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> |
5f3c8d62 URS |
60 | #include <linux/vmalloc.h> |
61 | #include <linux/mm.h> | |
26e760c9 | 62 | #include <linux/kasan.h> |
48d07c04 | 63 | #include "../time/tick-internal.h" |
64db4cff | 64 | |
4102adab | 65 | #include "tree.h" |
29c00b4a | 66 | #include "rcu.h" |
9f77da9f | 67 | |
4102adab PM |
68 | #ifdef MODULE_PARAM_PREFIX |
69 | #undef MODULE_PARAM_PREFIX | |
70 | #endif | |
71 | #define MODULE_PARAM_PREFIX "rcutree." | |
72 | ||
64db4cff PM |
73 | /* Data structures. */ |
74 | ||
f7f7bac9 | 75 | /* |
dc5a4f29 PM |
76 | * Steal a bit from the bottom of ->dynticks for idle entry/exit |
77 | * control. Initially this is for TLB flushing. | |
f7f7bac9 | 78 | */ |
dc5a4f29 PM |
79 | #define RCU_DYNTICK_CTRL_MASK 0x1 |
80 | #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1) | |
a8a29b3b | 81 | |
4c5273bf PM |
82 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = { |
83 | .dynticks_nesting = 1, | |
84 | .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE, | |
dc5a4f29 | 85 | .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR), |
8d346d43 FW |
86 | #ifdef CONFIG_RCU_NOCB_CPU |
87 | .cblist.flags = SEGCBLIST_SOFTIRQ_ONLY, | |
88 | #endif | |
4c5273bf | 89 | }; |
c30fe541 | 90 | static struct rcu_state rcu_state = { |
358be2d3 | 91 | .level = { &rcu_state.node[0] }, |
358be2d3 PM |
92 | .gp_state = RCU_GP_IDLE, |
93 | .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT, | |
94 | .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex), | |
95 | .name = RCU_NAME, | |
96 | .abbr = RCU_ABBR, | |
97 | .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex), | |
98 | .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex), | |
894d45bb | 99 | .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock), |
358be2d3 | 100 | }; |
27f4d280 | 101 | |
a3dc2948 PM |
102 | /* Dump rcu_node combining tree at boot to verify correct setup. */ |
103 | static bool dump_tree; | |
104 | module_param(dump_tree, bool, 0444); | |
48d07c04 | 105 | /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */ |
8b9a0ecc SW |
106 | static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT); |
107 | #ifndef CONFIG_PREEMPT_RT | |
48d07c04 | 108 | module_param(use_softirq, bool, 0444); |
8b9a0ecc | 109 | #endif |
7fa27001 PM |
110 | /* Control rcu_node-tree auto-balancing at boot time. */ |
111 | static bool rcu_fanout_exact; | |
112 | module_param(rcu_fanout_exact, bool, 0444); | |
47d631af PM |
113 | /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */ |
114 | static int rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
7e5c2dfb | 115 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 | 116 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
cb007102 | 117 | /* Number of rcu_nodes at specified level. */ |
e95d68d2 | 118 | int num_rcu_lvl[] = NUM_RCU_LVL_INIT; |
f885b7f2 PM |
119 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ |
120 | ||
b0d30417 | 121 | /* |
52d7e48b PM |
122 | * The rcu_scheduler_active variable is initialized to the value |
123 | * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the | |
124 | * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE, | |
125 | * RCU can assume that there is but one task, allowing RCU to (for example) | |
0d95092c | 126 | * optimize synchronize_rcu() to a simple barrier(). When this variable |
52d7e48b PM |
127 | * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required |
128 | * to detect real grace periods. This variable is also used to suppress | |
129 | * boot-time false positives from lockdep-RCU error checking. Finally, it | |
130 | * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU | |
131 | * is fully initialized, including all of its kthreads having been spawned. | |
b0d30417 | 132 | */ |
bbad9379 PM |
133 | int rcu_scheduler_active __read_mostly; |
134 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
135 | ||
b0d30417 PM |
136 | /* |
137 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
138 | * during the early_initcall() processing, which is after the scheduler | |
139 | * is capable of creating new tasks. So RCU processing (for example, | |
140 | * creating tasks for RCU priority boosting) must be delayed until after | |
141 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
142 | * currently delay invocation of any RCU callbacks until after this point. | |
143 | * | |
144 | * It might later prove better for people registering RCU callbacks during | |
145 | * early boot to take responsibility for these callbacks, but one step at | |
146 | * a time. | |
147 | */ | |
148 | static int rcu_scheduler_fully_active __read_mostly; | |
149 | ||
b50912d0 PM |
150 | static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp, |
151 | unsigned long gps, unsigned long flags); | |
0aa04b05 PM |
152 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); |
153 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); | |
5d01bbd1 | 154 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
a46e0899 | 155 | static void invoke_rcu_core(void); |
63d4c8c9 | 156 | static void rcu_report_exp_rdp(struct rcu_data *rdp); |
3549c2bc | 157 | static void sync_sched_exp_online_cleanup(int cpu); |
b2b00ddf | 158 | static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp); |
3820b513 | 159 | static bool rcu_rdp_is_offloaded(struct rcu_data *rdp); |
a26ac245 | 160 | |
a94844b2 | 161 | /* rcuc/rcub kthread realtime priority */ |
26730f55 | 162 | static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0; |
3ffe3d1a | 163 | module_param(kthread_prio, int, 0444); |
a94844b2 | 164 | |
8d7dc928 | 165 | /* Delay in jiffies for grace-period initialization delays, debug only. */ |
0f41c0dd | 166 | |
90040c9e PM |
167 | static int gp_preinit_delay; |
168 | module_param(gp_preinit_delay, int, 0444); | |
169 | static int gp_init_delay; | |
170 | module_param(gp_init_delay, int, 0444); | |
171 | static int gp_cleanup_delay; | |
172 | module_param(gp_cleanup_delay, int, 0444); | |
0f41c0dd | 173 | |
aa40c138 PM |
174 | // Add delay to rcu_read_unlock() for strict grace periods. |
175 | static int rcu_unlock_delay; | |
176 | #ifdef CONFIG_RCU_STRICT_GRACE_PERIOD | |
177 | module_param(rcu_unlock_delay, int, 0444); | |
178 | #endif | |
179 | ||
53c72b59 URS |
180 | /* |
181 | * This rcu parameter is runtime-read-only. It reflects | |
182 | * a minimum allowed number of objects which can be cached | |
183 | * per-CPU. Object size is equal to one page. This value | |
184 | * can be changed at boot time. | |
185 | */ | |
56292e86 | 186 | static int rcu_min_cached_objs = 5; |
53c72b59 URS |
187 | module_param(rcu_min_cached_objs, int, 0444); |
188 | ||
4cf439a2 | 189 | /* Retrieve RCU kthreads priority for rcutorture */ |
4babd855 JFG |
190 | int rcu_get_gp_kthreads_prio(void) |
191 | { | |
192 | return kthread_prio; | |
193 | } | |
194 | EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio); | |
195 | ||
eab128e8 PM |
196 | /* |
197 | * Number of grace periods between delays, normalized by the duration of | |
bfd090be | 198 | * the delay. The longer the delay, the more the grace periods between |
eab128e8 PM |
199 | * each delay. The reason for this normalization is that it means that, |
200 | * for non-zero delays, the overall slowdown of grace periods is constant | |
201 | * regardless of the duration of the delay. This arrangement balances | |
202 | * the need for long delays to increase some race probabilities with the | |
203 | * need for fast grace periods to increase other race probabilities. | |
204 | */ | |
205 | #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */ | |
37745d28 | 206 | |
0aa04b05 PM |
207 | /* |
208 | * Compute the mask of online CPUs for the specified rcu_node structure. | |
209 | * This will not be stable unless the rcu_node structure's ->lock is | |
210 | * held, but the bit corresponding to the current CPU will be stable | |
211 | * in most contexts. | |
212 | */ | |
c30fe541 | 213 | static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) |
0aa04b05 | 214 | { |
7d0ae808 | 215 | return READ_ONCE(rnp->qsmaskinitnext); |
0aa04b05 PM |
216 | } |
217 | ||
fc2219d4 | 218 | /* |
7d0ae808 | 219 | * Return true if an RCU grace period is in progress. The READ_ONCE()s |
fc2219d4 PM |
220 | * permit this function to be invoked without holding the root rcu_node |
221 | * structure's ->lock, but of course results can be subject to change. | |
222 | */ | |
de8e8730 | 223 | static int rcu_gp_in_progress(void) |
fc2219d4 | 224 | { |
de8e8730 | 225 | return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq)); |
b1f77b05 IM |
226 | } |
227 | ||
903ee83d PM |
228 | /* |
229 | * Return the number of callbacks queued on the specified CPU. | |
230 | * Handles both the nocbs and normal cases. | |
231 | */ | |
232 | static long rcu_get_n_cbs_cpu(int cpu) | |
233 | { | |
234 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
235 | ||
c035280f | 236 | if (rcu_segcblist_is_enabled(&rdp->cblist)) |
903ee83d | 237 | return rcu_segcblist_n_cbs(&rdp->cblist); |
c035280f | 238 | return 0; |
903ee83d PM |
239 | } |
240 | ||
d28139c4 | 241 | void rcu_softirq_qs(void) |
b1f77b05 | 242 | { |
45975c7d | 243 | rcu_qs(); |
d28139c4 | 244 | rcu_preempt_deferred_qs(current); |
b1f77b05 | 245 | } |
64db4cff | 246 | |
2625d469 PM |
247 | /* |
248 | * Record entry into an extended quiescent state. This is only to be | |
ac3caf82 PM |
249 | * called when not already in an extended quiescent state, that is, |
250 | * RCU is watching prior to the call to this function and is no longer | |
251 | * watching upon return. | |
2625d469 | 252 | */ |
ff5c4f5c | 253 | static noinstr void rcu_dynticks_eqs_enter(void) |
2625d469 | 254 | { |
dc5a4f29 | 255 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
b8c17e66 | 256 | int seq; |
2625d469 PM |
257 | |
258 | /* | |
b8c17e66 | 259 | * CPUs seeing atomic_add_return() must see prior RCU read-side |
2625d469 PM |
260 | * critical sections, and we also must force ordering with the |
261 | * next idle sojourn. | |
262 | */ | |
7d0c9c50 | 263 | rcu_dynticks_task_trace_enter(); // Before ->dynticks update! |
b58e733f | 264 | seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks); |
ac3caf82 | 265 | // RCU is no longer watching. Better be in extended quiescent state! |
b8c17e66 PM |
266 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
267 | (seq & RCU_DYNTICK_CTRL_CTR)); | |
268 | /* Better not have special action (TLB flush) pending! */ | |
269 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && | |
270 | (seq & RCU_DYNTICK_CTRL_MASK)); | |
2625d469 PM |
271 | } |
272 | ||
273 | /* | |
274 | * Record exit from an extended quiescent state. This is only to be | |
ac3caf82 PM |
275 | * called from an extended quiescent state, that is, RCU is not watching |
276 | * prior to the call to this function and is watching upon return. | |
2625d469 | 277 | */ |
ff5c4f5c | 278 | static noinstr void rcu_dynticks_eqs_exit(void) |
2625d469 | 279 | { |
dc5a4f29 | 280 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
b8c17e66 | 281 | int seq; |
2625d469 PM |
282 | |
283 | /* | |
b8c17e66 | 284 | * CPUs seeing atomic_add_return() must see prior idle sojourns, |
2625d469 PM |
285 | * and we also must force ordering with the next RCU read-side |
286 | * critical section. | |
287 | */ | |
b58e733f | 288 | seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks); |
ac3caf82 | 289 | // RCU is now watching. Better not be in an extended quiescent state! |
7d0c9c50 | 290 | rcu_dynticks_task_trace_exit(); // After ->dynticks update! |
b8c17e66 PM |
291 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
292 | !(seq & RCU_DYNTICK_CTRL_CTR)); | |
293 | if (seq & RCU_DYNTICK_CTRL_MASK) { | |
b58e733f | 294 | arch_atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks); |
b8c17e66 | 295 | smp_mb__after_atomic(); /* _exit after clearing mask. */ |
b8c17e66 | 296 | } |
2625d469 PM |
297 | } |
298 | ||
299 | /* | |
300 | * Reset the current CPU's ->dynticks counter to indicate that the | |
301 | * newly onlined CPU is no longer in an extended quiescent state. | |
302 | * This will either leave the counter unchanged, or increment it | |
303 | * to the next non-quiescent value. | |
304 | * | |
305 | * The non-atomic test/increment sequence works because the upper bits | |
306 | * of the ->dynticks counter are manipulated only by the corresponding CPU, | |
307 | * or when the corresponding CPU is offline. | |
308 | */ | |
309 | static void rcu_dynticks_eqs_online(void) | |
310 | { | |
dc5a4f29 | 311 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
2625d469 | 312 | |
dc5a4f29 | 313 | if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR) |
2625d469 | 314 | return; |
dc5a4f29 | 315 | atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks); |
2625d469 PM |
316 | } |
317 | ||
02a5c550 PM |
318 | /* |
319 | * Is the current CPU in an extended quiescent state? | |
320 | * | |
321 | * No ordering, as we are sampling CPU-local information. | |
322 | */ | |
ff5c4f5c | 323 | static __always_inline bool rcu_dynticks_curr_cpu_in_eqs(void) |
02a5c550 | 324 | { |
dc5a4f29 | 325 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
02a5c550 | 326 | |
b58e733f | 327 | return !(arch_atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR); |
02a5c550 PM |
328 | } |
329 | ||
8b2f63ab PM |
330 | /* |
331 | * Snapshot the ->dynticks counter with full ordering so as to allow | |
332 | * stable comparison of this counter with past and future snapshots. | |
333 | */ | |
c30fe541 | 334 | static int rcu_dynticks_snap(struct rcu_data *rdp) |
8b2f63ab | 335 | { |
dc5a4f29 | 336 | int snap = atomic_add_return(0, &rdp->dynticks); |
8b2f63ab | 337 | |
b8c17e66 | 338 | return snap & ~RCU_DYNTICK_CTRL_MASK; |
8b2f63ab PM |
339 | } |
340 | ||
02a5c550 PM |
341 | /* |
342 | * Return true if the snapshot returned from rcu_dynticks_snap() | |
343 | * indicates that RCU is in an extended quiescent state. | |
344 | */ | |
345 | static bool rcu_dynticks_in_eqs(int snap) | |
346 | { | |
b8c17e66 | 347 | return !(snap & RCU_DYNTICK_CTRL_CTR); |
02a5c550 PM |
348 | } |
349 | ||
3fcd6a23 PM |
350 | /* Return true if the specified CPU is currently idle from an RCU viewpoint. */ |
351 | bool rcu_is_idle_cpu(int cpu) | |
352 | { | |
353 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
354 | ||
355 | return rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)); | |
356 | } | |
357 | ||
02a5c550 | 358 | /* |
dc5a4f29 | 359 | * Return true if the CPU corresponding to the specified rcu_data |
02a5c550 PM |
360 | * structure has spent some time in an extended quiescent state since |
361 | * rcu_dynticks_snap() returned the specified snapshot. | |
362 | */ | |
dc5a4f29 | 363 | static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap) |
02a5c550 | 364 | { |
dc5a4f29 | 365 | return snap != rcu_dynticks_snap(rdp); |
02a5c550 PM |
366 | } |
367 | ||
7d0c9c50 PM |
368 | /* |
369 | * Return true if the referenced integer is zero while the specified | |
370 | * CPU remains within a single extended quiescent state. | |
371 | */ | |
372 | bool rcu_dynticks_zero_in_eqs(int cpu, int *vp) | |
373 | { | |
374 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
375 | int snap; | |
376 | ||
377 | // If not quiescent, force back to earlier extended quiescent state. | |
378 | snap = atomic_read(&rdp->dynticks) & ~(RCU_DYNTICK_CTRL_MASK | | |
379 | RCU_DYNTICK_CTRL_CTR); | |
380 | ||
381 | smp_rmb(); // Order ->dynticks and *vp reads. | |
382 | if (READ_ONCE(*vp)) | |
383 | return false; // Non-zero, so report failure; | |
384 | smp_rmb(); // Order *vp read and ->dynticks re-read. | |
385 | ||
386 | // If still in the same extended quiescent state, we are good! | |
387 | return snap == (atomic_read(&rdp->dynticks) & ~RCU_DYNTICK_CTRL_MASK); | |
388 | } | |
389 | ||
b8c17e66 PM |
390 | /* |
391 | * Set the special (bottom) bit of the specified CPU so that it | |
392 | * will take special action (such as flushing its TLB) on the | |
393 | * next exit from an extended quiescent state. Returns true if | |
394 | * the bit was successfully set, or false if the CPU was not in | |
395 | * an extended quiescent state. | |
396 | */ | |
397 | bool rcu_eqs_special_set(int cpu) | |
398 | { | |
399 | int old; | |
400 | int new; | |
faa059c3 | 401 | int new_old; |
dc5a4f29 | 402 | struct rcu_data *rdp = &per_cpu(rcu_data, cpu); |
b8c17e66 | 403 | |
faa059c3 | 404 | new_old = atomic_read(&rdp->dynticks); |
b8c17e66 | 405 | do { |
faa059c3 | 406 | old = new_old; |
b8c17e66 PM |
407 | if (old & RCU_DYNTICK_CTRL_CTR) |
408 | return false; | |
409 | new = old | RCU_DYNTICK_CTRL_MASK; | |
faa059c3 PM |
410 | new_old = atomic_cmpxchg(&rdp->dynticks, old, new); |
411 | } while (new_old != old); | |
b8c17e66 | 412 | return true; |
6563de9d | 413 | } |
5cd37193 | 414 | |
4a81e832 PM |
415 | /* |
416 | * Let the RCU core know that this CPU has gone through the scheduler, | |
417 | * which is a quiescent state. This is called when the need for a | |
418 | * quiescent state is urgent, so we burn an atomic operation and full | |
419 | * memory barriers to let the RCU core know about it, regardless of what | |
420 | * this CPU might (or might not) do in the near future. | |
421 | * | |
0f9be8ca | 422 | * We inform the RCU core by emulating a zero-duration dyntick-idle period. |
46a5d164 | 423 | * |
3b57a399 | 424 | * The caller must have disabled interrupts and must not be idle. |
4a81e832 | 425 | */ |
4230e2de | 426 | notrace void rcu_momentary_dyntick_idle(void) |
4a81e832 | 427 | { |
3b57a399 PM |
428 | int special; |
429 | ||
2dba13f0 | 430 | raw_cpu_write(rcu_data.rcu_need_heavy_qs, false); |
dc5a4f29 PM |
431 | special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR, |
432 | &this_cpu_ptr(&rcu_data)->dynticks); | |
3b57a399 PM |
433 | /* It is illegal to call this from idle state. */ |
434 | WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR)); | |
3e310098 | 435 | rcu_preempt_deferred_qs(current); |
4a81e832 | 436 | } |
79ba7ff5 | 437 | EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle); |
4a81e832 | 438 | |
45975c7d | 439 | /** |
806f04e9 | 440 | * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle |
bb73c52b | 441 | * |
eddded80 | 442 | * If the current CPU is idle and running at a first-level (not nested) |
806f04e9 PZ |
443 | * interrupt, or directly, from idle, return true. |
444 | * | |
445 | * The caller must have at least disabled IRQs. | |
5cd37193 | 446 | */ |
45975c7d | 447 | static int rcu_is_cpu_rrupt_from_idle(void) |
5cd37193 | 448 | { |
806f04e9 PZ |
449 | long nesting; |
450 | ||
451 | /* | |
452 | * Usually called from the tick; but also used from smp_function_call() | |
453 | * for expedited grace periods. This latter can result in running from | |
454 | * the idle task, instead of an actual IPI. | |
455 | */ | |
456 | lockdep_assert_irqs_disabled(); | |
eddded80 JFG |
457 | |
458 | /* Check for counter underflows */ | |
459 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0, | |
460 | "RCU dynticks_nesting counter underflow!"); | |
461 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0, | |
462 | "RCU dynticks_nmi_nesting counter underflow/zero!"); | |
463 | ||
464 | /* Are we at first interrupt nesting level? */ | |
806f04e9 PZ |
465 | nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting); |
466 | if (nesting > 1) | |
eddded80 JFG |
467 | return false; |
468 | ||
806f04e9 PZ |
469 | /* |
470 | * If we're not in an interrupt, we must be in the idle task! | |
471 | */ | |
472 | WARN_ON_ONCE(!nesting && !is_idle_task(current)); | |
473 | ||
eddded80 JFG |
474 | /* Does CPU appear to be idle from an RCU standpoint? */ |
475 | return __this_cpu_read(rcu_data.dynticks_nesting) == 0; | |
5cd37193 | 476 | } |
5cd37193 | 477 | |
29fc5f93 PM |
478 | #define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10) |
479 | // Maximum callbacks per rcu_do_batch ... | |
480 | #define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood. | |
17c7798b | 481 | static long blimit = DEFAULT_RCU_BLIMIT; |
29fc5f93 | 482 | #define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit. |
17c7798b | 483 | static long qhimark = DEFAULT_RCU_QHIMARK; |
29fc5f93 | 484 | #define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit. |
17c7798b | 485 | static long qlowmark = DEFAULT_RCU_QLOMARK; |
b2b00ddf PM |
486 | #define DEFAULT_RCU_QOVLD_MULT 2 |
487 | #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK) | |
29fc5f93 PM |
488 | static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS. |
489 | static long qovld_calc = -1; // No pre-initialization lock acquisitions! | |
64db4cff | 490 | |
878d7439 ED |
491 | module_param(blimit, long, 0444); |
492 | module_param(qhimark, long, 0444); | |
493 | module_param(qlowmark, long, 0444); | |
b2b00ddf | 494 | module_param(qovld, long, 0444); |
3d76c082 | 495 | |
aecd34b9 | 496 | static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX; |
026ad283 | 497 | static ulong jiffies_till_next_fqs = ULONG_MAX; |
8c7c4829 | 498 | static bool rcu_kick_kthreads; |
cfcdef5e ED |
499 | static int rcu_divisor = 7; |
500 | module_param(rcu_divisor, int, 0644); | |
501 | ||
502 | /* Force an exit from rcu_do_batch() after 3 milliseconds. */ | |
503 | static long rcu_resched_ns = 3 * NSEC_PER_MSEC; | |
504 | module_param(rcu_resched_ns, long, 0644); | |
d40011f6 | 505 | |
c06aed0e PM |
506 | /* |
507 | * How long the grace period must be before we start recruiting | |
508 | * quiescent-state help from rcu_note_context_switch(). | |
509 | */ | |
510 | static ulong jiffies_till_sched_qs = ULONG_MAX; | |
511 | module_param(jiffies_till_sched_qs, ulong, 0444); | |
85f2b60c | 512 | static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */ |
c06aed0e PM |
513 | module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */ |
514 | ||
515 | /* | |
516 | * Make sure that we give the grace-period kthread time to detect any | |
517 | * idle CPUs before taking active measures to force quiescent states. | |
518 | * However, don't go below 100 milliseconds, adjusted upwards for really | |
519 | * large systems. | |
520 | */ | |
521 | static void adjust_jiffies_till_sched_qs(void) | |
522 | { | |
523 | unsigned long j; | |
524 | ||
525 | /* If jiffies_till_sched_qs was specified, respect the request. */ | |
526 | if (jiffies_till_sched_qs != ULONG_MAX) { | |
527 | WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs); | |
528 | return; | |
529 | } | |
85f2b60c | 530 | /* Otherwise, set to third fqs scan, but bound below on large system. */ |
c06aed0e PM |
531 | j = READ_ONCE(jiffies_till_first_fqs) + |
532 | 2 * READ_ONCE(jiffies_till_next_fqs); | |
533 | if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV) | |
534 | j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
535 | pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j); | |
536 | WRITE_ONCE(jiffies_to_sched_qs, j); | |
537 | } | |
538 | ||
67abb96c BP |
539 | static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp) |
540 | { | |
541 | ulong j; | |
542 | int ret = kstrtoul(val, 0, &j); | |
543 | ||
c06aed0e | 544 | if (!ret) { |
67abb96c | 545 | WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j); |
c06aed0e PM |
546 | adjust_jiffies_till_sched_qs(); |
547 | } | |
67abb96c BP |
548 | return ret; |
549 | } | |
550 | ||
551 | static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp) | |
552 | { | |
553 | ulong j; | |
554 | int ret = kstrtoul(val, 0, &j); | |
555 | ||
c06aed0e | 556 | if (!ret) { |
67abb96c | 557 | WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1)); |
c06aed0e PM |
558 | adjust_jiffies_till_sched_qs(); |
559 | } | |
67abb96c BP |
560 | return ret; |
561 | } | |
562 | ||
7c47ee5a | 563 | static const struct kernel_param_ops first_fqs_jiffies_ops = { |
67abb96c BP |
564 | .set = param_set_first_fqs_jiffies, |
565 | .get = param_get_ulong, | |
566 | }; | |
567 | ||
7c47ee5a | 568 | static const struct kernel_param_ops next_fqs_jiffies_ops = { |
67abb96c BP |
569 | .set = param_set_next_fqs_jiffies, |
570 | .get = param_get_ulong, | |
571 | }; | |
572 | ||
573 | module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644); | |
574 | module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644); | |
8c7c4829 | 575 | module_param(rcu_kick_kthreads, bool, 0644); |
d40011f6 | 576 | |
8ff0b907 | 577 | static void force_qs_rnp(int (*f)(struct rcu_data *rdp)); |
dd7dafd1 | 578 | static int rcu_pending(int user); |
64db4cff PM |
579 | |
580 | /* | |
17ef2fe9 | 581 | * Return the number of RCU GPs completed thus far for debug & stats. |
64db4cff | 582 | */ |
17ef2fe9 | 583 | unsigned long rcu_get_gp_seq(void) |
917963d0 | 584 | { |
16fc9c60 | 585 | return READ_ONCE(rcu_state.gp_seq); |
917963d0 | 586 | } |
17ef2fe9 | 587 | EXPORT_SYMBOL_GPL(rcu_get_gp_seq); |
917963d0 | 588 | |
291783b8 PM |
589 | /* |
590 | * Return the number of RCU expedited batches completed thus far for | |
591 | * debug & stats. Odd numbers mean that a batch is in progress, even | |
592 | * numbers mean idle. The value returned will thus be roughly double | |
593 | * the cumulative batches since boot. | |
594 | */ | |
595 | unsigned long rcu_exp_batches_completed(void) | |
596 | { | |
16fc9c60 | 597 | return rcu_state.expedited_sequence; |
291783b8 PM |
598 | } |
599 | EXPORT_SYMBOL_GPL(rcu_exp_batches_completed); | |
600 | ||
fd897573 PM |
601 | /* |
602 | * Return the root node of the rcu_state structure. | |
603 | */ | |
604 | static struct rcu_node *rcu_get_root(void) | |
605 | { | |
606 | return &rcu_state.node[0]; | |
607 | } | |
608 | ||
ad0dc7f9 PM |
609 | /* |
610 | * Send along grace-period-related data for rcutorture diagnostics. | |
611 | */ | |
612 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, | |
aebc8264 | 613 | unsigned long *gp_seq) |
ad0dc7f9 | 614 | { |
ad0dc7f9 PM |
615 | switch (test_type) { |
616 | case RCU_FLAVOR: | |
f7dd7d44 PM |
617 | *flags = READ_ONCE(rcu_state.gp_flags); |
618 | *gp_seq = rcu_seq_current(&rcu_state.gp_seq); | |
ad0dc7f9 PM |
619 | break; |
620 | default: | |
621 | break; | |
622 | } | |
ad0dc7f9 PM |
623 | } |
624 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
625 | ||
9b2e4f18 | 626 | /* |
215bba9f PM |
627 | * Enter an RCU extended quiescent state, which can be either the |
628 | * idle loop or adaptive-tickless usermode execution. | |
9b2e4f18 | 629 | * |
215bba9f PM |
630 | * We crowbar the ->dynticks_nmi_nesting field to zero to allow for |
631 | * the possibility of usermode upcalls having messed up our count | |
632 | * of interrupt nesting level during the prior busy period. | |
9b2e4f18 | 633 | */ |
ff5c4f5c | 634 | static noinstr void rcu_eqs_enter(bool user) |
9b2e4f18 | 635 | { |
4c5273bf | 636 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
96d3fd0d | 637 | |
4c5273bf PM |
638 | WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE); |
639 | WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); | |
215bba9f | 640 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && |
4c5273bf PM |
641 | rdp->dynticks_nesting == 0); |
642 | if (rdp->dynticks_nesting != 1) { | |
ac3caf82 | 643 | // RCU will still be watching, so just do accounting and leave. |
4c5273bf | 644 | rdp->dynticks_nesting--; |
215bba9f | 645 | return; |
9b2e4f18 | 646 | } |
96d3fd0d | 647 | |
b04db8e1 | 648 | lockdep_assert_irqs_disabled(); |
ff5c4f5c | 649 | instrumentation_begin(); |
6cf539a8 | 650 | trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks)); |
e68bbb26 | 651 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current)); |
198bbf81 | 652 | rcu_prepare_for_idle(); |
3e310098 | 653 | rcu_preempt_deferred_qs(current); |
b58e733f PZ |
654 | |
655 | // instrumentation for the noinstr rcu_dynticks_eqs_enter() | |
656 | instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks)); | |
657 | ||
ff5c4f5c | 658 | instrumentation_end(); |
4c5273bf | 659 | WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */ |
ac3caf82 | 660 | // RCU is watching here ... |
844ccdd7 | 661 | rcu_dynticks_eqs_enter(); |
ac3caf82 | 662 | // ... but is no longer watching here. |
176f8f7a | 663 | rcu_dynticks_task_enter(); |
64db4cff | 664 | } |
adf5091e FW |
665 | |
666 | /** | |
667 | * rcu_idle_enter - inform RCU that current CPU is entering idle | |
668 | * | |
669 | * Enter idle mode, in other words, -leave- the mode in which RCU | |
670 | * read-side critical sections can occur. (Though RCU read-side | |
671 | * critical sections can occur in irq handlers in idle, a possibility | |
672 | * handled by irq_enter() and irq_exit().) | |
673 | * | |
c0da313e PM |
674 | * If you add or remove a call to rcu_idle_enter(), be sure to test with |
675 | * CONFIG_RCU_EQS_DEBUG=y. | |
adf5091e FW |
676 | */ |
677 | void rcu_idle_enter(void) | |
678 | { | |
b04db8e1 | 679 | lockdep_assert_irqs_disabled(); |
cb349ca9 | 680 | rcu_eqs_enter(false); |
adf5091e | 681 | } |
3ad1c8ef | 682 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
64db4cff | 683 | |
d1ec4c34 | 684 | #ifdef CONFIG_NO_HZ_FULL |
f8bb5cae | 685 | |
4ae7dc97 | 686 | #if !defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK) |
f8bb5cae FW |
687 | /* |
688 | * An empty function that will trigger a reschedule on | |
4ae7dc97 | 689 | * IRQ tail once IRQs get re-enabled on userspace/guest resume. |
f8bb5cae FW |
690 | */ |
691 | static void late_wakeup_func(struct irq_work *work) | |
692 | { | |
693 | } | |
694 | ||
695 | static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) = | |
696 | IRQ_WORK_INIT(late_wakeup_func); | |
697 | ||
4ae7dc97 FW |
698 | /* |
699 | * If either: | |
700 | * | |
701 | * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work | |
702 | * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry. | |
703 | * | |
704 | * In these cases the late RCU wake ups aren't supported in the resched loops and our | |
705 | * last resort is to fire a local irq_work that will trigger a reschedule once IRQs | |
706 | * get re-enabled again. | |
707 | */ | |
708 | noinstr static void rcu_irq_work_resched(void) | |
709 | { | |
710 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); | |
711 | ||
712 | if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU)) | |
713 | return; | |
714 | ||
715 | if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU)) | |
716 | return; | |
717 | ||
718 | instrumentation_begin(); | |
719 | if (do_nocb_deferred_wakeup(rdp) && need_resched()) { | |
720 | irq_work_queue(this_cpu_ptr(&late_wakeup_work)); | |
721 | } | |
722 | instrumentation_end(); | |
723 | } | |
724 | ||
725 | #else | |
726 | static inline void rcu_irq_work_resched(void) { } | |
727 | #endif | |
728 | ||
adf5091e FW |
729 | /** |
730 | * rcu_user_enter - inform RCU that we are resuming userspace. | |
731 | * | |
732 | * Enter RCU idle mode right before resuming userspace. No use of RCU | |
733 | * is permitted between this call and rcu_user_exit(). This way the | |
734 | * CPU doesn't need to maintain the tick for RCU maintenance purposes | |
735 | * when the CPU runs in userspace. | |
c0da313e PM |
736 | * |
737 | * If you add or remove a call to rcu_user_enter(), be sure to test with | |
738 | * CONFIG_RCU_EQS_DEBUG=y. | |
adf5091e | 739 | */ |
ff5c4f5c | 740 | noinstr void rcu_user_enter(void) |
adf5091e | 741 | { |
b04db8e1 | 742 | lockdep_assert_irqs_disabled(); |
54b7429e | 743 | |
f8bb5cae | 744 | /* |
47b8ff19 FW |
745 | * Other than generic entry implementation, we may be past the last |
746 | * rescheduling opportunity in the entry code. Trigger a self IPI | |
747 | * that will fire and reschedule once we resume in user/guest mode. | |
f8bb5cae | 748 | */ |
4ae7dc97 | 749 | rcu_irq_work_resched(); |
d4db30af | 750 | rcu_eqs_enter(true); |
adf5091e | 751 | } |
f8bb5cae | 752 | |
d1ec4c34 | 753 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 754 | |
9ea366f6 PM |
755 | /** |
756 | * rcu_nmi_exit - inform RCU of exit from NMI context | |
757 | * | |
fd581a91 | 758 | * If we are returning from the outermost NMI handler that interrupted an |
dc5a4f29 | 759 | * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting |
fd581a91 PM |
760 | * to let the RCU grace-period handling know that the CPU is back to |
761 | * being RCU-idle. | |
762 | * | |
9ea366f6 | 763 | * If you add or remove a call to rcu_nmi_exit(), be sure to test |
fd581a91 PM |
764 | * with CONFIG_RCU_EQS_DEBUG=y. |
765 | */ | |
9ea366f6 | 766 | noinstr void rcu_nmi_exit(void) |
fd581a91 | 767 | { |
4c5273bf | 768 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
fd581a91 | 769 | |
b58e733f | 770 | instrumentation_begin(); |
fd581a91 PM |
771 | /* |
772 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. | |
773 | * (We are exiting an NMI handler, so RCU better be paying attention | |
774 | * to us!) | |
775 | */ | |
4c5273bf | 776 | WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0); |
fd581a91 PM |
777 | WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs()); |
778 | ||
779 | /* | |
780 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so | |
781 | * leave it in non-RCU-idle state. | |
782 | */ | |
4c5273bf | 783 | if (rdp->dynticks_nmi_nesting != 1) { |
6cf539a8 ME |
784 | trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2, |
785 | atomic_read(&rdp->dynticks)); | |
4c5273bf PM |
786 | WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */ |
787 | rdp->dynticks_nmi_nesting - 2); | |
ff5c4f5c | 788 | instrumentation_end(); |
fd581a91 PM |
789 | return; |
790 | } | |
791 | ||
792 | /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ | |
6cf539a8 | 793 | trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks)); |
4c5273bf | 794 | WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */ |
cf7614e1 | 795 | |
9ea366f6 | 796 | if (!in_nmi()) |
cf7614e1 | 797 | rcu_prepare_for_idle(); |
b58e733f PZ |
798 | |
799 | // instrumentation for the noinstr rcu_dynticks_eqs_enter() | |
800 | instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks)); | |
ff5c4f5c | 801 | instrumentation_end(); |
cf7614e1 | 802 | |
ac3caf82 | 803 | // RCU is watching here ... |
fd581a91 | 804 | rcu_dynticks_eqs_enter(); |
ac3caf82 | 805 | // ... but is no longer watching here. |
cf7614e1 | 806 | |
9ea366f6 | 807 | if (!in_nmi()) |
cf7614e1 BP |
808 | rcu_dynticks_task_enter(); |
809 | } | |
810 | ||
9b2e4f18 PM |
811 | /** |
812 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle | |
813 | * | |
814 | * Exit from an interrupt handler, which might possibly result in entering | |
815 | * idle mode, in other words, leaving the mode in which read-side critical | |
7c9906ca | 816 | * sections can occur. The caller must have disabled interrupts. |
64db4cff | 817 | * |
9b2e4f18 PM |
818 | * This code assumes that the idle loop never does anything that might |
819 | * result in unbalanced calls to irq_enter() and irq_exit(). If your | |
58721f5d PM |
820 | * architecture's idle loop violates this assumption, RCU will give you what |
821 | * you deserve, good and hard. But very infrequently and irreproducibly. | |
9b2e4f18 PM |
822 | * |
823 | * Use things like work queues to work around this limitation. | |
824 | * | |
825 | * You have been warned. | |
c0da313e PM |
826 | * |
827 | * If you add or remove a call to rcu_irq_exit(), be sure to test with | |
828 | * CONFIG_RCU_EQS_DEBUG=y. | |
64db4cff | 829 | */ |
ff5c4f5c | 830 | void noinstr rcu_irq_exit(void) |
64db4cff | 831 | { |
b04db8e1 | 832 | lockdep_assert_irqs_disabled(); |
9ea366f6 | 833 | rcu_nmi_exit(); |
7c9906ca PM |
834 | } |
835 | ||
8ae0ae67 TG |
836 | /** |
837 | * rcu_irq_exit_preempt - Inform RCU that current CPU is exiting irq | |
838 | * towards in kernel preemption | |
839 | * | |
840 | * Same as rcu_irq_exit() but has a sanity check that scheduling is safe | |
841 | * from RCU point of view. Invoked from return from interrupt before kernel | |
842 | * preemption. | |
843 | */ | |
844 | void rcu_irq_exit_preempt(void) | |
845 | { | |
846 | lockdep_assert_irqs_disabled(); | |
847 | rcu_nmi_exit(); | |
848 | ||
849 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0, | |
850 | "RCU dynticks_nesting counter underflow/zero!"); | |
851 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) != | |
852 | DYNTICK_IRQ_NONIDLE, | |
853 | "Bad RCU dynticks_nmi_nesting counter\n"); | |
854 | RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(), | |
855 | "RCU in extended quiescent state!"); | |
856 | } | |
857 | ||
07325d4a TG |
858 | #ifdef CONFIG_PROVE_RCU |
859 | /** | |
860 | * rcu_irq_exit_check_preempt - Validate that scheduling is possible | |
861 | */ | |
862 | void rcu_irq_exit_check_preempt(void) | |
863 | { | |
864 | lockdep_assert_irqs_disabled(); | |
865 | ||
866 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0, | |
867 | "RCU dynticks_nesting counter underflow/zero!"); | |
868 | RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) != | |
869 | DYNTICK_IRQ_NONIDLE, | |
870 | "Bad RCU dynticks_nmi_nesting counter\n"); | |
871 | RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(), | |
872 | "RCU in extended quiescent state!"); | |
873 | } | |
874 | #endif /* #ifdef CONFIG_PROVE_RCU */ | |
875 | ||
7c9906ca PM |
876 | /* |
877 | * Wrapper for rcu_irq_exit() where interrupts are enabled. | |
c0da313e PM |
878 | * |
879 | * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test | |
880 | * with CONFIG_RCU_EQS_DEBUG=y. | |
7c9906ca PM |
881 | */ |
882 | void rcu_irq_exit_irqson(void) | |
883 | { | |
884 | unsigned long flags; | |
885 | ||
886 | local_irq_save(flags); | |
887 | rcu_irq_exit(); | |
9b2e4f18 PM |
888 | local_irq_restore(flags); |
889 | } | |
890 | ||
adf5091e FW |
891 | /* |
892 | * Exit an RCU extended quiescent state, which can be either the | |
893 | * idle loop or adaptive-tickless usermode execution. | |
51a1fd30 PM |
894 | * |
895 | * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to | |
896 | * allow for the possibility of usermode upcalls messing up our count of | |
897 | * interrupt nesting level during the busy period that is just now starting. | |
9b2e4f18 | 898 | */ |
ff5c4f5c | 899 | static void noinstr rcu_eqs_exit(bool user) |
9b2e4f18 | 900 | { |
4c5273bf | 901 | struct rcu_data *rdp; |
84585aa8 | 902 | long oldval; |
9b2e4f18 | 903 | |
b04db8e1 | 904 | lockdep_assert_irqs_disabled(); |
4c5273bf PM |
905 | rdp = this_cpu_ptr(&rcu_data); |
906 | oldval = rdp->dynticks_nesting; | |
1ce46ee5 | 907 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0); |
51a1fd30 | 908 | if (oldval) { |
ac3caf82 | 909 | // RCU was already watching, so just do accounting and leave. |
4c5273bf | 910 | rdp->dynticks_nesting++; |
9dd238e2 | 911 | return; |
3a592405 | 912 | } |
9dd238e2 | 913 | rcu_dynticks_task_exit(); |
ac3caf82 | 914 | // RCU is not watching here ... |
9dd238e2 | 915 | rcu_dynticks_eqs_exit(); |
ac3caf82 | 916 | // ... but is watching here. |
ff5c4f5c | 917 | instrumentation_begin(); |
b58e733f PZ |
918 | |
919 | // instrumentation for the noinstr rcu_dynticks_eqs_exit() | |
920 | instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks)); | |
921 | ||
9dd238e2 | 922 | rcu_cleanup_after_idle(); |
6cf539a8 | 923 | trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks)); |
e68bbb26 | 924 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current)); |
4c5273bf PM |
925 | WRITE_ONCE(rdp->dynticks_nesting, 1); |
926 | WARN_ON_ONCE(rdp->dynticks_nmi_nesting); | |
927 | WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE); | |
ff5c4f5c | 928 | instrumentation_end(); |
9b2e4f18 | 929 | } |
adf5091e FW |
930 | |
931 | /** | |
932 | * rcu_idle_exit - inform RCU that current CPU is leaving idle | |
933 | * | |
934 | * Exit idle mode, in other words, -enter- the mode in which RCU | |
935 | * read-side critical sections can occur. | |
936 | * | |
c0da313e PM |
937 | * If you add or remove a call to rcu_idle_exit(), be sure to test with |
938 | * CONFIG_RCU_EQS_DEBUG=y. | |
adf5091e FW |
939 | */ |
940 | void rcu_idle_exit(void) | |
941 | { | |
c5d900bf FW |
942 | unsigned long flags; |
943 | ||
944 | local_irq_save(flags); | |
cb349ca9 | 945 | rcu_eqs_exit(false); |
c5d900bf | 946 | local_irq_restore(flags); |
adf5091e | 947 | } |
3ad1c8ef | 948 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
9b2e4f18 | 949 | |
d1ec4c34 | 950 | #ifdef CONFIG_NO_HZ_FULL |
adf5091e FW |
951 | /** |
952 | * rcu_user_exit - inform RCU that we are exiting userspace. | |
953 | * | |
954 | * Exit RCU idle mode while entering the kernel because it can | |
955 | * run a RCU read side critical section anytime. | |
c0da313e PM |
956 | * |
957 | * If you add or remove a call to rcu_user_exit(), be sure to test with | |
958 | * CONFIG_RCU_EQS_DEBUG=y. | |
adf5091e | 959 | */ |
ff5c4f5c | 960 | void noinstr rcu_user_exit(void) |
adf5091e | 961 | { |
91d1aa43 | 962 | rcu_eqs_exit(1); |
adf5091e | 963 | } |
aaf2bc50 PM |
964 | |
965 | /** | |
966 | * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it. | |
967 | * | |
968 | * The scheduler tick is not normally enabled when CPUs enter the kernel | |
969 | * from nohz_full userspace execution. After all, nohz_full userspace | |
970 | * execution is an RCU quiescent state and the time executing in the kernel | |
971 | * is quite short. Except of course when it isn't. And it is not hard to | |
972 | * cause a large system to spend tens of seconds or even minutes looping | |
973 | * in the kernel, which can cause a number of problems, include RCU CPU | |
974 | * stall warnings. | |
975 | * | |
976 | * Therefore, if a nohz_full CPU fails to report a quiescent state | |
977 | * in a timely manner, the RCU grace-period kthread sets that CPU's | |
978 | * ->rcu_urgent_qs flag with the expectation that the next interrupt or | |
979 | * exception will invoke this function, which will turn on the scheduler | |
980 | * tick, which will enable RCU to detect that CPU's quiescent states, | |
981 | * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels. | |
982 | * The tick will be disabled once a quiescent state is reported for | |
983 | * this CPU. | |
984 | * | |
985 | * Of course, in carefully tuned systems, there might never be an | |
986 | * interrupt or exception. In that case, the RCU grace-period kthread | |
987 | * will eventually cause one to happen. However, in less carefully | |
988 | * controlled environments, this function allows RCU to get what it | |
989 | * needs without creating otherwise useless interruptions. | |
990 | */ | |
991 | void __rcu_irq_enter_check_tick(void) | |
992 | { | |
993 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); | |
994 | ||
6dbce04d PZ |
995 | // If we're here from NMI there's nothing to do. |
996 | if (in_nmi()) | |
aaf2bc50 PM |
997 | return; |
998 | ||
999 | RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(), | |
1000 | "Illegal rcu_irq_enter_check_tick() from extended quiescent state"); | |
1001 | ||
1002 | if (!tick_nohz_full_cpu(rdp->cpu) || | |
1003 | !READ_ONCE(rdp->rcu_urgent_qs) || | |
1004 | READ_ONCE(rdp->rcu_forced_tick)) { | |
1005 | // RCU doesn't need nohz_full help from this CPU, or it is | |
1006 | // already getting that help. | |
1007 | return; | |
1008 | } | |
1009 | ||
1010 | // We get here only when not in an extended quiescent state and | |
1011 | // from interrupts (as opposed to NMIs). Therefore, (1) RCU is | |
1012 | // already watching and (2) The fact that we are in an interrupt | |
1013 | // handler and that the rcu_node lock is an irq-disabled lock | |
1014 | // prevents self-deadlock. So we can safely recheck under the lock. | |
1015 | // Note that the nohz_full state currently cannot change. | |
1016 | raw_spin_lock_rcu_node(rdp->mynode); | |
1017 | if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) { | |
1018 | // A nohz_full CPU is in the kernel and RCU needs a | |
1019 | // quiescent state. Turn on the tick! | |
1020 | WRITE_ONCE(rdp->rcu_forced_tick, true); | |
1021 | tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU); | |
1022 | } | |
1023 | raw_spin_unlock_rcu_node(rdp->mynode); | |
1024 | } | |
d1ec4c34 | 1025 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 1026 | |
64db4cff | 1027 | /** |
9ea366f6 | 1028 | * rcu_nmi_enter - inform RCU of entry to NMI context |
64db4cff | 1029 | * |
dc5a4f29 | 1030 | * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and |
4c5273bf | 1031 | * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know |
734d1680 PM |
1032 | * that the CPU is active. This implementation permits nested NMIs, as |
1033 | * long as the nesting level does not overflow an int. (You will probably | |
1034 | * run out of stack space first.) | |
c0da313e | 1035 | * |
9ea366f6 | 1036 | * If you add or remove a call to rcu_nmi_enter(), be sure to test |
c0da313e | 1037 | * with CONFIG_RCU_EQS_DEBUG=y. |
64db4cff | 1038 | */ |
9ea366f6 | 1039 | noinstr void rcu_nmi_enter(void) |
64db4cff | 1040 | { |
84585aa8 | 1041 | long incby = 2; |
5b14557b | 1042 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
64db4cff | 1043 | |
734d1680 | 1044 | /* Complain about underflow. */ |
4c5273bf | 1045 | WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0); |
734d1680 PM |
1046 | |
1047 | /* | |
1048 | * If idle from RCU viewpoint, atomically increment ->dynticks | |
1049 | * to mark non-idle and increment ->dynticks_nmi_nesting by one. | |
1050 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means | |
1051 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed | |
1052 | * to be in the outermost NMI handler that interrupted an RCU-idle | |
1053 | * period (observation due to Andy Lutomirski). | |
1054 | */ | |
02a5c550 | 1055 | if (rcu_dynticks_curr_cpu_in_eqs()) { |
cf7614e1 | 1056 | |
9ea366f6 | 1057 | if (!in_nmi()) |
cf7614e1 BP |
1058 | rcu_dynticks_task_exit(); |
1059 | ||
ac3caf82 | 1060 | // RCU is not watching here ... |
2625d469 | 1061 | rcu_dynticks_eqs_exit(); |
ac3caf82 | 1062 | // ... but is watching here. |
cf7614e1 | 1063 | |
04b25a49 PM |
1064 | if (!in_nmi()) { |
1065 | instrumentation_begin(); | |
cf7614e1 | 1066 | rcu_cleanup_after_idle(); |
04b25a49 PM |
1067 | instrumentation_end(); |
1068 | } | |
cf7614e1 | 1069 | |
b58e733f PZ |
1070 | instrumentation_begin(); |
1071 | // instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs() | |
1072 | instrument_atomic_read(&rdp->dynticks, sizeof(rdp->dynticks)); | |
1073 | // instrumentation for the noinstr rcu_dynticks_eqs_exit() | |
1074 | instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks)); | |
1075 | ||
734d1680 | 1076 | incby = 1; |
9ea366f6 | 1077 | } else if (!in_nmi()) { |
ff5c4f5c | 1078 | instrumentation_begin(); |
aaf2bc50 | 1079 | rcu_irq_enter_check_tick(); |
b58e733f PZ |
1080 | } else { |
1081 | instrumentation_begin(); | |
734d1680 | 1082 | } |
b58e733f | 1083 | |
bd2b879a | 1084 | trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="), |
4c5273bf | 1085 | rdp->dynticks_nmi_nesting, |
6cf539a8 | 1086 | rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks)); |
ff5c4f5c | 1087 | instrumentation_end(); |
4c5273bf PM |
1088 | WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */ |
1089 | rdp->dynticks_nmi_nesting + incby); | |
734d1680 | 1090 | barrier(); |
64db4cff PM |
1091 | } |
1092 | ||
1093 | /** | |
9b2e4f18 | 1094 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle |
64db4cff | 1095 | * |
9b2e4f18 PM |
1096 | * Enter an interrupt handler, which might possibly result in exiting |
1097 | * idle mode, in other words, entering the mode in which read-side critical | |
7c9906ca | 1098 | * sections can occur. The caller must have disabled interrupts. |
c0da313e | 1099 | * |
9b2e4f18 | 1100 | * Note that the Linux kernel is fully capable of entering an interrupt |
58721f5d PM |
1101 | * handler that it never exits, for example when doing upcalls to user mode! |
1102 | * This code assumes that the idle loop never does upcalls to user mode. | |
1103 | * If your architecture's idle loop does do upcalls to user mode (or does | |
1104 | * anything else that results in unbalanced calls to the irq_enter() and | |
1105 | * irq_exit() functions), RCU will give you what you deserve, good and hard. | |
1106 | * But very infrequently and irreproducibly. | |
9b2e4f18 PM |
1107 | * |
1108 | * Use things like work queues to work around this limitation. | |
1109 | * | |
1110 | * You have been warned. | |
c0da313e PM |
1111 | * |
1112 | * If you add or remove a call to rcu_irq_enter(), be sure to test with | |
1113 | * CONFIG_RCU_EQS_DEBUG=y. | |
64db4cff | 1114 | */ |
ff5c4f5c | 1115 | noinstr void rcu_irq_enter(void) |
64db4cff | 1116 | { |
b04db8e1 | 1117 | lockdep_assert_irqs_disabled(); |
9ea366f6 | 1118 | rcu_nmi_enter(); |
7c9906ca | 1119 | } |
734d1680 | 1120 | |
7c9906ca PM |
1121 | /* |
1122 | * Wrapper for rcu_irq_enter() where interrupts are enabled. | |
c0da313e PM |
1123 | * |
1124 | * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test | |
1125 | * with CONFIG_RCU_EQS_DEBUG=y. | |
7c9906ca PM |
1126 | */ |
1127 | void rcu_irq_enter_irqson(void) | |
1128 | { | |
1129 | unsigned long flags; | |
734d1680 | 1130 | |
7c9906ca PM |
1131 | local_irq_save(flags); |
1132 | rcu_irq_enter(); | |
64db4cff | 1133 | local_irq_restore(flags); |
64db4cff PM |
1134 | } |
1135 | ||
66e4c33b | 1136 | /* |
516e5ae0 JFG |
1137 | * If any sort of urgency was applied to the current CPU (for example, |
1138 | * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order | |
1139 | * to get to a quiescent state, disable it. | |
66e4c33b | 1140 | */ |
516e5ae0 | 1141 | static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp) |
66e4c33b | 1142 | { |
5b14557b | 1143 | raw_lockdep_assert_held_rcu_node(rdp->mynode); |
516e5ae0 JFG |
1144 | WRITE_ONCE(rdp->rcu_urgent_qs, false); |
1145 | WRITE_ONCE(rdp->rcu_need_heavy_qs, false); | |
66e4c33b PM |
1146 | if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) { |
1147 | tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU); | |
2a2ae872 | 1148 | WRITE_ONCE(rdp->rcu_forced_tick, false); |
66e4c33b PM |
1149 | } |
1150 | } | |
1151 | ||
5c173eb8 | 1152 | /** |
2320bda2 | 1153 | * rcu_is_watching - see if RCU thinks that the current CPU is not idle |
64db4cff | 1154 | * |
791875d1 PM |
1155 | * Return true if RCU is watching the running CPU, which means that this |
1156 | * CPU can safely enter RCU read-side critical sections. In other words, | |
2320bda2 ZZ |
1157 | * if the current CPU is not in its idle loop or is in an interrupt or |
1158 | * NMI handler, return true. | |
d2098b44 PZ |
1159 | * |
1160 | * Make notrace because it can be called by the internal functions of | |
1161 | * ftrace, and making this notrace removes unnecessary recursion calls. | |
64db4cff | 1162 | */ |
d2098b44 | 1163 | notrace bool rcu_is_watching(void) |
64db4cff | 1164 | { |
f534ed1f | 1165 | bool ret; |
34240697 | 1166 | |
46f00d18 | 1167 | preempt_disable_notrace(); |
791875d1 | 1168 | ret = !rcu_dynticks_curr_cpu_in_eqs(); |
46f00d18 | 1169 | preempt_enable_notrace(); |
34240697 | 1170 | return ret; |
64db4cff | 1171 | } |
5c173eb8 | 1172 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 1173 | |
bcbfdd01 PM |
1174 | /* |
1175 | * If a holdout task is actually running, request an urgent quiescent | |
1176 | * state from its CPU. This is unsynchronized, so migrations can cause | |
1177 | * the request to go to the wrong CPU. Which is OK, all that will happen | |
1178 | * is that the CPU's next context switch will be a bit slower and next | |
1179 | * time around this task will generate another request. | |
1180 | */ | |
1181 | void rcu_request_urgent_qs_task(struct task_struct *t) | |
1182 | { | |
1183 | int cpu; | |
1184 | ||
1185 | barrier(); | |
1186 | cpu = task_cpu(t); | |
1187 | if (!task_curr(t)) | |
1188 | return; /* This task is not running on that CPU. */ | |
2dba13f0 | 1189 | smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true); |
bcbfdd01 PM |
1190 | } |
1191 | ||
62fde6ed | 1192 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
c0d6d01b PM |
1193 | |
1194 | /* | |
5554788e | 1195 | * Is the current CPU online as far as RCU is concerned? |
2036d94a | 1196 | * |
5554788e PM |
1197 | * Disable preemption to avoid false positives that could otherwise |
1198 | * happen due to the current CPU number being sampled, this task being | |
1199 | * preempted, its old CPU being taken offline, resuming on some other CPU, | |
49918a54 | 1200 | * then determining that its old CPU is now offline. |
c0d6d01b | 1201 | * |
5554788e PM |
1202 | * Disable checking if in an NMI handler because we cannot safely |
1203 | * report errors from NMI handlers anyway. In addition, it is OK to use | |
1204 | * RCU on an offline processor during initial boot, hence the check for | |
1205 | * rcu_scheduler_fully_active. | |
c0d6d01b PM |
1206 | */ |
1207 | bool rcu_lockdep_current_cpu_online(void) | |
1208 | { | |
2036d94a PM |
1209 | struct rcu_data *rdp; |
1210 | struct rcu_node *rnp; | |
b97d23c5 | 1211 | bool ret = false; |
c0d6d01b | 1212 | |
5554788e | 1213 | if (in_nmi() || !rcu_scheduler_fully_active) |
f6f7ee9a | 1214 | return true; |
ff5c4f5c | 1215 | preempt_disable_notrace(); |
b97d23c5 PM |
1216 | rdp = this_cpu_ptr(&rcu_data); |
1217 | rnp = rdp->mynode; | |
4d60b475 | 1218 | if (rdp->grpmask & rcu_rnp_online_cpus(rnp) || READ_ONCE(rnp->ofl_seq) & 0x1) |
b97d23c5 | 1219 | ret = true; |
ff5c4f5c | 1220 | preempt_enable_notrace(); |
b97d23c5 | 1221 | return ret; |
c0d6d01b PM |
1222 | } |
1223 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
1224 | ||
62fde6ed | 1225 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
9b2e4f18 | 1226 | |
9b9500da PM |
1227 | /* |
1228 | * We are reporting a quiescent state on behalf of some other CPU, so | |
1229 | * it is our responsibility to check for and handle potential overflow | |
a66ae8ae | 1230 | * of the rcu_node ->gp_seq counter with respect to the rcu_data counters. |
9b9500da PM |
1231 | * After all, the CPU might be in deep idle state, and thus executing no |
1232 | * code whatsoever. | |
1233 | */ | |
1234 | static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp) | |
1235 | { | |
a32e01ee | 1236 | raw_lockdep_assert_held_rcu_node(rnp); |
a66ae8ae PM |
1237 | if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4, |
1238 | rnp->gp_seq)) | |
9b9500da | 1239 | WRITE_ONCE(rdp->gpwrap, true); |
8aa670cd PM |
1240 | if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq)) |
1241 | rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4; | |
9b9500da PM |
1242 | } |
1243 | ||
64db4cff PM |
1244 | /* |
1245 | * Snapshot the specified CPU's dynticks counter so that we can later | |
1246 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 1247 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 1248 | */ |
fe5ac724 | 1249 | static int dyntick_save_progress_counter(struct rcu_data *rdp) |
64db4cff | 1250 | { |
dc5a4f29 | 1251 | rdp->dynticks_snap = rcu_dynticks_snap(rdp); |
02a5c550 | 1252 | if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) { |
88d1bead | 1253 | trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti")); |
9b9500da | 1254 | rcu_gpnum_ovf(rdp->mynode, rdp); |
23a9bacd | 1255 | return 1; |
7941dbde | 1256 | } |
23a9bacd | 1257 | return 0; |
64db4cff PM |
1258 | } |
1259 | ||
1260 | /* | |
1261 | * Return true if the specified CPU has passed through a quiescent | |
1262 | * state by virtue of being in or having passed through an dynticks | |
1263 | * idle state since the last call to dyntick_save_progress_counter() | |
a82dcc76 | 1264 | * for this same CPU, or by virtue of having been offline. |
64db4cff | 1265 | */ |
fe5ac724 | 1266 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp) |
64db4cff | 1267 | { |
3a19b46a | 1268 | unsigned long jtsq; |
0f9be8ca | 1269 | bool *rnhqp; |
9226b10d | 1270 | bool *ruqp; |
9b9500da | 1271 | struct rcu_node *rnp = rdp->mynode; |
64db4cff PM |
1272 | |
1273 | /* | |
1274 | * If the CPU passed through or entered a dynticks idle phase with | |
1275 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
1276 | * already acknowledged the request to pass through a quiescent | |
1277 | * state. Either way, that CPU cannot possibly be in an RCU | |
1278 | * read-side critical section that started before the beginning | |
1279 | * of the current RCU grace period. | |
1280 | */ | |
dc5a4f29 | 1281 | if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) { |
88d1bead | 1282 | trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti")); |
9b9500da | 1283 | rcu_gpnum_ovf(rnp, rdp); |
3a19b46a PM |
1284 | return 1; |
1285 | } | |
1286 | ||
666ca290 JFG |
1287 | /* |
1288 | * Complain if a CPU that is considered to be offline from RCU's | |
1289 | * perspective has not yet reported a quiescent state. After all, | |
1290 | * the offline CPU should have reported a quiescent state during | |
1291 | * the CPU-offline process, or, failing that, by rcu_gp_init() | |
1292 | * if it ran concurrently with either the CPU going offline or the | |
1293 | * last task on a leaf rcu_node structure exiting its RCU read-side | |
1294 | * critical section while all CPUs corresponding to that structure | |
1295 | * are offline. This added warning detects bugs in any of these | |
1296 | * code paths. | |
1297 | * | |
1298 | * The rcu_node structure's ->lock is held here, which excludes | |
1299 | * the relevant portions the CPU-hotplug code, the grace-period | |
1300 | * initialization code, and the rcu_read_unlock() code paths. | |
1301 | * | |
1302 | * For more detail, please refer to the "Hotplug CPU" section | |
1303 | * of RCU's Requirements documentation. | |
1304 | */ | |
1305 | if (WARN_ON_ONCE(!(rdp->grpmask & rcu_rnp_online_cpus(rnp)))) { | |
f2e2df59 PM |
1306 | bool onl; |
1307 | struct rcu_node *rnp1; | |
1308 | ||
f2e2df59 PM |
1309 | pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n", |
1310 | __func__, rnp->grplo, rnp->grphi, rnp->level, | |
1311 | (long)rnp->gp_seq, (long)rnp->completedqs); | |
1312 | for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent) | |
1313 | pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n", | |
1314 | __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask); | |
1315 | onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp)); | |
1316 | pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n", | |
1317 | __func__, rdp->cpu, ".o"[onl], | |
1318 | (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags, | |
1319 | (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags); | |
1320 | return 1; /* Break things loose after complaining. */ | |
1321 | } | |
1322 | ||
65d798f0 | 1323 | /* |
4a81e832 | 1324 | * A CPU running for an extended time within the kernel can |
c06aed0e PM |
1325 | * delay RCU grace periods: (1) At age jiffies_to_sched_qs, |
1326 | * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set | |
7e28c5af PM |
1327 | * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the |
1328 | * unsynchronized assignments to the per-CPU rcu_need_heavy_qs | |
1329 | * variable are safe because the assignments are repeated if this | |
1330 | * CPU failed to pass through a quiescent state. This code | |
c06aed0e | 1331 | * also checks .jiffies_resched in case jiffies_to_sched_qs |
7e28c5af | 1332 | * is set way high. |
6193c76a | 1333 | */ |
c06aed0e | 1334 | jtsq = READ_ONCE(jiffies_to_sched_qs); |
2dba13f0 PM |
1335 | ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu); |
1336 | rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu); | |
0f9be8ca | 1337 | if (!READ_ONCE(*rnhqp) && |
7e28c5af | 1338 | (time_after(jiffies, rcu_state.gp_start + jtsq * 2) || |
b2b00ddf PM |
1339 | time_after(jiffies, rcu_state.jiffies_resched) || |
1340 | rcu_state.cbovld)) { | |
0f9be8ca | 1341 | WRITE_ONCE(*rnhqp, true); |
9226b10d PM |
1342 | /* Store rcu_need_heavy_qs before rcu_urgent_qs. */ |
1343 | smp_store_release(ruqp, true); | |
7e28c5af PM |
1344 | } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) { |
1345 | WRITE_ONCE(*ruqp, true); | |
6193c76a PM |
1346 | } |
1347 | ||
28053bc7 | 1348 | /* |
c98cac60 | 1349 | * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq! |
d3052109 PM |
1350 | * The above code handles this, but only for straight cond_resched(). |
1351 | * And some in-kernel loops check need_resched() before calling | |
1352 | * cond_resched(), which defeats the above code for CPUs that are | |
1353 | * running in-kernel with scheduling-clock interrupts disabled. | |
1354 | * So hit them over the head with the resched_cpu() hammer! | |
28053bc7 | 1355 | */ |
d3052109 | 1356 | if (tick_nohz_full_cpu(rdp->cpu) && |
b2b00ddf PM |
1357 | (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) || |
1358 | rcu_state.cbovld)) { | |
05ef9e9e | 1359 | WRITE_ONCE(*ruqp, true); |
28053bc7 | 1360 | resched_cpu(rdp->cpu); |
d3052109 PM |
1361 | WRITE_ONCE(rdp->last_fqs_resched, jiffies); |
1362 | } | |
1363 | ||
1364 | /* | |
1365 | * If more than halfway to RCU CPU stall-warning time, invoke | |
1366 | * resched_cpu() more frequently to try to loosen things up a bit. | |
1367 | * Also check to see if the CPU is getting hammered with interrupts, | |
1368 | * but only once per grace period, just to keep the IPIs down to | |
1369 | * a dull roar. | |
1370 | */ | |
1371 | if (time_after(jiffies, rcu_state.jiffies_resched)) { | |
1372 | if (time_after(jiffies, | |
1373 | READ_ONCE(rdp->last_fqs_resched) + jtsq)) { | |
1374 | resched_cpu(rdp->cpu); | |
1375 | WRITE_ONCE(rdp->last_fqs_resched, jiffies); | |
1376 | } | |
9b9500da | 1377 | if (IS_ENABLED(CONFIG_IRQ_WORK) && |
8aa670cd | 1378 | !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq && |
9b9500da | 1379 | (rnp->ffmask & rdp->grpmask)) { |
9b9500da | 1380 | rdp->rcu_iw_pending = true; |
8aa670cd | 1381 | rdp->rcu_iw_gp_seq = rnp->gp_seq; |
9b9500da PM |
1382 | irq_work_queue_on(&rdp->rcu_iw, rdp->cpu); |
1383 | } | |
1384 | } | |
4914950a | 1385 | |
a82dcc76 | 1386 | return 0; |
64db4cff PM |
1387 | } |
1388 | ||
41e80595 PM |
1389 | /* Trace-event wrapper function for trace_rcu_future_grace_period. */ |
1390 | static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
b73de91d | 1391 | unsigned long gp_seq_req, const char *s) |
0446be48 | 1392 | { |
0937d045 PM |
1393 | trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq), |
1394 | gp_seq_req, rnp->level, | |
1395 | rnp->grplo, rnp->grphi, s); | |
0446be48 PM |
1396 | } |
1397 | ||
1398 | /* | |
b73de91d | 1399 | * rcu_start_this_gp - Request the start of a particular grace period |
df2bf8f7 | 1400 | * @rnp_start: The leaf node of the CPU from which to start. |
b73de91d JF |
1401 | * @rdp: The rcu_data corresponding to the CPU from which to start. |
1402 | * @gp_seq_req: The gp_seq of the grace period to start. | |
1403 | * | |
41e80595 | 1404 | * Start the specified grace period, as needed to handle newly arrived |
0446be48 | 1405 | * callbacks. The required future grace periods are recorded in each |
7a1d0f23 | 1406 | * rcu_node structure's ->gp_seq_needed field. Returns true if there |
48a7639c | 1407 | * is reason to awaken the grace-period kthread. |
0446be48 | 1408 | * |
d5cd9685 PM |
1409 | * The caller must hold the specified rcu_node structure's ->lock, which |
1410 | * is why the caller is responsible for waking the grace-period kthread. | |
b73de91d JF |
1411 | * |
1412 | * Returns true if the GP thread needs to be awakened else false. | |
0446be48 | 1413 | */ |
df2bf8f7 | 1414 | static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp, |
b73de91d | 1415 | unsigned long gp_seq_req) |
0446be48 | 1416 | { |
48a7639c | 1417 | bool ret = false; |
df2bf8f7 | 1418 | struct rcu_node *rnp; |
0446be48 PM |
1419 | |
1420 | /* | |
360e0da6 PM |
1421 | * Use funnel locking to either acquire the root rcu_node |
1422 | * structure's lock or bail out if the need for this grace period | |
df2bf8f7 JFG |
1423 | * has already been recorded -- or if that grace period has in |
1424 | * fact already started. If there is already a grace period in | |
1425 | * progress in a non-leaf node, no recording is needed because the | |
1426 | * end of the grace period will scan the leaf rcu_node structures. | |
1427 | * Note that rnp_start->lock must not be released. | |
0446be48 | 1428 | */ |
df2bf8f7 JFG |
1429 | raw_lockdep_assert_held_rcu_node(rnp_start); |
1430 | trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf")); | |
1431 | for (rnp = rnp_start; 1; rnp = rnp->parent) { | |
1432 | if (rnp != rnp_start) | |
1433 | raw_spin_lock_rcu_node(rnp); | |
1434 | if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) || | |
1435 | rcu_seq_started(&rnp->gp_seq, gp_seq_req) || | |
1436 | (rnp != rnp_start && | |
1437 | rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) { | |
1438 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, | |
b73de91d | 1439 | TPS("Prestarted")); |
360e0da6 PM |
1440 | goto unlock_out; |
1441 | } | |
8ff37290 | 1442 | WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req); |
226ca5e7 | 1443 | if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) { |
a2165e41 | 1444 | /* |
226ca5e7 JFG |
1445 | * We just marked the leaf or internal node, and a |
1446 | * grace period is in progress, which means that | |
1447 | * rcu_gp_cleanup() will see the marking. Bail to | |
1448 | * reduce contention. | |
a2165e41 | 1449 | */ |
df2bf8f7 | 1450 | trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, |
b73de91d | 1451 | TPS("Startedleaf")); |
a2165e41 PM |
1452 | goto unlock_out; |
1453 | } | |
df2bf8f7 JFG |
1454 | if (rnp != rnp_start && rnp->parent != NULL) |
1455 | raw_spin_unlock_rcu_node(rnp); | |
1456 | if (!rnp->parent) | |
360e0da6 | 1457 | break; /* At root, and perhaps also leaf. */ |
0446be48 PM |
1458 | } |
1459 | ||
360e0da6 | 1460 | /* If GP already in progress, just leave, otherwise start one. */ |
de8e8730 | 1461 | if (rcu_gp_in_progress()) { |
df2bf8f7 | 1462 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot")); |
0446be48 PM |
1463 | goto unlock_out; |
1464 | } | |
df2bf8f7 | 1465 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot")); |
9cbc5b97 | 1466 | WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT); |
2906d215 | 1467 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); |
5648d659 | 1468 | if (!READ_ONCE(rcu_state.gp_kthread)) { |
df2bf8f7 | 1469 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread")); |
360e0da6 | 1470 | goto unlock_out; |
0446be48 | 1471 | } |
62ae1951 | 1472 | trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq")); |
360e0da6 | 1473 | ret = true; /* Caller must wake GP kthread. */ |
0446be48 | 1474 | unlock_out: |
ab5e869c | 1475 | /* Push furthest requested GP to leaf node and rcu_data structure. */ |
df2bf8f7 | 1476 | if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) { |
8ff37290 PM |
1477 | WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed); |
1478 | WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed); | |
ab5e869c | 1479 | } |
df2bf8f7 JFG |
1480 | if (rnp != rnp_start) |
1481 | raw_spin_unlock_rcu_node(rnp); | |
48a7639c | 1482 | return ret; |
0446be48 PM |
1483 | } |
1484 | ||
1485 | /* | |
1486 | * Clean up any old requests for the just-ended grace period. Also return | |
d1e4f01d | 1487 | * whether any additional grace periods have been requested. |
0446be48 | 1488 | */ |
3481f2ea | 1489 | static bool rcu_future_gp_cleanup(struct rcu_node *rnp) |
0446be48 | 1490 | { |
fb31340f | 1491 | bool needmore; |
da1df50d | 1492 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
0446be48 | 1493 | |
7a1d0f23 PM |
1494 | needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed); |
1495 | if (!needmore) | |
1496 | rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */ | |
b73de91d | 1497 | trace_rcu_this_gp(rnp, rdp, rnp->gp_seq, |
41e80595 | 1498 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); |
0446be48 PM |
1499 | return needmore; |
1500 | } | |
1501 | ||
48a7639c | 1502 | /* |
5648d659 PM |
1503 | * Awaken the grace-period kthread. Don't do a self-awaken (unless in an |
1504 | * interrupt or softirq handler, in which case we just might immediately | |
1505 | * sleep upon return, resulting in a grace-period hang), and don't bother | |
1506 | * awakening when there is nothing for the grace-period kthread to do | |
1507 | * (as in several CPUs raced to awaken, we lost), and finally don't try | |
1508 | * to awaken a kthread that has not yet been created. If all those checks | |
1509 | * are passed, track some debug information and awaken. | |
1d1f898d ZJ |
1510 | * |
1511 | * So why do the self-wakeup when in an interrupt or softirq handler | |
1512 | * in the grace-period kthread's context? Because the kthread might have | |
1513 | * been interrupted just as it was going to sleep, and just after the final | |
1514 | * pre-sleep check of the awaken condition. In this case, a wakeup really | |
1515 | * is required, and is therefore supplied. | |
48a7639c | 1516 | */ |
532c00c9 | 1517 | static void rcu_gp_kthread_wake(void) |
48a7639c | 1518 | { |
5648d659 PM |
1519 | struct task_struct *t = READ_ONCE(rcu_state.gp_kthread); |
1520 | ||
1521 | if ((current == t && !in_irq() && !in_serving_softirq()) || | |
1522 | !READ_ONCE(rcu_state.gp_flags) || !t) | |
48a7639c | 1523 | return; |
fd897573 PM |
1524 | WRITE_ONCE(rcu_state.gp_wake_time, jiffies); |
1525 | WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq)); | |
532c00c9 | 1526 | swake_up_one(&rcu_state.gp_wq); |
48a7639c PM |
1527 | } |
1528 | ||
dc35c893 | 1529 | /* |
29365e56 PM |
1530 | * If there is room, assign a ->gp_seq number to any callbacks on this |
1531 | * CPU that have not already been assigned. Also accelerate any callbacks | |
1532 | * that were previously assigned a ->gp_seq number that has since proven | |
1533 | * to be too conservative, which can happen if callbacks get assigned a | |
1534 | * ->gp_seq number while RCU is idle, but with reference to a non-root | |
1535 | * rcu_node structure. This function is idempotent, so it does not hurt | |
1536 | * to call it repeatedly. Returns an flag saying that we should awaken | |
1537 | * the RCU grace-period kthread. | |
dc35c893 PM |
1538 | * |
1539 | * The caller must hold rnp->lock with interrupts disabled. | |
1540 | */ | |
02f50142 | 1541 | static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp) |
dc35c893 | 1542 | { |
b73de91d | 1543 | unsigned long gp_seq_req; |
15fecf89 | 1544 | bool ret = false; |
dc35c893 | 1545 | |
d1b222c6 | 1546 | rcu_lockdep_assert_cblist_protected(rdp); |
a32e01ee | 1547 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1548 | |
15fecf89 PM |
1549 | /* If no pending (not yet ready to invoke) callbacks, nothing to do. */ |
1550 | if (!rcu_segcblist_pend_cbs(&rdp->cblist)) | |
48a7639c | 1551 | return false; |
dc35c893 | 1552 | |
3afe7fa5 JFG |
1553 | trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc")); |
1554 | ||
dc35c893 | 1555 | /* |
15fecf89 PM |
1556 | * Callbacks are often registered with incomplete grace-period |
1557 | * information. Something about the fact that getting exact | |
1558 | * information requires acquiring a global lock... RCU therefore | |
1559 | * makes a conservative estimate of the grace period number at which | |
1560 | * a given callback will become ready to invoke. The following | |
1561 | * code checks this estimate and improves it when possible, thus | |
1562 | * accelerating callback invocation to an earlier grace-period | |
1563 | * number. | |
dc35c893 | 1564 | */ |
9cbc5b97 | 1565 | gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq); |
b73de91d JF |
1566 | if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req)) |
1567 | ret = rcu_start_this_gp(rnp, rdp, gp_seq_req); | |
6d4b418c PM |
1568 | |
1569 | /* Trace depending on how much we were able to accelerate. */ | |
15fecf89 | 1570 | if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL)) |
a7886e89 | 1571 | trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB")); |
6d4b418c | 1572 | else |
a7886e89 JFG |
1573 | trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB")); |
1574 | ||
3afe7fa5 JFG |
1575 | trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc")); |
1576 | ||
48a7639c | 1577 | return ret; |
dc35c893 PM |
1578 | } |
1579 | ||
e44e73ca PM |
1580 | /* |
1581 | * Similar to rcu_accelerate_cbs(), but does not require that the leaf | |
1582 | * rcu_node structure's ->lock be held. It consults the cached value | |
1583 | * of ->gp_seq_needed in the rcu_data structure, and if that indicates | |
1584 | * that a new grace-period request be made, invokes rcu_accelerate_cbs() | |
1585 | * while holding the leaf rcu_node structure's ->lock. | |
1586 | */ | |
c6e09b97 | 1587 | static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp, |
e44e73ca PM |
1588 | struct rcu_data *rdp) |
1589 | { | |
1590 | unsigned long c; | |
1591 | bool needwake; | |
1592 | ||
d1b222c6 | 1593 | rcu_lockdep_assert_cblist_protected(rdp); |
c6e09b97 | 1594 | c = rcu_seq_snap(&rcu_state.gp_seq); |
a5b89501 | 1595 | if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) { |
e44e73ca PM |
1596 | /* Old request still live, so mark recent callbacks. */ |
1597 | (void)rcu_segcblist_accelerate(&rdp->cblist, c); | |
1598 | return; | |
1599 | } | |
1600 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ | |
02f50142 | 1601 | needwake = rcu_accelerate_cbs(rnp, rdp); |
e44e73ca PM |
1602 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
1603 | if (needwake) | |
532c00c9 | 1604 | rcu_gp_kthread_wake(); |
e44e73ca PM |
1605 | } |
1606 | ||
dc35c893 PM |
1607 | /* |
1608 | * Move any callbacks whose grace period has completed to the | |
1609 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
29365e56 | 1610 | * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL |
dc35c893 PM |
1611 | * sublist. This function is idempotent, so it does not hurt to |
1612 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1613 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1614 | * |
1615 | * The caller must hold rnp->lock with interrupts disabled. | |
1616 | */ | |
834f56bf | 1617 | static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp) |
dc35c893 | 1618 | { |
d1b222c6 | 1619 | rcu_lockdep_assert_cblist_protected(rdp); |
a32e01ee | 1620 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1621 | |
15fecf89 PM |
1622 | /* If no pending (not yet ready to invoke) callbacks, nothing to do. */ |
1623 | if (!rcu_segcblist_pend_cbs(&rdp->cblist)) | |
48a7639c | 1624 | return false; |
dc35c893 PM |
1625 | |
1626 | /* | |
29365e56 | 1627 | * Find all callbacks whose ->gp_seq numbers indicate that they |
dc35c893 PM |
1628 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. |
1629 | */ | |
29365e56 | 1630 | rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq); |
dc35c893 PM |
1631 | |
1632 | /* Classify any remaining callbacks. */ | |
02f50142 | 1633 | return rcu_accelerate_cbs(rnp, rdp); |
dc35c893 PM |
1634 | } |
1635 | ||
7f36ef82 PM |
1636 | /* |
1637 | * Move and classify callbacks, but only if doing so won't require | |
1638 | * that the RCU grace-period kthread be awakened. | |
1639 | */ | |
1640 | static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp, | |
1641 | struct rcu_data *rdp) | |
1642 | { | |
d1b222c6 | 1643 | rcu_lockdep_assert_cblist_protected(rdp); |
6608c3a0 PM |
1644 | if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || |
1645 | !raw_spin_trylock_rcu_node(rnp)) | |
7f36ef82 PM |
1646 | return; |
1647 | WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp)); | |
6608c3a0 | 1648 | raw_spin_unlock_rcu_node(rnp); |
7f36ef82 PM |
1649 | } |
1650 | ||
1a2f5d57 PM |
1651 | /* |
1652 | * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a | |
1653 | * quiescent state. This is intended to be invoked when the CPU notices | |
1654 | * a new grace period. | |
1655 | */ | |
1656 | static void rcu_strict_gp_check_qs(void) | |
1657 | { | |
1658 | if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) { | |
1659 | rcu_read_lock(); | |
1660 | rcu_read_unlock(); | |
1661 | } | |
1662 | } | |
1663 | ||
d09b62df | 1664 | /* |
ba9fbe95 PM |
1665 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1666 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1667 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1668 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1669 | */ |
c7e48f7b | 1670 | static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp) |
d09b62df | 1671 | { |
5d6742b3 | 1672 | bool ret = false; |
b5ea0370 | 1673 | bool need_qs; |
3820b513 | 1674 | const bool offloaded = rcu_rdp_is_offloaded(rdp); |
48a7639c | 1675 | |
a32e01ee | 1676 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1677 | |
67e14c1e PM |
1678 | if (rdp->gp_seq == rnp->gp_seq) |
1679 | return false; /* Nothing to do. */ | |
d09b62df | 1680 | |
67e14c1e PM |
1681 | /* Handle the ends of any preceding grace periods first. */ |
1682 | if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) || | |
1683 | unlikely(READ_ONCE(rdp->gpwrap))) { | |
5d6742b3 PM |
1684 | if (!offloaded) |
1685 | ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */ | |
b5ea0370 | 1686 | rdp->core_needs_qs = false; |
9cbc5b97 | 1687 | trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend")); |
67e14c1e | 1688 | } else { |
5d6742b3 PM |
1689 | if (!offloaded) |
1690 | ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */ | |
b5ea0370 PM |
1691 | if (rdp->core_needs_qs) |
1692 | rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask); | |
d09b62df | 1693 | } |
398ebe60 | 1694 | |
67e14c1e PM |
1695 | /* Now handle the beginnings of any new-to-this-CPU grace periods. */ |
1696 | if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) || | |
1697 | unlikely(READ_ONCE(rdp->gpwrap))) { | |
6eaef633 PM |
1698 | /* |
1699 | * If the current grace period is waiting for this CPU, | |
1700 | * set up to detect a quiescent state, otherwise don't | |
1701 | * go looking for one. | |
1702 | */ | |
9cbc5b97 | 1703 | trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart")); |
b5ea0370 PM |
1704 | need_qs = !!(rnp->qsmask & rdp->grpmask); |
1705 | rdp->cpu_no_qs.b.norm = need_qs; | |
1706 | rdp->core_needs_qs = need_qs; | |
6eaef633 PM |
1707 | zero_cpu_stall_ticks(rdp); |
1708 | } | |
67e14c1e | 1709 | rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */ |
13dc7d0c | 1710 | if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap) |
8ff37290 | 1711 | WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed); |
3d18469a PM |
1712 | WRITE_ONCE(rdp->gpwrap, false); |
1713 | rcu_gpnum_ovf(rnp, rdp); | |
48a7639c | 1714 | return ret; |
6eaef633 PM |
1715 | } |
1716 | ||
15cabdff | 1717 | static void note_gp_changes(struct rcu_data *rdp) |
6eaef633 PM |
1718 | { |
1719 | unsigned long flags; | |
48a7639c | 1720 | bool needwake; |
6eaef633 PM |
1721 | struct rcu_node *rnp; |
1722 | ||
1723 | local_irq_save(flags); | |
1724 | rnp = rdp->mynode; | |
67e14c1e | 1725 | if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) && |
7d0ae808 | 1726 | !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */ |
2a67e741 | 1727 | !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */ |
6eaef633 PM |
1728 | local_irq_restore(flags); |
1729 | return; | |
1730 | } | |
c7e48f7b | 1731 | needwake = __note_gp_changes(rnp, rdp); |
67c583a7 | 1732 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1a2f5d57 | 1733 | rcu_strict_gp_check_qs(); |
48a7639c | 1734 | if (needwake) |
532c00c9 | 1735 | rcu_gp_kthread_wake(); |
6eaef633 PM |
1736 | } |
1737 | ||
22212332 | 1738 | static void rcu_gp_slow(int delay) |
0f41c0dd PM |
1739 | { |
1740 | if (delay > 0 && | |
22212332 | 1741 | !(rcu_seq_ctr(rcu_state.gp_seq) % |
dee4f422 | 1742 | (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay))) |
77865dea | 1743 | schedule_timeout_idle(delay); |
0f41c0dd PM |
1744 | } |
1745 | ||
55b2dcf5 PM |
1746 | static unsigned long sleep_duration; |
1747 | ||
1748 | /* Allow rcutorture to stall the grace-period kthread. */ | |
1749 | void rcu_gp_set_torture_wait(int duration) | |
1750 | { | |
1751 | if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0) | |
1752 | WRITE_ONCE(sleep_duration, duration); | |
1753 | } | |
1754 | EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait); | |
1755 | ||
1756 | /* Actually implement the aforementioned wait. */ | |
1757 | static void rcu_gp_torture_wait(void) | |
1758 | { | |
1759 | unsigned long duration; | |
1760 | ||
1761 | if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST)) | |
1762 | return; | |
1763 | duration = xchg(&sleep_duration, 0UL); | |
1764 | if (duration > 0) { | |
1765 | pr_alert("%s: Waiting %lu jiffies\n", __func__, duration); | |
77865dea | 1766 | schedule_timeout_idle(duration); |
55b2dcf5 PM |
1767 | pr_alert("%s: Wait complete\n", __func__); |
1768 | } | |
1769 | } | |
1770 | ||
933ada2c PM |
1771 | /* |
1772 | * Handler for on_each_cpu() to invoke the target CPU's RCU core | |
1773 | * processing. | |
1774 | */ | |
1775 | static void rcu_strict_gp_boundary(void *unused) | |
1776 | { | |
1777 | invoke_rcu_core(); | |
1778 | } | |
1779 | ||
b3dbec76 | 1780 | /* |
45fed3e7 | 1781 | * Initialize a new grace period. Return false if no grace period required. |
b3dbec76 | 1782 | */ |
0854a05c | 1783 | static bool rcu_gp_init(void) |
b3dbec76 | 1784 | { |
4d60b475 | 1785 | unsigned long firstseq; |
ec2c2976 | 1786 | unsigned long flags; |
0aa04b05 | 1787 | unsigned long oldmask; |
ec2c2976 | 1788 | unsigned long mask; |
b3dbec76 | 1789 | struct rcu_data *rdp; |
336a4f6c | 1790 | struct rcu_node *rnp = rcu_get_root(); |
b3dbec76 | 1791 | |
9cbc5b97 | 1792 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
2a67e741 | 1793 | raw_spin_lock_irq_rcu_node(rnp); |
9cbc5b97 | 1794 | if (!READ_ONCE(rcu_state.gp_flags)) { |
f7be8209 | 1795 | /* Spurious wakeup, tell caller to go back to sleep. */ |
67c583a7 | 1796 | raw_spin_unlock_irq_rcu_node(rnp); |
45fed3e7 | 1797 | return false; |
f7be8209 | 1798 | } |
9cbc5b97 | 1799 | WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */ |
b3dbec76 | 1800 | |
de8e8730 | 1801 | if (WARN_ON_ONCE(rcu_gp_in_progress())) { |
f7be8209 PM |
1802 | /* |
1803 | * Grace period already in progress, don't start another. | |
1804 | * Not supposed to be able to happen. | |
1805 | */ | |
67c583a7 | 1806 | raw_spin_unlock_irq_rcu_node(rnp); |
45fed3e7 | 1807 | return false; |
7fdefc10 PM |
1808 | } |
1809 | ||
7fdefc10 | 1810 | /* Advance to a new grace period and initialize state. */ |
ad3832e9 | 1811 | record_gp_stall_check_time(); |
ff3bb6f4 | 1812 | /* Record GP times before starting GP, hence rcu_seq_start(). */ |
9cbc5b97 | 1813 | rcu_seq_start(&rcu_state.gp_seq); |
62ae1951 | 1814 | ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); |
9cbc5b97 | 1815 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); |
67c583a7 | 1816 | raw_spin_unlock_irq_rcu_node(rnp); |
7fdefc10 | 1817 | |
0aa04b05 | 1818 | /* |
f37599e6 JFG |
1819 | * Apply per-leaf buffered online and offline operations to |
1820 | * the rcu_node tree. Note that this new grace period need not | |
1821 | * wait for subsequent online CPUs, and that RCU hooks in the CPU | |
1822 | * offlining path, when combined with checks in this function, | |
1823 | * will handle CPUs that are currently going offline or that will | |
1824 | * go offline later. Please also refer to "Hotplug CPU" section | |
1825 | * of RCU's Requirements documentation. | |
0aa04b05 | 1826 | */ |
683954e5 | 1827 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF); |
aedf4ba9 | 1828 | rcu_for_each_leaf_node(rnp) { |
4d60b475 PM |
1829 | smp_mb(); // Pair with barriers used when updating ->ofl_seq to odd values. |
1830 | firstseq = READ_ONCE(rnp->ofl_seq); | |
1831 | if (firstseq & 0x1) | |
1832 | while (firstseq == READ_ONCE(rnp->ofl_seq)) | |
1833 | schedule_timeout_idle(1); // Can't wake unless RCU is watching. | |
1834 | smp_mb(); // Pair with barriers used when updating ->ofl_seq to even values. | |
894d45bb | 1835 | raw_spin_lock(&rcu_state.ofl_lock); |
2a67e741 | 1836 | raw_spin_lock_irq_rcu_node(rnp); |
0aa04b05 PM |
1837 | if (rnp->qsmaskinit == rnp->qsmaskinitnext && |
1838 | !rnp->wait_blkd_tasks) { | |
1839 | /* Nothing to do on this leaf rcu_node structure. */ | |
67c583a7 | 1840 | raw_spin_unlock_irq_rcu_node(rnp); |
894d45bb | 1841 | raw_spin_unlock(&rcu_state.ofl_lock); |
0aa04b05 PM |
1842 | continue; |
1843 | } | |
1844 | ||
1845 | /* Record old state, apply changes to ->qsmaskinit field. */ | |
1846 | oldmask = rnp->qsmaskinit; | |
1847 | rnp->qsmaskinit = rnp->qsmaskinitnext; | |
1848 | ||
1849 | /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ | |
1850 | if (!oldmask != !rnp->qsmaskinit) { | |
962aff03 PM |
1851 | if (!oldmask) { /* First online CPU for rcu_node. */ |
1852 | if (!rnp->wait_blkd_tasks) /* Ever offline? */ | |
1853 | rcu_init_new_rnp(rnp); | |
1854 | } else if (rcu_preempt_has_tasks(rnp)) { | |
1855 | rnp->wait_blkd_tasks = true; /* blocked tasks */ | |
1856 | } else { /* Last offline CPU and can propagate. */ | |
0aa04b05 | 1857 | rcu_cleanup_dead_rnp(rnp); |
962aff03 | 1858 | } |
0aa04b05 PM |
1859 | } |
1860 | ||
1861 | /* | |
1862 | * If all waited-on tasks from prior grace period are | |
1863 | * done, and if all this rcu_node structure's CPUs are | |
1864 | * still offline, propagate up the rcu_node tree and | |
1865 | * clear ->wait_blkd_tasks. Otherwise, if one of this | |
1866 | * rcu_node structure's CPUs has since come back online, | |
962aff03 | 1867 | * simply clear ->wait_blkd_tasks. |
0aa04b05 PM |
1868 | */ |
1869 | if (rnp->wait_blkd_tasks && | |
962aff03 | 1870 | (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) { |
0aa04b05 | 1871 | rnp->wait_blkd_tasks = false; |
962aff03 PM |
1872 | if (!rnp->qsmaskinit) |
1873 | rcu_cleanup_dead_rnp(rnp); | |
0aa04b05 PM |
1874 | } |
1875 | ||
67c583a7 | 1876 | raw_spin_unlock_irq_rcu_node(rnp); |
894d45bb | 1877 | raw_spin_unlock(&rcu_state.ofl_lock); |
0aa04b05 | 1878 | } |
22212332 | 1879 | rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */ |
7fdefc10 PM |
1880 | |
1881 | /* | |
1882 | * Set the quiescent-state-needed bits in all the rcu_node | |
9cbc5b97 PM |
1883 | * structures for all currently online CPUs in breadth-first |
1884 | * order, starting from the root rcu_node structure, relying on the | |
1885 | * layout of the tree within the rcu_state.node[] array. Note that | |
1886 | * other CPUs will access only the leaves of the hierarchy, thus | |
1887 | * seeing that no grace period is in progress, at least until the | |
1888 | * corresponding leaf node has been initialized. | |
7fdefc10 PM |
1889 | * |
1890 | * The grace period cannot complete until the initialization | |
1891 | * process finishes, because this kthread handles both. | |
1892 | */ | |
683954e5 | 1893 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT); |
aedf4ba9 | 1894 | rcu_for_each_node_breadth_first(rnp) { |
22212332 | 1895 | rcu_gp_slow(gp_init_delay); |
ec2c2976 | 1896 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
da1df50d | 1897 | rdp = this_cpu_ptr(&rcu_data); |
81ab59a3 | 1898 | rcu_preempt_check_blocked_tasks(rnp); |
7fdefc10 | 1899 | rnp->qsmask = rnp->qsmaskinit; |
9cbc5b97 | 1900 | WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq); |
7fdefc10 | 1901 | if (rnp == rdp->mynode) |
c7e48f7b | 1902 | (void)__note_gp_changes(rnp, rdp); |
7fdefc10 | 1903 | rcu_preempt_boost_start_gp(rnp); |
9cbc5b97 | 1904 | trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq, |
7fdefc10 PM |
1905 | rnp->level, rnp->grplo, |
1906 | rnp->grphi, rnp->qsmask); | |
ec2c2976 PM |
1907 | /* Quiescent states for tasks on any now-offline CPUs. */ |
1908 | mask = rnp->qsmask & ~rnp->qsmaskinitnext; | |
f2e2df59 | 1909 | rnp->rcu_gp_init_mask = mask; |
ec2c2976 | 1910 | if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp)) |
b50912d0 | 1911 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
ec2c2976 PM |
1912 | else |
1913 | raw_spin_unlock_irq_rcu_node(rnp); | |
cee43939 | 1914 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 1915 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
7fdefc10 | 1916 | } |
b3dbec76 | 1917 | |
933ada2c PM |
1918 | // If strict, make all CPUs aware of new grace period. |
1919 | if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) | |
1920 | on_each_cpu(rcu_strict_gp_boundary, NULL, 0); | |
1921 | ||
45fed3e7 | 1922 | return true; |
7fdefc10 | 1923 | } |
b3dbec76 | 1924 | |
b9a425cf | 1925 | /* |
b3dae109 | 1926 | * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state |
d5374226 | 1927 | * time. |
b9a425cf | 1928 | */ |
0854a05c | 1929 | static bool rcu_gp_fqs_check_wake(int *gfp) |
b9a425cf | 1930 | { |
336a4f6c | 1931 | struct rcu_node *rnp = rcu_get_root(); |
b9a425cf | 1932 | |
1fca4d12 PM |
1933 | // If under overload conditions, force an immediate FQS scan. |
1934 | if (*gfp & RCU_GP_FLAG_OVLD) | |
1935 | return true; | |
1936 | ||
1937 | // Someone like call_rcu() requested a force-quiescent-state scan. | |
0854a05c | 1938 | *gfp = READ_ONCE(rcu_state.gp_flags); |
b9a425cf PM |
1939 | if (*gfp & RCU_GP_FLAG_FQS) |
1940 | return true; | |
1941 | ||
1fca4d12 | 1942 | // The current grace period has completed. |
b9a425cf PM |
1943 | if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp)) |
1944 | return true; | |
1945 | ||
1946 | return false; | |
1947 | } | |
1948 | ||
4cdfc175 PM |
1949 | /* |
1950 | * Do one round of quiescent-state forcing. | |
1951 | */ | |
0854a05c | 1952 | static void rcu_gp_fqs(bool first_time) |
4cdfc175 | 1953 | { |
336a4f6c | 1954 | struct rcu_node *rnp = rcu_get_root(); |
4cdfc175 | 1955 | |
9cbc5b97 PM |
1956 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
1957 | rcu_state.n_force_qs++; | |
77f81fe0 | 1958 | if (first_time) { |
4cdfc175 | 1959 | /* Collect dyntick-idle snapshots. */ |
e9ecb780 | 1960 | force_qs_rnp(dyntick_save_progress_counter); |
4cdfc175 PM |
1961 | } else { |
1962 | /* Handle dyntick-idle and offline CPUs. */ | |
e9ecb780 | 1963 | force_qs_rnp(rcu_implicit_dynticks_qs); |
4cdfc175 PM |
1964 | } |
1965 | /* Clear flag to prevent immediate re-entry. */ | |
9cbc5b97 | 1966 | if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) { |
2a67e741 | 1967 | raw_spin_lock_irq_rcu_node(rnp); |
9cbc5b97 PM |
1968 | WRITE_ONCE(rcu_state.gp_flags, |
1969 | READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS); | |
67c583a7 | 1970 | raw_spin_unlock_irq_rcu_node(rnp); |
4cdfc175 | 1971 | } |
4cdfc175 PM |
1972 | } |
1973 | ||
c3854a05 PM |
1974 | /* |
1975 | * Loop doing repeated quiescent-state forcing until the grace period ends. | |
1976 | */ | |
1977 | static void rcu_gp_fqs_loop(void) | |
1978 | { | |
1979 | bool first_gp_fqs; | |
1fca4d12 | 1980 | int gf = 0; |
c3854a05 PM |
1981 | unsigned long j; |
1982 | int ret; | |
1983 | struct rcu_node *rnp = rcu_get_root(); | |
1984 | ||
1985 | first_gp_fqs = true; | |
c06aed0e | 1986 | j = READ_ONCE(jiffies_till_first_fqs); |
1fca4d12 PM |
1987 | if (rcu_state.cbovld) |
1988 | gf = RCU_GP_FLAG_OVLD; | |
c3854a05 PM |
1989 | ret = 0; |
1990 | for (;;) { | |
1991 | if (!ret) { | |
683954e5 NU |
1992 | WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j); |
1993 | /* | |
1994 | * jiffies_force_qs before RCU_GP_WAIT_FQS state | |
1995 | * update; required for stall checks. | |
1996 | */ | |
1997 | smp_wmb(); | |
c3854a05 | 1998 | WRITE_ONCE(rcu_state.jiffies_kick_kthreads, |
9cf422a8 | 1999 | jiffies + (j ? 3 * j : 2)); |
c3854a05 | 2000 | } |
0f11ad32 | 2001 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 | 2002 | TPS("fqswait")); |
683954e5 | 2003 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS); |
c3854a05 PM |
2004 | ret = swait_event_idle_timeout_exclusive( |
2005 | rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j); | |
55b2dcf5 | 2006 | rcu_gp_torture_wait(); |
683954e5 | 2007 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS); |
c3854a05 PM |
2008 | /* Locking provides needed memory barriers. */ |
2009 | /* If grace period done, leave loop. */ | |
2010 | if (!READ_ONCE(rnp->qsmask) && | |
2011 | !rcu_preempt_blocked_readers_cgp(rnp)) | |
2012 | break; | |
2013 | /* If time for quiescent-state forcing, do it. */ | |
29ffebc5 | 2014 | if (!time_after(rcu_state.jiffies_force_qs, jiffies) || |
9c392453 | 2015 | (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) { |
0f11ad32 | 2016 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
2017 | TPS("fqsstart")); |
2018 | rcu_gp_fqs(first_gp_fqs); | |
1fca4d12 PM |
2019 | gf = 0; |
2020 | if (first_gp_fqs) { | |
2021 | first_gp_fqs = false; | |
2022 | gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0; | |
2023 | } | |
0f11ad32 | 2024 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
2025 | TPS("fqsend")); |
2026 | cond_resched_tasks_rcu_qs(); | |
2027 | WRITE_ONCE(rcu_state.gp_activity, jiffies); | |
2028 | ret = 0; /* Force full wait till next FQS. */ | |
c06aed0e | 2029 | j = READ_ONCE(jiffies_till_next_fqs); |
c3854a05 PM |
2030 | } else { |
2031 | /* Deal with stray signal. */ | |
2032 | cond_resched_tasks_rcu_qs(); | |
2033 | WRITE_ONCE(rcu_state.gp_activity, jiffies); | |
2034 | WARN_ON(signal_pending(current)); | |
0f11ad32 | 2035 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
2036 | TPS("fqswaitsig")); |
2037 | ret = 1; /* Keep old FQS timing. */ | |
2038 | j = jiffies; | |
2039 | if (time_after(jiffies, rcu_state.jiffies_force_qs)) | |
2040 | j = 1; | |
2041 | else | |
2042 | j = rcu_state.jiffies_force_qs - j; | |
1fca4d12 | 2043 | gf = 0; |
c3854a05 PM |
2044 | } |
2045 | } | |
2046 | } | |
2047 | ||
7fdefc10 PM |
2048 | /* |
2049 | * Clean up after the old grace period. | |
2050 | */ | |
0854a05c | 2051 | static void rcu_gp_cleanup(void) |
7fdefc10 | 2052 | { |
b2b00ddf | 2053 | int cpu; |
48a7639c | 2054 | bool needgp = false; |
b2b00ddf | 2055 | unsigned long gp_duration; |
de30ad51 | 2056 | unsigned long new_gp_seq; |
5d6742b3 | 2057 | bool offloaded; |
7fdefc10 | 2058 | struct rcu_data *rdp; |
336a4f6c | 2059 | struct rcu_node *rnp = rcu_get_root(); |
abedf8e2 | 2060 | struct swait_queue_head *sq; |
b3dbec76 | 2061 | |
9cbc5b97 | 2062 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
2a67e741 | 2063 | raw_spin_lock_irq_rcu_node(rnp); |
c51d7b5e PM |
2064 | rcu_state.gp_end = jiffies; |
2065 | gp_duration = rcu_state.gp_end - rcu_state.gp_start; | |
9cbc5b97 PM |
2066 | if (gp_duration > rcu_state.gp_max) |
2067 | rcu_state.gp_max = gp_duration; | |
b3dbec76 | 2068 | |
7fdefc10 PM |
2069 | /* |
2070 | * We know the grace period is complete, but to everyone else | |
2071 | * it appears to still be ongoing. But it is also the case | |
2072 | * that to everyone else it looks like there is nothing that | |
2073 | * they can do to advance the grace period. It is therefore | |
2074 | * safe for us to drop the lock in order to mark the grace | |
2075 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 2076 | */ |
67c583a7 | 2077 | raw_spin_unlock_irq_rcu_node(rnp); |
b3dbec76 | 2078 | |
5d4b8659 | 2079 | /* |
ff3bb6f4 PM |
2080 | * Propagate new ->gp_seq value to rcu_node structures so that |
2081 | * other CPUs don't have to wait until the start of the next grace | |
2082 | * period to process their callbacks. This also avoids some nasty | |
2083 | * RCU grace-period initialization races by forcing the end of | |
2084 | * the current grace period to be completely recorded in all of | |
2085 | * the rcu_node structures before the beginning of the next grace | |
2086 | * period is recorded in any of the rcu_node structures. | |
5d4b8659 | 2087 | */ |
9cbc5b97 | 2088 | new_gp_seq = rcu_state.gp_seq; |
de30ad51 | 2089 | rcu_seq_end(&new_gp_seq); |
aedf4ba9 | 2090 | rcu_for_each_node_breadth_first(rnp) { |
2a67e741 | 2091 | raw_spin_lock_irq_rcu_node(rnp); |
4bc8d555 | 2092 | if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp))) |
81ab59a3 | 2093 | dump_blkd_tasks(rnp, 10); |
5c60d25f | 2094 | WARN_ON_ONCE(rnp->qsmask); |
de30ad51 | 2095 | WRITE_ONCE(rnp->gp_seq, new_gp_seq); |
da1df50d | 2096 | rdp = this_cpu_ptr(&rcu_data); |
b11cc576 | 2097 | if (rnp == rdp->mynode) |
c7e48f7b | 2098 | needgp = __note_gp_changes(rnp, rdp) || needgp; |
78e4bc34 | 2099 | /* smp_mb() provided by prior unlock-lock pair. */ |
3481f2ea | 2100 | needgp = rcu_future_gp_cleanup(rnp) || needgp; |
b2b00ddf PM |
2101 | // Reset overload indication for CPUs no longer overloaded |
2102 | if (rcu_is_leaf_node(rnp)) | |
2103 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) { | |
2104 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
2105 | check_cb_ovld_locked(rdp, rnp); | |
2106 | } | |
065bb78c | 2107 | sq = rcu_nocb_gp_get(rnp); |
67c583a7 | 2108 | raw_spin_unlock_irq_rcu_node(rnp); |
065bb78c | 2109 | rcu_nocb_gp_cleanup(sq); |
cee43939 | 2110 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 2111 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
22212332 | 2112 | rcu_gp_slow(gp_cleanup_delay); |
7fdefc10 | 2113 | } |
336a4f6c | 2114 | rnp = rcu_get_root(); |
9cbc5b97 | 2115 | raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */ |
7fdefc10 | 2116 | |
0a89e5a4 | 2117 | /* Declare grace period done, trace first to use old GP number. */ |
9cbc5b97 | 2118 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end")); |
0a89e5a4 | 2119 | rcu_seq_end(&rcu_state.gp_seq); |
62ae1951 | 2120 | ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); |
683954e5 | 2121 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE); |
fb31340f | 2122 | /* Check for GP requests since above loop. */ |
da1df50d | 2123 | rdp = this_cpu_ptr(&rcu_data); |
5b55072f | 2124 | if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) { |
abd13fdd | 2125 | trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed, |
41e80595 | 2126 | TPS("CleanupMore")); |
fb31340f PM |
2127 | needgp = true; |
2128 | } | |
48a7639c | 2129 | /* Advance CBs to reduce false positives below. */ |
3820b513 | 2130 | offloaded = rcu_rdp_is_offloaded(rdp); |
5d6742b3 | 2131 | if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) { |
9cbc5b97 | 2132 | WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT); |
2906d215 | 2133 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); |
9cbc5b97 | 2134 | trace_rcu_grace_period(rcu_state.name, |
0f11ad32 | 2135 | rcu_state.gp_seq, |
bb311ecc | 2136 | TPS("newreq")); |
18390aea | 2137 | } else { |
9cbc5b97 PM |
2138 | WRITE_ONCE(rcu_state.gp_flags, |
2139 | rcu_state.gp_flags & RCU_GP_FLAG_INIT); | |
bb311ecc | 2140 | } |
67c583a7 | 2141 | raw_spin_unlock_irq_rcu_node(rnp); |
4e025f52 PM |
2142 | |
2143 | // If strict, make all CPUs aware of the end of the old grace period. | |
2144 | if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) | |
2145 | on_each_cpu(rcu_strict_gp_boundary, NULL, 0); | |
7fdefc10 PM |
2146 | } |
2147 | ||
2148 | /* | |
2149 | * Body of kthread that handles grace periods. | |
2150 | */ | |
0854a05c | 2151 | static int __noreturn rcu_gp_kthread(void *unused) |
7fdefc10 | 2152 | { |
5871968d | 2153 | rcu_bind_gp_kthread(); |
7fdefc10 PM |
2154 | for (;;) { |
2155 | ||
2156 | /* Handle grace-period start. */ | |
2157 | for (;;) { | |
0f11ad32 | 2158 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
63c4db78 | 2159 | TPS("reqwait")); |
683954e5 | 2160 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS); |
9cbc5b97 PM |
2161 | swait_event_idle_exclusive(rcu_state.gp_wq, |
2162 | READ_ONCE(rcu_state.gp_flags) & | |
2163 | RCU_GP_FLAG_INIT); | |
55b2dcf5 | 2164 | rcu_gp_torture_wait(); |
683954e5 | 2165 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS); |
78e4bc34 | 2166 | /* Locking provides needed memory barrier. */ |
0854a05c | 2167 | if (rcu_gp_init()) |
7fdefc10 | 2168 | break; |
cee43939 | 2169 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 2170 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
73a860cd | 2171 | WARN_ON(signal_pending(current)); |
0f11ad32 | 2172 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
63c4db78 | 2173 | TPS("reqwaitsig")); |
7fdefc10 | 2174 | } |
cabc49c1 | 2175 | |
4cdfc175 | 2176 | /* Handle quiescent-state forcing. */ |
c3854a05 | 2177 | rcu_gp_fqs_loop(); |
4cdfc175 PM |
2178 | |
2179 | /* Handle grace-period end. */ | |
683954e5 | 2180 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP); |
0854a05c | 2181 | rcu_gp_cleanup(); |
683954e5 | 2182 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED); |
b3dbec76 | 2183 | } |
b3dbec76 PM |
2184 | } |
2185 | ||
f41d911f | 2186 | /* |
49918a54 PM |
2187 | * Report a full set of quiescent states to the rcu_state data structure. |
2188 | * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if | |
2189 | * another grace period is required. Whether we wake the grace-period | |
2190 | * kthread or it awakens itself for the next round of quiescent-state | |
2191 | * forcing, that kthread will clean up after the just-completed grace | |
2192 | * period. Note that the caller must hold rnp->lock, which is released | |
2193 | * before return. | |
f41d911f | 2194 | */ |
aff4e9ed | 2195 | static void rcu_report_qs_rsp(unsigned long flags) |
336a4f6c | 2196 | __releases(rcu_get_root()->lock) |
f41d911f | 2197 | { |
336a4f6c | 2198 | raw_lockdep_assert_held_rcu_node(rcu_get_root()); |
de8e8730 | 2199 | WARN_ON_ONCE(!rcu_gp_in_progress()); |
9cbc5b97 PM |
2200 | WRITE_ONCE(rcu_state.gp_flags, |
2201 | READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS); | |
336a4f6c | 2202 | raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags); |
532c00c9 | 2203 | rcu_gp_kthread_wake(); |
f41d911f PM |
2204 | } |
2205 | ||
64db4cff | 2206 | /* |
d3f6bad3 PM |
2207 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
2208 | * Allows quiescent states for a group of CPUs to be reported at one go | |
2209 | * to the specified rcu_node structure, though all the CPUs in the group | |
654e9533 PM |
2210 | * must be represented by the same rcu_node structure (which need not be a |
2211 | * leaf rcu_node structure, though it often will be). The gps parameter | |
2212 | * is the grace-period snapshot, which means that the quiescent states | |
c9a24e2d | 2213 | * are valid only if rnp->gp_seq is equal to gps. That structure's lock |
654e9533 | 2214 | * must be held upon entry, and it is released before return. |
ec2c2976 PM |
2215 | * |
2216 | * As a special case, if mask is zero, the bit-already-cleared check is | |
2217 | * disabled. This allows propagating quiescent state due to resumed tasks | |
2218 | * during grace-period initialization. | |
64db4cff | 2219 | */ |
b50912d0 PM |
2220 | static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp, |
2221 | unsigned long gps, unsigned long flags) | |
64db4cff PM |
2222 | __releases(rnp->lock) |
2223 | { | |
654e9533 | 2224 | unsigned long oldmask = 0; |
28ecd580 PM |
2225 | struct rcu_node *rnp_c; |
2226 | ||
a32e01ee | 2227 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 2228 | |
64db4cff PM |
2229 | /* Walk up the rcu_node hierarchy. */ |
2230 | for (;;) { | |
ec2c2976 | 2231 | if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) { |
64db4cff | 2232 | |
654e9533 PM |
2233 | /* |
2234 | * Our bit has already been cleared, or the | |
2235 | * relevant grace period is already over, so done. | |
2236 | */ | |
67c583a7 | 2237 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2238 | return; |
2239 | } | |
654e9533 | 2240 | WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ |
5b4c11d5 | 2241 | WARN_ON_ONCE(!rcu_is_leaf_node(rnp) && |
2dee9404 | 2242 | rcu_preempt_blocked_readers_cgp(rnp)); |
7672d647 | 2243 | WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask); |
67a0edbf | 2244 | trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq, |
d4c08f2a PM |
2245 | mask, rnp->qsmask, rnp->level, |
2246 | rnp->grplo, rnp->grphi, | |
2247 | !!rnp->gp_tasks); | |
27f4d280 | 2248 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
2249 | |
2250 | /* Other bits still set at this level, so done. */ | |
67c583a7 | 2251 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2252 | return; |
2253 | } | |
d43a5d32 | 2254 | rnp->completedqs = rnp->gp_seq; |
64db4cff PM |
2255 | mask = rnp->grpmask; |
2256 | if (rnp->parent == NULL) { | |
2257 | ||
2258 | /* No more levels. Exit loop holding root lock. */ | |
2259 | ||
2260 | break; | |
2261 | } | |
67c583a7 | 2262 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
28ecd580 | 2263 | rnp_c = rnp; |
64db4cff | 2264 | rnp = rnp->parent; |
2a67e741 | 2265 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
0937d045 | 2266 | oldmask = READ_ONCE(rnp_c->qsmask); |
64db4cff PM |
2267 | } |
2268 | ||
2269 | /* | |
2270 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 2271 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 2272 | * to clean up and start the next grace period if one is needed. |
64db4cff | 2273 | */ |
aff4e9ed | 2274 | rcu_report_qs_rsp(flags); /* releases rnp->lock. */ |
64db4cff PM |
2275 | } |
2276 | ||
cc99a310 PM |
2277 | /* |
2278 | * Record a quiescent state for all tasks that were previously queued | |
2279 | * on the specified rcu_node structure and that were blocking the current | |
49918a54 | 2280 | * RCU grace period. The caller must hold the corresponding rnp->lock with |
cc99a310 PM |
2281 | * irqs disabled, and this lock is released upon return, but irqs remain |
2282 | * disabled. | |
2283 | */ | |
17a8212b | 2284 | static void __maybe_unused |
139ad4da | 2285 | rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
cc99a310 PM |
2286 | __releases(rnp->lock) |
2287 | { | |
654e9533 | 2288 | unsigned long gps; |
cc99a310 PM |
2289 | unsigned long mask; |
2290 | struct rcu_node *rnp_p; | |
2291 | ||
a32e01ee | 2292 | raw_lockdep_assert_held_rcu_node(rnp); |
c130d2dc | 2293 | if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) || |
c74859d1 PM |
2294 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) || |
2295 | rnp->qsmask != 0) { | |
67c583a7 | 2296 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
cc99a310 PM |
2297 | return; /* Still need more quiescent states! */ |
2298 | } | |
2299 | ||
77cfc7bf | 2300 | rnp->completedqs = rnp->gp_seq; |
cc99a310 PM |
2301 | rnp_p = rnp->parent; |
2302 | if (rnp_p == NULL) { | |
2303 | /* | |
a77da14c PM |
2304 | * Only one rcu_node structure in the tree, so don't |
2305 | * try to report up to its nonexistent parent! | |
cc99a310 | 2306 | */ |
aff4e9ed | 2307 | rcu_report_qs_rsp(flags); |
cc99a310 PM |
2308 | return; |
2309 | } | |
2310 | ||
c9a24e2d PM |
2311 | /* Report up the rest of the hierarchy, tracking current ->gp_seq. */ |
2312 | gps = rnp->gp_seq; | |
cc99a310 | 2313 | mask = rnp->grpmask; |
67c583a7 | 2314 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
2a67e741 | 2315 | raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */ |
b50912d0 | 2316 | rcu_report_qs_rnp(mask, rnp_p, gps, flags); |
cc99a310 PM |
2317 | } |
2318 | ||
64db4cff | 2319 | /* |
d3f6bad3 | 2320 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
4b455dc3 | 2321 | * structure. This must be called from the specified CPU. |
64db4cff PM |
2322 | */ |
2323 | static void | |
cfeac397 | 2324 | rcu_report_qs_rdp(struct rcu_data *rdp) |
64db4cff PM |
2325 | { |
2326 | unsigned long flags; | |
2327 | unsigned long mask; | |
5d6742b3 | 2328 | bool needwake = false; |
3820b513 | 2329 | const bool offloaded = rcu_rdp_is_offloaded(rdp); |
64db4cff PM |
2330 | struct rcu_node *rnp; |
2331 | ||
cfeac397 | 2332 | WARN_ON_ONCE(rdp->cpu != smp_processor_id()); |
64db4cff | 2333 | rnp = rdp->mynode; |
2a67e741 | 2334 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
c9a24e2d PM |
2335 | if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq || |
2336 | rdp->gpwrap) { | |
64db4cff PM |
2337 | |
2338 | /* | |
e4cc1f22 PM |
2339 | * The grace period in which this quiescent state was |
2340 | * recorded has ended, so don't report it upwards. | |
2341 | * We will instead need a new quiescent state that lies | |
2342 | * within the current grace period. | |
64db4cff | 2343 | */ |
5b74c458 | 2344 | rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */ |
67c583a7 | 2345 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2346 | return; |
2347 | } | |
2348 | mask = rdp->grpmask; | |
cfeac397 | 2349 | rdp->core_needs_qs = false; |
64db4cff | 2350 | if ((rnp->qsmask & mask) == 0) { |
67c583a7 | 2351 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff | 2352 | } else { |
64db4cff PM |
2353 | /* |
2354 | * This GP can't end until cpu checks in, so all of our | |
2355 | * callbacks can be processed during the next GP. | |
2356 | */ | |
5d6742b3 PM |
2357 | if (!offloaded) |
2358 | needwake = rcu_accelerate_cbs(rnp, rdp); | |
64db4cff | 2359 | |
516e5ae0 | 2360 | rcu_disable_urgency_upon_qs(rdp); |
b50912d0 | 2361 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
654e9533 | 2362 | /* ^^^ Released rnp->lock */ |
48a7639c | 2363 | if (needwake) |
532c00c9 | 2364 | rcu_gp_kthread_wake(); |
64db4cff PM |
2365 | } |
2366 | } | |
2367 | ||
2368 | /* | |
2369 | * Check to see if there is a new grace period of which this CPU | |
2370 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2371 | * Otherwise, see if this CPU has just passed through its first | |
2372 | * quiescent state for this grace period, and record that fact if so. | |
2373 | */ | |
2374 | static void | |
8087d3e3 | 2375 | rcu_check_quiescent_state(struct rcu_data *rdp) |
64db4cff | 2376 | { |
05eb552b | 2377 | /* Check for grace-period ends and beginnings. */ |
15cabdff | 2378 | note_gp_changes(rdp); |
64db4cff PM |
2379 | |
2380 | /* | |
2381 | * Does this CPU still need to do its part for current grace period? | |
2382 | * If no, return and let the other CPUs do their part as well. | |
2383 | */ | |
97c668b8 | 2384 | if (!rdp->core_needs_qs) |
64db4cff PM |
2385 | return; |
2386 | ||
2387 | /* | |
2388 | * Was there a quiescent state since the beginning of the grace | |
2389 | * period? If no, then exit and wait for the next call. | |
2390 | */ | |
3a19b46a | 2391 | if (rdp->cpu_no_qs.b.norm) |
64db4cff PM |
2392 | return; |
2393 | ||
d3f6bad3 PM |
2394 | /* |
2395 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2396 | * judge of that). | |
2397 | */ | |
cfeac397 | 2398 | rcu_report_qs_rdp(rdp); |
64db4cff PM |
2399 | } |
2400 | ||
b1420f1c | 2401 | /* |
780cd590 PM |
2402 | * Near the end of the offline process. Trace the fact that this CPU |
2403 | * is going offline. | |
b1420f1c | 2404 | */ |
780cd590 | 2405 | int rcutree_dying_cpu(unsigned int cpu) |
b1420f1c | 2406 | { |
4f5fbd78 YS |
2407 | bool blkd; |
2408 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); | |
2409 | struct rcu_node *rnp = rdp->mynode; | |
b1420f1c | 2410 | |
ea46351c | 2411 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
780cd590 | 2412 | return 0; |
ea46351c | 2413 | |
4f5fbd78 | 2414 | blkd = !!(rnp->qsmask & rdp->grpmask); |
0937d045 | 2415 | trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq), |
47fcbc8d | 2416 | blkd ? TPS("cpuofl-bgp") : TPS("cpuofl")); |
780cd590 | 2417 | return 0; |
64db4cff PM |
2418 | } |
2419 | ||
8af3a5e7 PM |
2420 | /* |
2421 | * All CPUs for the specified rcu_node structure have gone offline, | |
2422 | * and all tasks that were preempted within an RCU read-side critical | |
2423 | * section while running on one of those CPUs have since exited their RCU | |
2424 | * read-side critical section. Some other CPU is reporting this fact with | |
2425 | * the specified rcu_node structure's ->lock held and interrupts disabled. | |
2426 | * This function therefore goes up the tree of rcu_node structures, | |
2427 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that | |
2428 | * the leaf rcu_node structure's ->qsmaskinit field has already been | |
c50cbe53 | 2429 | * updated. |
8af3a5e7 PM |
2430 | * |
2431 | * This function does check that the specified rcu_node structure has | |
2432 | * all CPUs offline and no blocked tasks, so it is OK to invoke it | |
2433 | * prematurely. That said, invoking it after the fact will cost you | |
2434 | * a needless lock acquisition. So once it has done its work, don't | |
2435 | * invoke it again. | |
2436 | */ | |
2437 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) | |
2438 | { | |
2439 | long mask; | |
2440 | struct rcu_node *rnp = rnp_leaf; | |
2441 | ||
962aff03 | 2442 | raw_lockdep_assert_held_rcu_node(rnp_leaf); |
ea46351c | 2443 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || |
962aff03 PM |
2444 | WARN_ON_ONCE(rnp_leaf->qsmaskinit) || |
2445 | WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf))) | |
8af3a5e7 PM |
2446 | return; |
2447 | for (;;) { | |
2448 | mask = rnp->grpmask; | |
2449 | rnp = rnp->parent; | |
2450 | if (!rnp) | |
2451 | break; | |
2a67e741 | 2452 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
8af3a5e7 | 2453 | rnp->qsmaskinit &= ~mask; |
962aff03 PM |
2454 | /* Between grace periods, so better already be zero! */ |
2455 | WARN_ON_ONCE(rnp->qsmask); | |
8af3a5e7 | 2456 | if (rnp->qsmaskinit) { |
67c583a7 BF |
2457 | raw_spin_unlock_rcu_node(rnp); |
2458 | /* irqs remain disabled. */ | |
8af3a5e7 PM |
2459 | return; |
2460 | } | |
67c583a7 | 2461 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
8af3a5e7 PM |
2462 | } |
2463 | } | |
2464 | ||
64db4cff | 2465 | /* |
e5601400 | 2466 | * The CPU has been completely removed, and some other CPU is reporting |
a58163d8 PM |
2467 | * this fact from process context. Do the remainder of the cleanup. |
2468 | * There can only be one CPU hotplug operation at a time, so no need for | |
2469 | * explicit locking. | |
64db4cff | 2470 | */ |
780cd590 | 2471 | int rcutree_dead_cpu(unsigned int cpu) |
64db4cff | 2472 | { |
da1df50d | 2473 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
b1420f1c | 2474 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
e5601400 | 2475 | |
ea46351c | 2476 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) |
780cd590 | 2477 | return 0; |
ea46351c | 2478 | |
ed73860c | 2479 | WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1); |
2036d94a | 2480 | /* Adjust any no-longer-needed kthreads. */ |
5d01bbd1 | 2481 | rcu_boost_kthread_setaffinity(rnp, -1); |
780cd590 PM |
2482 | /* Do any needed no-CB deferred wakeups from this CPU. */ |
2483 | do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu)); | |
96926686 PM |
2484 | |
2485 | // Stop-machine done, so allow nohz_full to disable tick. | |
2486 | tick_dep_clear(TICK_DEP_BIT_RCU); | |
780cd590 | 2487 | return 0; |
64db4cff PM |
2488 | } |
2489 | ||
64db4cff PM |
2490 | /* |
2491 | * Invoke any RCU callbacks that have made it to the end of their grace | |
a616aec9 | 2492 | * period. Throttle as specified by rdp->blimit. |
64db4cff | 2493 | */ |
5bb5d09c | 2494 | static void rcu_do_batch(struct rcu_data *rdp) |
64db4cff | 2495 | { |
b5374b2d | 2496 | int div; |
b4e6039e | 2497 | bool __maybe_unused empty; |
64db4cff | 2498 | unsigned long flags; |
3820b513 | 2499 | const bool offloaded = rcu_rdp_is_offloaded(rdp); |
15fecf89 PM |
2500 | struct rcu_head *rhp; |
2501 | struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl); | |
6bc33582 | 2502 | long bl, count = 0; |
cfcdef5e | 2503 | long pending, tlimit = 0; |
64db4cff | 2504 | |
dc35c893 | 2505 | /* If no callbacks are ready, just return. */ |
15fecf89 | 2506 | if (!rcu_segcblist_ready_cbs(&rdp->cblist)) { |
3c779dfe | 2507 | trace_rcu_batch_start(rcu_state.name, |
15fecf89 | 2508 | rcu_segcblist_n_cbs(&rdp->cblist), 0); |
3c779dfe | 2509 | trace_rcu_batch_end(rcu_state.name, 0, |
15fecf89 | 2510 | !rcu_segcblist_empty(&rdp->cblist), |
4968c300 PM |
2511 | need_resched(), is_idle_task(current), |
2512 | rcu_is_callbacks_kthread()); | |
64db4cff | 2513 | return; |
29c00b4a | 2514 | } |
64db4cff PM |
2515 | |
2516 | /* | |
2517 | * Extract the list of ready callbacks, disabling to prevent | |
15fecf89 PM |
2518 | * races with call_rcu() from interrupt handlers. Leave the |
2519 | * callback counts, as rcu_barrier() needs to be conservative. | |
64db4cff PM |
2520 | */ |
2521 | local_irq_save(flags); | |
5d6742b3 | 2522 | rcu_nocb_lock(rdp); |
8146c4e2 | 2523 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
cfcdef5e | 2524 | pending = rcu_segcblist_n_cbs(&rdp->cblist); |
b5374b2d PM |
2525 | div = READ_ONCE(rcu_divisor); |
2526 | div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div; | |
2527 | bl = max(rdp->blimit, pending >> div); | |
a2b354b9 PM |
2528 | if (unlikely(bl > 100)) { |
2529 | long rrn = READ_ONCE(rcu_resched_ns); | |
2530 | ||
2531 | rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn; | |
2532 | tlimit = local_clock() + rrn; | |
2533 | } | |
3c779dfe | 2534 | trace_rcu_batch_start(rcu_state.name, |
15fecf89 PM |
2535 | rcu_segcblist_n_cbs(&rdp->cblist), bl); |
2536 | rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl); | |
7f36ef82 PM |
2537 | if (offloaded) |
2538 | rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist); | |
3afe7fa5 JFG |
2539 | |
2540 | trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued")); | |
5d6742b3 | 2541 | rcu_nocb_unlock_irqrestore(rdp, flags); |
64db4cff PM |
2542 | |
2543 | /* Invoke callbacks. */ | |
6a949b7a | 2544 | tick_dep_set_task(current, TICK_DEP_BIT_RCU); |
15fecf89 | 2545 | rhp = rcu_cblist_dequeue(&rcl); |
3afe7fa5 | 2546 | |
15fecf89 | 2547 | for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) { |
77a40f97 JFG |
2548 | rcu_callback_t f; |
2549 | ||
6bc33582 | 2550 | count++; |
15fecf89 | 2551 | debug_rcu_head_unqueue(rhp); |
77a40f97 JFG |
2552 | |
2553 | rcu_lock_acquire(&rcu_callback_map); | |
2554 | trace_rcu_invoke_callback(rcu_state.name, rhp); | |
2555 | ||
2556 | f = rhp->func; | |
2557 | WRITE_ONCE(rhp->func, (rcu_callback_t)0L); | |
2558 | f(rhp); | |
2559 | ||
2560 | rcu_lock_release(&rcu_callback_map); | |
2561 | ||
15fecf89 PM |
2562 | /* |
2563 | * Stop only if limit reached and CPU has something to do. | |
15fecf89 | 2564 | */ |
6bc33582 | 2565 | if (count >= bl && !offloaded && |
dff1672d PM |
2566 | (need_resched() || |
2567 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) | |
64db4cff | 2568 | break; |
cfcdef5e ED |
2569 | if (unlikely(tlimit)) { |
2570 | /* only call local_clock() every 32 callbacks */ | |
6bc33582 | 2571 | if (likely((count & 31) || local_clock() < tlimit)) |
cfcdef5e ED |
2572 | continue; |
2573 | /* Exceeded the time limit, so leave. */ | |
2574 | break; | |
2575 | } | |
e3abe959 | 2576 | if (!in_serving_softirq()) { |
5d6742b3 PM |
2577 | local_bh_enable(); |
2578 | lockdep_assert_irqs_enabled(); | |
2579 | cond_resched_tasks_rcu_qs(); | |
2580 | lockdep_assert_irqs_enabled(); | |
2581 | local_bh_disable(); | |
2582 | } | |
64db4cff PM |
2583 | } |
2584 | ||
2585 | local_irq_save(flags); | |
5d6742b3 | 2586 | rcu_nocb_lock(rdp); |
e816d56f | 2587 | rdp->n_cbs_invoked += count; |
3c779dfe | 2588 | trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(), |
8ef0f37e | 2589 | is_idle_task(current), rcu_is_callbacks_kthread()); |
64db4cff | 2590 | |
15fecf89 PM |
2591 | /* Update counts and requeue any remaining callbacks. */ |
2592 | rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl); | |
6bc33582 | 2593 | rcu_segcblist_add_len(&rdp->cblist, -count); |
64db4cff PM |
2594 | |
2595 | /* Reinstate batch limit if we have worked down the excess. */ | |
15fecf89 | 2596 | count = rcu_segcblist_n_cbs(&rdp->cblist); |
d5a9a8c3 | 2597 | if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark) |
64db4cff PM |
2598 | rdp->blimit = blimit; |
2599 | ||
37c72e56 | 2600 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
15fecf89 | 2601 | if (count == 0 && rdp->qlen_last_fqs_check != 0) { |
37c72e56 | 2602 | rdp->qlen_last_fqs_check = 0; |
3c779dfe | 2603 | rdp->n_force_qs_snap = rcu_state.n_force_qs; |
15fecf89 PM |
2604 | } else if (count < rdp->qlen_last_fqs_check - qhimark) |
2605 | rdp->qlen_last_fqs_check = count; | |
efd88b02 PM |
2606 | |
2607 | /* | |
2608 | * The following usually indicates a double call_rcu(). To track | |
2609 | * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y. | |
2610 | */ | |
b4e6039e JFG |
2611 | empty = rcu_segcblist_empty(&rdp->cblist); |
2612 | WARN_ON_ONCE(count == 0 && !empty); | |
d1b222c6 | 2613 | WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) && |
b4e6039e JFG |
2614 | count != 0 && empty); |
2615 | WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0); | |
2616 | WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0); | |
37c72e56 | 2617 | |
5d6742b3 | 2618 | rcu_nocb_unlock_irqrestore(rdp, flags); |
64db4cff | 2619 | |
e0f23060 | 2620 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
ec5ef87b | 2621 | if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist)) |
a46e0899 | 2622 | invoke_rcu_core(); |
6a949b7a | 2623 | tick_dep_clear_task(current, TICK_DEP_BIT_RCU); |
64db4cff PM |
2624 | } |
2625 | ||
2626 | /* | |
c98cac60 PM |
2627 | * This function is invoked from each scheduling-clock interrupt, |
2628 | * and checks to see if this CPU is in a non-context-switch quiescent | |
2629 | * state, for example, user mode or idle loop. It also schedules RCU | |
2630 | * core processing. If the current grace period has gone on too long, | |
2631 | * it will ask the scheduler to manufacture a context switch for the sole | |
2632 | * purpose of providing a providing the needed quiescent state. | |
64db4cff | 2633 | */ |
c98cac60 | 2634 | void rcu_sched_clock_irq(int user) |
64db4cff | 2635 | { |
f7f7bac9 | 2636 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a649d25d | 2637 | lockdep_assert_irqs_disabled(); |
4e95020c | 2638 | raw_cpu_inc(rcu_data.ticks_this_gp); |
92aa39e9 | 2639 | /* The load-acquire pairs with the store-release setting to true. */ |
2dba13f0 | 2640 | if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) { |
92aa39e9 | 2641 | /* Idle and userspace execution already are quiescent states. */ |
a0ef9ec2 | 2642 | if (!rcu_is_cpu_rrupt_from_idle() && !user) { |
92aa39e9 PM |
2643 | set_tsk_need_resched(current); |
2644 | set_preempt_need_resched(); | |
2645 | } | |
2dba13f0 | 2646 | __this_cpu_write(rcu_data.rcu_urgent_qs, false); |
64db4cff | 2647 | } |
c98cac60 | 2648 | rcu_flavor_sched_clock_irq(user); |
dd7dafd1 | 2649 | if (rcu_pending(user)) |
a46e0899 | 2650 | invoke_rcu_core(); |
a649d25d | 2651 | lockdep_assert_irqs_disabled(); |
07f27570 | 2652 | |
f7f7bac9 | 2653 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2654 | } |
2655 | ||
64db4cff | 2656 | /* |
5d8a752e ZZ |
2657 | * Scan the leaf rcu_node structures. For each structure on which all |
2658 | * CPUs have reported a quiescent state and on which there are tasks | |
2659 | * blocking the current grace period, initiate RCU priority boosting. | |
2660 | * Otherwise, invoke the specified function to check dyntick state for | |
2661 | * each CPU that has not yet reported a quiescent state. | |
64db4cff | 2662 | */ |
8ff0b907 | 2663 | static void force_qs_rnp(int (*f)(struct rcu_data *rdp)) |
64db4cff | 2664 | { |
64db4cff PM |
2665 | int cpu; |
2666 | unsigned long flags; | |
2667 | unsigned long mask; | |
66e4c33b | 2668 | struct rcu_data *rdp; |
a0b6c9a7 | 2669 | struct rcu_node *rnp; |
64db4cff | 2670 | |
b2b00ddf PM |
2671 | rcu_state.cbovld = rcu_state.cbovldnext; |
2672 | rcu_state.cbovldnext = false; | |
aedf4ba9 | 2673 | rcu_for_each_leaf_node(rnp) { |
cee43939 | 2674 | cond_resched_tasks_rcu_qs(); |
64db4cff | 2675 | mask = 0; |
2a67e741 | 2676 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
b2b00ddf | 2677 | rcu_state.cbovldnext |= !!rnp->cbovldmask; |
a0b6c9a7 | 2678 | if (rnp->qsmask == 0) { |
9b1ce0ac | 2679 | if (rcu_preempt_blocked_readers_cgp(rnp)) { |
a77da14c PM |
2680 | /* |
2681 | * No point in scanning bits because they | |
2682 | * are all zero. But we might need to | |
2683 | * priority-boost blocked readers. | |
2684 | */ | |
2685 | rcu_initiate_boost(rnp, flags); | |
2686 | /* rcu_initiate_boost() releases rnp->lock */ | |
2687 | continue; | |
2688 | } | |
92816435 PM |
2689 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
2690 | continue; | |
64db4cff | 2691 | } |
7441e766 PM |
2692 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) { |
2693 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
2694 | if (f(rdp)) { | |
2695 | mask |= rdp->grpmask; | |
2696 | rcu_disable_urgency_upon_qs(rdp); | |
0edd1b17 | 2697 | } |
64db4cff | 2698 | } |
45f014c5 | 2699 | if (mask != 0) { |
c9a24e2d | 2700 | /* Idle/offline CPUs, report (releases rnp->lock). */ |
b50912d0 | 2701 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
0aa04b05 PM |
2702 | } else { |
2703 | /* Nothing to do here, so just drop the lock. */ | |
67c583a7 | 2704 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff | 2705 | } |
64db4cff | 2706 | } |
64db4cff PM |
2707 | } |
2708 | ||
2709 | /* | |
2710 | * Force quiescent states on reluctant CPUs, and also detect which | |
2711 | * CPUs are in dyntick-idle mode. | |
2712 | */ | |
cd920e5a | 2713 | void rcu_force_quiescent_state(void) |
64db4cff PM |
2714 | { |
2715 | unsigned long flags; | |
394f2769 PM |
2716 | bool ret; |
2717 | struct rcu_node *rnp; | |
2718 | struct rcu_node *rnp_old = NULL; | |
2719 | ||
2720 | /* Funnel through hierarchy to reduce memory contention. */ | |
da1df50d | 2721 | rnp = __this_cpu_read(rcu_data.mynode); |
394f2769 | 2722 | for (; rnp != NULL; rnp = rnp->parent) { |
67a0edbf | 2723 | ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) || |
66e4c33b | 2724 | !raw_spin_trylock(&rnp->fqslock); |
394f2769 PM |
2725 | if (rnp_old != NULL) |
2726 | raw_spin_unlock(&rnp_old->fqslock); | |
d62df573 | 2727 | if (ret) |
394f2769 | 2728 | return; |
394f2769 PM |
2729 | rnp_old = rnp; |
2730 | } | |
336a4f6c | 2731 | /* rnp_old == rcu_get_root(), rnp == NULL. */ |
64db4cff | 2732 | |
394f2769 | 2733 | /* Reached the root of the rcu_node tree, acquire lock. */ |
2a67e741 | 2734 | raw_spin_lock_irqsave_rcu_node(rnp_old, flags); |
394f2769 | 2735 | raw_spin_unlock(&rnp_old->fqslock); |
67a0edbf | 2736 | if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) { |
67c583a7 | 2737 | raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags); |
4cdfc175 | 2738 | return; /* Someone beat us to it. */ |
46a1e34e | 2739 | } |
67a0edbf PM |
2740 | WRITE_ONCE(rcu_state.gp_flags, |
2741 | READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS); | |
67c583a7 | 2742 | raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags); |
532c00c9 | 2743 | rcu_gp_kthread_wake(); |
64db4cff | 2744 | } |
cd920e5a | 2745 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); |
64db4cff | 2746 | |
a657f261 PM |
2747 | // Workqueue handler for an RCU reader for kernels enforcing struct RCU |
2748 | // grace periods. | |
2749 | static void strict_work_handler(struct work_struct *work) | |
2750 | { | |
2751 | rcu_read_lock(); | |
2752 | rcu_read_unlock(); | |
2753 | } | |
2754 | ||
fb60e533 | 2755 | /* Perform RCU core processing work for the current CPU. */ |
48d07c04 | 2756 | static __latent_entropy void rcu_core(void) |
64db4cff PM |
2757 | { |
2758 | unsigned long flags; | |
da1df50d | 2759 | struct rcu_data *rdp = raw_cpu_ptr(&rcu_data); |
26d950a9 | 2760 | struct rcu_node *rnp = rdp->mynode; |
32aa2f41 | 2761 | const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist); |
64db4cff | 2762 | |
b049fdf8 PM |
2763 | if (cpu_is_offline(smp_processor_id())) |
2764 | return; | |
2765 | trace_rcu_utilization(TPS("Start RCU core")); | |
50dc7def | 2766 | WARN_ON_ONCE(!rdp->beenonline); |
2e597558 | 2767 | |
3e310098 | 2768 | /* Report any deferred quiescent states if preemption enabled. */ |
fced9c8c | 2769 | if (!(preempt_count() & PREEMPT_MASK)) { |
3e310098 | 2770 | rcu_preempt_deferred_qs(current); |
fced9c8c PM |
2771 | } else if (rcu_preempt_need_deferred_qs(current)) { |
2772 | set_tsk_need_resched(current); | |
2773 | set_preempt_need_resched(); | |
2774 | } | |
3e310098 | 2775 | |
64db4cff | 2776 | /* Update RCU state based on any recent quiescent states. */ |
8087d3e3 | 2777 | rcu_check_quiescent_state(rdp); |
64db4cff | 2778 | |
bd7af846 | 2779 | /* No grace period and unregistered callbacks? */ |
de8e8730 | 2780 | if (!rcu_gp_in_progress() && |
634954c2 FW |
2781 | rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) { |
2782 | rcu_nocb_lock_irqsave(rdp, flags); | |
e44e73ca | 2783 | if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL)) |
c6e09b97 | 2784 | rcu_accelerate_cbs_unlocked(rnp, rdp); |
634954c2 | 2785 | rcu_nocb_unlock_irqrestore(rdp, flags); |
64db4cff PM |
2786 | } |
2787 | ||
791416c4 | 2788 | rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check()); |
26d950a9 | 2789 | |
64db4cff | 2790 | /* If there are callbacks ready, invoke them. */ |
32aa2f41 | 2791 | if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) && |
43e903ad PM |
2792 | likely(READ_ONCE(rcu_scheduler_fully_active))) |
2793 | rcu_do_batch(rdp); | |
96d3fd0d PM |
2794 | |
2795 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2796 | do_nocb_deferred_wakeup(rdp); | |
f7f7bac9 | 2797 | trace_rcu_utilization(TPS("End RCU core")); |
a657f261 PM |
2798 | |
2799 | // If strict GPs, schedule an RCU reader in a clean environment. | |
2800 | if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) | |
2801 | queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work); | |
64db4cff PM |
2802 | } |
2803 | ||
48d07c04 SAS |
2804 | static void rcu_core_si(struct softirq_action *h) |
2805 | { | |
2806 | rcu_core(); | |
2807 | } | |
2808 | ||
2809 | static void rcu_wake_cond(struct task_struct *t, int status) | |
2810 | { | |
2811 | /* | |
2812 | * If the thread is yielding, only wake it when this | |
2813 | * is invoked from idle | |
2814 | */ | |
2815 | if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current))) | |
2816 | wake_up_process(t); | |
2817 | } | |
2818 | ||
2819 | static void invoke_rcu_core_kthread(void) | |
2820 | { | |
2821 | struct task_struct *t; | |
2822 | unsigned long flags; | |
2823 | ||
2824 | local_irq_save(flags); | |
2825 | __this_cpu_write(rcu_data.rcu_cpu_has_work, 1); | |
2826 | t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task); | |
2827 | if (t != NULL && t != current) | |
2828 | rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status)); | |
2829 | local_irq_restore(flags); | |
2830 | } | |
2831 | ||
48d07c04 SAS |
2832 | /* |
2833 | * Wake up this CPU's rcuc kthread to do RCU core processing. | |
2834 | */ | |
a46e0899 | 2835 | static void invoke_rcu_core(void) |
09223371 | 2836 | { |
48d07c04 SAS |
2837 | if (!cpu_online(smp_processor_id())) |
2838 | return; | |
2839 | if (use_softirq) | |
b0f74036 | 2840 | raise_softirq(RCU_SOFTIRQ); |
48d07c04 SAS |
2841 | else |
2842 | invoke_rcu_core_kthread(); | |
2843 | } | |
2844 | ||
2845 | static void rcu_cpu_kthread_park(unsigned int cpu) | |
2846 | { | |
2847 | per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; | |
2848 | } | |
2849 | ||
2850 | static int rcu_cpu_kthread_should_run(unsigned int cpu) | |
2851 | { | |
2852 | return __this_cpu_read(rcu_data.rcu_cpu_has_work); | |
2853 | } | |
2854 | ||
2855 | /* | |
2856 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces | |
2857 | * the RCU softirq used in configurations of RCU that do not support RCU | |
2858 | * priority boosting. | |
2859 | */ | |
2860 | static void rcu_cpu_kthread(unsigned int cpu) | |
2861 | { | |
2862 | unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status); | |
2863 | char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work); | |
2864 | int spincnt; | |
2865 | ||
2488a5e6 | 2866 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_run")); |
48d07c04 | 2867 | for (spincnt = 0; spincnt < 10; spincnt++) { |
48d07c04 SAS |
2868 | local_bh_disable(); |
2869 | *statusp = RCU_KTHREAD_RUNNING; | |
2870 | local_irq_disable(); | |
2871 | work = *workp; | |
2872 | *workp = 0; | |
2873 | local_irq_enable(); | |
2874 | if (work) | |
2875 | rcu_core(); | |
2876 | local_bh_enable(); | |
2877 | if (*workp == 0) { | |
2878 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); | |
2879 | *statusp = RCU_KTHREAD_WAITING; | |
2880 | return; | |
2881 | } | |
2882 | } | |
2883 | *statusp = RCU_KTHREAD_YIELDING; | |
2884 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); | |
77865dea | 2885 | schedule_timeout_idle(2); |
48d07c04 SAS |
2886 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); |
2887 | *statusp = RCU_KTHREAD_WAITING; | |
2888 | } | |
2889 | ||
2890 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { | |
2891 | .store = &rcu_data.rcu_cpu_kthread_task, | |
2892 | .thread_should_run = rcu_cpu_kthread_should_run, | |
2893 | .thread_fn = rcu_cpu_kthread, | |
2894 | .thread_comm = "rcuc/%u", | |
2895 | .setup = rcu_cpu_kthread_setup, | |
2896 | .park = rcu_cpu_kthread_park, | |
2897 | }; | |
2898 | ||
2899 | /* | |
2900 | * Spawn per-CPU RCU core processing kthreads. | |
2901 | */ | |
2902 | static int __init rcu_spawn_core_kthreads(void) | |
2903 | { | |
2904 | int cpu; | |
2905 | ||
2906 | for_each_possible_cpu(cpu) | |
2907 | per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0; | |
2908 | if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq) | |
2909 | return 0; | |
2910 | WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec), | |
2911 | "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__); | |
2912 | return 0; | |
09223371 | 2913 | } |
48d07c04 | 2914 | early_initcall(rcu_spawn_core_kthreads); |
09223371 | 2915 | |
29154c57 PM |
2916 | /* |
2917 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2918 | */ | |
5c7d8967 PM |
2919 | static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head, |
2920 | unsigned long flags) | |
64db4cff | 2921 | { |
62fde6ed PM |
2922 | /* |
2923 | * If called from an extended quiescent state, invoke the RCU | |
2924 | * core in order to force a re-evaluation of RCU's idleness. | |
2925 | */ | |
9910affa | 2926 | if (!rcu_is_watching()) |
62fde6ed PM |
2927 | invoke_rcu_core(); |
2928 | ||
a16b7a69 | 2929 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2930 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2931 | return; |
64db4cff | 2932 | |
37c72e56 PM |
2933 | /* |
2934 | * Force the grace period if too many callbacks or too long waiting. | |
cd920e5a | 2935 | * Enforce hysteresis, and don't invoke rcu_force_quiescent_state() |
37c72e56 | 2936 | * if some other CPU has recently done so. Also, don't bother |
cd920e5a | 2937 | * invoking rcu_force_quiescent_state() if the newly enqueued callback |
37c72e56 PM |
2938 | * is the only one waiting for a grace period to complete. |
2939 | */ | |
15fecf89 PM |
2940 | if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) > |
2941 | rdp->qlen_last_fqs_check + qhimark)) { | |
b52573d2 PM |
2942 | |
2943 | /* Are we ignoring a completed grace period? */ | |
15cabdff | 2944 | note_gp_changes(rdp); |
b52573d2 PM |
2945 | |
2946 | /* Start a new grace period if one not already started. */ | |
de8e8730 | 2947 | if (!rcu_gp_in_progress()) { |
c6e09b97 | 2948 | rcu_accelerate_cbs_unlocked(rdp->mynode, rdp); |
b52573d2 PM |
2949 | } else { |
2950 | /* Give the grace period a kick. */ | |
d5a9a8c3 | 2951 | rdp->blimit = DEFAULT_MAX_RCU_BLIMIT; |
5c7d8967 | 2952 | if (rcu_state.n_force_qs == rdp->n_force_qs_snap && |
15fecf89 | 2953 | rcu_segcblist_first_pend_cb(&rdp->cblist) != head) |
cd920e5a | 2954 | rcu_force_quiescent_state(); |
5c7d8967 | 2955 | rdp->n_force_qs_snap = rcu_state.n_force_qs; |
15fecf89 | 2956 | rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist); |
b52573d2 | 2957 | } |
4cdfc175 | 2958 | } |
29154c57 PM |
2959 | } |
2960 | ||
ae150184 PM |
2961 | /* |
2962 | * RCU callback function to leak a callback. | |
2963 | */ | |
2964 | static void rcu_leak_callback(struct rcu_head *rhp) | |
2965 | { | |
2966 | } | |
2967 | ||
3fbfbf7a | 2968 | /* |
b2b00ddf PM |
2969 | * Check and if necessary update the leaf rcu_node structure's |
2970 | * ->cbovldmask bit corresponding to the current CPU based on that CPU's | |
2971 | * number of queued RCU callbacks. The caller must hold the leaf rcu_node | |
2972 | * structure's ->lock. | |
3fbfbf7a | 2973 | */ |
b2b00ddf PM |
2974 | static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp) |
2975 | { | |
2976 | raw_lockdep_assert_held_rcu_node(rnp); | |
2977 | if (qovld_calc <= 0) | |
2978 | return; // Early boot and wildcard value set. | |
2979 | if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) | |
2980 | WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask); | |
2981 | else | |
2982 | WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask); | |
2983 | } | |
2984 | ||
2985 | /* | |
2986 | * Check and if necessary update the leaf rcu_node structure's | |
2987 | * ->cbovldmask bit corresponding to the current CPU based on that CPU's | |
2988 | * number of queued RCU callbacks. No locks need be held, but the | |
2989 | * caller must have disabled interrupts. | |
2990 | * | |
2991 | * Note that this function ignores the possibility that there are a lot | |
2992 | * of callbacks all of which have already seen the end of their respective | |
2993 | * grace periods. This omission is due to the need for no-CBs CPUs to | |
2994 | * be holding ->nocb_lock to do this check, which is too heavy for a | |
2995 | * common-case operation. | |
3fbfbf7a | 2996 | */ |
b2b00ddf PM |
2997 | static void check_cb_ovld(struct rcu_data *rdp) |
2998 | { | |
2999 | struct rcu_node *const rnp = rdp->mynode; | |
3000 | ||
3001 | if (qovld_calc <= 0 || | |
3002 | ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) == | |
3003 | !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask))) | |
3004 | return; // Early boot wildcard value or already set correctly. | |
3005 | raw_spin_lock_rcu_node(rnp); | |
3006 | check_cb_ovld_locked(rdp, rnp); | |
3007 | raw_spin_unlock_rcu_node(rnp); | |
3008 | } | |
3009 | ||
b692dc4a | 3010 | /* Helper function for call_rcu() and friends. */ |
64db4cff | 3011 | static void |
77a40f97 | 3012 | __call_rcu(struct rcu_head *head, rcu_callback_t func) |
64db4cff | 3013 | { |
b4b7914a | 3014 | static atomic_t doublefrees; |
64db4cff PM |
3015 | unsigned long flags; |
3016 | struct rcu_data *rdp; | |
5d6742b3 | 3017 | bool was_alldone; |
64db4cff | 3018 | |
b8f2ed53 PM |
3019 | /* Misaligned rcu_head! */ |
3020 | WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1)); | |
3021 | ||
ae150184 | 3022 | if (debug_rcu_head_queue(head)) { |
fa3c6647 PM |
3023 | /* |
3024 | * Probable double call_rcu(), so leak the callback. | |
3025 | * Use rcu:rcu_callback trace event to find the previous | |
3026 | * time callback was passed to __call_rcu(). | |
3027 | */ | |
b4b7914a PM |
3028 | if (atomic_inc_return(&doublefrees) < 4) { |
3029 | pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func); | |
3030 | mem_dump_obj(head); | |
3031 | } | |
7d0ae808 | 3032 | WRITE_ONCE(head->func, rcu_leak_callback); |
ae150184 PM |
3033 | return; |
3034 | } | |
64db4cff PM |
3035 | head->func = func; |
3036 | head->next = NULL; | |
64db4cff | 3037 | local_irq_save(flags); |
26e760c9 | 3038 | kasan_record_aux_stack(head); |
da1df50d | 3039 | rdp = this_cpu_ptr(&rcu_data); |
64db4cff PM |
3040 | |
3041 | /* Add the callback to our list. */ | |
5d6742b3 PM |
3042 | if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) { |
3043 | // This can trigger due to call_rcu() from offline CPU: | |
3044 | WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE); | |
34404ca8 | 3045 | WARN_ON_ONCE(!rcu_is_watching()); |
5d6742b3 PM |
3046 | // Very early boot, before rcu_init(). Initialize if needed |
3047 | // and then drop through to queue the callback. | |
15fecf89 PM |
3048 | if (rcu_segcblist_empty(&rdp->cblist)) |
3049 | rcu_segcblist_init(&rdp->cblist); | |
0d8ee37e | 3050 | } |
77a40f97 | 3051 | |
b2b00ddf | 3052 | check_cb_ovld(rdp); |
d1b222c6 PM |
3053 | if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags)) |
3054 | return; // Enqueued onto ->nocb_bypass, so just leave. | |
b692dc4a | 3055 | // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock. |
77a40f97 | 3056 | rcu_segcblist_enqueue(&rdp->cblist, head); |
c408b215 URS |
3057 | if (__is_kvfree_rcu_offset((unsigned long)func)) |
3058 | trace_rcu_kvfree_callback(rcu_state.name, head, | |
3c779dfe | 3059 | (unsigned long)func, |
15fecf89 | 3060 | rcu_segcblist_n_cbs(&rdp->cblist)); |
d4c08f2a | 3061 | else |
3c779dfe | 3062 | trace_rcu_callback(rcu_state.name, head, |
15fecf89 | 3063 | rcu_segcblist_n_cbs(&rdp->cblist)); |
d4c08f2a | 3064 | |
3afe7fa5 JFG |
3065 | trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued")); |
3066 | ||
29154c57 | 3067 | /* Go handle any RCU core processing required. */ |
3820b513 | 3068 | if (unlikely(rcu_rdp_is_offloaded(rdp))) { |
5d6742b3 PM |
3069 | __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */ |
3070 | } else { | |
3071 | __call_rcu_core(rdp, head, flags); | |
3072 | local_irq_restore(flags); | |
3073 | } | |
64db4cff PM |
3074 | } |
3075 | ||
a68a2bb2 | 3076 | /** |
45975c7d | 3077 | * call_rcu() - Queue an RCU callback for invocation after a grace period. |
a68a2bb2 PM |
3078 | * @head: structure to be used for queueing the RCU updates. |
3079 | * @func: actual callback function to be invoked after the grace period | |
3080 | * | |
3081 | * The callback function will be invoked some time after a full grace | |
45975c7d PM |
3082 | * period elapses, in other words after all pre-existing RCU read-side |
3083 | * critical sections have completed. However, the callback function | |
3084 | * might well execute concurrently with RCU read-side critical sections | |
3085 | * that started after call_rcu() was invoked. RCU read-side critical | |
3086 | * sections are delimited by rcu_read_lock() and rcu_read_unlock(), and | |
3087 | * may be nested. In addition, regions of code across which interrupts, | |
3088 | * preemption, or softirqs have been disabled also serve as RCU read-side | |
3089 | * critical sections. This includes hardware interrupt handlers, softirq | |
3090 | * handlers, and NMI handlers. | |
3091 | * | |
3092 | * Note that all CPUs must agree that the grace period extended beyond | |
3093 | * all pre-existing RCU read-side critical section. On systems with more | |
3094 | * than one CPU, this means that when "func()" is invoked, each CPU is | |
3095 | * guaranteed to have executed a full memory barrier since the end of its | |
3096 | * last RCU read-side critical section whose beginning preceded the call | |
3097 | * to call_rcu(). It also means that each CPU executing an RCU read-side | |
3098 | * critical section that continues beyond the start of "func()" must have | |
3099 | * executed a memory barrier after the call_rcu() but before the beginning | |
3100 | * of that RCU read-side critical section. Note that these guarantees | |
3101 | * include CPUs that are offline, idle, or executing in user mode, as | |
3102 | * well as CPUs that are executing in the kernel. | |
3103 | * | |
3104 | * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the | |
3105 | * resulting RCU callback function "func()", then both CPU A and CPU B are | |
3106 | * guaranteed to execute a full memory barrier during the time interval | |
3107 | * between the call to call_rcu() and the invocation of "func()" -- even | |
3108 | * if CPU A and CPU B are the same CPU (but again only if the system has | |
3109 | * more than one CPU). | |
3110 | */ | |
3111 | void call_rcu(struct rcu_head *head, rcu_callback_t func) | |
3112 | { | |
77a40f97 | 3113 | __call_rcu(head, func); |
45975c7d PM |
3114 | } |
3115 | EXPORT_SYMBOL_GPL(call_rcu); | |
64db4cff | 3116 | |
a35d1690 BP |
3117 | |
3118 | /* Maximum number of jiffies to wait before draining a batch. */ | |
3119 | #define KFREE_DRAIN_JIFFIES (HZ / 50) | |
0392bebe | 3120 | #define KFREE_N_BATCHES 2 |
5f3c8d62 | 3121 | #define FREE_N_CHANNELS 2 |
34c88174 URS |
3122 | |
3123 | /** | |
5f3c8d62 | 3124 | * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers |
34c88174 | 3125 | * @nr_records: Number of active pointers in the array |
34c88174 | 3126 | * @next: Next bulk object in the block chain |
5f3c8d62 | 3127 | * @records: Array of the kvfree_rcu() pointers |
34c88174 | 3128 | */ |
5f3c8d62 | 3129 | struct kvfree_rcu_bulk_data { |
34c88174 | 3130 | unsigned long nr_records; |
5f3c8d62 | 3131 | struct kvfree_rcu_bulk_data *next; |
3af84862 | 3132 | void *records[]; |
34c88174 URS |
3133 | }; |
3134 | ||
3af84862 URS |
3135 | /* |
3136 | * This macro defines how many entries the "records" array | |
3137 | * will contain. It is based on the fact that the size of | |
5f3c8d62 | 3138 | * kvfree_rcu_bulk_data structure becomes exactly one page. |
3af84862 | 3139 | */ |
5f3c8d62 URS |
3140 | #define KVFREE_BULK_MAX_ENTR \ |
3141 | ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *)) | |
3af84862 | 3142 | |
a35d1690 | 3143 | /** |
0392bebe | 3144 | * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests |
a35d1690 | 3145 | * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period |
0392bebe | 3146 | * @head_free: List of kfree_rcu() objects waiting for a grace period |
5f3c8d62 | 3147 | * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period |
0392bebe JFG |
3148 | * @krcp: Pointer to @kfree_rcu_cpu structure |
3149 | */ | |
3150 | ||
3151 | struct kfree_rcu_cpu_work { | |
3152 | struct rcu_work rcu_work; | |
3153 | struct rcu_head *head_free; | |
5f3c8d62 | 3154 | struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS]; |
0392bebe JFG |
3155 | struct kfree_rcu_cpu *krcp; |
3156 | }; | |
3157 | ||
3158 | /** | |
3159 | * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period | |
a35d1690 | 3160 | * @head: List of kfree_rcu() objects not yet waiting for a grace period |
5f3c8d62 | 3161 | * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period |
0392bebe | 3162 | * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period |
a35d1690 BP |
3163 | * @lock: Synchronize access to this structure |
3164 | * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES | |
3165 | * @monitor_todo: Tracks whether a @monitor_work delayed work is pending | |
69f08d39 | 3166 | * @initialized: The @rcu_work fields have been initialized |
8e11690d | 3167 | * @count: Number of objects for which GP not started |
72a2fbda MCC |
3168 | * @bkvcache: |
3169 | * A simple cache list that contains objects for reuse purpose. | |
3170 | * In order to save some per-cpu space the list is singular. | |
3171 | * Even though it is lockless an access has to be protected by the | |
3172 | * per-cpu lock. | |
56292e86 URS |
3173 | * @page_cache_work: A work to refill the cache when it is empty |
3174 | * @work_in_progress: Indicates that page_cache_work is running | |
3175 | * @hrtimer: A hrtimer for scheduling a page_cache_work | |
72a2fbda | 3176 | * @nr_bkv_objs: number of allocated objects at @bkvcache. |
a35d1690 BP |
3177 | * |
3178 | * This is a per-CPU structure. The reason that it is not included in | |
3179 | * the rcu_data structure is to permit this code to be extracted from | |
3180 | * the RCU files. Such extraction could allow further optimization of | |
3181 | * the interactions with the slab allocators. | |
3182 | */ | |
3183 | struct kfree_rcu_cpu { | |
a35d1690 | 3184 | struct rcu_head *head; |
5f3c8d62 | 3185 | struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS]; |
0392bebe | 3186 | struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES]; |
8ac88f71 | 3187 | raw_spinlock_t lock; |
a35d1690 | 3188 | struct delayed_work monitor_work; |
569d7670 | 3189 | bool monitor_todo; |
a35d1690 | 3190 | bool initialized; |
9154244c | 3191 | int count; |
56292e86 URS |
3192 | |
3193 | struct work_struct page_cache_work; | |
3194 | atomic_t work_in_progress; | |
3195 | struct hrtimer hrtimer; | |
3196 | ||
53c72b59 URS |
3197 | struct llist_head bkvcache; |
3198 | int nr_bkv_objs; | |
a35d1690 BP |
3199 | }; |
3200 | ||
69f08d39 SAS |
3201 | static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = { |
3202 | .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock), | |
3203 | }; | |
a35d1690 | 3204 | |
34c88174 | 3205 | static __always_inline void |
5f3c8d62 | 3206 | debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead) |
34c88174 URS |
3207 | { |
3208 | #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD | |
446044eb JFG |
3209 | int i; |
3210 | ||
3211 | for (i = 0; i < bhead->nr_records; i++) | |
3212 | debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i])); | |
34c88174 URS |
3213 | #endif |
3214 | } | |
3215 | ||
952371d6 URS |
3216 | static inline struct kfree_rcu_cpu * |
3217 | krc_this_cpu_lock(unsigned long *flags) | |
3218 | { | |
3219 | struct kfree_rcu_cpu *krcp; | |
3220 | ||
3221 | local_irq_save(*flags); // For safely calling this_cpu_ptr(). | |
3222 | krcp = this_cpu_ptr(&krc); | |
69f08d39 | 3223 | raw_spin_lock(&krcp->lock); |
952371d6 URS |
3224 | |
3225 | return krcp; | |
3226 | } | |
3227 | ||
3228 | static inline void | |
3229 | krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags) | |
3230 | { | |
7ffc9ec8 | 3231 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
952371d6 URS |
3232 | } |
3233 | ||
5f3c8d62 | 3234 | static inline struct kvfree_rcu_bulk_data * |
53c72b59 URS |
3235 | get_cached_bnode(struct kfree_rcu_cpu *krcp) |
3236 | { | |
3237 | if (!krcp->nr_bkv_objs) | |
3238 | return NULL; | |
3239 | ||
3240 | krcp->nr_bkv_objs--; | |
5f3c8d62 | 3241 | return (struct kvfree_rcu_bulk_data *) |
53c72b59 URS |
3242 | llist_del_first(&krcp->bkvcache); |
3243 | } | |
3244 | ||
3245 | static inline bool | |
3246 | put_cached_bnode(struct kfree_rcu_cpu *krcp, | |
5f3c8d62 | 3247 | struct kvfree_rcu_bulk_data *bnode) |
53c72b59 URS |
3248 | { |
3249 | // Check the limit. | |
3250 | if (krcp->nr_bkv_objs >= rcu_min_cached_objs) | |
3251 | return false; | |
3252 | ||
3253 | llist_add((struct llist_node *) bnode, &krcp->bkvcache); | |
3254 | krcp->nr_bkv_objs++; | |
3255 | return true; | |
3256 | ||
3257 | } | |
3258 | ||
495aa969 | 3259 | /* |
a35d1690 | 3260 | * This function is invoked in workqueue context after a grace period. |
34c88174 | 3261 | * It frees all the objects queued on ->bhead_free or ->head_free. |
495aa969 | 3262 | */ |
a35d1690 BP |
3263 | static void kfree_rcu_work(struct work_struct *work) |
3264 | { | |
3265 | unsigned long flags; | |
5f3c8d62 | 3266 | struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext; |
a35d1690 BP |
3267 | struct rcu_head *head, *next; |
3268 | struct kfree_rcu_cpu *krcp; | |
0392bebe | 3269 | struct kfree_rcu_cpu_work *krwp; |
5f3c8d62 | 3270 | int i, j; |
a35d1690 | 3271 | |
0392bebe JFG |
3272 | krwp = container_of(to_rcu_work(work), |
3273 | struct kfree_rcu_cpu_work, rcu_work); | |
3274 | krcp = krwp->krcp; | |
34c88174 | 3275 | |
8ac88f71 | 3276 | raw_spin_lock_irqsave(&krcp->lock, flags); |
5f3c8d62 URS |
3277 | // Channels 1 and 2. |
3278 | for (i = 0; i < FREE_N_CHANNELS; i++) { | |
3279 | bkvhead[i] = krwp->bkvhead_free[i]; | |
3280 | krwp->bkvhead_free[i] = NULL; | |
3281 | } | |
34c88174 | 3282 | |
5f3c8d62 | 3283 | // Channel 3. |
0392bebe JFG |
3284 | head = krwp->head_free; |
3285 | krwp->head_free = NULL; | |
8ac88f71 | 3286 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
a35d1690 | 3287 | |
5f3c8d62 URS |
3288 | // Handle two first channels. |
3289 | for (i = 0; i < FREE_N_CHANNELS; i++) { | |
3290 | for (; bkvhead[i]; bkvhead[i] = bnext) { | |
3291 | bnext = bkvhead[i]->next; | |
3292 | debug_rcu_bhead_unqueue(bkvhead[i]); | |
3293 | ||
3294 | rcu_lock_acquire(&rcu_callback_map); | |
3295 | if (i == 0) { // kmalloc() / kfree(). | |
3296 | trace_rcu_invoke_kfree_bulk_callback( | |
3297 | rcu_state.name, bkvhead[i]->nr_records, | |
3298 | bkvhead[i]->records); | |
3299 | ||
3300 | kfree_bulk(bkvhead[i]->nr_records, | |
3301 | bkvhead[i]->records); | |
3302 | } else { // vmalloc() / vfree(). | |
3303 | for (j = 0; j < bkvhead[i]->nr_records; j++) { | |
c408b215 | 3304 | trace_rcu_invoke_kvfree_callback( |
5f3c8d62 URS |
3305 | rcu_state.name, |
3306 | bkvhead[i]->records[j], 0); | |
3307 | ||
3308 | vfree(bkvhead[i]->records[j]); | |
3309 | } | |
3310 | } | |
3311 | rcu_lock_release(&rcu_callback_map); | |
61370792 | 3312 | |
56292e86 | 3313 | raw_spin_lock_irqsave(&krcp->lock, flags); |
5f3c8d62 URS |
3314 | if (put_cached_bnode(krcp, bkvhead[i])) |
3315 | bkvhead[i] = NULL; | |
56292e86 | 3316 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
34c88174 | 3317 | |
5f3c8d62 URS |
3318 | if (bkvhead[i]) |
3319 | free_page((unsigned long) bkvhead[i]); | |
34c88174 | 3320 | |
5f3c8d62 URS |
3321 | cond_resched_tasks_rcu_qs(); |
3322 | } | |
34c88174 URS |
3323 | } |
3324 | ||
3325 | /* | |
3326 | * Emergency case only. It can happen under low memory | |
3327 | * condition when an allocation gets failed, so the "bulk" | |
3328 | * path can not be temporary maintained. | |
3329 | */ | |
a35d1690 | 3330 | for (; head; head = next) { |
77a40f97 | 3331 | unsigned long offset = (unsigned long)head->func; |
446044eb | 3332 | void *ptr = (void *)head - offset; |
77a40f97 | 3333 | |
a35d1690 | 3334 | next = head->next; |
446044eb | 3335 | debug_rcu_head_unqueue((struct rcu_head *)ptr); |
77a40f97 | 3336 | rcu_lock_acquire(&rcu_callback_map); |
c408b215 | 3337 | trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset); |
77a40f97 | 3338 | |
c408b215 | 3339 | if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset))) |
5f3c8d62 | 3340 | kvfree(ptr); |
77a40f97 JFG |
3341 | |
3342 | rcu_lock_release(&rcu_callback_map); | |
a35d1690 BP |
3343 | cond_resched_tasks_rcu_qs(); |
3344 | } | |
3345 | } | |
3346 | ||
495aa969 | 3347 | /* |
a35d1690 BP |
3348 | * Schedule the kfree batch RCU work to run in workqueue context after a GP. |
3349 | * | |
3350 | * This function is invoked by kfree_rcu_monitor() when the KFREE_DRAIN_JIFFIES | |
3351 | * timeout has been reached. | |
3352 | */ | |
3353 | static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp) | |
3354 | { | |
34c88174 | 3355 | struct kfree_rcu_cpu_work *krwp; |
594aa597 | 3356 | bool repeat = false; |
5f3c8d62 | 3357 | int i, j; |
0392bebe | 3358 | |
a35d1690 BP |
3359 | lockdep_assert_held(&krcp->lock); |
3360 | ||
34c88174 URS |
3361 | for (i = 0; i < KFREE_N_BATCHES; i++) { |
3362 | krwp = &(krcp->krw_arr[i]); | |
a35d1690 | 3363 | |
34c88174 | 3364 | /* |
5f3c8d62 | 3365 | * Try to detach bkvhead or head and attach it over any |
34c88174 URS |
3366 | * available corresponding free channel. It can be that |
3367 | * a previous RCU batch is in progress, it means that | |
3368 | * immediately to queue another one is not possible so | |
3369 | * return false to tell caller to retry. | |
3370 | */ | |
5f3c8d62 URS |
3371 | if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) || |
3372 | (krcp->bkvhead[1] && !krwp->bkvhead_free[1]) || | |
34c88174 | 3373 | (krcp->head && !krwp->head_free)) { |
5f3c8d62 URS |
3374 | // Channel 1 corresponds to SLAB ptrs. |
3375 | // Channel 2 corresponds to vmalloc ptrs. | |
3376 | for (j = 0; j < FREE_N_CHANNELS; j++) { | |
3377 | if (!krwp->bkvhead_free[j]) { | |
3378 | krwp->bkvhead_free[j] = krcp->bkvhead[j]; | |
3379 | krcp->bkvhead[j] = NULL; | |
3380 | } | |
34c88174 URS |
3381 | } |
3382 | ||
5f3c8d62 | 3383 | // Channel 3 corresponds to emergency path. |
34c88174 URS |
3384 | if (!krwp->head_free) { |
3385 | krwp->head_free = krcp->head; | |
3386 | krcp->head = NULL; | |
3387 | } | |
3388 | ||
a6a82ce1 | 3389 | WRITE_ONCE(krcp->count, 0); |
9154244c | 3390 | |
34c88174 | 3391 | /* |
5f3c8d62 URS |
3392 | * One work is per one batch, so there are three |
3393 | * "free channels", the batch can handle. It can | |
3394 | * be that the work is in the pending state when | |
3395 | * channels have been detached following by each | |
3396 | * other. | |
34c88174 URS |
3397 | */ |
3398 | queue_rcu_work(system_wq, &krwp->rcu_work); | |
34c88174 | 3399 | } |
594aa597 | 3400 | |
5f3c8d62 URS |
3401 | // Repeat if any "free" corresponding channel is still busy. |
3402 | if (krcp->bkvhead[0] || krcp->bkvhead[1] || krcp->head) | |
594aa597 | 3403 | repeat = true; |
34c88174 URS |
3404 | } |
3405 | ||
594aa597 | 3406 | return !repeat; |
a35d1690 BP |
3407 | } |
3408 | ||
3409 | static inline void kfree_rcu_drain_unlock(struct kfree_rcu_cpu *krcp, | |
3410 | unsigned long flags) | |
3411 | { | |
3412 | // Attempt to start a new batch. | |
569d7670 | 3413 | krcp->monitor_todo = false; |
a35d1690 BP |
3414 | if (queue_kfree_rcu_work(krcp)) { |
3415 | // Success! Our job is done here. | |
8ac88f71 | 3416 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
a35d1690 BP |
3417 | return; |
3418 | } | |
3419 | ||
3420 | // Previous RCU batch still in progress, try again later. | |
569d7670 JF |
3421 | krcp->monitor_todo = true; |
3422 | schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES); | |
8ac88f71 | 3423 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
a35d1690 BP |
3424 | } |
3425 | ||
495aa969 | 3426 | /* |
a35d1690 BP |
3427 | * This function is invoked after the KFREE_DRAIN_JIFFIES timeout. |
3428 | * It invokes kfree_rcu_drain_unlock() to attempt to start another batch. | |
3429 | */ | |
3430 | static void kfree_rcu_monitor(struct work_struct *work) | |
3431 | { | |
3432 | unsigned long flags; | |
3433 | struct kfree_rcu_cpu *krcp = container_of(work, struct kfree_rcu_cpu, | |
3434 | monitor_work.work); | |
3435 | ||
8ac88f71 | 3436 | raw_spin_lock_irqsave(&krcp->lock, flags); |
569d7670 | 3437 | if (krcp->monitor_todo) |
a35d1690 BP |
3438 | kfree_rcu_drain_unlock(krcp, flags); |
3439 | else | |
8ac88f71 | 3440 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
a35d1690 BP |
3441 | } |
3442 | ||
56292e86 URS |
3443 | static enum hrtimer_restart |
3444 | schedule_page_work_fn(struct hrtimer *t) | |
3445 | { | |
3446 | struct kfree_rcu_cpu *krcp = | |
3447 | container_of(t, struct kfree_rcu_cpu, hrtimer); | |
3448 | ||
3449 | queue_work(system_highpri_wq, &krcp->page_cache_work); | |
3450 | return HRTIMER_NORESTART; | |
3451 | } | |
3452 | ||
3453 | static void fill_page_cache_func(struct work_struct *work) | |
3454 | { | |
3455 | struct kvfree_rcu_bulk_data *bnode; | |
3456 | struct kfree_rcu_cpu *krcp = | |
3457 | container_of(work, struct kfree_rcu_cpu, | |
3458 | page_cache_work); | |
3459 | unsigned long flags; | |
3460 | bool pushed; | |
3461 | int i; | |
3462 | ||
3463 | for (i = 0; i < rcu_min_cached_objs; i++) { | |
3464 | bnode = (struct kvfree_rcu_bulk_data *) | |
ee6ddf58 | 3465 | __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); |
56292e86 URS |
3466 | |
3467 | if (bnode) { | |
3468 | raw_spin_lock_irqsave(&krcp->lock, flags); | |
3469 | pushed = put_cached_bnode(krcp, bnode); | |
3470 | raw_spin_unlock_irqrestore(&krcp->lock, flags); | |
3471 | ||
3472 | if (!pushed) { | |
3473 | free_page((unsigned long) bnode); | |
3474 | break; | |
3475 | } | |
3476 | } | |
3477 | } | |
3478 | ||
3479 | atomic_set(&krcp->work_in_progress, 0); | |
3480 | } | |
3481 | ||
3482 | static void | |
3483 | run_page_cache_worker(struct kfree_rcu_cpu *krcp) | |
3484 | { | |
3485 | if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING && | |
3486 | !atomic_xchg(&krcp->work_in_progress, 1)) { | |
3487 | hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, | |
3488 | HRTIMER_MODE_REL); | |
3489 | krcp->hrtimer.function = schedule_page_work_fn; | |
3490 | hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL); | |
3491 | } | |
3492 | } | |
3493 | ||
148e3731 URS |
3494 | // Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock() |
3495 | // state specified by flags. If can_alloc is true, the caller must | |
3496 | // be schedulable and not be holding any locks or mutexes that might be | |
3497 | // acquired by the memory allocator or anything that it might invoke. | |
3498 | // Returns true if ptr was successfully recorded, else the caller must | |
3499 | // use a fallback. | |
34c88174 | 3500 | static inline bool |
148e3731 URS |
3501 | add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp, |
3502 | unsigned long *flags, void *ptr, bool can_alloc) | |
34c88174 | 3503 | { |
5f3c8d62 URS |
3504 | struct kvfree_rcu_bulk_data *bnode; |
3505 | int idx; | |
34c88174 | 3506 | |
148e3731 URS |
3507 | *krcp = krc_this_cpu_lock(flags); |
3508 | if (unlikely(!(*krcp)->initialized)) | |
34c88174 URS |
3509 | return false; |
3510 | ||
5f3c8d62 | 3511 | idx = !!is_vmalloc_addr(ptr); |
34c88174 URS |
3512 | |
3513 | /* Check if a new block is required. */ | |
148e3731 URS |
3514 | if (!(*krcp)->bkvhead[idx] || |
3515 | (*krcp)->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) { | |
3516 | bnode = get_cached_bnode(*krcp); | |
3517 | if (!bnode && can_alloc) { | |
3518 | krc_this_cpu_unlock(*krcp, *flags); | |
3e7ce7a1 URS |
3519 | |
3520 | // __GFP_NORETRY - allows a light-weight direct reclaim | |
3521 | // what is OK from minimizing of fallback hitting point of | |
3522 | // view. Apart of that it forbids any OOM invoking what is | |
3523 | // also beneficial since we are about to release memory soon. | |
3524 | // | |
3525 | // __GFP_NOMEMALLOC - prevents from consuming of all the | |
3526 | // memory reserves. Please note we have a fallback path. | |
3527 | // | |
3528 | // __GFP_NOWARN - it is supposed that an allocation can | |
3529 | // be failed under low memory or high memory pressure | |
3530 | // scenarios. | |
148e3731 | 3531 | bnode = (struct kvfree_rcu_bulk_data *) |
3e7ce7a1 | 3532 | __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); |
148e3731 URS |
3533 | *krcp = krc_this_cpu_lock(flags); |
3534 | } | |
3535 | ||
56292e86 | 3536 | if (!bnode) |
34c88174 URS |
3537 | return false; |
3538 | ||
3539 | /* Initialize the new block. */ | |
3540 | bnode->nr_records = 0; | |
148e3731 | 3541 | bnode->next = (*krcp)->bkvhead[idx]; |
34c88174 URS |
3542 | |
3543 | /* Attach it to the head. */ | |
148e3731 | 3544 | (*krcp)->bkvhead[idx] = bnode; |
34c88174 URS |
3545 | } |
3546 | ||
34c88174 | 3547 | /* Finally insert. */ |
148e3731 URS |
3548 | (*krcp)->bkvhead[idx]->records |
3549 | [(*krcp)->bkvhead[idx]->nr_records++] = ptr; | |
34c88174 URS |
3550 | |
3551 | return true; | |
3552 | } | |
3553 | ||
a35d1690 | 3554 | /* |
5f3c8d62 URS |
3555 | * Queue a request for lazy invocation of appropriate free routine after a |
3556 | * grace period. Please note there are three paths are maintained, two are the | |
3557 | * main ones that use array of pointers interface and third one is emergency | |
3558 | * one, that is used only when the main path can not be maintained temporary, | |
3559 | * due to memory pressure. | |
a35d1690 | 3560 | * |
c408b215 | 3561 | * Each kvfree_call_rcu() request is added to a batch. The batch will be drained |
34c88174 URS |
3562 | * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will |
3563 | * be free'd in workqueue context. This allows us to: batch requests together to | |
5f3c8d62 | 3564 | * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load. |
495aa969 | 3565 | */ |
c408b215 | 3566 | void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func) |
495aa969 | 3567 | { |
a35d1690 BP |
3568 | unsigned long flags; |
3569 | struct kfree_rcu_cpu *krcp; | |
3042f83f | 3570 | bool success; |
446044eb | 3571 | void *ptr; |
a35d1690 | 3572 | |
3042f83f URS |
3573 | if (head) { |
3574 | ptr = (void *) head - (unsigned long) func; | |
3575 | } else { | |
3576 | /* | |
3577 | * Please note there is a limitation for the head-less | |
3578 | * variant, that is why there is a clear rule for such | |
3579 | * objects: it can be used from might_sleep() context | |
3580 | * only. For other places please embed an rcu_head to | |
3581 | * your data. | |
3582 | */ | |
3583 | might_sleep(); | |
3584 | ptr = (unsigned long *) func; | |
3585 | } | |
3586 | ||
a35d1690 | 3587 | // Queue the object but don't yet schedule the batch. |
446044eb | 3588 | if (debug_rcu_head_queue(ptr)) { |
e99637be JFG |
3589 | // Probable double kfree_rcu(), just leak. |
3590 | WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n", | |
3591 | __func__, head); | |
3042f83f URS |
3592 | |
3593 | // Mark as success and leave. | |
148e3731 | 3594 | return; |
e99637be | 3595 | } |
34c88174 | 3596 | |
84109ab5 | 3597 | kasan_record_aux_stack(ptr); |
148e3731 | 3598 | success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head); |
3042f83f | 3599 | if (!success) { |
56292e86 URS |
3600 | run_page_cache_worker(krcp); |
3601 | ||
3042f83f URS |
3602 | if (head == NULL) |
3603 | // Inline if kvfree_rcu(one_arg) call. | |
3604 | goto unlock_return; | |
3605 | ||
34c88174 URS |
3606 | head->func = func; |
3607 | head->next = krcp->head; | |
3608 | krcp->head = head; | |
3042f83f | 3609 | success = true; |
34c88174 | 3610 | } |
a35d1690 | 3611 | |
a6a82ce1 | 3612 | WRITE_ONCE(krcp->count, krcp->count + 1); |
9154244c | 3613 | |
a35d1690 BP |
3614 | // Set timer to drain after KFREE_DRAIN_JIFFIES. |
3615 | if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING && | |
569d7670 JF |
3616 | !krcp->monitor_todo) { |
3617 | krcp->monitor_todo = true; | |
a35d1690 | 3618 | schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES); |
569d7670 | 3619 | } |
a35d1690 | 3620 | |
e99637be | 3621 | unlock_return: |
952371d6 | 3622 | krc_this_cpu_unlock(krcp, flags); |
3042f83f URS |
3623 | |
3624 | /* | |
3625 | * Inline kvfree() after synchronize_rcu(). We can do | |
3626 | * it from might_sleep() context only, so the current | |
3627 | * CPU can pass the QS state. | |
3628 | */ | |
3629 | if (!success) { | |
3630 | debug_rcu_head_unqueue((struct rcu_head *) ptr); | |
3631 | synchronize_rcu(); | |
3632 | kvfree(ptr); | |
3633 | } | |
495aa969 | 3634 | } |
c408b215 | 3635 | EXPORT_SYMBOL_GPL(kvfree_call_rcu); |
495aa969 | 3636 | |
9154244c JFG |
3637 | static unsigned long |
3638 | kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc) | |
3639 | { | |
3640 | int cpu; | |
a6a82ce1 | 3641 | unsigned long count = 0; |
9154244c JFG |
3642 | |
3643 | /* Snapshot count of all CPUs */ | |
70060b87 | 3644 | for_each_possible_cpu(cpu) { |
9154244c JFG |
3645 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); |
3646 | ||
a6a82ce1 | 3647 | count += READ_ONCE(krcp->count); |
9154244c JFG |
3648 | } |
3649 | ||
3650 | return count; | |
3651 | } | |
3652 | ||
3653 | static unsigned long | |
3654 | kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc) | |
3655 | { | |
3656 | int cpu, freed = 0; | |
3657 | unsigned long flags; | |
3658 | ||
70060b87 | 3659 | for_each_possible_cpu(cpu) { |
9154244c JFG |
3660 | int count; |
3661 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); | |
3662 | ||
3663 | count = krcp->count; | |
8ac88f71 | 3664 | raw_spin_lock_irqsave(&krcp->lock, flags); |
9154244c JFG |
3665 | if (krcp->monitor_todo) |
3666 | kfree_rcu_drain_unlock(krcp, flags); | |
3667 | else | |
8ac88f71 | 3668 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
9154244c JFG |
3669 | |
3670 | sc->nr_to_scan -= count; | |
3671 | freed += count; | |
3672 | ||
3673 | if (sc->nr_to_scan <= 0) | |
3674 | break; | |
3675 | } | |
3676 | ||
c6dfd72b | 3677 | return freed == 0 ? SHRINK_STOP : freed; |
9154244c JFG |
3678 | } |
3679 | ||
3680 | static struct shrinker kfree_rcu_shrinker = { | |
3681 | .count_objects = kfree_rcu_shrink_count, | |
3682 | .scan_objects = kfree_rcu_shrink_scan, | |
3683 | .batch = 0, | |
3684 | .seeks = DEFAULT_SEEKS, | |
3685 | }; | |
3686 | ||
a35d1690 BP |
3687 | void __init kfree_rcu_scheduler_running(void) |
3688 | { | |
3689 | int cpu; | |
3690 | unsigned long flags; | |
3691 | ||
70060b87 | 3692 | for_each_possible_cpu(cpu) { |
a35d1690 BP |
3693 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); |
3694 | ||
8ac88f71 | 3695 | raw_spin_lock_irqsave(&krcp->lock, flags); |
569d7670 | 3696 | if (!krcp->head || krcp->monitor_todo) { |
8ac88f71 | 3697 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
a35d1690 BP |
3698 | continue; |
3699 | } | |
569d7670 | 3700 | krcp->monitor_todo = true; |
0392bebe JFG |
3701 | schedule_delayed_work_on(cpu, &krcp->monitor_work, |
3702 | KFREE_DRAIN_JIFFIES); | |
8ac88f71 | 3703 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
a35d1690 BP |
3704 | } |
3705 | } | |
3706 | ||
e5bc3af7 PM |
3707 | /* |
3708 | * During early boot, any blocking grace-period wait automatically | |
90326f05 | 3709 | * implies a grace period. Later on, this is never the case for PREEMPTION. |
e5bc3af7 | 3710 | * |
354c3f0e | 3711 | * However, because a context switch is a grace period for !PREEMPTION, any |
e5bc3af7 PM |
3712 | * blocking grace-period wait automatically implies a grace period if |
3713 | * there is only one CPU online at any point time during execution of | |
3714 | * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to | |
3715 | * occasionally incorrectly indicate that there are multiple CPUs online | |
3716 | * when there was in fact only one the whole time, as this just adds some | |
3717 | * overhead: RCU still operates correctly. | |
3718 | */ | |
3719 | static int rcu_blocking_is_gp(void) | |
3720 | { | |
3721 | int ret; | |
3722 | ||
01b1d88b | 3723 | if (IS_ENABLED(CONFIG_PREEMPTION)) |
e5bc3af7 PM |
3724 | return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE; |
3725 | might_sleep(); /* Check for RCU read-side critical section. */ | |
3726 | preempt_disable(); | |
ed73860c NU |
3727 | /* |
3728 | * If the rcu_state.n_online_cpus counter is equal to one, | |
3729 | * there is only one CPU, and that CPU sees all prior accesses | |
3730 | * made by any CPU that was online at the time of its access. | |
3731 | * Furthermore, if this counter is equal to one, its value cannot | |
3732 | * change until after the preempt_enable() below. | |
3733 | * | |
3734 | * Furthermore, if rcu_state.n_online_cpus is equal to one here, | |
3735 | * all later CPUs (both this one and any that come online later | |
3736 | * on) are guaranteed to see all accesses prior to this point | |
3737 | * in the code, without the need for additional memory barriers. | |
3738 | * Those memory barriers are provided by CPU-hotplug code. | |
3739 | */ | |
3740 | ret = READ_ONCE(rcu_state.n_online_cpus) <= 1; | |
e5bc3af7 PM |
3741 | preempt_enable(); |
3742 | return ret; | |
3743 | } | |
3744 | ||
3745 | /** | |
3746 | * synchronize_rcu - wait until a grace period has elapsed. | |
3747 | * | |
3748 | * Control will return to the caller some time after a full grace | |
3749 | * period has elapsed, in other words after all currently executing RCU | |
3750 | * read-side critical sections have completed. Note, however, that | |
3751 | * upon return from synchronize_rcu(), the caller might well be executing | |
3752 | * concurrently with new RCU read-side critical sections that began while | |
3753 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
3754 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
3755 | * In addition, regions of code across which interrupts, preemption, or | |
3756 | * softirqs have been disabled also serve as RCU read-side critical | |
3757 | * sections. This includes hardware interrupt handlers, softirq handlers, | |
3758 | * and NMI handlers. | |
3759 | * | |
3760 | * Note that this guarantee implies further memory-ordering guarantees. | |
3761 | * On systems with more than one CPU, when synchronize_rcu() returns, | |
3762 | * each CPU is guaranteed to have executed a full memory barrier since | |
3763 | * the end of its last RCU read-side critical section whose beginning | |
3764 | * preceded the call to synchronize_rcu(). In addition, each CPU having | |
3765 | * an RCU read-side critical section that extends beyond the return from | |
3766 | * synchronize_rcu() is guaranteed to have executed a full memory barrier | |
3767 | * after the beginning of synchronize_rcu() and before the beginning of | |
3768 | * that RCU read-side critical section. Note that these guarantees include | |
3769 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
3770 | * that are executing in the kernel. | |
3771 | * | |
3772 | * Furthermore, if CPU A invoked synchronize_rcu(), which returned | |
3773 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
3774 | * to have executed a full memory barrier during the execution of | |
3775 | * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but | |
3776 | * again only if the system has more than one CPU). | |
3777 | */ | |
3778 | void synchronize_rcu(void) | |
3779 | { | |
3780 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || | |
3781 | lock_is_held(&rcu_lock_map) || | |
3782 | lock_is_held(&rcu_sched_lock_map), | |
3783 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
3784 | if (rcu_blocking_is_gp()) | |
ed73860c | 3785 | return; // Context allows vacuous grace periods. |
e5bc3af7 PM |
3786 | if (rcu_gp_is_expedited()) |
3787 | synchronize_rcu_expedited(); | |
3788 | else | |
3789 | wait_rcu_gp(call_rcu); | |
3790 | } | |
3791 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
3792 | ||
765a3f4f PM |
3793 | /** |
3794 | * get_state_synchronize_rcu - Snapshot current RCU state | |
3795 | * | |
3796 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
7abb18bd PM |
3797 | * or poll_state_synchronize_rcu() to determine whether or not a full |
3798 | * grace period has elapsed in the meantime. | |
765a3f4f PM |
3799 | */ |
3800 | unsigned long get_state_synchronize_rcu(void) | |
3801 | { | |
3802 | /* | |
3803 | * Any prior manipulation of RCU-protected data must happen | |
e4be81a2 | 3804 | * before the load from ->gp_seq. |
765a3f4f PM |
3805 | */ |
3806 | smp_mb(); /* ^^^ */ | |
16fc9c60 | 3807 | return rcu_seq_snap(&rcu_state.gp_seq); |
765a3f4f PM |
3808 | } |
3809 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
3810 | ||
7abb18bd PM |
3811 | /** |
3812 | * start_poll_synchronize_rcu - Snapshot and start RCU grace period | |
3813 | * | |
3814 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
3815 | * or poll_state_synchronize_rcu() to determine whether or not a full | |
3816 | * grace period has elapsed in the meantime. If the needed grace period | |
3817 | * is not already slated to start, notifies RCU core of the need for that | |
3818 | * grace period. | |
3819 | * | |
3820 | * Interrupts must be enabled for the case where it is necessary to awaken | |
3821 | * the grace-period kthread. | |
3822 | */ | |
3823 | unsigned long start_poll_synchronize_rcu(void) | |
3824 | { | |
3825 | unsigned long flags; | |
3826 | unsigned long gp_seq = get_state_synchronize_rcu(); | |
3827 | bool needwake; | |
3828 | struct rcu_data *rdp; | |
3829 | struct rcu_node *rnp; | |
3830 | ||
3831 | lockdep_assert_irqs_enabled(); | |
3832 | local_irq_save(flags); | |
3833 | rdp = this_cpu_ptr(&rcu_data); | |
3834 | rnp = rdp->mynode; | |
3835 | raw_spin_lock_rcu_node(rnp); // irqs already disabled. | |
3836 | needwake = rcu_start_this_gp(rnp, rdp, gp_seq); | |
3837 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
3838 | if (needwake) | |
3839 | rcu_gp_kthread_wake(); | |
3840 | return gp_seq; | |
3841 | } | |
3842 | EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu); | |
3843 | ||
3844 | /** | |
3845 | * poll_state_synchronize_rcu - Conditionally wait for an RCU grace period | |
3846 | * | |
3847 | * @oldstate: return from call to get_state_synchronize_rcu() or start_poll_synchronize_rcu() | |
3848 | * | |
3849 | * If a full RCU grace period has elapsed since the earlier call from | |
3850 | * which oldstate was obtained, return @true, otherwise return @false. | |
a616aec9 | 3851 | * If @false is returned, it is the caller's responsibility to invoke this |
7abb18bd PM |
3852 | * function later on until it does return @true. Alternatively, the caller |
3853 | * can explicitly wait for a grace period, for example, by passing @oldstate | |
3854 | * to cond_synchronize_rcu() or by directly invoking synchronize_rcu(). | |
3855 | * | |
3856 | * Yes, this function does not take counter wrap into account. | |
3857 | * But counter wrap is harmless. If the counter wraps, we have waited for | |
3858 | * more than 2 billion grace periods (and way more on a 64-bit system!). | |
3859 | * Those needing to keep oldstate values for very long time periods | |
3860 | * (many hours even on 32-bit systems) should check them occasionally | |
3861 | * and either refresh them or set a flag indicating that the grace period | |
3862 | * has completed. | |
3863 | */ | |
3864 | bool poll_state_synchronize_rcu(unsigned long oldstate) | |
3865 | { | |
3866 | if (rcu_seq_done(&rcu_state.gp_seq, oldstate)) { | |
3867 | smp_mb(); /* Ensure GP ends before subsequent accesses. */ | |
3868 | return true; | |
3869 | } | |
3870 | return false; | |
3871 | } | |
3872 | EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu); | |
3873 | ||
765a3f4f PM |
3874 | /** |
3875 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
3876 | * | |
3877 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() | |
3878 | * | |
3879 | * If a full RCU grace period has elapsed since the earlier call to | |
7abb18bd PM |
3880 | * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return. |
3881 | * Otherwise, invoke synchronize_rcu() to wait for a full grace period. | |
765a3f4f PM |
3882 | * |
3883 | * Yes, this function does not take counter wrap into account. But | |
3884 | * counter wrap is harmless. If the counter wraps, we have waited for | |
3885 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
3886 | * so waiting for one additional grace period should be just fine. | |
3887 | */ | |
3888 | void cond_synchronize_rcu(unsigned long oldstate) | |
3889 | { | |
7abb18bd | 3890 | if (!poll_state_synchronize_rcu(oldstate)) |
765a3f4f PM |
3891 | synchronize_rcu(); |
3892 | } | |
3893 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
3894 | ||
64db4cff | 3895 | /* |
98ece508 | 3896 | * Check to see if there is any immediate RCU-related work to be done by |
49918a54 PM |
3897 | * the current CPU, returning 1 if so and zero otherwise. The checks are |
3898 | * in order of increasing expense: checks that can be carried out against | |
3899 | * CPU-local state are performed first. However, we must check for CPU | |
3900 | * stalls first, else we might not get a chance. | |
64db4cff | 3901 | */ |
dd7dafd1 | 3902 | static int rcu_pending(int user) |
64db4cff | 3903 | { |
ed93dfc6 | 3904 | bool gp_in_progress; |
98ece508 | 3905 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
2f51f988 PM |
3906 | struct rcu_node *rnp = rdp->mynode; |
3907 | ||
a649d25d PM |
3908 | lockdep_assert_irqs_disabled(); |
3909 | ||
64db4cff | 3910 | /* Check for CPU stalls, if enabled. */ |
ea12ff2b | 3911 | check_cpu_stall(rdp); |
64db4cff | 3912 | |
85f69b32 | 3913 | /* Does this CPU need a deferred NOCB wakeup? */ |
87090516 | 3914 | if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE)) |
85f69b32 PM |
3915 | return 1; |
3916 | ||
dd7dafd1 PM |
3917 | /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */ |
3918 | if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu()) | |
a096932f PM |
3919 | return 0; |
3920 | ||
64db4cff | 3921 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
ed93dfc6 PM |
3922 | gp_in_progress = rcu_gp_in_progress(); |
3923 | if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress) | |
64db4cff PM |
3924 | return 1; |
3925 | ||
3926 | /* Does this CPU have callbacks ready to invoke? */ | |
3820b513 | 3927 | if (!rcu_rdp_is_offloaded(rdp) && |
bd56e0a4 | 3928 | rcu_segcblist_ready_cbs(&rdp->cblist)) |
64db4cff PM |
3929 | return 1; |
3930 | ||
3931 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
ed93dfc6 | 3932 | if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) && |
3820b513 | 3933 | !rcu_rdp_is_offloaded(rdp) && |
c1935209 | 3934 | !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL)) |
64db4cff PM |
3935 | return 1; |
3936 | ||
67e14c1e PM |
3937 | /* Have RCU grace period completed or started? */ |
3938 | if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq || | |
01c495f7 | 3939 | unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */ |
64db4cff PM |
3940 | return 1; |
3941 | ||
64db4cff PM |
3942 | /* nothing to do */ |
3943 | return 0; | |
3944 | } | |
3945 | ||
a83eff0a | 3946 | /* |
dd46a788 | 3947 | * Helper function for rcu_barrier() tracing. If tracing is disabled, |
a83eff0a PM |
3948 | * the compiler is expected to optimize this away. |
3949 | */ | |
dd46a788 | 3950 | static void rcu_barrier_trace(const char *s, int cpu, unsigned long done) |
a83eff0a | 3951 | { |
8344b871 PM |
3952 | trace_rcu_barrier(rcu_state.name, s, cpu, |
3953 | atomic_read(&rcu_state.barrier_cpu_count), done); | |
a83eff0a PM |
3954 | } |
3955 | ||
b1420f1c | 3956 | /* |
dd46a788 PM |
3957 | * RCU callback function for rcu_barrier(). If we are last, wake |
3958 | * up the task executing rcu_barrier(). | |
aa24f937 PM |
3959 | * |
3960 | * Note that the value of rcu_state.barrier_sequence must be captured | |
3961 | * before the atomic_dec_and_test(). Otherwise, if this CPU is not last, | |
3962 | * other CPUs might count the value down to zero before this CPU gets | |
3963 | * around to invoking rcu_barrier_trace(), which might result in bogus | |
3964 | * data from the next instance of rcu_barrier(). | |
b1420f1c | 3965 | */ |
24ebbca8 | 3966 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 3967 | { |
aa24f937 PM |
3968 | unsigned long __maybe_unused s = rcu_state.barrier_sequence; |
3969 | ||
ec9f5835 | 3970 | if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) { |
aa24f937 | 3971 | rcu_barrier_trace(TPS("LastCB"), -1, s); |
ec9f5835 | 3972 | complete(&rcu_state.barrier_completion); |
a83eff0a | 3973 | } else { |
aa24f937 | 3974 | rcu_barrier_trace(TPS("CB"), -1, s); |
a83eff0a | 3975 | } |
d0ec774c PM |
3976 | } |
3977 | ||
3978 | /* | |
3979 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
3980 | */ | |
127e2981 | 3981 | static void rcu_barrier_func(void *cpu_in) |
d0ec774c | 3982 | { |
127e2981 PM |
3983 | uintptr_t cpu = (uintptr_t)cpu_in; |
3984 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
d0ec774c | 3985 | |
dd46a788 | 3986 | rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence); |
f92c734f PM |
3987 | rdp->barrier_head.func = rcu_barrier_callback; |
3988 | debug_rcu_head_queue(&rdp->barrier_head); | |
5d6742b3 | 3989 | rcu_nocb_lock(rdp); |
d1b222c6 | 3990 | WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies)); |
77a40f97 | 3991 | if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) { |
ec9f5835 | 3992 | atomic_inc(&rcu_state.barrier_cpu_count); |
f92c734f PM |
3993 | } else { |
3994 | debug_rcu_head_unqueue(&rdp->barrier_head); | |
dd46a788 | 3995 | rcu_barrier_trace(TPS("IRQNQ"), -1, |
66e4c33b | 3996 | rcu_state.barrier_sequence); |
f92c734f | 3997 | } |
5d6742b3 | 3998 | rcu_nocb_unlock(rdp); |
d0ec774c PM |
3999 | } |
4000 | ||
dd46a788 PM |
4001 | /** |
4002 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
4003 | * | |
4004 | * Note that this primitive does not necessarily wait for an RCU grace period | |
4005 | * to complete. For example, if there are no RCU callbacks queued anywhere | |
4006 | * in the system, then rcu_barrier() is within its rights to return | |
4007 | * immediately, without waiting for anything, much less an RCU grace period. | |
d0ec774c | 4008 | */ |
dd46a788 | 4009 | void rcu_barrier(void) |
d0ec774c | 4010 | { |
127e2981 | 4011 | uintptr_t cpu; |
b1420f1c | 4012 | struct rcu_data *rdp; |
ec9f5835 | 4013 | unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence); |
b1420f1c | 4014 | |
dd46a788 | 4015 | rcu_barrier_trace(TPS("Begin"), -1, s); |
b1420f1c | 4016 | |
e74f4c45 | 4017 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
ec9f5835 | 4018 | mutex_lock(&rcu_state.barrier_mutex); |
b1420f1c | 4019 | |
4f525a52 | 4020 | /* Did someone else do our work for us? */ |
ec9f5835 | 4021 | if (rcu_seq_done(&rcu_state.barrier_sequence, s)) { |
dd46a788 | 4022 | rcu_barrier_trace(TPS("EarlyExit"), -1, |
66e4c33b | 4023 | rcu_state.barrier_sequence); |
cf3a9c48 | 4024 | smp_mb(); /* caller's subsequent code after above check. */ |
ec9f5835 | 4025 | mutex_unlock(&rcu_state.barrier_mutex); |
cf3a9c48 PM |
4026 | return; |
4027 | } | |
4028 | ||
4f525a52 | 4029 | /* Mark the start of the barrier operation. */ |
ec9f5835 | 4030 | rcu_seq_start(&rcu_state.barrier_sequence); |
dd46a788 | 4031 | rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence); |
b1420f1c | 4032 | |
d0ec774c | 4033 | /* |
127e2981 PM |
4034 | * Initialize the count to two rather than to zero in order |
4035 | * to avoid a too-soon return to zero in case of an immediate | |
4036 | * invocation of the just-enqueued callback (or preemption of | |
4037 | * this task). Exclude CPU-hotplug operations to ensure that no | |
4038 | * offline non-offloaded CPU has callbacks queued. | |
d0ec774c | 4039 | */ |
ec9f5835 | 4040 | init_completion(&rcu_state.barrier_completion); |
127e2981 | 4041 | atomic_set(&rcu_state.barrier_cpu_count, 2); |
1331e7a1 | 4042 | get_online_cpus(); |
b1420f1c PM |
4043 | |
4044 | /* | |
1331e7a1 PM |
4045 | * Force each CPU with callbacks to register a new callback. |
4046 | * When that callback is invoked, we will know that all of the | |
4047 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 4048 | */ |
3fbfbf7a | 4049 | for_each_possible_cpu(cpu) { |
da1df50d | 4050 | rdp = per_cpu_ptr(&rcu_data, cpu); |
127e2981 | 4051 | if (cpu_is_offline(cpu) && |
3820b513 | 4052 | !rcu_rdp_is_offloaded(rdp)) |
ce5215c1 | 4053 | continue; |
127e2981 | 4054 | if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) { |
dd46a788 | 4055 | rcu_barrier_trace(TPS("OnlineQ"), cpu, |
66e4c33b | 4056 | rcu_state.barrier_sequence); |
127e2981 PM |
4057 | smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1); |
4058 | } else if (rcu_segcblist_n_cbs(&rdp->cblist) && | |
4059 | cpu_is_offline(cpu)) { | |
4060 | rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, | |
4061 | rcu_state.barrier_sequence); | |
4062 | local_irq_disable(); | |
4063 | rcu_barrier_func((void *)cpu); | |
4064 | local_irq_enable(); | |
4065 | } else if (cpu_is_offline(cpu)) { | |
4066 | rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu, | |
4067 | rcu_state.barrier_sequence); | |
b1420f1c | 4068 | } else { |
dd46a788 | 4069 | rcu_barrier_trace(TPS("OnlineNQ"), cpu, |
66e4c33b | 4070 | rcu_state.barrier_sequence); |
b1420f1c PM |
4071 | } |
4072 | } | |
1331e7a1 | 4073 | put_online_cpus(); |
b1420f1c PM |
4074 | |
4075 | /* | |
4076 | * Now that we have an rcu_barrier_callback() callback on each | |
4077 | * CPU, and thus each counted, remove the initial count. | |
4078 | */ | |
127e2981 | 4079 | if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count)) |
ec9f5835 | 4080 | complete(&rcu_state.barrier_completion); |
b1420f1c PM |
4081 | |
4082 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ | |
ec9f5835 | 4083 | wait_for_completion(&rcu_state.barrier_completion); |
b1420f1c | 4084 | |
4f525a52 | 4085 | /* Mark the end of the barrier operation. */ |
dd46a788 | 4086 | rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence); |
ec9f5835 | 4087 | rcu_seq_end(&rcu_state.barrier_sequence); |
4f525a52 | 4088 | |
b1420f1c | 4089 | /* Other rcu_barrier() invocations can now safely proceed. */ |
ec9f5835 | 4090 | mutex_unlock(&rcu_state.barrier_mutex); |
d0ec774c | 4091 | } |
45975c7d | 4092 | EXPORT_SYMBOL_GPL(rcu_barrier); |
d0ec774c | 4093 | |
0aa04b05 PM |
4094 | /* |
4095 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the | |
4096 | * first CPU in a given leaf rcu_node structure coming online. The caller | |
a616aec9 | 4097 | * must hold the corresponding leaf rcu_node ->lock with interrupts |
0aa04b05 PM |
4098 | * disabled. |
4099 | */ | |
4100 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) | |
4101 | { | |
4102 | long mask; | |
8d672fa6 | 4103 | long oldmask; |
0aa04b05 PM |
4104 | struct rcu_node *rnp = rnp_leaf; |
4105 | ||
8d672fa6 | 4106 | raw_lockdep_assert_held_rcu_node(rnp_leaf); |
962aff03 | 4107 | WARN_ON_ONCE(rnp->wait_blkd_tasks); |
0aa04b05 PM |
4108 | for (;;) { |
4109 | mask = rnp->grpmask; | |
4110 | rnp = rnp->parent; | |
4111 | if (rnp == NULL) | |
4112 | return; | |
6cf10081 | 4113 | raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */ |
8d672fa6 | 4114 | oldmask = rnp->qsmaskinit; |
0aa04b05 | 4115 | rnp->qsmaskinit |= mask; |
67c583a7 | 4116 | raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */ |
8d672fa6 PM |
4117 | if (oldmask) |
4118 | return; | |
0aa04b05 PM |
4119 | } |
4120 | } | |
4121 | ||
64db4cff | 4122 | /* |
27569620 | 4123 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 4124 | */ |
27569620 | 4125 | static void __init |
53b46303 | 4126 | rcu_boot_init_percpu_data(int cpu) |
64db4cff | 4127 | { |
da1df50d | 4128 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
27569620 PM |
4129 | |
4130 | /* Set up local state, ensuring consistent view of global state. */ | |
bc75e999 | 4131 | rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu); |
a657f261 | 4132 | INIT_WORK(&rdp->strict_work, strict_work_handler); |
4c5273bf | 4133 | WARN_ON_ONCE(rdp->dynticks_nesting != 1); |
dc5a4f29 | 4134 | WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp))); |
53b46303 | 4135 | rdp->rcu_ofl_gp_seq = rcu_state.gp_seq; |
57738942 | 4136 | rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED; |
53b46303 | 4137 | rdp->rcu_onl_gp_seq = rcu_state.gp_seq; |
57738942 | 4138 | rdp->rcu_onl_gp_flags = RCU_GP_CLEANED; |
27569620 | 4139 | rdp->cpu = cpu; |
3fbfbf7a | 4140 | rcu_boot_init_nocb_percpu_data(rdp); |
27569620 PM |
4141 | } |
4142 | ||
4143 | /* | |
53b46303 PM |
4144 | * Invoked early in the CPU-online process, when pretty much all services |
4145 | * are available. The incoming CPU is not present. | |
4146 | * | |
4147 | * Initializes a CPU's per-CPU RCU data. Note that only one online or | |
ff3bb6f4 PM |
4148 | * offline event can be happening at a given time. Note also that we can |
4149 | * accept some slop in the rsp->gp_seq access due to the fact that this | |
e83e73f5 PM |
4150 | * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet. |
4151 | * And any offloaded callbacks are being numbered elsewhere. | |
64db4cff | 4152 | */ |
53b46303 | 4153 | int rcutree_prepare_cpu(unsigned int cpu) |
64db4cff PM |
4154 | { |
4155 | unsigned long flags; | |
da1df50d | 4156 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
336a4f6c | 4157 | struct rcu_node *rnp = rcu_get_root(); |
64db4cff PM |
4158 | |
4159 | /* Set up local state, ensuring consistent view of global state. */ | |
6cf10081 | 4160 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
37c72e56 | 4161 | rdp->qlen_last_fqs_check = 0; |
53b46303 | 4162 | rdp->n_force_qs_snap = rcu_state.n_force_qs; |
64db4cff | 4163 | rdp->blimit = blimit; |
4c5273bf | 4164 | rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */ |
2625d469 | 4165 | rcu_dynticks_eqs_online(); |
67c583a7 | 4166 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
ec711bc1 | 4167 | |
126d9d49 | 4168 | /* |
ec711bc1 FW |
4169 | * Only non-NOCB CPUs that didn't have early-boot callbacks need to be |
4170 | * (re-)initialized. | |
126d9d49 | 4171 | */ |
ec711bc1 | 4172 | if (!rcu_segcblist_is_enabled(&rdp->cblist)) |
126d9d49 | 4173 | rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */ |
64db4cff | 4174 | |
0aa04b05 PM |
4175 | /* |
4176 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed | |
4177 | * propagation up the rcu_node tree will happen at the beginning | |
4178 | * of the next grace period. | |
4179 | */ | |
64db4cff | 4180 | rnp = rdp->mynode; |
2a67e741 | 4181 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
b9585e94 | 4182 | rdp->beenonline = true; /* We have now been online. */ |
8ff37290 PM |
4183 | rdp->gp_seq = READ_ONCE(rnp->gp_seq); |
4184 | rdp->gp_seq_needed = rdp->gp_seq; | |
5b74c458 | 4185 | rdp->cpu_no_qs.b.norm = true; |
97c668b8 | 4186 | rdp->core_needs_qs = false; |
9b9500da | 4187 | rdp->rcu_iw_pending = false; |
7a9f50a0 | 4188 | rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler); |
8ff37290 | 4189 | rdp->rcu_iw_gp_seq = rdp->gp_seq - 1; |
53b46303 | 4190 | trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl")); |
67c583a7 | 4191 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
4df83742 | 4192 | rcu_prepare_kthreads(cpu); |
ad368d15 | 4193 | rcu_spawn_cpu_nocb_kthread(cpu); |
ed73860c | 4194 | WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1); |
4df83742 TG |
4195 | |
4196 | return 0; | |
4197 | } | |
4198 | ||
deb34f36 PM |
4199 | /* |
4200 | * Update RCU priority boot kthread affinity for CPU-hotplug changes. | |
4201 | */ | |
4df83742 TG |
4202 | static void rcutree_affinity_setting(unsigned int cpu, int outgoing) |
4203 | { | |
da1df50d | 4204 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
4df83742 TG |
4205 | |
4206 | rcu_boost_kthread_setaffinity(rdp->mynode, outgoing); | |
4207 | } | |
4208 | ||
deb34f36 PM |
4209 | /* |
4210 | * Near the end of the CPU-online process. Pretty much all services | |
4211 | * enabled, and the CPU is now very much alive. | |
4212 | */ | |
4df83742 TG |
4213 | int rcutree_online_cpu(unsigned int cpu) |
4214 | { | |
9b9500da PM |
4215 | unsigned long flags; |
4216 | struct rcu_data *rdp; | |
4217 | struct rcu_node *rnp; | |
9b9500da | 4218 | |
b97d23c5 PM |
4219 | rdp = per_cpu_ptr(&rcu_data, cpu); |
4220 | rnp = rdp->mynode; | |
4221 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
4222 | rnp->ffmask |= rdp->grpmask; | |
4223 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
9b9500da PM |
4224 | if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE) |
4225 | return 0; /* Too early in boot for scheduler work. */ | |
4226 | sync_sched_exp_online_cleanup(cpu); | |
4227 | rcutree_affinity_setting(cpu, -1); | |
96926686 PM |
4228 | |
4229 | // Stop-machine done, so allow nohz_full to disable tick. | |
4230 | tick_dep_clear(TICK_DEP_BIT_RCU); | |
4df83742 TG |
4231 | return 0; |
4232 | } | |
4233 | ||
deb34f36 PM |
4234 | /* |
4235 | * Near the beginning of the process. The CPU is still very much alive | |
4236 | * with pretty much all services enabled. | |
4237 | */ | |
4df83742 TG |
4238 | int rcutree_offline_cpu(unsigned int cpu) |
4239 | { | |
9b9500da PM |
4240 | unsigned long flags; |
4241 | struct rcu_data *rdp; | |
4242 | struct rcu_node *rnp; | |
9b9500da | 4243 | |
b97d23c5 PM |
4244 | rdp = per_cpu_ptr(&rcu_data, cpu); |
4245 | rnp = rdp->mynode; | |
4246 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
4247 | rnp->ffmask &= ~rdp->grpmask; | |
4248 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
9b9500da | 4249 | |
4df83742 | 4250 | rcutree_affinity_setting(cpu, cpu); |
96926686 PM |
4251 | |
4252 | // nohz_full CPUs need the tick for stop-machine to work quickly | |
4253 | tick_dep_set(TICK_DEP_BIT_RCU); | |
4df83742 TG |
4254 | return 0; |
4255 | } | |
4256 | ||
7ec99de3 PM |
4257 | /* |
4258 | * Mark the specified CPU as being online so that subsequent grace periods | |
4259 | * (both expedited and normal) will wait on it. Note that this means that | |
4260 | * incoming CPUs are not allowed to use RCU read-side critical sections | |
4261 | * until this function is called. Failing to observe this restriction | |
4262 | * will result in lockdep splats. | |
deb34f36 PM |
4263 | * |
4264 | * Note that this function is special in that it is invoked directly | |
4265 | * from the incoming CPU rather than from the cpuhp_step mechanism. | |
4266 | * This is because this function must be invoked at a precise location. | |
7ec99de3 PM |
4267 | */ |
4268 | void rcu_cpu_starting(unsigned int cpu) | |
4269 | { | |
4270 | unsigned long flags; | |
4271 | unsigned long mask; | |
4272 | struct rcu_data *rdp; | |
4273 | struct rcu_node *rnp; | |
abfce041 | 4274 | bool newcpu; |
7ec99de3 | 4275 | |
c0f97f20 PM |
4276 | rdp = per_cpu_ptr(&rcu_data, cpu); |
4277 | if (rdp->cpu_started) | |
f64c6013 | 4278 | return; |
c0f97f20 | 4279 | rdp->cpu_started = true; |
f64c6013 | 4280 | |
b97d23c5 PM |
4281 | rnp = rdp->mynode; |
4282 | mask = rdp->grpmask; | |
4d60b475 PM |
4283 | WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1); |
4284 | WARN_ON_ONCE(!(rnp->ofl_seq & 0x1)); | |
4285 | smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier(). | |
b97d23c5 | 4286 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
105abf82 | 4287 | WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask); |
abfce041 | 4288 | newcpu = !(rnp->expmaskinitnext & mask); |
b97d23c5 | 4289 | rnp->expmaskinitnext |= mask; |
b97d23c5 | 4290 | /* Allow lockless access for expedited grace periods. */ |
abfce041 | 4291 | smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */ |
2f084695 | 4292 | ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus); |
b97d23c5 | 4293 | rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */ |
eb7a6653 PM |
4294 | rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq); |
4295 | rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags); | |
9f866dac JFG |
4296 | |
4297 | /* An incoming CPU should never be blocking a grace period. */ | |
4298 | if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */ | |
516e5ae0 | 4299 | rcu_disable_urgency_upon_qs(rdp); |
b97d23c5 PM |
4300 | /* Report QS -after- changing ->qsmaskinitnext! */ |
4301 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); | |
4302 | } else { | |
4303 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
7ec99de3 | 4304 | } |
4d60b475 PM |
4305 | smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier(). |
4306 | WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1); | |
4307 | WARN_ON_ONCE(rnp->ofl_seq & 0x1); | |
313517fc | 4308 | smp_mb(); /* Ensure RCU read-side usage follows above initialization. */ |
7ec99de3 PM |
4309 | } |
4310 | ||
27d50c7e | 4311 | /* |
53b46303 PM |
4312 | * The outgoing function has no further need of RCU, so remove it from |
4313 | * the rcu_node tree's ->qsmaskinitnext bit masks. | |
4314 | * | |
4315 | * Note that this function is special in that it is invoked directly | |
4316 | * from the outgoing CPU rather than from the cpuhp_step mechanism. | |
4317 | * This is because this function must be invoked at a precise location. | |
27d50c7e | 4318 | */ |
53b46303 | 4319 | void rcu_report_dead(unsigned int cpu) |
27d50c7e TG |
4320 | { |
4321 | unsigned long flags; | |
4322 | unsigned long mask; | |
da1df50d | 4323 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
27d50c7e TG |
4324 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
4325 | ||
147c6852 PM |
4326 | // Do any dangling deferred wakeups. |
4327 | do_nocb_deferred_wakeup(rdp); | |
4328 | ||
49918a54 | 4329 | /* QS for any half-done expedited grace period. */ |
53b46303 | 4330 | preempt_disable(); |
63d4c8c9 | 4331 | rcu_report_exp_rdp(this_cpu_ptr(&rcu_data)); |
53b46303 PM |
4332 | preempt_enable(); |
4333 | rcu_preempt_deferred_qs(current); | |
4334 | ||
27d50c7e TG |
4335 | /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ |
4336 | mask = rdp->grpmask; | |
4d60b475 PM |
4337 | WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1); |
4338 | WARN_ON_ONCE(!(rnp->ofl_seq & 0x1)); | |
4339 | smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier(). | |
894d45bb | 4340 | raw_spin_lock(&rcu_state.ofl_lock); |
27d50c7e | 4341 | raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */ |
53b46303 PM |
4342 | rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq); |
4343 | rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags); | |
fece2776 PM |
4344 | if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */ |
4345 | /* Report quiescent state -before- changing ->qsmaskinitnext! */ | |
b50912d0 | 4346 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
fece2776 PM |
4347 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
4348 | } | |
105abf82 | 4349 | WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask); |
710d60cb | 4350 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
894d45bb | 4351 | raw_spin_unlock(&rcu_state.ofl_lock); |
4d60b475 PM |
4352 | smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier(). |
4353 | WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1); | |
4354 | WARN_ON_ONCE(rnp->ofl_seq & 0x1); | |
f64c6013 | 4355 | |
c0f97f20 | 4356 | rdp->cpu_started = false; |
27d50c7e | 4357 | } |
a58163d8 | 4358 | |
04e613de | 4359 | #ifdef CONFIG_HOTPLUG_CPU |
53b46303 PM |
4360 | /* |
4361 | * The outgoing CPU has just passed through the dying-idle state, and we | |
4362 | * are being invoked from the CPU that was IPIed to continue the offline | |
4363 | * operation. Migrate the outgoing CPU's callbacks to the current CPU. | |
4364 | */ | |
4365 | void rcutree_migrate_callbacks(int cpu) | |
a58163d8 PM |
4366 | { |
4367 | unsigned long flags; | |
b1a2d79f | 4368 | struct rcu_data *my_rdp; |
c00045be | 4369 | struct rcu_node *my_rnp; |
da1df50d | 4370 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
ec4eacce | 4371 | bool needwake; |
a58163d8 | 4372 | |
3820b513 | 4373 | if (rcu_rdp_is_offloaded(rdp) || |
ce5215c1 | 4374 | rcu_segcblist_empty(&rdp->cblist)) |
95335c03 PM |
4375 | return; /* No callbacks to migrate. */ |
4376 | ||
b1a2d79f | 4377 | local_irq_save(flags); |
da1df50d | 4378 | my_rdp = this_cpu_ptr(&rcu_data); |
c00045be | 4379 | my_rnp = my_rdp->mynode; |
5d6742b3 | 4380 | rcu_nocb_lock(my_rdp); /* irqs already disabled. */ |
d1b222c6 | 4381 | WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies)); |
c00045be | 4382 | raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */ |
ec4eacce | 4383 | /* Leverage recent GPs and set GP for new callbacks. */ |
c00045be PM |
4384 | needwake = rcu_advance_cbs(my_rnp, rdp) || |
4385 | rcu_advance_cbs(my_rnp, my_rdp); | |
f2dbe4a5 | 4386 | rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist); |
23651d9b | 4387 | needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp); |
c035280f | 4388 | rcu_segcblist_disable(&rdp->cblist); |
09efeeee PM |
4389 | WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != |
4390 | !rcu_segcblist_n_cbs(&my_rdp->cblist)); | |
3820b513 | 4391 | if (rcu_rdp_is_offloaded(my_rdp)) { |
5d6742b3 PM |
4392 | raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */ |
4393 | __call_rcu_nocb_wake(my_rdp, true, flags); | |
4394 | } else { | |
4395 | rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */ | |
4396 | raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags); | |
4397 | } | |
ec4eacce | 4398 | if (needwake) |
532c00c9 | 4399 | rcu_gp_kthread_wake(); |
5d6742b3 | 4400 | lockdep_assert_irqs_enabled(); |
a58163d8 PM |
4401 | WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 || |
4402 | !rcu_segcblist_empty(&rdp->cblist), | |
4403 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n", | |
4404 | cpu, rcu_segcblist_n_cbs(&rdp->cblist), | |
4405 | rcu_segcblist_first_cb(&rdp->cblist)); | |
4406 | } | |
27d50c7e TG |
4407 | #endif |
4408 | ||
deb34f36 PM |
4409 | /* |
4410 | * On non-huge systems, use expedited RCU grace periods to make suspend | |
4411 | * and hibernation run faster. | |
4412 | */ | |
d1d74d14 BP |
4413 | static int rcu_pm_notify(struct notifier_block *self, |
4414 | unsigned long action, void *hcpu) | |
4415 | { | |
4416 | switch (action) { | |
4417 | case PM_HIBERNATION_PREPARE: | |
4418 | case PM_SUSPEND_PREPARE: | |
e85e6a21 | 4419 | rcu_expedite_gp(); |
d1d74d14 BP |
4420 | break; |
4421 | case PM_POST_HIBERNATION: | |
4422 | case PM_POST_SUSPEND: | |
e85e6a21 | 4423 | rcu_unexpedite_gp(); |
d1d74d14 BP |
4424 | break; |
4425 | default: | |
4426 | break; | |
4427 | } | |
4428 | return NOTIFY_OK; | |
4429 | } | |
4430 | ||
b3dbec76 | 4431 | /* |
49918a54 | 4432 | * Spawn the kthreads that handle RCU's grace periods. |
b3dbec76 PM |
4433 | */ |
4434 | static int __init rcu_spawn_gp_kthread(void) | |
4435 | { | |
4436 | unsigned long flags; | |
a94844b2 | 4437 | int kthread_prio_in = kthread_prio; |
b3dbec76 | 4438 | struct rcu_node *rnp; |
a94844b2 | 4439 | struct sched_param sp; |
b3dbec76 PM |
4440 | struct task_struct *t; |
4441 | ||
a94844b2 | 4442 | /* Force priority into range. */ |
c7cd161e JFG |
4443 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2 |
4444 | && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)) | |
4445 | kthread_prio = 2; | |
4446 | else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) | |
a94844b2 PM |
4447 | kthread_prio = 1; |
4448 | else if (kthread_prio < 0) | |
4449 | kthread_prio = 0; | |
4450 | else if (kthread_prio > 99) | |
4451 | kthread_prio = 99; | |
c7cd161e | 4452 | |
a94844b2 PM |
4453 | if (kthread_prio != kthread_prio_in) |
4454 | pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n", | |
4455 | kthread_prio, kthread_prio_in); | |
4456 | ||
9386c0b7 | 4457 | rcu_scheduler_fully_active = 1; |
b97d23c5 | 4458 | t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name); |
08543bda PM |
4459 | if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__)) |
4460 | return 0; | |
b97d23c5 PM |
4461 | if (kthread_prio) { |
4462 | sp.sched_priority = kthread_prio; | |
4463 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
b3dbec76 | 4464 | } |
b97d23c5 PM |
4465 | rnp = rcu_get_root(); |
4466 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
5648d659 PM |
4467 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
4468 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); | |
4469 | // Reset .gp_activity and .gp_req_activity before setting .gp_kthread. | |
4470 | smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */ | |
b97d23c5 PM |
4471 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
4472 | wake_up_process(t); | |
35ce7f29 | 4473 | rcu_spawn_nocb_kthreads(); |
9386c0b7 | 4474 | rcu_spawn_boost_kthreads(); |
b3dbec76 PM |
4475 | return 0; |
4476 | } | |
4477 | early_initcall(rcu_spawn_gp_kthread); | |
4478 | ||
bbad9379 | 4479 | /* |
52d7e48b PM |
4480 | * This function is invoked towards the end of the scheduler's |
4481 | * initialization process. Before this is called, the idle task might | |
4482 | * contain synchronous grace-period primitives (during which time, this idle | |
4483 | * task is booting the system, and such primitives are no-ops). After this | |
4484 | * function is called, any synchronous grace-period primitives are run as | |
4485 | * expedited, with the requesting task driving the grace period forward. | |
900b1028 | 4486 | * A later core_initcall() rcu_set_runtime_mode() will switch to full |
52d7e48b | 4487 | * runtime RCU functionality. |
bbad9379 PM |
4488 | */ |
4489 | void rcu_scheduler_starting(void) | |
4490 | { | |
4491 | WARN_ON(num_online_cpus() != 1); | |
4492 | WARN_ON(nr_context_switches() > 0); | |
52d7e48b PM |
4493 | rcu_test_sync_prims(); |
4494 | rcu_scheduler_active = RCU_SCHEDULER_INIT; | |
4495 | rcu_test_sync_prims(); | |
bbad9379 PM |
4496 | } |
4497 | ||
64db4cff | 4498 | /* |
49918a54 | 4499 | * Helper function for rcu_init() that initializes the rcu_state structure. |
64db4cff | 4500 | */ |
b8bb1f63 | 4501 | static void __init rcu_init_one(void) |
64db4cff | 4502 | { |
cb007102 AG |
4503 | static const char * const buf[] = RCU_NODE_NAME_INIT; |
4504 | static const char * const fqs[] = RCU_FQS_NAME_INIT; | |
3dc5dbe9 PM |
4505 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
4506 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; | |
199977bf | 4507 | |
199977bf | 4508 | int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */ |
64db4cff PM |
4509 | int cpustride = 1; |
4510 | int i; | |
4511 | int j; | |
4512 | struct rcu_node *rnp; | |
4513 | ||
05b84aec | 4514 | BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
b6407e86 | 4515 | |
3eaaaf6c PM |
4516 | /* Silence gcc 4.8 false positive about array index out of range. */ |
4517 | if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS) | |
4518 | panic("rcu_init_one: rcu_num_lvls out of range"); | |
4930521a | 4519 | |
64db4cff PM |
4520 | /* Initialize the level-tracking arrays. */ |
4521 | ||
f885b7f2 | 4522 | for (i = 1; i < rcu_num_lvls; i++) |
eb7a6653 PM |
4523 | rcu_state.level[i] = |
4524 | rcu_state.level[i - 1] + num_rcu_lvl[i - 1]; | |
41f5c631 | 4525 | rcu_init_levelspread(levelspread, num_rcu_lvl); |
64db4cff PM |
4526 | |
4527 | /* Initialize the elements themselves, starting from the leaves. */ | |
4528 | ||
f885b7f2 | 4529 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
199977bf | 4530 | cpustride *= levelspread[i]; |
eb7a6653 | 4531 | rnp = rcu_state.level[i]; |
41f5c631 | 4532 | for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) { |
67c583a7 BF |
4533 | raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock)); |
4534 | lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock), | |
b6407e86 | 4535 | &rcu_node_class[i], buf[i]); |
394f2769 PM |
4536 | raw_spin_lock_init(&rnp->fqslock); |
4537 | lockdep_set_class_and_name(&rnp->fqslock, | |
4538 | &rcu_fqs_class[i], fqs[i]); | |
eb7a6653 PM |
4539 | rnp->gp_seq = rcu_state.gp_seq; |
4540 | rnp->gp_seq_needed = rcu_state.gp_seq; | |
4541 | rnp->completedqs = rcu_state.gp_seq; | |
64db4cff PM |
4542 | rnp->qsmask = 0; |
4543 | rnp->qsmaskinit = 0; | |
4544 | rnp->grplo = j * cpustride; | |
4545 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
4546 | if (rnp->grphi >= nr_cpu_ids) |
4547 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
4548 | if (i == 0) { |
4549 | rnp->grpnum = 0; | |
4550 | rnp->grpmask = 0; | |
4551 | rnp->parent = NULL; | |
4552 | } else { | |
199977bf | 4553 | rnp->grpnum = j % levelspread[i - 1]; |
df63fa5b | 4554 | rnp->grpmask = BIT(rnp->grpnum); |
eb7a6653 | 4555 | rnp->parent = rcu_state.level[i - 1] + |
199977bf | 4556 | j / levelspread[i - 1]; |
64db4cff PM |
4557 | } |
4558 | rnp->level = i; | |
12f5f524 | 4559 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 4560 | rcu_init_one_nocb(rnp); |
f6a12f34 PM |
4561 | init_waitqueue_head(&rnp->exp_wq[0]); |
4562 | init_waitqueue_head(&rnp->exp_wq[1]); | |
3b5f668e PM |
4563 | init_waitqueue_head(&rnp->exp_wq[2]); |
4564 | init_waitqueue_head(&rnp->exp_wq[3]); | |
f6a12f34 | 4565 | spin_lock_init(&rnp->exp_lock); |
64db4cff PM |
4566 | } |
4567 | } | |
0c34029a | 4568 | |
eb7a6653 PM |
4569 | init_swait_queue_head(&rcu_state.gp_wq); |
4570 | init_swait_queue_head(&rcu_state.expedited_wq); | |
aedf4ba9 | 4571 | rnp = rcu_first_leaf_node(); |
0c34029a | 4572 | for_each_possible_cpu(i) { |
4a90a068 | 4573 | while (i > rnp->grphi) |
0c34029a | 4574 | rnp++; |
da1df50d | 4575 | per_cpu_ptr(&rcu_data, i)->mynode = rnp; |
53b46303 | 4576 | rcu_boot_init_percpu_data(i); |
0c34029a | 4577 | } |
64db4cff PM |
4578 | } |
4579 | ||
f885b7f2 PM |
4580 | /* |
4581 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 4582 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
4583 | * the ->node array in the rcu_state structure. |
4584 | */ | |
4585 | static void __init rcu_init_geometry(void) | |
4586 | { | |
026ad283 | 4587 | ulong d; |
f885b7f2 | 4588 | int i; |
05b84aec | 4589 | int rcu_capacity[RCU_NUM_LVLS]; |
f885b7f2 | 4590 | |
026ad283 PM |
4591 | /* |
4592 | * Initialize any unspecified boot parameters. | |
4593 | * The default values of jiffies_till_first_fqs and | |
4594 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
4595 | * value, which is a function of HZ, then adding one for each | |
4596 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
4597 | */ | |
4598 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
4599 | if (jiffies_till_first_fqs == ULONG_MAX) | |
4600 | jiffies_till_first_fqs = d; | |
4601 | if (jiffies_till_next_fqs == ULONG_MAX) | |
4602 | jiffies_till_next_fqs = d; | |
6973032a | 4603 | adjust_jiffies_till_sched_qs(); |
026ad283 | 4604 | |
f885b7f2 | 4605 | /* If the compile-time values are accurate, just leave. */ |
47d631af | 4606 | if (rcu_fanout_leaf == RCU_FANOUT_LEAF && |
b17c7035 | 4607 | nr_cpu_ids == NR_CPUS) |
f885b7f2 | 4608 | return; |
a7538352 | 4609 | pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n", |
39479098 | 4610 | rcu_fanout_leaf, nr_cpu_ids); |
f885b7f2 | 4611 | |
f885b7f2 | 4612 | /* |
ee968ac6 PM |
4613 | * The boot-time rcu_fanout_leaf parameter must be at least two |
4614 | * and cannot exceed the number of bits in the rcu_node masks. | |
4615 | * Complain and fall back to the compile-time values if this | |
4616 | * limit is exceeded. | |
f885b7f2 | 4617 | */ |
ee968ac6 | 4618 | if (rcu_fanout_leaf < 2 || |
75cf15a4 | 4619 | rcu_fanout_leaf > sizeof(unsigned long) * 8) { |
13bd6494 | 4620 | rcu_fanout_leaf = RCU_FANOUT_LEAF; |
f885b7f2 PM |
4621 | WARN_ON(1); |
4622 | return; | |
4623 | } | |
4624 | ||
f885b7f2 PM |
4625 | /* |
4626 | * Compute number of nodes that can be handled an rcu_node tree | |
9618138b | 4627 | * with the given number of levels. |
f885b7f2 | 4628 | */ |
9618138b | 4629 | rcu_capacity[0] = rcu_fanout_leaf; |
05b84aec | 4630 | for (i = 1; i < RCU_NUM_LVLS; i++) |
05c5df31 | 4631 | rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT; |
f885b7f2 PM |
4632 | |
4633 | /* | |
75cf15a4 | 4634 | * The tree must be able to accommodate the configured number of CPUs. |
ee968ac6 | 4635 | * If this limit is exceeded, fall back to the compile-time values. |
f885b7f2 | 4636 | */ |
ee968ac6 PM |
4637 | if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) { |
4638 | rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
4639 | WARN_ON(1); | |
4640 | return; | |
4641 | } | |
f885b7f2 | 4642 | |
679f9858 | 4643 | /* Calculate the number of levels in the tree. */ |
9618138b | 4644 | for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) { |
679f9858 | 4645 | } |
9618138b | 4646 | rcu_num_lvls = i + 1; |
679f9858 | 4647 | |
f885b7f2 | 4648 | /* Calculate the number of rcu_nodes at each level of the tree. */ |
679f9858 | 4649 | for (i = 0; i < rcu_num_lvls; i++) { |
9618138b | 4650 | int cap = rcu_capacity[(rcu_num_lvls - 1) - i]; |
679f9858 AG |
4651 | num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap); |
4652 | } | |
f885b7f2 PM |
4653 | |
4654 | /* Calculate the total number of rcu_node structures. */ | |
4655 | rcu_num_nodes = 0; | |
679f9858 | 4656 | for (i = 0; i < rcu_num_lvls; i++) |
f885b7f2 | 4657 | rcu_num_nodes += num_rcu_lvl[i]; |
f885b7f2 PM |
4658 | } |
4659 | ||
a3dc2948 PM |
4660 | /* |
4661 | * Dump out the structure of the rcu_node combining tree associated | |
49918a54 | 4662 | * with the rcu_state structure. |
a3dc2948 | 4663 | */ |
b8bb1f63 | 4664 | static void __init rcu_dump_rcu_node_tree(void) |
a3dc2948 PM |
4665 | { |
4666 | int level = 0; | |
4667 | struct rcu_node *rnp; | |
4668 | ||
4669 | pr_info("rcu_node tree layout dump\n"); | |
4670 | pr_info(" "); | |
aedf4ba9 | 4671 | rcu_for_each_node_breadth_first(rnp) { |
a3dc2948 PM |
4672 | if (rnp->level != level) { |
4673 | pr_cont("\n"); | |
4674 | pr_info(" "); | |
4675 | level = rnp->level; | |
4676 | } | |
4677 | pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum); | |
4678 | } | |
4679 | pr_cont("\n"); | |
4680 | } | |
4681 | ||
ad7c946b | 4682 | struct workqueue_struct *rcu_gp_wq; |
25f3d7ef | 4683 | struct workqueue_struct *rcu_par_gp_wq; |
ad7c946b | 4684 | |
a35d1690 BP |
4685 | static void __init kfree_rcu_batch_init(void) |
4686 | { | |
4687 | int cpu; | |
0392bebe | 4688 | int i; |
a35d1690 BP |
4689 | |
4690 | for_each_possible_cpu(cpu) { | |
4691 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); | |
4692 | ||
34c88174 URS |
4693 | for (i = 0; i < KFREE_N_BATCHES; i++) { |
4694 | INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work); | |
0392bebe | 4695 | krcp->krw_arr[i].krcp = krcp; |
34c88174 URS |
4696 | } |
4697 | ||
a35d1690 | 4698 | INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor); |
56292e86 | 4699 | INIT_WORK(&krcp->page_cache_work, fill_page_cache_func); |
a35d1690 BP |
4700 | krcp->initialized = true; |
4701 | } | |
9154244c JFG |
4702 | if (register_shrinker(&kfree_rcu_shrinker)) |
4703 | pr_err("Failed to register kfree_rcu() shrinker!\n"); | |
a35d1690 BP |
4704 | } |
4705 | ||
9f680ab4 | 4706 | void __init rcu_init(void) |
64db4cff | 4707 | { |
017c4261 | 4708 | int cpu; |
9f680ab4 | 4709 | |
47627678 PM |
4710 | rcu_early_boot_tests(); |
4711 | ||
a35d1690 | 4712 | kfree_rcu_batch_init(); |
f41d911f | 4713 | rcu_bootup_announce(); |
f885b7f2 | 4714 | rcu_init_geometry(); |
b8bb1f63 | 4715 | rcu_init_one(); |
a3dc2948 | 4716 | if (dump_tree) |
b8bb1f63 | 4717 | rcu_dump_rcu_node_tree(); |
48d07c04 SAS |
4718 | if (use_softirq) |
4719 | open_softirq(RCU_SOFTIRQ, rcu_core_si); | |
9f680ab4 PM |
4720 | |
4721 | /* | |
4722 | * We don't need protection against CPU-hotplug here because | |
4723 | * this is called early in boot, before either interrupts | |
4724 | * or the scheduler are operational. | |
4725 | */ | |
d1d74d14 | 4726 | pm_notifier(rcu_pm_notify, 0); |
7ec99de3 | 4727 | for_each_online_cpu(cpu) { |
4df83742 | 4728 | rcutree_prepare_cpu(cpu); |
7ec99de3 | 4729 | rcu_cpu_starting(cpu); |
9b9500da | 4730 | rcutree_online_cpu(cpu); |
7ec99de3 | 4731 | } |
ad7c946b PM |
4732 | |
4733 | /* Create workqueue for expedited GPs and for Tree SRCU. */ | |
4734 | rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0); | |
4735 | WARN_ON(!rcu_gp_wq); | |
25f3d7ef PM |
4736 | rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0); |
4737 | WARN_ON(!rcu_par_gp_wq); | |
e0fcba9a | 4738 | srcu_init(); |
b2b00ddf PM |
4739 | |
4740 | /* Fill in default value for rcutree.qovld boot parameter. */ | |
4741 | /* -After- the rcu_node ->lock fields are initialized! */ | |
4742 | if (qovld < 0) | |
4743 | qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark; | |
4744 | else | |
4745 | qovld_calc = qovld; | |
64db4cff PM |
4746 | } |
4747 | ||
10462d6f | 4748 | #include "tree_stall.h" |
3549c2bc | 4749 | #include "tree_exp.h" |
4102adab | 4750 | #include "tree_plugin.h" |