4cff PM |
2071 | __releases(rnp->lock) |
2072 | { | |
654e9533 | 2073 | unsigned long oldmask = 0; |
28ecd580 PM |
2074 | struct rcu_node *rnp_c; |
2075 | ||
a32e01ee | 2076 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 2077 | |
64db4cff PM |
2078 | /* Walk up the rcu_node hierarchy. */ |
2079 | for (;;) { | |
ec2c2976 | 2080 | if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) { |
64db4cff | 2081 | |
654e9533 PM |
2082 | /* |
2083 | * Our bit has already been cleared, or the | |
2084 | * relevant grace period is already over, so done. | |
2085 | */ | |
67c583a7 | 2086 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2087 | return; |
2088 | } | |
654e9533 | 2089 | WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ |
5b4c11d5 | 2090 | WARN_ON_ONCE(!rcu_is_leaf_node(rnp) && |
2dee9404 | 2091 | rcu_preempt_blocked_readers_cgp(rnp)); |
7672d647 | 2092 | WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask); |
67a0edbf | 2093 | trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq, |
d4c08f2a PM |
2094 | mask, rnp->qsmask, rnp->level, |
2095 | rnp->grplo, rnp->grphi, | |
2096 | !!rnp->gp_tasks); | |
27f4d280 | 2097 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
2098 | |
2099 | /* Other bits still set at this level, so done. */ | |
67c583a7 | 2100 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2101 | return; |
2102 | } | |
d43a5d32 | 2103 | rnp->completedqs = rnp->gp_seq; |
64db4cff PM |
2104 | mask = rnp->grpmask; |
2105 | if (rnp->parent == NULL) { | |
2106 | ||
2107 | /* No more levels. Exit loop holding root lock. */ | |
2108 | ||
2109 | break; | |
2110 | } | |
67c583a7 | 2111 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
28ecd580 | 2112 | rnp_c = rnp; |
64db4cff | 2113 | rnp = rnp->parent; |
2a67e741 | 2114 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
0937d045 | 2115 | oldmask = READ_ONCE(rnp_c->qsmask); |
64db4cff PM |
2116 | } |
2117 | ||
2118 | /* | |
2119 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 2120 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 2121 | * to clean up and start the next grace period if one is needed. |
64db4cff | 2122 | */ |
aff4e9ed | 2123 | rcu_report_qs_rsp(flags); /* releases rnp->lock. */ |
64db4cff PM |
2124 | } |
2125 | ||
cc99a310 PM |
2126 | /* |
2127 | * Record a quiescent state for all tasks that were previously queued | |
2128 | * on the specified rcu_node structure and that were blocking the current | |
49918a54 | 2129 | * RCU grace period. The caller must hold the corresponding rnp->lock with |
cc99a310 PM |
2130 | * irqs disabled, and this lock is released upon return, but irqs remain |
2131 | * disabled. | |
2132 | */ | |
17a8212b | 2133 | static void __maybe_unused |
139ad4da | 2134 | rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
cc99a310 PM |
2135 | __releases(rnp->lock) |
2136 | { | |
654e9533 | 2137 | unsigned long gps; |
cc99a310 PM |
2138 | unsigned long mask; |
2139 | struct rcu_node *rnp_p; | |
2140 | ||
a32e01ee | 2141 | raw_lockdep_assert_held_rcu_node(rnp); |
c130d2dc | 2142 | if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) || |
c74859d1 PM |
2143 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) || |
2144 | rnp->qsmask != 0) { | |
67c583a7 | 2145 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
cc99a310 PM |
2146 | return; /* Still need more quiescent states! */ |
2147 | } | |
2148 | ||
77cfc7bf | 2149 | rnp->completedqs = rnp->gp_seq; |
cc99a310 PM |
2150 | rnp_p = rnp->parent; |
2151 | if (rnp_p == NULL) { | |
2152 | /* | |
a77da14c PM |
2153 | * Only one rcu_node structure in the tree, so don't |
2154 | * try to report up to its nonexistent parent! | |
cc99a310 | 2155 | */ |
aff4e9ed | 2156 | rcu_report_qs_rsp(flags); |
cc99a310 PM |
2157 | return; |
2158 | } | |
2159 | ||
c9a24e2d PM |
2160 | /* Report up the rest of the hierarchy, tracking current ->gp_seq. */ |
2161 | gps = rnp->gp_seq; | |
cc99a310 | 2162 | mask = rnp->grpmask; |
67c583a7 | 2163 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
2a67e741 | 2164 | raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */ |
b50912d0 | 2165 | rcu_report_qs_rnp(mask, rnp_p, gps, flags); |
cc99a310 PM |
2166 | } |
2167 | ||
64db4cff | 2168 | /* |
d3f6bad3 | 2169 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
4b455dc3 | 2170 | * structure. This must be called from the specified CPU. |
64db4cff PM |
2171 | */ |
2172 | static void | |
33085c46 | 2173 | rcu_report_qs_rdp(int cpu, struct rcu_data *rdp) |
64db4cff PM |
2174 | { |
2175 | unsigned long flags; | |
2176 | unsigned long mask; | |
5d6742b3 PM |
2177 | bool needwake = false; |
2178 | const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) && | |
2179 | rcu_segcblist_is_offloaded(&rdp->cblist); | |
64db4cff PM |
2180 | struct rcu_node *rnp; |
2181 | ||
2182 | rnp = rdp->mynode; | |
2a67e741 | 2183 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
c9a24e2d PM |
2184 | if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq || |
2185 | rdp->gpwrap) { | |
64db4cff PM |
2186 | |
2187 | /* | |
e4cc1f22 PM |
2188 | * The grace period in which this quiescent state was |
2189 | * recorded has ended, so don't report it upwards. | |
2190 | * We will instead need a new quiescent state that lies | |
2191 | * within the current grace period. | |
64db4cff | 2192 | */ |
5b74c458 | 2193 | rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */ |
67c583a7 | 2194 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2195 | return; |
2196 | } | |
2197 | mask = rdp->grpmask; | |
b5ea0370 PM |
2198 | if (rdp->cpu == smp_processor_id()) |
2199 | rdp->core_needs_qs = false; | |
64db4cff | 2200 | if ((rnp->qsmask & mask) == 0) { |
67c583a7 | 2201 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff | 2202 | } else { |
64db4cff PM |
2203 | /* |
2204 | * This GP can't end until cpu checks in, so all of our | |
2205 | * callbacks can be processed during the next GP. | |
2206 | */ | |
5d6742b3 PM |
2207 | if (!offloaded) |
2208 | needwake = rcu_accelerate_cbs(rnp, rdp); | |
64db4cff | 2209 | |
516e5ae0 | 2210 | rcu_disable_urgency_upon_qs(rdp); |
b50912d0 | 2211 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
654e9533 | 2212 | /* ^^^ Released rnp->lock */ |
48a7639c | 2213 | if (needwake) |
532c00c9 | 2214 | rcu_gp_kthread_wake(); |
64db4cff PM |
2215 | } |
2216 | } | |
2217 | ||
2218 | /* | |
2219 | * Check to see if there is a new grace period of which this CPU | |
2220 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2221 | * Otherwise, see if this CPU has just passed through its first | |
2222 | * quiescent state for this grace period, and record that fact if so. | |
2223 | */ | |
2224 | static void | |
8087d3e3 | 2225 | rcu_check_quiescent_state(struct rcu_data *rdp) |
64db4cff | 2226 | { |
05eb552b | 2227 | /* Check for grace-period ends and beginnings. */ |
15cabdff | 2228 | note_gp_changes(rdp); |
64db4cff PM |
2229 | |
2230 | /* | |
2231 | * Does this CPU still need to do its part for current grace period? | |
2232 | * If no, return and let the other CPUs do their part as well. | |
2233 | */ | |
97c668b8 | 2234 | if (!rdp->core_needs_qs) |
64db4cff PM |
2235 | return; |
2236 | ||
2237 | /* | |
2238 | * Was there a quiescent state since the beginning of the grace | |
2239 | * period? If no, then exit and wait for the next call. | |
2240 | */ | |
3a19b46a | 2241 | if (rdp->cpu_no_qs.b.norm) |
64db4cff PM |
2242 | return; |
2243 | ||
d3f6bad3 PM |
2244 | /* |
2245 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2246 | * judge of that). | |
2247 | */ | |
33085c46 | 2248 | rcu_report_qs_rdp(rdp->cpu, rdp); |
64db4cff PM |
2249 | } |
2250 | ||
b1420f1c | 2251 | /* |
780cd590 PM |
2252 | * Near the end of the offline process. Trace the fact that this CPU |
2253 | * is going offline. | |
b1420f1c | 2254 | */ |
780cd590 | 2255 | int rcutree_dying_cpu(unsigned int cpu) |
b1420f1c | 2256 | { |
4f5fbd78 YS |
2257 | bool blkd; |
2258 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); | |
2259 | struct rcu_node *rnp = rdp->mynode; | |
b1420f1c | 2260 | |
ea46351c | 2261 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
780cd590 | 2262 | return 0; |
ea46351c | 2263 | |
4f5fbd78 | 2264 | blkd = !!(rnp->qsmask & rdp->grpmask); |
0937d045 | 2265 | trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq), |
477351f7 | 2266 | blkd ? TPS("cpuofl") : TPS("cpuofl-bgp")); |
780cd590 | 2267 | return 0; |
64db4cff PM |
2268 | } |
2269 | ||
8af3a5e7 PM |
2270 | /* |
2271 | * All CPUs for the specified rcu_node structure have gone offline, | |
2272 | * and all tasks that were preempted within an RCU read-side critical | |
2273 | * section while running on one of those CPUs have since exited their RCU | |
2274 | * read-side critical section. Some other CPU is reporting this fact with | |
2275 | * the specified rcu_node structure's ->lock held and interrupts disabled. | |
2276 | * This function therefore goes up the tree of rcu_node structures, | |
2277 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that | |
2278 | * the leaf rcu_node structure's ->qsmaskinit field has already been | |
c50cbe53 | 2279 | * updated. |
8af3a5e7 PM |
2280 | * |
2281 | * This function does check that the specified rcu_node structure has | |
2282 | * all CPUs offline and no blocked tasks, so it is OK to invoke it | |
2283 | * prematurely. That said, invoking it after the fact will cost you | |
2284 | * a needless lock acquisition. So once it has done its work, don't | |
2285 | * invoke it again. | |
2286 | */ | |
2287 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) | |
2288 | { | |
2289 | long mask; | |
2290 | struct rcu_node *rnp = rnp_leaf; | |
2291 | ||
962aff03 | 2292 | raw_lockdep_assert_held_rcu_node(rnp_leaf); |
ea46351c | 2293 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || |
962aff03 PM |
2294 | WARN_ON_ONCE(rnp_leaf->qsmaskinit) || |
2295 | WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf))) | |
8af3a5e7 PM |
2296 | return; |
2297 | for (;;) { | |
2298 | mask = rnp->grpmask; | |
2299 | rnp = rnp->parent; | |
2300 | if (!rnp) | |
2301 | break; | |
2a67e741 | 2302 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
8af3a5e7 | 2303 | rnp->qsmaskinit &= ~mask; |
962aff03 PM |
2304 | /* Between grace periods, so better already be zero! */ |
2305 | WARN_ON_ONCE(rnp->qsmask); | |
8af3a5e7 | 2306 | if (rnp->qsmaskinit) { |
67c583a7 BF |
2307 | raw_spin_unlock_rcu_node(rnp); |
2308 | /* irqs remain disabled. */ | |
8af3a5e7 PM |
2309 | return; |
2310 | } | |
67c583a7 | 2311 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
8af3a5e7 PM |
2312 | } |
2313 | } | |
2314 | ||
64db4cff | 2315 | /* |
e5601400 | 2316 | * The CPU has been completely removed, and some other CPU is reporting |
a58163d8 PM |
2317 | * this fact from process context. Do the remainder of the cleanup. |
2318 | * There can only be one CPU hotplug operation at a time, so no need for | |
2319 | * explicit locking. | |
64db4cff | 2320 | */ |
780cd590 | 2321 | int rcutree_dead_cpu(unsigned int cpu) |
64db4cff | 2322 | { |
da1df50d | 2323 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
b1420f1c | 2324 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2325 | |
ea46351c | 2326 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
780cd590 | 2327 | return 0; |
ea46351c | 2328 | |
2036d94a | 2329 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2330 | rcu_boost_kthread_setaffinity(rnp, -1); |
780cd590 PM |
2331 | /* Do any needed no-CB deferred wakeups from this CPU. */ |
2332 | do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu)); | |
96926686 PM |
2333 | |
2334 | // Stop-machine done, so allow nohz_full to disable tick. | |
2335 | tick_dep_clear(TICK_DEP_BIT_RCU); | |
780cd590 | 2336 | return 0; |
64db4cff PM |
2337 | } |
2338 | ||
64db4cff PM |
2339 | /* |
2340 | * Invoke any RCU callbacks that have made it to the end of their grace | |
2341 | * period. Thottle as specified by rdp->blimit. | |
2342 | */ | |
5bb5d09c | 2343 | static void rcu_do_batch(struct rcu_data *rdp) |
64db4cff PM |
2344 | { |
2345 | unsigned long flags; | |
ec5ef87b PM |
2346 | const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) && |
2347 | rcu_segcblist_is_offloaded(&rdp->cblist); | |
15fecf89 PM |
2348 | struct rcu_head *rhp; |
2349 | struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl); | |
2350 | long bl, count; | |
cfcdef5e | 2351 | long pending, tlimit = 0; |
64db4cff | 2352 | |
dc35c893 | 2353 | /* If no callbacks are ready, just return. */ |
15fecf89 | 2354 | if (!rcu_segcblist_ready_cbs(&rdp->cblist)) { |
3c779dfe | 2355 | trace_rcu_batch_start(rcu_state.name, |
15fecf89 | 2356 | rcu_segcblist_n_cbs(&rdp->cblist), 0); |
3c779dfe | 2357 | trace_rcu_batch_end(rcu_state.name, 0, |
15fecf89 | 2358 | !rcu_segcblist_empty(&rdp->cblist), |
4968c300 PM |
2359 | need_resched(), is_idle_task(current), |
2360 | rcu_is_callbacks_kthread()); | |
64db4cff | 2361 | return; |
29c00b4a | 2362 | } |
64db4cff PM |
2363 | |
2364 | /* | |
2365 | * Extract the list of ready callbacks, disabling to prevent | |
15fecf89 PM |
2366 | * races with call_rcu() from interrupt handlers. Leave the |
2367 | * callback counts, as rcu_barrier() needs to be conservative. | |
64db4cff PM |
2368 | */ |
2369 | local_irq_save(flags); | |
5d6742b3 | 2370 | rcu_nocb_lock(rdp); |
8146c4e2 | 2371 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
cfcdef5e ED |
2372 | pending = rcu_segcblist_n_cbs(&rdp->cblist); |
2373 | bl = max(rdp->blimit, pending >> rcu_divisor); | |
2374 | if (unlikely(bl > 100)) | |
2375 | tlimit = local_clock() + rcu_resched_ns; | |
3c779dfe | 2376 | trace_rcu_batch_start(rcu_state.name, |
15fecf89 PM |
2377 | rcu_segcblist_n_cbs(&rdp->cblist), bl); |
2378 | rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl); | |
7f36ef82 PM |
2379 | if (offloaded) |
2380 | rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist); | |
5d6742b3 | 2381 | rcu_nocb_unlock_irqrestore(rdp, flags); |
64db4cff PM |
2382 | |
2383 | /* Invoke callbacks. */ | |
6a949b7a | 2384 | tick_dep_set_task(current, TICK_DEP_BIT_RCU); |
15fecf89 PM |
2385 | rhp = rcu_cblist_dequeue(&rcl); |
2386 | for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) { | |
77a40f97 JFG |
2387 | rcu_callback_t f; |
2388 | ||
15fecf89 | 2389 | debug_rcu_head_unqueue(rhp); |
77a40f97 JFG |
2390 | |
2391 | rcu_lock_acquire(&rcu_callback_map); | |
2392 | trace_rcu_invoke_callback(rcu_state.name, rhp); | |
2393 | ||
2394 | f = rhp->func; | |
2395 | WRITE_ONCE(rhp->func, (rcu_callback_t)0L); | |
2396 | f(rhp); | |
2397 | ||
2398 | rcu_lock_release(&rcu_callback_map); | |
2399 | ||
15fecf89 PM |
2400 | /* |
2401 | * Stop only if limit reached and CPU has something to do. | |
2402 | * Note: The rcl structure counts down from zero. | |
2403 | */ | |
ec5ef87b | 2404 | if (-rcl.len >= bl && !offloaded && |
dff1672d PM |
2405 | (need_resched() || |
2406 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff | 2407 | break; |
cfcdef5e ED |
2408 | if (unlikely(tlimit)) { |
2409 | /* only call local_clock() every 32 callbacks */ | |
2410 | if (likely((-rcl.len & 31) || local_clock() < tlimit)) | |
2411 | continue; | |
2412 | /* Exceeded the time limit, so leave. */ | |
2413 | break; | |
2414 | } | |
ec5ef87b | 2415 | if (offloaded) { |
5d6742b3 PM |
2416 | WARN_ON_ONCE(in_serving_softirq()); |
2417 | local_bh_enable(); | |
2418 | lockdep_assert_irqs_enabled(); | |
2419 | cond_resched_tasks_rcu_qs(); | |
2420 | lockdep_assert_irqs_enabled(); | |
2421 | local_bh_disable(); | |
2422 | } | |
64db4cff PM |
2423 | } |
2424 | ||
2425 | local_irq_save(flags); | |
5d6742b3 | 2426 | rcu_nocb_lock(rdp); |
4b27f20b | 2427 | count = -rcl.len; |
3c779dfe | 2428 | trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(), |
8ef0f37e | 2429 | is_idle_task(current), rcu_is_callbacks_kthread()); |
64db4cff | 2430 | |
15fecf89 PM |
2431 | /* Update counts and requeue any remaining callbacks. */ |
2432 | rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl); | |
b1420f1c | 2433 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
15fecf89 | 2434 | rcu_segcblist_insert_count(&rdp->cblist, &rcl); |
64db4cff PM |
2435 | |
2436 | /* Reinstate batch limit if we have worked down the excess. */ | |
15fecf89 | 2437 | count = rcu_segcblist_n_cbs(&rdp->cblist); |
d5a9a8c3 | 2438 | if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark) |
64db4cff PM |
2439 | rdp->blimit = blimit; |
2440 | ||
37c72e56 | 2441 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
15fecf89 | 2442 | if (count == 0 && rdp->qlen_last_fqs_check != 0) { |
37c72e56 | 2443 | rdp->qlen_last_fqs_check = 0; |
3c779dfe | 2444 | rdp->n_force_qs_snap = rcu_state.n_force_qs; |
15fecf89 PM |
2445 | } else if (count < rdp->qlen_last_fqs_check - qhimark) |
2446 | rdp->qlen_last_fqs_check = count; | |
efd88b02 PM |
2447 | |
2448 | /* | |
2449 | * The following usually indicates a double call_rcu(). To track | |
2450 | * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y. | |
2451 | */ | |
d1b222c6 PM |
2452 | WARN_ON_ONCE(count == 0 && !rcu_segcblist_empty(&rdp->cblist)); |
2453 | WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) && | |
2454 | count != 0 && rcu_segcblist_empty(&rdp->cblist)); | |
37c72e56 | 2455 | |
5d6742b3 | 2456 | rcu_nocb_unlock_irqrestore(rdp, flags); |
64db4cff | 2457 | |
e0f23060 | 2458 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
ec5ef87b | 2459 | if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist)) |
a46e0899 | 2460 | invoke_rcu_core(); |
6a949b7a | 2461 | tick_dep_clear_task(current, TICK_DEP_BIT_RCU); |
64db4cff PM |
2462 | } |
2463 | ||
2464 | /* | |
c98cac60 PM |
2465 | * This function is invoked from each scheduling-clock interrupt, |
2466 | * and checks to see if this CPU is in a non-context-switch quiescent | |
2467 | * state, for example, user mode or idle loop. It also schedules RCU | |
2468 | * core processing. If the current grace period has gone on too long, | |
2469 | * it will ask the scheduler to manufacture a context switch for the sole | |
2470 | * purpose of providing a providing the needed quiescent state. | |
64db4cff | 2471 | */ |
c98cac60 | 2472 | void rcu_sched_clock_irq(int user) |
64db4cff | 2473 | { |
f7f7bac9 | 2474 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
4e95020c | 2475 | raw_cpu_inc(rcu_data.ticks_this_gp); |
92aa39e9 | 2476 | /* The load-acquire pairs with the store-release setting to true. */ |
2dba13f0 | 2477 | if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) { |
92aa39e9 | 2478 | /* Idle and userspace execution already are quiescent states. */ |
a0ef9ec2 | 2479 | if (!rcu_is_cpu_rrupt_from_idle() && !user) { |
92aa39e9 PM |
2480 | set_tsk_need_resched(current); |
2481 | set_preempt_need_resched(); | |
2482 | } | |
2dba13f0 | 2483 | __this_cpu_write(rcu_data.rcu_urgent_qs, false); |
64db4cff | 2484 | } |
c98cac60 | 2485 | rcu_flavor_sched_clock_irq(user); |
dd7dafd1 | 2486 | if (rcu_pending(user)) |
a46e0899 | 2487 | invoke_rcu_core(); |
07f27570 | 2488 | |
f7f7bac9 | 2489 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2490 | } |
2491 | ||
64db4cff | 2492 | /* |
5d8a752e ZZ |
2493 | * Scan the leaf rcu_node structures. For each structure on which all |
2494 | * CPUs have reported a quiescent state and on which there are tasks | |
2495 | * blocking the current grace period, initiate RCU priority boosting. | |
2496 | * Otherwise, invoke the specified function to check dyntick state for | |
2497 | * each CPU that has not yet reported a quiescent state. | |
64db4cff | 2498 | */ |
8ff0b907 | 2499 | static void force_qs_rnp(int (*f)(struct rcu_data *rdp)) |
64db4cff | 2500 | { |
64db4cff PM |
2501 | int cpu; |
2502 | unsigned long flags; | |
2503 | unsigned long mask; | |
66e4c33b | 2504 | struct rcu_data *rdp; |
a0b6c9a7 | 2505 | struct rcu_node *rnp; |
64db4cff | 2506 | |
b2b00ddf PM |
2507 | rcu_state.cbovld = rcu_state.cbovldnext; |
2508 | rcu_state.cbovldnext = false; | |
aedf4ba9 | 2509 | rcu_for_each_leaf_node(rnp) { |
cee43939 | 2510 | cond_resched_tasks_rcu_qs(); |
64db4cff | 2511 | mask = 0; |
2a67e741 | 2512 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
b2b00ddf | 2513 | rcu_state.cbovldnext |= !!rnp->cbovldmask; |
a0b6c9a7 | 2514 | if (rnp->qsmask == 0) { |
c130d2dc | 2515 | if (!IS_ENABLED(CONFIG_PREEMPT_RCU) || |
a77da14c PM |
2516 | rcu_preempt_blocked_readers_cgp(rnp)) { |
2517 | /* | |
2518 | * No point in scanning bits because they | |
2519 | * are all zero. But we might need to | |
2520 | * priority-boost blocked readers. | |
2521 | */ | |
2522 | rcu_initiate_boost(rnp, flags); | |
2523 | /* rcu_initiate_boost() releases rnp->lock */ | |
2524 | continue; | |
2525 | } | |
92816435 PM |
2526 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
2527 | continue; | |
64db4cff | 2528 | } |
7441e766 PM |
2529 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) { |
2530 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
2531 | if (f(rdp)) { | |
2532 | mask |= rdp->grpmask; | |
2533 | rcu_disable_urgency_upon_qs(rdp); | |
0edd1b17 | 2534 | } |
64db4cff | 2535 | } |
45f014c5 | 2536 | if (mask != 0) { |
c9a24e2d | 2537 | /* Idle/offline CPUs, report (releases rnp->lock). */ |
b50912d0 | 2538 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
0aa04b05 PM |
2539 | } else { |
2540 | /* Nothing to do here, so just drop the lock. */ | |
67c583a7 | 2541 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff | 2542 | } |
64db4cff | 2543 | } |
64db4cff PM |
2544 | } |
2545 | ||
2546 | /* | |
2547 | * Force quiescent states on reluctant CPUs, and also detect which | |
2548 | * CPUs are in dyntick-idle mode. | |
2549 | */ | |
cd920e5a | 2550 | void rcu_force_quiescent_state(void) |
64db4cff PM |
2551 | { |
2552 | unsigned long flags; | |
394f2769 PM |
2553 | bool ret; |
2554 | struct rcu_node *rnp; | |
2555 | struct rcu_node *rnp_old = NULL; | |
2556 | ||
2557 | /* Funnel through hierarchy to reduce memory contention. */ | |
da1df50d | 2558 | rnp = __this_cpu_read(rcu_data.mynode); |
394f2769 | 2559 | for (; rnp != NULL; rnp = rnp->parent) { |
67a0edbf | 2560 | ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) || |
66e4c33b | 2561 | !raw_spin_trylock(&rnp->fqslock); |
394f2769 PM |
2562 | if (rnp_old != NULL) |
2563 | raw_spin_unlock(&rnp_old->fqslock); | |
d62df573 | 2564 | if (ret) |
394f2769 | 2565 | return; |
394f2769 PM |
2566 | rnp_old = rnp; |
2567 | } | |
336a4f6c | 2568 | /* rnp_old == rcu_get_root(), rnp == NULL. */ |
64db4cff | 2569 | |
394f2769 | 2570 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2a67e741 | 2571 | raw_spin_lock_irqsave_rcu_node(rnp_old, flags); |
394f2769 | 2572 | raw_spin_unlock(&rnp_old->fqslock); |
67a0edbf | 2573 | if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) { |
67c583a7 | 2574 | raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags); |
4cdfc175 | 2575 | return; /* Someone beat us to it. */ |
46a1e34e | 2576 | } |
67a0edbf PM |
2577 | WRITE_ONCE(rcu_state.gp_flags, |
2578 | READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS); | |
67c583a7 | 2579 | raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags); |
532c00c9 | 2580 | rcu_gp_kthread_wake(); |
64db4cff | 2581 | } |
cd920e5a | 2582 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); |
64db4cff | 2583 | |
fb60e533 | 2584 | /* Perform RCU core processing work for the current CPU. */ |
48d07c04 | 2585 | static __latent_entropy void rcu_core(void) |
64db4cff PM |
2586 | { |
2587 | unsigned long flags; | |
da1df50d | 2588 | struct rcu_data *rdp = raw_cpu_ptr(&rcu_data); |
26d950a9 | 2589 | struct rcu_node *rnp = rdp->mynode; |
c1ab99d6 PM |
2590 | const bool offloaded = IS_ENABLED(CONFIG_RCU_NOCB_CPU) && |
2591 | rcu_segcblist_is_offloaded(&rdp->cblist); | |
64db4cff | 2592 | |
b049fdf8 PM |
2593 | if (cpu_is_offline(smp_processor_id())) |
2594 | return; | |
2595 | trace_rcu_utilization(TPS("Start RCU core")); | |
50dc7def | 2596 | WARN_ON_ONCE(!rdp->beenonline); |
2e597558 | 2597 | |
3e310098 | 2598 | /* Report any deferred quiescent states if preemption enabled. */ |
fced9c8c | 2599 | if (!(preempt_count() & PREEMPT_MASK)) { |
3e310098 | 2600 | rcu_preempt_deferred_qs(current); |
fced9c8c PM |
2601 | } else if (rcu_preempt_need_deferred_qs(current)) { |
2602 | set_tsk_need_resched(current); | |
2603 | set_preempt_need_resched(); | |
2604 | } | |
3e310098 | 2605 | |
64db4cff | 2606 | /* Update RCU state based on any recent quiescent states. */ |
8087d3e3 | 2607 | rcu_check_quiescent_state(rdp); |
64db4cff | 2608 | |
bd7af846 | 2609 | /* No grace period and unregistered callbacks? */ |
de8e8730 | 2610 | if (!rcu_gp_in_progress() && |
c1ab99d6 | 2611 | rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) { |
bd7af846 | 2612 | local_irq_save(flags); |
e44e73ca | 2613 | if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL)) |
c6e09b97 | 2614 | rcu_accelerate_cbs_unlocked(rnp, rdp); |
e44e73ca | 2615 | local_irq_restore(flags); |
64db4cff PM |
2616 | } |
2617 | ||
791416c4 | 2618 | rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check()); |
26d950a9 | 2619 | |
64db4cff | 2620 | /* If there are callbacks ready, invoke them. */ |
c1ab99d6 | 2621 | if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist) && |
43e903ad PM |
2622 | likely(READ_ONCE(rcu_scheduler_fully_active))) |
2623 | rcu_do_batch(rdp); | |
96d3fd0d PM |
2624 | |
2625 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2626 | do_nocb_deferred_wakeup(rdp); | |
f7f7bac9 | 2627 | trace_rcu_utilization(TPS("End RCU core")); |
64db4cff PM |
2628 | } |
2629 | ||
48d07c04 SAS |
2630 | static void rcu_core_si(struct softirq_action *h) |
2631 | { | |
2632 | rcu_core(); | |
2633 | } | |
2634 | ||
2635 | static void rcu_wake_cond(struct task_struct *t, int status) | |
2636 | { | |
2637 | /* | |
2638 | * If the thread is yielding, only wake it when this | |
2639 | * is invoked from idle | |
2640 | */ | |
2641 | if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current))) | |
2642 | wake_up_process(t); | |
2643 | } | |
2644 | ||
2645 | static void invoke_rcu_core_kthread(void) | |
2646 | { | |
2647 | struct task_struct *t; | |
2648 | unsigned long flags; | |
2649 | ||
2650 | local_irq_save(flags); | |
2651 | __this_cpu_write(rcu_data.rcu_cpu_has_work, 1); | |
2652 | t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task); | |
2653 | if (t != NULL && t != current) | |
2654 | rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status)); | |
2655 | local_irq_restore(flags); | |
2656 | } | |
2657 | ||
48d07c04 SAS |
2658 | /* |
2659 | * Wake up this CPU's rcuc kthread to do RCU core processing. | |
2660 | */ | |
a46e0899 | 2661 | static void invoke_rcu_core(void) |
09223371 | 2662 | { |
48d07c04 SAS |
2663 | if (!cpu_online(smp_processor_id())) |
2664 | return; | |
2665 | if (use_softirq) | |
b0f74036 | 2666 | raise_softirq(RCU_SOFTIRQ); |
48d07c04 SAS |
2667 | else |
2668 | invoke_rcu_core_kthread(); | |
2669 | } | |
2670 | ||
2671 | static void rcu_cpu_kthread_park(unsigned int cpu) | |
2672 | { | |
2673 | per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; | |
2674 | } | |
2675 | ||
2676 | static int rcu_cpu_kthread_should_run(unsigned int cpu) | |
2677 | { | |
2678 | return __this_cpu_read(rcu_data.rcu_cpu_has_work); | |
2679 | } | |
2680 | ||
2681 | /* | |
2682 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces | |
2683 | * the RCU softirq used in configurations of RCU that do not support RCU | |
2684 | * priority boosting. | |
2685 | */ | |
2686 | static void rcu_cpu_kthread(unsigned int cpu) | |
2687 | { | |
2688 | unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status); | |
2689 | char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work); | |
2690 | int spincnt; | |
2691 | ||
2488a5e6 | 2692 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_run")); |
48d07c04 | 2693 | for (spincnt = 0; spincnt < 10; spincnt++) { |
48d07c04 SAS |
2694 | local_bh_disable(); |
2695 | *statusp = RCU_KTHREAD_RUNNING; | |
2696 | local_irq_disable(); | |
2697 | work = *workp; | |
2698 | *workp = 0; | |
2699 | local_irq_enable(); | |
2700 | if (work) | |
2701 | rcu_core(); | |
2702 | local_bh_enable(); | |
2703 | if (*workp == 0) { | |
2704 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); | |
2705 | *statusp = RCU_KTHREAD_WAITING; | |
2706 | return; | |
2707 | } | |
2708 | } | |
2709 | *statusp = RCU_KTHREAD_YIELDING; | |
2710 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); | |
2711 | schedule_timeout_interruptible(2); | |
2712 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); | |
2713 | *statusp = RCU_KTHREAD_WAITING; | |
2714 | } | |
2715 | ||
2716 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { | |
2717 | .store = &rcu_data.rcu_cpu_kthread_task, | |
2718 | .thread_should_run = rcu_cpu_kthread_should_run, | |
2719 | .thread_fn = rcu_cpu_kthread, | |
2720 | .thread_comm = "rcuc/%u", | |
2721 | .setup = rcu_cpu_kthread_setup, | |
2722 | .park = rcu_cpu_kthread_park, | |
2723 | }; | |
2724 | ||
2725 | /* | |
2726 | * Spawn per-CPU RCU core processing kthreads. | |
2727 | */ | |
2728 | static int __init rcu_spawn_core_kthreads(void) | |
2729 | { | |
2730 | int cpu; | |
2731 | ||
2732 | for_each_possible_cpu(cpu) | |
2733 | per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0; | |
2734 | if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq) | |
2735 | return 0; | |
2736 | WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec), | |
2737 | "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__); | |
2738 | return 0; | |
09223371 | 2739 | } |
48d07c04 | 2740 | early_initcall(rcu_spawn_core_kthreads); |
09223371 | 2741 | |
29154c57 PM |
2742 | /* |
2743 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2744 | */ | |
5c7d8967 PM |
2745 | static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head, |
2746 | unsigned long flags) | |
64db4cff | 2747 | { |
62fde6ed PM |
2748 | /* |
2749 | * If called from an extended quiescent state, invoke the RCU | |
2750 | * core in order to force a re-evaluation of RCU's idleness. | |
2751 | */ | |
9910affa | 2752 | if (!rcu_is_watching()) |
62fde6ed PM |
2753 | invoke_rcu_core(); |
2754 | ||
a16b7a69 | 2755 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2756 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2757 | return; |
64db4cff | 2758 | |
37c72e56 PM |
2759 | /* |
2760 | * Force the grace period if too many callbacks or too long waiting. | |
cd920e5a | 2761 | * Enforce hysteresis, and don't invoke rcu_force_quiescent_state() |
37c72e56 | 2762 | * if some other CPU has recently done so. Also, don't bother |
cd920e5a | 2763 | * invoking rcu_force_quiescent_state() if the newly enqueued callback |
37c72e56 PM |
2764 | * is the only one waiting for a grace period to complete. |
2765 | */ | |
15fecf89 PM |
2766 | if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) > |
2767 | rdp->qlen_last_fqs_check + qhimark)) { | |
b52573d2 PM |
2768 | |
2769 | /* Are we ignoring a completed grace period? */ | |
15cabdff | 2770 | note_gp_changes(rdp); |
b52573d2 PM |
2771 | |
2772 | /* Start a new grace period if one not already started. */ | |
de8e8730 | 2773 | if (!rcu_gp_in_progress()) { |
c6e09b97 | 2774 | rcu_accelerate_cbs_unlocked(rdp->mynode, rdp); |
b52573d2 PM |
2775 | } else { |
2776 | /* Give the grace period a kick. */ | |
d5a9a8c3 | 2777 | rdp->blimit = DEFAULT_MAX_RCU_BLIMIT; |
5c7d8967 | 2778 | if (rcu_state.n_force_qs == rdp->n_force_qs_snap && |
15fecf89 | 2779 | rcu_segcblist_first_pend_cb(&rdp->cblist) != head) |
cd920e5a | 2780 | rcu_force_quiescent_state(); |
5c7d8967 | 2781 | rdp->n_force_qs_snap = rcu_state.n_force_qs; |
15fecf89 | 2782 | rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist); |
b52573d2 | 2783 | } |
4cdfc175 | 2784 | } |
29154c57 PM |
2785 | } |
2786 | ||
ae150184 PM |
2787 | /* |
2788 | * RCU callback function to leak a callback. | |
2789 | */ | |
2790 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2791 | { | |
2792 | } | |
2793 | ||
3fbfbf7a | 2794 | /* |
b2b00ddf PM |
2795 | * Check and if necessary update the leaf rcu_node structure's |
2796 | * ->cbovldmask bit corresponding to the current CPU based on that CPU's | |
2797 | * number of queued RCU callbacks. The caller must hold the leaf rcu_node | |
2798 | * structure's ->lock. | |
3fbfbf7a | 2799 | */ |
b2b00ddf PM |
2800 | static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp) |
2801 | { | |
2802 | raw_lockdep_assert_held_rcu_node(rnp); | |
2803 | if (qovld_calc <= 0) | |
2804 | return; // Early boot and wildcard value set. | |
2805 | if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) | |
2806 | WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask); | |
2807 | else | |
2808 | WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask); | |
2809 | } | |
2810 | ||
2811 | /* | |
2812 | * Check and if necessary update the leaf rcu_node structure's | |
2813 | * ->cbovldmask bit corresponding to the current CPU based on that CPU's | |
2814 | * number of queued RCU callbacks. No locks need be held, but the | |
2815 | * caller must have disabled interrupts. | |
2816 | * | |
2817 | * Note that this function ignores the possibility that there are a lot | |
2818 | * of callbacks all of which have already seen the end of their respective | |
2819 | * grace periods. This omission is due to the need for no-CBs CPUs to | |
2820 | * be holding ->nocb_lock to do this check, which is too heavy for a | |
2821 | * common-case operation. | |
3fbfbf7a | 2822 | */ |
b2b00ddf PM |
2823 | static void check_cb_ovld(struct rcu_data *rdp) |
2824 | { | |
2825 | struct rcu_node *const rnp = rdp->mynode; | |
2826 | ||
2827 | if (qovld_calc <= 0 || | |
2828 | ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) == | |
2829 | !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask))) | |
2830 | return; // Early boot wildcard value or already set correctly. | |
2831 | raw_spin_lock_rcu_node(rnp); | |
2832 | check_cb_ovld_locked(rdp, rnp); | |
2833 | raw_spin_unlock_rcu_node(rnp); | |
2834 | } | |
2835 | ||
b692dc4a | 2836 | /* Helper function for call_rcu() and friends. */ |
64db4cff | 2837 | static void |
77a40f97 | 2838 | __call_rcu(struct rcu_head *head, rcu_callback_t func) |
64db4cff PM |
2839 | { |
2840 | unsigned long flags; | |
2841 | struct rcu_data *rdp; | |
5d6742b3 | 2842 | bool was_alldone; |
64db4cff | 2843 | |
b8f2ed53 PM |
2844 | /* Misaligned rcu_head! */ |
2845 | WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1)); | |
2846 | ||
ae150184 | 2847 | if (debug_rcu_head_queue(head)) { |
fa3c6647 PM |
2848 | /* |
2849 | * Probable double call_rcu(), so leak the callback. | |
2850 | * Use rcu:rcu_callback trace event to find the previous | |
2851 | * time callback was passed to __call_rcu(). | |
2852 | */ | |
d75f773c | 2853 | WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pS()!!!\n", |
fa3c6647 | 2854 | head, head->func); |
7d0ae808 | 2855 | WRITE_ONCE(head->func, rcu_leak_callback); |
ae150184 PM |
2856 | return; |
2857 | } | |
64db4cff PM |
2858 | head->func = func; |
2859 | head->next = NULL; | |
64db4cff | 2860 | local_irq_save(flags); |
da1df50d | 2861 | rdp = this_cpu_ptr(&rcu_data); |
64db4cff PM |
2862 | |
2863 | /* Add the callback to our list. */ | |
5d6742b3 PM |
2864 | if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) { |
2865 | // This can trigger due to call_rcu() from offline CPU: | |
2866 | WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE); | |
34404ca8 | 2867 | WARN_ON_ONCE(!rcu_is_watching()); |
5d6742b3 PM |
2868 | // Very early boot, before rcu_init(). Initialize if needed |
2869 | // and then drop through to queue the callback. | |
15fecf89 PM |
2870 | if (rcu_segcblist_empty(&rdp->cblist)) |
2871 | rcu_segcblist_init(&rdp->cblist); | |
0d8ee37e | 2872 | } |
77a40f97 | 2873 | |
b2b00ddf | 2874 | check_cb_ovld(rdp); |
d1b222c6 PM |
2875 | if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags)) |
2876 | return; // Enqueued onto ->nocb_bypass, so just leave. | |
b692dc4a | 2877 | // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock. |
77a40f97 | 2878 | rcu_segcblist_enqueue(&rdp->cblist, head); |
d4c08f2a | 2879 | if (__is_kfree_rcu_offset((unsigned long)func)) |
3c779dfe PM |
2880 | trace_rcu_kfree_callback(rcu_state.name, head, |
2881 | (unsigned long)func, | |
15fecf89 | 2882 | rcu_segcblist_n_cbs(&rdp->cblist)); |
d4c08f2a | 2883 | else |
3c779dfe | 2884 | trace_rcu_callback(rcu_state.name, head, |
15fecf89 | 2885 | rcu_segcblist_n_cbs(&rdp->cblist)); |
d4c08f2a | 2886 | |
29154c57 | 2887 | /* Go handle any RCU core processing required. */ |
5d6742b3 PM |
2888 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU) && |
2889 | unlikely(rcu_segcblist_is_offloaded(&rdp->cblist))) { | |
2890 | __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */ | |
2891 | } else { | |
2892 | __call_rcu_core(rdp, head, flags); | |
2893 | local_irq_restore(flags); | |
2894 | } | |
64db4cff PM |
2895 | } |
2896 | ||
a68a2bb2 | 2897 | /** |
45975c7d | 2898 | * call_rcu() - Queue an RCU callback for invocation after a grace period. |
a68a2bb2 PM |
2899 | * @head: structure to be used for queueing the RCU updates. |
2900 | * @func: actual callback function to be invoked after the grace period | |
2901 | * | |
2902 | * The callback function will be invoked some time after a full grace | |
45975c7d PM |
2903 | * period elapses, in other words after all pre-existing RCU read-side |
2904 | * critical sections have completed. However, the callback function | |
2905 | * might well execute concurrently with RCU read-side critical sections | |
2906 | * that started after call_rcu() was invoked. RCU read-side critical | |
2907 | * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and | |
2908 | * may be nested. In addition, regions of code across which interrupts, | |
2909 | * preemption, or softirqs have been disabled also serve as RCU read-side | |
2910 | * critical sections. This includes hardware interrupt handlers, softirq | |
2911 | * handlers, and NMI handlers. | |
2912 | * | |
2913 | * Note that all CPUs must agree that the grace period extended beyond | |
2914 | * all pre-existing RCU read-side critical section. On systems with more | |
2915 | * than one CPU, this means that when "func()" is invoked, each CPU is | |
2916 | * guaranteed to have executed a full memory barrier since the end of its | |
2917 | * last RCU read-side critical section whose beginning preceded the call | |
2918 | * to call_rcu(). It also means that each CPU executing an RCU read-side | |
2919 | * critical section that continues beyond the start of "func()" must have | |
2920 | * executed a memory barrier after the call_rcu() but before the beginning | |
2921 | * of that RCU read-side critical section. Note that these guarantees | |
2922 | * include CPUs that are offline, idle, or executing in user mode, as | |
2923 | * well as CPUs that are executing in the kernel. | |
2924 | * | |
2925 | * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the | |
2926 | * resulting RCU callback function "func()", then both CPU A and CPU B are | |
2927 | * guaranteed to execute a full memory barrier during the time interval | |
2928 | * between the call to call_rcu() and the invocation of "func()" -- even | |
2929 | * if CPU A and CPU B are the same CPU (but again only if the system has | |
2930 | * more than one CPU). | |
2931 | */ | |
2932 | void call_rcu(struct rcu_head *head, rcu_callback_t func) | |
2933 | { | |
77a40f97 | 2934 | __call_rcu(head, func); |
45975c7d PM |
2935 | } |
2936 | EXPORT_SYMBOL_GPL(call_rcu); | |
64db4cff | 2937 | |
a35d1690 BP |
2938 | |
2939 | /* Maximum number of jiffies to wait before draining a batch. */ | |
2940 | #define KFREE_DRAIN_JIFFIES (HZ / 50) | |
0392bebe | 2941 | #define KFREE_N_BATCHES 2 |
a35d1690 | 2942 | |
34c88174 URS |
2943 | /* |
2944 | * This macro defines how many entries the "records" array | |
2945 | * will contain. It is based on the fact that the size of | |
2946 | * kfree_rcu_bulk_data structure becomes exactly one page. | |
2947 | */ | |
2948 | #define KFREE_BULK_MAX_ENTR ((PAGE_SIZE / sizeof(void *)) - 3) | |
2949 | ||
2950 | /** | |
2951 | * struct kfree_rcu_bulk_data - single block to store kfree_rcu() pointers | |
2952 | * @nr_records: Number of active pointers in the array | |
2953 | * @records: Array of the kfree_rcu() pointers | |
2954 | * @next: Next bulk object in the block chain | |
2955 | * @head_free_debug: For debug, when CONFIG_DEBUG_OBJECTS_RCU_HEAD is set | |
2956 | */ | |
2957 | struct kfree_rcu_bulk_data { | |
2958 | unsigned long nr_records; | |
2959 | void *records[KFREE_BULK_MAX_ENTR]; | |
2960 | struct kfree_rcu_bulk_data *next; | |
2961 | struct rcu_head *head_free_debug; | |
2962 | }; | |
2963 | ||
a35d1690 | 2964 | /** |
0392bebe | 2965 | * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests |
a35d1690 | 2966 | * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period |
0392bebe | 2967 | * @head_free: List of kfree_rcu() objects waiting for a grace period |
34c88174 | 2968 | * @bhead_free: Bulk-List of kfree_rcu() objects waiting for a grace period |
0392bebe JFG |
2969 | * @krcp: Pointer to @kfree_rcu_cpu structure |
2970 | */ | |
2971 | ||
2972 | struct kfree_rcu_cpu_work { | |
2973 | struct rcu_work rcu_work; | |
2974 | struct rcu_head *head_free; | |
34c88174 | 2975 | struct kfree_rcu_bulk_data *bhead_free; |
0392bebe JFG |
2976 | struct kfree_rcu_cpu *krcp; |
2977 | }; | |
2978 | ||
2979 | /** | |
2980 | * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period | |
a35d1690 | 2981 | * @head: List of kfree_rcu() objects not yet waiting for a grace period |
34c88174 URS |
2982 | * @bhead: Bulk-List of kfree_rcu() objects not yet waiting for a grace period |
2983 | * @bcached: Keeps at most one object for later reuse when build chain blocks | |
0392bebe | 2984 | * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period |
a35d1690 BP |
2985 | * @lock: Synchronize access to this structure |
2986 | * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES | |
2987 | * @monitor_todo: Tracks whether a @monitor_work delayed work is pending | |
2988 | * @initialized: The @lock and @rcu_work fields have been initialized | |
2989 | * | |
2990 | * This is a per-CPU structure. The reason that it is not included in | |
2991 | * the rcu_data structure is to permit this code to be extracted from | |
2992 | * the RCU files. Such extraction could allow further optimization of | |
2993 | * the interactions with the slab allocators. | |
2994 | */ | |
2995 | struct kfree_rcu_cpu { | |
a35d1690 | 2996 | struct rcu_head *head; |
34c88174 URS |
2997 | struct kfree_rcu_bulk_data *bhead; |
2998 | struct kfree_rcu_bulk_data *bcached; | |
0392bebe | 2999 | struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES]; |
a35d1690 BP |
3000 | spinlock_t lock; |
3001 | struct delayed_work monitor_work; | |
569d7670 | 3002 | bool monitor_todo; |
a35d1690 | 3003 | bool initialized; |
9154244c JFG |
3004 | // Number of objects for which GP not started |
3005 | int count; | |
a35d1690 BP |
3006 | }; |
3007 | ||
3008 | static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc); | |
3009 | ||
34c88174 URS |
3010 | static __always_inline void |
3011 | debug_rcu_head_unqueue_bulk(struct rcu_head *head) | |
3012 | { | |
3013 | #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD | |
3014 | for (; head; head = head->next) | |
3015 | debug_rcu_head_unqueue(head); | |
3016 | #endif | |
3017 | } | |
3018 | ||
495aa969 | 3019 | /* |
a35d1690 | 3020 | * This function is invoked in workqueue context after a grace period. |
34c88174 | 3021 | * It frees all the objects queued on ->bhead_free or ->head_free. |
495aa969 | 3022 | */ |
a35d1690 BP |
3023 | static void kfree_rcu_work(struct work_struct *work) |
3024 | { | |
3025 | unsigned long flags; | |
3026 | struct rcu_head *head, *next; | |
34c88174 | 3027 | struct kfree_rcu_bulk_data *bhead, *bnext; |
a35d1690 | 3028 | struct kfree_rcu_cpu *krcp; |
0392bebe | 3029 | struct kfree_rcu_cpu_work *krwp; |
a35d1690 | 3030 | |
0392bebe JFG |
3031 | krwp = container_of(to_rcu_work(work), |
3032 | struct kfree_rcu_cpu_work, rcu_work); | |
3033 | krcp = krwp->krcp; | |
a35d1690 | 3034 | spin_lock_irqsave(&krcp->lock, flags); |
0392bebe JFG |
3035 | head = krwp->head_free; |
3036 | krwp->head_free = NULL; | |
34c88174 URS |
3037 | bhead = krwp->bhead_free; |
3038 | krwp->bhead_free = NULL; | |
a35d1690 BP |
3039 | spin_unlock_irqrestore(&krcp->lock, flags); |
3040 | ||
34c88174 URS |
3041 | /* "bhead" is now private, so traverse locklessly. */ |
3042 | for (; bhead; bhead = bnext) { | |
3043 | bnext = bhead->next; | |
3044 | ||
3045 | debug_rcu_head_unqueue_bulk(bhead->head_free_debug); | |
3046 | ||
3047 | rcu_lock_acquire(&rcu_callback_map); | |
61370792 URS |
3048 | trace_rcu_invoke_kfree_bulk_callback(rcu_state.name, |
3049 | bhead->nr_records, bhead->records); | |
3050 | ||
34c88174 URS |
3051 | kfree_bulk(bhead->nr_records, bhead->records); |
3052 | rcu_lock_release(&rcu_callback_map); | |
3053 | ||
3054 | if (cmpxchg(&krcp->bcached, NULL, bhead)) | |
3055 | free_page((unsigned long) bhead); | |
3056 | ||
3057 | cond_resched_tasks_rcu_qs(); | |
3058 | } | |
3059 | ||
3060 | /* | |
3061 | * Emergency case only. It can happen under low memory | |
3062 | * condition when an allocation gets failed, so the "bulk" | |
3063 | * path can not be temporary maintained. | |
3064 | */ | |
a35d1690 | 3065 | for (; head; head = next) { |
77a40f97 JFG |
3066 | unsigned long offset = (unsigned long)head->func; |
3067 | ||
a35d1690 | 3068 | next = head->next; |
e99637be | 3069 | debug_rcu_head_unqueue(head); |
77a40f97 JFG |
3070 | rcu_lock_acquire(&rcu_callback_map); |
3071 | trace_rcu_invoke_kfree_callback(rcu_state.name, head, offset); | |
3072 | ||
34c88174 | 3073 | if (!WARN_ON_ONCE(!__is_kfree_rcu_offset(offset))) |
189a6883 | 3074 | kfree((void *)head - offset); |
77a40f97 JFG |
3075 | |
3076 | rcu_lock_release(&rcu_callback_map); | |
a35d1690 BP |
3077 | cond_resched_tasks_rcu_qs(); |
3078 | } | |
3079 | } | |
3080 | ||
495aa969 | 3081 | /* |
a35d1690 BP |
3082 | * Schedule the kfree batch RCU work to run in workqueue context after a GP. |
3083 | * | |
3084 | * This function is invoked by kfree_rcu_monitor() when the KFREE_DRAIN_JIFFIES | |
3085 | * timeout has been reached. | |
3086 | */ | |
3087 | static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp) | |
3088 | { | |
34c88174 URS |
3089 | struct kfree_rcu_cpu_work *krwp; |
3090 | bool queued = false; | |
0392bebe | 3091 | int i; |
0392bebe | 3092 | |
a35d1690 BP |
3093 | lockdep_assert_held(&krcp->lock); |
3094 | ||
34c88174 URS |
3095 | for (i = 0; i < KFREE_N_BATCHES; i++) { |
3096 | krwp = &(krcp->krw_arr[i]); | |
a35d1690 | 3097 | |
34c88174 URS |
3098 | /* |
3099 | * Try to detach bhead or head and attach it over any | |
3100 | * available corresponding free channel. It can be that | |
3101 | * a previous RCU batch is in progress, it means that | |
3102 | * immediately to queue another one is not possible so | |
3103 | * return false to tell caller to retry. | |
3104 | */ | |
3105 | if ((krcp->bhead && !krwp->bhead_free) || | |
3106 | (krcp->head && !krwp->head_free)) { | |
3107 | /* Channel 1. */ | |
3108 | if (!krwp->bhead_free) { | |
3109 | krwp->bhead_free = krcp->bhead; | |
3110 | krcp->bhead = NULL; | |
3111 | } | |
3112 | ||
3113 | /* Channel 2. */ | |
3114 | if (!krwp->head_free) { | |
3115 | krwp->head_free = krcp->head; | |
3116 | krcp->head = NULL; | |
3117 | } | |
3118 | ||
a6a82ce1 | 3119 | WRITE_ONCE(krcp->count, 0); |
9154244c | 3120 | |
34c88174 URS |
3121 | /* |
3122 | * One work is per one batch, so there are two "free channels", | |
3123 | * "bhead_free" and "head_free" the batch can handle. It can be | |
3124 | * that the work is in the pending state when two channels have | |
3125 | * been detached following each other, one by one. | |
3126 | */ | |
3127 | queue_rcu_work(system_wq, &krwp->rcu_work); | |
3128 | queued = true; | |
3129 | } | |
3130 | } | |
3131 | ||
3132 | return queued; | |
a35d1690 BP |
3133 | } |
3134 | ||
3135 | static inline void kfree_rcu_drain_unlock(struct kfree_rcu_cpu *krcp, | |
3136 | unsigned long flags) | |
3137 | { | |
3138 | // Attempt to start a new batch. | |
569d7670 | 3139 | krcp->monitor_todo = false; |
a35d1690 BP |
3140 | if (queue_kfree_rcu_work(krcp)) { |
3141 | // Success! Our job is done here. | |
3142 | spin_unlock_irqrestore(&krcp->lock, flags); | |
3143 | return; | |
3144 | } | |
3145 | ||
3146 | // Previous RCU batch still in progress, try again later. | |
569d7670 JF |
3147 | krcp->monitor_todo = true; |
3148 | schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES); | |
a35d1690 BP |
3149 | spin_unlock_irqrestore(&krcp->lock, flags); |
3150 | } | |
3151 | ||
495aa969 | 3152 | /* |
a35d1690 BP |
3153 | * This function is invoked after the KFREE_DRAIN_JIFFIES timeout. |
3154 | * It invokes kfree_rcu_drain_unlock() to attempt to start another batch. | |
3155 | */ | |
3156 | static void kfree_rcu_monitor(struct work_struct *work) | |
3157 | { | |
3158 | unsigned long flags; | |
3159 | struct kfree_rcu_cpu *krcp = container_of(work, struct kfree_rcu_cpu, | |
3160 | monitor_work.work); | |
3161 | ||
3162 | spin_lock_irqsave(&krcp->lock, flags); | |
569d7670 | 3163 | if (krcp->monitor_todo) |
a35d1690 BP |
3164 | kfree_rcu_drain_unlock(krcp, flags); |
3165 | else | |
3166 | spin_unlock_irqrestore(&krcp->lock, flags); | |
3167 | } | |
3168 | ||
34c88174 URS |
3169 | static inline bool |
3170 | kfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, | |
3171 | struct rcu_head *head, rcu_callback_t func) | |
3172 | { | |
3173 | struct kfree_rcu_bulk_data *bnode; | |
3174 | ||
3175 | if (unlikely(!krcp->initialized)) | |
3176 | return false; | |
3177 | ||
3178 | lockdep_assert_held(&krcp->lock); | |
3179 | ||
3180 | /* Check if a new block is required. */ | |
3181 | if (!krcp->bhead || | |
3182 | krcp->bhead->nr_records == KFREE_BULK_MAX_ENTR) { | |
3183 | bnode = xchg(&krcp->bcached, NULL); | |
3184 | if (!bnode) { | |
3185 | WARN_ON_ONCE(sizeof(struct kfree_rcu_bulk_data) > PAGE_SIZE); | |
3186 | ||
3187 | bnode = (struct kfree_rcu_bulk_data *) | |
3188 | __get_free_page(GFP_NOWAIT | __GFP_NOWARN); | |
3189 | } | |
3190 | ||
3191 | /* Switch to emergency path. */ | |
3192 | if (unlikely(!bnode)) | |
3193 | return false; | |
3194 | ||
3195 | /* Initialize the new block. */ | |
3196 | bnode->nr_records = 0; | |
3197 | bnode->next = krcp->bhead; | |
3198 | bnode->head_free_debug = NULL; | |
3199 | ||
3200 | /* Attach it to the head. */ | |
3201 | krcp->bhead = bnode; | |
3202 | } | |
3203 | ||
3204 | #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD | |
3205 | head->func = func; | |
3206 | head->next = krcp->bhead->head_free_debug; | |
3207 | krcp->bhead->head_free_debug = head; | |
3208 | #endif | |
3209 | ||
3210 | /* Finally insert. */ | |
3211 | krcp->bhead->records[krcp->bhead->nr_records++] = | |
3212 | (void *) head - (unsigned long) func; | |
3213 | ||
3214 | return true; | |
3215 | } | |
3216 | ||
a35d1690 | 3217 | /* |
34c88174 URS |
3218 | * Queue a request for lazy invocation of kfree_bulk()/kfree() after a grace |
3219 | * period. Please note there are two paths are maintained, one is the main one | |
3220 | * that uses kfree_bulk() interface and second one is emergency one, that is | |
3221 | * used only when the main path can not be maintained temporary, due to memory | |
3222 | * pressure. | |
a35d1690 BP |
3223 | * |
3224 | * Each kfree_call_rcu() request is added to a batch. The batch will be drained | |
34c88174 URS |
3225 | * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will |
3226 | * be free'd in workqueue context. This allows us to: batch requests together to | |
3227 | * reduce the number of grace periods during heavy kfree_rcu() load. | |
495aa969 | 3228 | */ |
98ece508 | 3229 | void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func) |
495aa969 | 3230 | { |
a35d1690 BP |
3231 | unsigned long flags; |
3232 | struct kfree_rcu_cpu *krcp; | |
3233 | ||
a35d1690 BP |
3234 | local_irq_save(flags); // For safely calling this_cpu_ptr(). |
3235 | krcp = this_cpu_ptr(&krc); | |
3236 | if (krcp->initialized) | |
3237 | spin_lock(&krcp->lock); | |
3238 | ||
3239 | // Queue the object but don't yet schedule the batch. | |
e99637be JFG |
3240 | if (debug_rcu_head_queue(head)) { |
3241 | // Probable double kfree_rcu(), just leak. | |
3242 | WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n", | |
3243 | __func__, head); | |
3244 | goto unlock_return; | |
3245 | } | |
34c88174 URS |
3246 | |
3247 | /* | |
3248 | * Under high memory pressure GFP_NOWAIT can fail, | |
3249 | * in that case the emergency path is maintained. | |
3250 | */ | |
3251 | if (unlikely(!kfree_call_rcu_add_ptr_to_bulk(krcp, head, func))) { | |
3252 | head->func = func; | |
3253 | head->next = krcp->head; | |
3254 | krcp->head = head; | |
3255 | } | |
a35d1690 | 3256 | |
a6a82ce1 | 3257 | WRITE_ONCE(krcp->count, krcp->count + 1); |
9154244c | 3258 | |
a35d1690 BP |
3259 | // Set timer to drain after KFREE_DRAIN_JIFFIES. |
3260 | if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING && | |
569d7670 JF |
3261 | !krcp->monitor_todo) { |
3262 | krcp->monitor_todo = true; | |
a35d1690 | 3263 | schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES); |
569d7670 | 3264 | } |
a35d1690 | 3265 | |
e99637be | 3266 | unlock_return: |
a35d1690 BP |
3267 | if (krcp->initialized) |
3268 | spin_unlock(&krcp->lock); | |
3269 | local_irq_restore(flags); | |
495aa969 ACB |
3270 | } |
3271 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
3272 | ||
9154244c JFG |
3273 | static unsigned long |
3274 | kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc) | |
3275 | { | |
3276 | int cpu; | |
a6a82ce1 | 3277 | unsigned long count = 0; |
9154244c JFG |
3278 | |
3279 | /* Snapshot count of all CPUs */ | |
3280 | for_each_online_cpu(cpu) { | |
3281 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); | |
3282 | ||
a6a82ce1 | 3283 | count += READ_ONCE(krcp->count); |
9154244c JFG |
3284 | } |
3285 | ||
3286 | return count; | |
3287 | } | |
3288 | ||
3289 | static unsigned long | |
3290 | kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc) | |
3291 | { | |
3292 | int cpu, freed = 0; | |
3293 | unsigned long flags; | |
3294 | ||
3295 | for_each_online_cpu(cpu) { | |
3296 | int count; | |
3297 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); | |
3298 | ||
3299 | count = krcp->count; | |
3300 | spin_lock_irqsave(&krcp->lock, flags); | |
3301 | if (krcp->monitor_todo) | |
3302 | kfree_rcu_drain_unlock(krcp, flags); | |
3303 | else | |
3304 | spin_unlock_irqrestore(&krcp->lock, flags); | |
3305 | ||
3306 | sc->nr_to_scan -= count; | |
3307 | freed += count; | |
3308 | ||
3309 | if (sc->nr_to_scan <= 0) | |
3310 | break; | |
3311 | } | |
3312 | ||
3313 | return freed; | |
3314 | } | |
3315 | ||
3316 | static struct shrinker kfree_rcu_shrinker = { | |
3317 | .count_objects = kfree_rcu_shrink_count, | |
3318 | .scan_objects = kfree_rcu_shrink_scan, | |
3319 | .batch = 0, | |
3320 | .seeks = DEFAULT_SEEKS, | |
3321 | }; | |
3322 | ||
a35d1690 BP |
3323 | void __init kfree_rcu_scheduler_running(void) |
3324 | { | |
3325 | int cpu; | |
3326 | unsigned long flags; | |
3327 | ||
3328 | for_each_online_cpu(cpu) { | |
3329 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); | |
3330 | ||
3331 | spin_lock_irqsave(&krcp->lock, flags); | |
569d7670 | 3332 | if (!krcp->head || krcp->monitor_todo) { |
a35d1690 BP |
3333 | spin_unlock_irqrestore(&krcp->lock, flags); |
3334 | continue; | |
3335 | } | |
569d7670 | 3336 | krcp->monitor_todo = true; |
0392bebe JFG |
3337 | schedule_delayed_work_on(cpu, &krcp->monitor_work, |
3338 | KFREE_DRAIN_JIFFIES); | |
a35d1690 BP |
3339 | spin_unlock_irqrestore(&krcp->lock, flags); |
3340 | } | |
3341 | } | |
3342 | ||
e5bc3af7 PM |
3343 | /* |
3344 | * During early boot, any blocking grace-period wait automatically | |
90326f05 | 3345 | * implies a grace period. Later on, this is never the case for PREEMPTION. |
e5bc3af7 | 3346 | * |
90326f05 | 3347 | * Howevr, because a context switch is a grace period for !PREEMPTION, any |
e5bc3af7 PM |
3348 | * blocking grace-period wait automatically implies a grace period if |
3349 | * there is only one CPU online at any point time during execution of | |
3350 | * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to | |
3351 | * occasionally incorrectly indicate that there are multiple CPUs online | |
3352 | * when there was in fact only one the whole time, as this just adds some | |
3353 | * overhead: RCU still operates correctly. | |
3354 | */ | |
3355 | static int rcu_blocking_is_gp(void) | |
3356 | { | |
3357 | int ret; | |
3358 | ||
01b1d88b | 3359 | if (IS_ENABLED(CONFIG_PREEMPTION)) |
e5bc3af7 PM |
3360 | return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE; |
3361 | might_sleep(); /* Check for RCU read-side critical section. */ | |
3362 | preempt_disable(); | |
3363 | ret = num_online_cpus() <= 1; | |
3364 | preempt_enable(); | |
3365 | return ret; | |
3366 | } | |
3367 | ||
3368 | /** | |
3369 | * synchronize_rcu - wait until a grace period has elapsed. | |
3370 | * | |
3371 | * Control will return to the caller some time after a full grace | |
3372 | * period has elapsed, in other words after all currently executing RCU | |
3373 | * read-side critical sections have completed. Note, however, that | |
3374 | * upon return from synchronize_rcu(), the caller might well be executing | |
3375 | * concurrently with new RCU read-side critical sections that began while | |
3376 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
3377 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
3378 | * In addition, regions of code across which interrupts, preemption, or | |
3379 | * softirqs have been disabled also serve as RCU read-side critical | |
3380 | * sections. This includes hardware interrupt handlers, softirq handlers, | |
3381 | * and NMI handlers. | |
3382 | * | |
3383 | * Note that this guarantee implies further memory-ordering guarantees. | |
3384 | * On systems with more than one CPU, when synchronize_rcu() returns, | |
3385 | * each CPU is guaranteed to have executed a full memory barrier since | |
3386 | * the end of its last RCU read-side critical section whose beginning | |
3387 | * preceded the call to synchronize_rcu(). In addition, each CPU having | |
3388 | * an RCU read-side critical section that extends beyond the return from | |
3389 | * synchronize_rcu() is guaranteed to have executed a full memory barrier | |
3390 | * after the beginning of synchronize_rcu() and before the beginning of | |
3391 | * that RCU read-side critical section. Note that these guarantees include | |
3392 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
3393 | * that are executing in the kernel. | |
3394 | * | |
3395 | * Furthermore, if CPU A invoked synchronize_rcu(), which returned | |
3396 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
3397 | * to have executed a full memory barrier during the execution of | |
3398 | * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but | |
3399 | * again only if the system has more than one CPU). | |
3400 | */ | |
3401 | void synchronize_rcu(void) | |
3402 | { | |
3403 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || | |
3404 | lock_is_held(&rcu_lock_map) || | |
3405 | lock_is_held(&rcu_sched_lock_map), | |
3406 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
3407 | if (rcu_blocking_is_gp()) | |
3408 | return; | |
3409 | if (rcu_gp_is_expedited()) | |
3410 | synchronize_rcu_expedited(); | |
3411 | else | |
3412 | wait_rcu_gp(call_rcu); | |
3413 | } | |
3414 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
3415 | ||
765a3f4f PM |
3416 | /** |
3417 | * get_state_synchronize_rcu - Snapshot current RCU state | |
3418 | * | |
3419 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
3420 | * to determine whether or not a full grace period has elapsed in the | |
3421 | * meantime. | |
3422 | */ | |
3423 | unsigned long get_state_synchronize_rcu(void) | |
3424 | { | |
3425 | /* | |
3426 | * Any prior manipulation of RCU-protected data must happen | |
e4be81a2 | 3427 | * before the load from ->gp_seq. |
765a3f4f PM |
3428 | */ |
3429 | smp_mb(); /* ^^^ */ | |
16fc9c60 | 3430 | return rcu_seq_snap(&rcu_state.gp_seq); |
765a3f4f PM |
3431 | } |
3432 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
3433 | ||
3434 | /** | |
3435 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
3436 | * | |
3437 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
3438 | * | |
3439 | * If a full RCU grace period has elapsed since the earlier call to | |
3440 | * get_state_synchronize_rcu(), just return. Otherwise, invoke | |
3441 | * synchronize_rcu() to wait for a full grace period. | |
3442 | * | |
3443 | * Yes, this function does not take counter wrap into account. But | |
3444 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3445 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3446 | * so waiting for one additional grace period should be just fine. | |
3447 | */ | |
3448 | void cond_synchronize_rcu(unsigned long oldstate) | |
3449 | { | |
16fc9c60 | 3450 | if (!rcu_seq_done(&rcu_state.gp_seq, oldstate)) |
765a3f4f | 3451 | synchronize_rcu(); |
e4be81a2 PM |
3452 | else |
3453 | smp_mb(); /* Ensure GP ends before subsequent accesses. */ | |
765a3f4f PM |
3454 | } |
3455 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
3456 | ||
64db4cff | 3457 | /* |
98ece508 | 3458 | * Check to see if there is any immediate RCU-related work to be done by |
49918a54 PM |
3459 | * the current CPU, returning 1 if so and zero otherwise. The checks are |
3460 | * in order of increasing expense: checks that can be carried out against | |
3461 | * CPU-local state are performed first. However, we must check for CPU | |
3462 | * stalls first, else we might not get a chance. | |
64db4cff | 3463 | */ |
dd7dafd1 | 3464 | static int rcu_pending(int user) |
64db4cff | 3465 | { |
ed93dfc6 | 3466 | bool gp_in_progress; |
98ece508 | 3467 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
2f51f988 PM |
3468 | struct rcu_node *rnp = rdp->mynode; |
3469 | ||
64db4cff | 3470 | /* Check for CPU stalls, if enabled. */ |
ea12ff2b | 3471 | check_cpu_stall(rdp); |
64db4cff | 3472 | |
85f69b32 PM |
3473 | /* Does this CPU need a deferred NOCB wakeup? */ |
3474 | if (rcu_nocb_need_deferred_wakeup(rdp)) | |
3475 | return 1; | |
3476 | ||
dd7dafd1 PM |
3477 | /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */ |
3478 | if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu()) | |
a096932f PM |
3479 | return 0; |
3480 | ||
64db4cff | 3481 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
ed93dfc6 PM |
3482 | gp_in_progress = rcu_gp_in_progress(); |
3483 | if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress) | |
64db4cff PM |
3484 | return 1; |
3485 | ||
3486 | /* Does this CPU have callbacks ready to invoke? */ | |
01c495f7 | 3487 | if (rcu_segcblist_ready_cbs(&rdp->cblist)) |
64db4cff PM |
3488 | return 1; |
3489 | ||
3490 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
ed93dfc6 | 3491 | if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) && |
921bb5fa PM |
3492 | (!IS_ENABLED(CONFIG_RCU_NOCB_CPU) || |
3493 | !rcu_segcblist_is_offloaded(&rdp->cblist)) && | |
c1935209 | 3494 | !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL)) |
64db4cff PM |
3495 | return 1; |
3496 | ||
67e14c1e PM |
3497 | /* Have RCU grace period completed or started? */ |
3498 | if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq || | |
01c495f7 | 3499 | unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */ |
64db4cff PM |
3500 | return 1; |
3501 | ||
64db4cff PM |
3502 | /* nothing to do */ |
3503 | return 0; | |
3504 | } | |
3505 | ||
a83eff0a | 3506 | /* |
dd46a788 | 3507 | * Helper function for rcu_barrier() tracing. If tracing is disabled, |
a83eff0a PM |
3508 | * the compiler is expected to optimize this away. |
3509 | */ | |
dd46a788 | 3510 | static void rcu_barrier_trace(const char *s, int cpu, unsigned long done) |
a83eff0a | 3511 | { |
8344b871 PM |
3512 | trace_rcu_barrier(rcu_state.name, s, cpu, |
3513 | atomic_read(&rcu_state.barrier_cpu_count), done); | |
a83eff0a PM |
3514 | } |
3515 | ||
b1420f1c | 3516 | /* |
dd46a788 PM |
3517 | * RCU callback function for rcu_barrier(). If we are last, wake |
3518 | * up the task executing rcu_barrier(). | |
aa24f937 PM |
3519 | * |
3520 | * Note that the value of rcu_state.barrier_sequence must be captured | |
3521 | * before the atomic_dec_and_test(). Otherwise, if this CPU is not last, | |
3522 | * other CPUs might count the value down to zero before this CPU gets | |
3523 | * around to invoking rcu_barrier_trace(), which might result in bogus | |
3524 | * data from the next instance of rcu_barrier(). | |
b1420f1c | 3525 | */ |
24ebbca8 | 3526 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3527 | { |
aa24f937 PM |
3528 | unsigned long __maybe_unused s = rcu_state.barrier_sequence; |
3529 | ||
ec9f5835 | 3530 | if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) { |
aa24f937 | 3531 | rcu_barrier_trace(TPS("LastCB"), -1, s); |
ec9f5835 | 3532 | complete(&rcu_state.barrier_completion); |
a83eff0a | 3533 | } else { |
aa24f937 | 3534 | rcu_barrier_trace(TPS("CB"), -1, s); |
a83eff0a | 3535 | } |
d0ec774c PM |
3536 | } |
3537 | ||
3538 | /* | |
3539 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3540 | */ | |
127e2981 | 3541 | static void rcu_barrier_func(void *cpu_in) |
d0ec774c | 3542 | { |
127e2981 PM |
3543 | uintptr_t cpu = (uintptr_t)cpu_in; |
3544 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
d0ec774c | 3545 | |
dd46a788 | 3546 | rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence); |
f92c734f PM |
3547 | rdp->barrier_head.func = rcu_barrier_callback; |
3548 | debug_rcu_head_queue(&rdp->barrier_head); | |
5d6742b3 | 3549 | rcu_nocb_lock(rdp); |
d1b222c6 | 3550 | WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies)); |
77a40f97 | 3551 | if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) { |
ec9f5835 | 3552 | atomic_inc(&rcu_state.barrier_cpu_count); |
f92c734f PM |
3553 | } else { |
3554 | debug_rcu_head_unqueue(&rdp->barrier_head); | |
dd46a788 | 3555 | rcu_barrier_trace(TPS("IRQNQ"), -1, |
66e4c33b | 3556 | rcu_state.barrier_sequence); |
f92c734f | 3557 | } |
5d6742b3 | 3558 | rcu_nocb_unlock(rdp); |
d0ec774c PM |
3559 | } |
3560 | ||
dd46a788 PM |
3561 | /** |
3562 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
3563 | * | |
3564 | * Note that this primitive does not necessarily wait for an RCU grace period | |
3565 | * to complete. For example, if there are no RCU callbacks queued anywhere | |
3566 | * in the system, then rcu_barrier() is within its rights to return | |
3567 | * immediately, without waiting for anything, much less an RCU grace period. | |
d0ec774c | 3568 | */ |
dd46a788 | 3569 | void rcu_barrier(void) |
d0ec774c | 3570 | { |
127e2981 | 3571 | uintptr_t cpu; |
b1420f1c | 3572 | struct rcu_data *rdp; |
ec9f5835 | 3573 | unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence); |
b1420f1c | 3574 | |
dd46a788 | 3575 | rcu_barrier_trace(TPS("Begin"), -1, s); |
b1420f1c | 3576 | |
e74f4c45 | 3577 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
ec9f5835 | 3578 | mutex_lock(&rcu_state.barrier_mutex); |
b1420f1c | 3579 | |
4f525a52 | 3580 | /* Did someone else do our work for us? */ |
ec9f5835 | 3581 | if (rcu_seq_done(&rcu_state.barrier_sequence, s)) { |
dd46a788 | 3582 | rcu_barrier_trace(TPS("EarlyExit"), -1, |
66e4c33b | 3583 | rcu_state.barrier_sequence); |
cf3a9c48 | 3584 | smp_mb(); /* caller's subsequent code after above check. */ |
ec9f5835 | 3585 | mutex_unlock(&rcu_state.barrier_mutex); |
cf3a9c48 PM |
3586 | return; |
3587 | } | |
3588 | ||
4f525a52 | 3589 | /* Mark the start of the barrier operation. */ |
ec9f5835 | 3590 | rcu_seq_start(&rcu_state.barrier_sequence); |
dd46a788 | 3591 | rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence); |
b1420f1c | 3592 | |
d0ec774c | 3593 | /* |
127e2981 PM |
3594 | * Initialize the count to two rather than to zero in order |
3595 | * to avoid a too-soon return to zero in case of an immediate | |
3596 | * invocation of the just-enqueued callback (or preemption of | |
3597 | * this task). Exclude CPU-hotplug operations to ensure that no | |
3598 | * offline non-offloaded CPU has callbacks queued. | |
d0ec774c | 3599 | */ |
ec9f5835 | 3600 | init_completion(&rcu_state.barrier_completion); |
127e2981 | 3601 | atomic_set(&rcu_state.barrier_cpu_count, 2); |
1331e7a1 | 3602 | get_online_cpus(); |
b1420f1c PM |
3603 | |
3604 | /* | |
1331e7a1 PM |
3605 | * Force each CPU with callbacks to register a new callback. |
3606 | * When that callback is invoked, we will know that all of the | |
3607 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 3608 | */ |
3fbfbf7a | 3609 | for_each_possible_cpu(cpu) { |
da1df50d | 3610 | rdp = per_cpu_ptr(&rcu_data, cpu); |
127e2981 | 3611 | if (cpu_is_offline(cpu) && |
ce5215c1 PM |
3612 | !rcu_segcblist_is_offloaded(&rdp->cblist)) |
3613 | continue; | |
127e2981 | 3614 | if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) { |
dd46a788 | 3615 | rcu_barrier_trace(TPS("OnlineQ"), cpu, |
66e4c33b | 3616 | rcu_state.barrier_sequence); |
127e2981 PM |
3617 | smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1); |
3618 | } else if (rcu_segcblist_n_cbs(&rdp->cblist) && | |
3619 | cpu_is_offline(cpu)) { | |
3620 | rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, | |
3621 | rcu_state.barrier_sequence); | |
3622 | local_irq_disable(); | |
3623 | rcu_barrier_func((void *)cpu); | |
3624 | local_irq_enable(); | |
3625 | } else if (cpu_is_offline(cpu)) { | |
3626 | rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu, | |
3627 | rcu_state.barrier_sequence); | |
b1420f1c | 3628 | } else { |
dd46a788 | 3629 | rcu_barrier_trace(TPS("OnlineNQ"), cpu, |
66e4c33b | 3630 | rcu_state.barrier_sequence); |
b1420f1c PM |
3631 | } |
3632 | } | |
1331e7a1 | 3633 | put_online_cpus(); |
b1420f1c PM |
3634 | |
3635 | /* | |
3636 | * Now that we have an rcu_barrier_callback() callback on each | |
3637 | * CPU, and thus each counted, remove the initial count. | |
3638 | */ | |
127e2981 | 3639 | if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count)) |
ec9f5835 | 3640 | complete(&rcu_state.barrier_completion); |
b1420f1c PM |
3641 | |
3642 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ | |
ec9f5835 | 3643 | wait_for_completion(&rcu_state.barrier_completion); |
b1420f1c | 3644 | |
4f525a52 | 3645 | /* Mark the end of the barrier operation. */ |
dd46a788 | 3646 | rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence); |
ec9f5835 | 3647 | rcu_seq_end(&rcu_state.barrier_sequence); |
4f525a52 | 3648 | |
b1420f1c | 3649 | /* Other rcu_barrier() invocations can now safely proceed. */ |
ec9f5835 | 3650 | mutex_unlock(&rcu_state.barrier_mutex); |
d0ec774c | 3651 | } |
45975c7d | 3652 | EXPORT_SYMBOL_GPL(rcu_barrier); |
d0ec774c | 3653 | |
0aa04b05 PM |
3654 | /* |
3655 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the | |
3656 | * first CPU in a given leaf rcu_node structure coming online. The caller | |
3657 | * must hold the corresponding leaf rcu_node ->lock with interrrupts | |
3658 | * disabled. | |
3659 | */ | |
3660 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) | |
3661 | { | |
3662 | long mask; | |
8d672fa6 | 3663 | long oldmask; |
0aa04b05 PM |
3664 | struct rcu_node *rnp = rnp_leaf; |
3665 | ||
8d672fa6 | 3666 | raw_lockdep_assert_held_rcu_node(rnp_leaf); |
962aff03 | 3667 | WARN_ON_ONCE(rnp->wait_blkd_tasks); |
0aa04b05 PM |
3668 | for (;;) { |
3669 | mask = rnp->grpmask; | |
3670 | rnp = rnp->parent; | |
3671 | if (rnp == NULL) | |
3672 | return; | |
6cf10081 | 3673 | raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */ |
8d672fa6 | 3674 | oldmask = rnp->qsmaskinit; |
0aa04b05 | 3675 | rnp->qsmaskinit |= mask; |
67c583a7 | 3676 | raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */ |
8d672fa6 PM |
3677 | if (oldmask) |
3678 | return; | |
0aa04b05 PM |
3679 | } |
3680 | } | |
3681 | ||
64db4cff | 3682 | /* |
27569620 | 3683 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 3684 | */ |
27569620 | 3685 | static void __init |
53b46303 | 3686 | rcu_boot_init_percpu_data(int cpu) |
64db4cff | 3687 | { |
da1df50d | 3688 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
27569620 PM |
3689 | |
3690 | /* Set up local state, ensuring consistent view of global state. */ | |
bc75e999 | 3691 | rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu); |
4c5273bf | 3692 | WARN_ON_ONCE(rdp->dynticks_nesting != 1); |
dc5a4f29 | 3693 | WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp))); |
53b46303 | 3694 | rdp->rcu_ofl_gp_seq = rcu_state.gp_seq; |
57738942 | 3695 | rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED; |
53b46303 | 3696 | rdp->rcu_onl_gp_seq = rcu_state.gp_seq; |
57738942 | 3697 | rdp->rcu_onl_gp_flags = RCU_GP_CLEANED; |
27569620 | 3698 | rdp->cpu = cpu; |
3fbfbf7a | 3699 | rcu_boot_init_nocb_percpu_data(rdp); |
27569620 PM |
3700 | } |
3701 | ||
3702 | /* | |
53b46303 PM |
3703 | * Invoked early in the CPU-online process, when pretty much all services |
3704 | * are available. The incoming CPU is not present. | |
3705 | * | |
3706 | * Initializes a CPU's per-CPU RCU data. Note that only one online or | |
ff3bb6f4 PM |
3707 | * offline event can be happening at a given time. Note also that we can |
3708 | * accept some slop in the rsp->gp_seq access due to the fact that this | |
e83e73f5 PM |
3709 | * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet. |
3710 | * And any offloaded callbacks are being numbered elsewhere. | |
64db4cff | 3711 | */ |
53b46303 | 3712 | int rcutree_prepare_cpu(unsigned int cpu) |
64db4cff PM |
3713 | { |
3714 | unsigned long flags; | |
da1df50d | 3715 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
336a4f6c | 3716 | struct rcu_node *rnp = rcu_get_root(); |
64db4cff PM |
3717 | |
3718 | /* Set up local state, ensuring consistent view of global state. */ | |
6cf10081 | 3719 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
37c72e56 | 3720 | rdp->qlen_last_fqs_check = 0; |
53b46303 | 3721 | rdp->n_force_qs_snap = rcu_state.n_force_qs; |
64db4cff | 3722 | rdp->blimit = blimit; |
15fecf89 | 3723 | if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */ |
e83e73f5 | 3724 | !rcu_segcblist_is_offloaded(&rdp->cblist)) |
15fecf89 | 3725 | rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */ |
4c5273bf | 3726 | rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */ |
2625d469 | 3727 | rcu_dynticks_eqs_online(); |
67c583a7 | 3728 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
64db4cff | 3729 | |
0aa04b05 PM |
3730 | /* |
3731 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed | |
3732 | * propagation up the rcu_node tree will happen at the beginning | |
3733 | * of the next grace period. | |
3734 | */ | |
64db4cff | 3735 | rnp = rdp->mynode; |
2a67e741 | 3736 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
b9585e94 | 3737 | rdp->beenonline = true; /* We have now been online. */ |
8ff37290 PM |
3738 | rdp->gp_seq = READ_ONCE(rnp->gp_seq); |
3739 | rdp->gp_seq_needed = rdp->gp_seq; | |
5b74c458 | 3740 | rdp->cpu_no_qs.b.norm = true; |
97c668b8 | 3741 | rdp->core_needs_qs = false; |
9b9500da | 3742 | rdp->rcu_iw_pending = false; |
8ff37290 | 3743 | rdp->rcu_iw_gp_seq = rdp->gp_seq - 1; |
53b46303 | 3744 | trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl")); |
67c583a7 | 3745 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
4df83742 | 3746 | rcu_prepare_kthreads(cpu); |
ad368d15 | 3747 | rcu_spawn_cpu_nocb_kthread(cpu); |
4df83742 TG |
3748 | |
3749 | return 0; | |
3750 | } | |
3751 | ||
deb34f36 PM |
3752 | /* |
3753 | * Update RCU priority boot kthread affinity for CPU-hotplug changes. | |
3754 | */ | |
4df83742 TG |
3755 | static void rcutree_affinity_setting(unsigned int cpu, int outgoing) |
3756 | { | |
da1df50d | 3757 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
4df83742 TG |
3758 | |
3759 | rcu_boost_kthread_setaffinity(rdp->mynode, outgoing); | |
3760 | } | |
3761 | ||
deb34f36 PM |
3762 | /* |
3763 | * Near the end of the CPU-online process. Pretty much all services | |
3764 | * enabled, and the CPU is now very much alive. | |
3765 | */ | |
4df83742 TG |
3766 | int rcutree_online_cpu(unsigned int cpu) |
3767 | { | |
9b9500da PM |
3768 | unsigned long flags; |
3769 | struct rcu_data *rdp; | |
3770 | struct rcu_node *rnp; | |
9b9500da | 3771 | |
b97d23c5 PM |
3772 | rdp = per_cpu_ptr(&rcu_data, cpu); |
3773 | rnp = rdp->mynode; | |
3774 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
3775 | rnp->ffmask |= rdp->grpmask; | |
3776 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
9b9500da PM |
3777 | if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE) |
3778 | return 0; /* Too early in boot for scheduler work. */ | |
3779 | sync_sched_exp_online_cleanup(cpu); | |
3780 | rcutree_affinity_setting(cpu, -1); | |
96926686 PM |
3781 | |
3782 | // Stop-machine done, so allow nohz_full to disable tick. | |
3783 | tick_dep_clear(TICK_DEP_BIT_RCU); | |
4df83742 TG |
3784 | return 0; |
3785 | } | |
3786 | ||
deb34f36 PM |
3787 | /* |
3788 | * Near the beginning of the process. The CPU is still very much alive | |
3789 | * with pretty much all services enabled. | |
3790 | */ | |
4df83742 TG |
3791 | int rcutree_offline_cpu(unsigned int cpu) |
3792 | { | |
9b9500da PM |
3793 | unsigned long flags; |
3794 | struct rcu_data *rdp; | |
3795 | struct rcu_node *rnp; | |
9b9500da | 3796 | |
b97d23c5 PM |
3797 | rdp = per_cpu_ptr(&rcu_data, cpu); |
3798 | rnp = rdp->mynode; | |
3799 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
3800 | rnp->ffmask &= ~rdp->grpmask; | |
3801 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
9b9500da | 3802 | |
4df83742 | 3803 | rcutree_affinity_setting(cpu, cpu); |
96926686 PM |
3804 | |
3805 | // nohz_full CPUs need the tick for stop-machine to work quickly | |
3806 | tick_dep_set(TICK_DEP_BIT_RCU); | |
4df83742 TG |
3807 | return 0; |
3808 | } | |
3809 | ||
f64c6013 PZ |
3810 | static DEFINE_PER_CPU(int, rcu_cpu_started); |
3811 | ||
7ec99de3 PM |
3812 | /* |
3813 | * Mark the specified CPU as being online so that subsequent grace periods | |
3814 | * (both expedited and normal) will wait on it. Note that this means that | |
3815 | * incoming CPUs are not allowed to use RCU read-side critical sections | |
3816 | * until this function is called. Failing to observe this restriction | |
3817 | * will result in lockdep splats. | |
deb34f36 PM |
3818 | * |
3819 | * Note that this function is special in that it is invoked directly | |
3820 | * from the incoming CPU rather than from the cpuhp_step mechanism. | |
3821 | * This is because this function must be invoked at a precise location. | |
7ec99de3 PM |
3822 | */ |
3823 | void rcu_cpu_starting(unsigned int cpu) | |
3824 | { | |
3825 | unsigned long flags; | |
3826 | unsigned long mask; | |
313517fc PM |
3827 | int nbits; |
3828 | unsigned long oldmask; | |
7ec99de3 PM |
3829 | struct rcu_data *rdp; |
3830 | struct rcu_node *rnp; | |
7ec99de3 | 3831 | |
f64c6013 PZ |
3832 | if (per_cpu(rcu_cpu_started, cpu)) |
3833 | return; | |
3834 | ||
3835 | per_cpu(rcu_cpu_started, cpu) = 1; | |
3836 | ||
b97d23c5 PM |
3837 | rdp = per_cpu_ptr(&rcu_data, cpu); |
3838 | rnp = rdp->mynode; | |
3839 | mask = rdp->grpmask; | |
3840 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
105abf82 | 3841 | WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask); |
b97d23c5 PM |
3842 | oldmask = rnp->expmaskinitnext; |
3843 | rnp->expmaskinitnext |= mask; | |
3844 | oldmask ^= rnp->expmaskinitnext; | |
3845 | nbits = bitmap_weight(&oldmask, BITS_PER_LONG); | |
3846 | /* Allow lockless access for expedited grace periods. */ | |
eb7a6653 | 3847 | smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + nbits); /* ^^^ */ |
2f084695 | 3848 | ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus); |
b97d23c5 | 3849 | rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */ |
eb7a6653 PM |
3850 | rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq); |
3851 | rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags); | |
b97d23c5 | 3852 | if (rnp->qsmask & mask) { /* RCU waiting on incoming CPU? */ |
516e5ae0 | 3853 | rcu_disable_urgency_upon_qs(rdp); |
b97d23c5 PM |
3854 | /* Report QS -after- changing ->qsmaskinitnext! */ |
3855 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); | |
3856 | } else { | |
3857 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
7ec99de3 | 3858 | } |
313517fc | 3859 | smp_mb(); /* Ensure RCU read-side usage follows above initialization. */ |
7ec99de3 PM |
3860 | } |
3861 | ||
27d50c7e TG |
3862 | #ifdef CONFIG_HOTPLUG_CPU |
3863 | /* | |
53b46303 PM |
3864 | * The outgoing function has no further need of RCU, so remove it from |
3865 | * the rcu_node tree's ->qsmaskinitnext bit masks. | |
3866 | * | |
3867 | * Note that this function is special in that it is invoked directly | |
3868 | * from the outgoing CPU rather than from the cpuhp_step mechanism. | |
3869 | * This is because this function must be invoked at a precise location. | |
27d50c7e | 3870 | */ |
53b46303 | 3871 | void rcu_report_dead(unsigned int cpu) |
27d50c7e TG |
3872 | { |
3873 | unsigned long flags; | |
3874 | unsigned long mask; | |
da1df50d | 3875 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
27d50c7e TG |
3876 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
3877 | ||
49918a54 | 3878 | /* QS for any half-done expedited grace period. */ |
53b46303 | 3879 | preempt_disable(); |
63d4c8c9 | 3880 | rcu_report_exp_rdp(this_cpu_ptr(&rcu_data)); |
53b46303 PM |
3881 | preempt_enable(); |
3882 | rcu_preempt_deferred_qs(current); | |
3883 | ||
27d50c7e TG |
3884 | /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ |
3885 | mask = rdp->grpmask; | |
894d45bb | 3886 | raw_spin_lock(&rcu_state.ofl_lock); |
27d50c7e | 3887 | raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */ |
53b46303 PM |
3888 | rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq); |
3889 | rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags); | |
fece2776 PM |
3890 | if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */ |
3891 | /* Report quiescent state -before- changing ->qsmaskinitnext! */ | |
b50912d0 | 3892 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
fece2776 PM |
3893 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
3894 | } | |
105abf82 | 3895 | WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask); |
710d60cb | 3896 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
894d45bb | 3897 | raw_spin_unlock(&rcu_state.ofl_lock); |
f64c6013 PZ |
3898 | |
3899 | per_cpu(rcu_cpu_started, cpu) = 0; | |
27d50c7e | 3900 | } |
a58163d8 | 3901 | |
53b46303 PM |
3902 | /* |
3903 | * The outgoing CPU has just passed through the dying-idle state, and we | |
3904 | * are being invoked from the CPU that was IPIed to continue the offline | |
3905 | * operation. Migrate the outgoing CPU's callbacks to the current CPU. | |
3906 | */ | |
3907 | void rcutree_migrate_callbacks(int cpu) | |
a58163d8 PM |
3908 | { |
3909 | unsigned long flags; | |
b1a2d79f | 3910 | struct rcu_data *my_rdp; |
c00045be | 3911 | struct rcu_node *my_rnp; |
da1df50d | 3912 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
ec4eacce | 3913 | bool needwake; |
a58163d8 | 3914 | |
ce5215c1 PM |
3915 | if (rcu_segcblist_is_offloaded(&rdp->cblist) || |
3916 | rcu_segcblist_empty(&rdp->cblist)) | |
95335c03 PM |
3917 | return; /* No callbacks to migrate. */ |
3918 | ||
b1a2d79f | 3919 | local_irq_save(flags); |
da1df50d | 3920 | my_rdp = this_cpu_ptr(&rcu_data); |
c00045be | 3921 | my_rnp = my_rdp->mynode; |
5d6742b3 | 3922 | rcu_nocb_lock(my_rdp); /* irqs already disabled. */ |
d1b222c6 | 3923 | WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies)); |
c00045be | 3924 | raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */ |
ec4eacce | 3925 | /* Leverage recent GPs and set GP for new callbacks. */ |
c00045be PM |
3926 | needwake = rcu_advance_cbs(my_rnp, rdp) || |
3927 | rcu_advance_cbs(my_rnp, my_rdp); | |
f2dbe4a5 | 3928 | rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist); |
23651d9b | 3929 | needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp); |
c035280f | 3930 | rcu_segcblist_disable(&rdp->cblist); |
09efeeee PM |
3931 | WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != |
3932 | !rcu_segcblist_n_cbs(&my_rdp->cblist)); | |
5d6742b3 PM |
3933 | if (rcu_segcblist_is_offloaded(&my_rdp->cblist)) { |
3934 | raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */ | |
3935 | __call_rcu_nocb_wake(my_rdp, true, flags); | |
3936 | } else { | |
3937 | rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */ | |
3938 | raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags); | |
3939 | } | |
ec4eacce | 3940 | if (needwake) |
532c00c9 | 3941 | rcu_gp_kthread_wake(); |
5d6742b3 | 3942 | lockdep_assert_irqs_enabled(); |
a58163d8 PM |
3943 | WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 || |
3944 | !rcu_segcblist_empty(&rdp->cblist), | |
3945 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n", | |
3946 | cpu, rcu_segcblist_n_cbs(&rdp->cblist), | |
3947 | rcu_segcblist_first_cb(&rdp->cblist)); | |
3948 | } | |
27d50c7e TG |
3949 | #endif |
3950 | ||
deb34f36 PM |
3951 | /* |
3952 | * On non-huge systems, use expedited RCU grace periods to make suspend | |
3953 | * and hibernation run faster. | |
3954 | */ | |
d1d74d14 BP |
3955 | static int rcu_pm_notify(struct notifier_block *self, |
3956 | unsigned long action, void *hcpu) | |
3957 | { | |
3958 | switch (action) { | |
3959 | case PM_HIBERNATION_PREPARE: | |
3960 | case PM_SUSPEND_PREPARE: | |
e85e6a21 | 3961 | rcu_expedite_gp(); |
d1d74d14 BP |
3962 | break; |
3963 | case PM_POST_HIBERNATION: | |
3964 | case PM_POST_SUSPEND: | |
e85e6a21 | 3965 | rcu_unexpedite_gp(); |
d1d74d14 BP |
3966 | break; |
3967 | default: | |
3968 | break; | |
3969 | } | |
3970 | return NOTIFY_OK; | |
3971 | } | |
3972 | ||
b3dbec76 | 3973 | /* |
49918a54 | 3974 | * Spawn the kthreads that handle RCU's grace periods. |
b3dbec76 PM |
3975 | */ |
3976 | static int __init rcu_spawn_gp_kthread(void) | |
3977 | { | |
3978 | unsigned long flags; | |
a94844b2 | 3979 | int kthread_prio_in = kthread_prio; |
b3dbec76 | 3980 | struct rcu_node *rnp; |
a94844b2 | 3981 | struct sched_param sp; |
b3dbec76 PM |
3982 | struct task_struct *t; |
3983 | ||
a94844b2 | 3984 | /* Force priority into range. */ |
c7cd161e JFG |
3985 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2 |
3986 | && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)) | |
3987 | kthread_prio = 2; | |
3988 | else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) | |
a94844b2 PM |
3989 | kthread_prio = 1; |
3990 | else if (kthread_prio < 0) | |
3991 | kthread_prio = 0; | |
3992 | else if (kthread_prio > 99) | |
3993 | kthread_prio = 99; | |
c7cd161e | 3994 | |
a94844b2 PM |
3995 | if (kthread_prio != kthread_prio_in) |
3996 | pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n", | |
3997 | kthread_prio, kthread_prio_in); | |
3998 | ||
9386c0b7 | 3999 | rcu_scheduler_fully_active = 1; |
b97d23c5 | 4000 | t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name); |
08543bda PM |
4001 | if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__)) |
4002 | return 0; | |
b97d23c5 PM |
4003 | if (kthread_prio) { |
4004 | sp.sched_priority = kthread_prio; | |
4005 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
b3dbec76 | 4006 | } |
b97d23c5 PM |
4007 | rnp = rcu_get_root(); |
4008 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
5648d659 PM |
4009 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
4010 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); | |
4011 | // Reset .gp_activity and .gp_req_activity before setting .gp_kthread. | |
4012 | smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */ | |
b97d23c5 PM |
4013 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
4014 | wake_up_process(t); | |
35ce7f29 | 4015 | rcu_spawn_nocb_kthreads(); |
9386c0b7 | 4016 | rcu_spawn_boost_kthreads(); |
b3dbec76 PM |
4017 | return 0; |
4018 | } | |
4019 | early_initcall(rcu_spawn_gp_kthread); | |
4020 | ||
bbad9379 | 4021 | /* |
52d7e48b PM |
4022 | * This function is invoked towards the end of the scheduler's |
4023 | * initialization process. Before this is called, the idle task might | |
4024 | * contain synchronous grace-period primitives (during which time, this idle | |
4025 | * task is booting the system, and such primitives are no-ops). After this | |
4026 | * function is called, any synchronous grace-period primitives are run as | |
4027 | * expedited, with the requesting task driving the grace period forward. | |
900b1028 | 4028 | * A later core_initcall() rcu_set_runtime_mode() will switch to full |
52d7e48b | 4029 | * runtime RCU functionality. |
bbad9379 PM |
4030 | */ |
4031 | void rcu_scheduler_starting(void) | |
4032 | { | |
4033 | WARN_ON(num_online_cpus() != 1); | |
4034 | WARN_ON(nr_context_switches() > 0); | |
52d7e48b PM |
4035 | rcu_test_sync_prims(); |
4036 | rcu_scheduler_active = RCU_SCHEDULER_INIT; | |
4037 | rcu_test_sync_prims(); | |
bbad9379 PM |
4038 | } |
4039 | ||
64db4cff | 4040 | /* |
49918a54 | 4041 | * Helper function for rcu_init() that initializes the rcu_state structure. |
64db4cff | 4042 | */ |
b8bb1f63 | 4043 | static void __init rcu_init_one(void) |
64db4cff | 4044 | { |
cb007102 AG |
4045 | static const char * const buf[] = RCU_NODE_NAME_INIT; |
4046 | static const char * const fqs[] = RCU_FQS_NAME_INIT; | |
3dc5dbe9 PM |
4047 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
4048 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; | |
199977bf | 4049 | |
199977bf | 4050 | int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */ |
64db4cff PM |
4051 | int cpustride = 1; |
4052 | int i; | |
4053 | int j; | |
4054 | struct rcu_node *rnp; | |
4055 | ||
05b84aec | 4056 | BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
b6407e86 | 4057 | |
3eaaaf6c PM |
4058 | /* Silence gcc 4.8 false positive about array index out of range. */ |
4059 | if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS) | |
4060 | panic("rcu_init_one: rcu_num_lvls out of range"); | |
4930521a | 4061 | |
64db4cff PM |
4062 | /* Initialize the level-tracking arrays. */ |
4063 | ||
f885b7f2 | 4064 | for (i = 1; i < rcu_num_lvls; i++) |
eb7a6653 PM |
4065 | rcu_state.level[i] = |
4066 | rcu_state.level[i - 1] + num_rcu_lvl[i - 1]; | |
41f5c631 | 4067 | rcu_init_levelspread(levelspread, num_rcu_lvl); |
64db4cff PM |
4068 | |
4069 | /* Initialize the elements themselves, starting from the leaves. */ | |
4070 | ||
f885b7f2 | 4071 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
199977bf | 4072 | cpustride *= levelspread[i]; |
eb7a6653 | 4073 | rnp = rcu_state.level[i]; |
41f5c631 | 4074 | for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) { |
67c583a7 BF |
4075 | raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock)); |
4076 | lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock), | |
b6407e86 | 4077 | &rcu_node_class[i], buf[i]); |
394f2769 PM |
4078 | raw_spin_lock_init(&rnp->fqslock); |
4079 | lockdep_set_class_and_name(&rnp->fqslock, | |
4080 | &rcu_fqs_class[i], fqs[i]); | |
eb7a6653 PM |
4081 | rnp->gp_seq = rcu_state.gp_seq; |
4082 | rnp->gp_seq_needed = rcu_state.gp_seq; | |
4083 | rnp->completedqs = rcu_state.gp_seq; | |
64db4cff PM |
4084 | rnp->qsmask = 0; |
4085 | rnp->qsmaskinit = 0; | |
4086 | rnp->grplo = j * cpustride; | |
4087 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
4088 | if (rnp->grphi >= nr_cpu_ids) |
4089 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
4090 | if (i == 0) { |
4091 | rnp->grpnum = 0; | |
4092 | rnp->grpmask = 0; | |
4093 | rnp->parent = NULL; | |
4094 | } else { | |
199977bf | 4095 | rnp->grpnum = j % levelspread[i - 1]; |
df63fa5b | 4096 | rnp->grpmask = BIT(rnp->grpnum); |
eb7a6653 | 4097 | rnp->parent = rcu_state.level[i - 1] + |
199977bf | 4098 | j / levelspread[i - 1]; |
64db4cff PM |
4099 | } |
4100 | rnp->level = i; | |
12f5f524 | 4101 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 4102 | rcu_init_one_nocb(rnp); |
f6a12f34 PM |
4103 | init_waitqueue_head(&rnp->exp_wq[0]); |
4104 | init_waitqueue_head(&rnp->exp_wq[1]); | |
3b5f668e PM |
4105 | init_waitqueue_head(&rnp->exp_wq[2]); |
4106 | init_waitqueue_head(&rnp->exp_wq[3]); | |
f6a12f34 | 4107 | spin_lock_init(&rnp->exp_lock); |
64db4cff PM |
4108 | } |
4109 | } | |
0c34029a | 4110 | |
eb7a6653 PM |
4111 | init_swait_queue_head(&rcu_state.gp_wq); |
4112 | init_swait_queue_head(&rcu_state.expedited_wq); | |
aedf4ba9 | 4113 | rnp = rcu_first_leaf_node(); |
0c34029a | 4114 | for_each_possible_cpu(i) { |
4a90a068 | 4115 | while (i > rnp->grphi) |
0c34029a | 4116 | rnp++; |
da1df50d | 4117 | per_cpu_ptr(&rcu_data, i)->mynode = rnp; |
53b46303 | 4118 | rcu_boot_init_percpu_data(i); |
0c34029a | 4119 | } |
64db4cff PM |
4120 | } |
4121 | ||
f885b7f2 PM |
4122 | /* |
4123 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 4124 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
4125 | * the ->node array in the rcu_state structure. |
4126 | */ | |
4127 | static void __init rcu_init_geometry(void) | |
4128 | { | |
026ad283 | 4129 | ulong d; |
f885b7f2 | 4130 | int i; |
05b84aec | 4131 | int rcu_capacity[RCU_NUM_LVLS]; |
f885b7f2 | 4132 | |
026ad283 PM |
4133 | /* |
4134 | * Initialize any unspecified boot parameters. | |
4135 | * The default values of jiffies_till_first_fqs and | |
4136 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
4137 | * value, which is a function of HZ, then adding one for each | |
4138 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
4139 | */ | |
4140 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
4141 | if (jiffies_till_first_fqs == ULONG_MAX) | |
4142 | jiffies_till_first_fqs = d; | |
4143 | if (jiffies_till_next_fqs == ULONG_MAX) | |
4144 | jiffies_till_next_fqs = d; | |
6973032a | 4145 | adjust_jiffies_till_sched_qs(); |
026ad283 | 4146 | |
f885b7f2 | 4147 | /* If the compile-time values are accurate, just leave. */ |
47d631af | 4148 | if (rcu_fanout_leaf == RCU_FANOUT_LEAF && |
b17c7035 | 4149 | nr_cpu_ids == NR_CPUS) |
f885b7f2 | 4150 | return; |
a7538352 | 4151 | pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n", |
39479098 | 4152 | rcu_fanout_leaf, nr_cpu_ids); |
f885b7f2 | 4153 | |
f885b7f2 | 4154 | /* |
ee968ac6 PM |
4155 | * The boot-time rcu_fanout_leaf parameter must be at least two |
4156 | * and cannot exceed the number of bits in the rcu_node masks. | |
4157 | * Complain and fall back to the compile-time values if this | |
4158 | * limit is exceeded. | |
f885b7f2 | 4159 | */ |
ee968ac6 | 4160 | if (rcu_fanout_leaf < 2 || |
75cf15a4 | 4161 | rcu_fanout_leaf > sizeof(unsigned long) * 8) { |
13bd6494 | 4162 | rcu_fanout_leaf = RCU_FANOUT_LEAF; |
f885b7f2 PM |
4163 | WARN_ON(1); |
4164 | return; | |
4165 | } | |
4166 | ||
f885b7f2 PM |
4167 | /* |
4168 | * Compute number of nodes that can be handled an rcu_node tree | |
9618138b | 4169 | * with the given number of levels. |
f885b7f2 | 4170 | */ |
9618138b | 4171 | rcu_capacity[0] = rcu_fanout_leaf; |
05b84aec | 4172 | for (i = 1; i < RCU_NUM_LVLS; i++) |
05c5df31 | 4173 | rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT; |
f885b7f2 PM |
4174 | |
4175 | /* | |
75cf15a4 | 4176 | * The tree must be able to accommodate the configured number of CPUs. |
ee968ac6 | 4177 | * If this limit is exceeded, fall back to the compile-time values. |
f885b7f2 | 4178 | */ |
ee968ac6 PM |
4179 | if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) { |
4180 | rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
4181 | WARN_ON(1); | |
4182 | return; | |
4183 | } | |
f885b7f2 | 4184 | |
679f9858 | 4185 | /* Calculate the number of levels in the tree. */ |
9618138b | 4186 | for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) { |
679f9858 | 4187 | } |
9618138b | 4188 | rcu_num_lvls = i + 1; |
679f9858 | 4189 | |
f885b7f2 | 4190 | /* Calculate the number of rcu_nodes at each level of the tree. */ |
679f9858 | 4191 | for (i = 0; i < rcu_num_lvls; i++) { |
9618138b | 4192 | int cap = rcu_capacity[(rcu_num_lvls - 1) - i]; |
679f9858 AG |
4193 | num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap); |
4194 | } | |
f885b7f2 PM |
4195 | |
4196 | /* Calculate the total number of rcu_node structures. */ | |
4197 | rcu_num_nodes = 0; | |
679f9858 | 4198 | for (i = 0; i < rcu_num_lvls; i++) |
f885b7f2 | 4199 | rcu_num_nodes += num_rcu_lvl[i]; |
f885b7f2 PM |
4200 | } |
4201 | ||
a3dc2948 PM |
4202 | /* |
4203 | * Dump out the structure of the rcu_node combining tree associated | |
49918a54 | 4204 | * with the rcu_state structure. |
a3dc2948 | 4205 | */ |
b8bb1f63 | 4206 | static void __init rcu_dump_rcu_node_tree(void) |
a3dc2948 PM |
4207 | { |
4208 | int level = 0; | |
4209 | struct rcu_node *rnp; | |
4210 | ||
4211 | pr_info("rcu_node tree layout dump\n"); | |
4212 | pr_info(" "); | |
aedf4ba9 | 4213 | rcu_for_each_node_breadth_first(rnp) { |
a3dc2948 PM |
4214 | if (rnp->level != level) { |
4215 | pr_cont("\n"); | |
4216 | pr_info(" "); | |
4217 | level = rnp->level; | |
4218 | } | |
4219 | pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum); | |
4220 | } | |
4221 | pr_cont("\n"); | |
4222 | } | |
4223 | ||
ad7c946b | 4224 | struct workqueue_struct *rcu_gp_wq; |
25f3d7ef | 4225 | struct workqueue_struct *rcu_par_gp_wq; |
ad7c946b | 4226 | |
a35d1690 BP |
4227 | static void __init kfree_rcu_batch_init(void) |
4228 | { | |
4229 | int cpu; | |
0392bebe | 4230 | int i; |
a35d1690 BP |
4231 | |
4232 | for_each_possible_cpu(cpu) { | |
4233 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); | |
4234 | ||
4235 | spin_lock_init(&krcp->lock); | |
34c88174 URS |
4236 | for (i = 0; i < KFREE_N_BATCHES; i++) { |
4237 | INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work); | |
0392bebe | 4238 | krcp->krw_arr[i].krcp = krcp; |
34c88174 URS |
4239 | } |
4240 | ||
a35d1690 BP |
4241 | INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor); |
4242 | krcp->initialized = true; | |
4243 | } | |
9154244c JFG |
4244 | if (register_shrinker(&kfree_rcu_shrinker)) |
4245 | pr_err("Failed to register kfree_rcu() shrinker!\n"); | |
a35d1690 BP |
4246 | } |
4247 | ||
9f680ab4 | 4248 | void __init rcu_init(void) |
64db4cff | 4249 | { |
017c4261 | 4250 | int cpu; |
9f680ab4 | 4251 | |
47627678 PM |
4252 | rcu_early_boot_tests(); |
4253 | ||
a35d1690 | 4254 | kfree_rcu_batch_init(); |
f41d911f | 4255 | rcu_bootup_announce(); |
f885b7f2 | 4256 | rcu_init_geometry(); |
b8bb1f63 | 4257 | rcu_init_one(); |
a3dc2948 | 4258 | if (dump_tree) |
b8bb1f63 | 4259 | rcu_dump_rcu_node_tree(); |
48d07c04 SAS |
4260 | if (use_softirq) |
4261 | open_softirq(RCU_SOFTIRQ, rcu_core_si); | |
9f680ab4 PM |
4262 | |
4263 | /* | |
4264 | * We don't need protection against CPU-hotplug here because | |
4265 | * this is called early in boot, before either interrupts | |
4266 | * or the scheduler are operational. | |
4267 | */ | |
d1d74d14 | 4268 | pm_notifier(rcu_pm_notify, 0); |
7ec99de3 | 4269 | for_each_online_cpu(cpu) { |
4df83742 | 4270 | rcutree_prepare_cpu(cpu); |
7ec99de3 | 4271 | rcu_cpu_starting(cpu); |
9b9500da | 4272 | rcutree_online_cpu(cpu); |
7ec99de3 | 4273 | } |
ad7c946b PM |
4274 | |
4275 | /* Create workqueue for expedited GPs and for Tree SRCU. */ | |
4276 | rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0); | |
4277 | WARN_ON(!rcu_gp_wq); | |
25f3d7ef PM |
4278 | rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0); |
4279 | WARN_ON(!rcu_par_gp_wq); | |
e0fcba9a | 4280 | srcu_init(); |
b2b00ddf PM |
4281 | |
4282 | /* Fill in default value for rcutree.qovld boot parameter. */ | |
4283 | /* -After- the rcu_node ->lock fields are initialized! */ | |
4284 | if (qovld < 0) | |
4285 | qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark; | |
4286 | else | |
4287 | qovld_calc = qovld; | |
64db4cff PM |
4288 | } |
4289 | ||
10462d6f | 4290 | #include "tree_stall.h" |
3549c2bc | 4291 | #include "tree_exp.h" |
4102adab | 4292 | #include "tree_plugin.h" |