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 | 33 | #include <linux/completion.h> |
5f98fd03 | 34 | #include <linux/kmemleak.h> |
64db4cff | 35 | #include <linux/moduleparam.h> |
f39650de AS |
36 | #include <linux/panic.h> |
37 | #include <linux/panic_notifier.h> | |
64db4cff PM |
38 | #include <linux/percpu.h> |
39 | #include <linux/notifier.h> | |
40 | #include <linux/cpu.h> | |
41 | #include <linux/mutex.h> | |
42 | #include <linux/time.h> | |
bbad9379 | 43 | #include <linux/kernel_stat.h> |
a26ac245 PM |
44 | #include <linux/wait.h> |
45 | #include <linux/kthread.h> | |
ae7e81c0 | 46 | #include <uapi/linux/sched/types.h> |
268bb0ce | 47 | #include <linux/prefetch.h> |
3d3b7db0 | 48 | #include <linux/delay.h> |
661a85dc | 49 | #include <linux/random.h> |
af658dca | 50 | #include <linux/trace_events.h> |
d1d74d14 | 51 | #include <linux/suspend.h> |
a278d471 | 52 | #include <linux/ftrace.h> |
d3052109 | 53 | #include <linux/tick.h> |
2ccaff10 | 54 | #include <linux/sysrq.h> |
c13324a5 | 55 | #include <linux/kprobes.h> |
48d07c04 SAS |
56 | #include <linux/gfp.h> |
57 | #include <linux/oom.h> | |
58 | #include <linux/smpboot.h> | |
59 | #include <linux/jiffies.h> | |
77a40f97 | 60 | #include <linux/slab.h> |
48d07c04 | 61 | #include <linux/sched/isolation.h> |
cfcdef5e | 62 | #include <linux/sched/clock.h> |
5f3c8d62 URS |
63 | #include <linux/vmalloc.h> |
64 | #include <linux/mm.h> | |
26e760c9 | 65 | #include <linux/kasan.h> |
17211455 | 66 | #include <linux/context_tracking.h> |
48d07c04 | 67 | #include "../time/tick-internal.h" |
64db4cff | 68 | |
4102adab | 69 | #include "tree.h" |
29c00b4a | 70 | #include "rcu.h" |
9f77da9f | 71 | |
4102adab PM |
72 | #ifdef MODULE_PARAM_PREFIX |
73 | #undef MODULE_PARAM_PREFIX | |
74 | #endif | |
75 | #define MODULE_PARAM_PREFIX "rcutree." | |
76 | ||
64db4cff | 77 | /* Data structures. */ |
988f569a | 78 | static void rcu_sr_normal_gp_cleanup_work(struct work_struct *); |
64db4cff | 79 | |
4c5273bf | 80 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = { |
a5d1b0b6 | 81 | .gpwrap = true, |
8d346d43 | 82 | #ifdef CONFIG_RCU_NOCB_CPU |
213d56bf | 83 | .cblist.flags = SEGCBLIST_RCU_CORE, |
8d346d43 | 84 | #endif |
4c5273bf | 85 | }; |
c30fe541 | 86 | static struct rcu_state rcu_state = { |
358be2d3 | 87 | .level = { &rcu_state.node[0] }, |
358be2d3 PM |
88 | .gp_state = RCU_GP_IDLE, |
89 | .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT, | |
90 | .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex), | |
80b3fd47 | 91 | .barrier_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.barrier_lock), |
358be2d3 PM |
92 | .name = RCU_NAME, |
93 | .abbr = RCU_ABBR, | |
94 | .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex), | |
95 | .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex), | |
82980b16 | 96 | .ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED, |
988f569a URS |
97 | .srs_cleanup_work = __WORK_INITIALIZER(rcu_state.srs_cleanup_work, |
98 | rcu_sr_normal_gp_cleanup_work), | |
358be2d3 | 99 | }; |
27f4d280 | 100 | |
a3dc2948 PM |
101 | /* Dump rcu_node combining tree at boot to verify correct setup. */ |
102 | static bool dump_tree; | |
103 | module_param(dump_tree, bool, 0444); | |
48d07c04 | 104 | /* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */ |
8b9a0ecc SW |
105 | static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT); |
106 | #ifndef CONFIG_PREEMPT_RT | |
48d07c04 | 107 | module_param(use_softirq, bool, 0444); |
8b9a0ecc | 108 | #endif |
7fa27001 PM |
109 | /* Control rcu_node-tree auto-balancing at boot time. */ |
110 | static bool rcu_fanout_exact; | |
111 | module_param(rcu_fanout_exact, bool, 0444); | |
47d631af PM |
112 | /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */ |
113 | static int rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
7e5c2dfb | 114 | module_param(rcu_fanout_leaf, int, 0444); |
f885b7f2 | 115 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
cb007102 | 116 | /* Number of rcu_nodes at specified level. */ |
e95d68d2 | 117 | int num_rcu_lvl[] = NUM_RCU_LVL_INIT; |
f885b7f2 PM |
118 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ |
119 | ||
b0d30417 | 120 | /* |
52d7e48b PM |
121 | * The rcu_scheduler_active variable is initialized to the value |
122 | * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the | |
123 | * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE, | |
124 | * RCU can assume that there is but one task, allowing RCU to (for example) | |
0d95092c | 125 | * optimize synchronize_rcu() to a simple barrier(). When this variable |
52d7e48b PM |
126 | * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required |
127 | * to detect real grace periods. This variable is also used to suppress | |
128 | * boot-time false positives from lockdep-RCU error checking. Finally, it | |
129 | * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU | |
130 | * is fully initialized, including all of its kthreads having been spawned. | |
b0d30417 | 131 | */ |
bbad9379 PM |
132 | int rcu_scheduler_active __read_mostly; |
133 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); | |
134 | ||
b0d30417 PM |
135 | /* |
136 | * The rcu_scheduler_fully_active variable transitions from zero to one | |
137 | * during the early_initcall() processing, which is after the scheduler | |
138 | * is capable of creating new tasks. So RCU processing (for example, | |
139 | * creating tasks for RCU priority boosting) must be delayed until after | |
140 | * rcu_scheduler_fully_active transitions from zero to one. We also | |
141 | * currently delay invocation of any RCU callbacks until after this point. | |
142 | * | |
143 | * It might later prove better for people registering RCU callbacks during | |
144 | * early boot to take responsibility for these callbacks, but one step at | |
145 | * a time. | |
146 | */ | |
147 | static int rcu_scheduler_fully_active __read_mostly; | |
148 | ||
b50912d0 PM |
149 | static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp, |
150 | unsigned long gps, unsigned long flags); | |
b67cffcb | 151 | static struct task_struct *rcu_boost_task(struct rcu_node *rnp); |
a46e0899 | 152 | static void invoke_rcu_core(void); |
63d4c8c9 | 153 | static void rcu_report_exp_rdp(struct rcu_data *rdp); |
3549c2bc | 154 | static void sync_sched_exp_online_cleanup(int cpu); |
b2b00ddf | 155 | static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp); |
3820b513 | 156 | static bool rcu_rdp_is_offloaded(struct rcu_data *rdp); |
5a04848d PM |
157 | static bool rcu_rdp_cpu_online(struct rcu_data *rdp); |
158 | static bool rcu_init_invoked(void); | |
159 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); | |
160 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); | |
a26ac245 | 161 | |
8f489b4d URS |
162 | /* |
163 | * rcuc/rcub/rcuop kthread realtime priority. The "rcuop" | |
164 | * real-time priority(enabling/disabling) is controlled by | |
165 | * the extra CONFIG_RCU_NOCB_CPU_CB_BOOST configuration. | |
166 | */ | |
26730f55 | 167 | static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0; |
3ffe3d1a | 168 | module_param(kthread_prio, int, 0444); |
a94844b2 | 169 | |
8d7dc928 | 170 | /* Delay in jiffies for grace-period initialization delays, debug only. */ |
0f41c0dd | 171 | |
90040c9e PM |
172 | static int gp_preinit_delay; |
173 | module_param(gp_preinit_delay, int, 0444); | |
174 | static int gp_init_delay; | |
175 | module_param(gp_init_delay, int, 0444); | |
176 | static int gp_cleanup_delay; | |
177 | module_param(gp_cleanup_delay, int, 0444); | |
0f41c0dd | 178 | |
aa40c138 PM |
179 | // Add delay to rcu_read_unlock() for strict grace periods. |
180 | static int rcu_unlock_delay; | |
181 | #ifdef CONFIG_RCU_STRICT_GRACE_PERIOD | |
182 | module_param(rcu_unlock_delay, int, 0444); | |
183 | #endif | |
184 | ||
53c72b59 URS |
185 | /* |
186 | * This rcu parameter is runtime-read-only. It reflects | |
187 | * a minimum allowed number of objects which can be cached | |
188 | * per-CPU. Object size is equal to one page. This value | |
189 | * can be changed at boot time. | |
190 | */ | |
56292e86 | 191 | static int rcu_min_cached_objs = 5; |
53c72b59 URS |
192 | module_param(rcu_min_cached_objs, int, 0444); |
193 | ||
d0bfa8b3 ZQ |
194 | // A page shrinker can ask for pages to be freed to make them |
195 | // available for other parts of the system. This usually happens | |
196 | // under low memory conditions, and in that case we should also | |
197 | // defer page-cache filling for a short time period. | |
198 | // | |
199 | // The default value is 5 seconds, which is long enough to reduce | |
200 | // interference with the shrinker while it asks other systems to | |
201 | // drain their caches. | |
202 | static int rcu_delay_page_cache_fill_msec = 5000; | |
203 | module_param(rcu_delay_page_cache_fill_msec, int, 0444); | |
204 | ||
4cf439a2 | 205 | /* Retrieve RCU kthreads priority for rcutorture */ |
4babd855 JFG |
206 | int rcu_get_gp_kthreads_prio(void) |
207 | { | |
208 | return kthread_prio; | |
209 | } | |
210 | EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio); | |
211 | ||
eab128e8 PM |
212 | /* |
213 | * Number of grace periods between delays, normalized by the duration of | |
bfd090be | 214 | * the delay. The longer the delay, the more the grace periods between |
eab128e8 PM |
215 | * each delay. The reason for this normalization is that it means that, |
216 | * for non-zero delays, the overall slowdown of grace periods is constant | |
217 | * regardless of the duration of the delay. This arrangement balances | |
218 | * the need for long delays to increase some race probabilities with the | |
219 | * need for fast grace periods to increase other race probabilities. | |
220 | */ | |
277ffe1b | 221 | #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */ |
37745d28 | 222 | |
fc2219d4 | 223 | /* |
7d0ae808 | 224 | * Return true if an RCU grace period is in progress. The READ_ONCE()s |
fc2219d4 PM |
225 | * permit this function to be invoked without holding the root rcu_node |
226 | * structure's ->lock, but of course results can be subject to change. | |
227 | */ | |
de8e8730 | 228 | static int rcu_gp_in_progress(void) |
fc2219d4 | 229 | { |
de8e8730 | 230 | return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq)); |
b1f77b05 IM |
231 | } |
232 | ||
903ee83d PM |
233 | /* |
234 | * Return the number of callbacks queued on the specified CPU. | |
235 | * Handles both the nocbs and normal cases. | |
236 | */ | |
237 | static long rcu_get_n_cbs_cpu(int cpu) | |
238 | { | |
239 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
240 | ||
c035280f | 241 | if (rcu_segcblist_is_enabled(&rdp->cblist)) |
903ee83d | 242 | return rcu_segcblist_n_cbs(&rdp->cblist); |
c035280f | 243 | return 0; |
903ee83d PM |
244 | } |
245 | ||
3183059a PM |
246 | /** |
247 | * rcu_softirq_qs - Provide a set of RCU quiescent states in softirq processing | |
248 | * | |
249 | * Mark a quiescent state for RCU, Tasks RCU, and Tasks Trace RCU. | |
250 | * This is a special-purpose function to be used in the softirq | |
251 | * infrastructure and perhaps the occasional long-running softirq | |
252 | * handler. | |
253 | * | |
254 | * Note that from RCU's viewpoint, a call to rcu_softirq_qs() is | |
255 | * equivalent to momentarily completely enabling preemption. For | |
256 | * example, given this code:: | |
257 | * | |
258 | * local_bh_disable(); | |
259 | * do_something(); | |
260 | * rcu_softirq_qs(); // A | |
261 | * do_something_else(); | |
262 | * local_bh_enable(); // B | |
263 | * | |
264 | * A call to synchronize_rcu() that began concurrently with the | |
265 | * call to do_something() would be guaranteed to wait only until | |
266 | * execution reached statement A. Without that rcu_softirq_qs(), | |
267 | * that same synchronize_rcu() would instead be guaranteed to wait | |
268 | * until execution reached statement B. | |
269 | */ | |
d28139c4 | 270 | void rcu_softirq_qs(void) |
b1f77b05 | 271 | { |
3183059a PM |
272 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
273 | lock_is_held(&rcu_lock_map) || | |
274 | lock_is_held(&rcu_sched_lock_map), | |
275 | "Illegal rcu_softirq_qs() in RCU read-side critical section"); | |
45975c7d | 276 | rcu_qs(); |
d28139c4 | 277 | rcu_preempt_deferred_qs(current); |
cf868c2a | 278 | rcu_tasks_qs(current, false); |
b1f77b05 | 279 | } |
64db4cff | 280 | |
2625d469 PM |
281 | /* |
282 | * Reset the current CPU's ->dynticks counter to indicate that the | |
283 | * newly onlined CPU is no longer in an extended quiescent state. | |
284 | * This will either leave the counter unchanged, or increment it | |
285 | * to the next non-quiescent value. | |
286 | * | |
287 | * The non-atomic test/increment sequence works because the upper bits | |
288 | * of the ->dynticks counter are manipulated only by the corresponding CPU, | |
289 | * or when the corresponding CPU is offline. | |
290 | */ | |
291 | static void rcu_dynticks_eqs_online(void) | |
292 | { | |
17147677 | 293 | if (ct_dynticks() & RCU_DYNTICKS_IDX) |
2625d469 | 294 | return; |
17147677 | 295 | ct_state_inc(RCU_DYNTICKS_IDX); |
02a5c550 PM |
296 | } |
297 | ||
8b2f63ab PM |
298 | /* |
299 | * Snapshot the ->dynticks counter with full ordering so as to allow | |
300 | * stable comparison of this counter with past and future snapshots. | |
301 | */ | |
62e2412d | 302 | static int rcu_dynticks_snap(int cpu) |
8b2f63ab | 303 | { |
2be57f73 | 304 | smp_mb(); // Fundamental RCU ordering guarantee. |
62e2412d | 305 | return ct_dynticks_cpu_acquire(cpu); |
8b2f63ab PM |
306 | } |
307 | ||
02a5c550 PM |
308 | /* |
309 | * Return true if the snapshot returned from rcu_dynticks_snap() | |
310 | * indicates that RCU is in an extended quiescent state. | |
311 | */ | |
312 | static bool rcu_dynticks_in_eqs(int snap) | |
313 | { | |
17147677 | 314 | return !(snap & RCU_DYNTICKS_IDX); |
02a5c550 PM |
315 | } |
316 | ||
317 | /* | |
dc5a4f29 | 318 | * Return true if the CPU corresponding to the specified rcu_data |
02a5c550 PM |
319 | * structure has spent some time in an extended quiescent state since |
320 | * rcu_dynticks_snap() returned the specified snapshot. | |
321 | */ | |
dc5a4f29 | 322 | static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap) |
02a5c550 | 323 | { |
62e2412d | 324 | return snap != rcu_dynticks_snap(rdp->cpu); |
02a5c550 PM |
325 | } |
326 | ||
7d0c9c50 PM |
327 | /* |
328 | * Return true if the referenced integer is zero while the specified | |
329 | * CPU remains within a single extended quiescent state. | |
330 | */ | |
331 | bool rcu_dynticks_zero_in_eqs(int cpu, int *vp) | |
332 | { | |
7d0c9c50 PM |
333 | int snap; |
334 | ||
335 | // If not quiescent, force back to earlier extended quiescent state. | |
17147677 | 336 | snap = ct_dynticks_cpu(cpu) & ~RCU_DYNTICKS_IDX; |
7d0c9c50 PM |
337 | smp_rmb(); // Order ->dynticks and *vp reads. |
338 | if (READ_ONCE(*vp)) | |
339 | return false; // Non-zero, so report failure; | |
340 | smp_rmb(); // Order *vp read and ->dynticks re-read. | |
341 | ||
342 | // If still in the same extended quiescent state, we are good! | |
62e2412d | 343 | return snap == ct_dynticks_cpu(cpu); |
6563de9d | 344 | } |
5cd37193 | 345 | |
4a81e832 PM |
346 | /* |
347 | * Let the RCU core know that this CPU has gone through the scheduler, | |
348 | * which is a quiescent state. This is called when the need for a | |
349 | * quiescent state is urgent, so we burn an atomic operation and full | |
350 | * memory barriers to let the RCU core know about it, regardless of what | |
351 | * this CPU might (or might not) do in the near future. | |
352 | * | |
0f9be8ca | 353 | * We inform the RCU core by emulating a zero-duration dyntick-idle period. |
46a5d164 | 354 | * |
3b57a399 | 355 | * The caller must have disabled interrupts and must not be idle. |
4a81e832 | 356 | */ |
4230e2de | 357 | notrace void rcu_momentary_dyntick_idle(void) |
4a81e832 | 358 | { |
2be57f73 | 359 | int seq; |
3b57a399 | 360 | |
2dba13f0 | 361 | raw_cpu_write(rcu_data.rcu_need_heavy_qs, false); |
17147677 | 362 | seq = ct_state_inc(2 * RCU_DYNTICKS_IDX); |
3b57a399 | 363 | /* It is illegal to call this from idle state. */ |
17147677 | 364 | WARN_ON_ONCE(!(seq & RCU_DYNTICKS_IDX)); |
3e310098 | 365 | rcu_preempt_deferred_qs(current); |
4a81e832 | 366 | } |
79ba7ff5 | 367 | EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle); |
4a81e832 | 368 | |
45975c7d | 369 | /** |
806f04e9 | 370 | * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle |
bb73c52b | 371 | * |
eddded80 | 372 | * If the current CPU is idle and running at a first-level (not nested) |
806f04e9 PZ |
373 | * interrupt, or directly, from idle, return true. |
374 | * | |
375 | * The caller must have at least disabled IRQs. | |
5cd37193 | 376 | */ |
45975c7d | 377 | static int rcu_is_cpu_rrupt_from_idle(void) |
5cd37193 | 378 | { |
806f04e9 PZ |
379 | long nesting; |
380 | ||
381 | /* | |
382 | * Usually called from the tick; but also used from smp_function_call() | |
383 | * for expedited grace periods. This latter can result in running from | |
384 | * the idle task, instead of an actual IPI. | |
385 | */ | |
386 | lockdep_assert_irqs_disabled(); | |
eddded80 JFG |
387 | |
388 | /* Check for counter underflows */ | |
904e600e | 389 | RCU_LOCKDEP_WARN(ct_dynticks_nesting() < 0, |
eddded80 | 390 | "RCU dynticks_nesting counter underflow!"); |
95e04f48 | 391 | RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() <= 0, |
eddded80 JFG |
392 | "RCU dynticks_nmi_nesting counter underflow/zero!"); |
393 | ||
394 | /* Are we at first interrupt nesting level? */ | |
95e04f48 | 395 | nesting = ct_dynticks_nmi_nesting(); |
806f04e9 | 396 | if (nesting > 1) |
eddded80 JFG |
397 | return false; |
398 | ||
806f04e9 PZ |
399 | /* |
400 | * If we're not in an interrupt, we must be in the idle task! | |
401 | */ | |
402 | WARN_ON_ONCE(!nesting && !is_idle_task(current)); | |
403 | ||
eddded80 | 404 | /* Does CPU appear to be idle from an RCU standpoint? */ |
904e600e | 405 | return ct_dynticks_nesting() == 0; |
5cd37193 | 406 | } |
5cd37193 | 407 | |
29fc5f93 PM |
408 | #define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10) |
409 | // Maximum callbacks per rcu_do_batch ... | |
410 | #define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood. | |
17c7798b | 411 | static long blimit = DEFAULT_RCU_BLIMIT; |
29fc5f93 | 412 | #define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit. |
17c7798b | 413 | static long qhimark = DEFAULT_RCU_QHIMARK; |
29fc5f93 | 414 | #define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit. |
17c7798b | 415 | static long qlowmark = DEFAULT_RCU_QLOMARK; |
b2b00ddf PM |
416 | #define DEFAULT_RCU_QOVLD_MULT 2 |
417 | #define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK) | |
29fc5f93 PM |
418 | static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS. |
419 | static long qovld_calc = -1; // No pre-initialization lock acquisitions! | |
64db4cff | 420 | |
878d7439 ED |
421 | module_param(blimit, long, 0444); |
422 | module_param(qhimark, long, 0444); | |
423 | module_param(qlowmark, long, 0444); | |
b2b00ddf | 424 | module_param(qovld, long, 0444); |
3d76c082 | 425 | |
aecd34b9 | 426 | static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX; |
026ad283 | 427 | static ulong jiffies_till_next_fqs = ULONG_MAX; |
8c7c4829 | 428 | static bool rcu_kick_kthreads; |
cfcdef5e ED |
429 | static int rcu_divisor = 7; |
430 | module_param(rcu_divisor, int, 0644); | |
431 | ||
432 | /* Force an exit from rcu_do_batch() after 3 milliseconds. */ | |
433 | static long rcu_resched_ns = 3 * NSEC_PER_MSEC; | |
434 | module_param(rcu_resched_ns, long, 0644); | |
d40011f6 | 435 | |
c06aed0e PM |
436 | /* |
437 | * How long the grace period must be before we start recruiting | |
438 | * quiescent-state help from rcu_note_context_switch(). | |
439 | */ | |
440 | static ulong jiffies_till_sched_qs = ULONG_MAX; | |
441 | module_param(jiffies_till_sched_qs, ulong, 0444); | |
85f2b60c | 442 | static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */ |
c06aed0e PM |
443 | module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */ |
444 | ||
445 | /* | |
446 | * Make sure that we give the grace-period kthread time to detect any | |
447 | * idle CPUs before taking active measures to force quiescent states. | |
448 | * However, don't go below 100 milliseconds, adjusted upwards for really | |
449 | * large systems. | |
450 | */ | |
451 | static void adjust_jiffies_till_sched_qs(void) | |
452 | { | |
453 | unsigned long j; | |
454 | ||
455 | /* If jiffies_till_sched_qs was specified, respect the request. */ | |
456 | if (jiffies_till_sched_qs != ULONG_MAX) { | |
457 | WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs); | |
458 | return; | |
459 | } | |
85f2b60c | 460 | /* Otherwise, set to third fqs scan, but bound below on large system. */ |
c06aed0e PM |
461 | j = READ_ONCE(jiffies_till_first_fqs) + |
462 | 2 * READ_ONCE(jiffies_till_next_fqs); | |
463 | if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV) | |
464 | j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
465 | pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j); | |
466 | WRITE_ONCE(jiffies_to_sched_qs, j); | |
467 | } | |
468 | ||
67abb96c BP |
469 | static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp) |
470 | { | |
471 | ulong j; | |
472 | int ret = kstrtoul(val, 0, &j); | |
473 | ||
c06aed0e | 474 | if (!ret) { |
67abb96c | 475 | WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j); |
c06aed0e PM |
476 | adjust_jiffies_till_sched_qs(); |
477 | } | |
67abb96c BP |
478 | return ret; |
479 | } | |
480 | ||
481 | static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp) | |
482 | { | |
483 | ulong j; | |
484 | int ret = kstrtoul(val, 0, &j); | |
485 | ||
c06aed0e | 486 | if (!ret) { |
67abb96c | 487 | WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1)); |
c06aed0e PM |
488 | adjust_jiffies_till_sched_qs(); |
489 | } | |
67abb96c BP |
490 | return ret; |
491 | } | |
492 | ||
7c47ee5a | 493 | static const struct kernel_param_ops first_fqs_jiffies_ops = { |
67abb96c BP |
494 | .set = param_set_first_fqs_jiffies, |
495 | .get = param_get_ulong, | |
496 | }; | |
497 | ||
7c47ee5a | 498 | static const struct kernel_param_ops next_fqs_jiffies_ops = { |
67abb96c BP |
499 | .set = param_set_next_fqs_jiffies, |
500 | .get = param_get_ulong, | |
501 | }; | |
502 | ||
503 | module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644); | |
504 | module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644); | |
8c7c4829 | 505 | module_param(rcu_kick_kthreads, bool, 0644); |
d40011f6 | 506 | |
8ff0b907 | 507 | static void force_qs_rnp(int (*f)(struct rcu_data *rdp)); |
dd7dafd1 | 508 | static int rcu_pending(int user); |
64db4cff PM |
509 | |
510 | /* | |
17ef2fe9 | 511 | * Return the number of RCU GPs completed thus far for debug & stats. |
64db4cff | 512 | */ |
17ef2fe9 | 513 | unsigned long rcu_get_gp_seq(void) |
917963d0 | 514 | { |
16fc9c60 | 515 | return READ_ONCE(rcu_state.gp_seq); |
917963d0 | 516 | } |
17ef2fe9 | 517 | EXPORT_SYMBOL_GPL(rcu_get_gp_seq); |
917963d0 | 518 | |
291783b8 PM |
519 | /* |
520 | * Return the number of RCU expedited batches completed thus far for | |
521 | * debug & stats. Odd numbers mean that a batch is in progress, even | |
522 | * numbers mean idle. The value returned will thus be roughly double | |
523 | * the cumulative batches since boot. | |
524 | */ | |
525 | unsigned long rcu_exp_batches_completed(void) | |
526 | { | |
16fc9c60 | 527 | return rcu_state.expedited_sequence; |
291783b8 PM |
528 | } |
529 | EXPORT_SYMBOL_GPL(rcu_exp_batches_completed); | |
530 | ||
fd897573 PM |
531 | /* |
532 | * Return the root node of the rcu_state structure. | |
533 | */ | |
534 | static struct rcu_node *rcu_get_root(void) | |
535 | { | |
536 | return &rcu_state.node[0]; | |
537 | } | |
538 | ||
ad0dc7f9 PM |
539 | /* |
540 | * Send along grace-period-related data for rcutorture diagnostics. | |
541 | */ | |
dddcddef | 542 | void rcutorture_get_gp_data(int *flags, unsigned long *gp_seq) |
ad0dc7f9 | 543 | { |
dddcddef Z |
544 | *flags = READ_ONCE(rcu_state.gp_flags); |
545 | *gp_seq = rcu_seq_current(&rcu_state.gp_seq); | |
ad0dc7f9 PM |
546 | } |
547 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); | |
548 | ||
17211455 | 549 | #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) |
f8bb5cae FW |
550 | /* |
551 | * An empty function that will trigger a reschedule on | |
4ae7dc97 | 552 | * IRQ tail once IRQs get re-enabled on userspace/guest resume. |
f8bb5cae FW |
553 | */ |
554 | static void late_wakeup_func(struct irq_work *work) | |
555 | { | |
556 | } | |
557 | ||
558 | static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) = | |
559 | IRQ_WORK_INIT(late_wakeup_func); | |
560 | ||
4ae7dc97 FW |
561 | /* |
562 | * If either: | |
563 | * | |
564 | * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work | |
565 | * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry. | |
566 | * | |
567 | * In these cases the late RCU wake ups aren't supported in the resched loops and our | |
568 | * last resort is to fire a local irq_work that will trigger a reschedule once IRQs | |
569 | * get re-enabled again. | |
570 | */ | |
56450649 | 571 | noinstr void rcu_irq_work_resched(void) |
4ae7dc97 FW |
572 | { |
573 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); | |
574 | ||
575 | if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU)) | |
576 | return; | |
577 | ||
578 | if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU)) | |
579 | return; | |
580 | ||
581 | instrumentation_begin(); | |
582 | if (do_nocb_deferred_wakeup(rdp) && need_resched()) { | |
583 | irq_work_queue(this_cpu_ptr(&late_wakeup_work)); | |
584 | } | |
585 | instrumentation_end(); | |
586 | } | |
17211455 | 587 | #endif /* #if defined(CONFIG_NO_HZ_FULL) && (!defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)) */ |
7c9906ca | 588 | |
07325d4a TG |
589 | #ifdef CONFIG_PROVE_RCU |
590 | /** | |
591 | * rcu_irq_exit_check_preempt - Validate that scheduling is possible | |
592 | */ | |
593 | void rcu_irq_exit_check_preempt(void) | |
594 | { | |
595 | lockdep_assert_irqs_disabled(); | |
596 | ||
904e600e | 597 | RCU_LOCKDEP_WARN(ct_dynticks_nesting() <= 0, |
07325d4a | 598 | "RCU dynticks_nesting counter underflow/zero!"); |
95e04f48 | 599 | RCU_LOCKDEP_WARN(ct_dynticks_nmi_nesting() != |
07325d4a TG |
600 | DYNTICK_IRQ_NONIDLE, |
601 | "Bad RCU dynticks_nmi_nesting counter\n"); | |
602 | RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(), | |
603 | "RCU in extended quiescent state!"); | |
604 | } | |
605 | #endif /* #ifdef CONFIG_PROVE_RCU */ | |
606 | ||
d1ec4c34 | 607 | #ifdef CONFIG_NO_HZ_FULL |
aaf2bc50 PM |
608 | /** |
609 | * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it. | |
610 | * | |
611 | * The scheduler tick is not normally enabled when CPUs enter the kernel | |
612 | * from nohz_full userspace execution. After all, nohz_full userspace | |
613 | * execution is an RCU quiescent state and the time executing in the kernel | |
614 | * is quite short. Except of course when it isn't. And it is not hard to | |
615 | * cause a large system to spend tens of seconds or even minutes looping | |
616 | * in the kernel, which can cause a number of problems, include RCU CPU | |
617 | * stall warnings. | |
618 | * | |
619 | * Therefore, if a nohz_full CPU fails to report a quiescent state | |
620 | * in a timely manner, the RCU grace-period kthread sets that CPU's | |
621 | * ->rcu_urgent_qs flag with the expectation that the next interrupt or | |
622 | * exception will invoke this function, which will turn on the scheduler | |
623 | * tick, which will enable RCU to detect that CPU's quiescent states, | |
624 | * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels. | |
625 | * The tick will be disabled once a quiescent state is reported for | |
626 | * this CPU. | |
627 | * | |
628 | * Of course, in carefully tuned systems, there might never be an | |
629 | * interrupt or exception. In that case, the RCU grace-period kthread | |
630 | * will eventually cause one to happen. However, in less carefully | |
631 | * controlled environments, this function allows RCU to get what it | |
632 | * needs without creating otherwise useless interruptions. | |
633 | */ | |
634 | void __rcu_irq_enter_check_tick(void) | |
635 | { | |
636 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); | |
637 | ||
6dbce04d PZ |
638 | // If we're here from NMI there's nothing to do. |
639 | if (in_nmi()) | |
aaf2bc50 PM |
640 | return; |
641 | ||
642 | RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(), | |
643 | "Illegal rcu_irq_enter_check_tick() from extended quiescent state"); | |
644 | ||
645 | if (!tick_nohz_full_cpu(rdp->cpu) || | |
646 | !READ_ONCE(rdp->rcu_urgent_qs) || | |
647 | READ_ONCE(rdp->rcu_forced_tick)) { | |
648 | // RCU doesn't need nohz_full help from this CPU, or it is | |
649 | // already getting that help. | |
650 | return; | |
651 | } | |
652 | ||
653 | // We get here only when not in an extended quiescent state and | |
654 | // from interrupts (as opposed to NMIs). Therefore, (1) RCU is | |
655 | // already watching and (2) The fact that we are in an interrupt | |
656 | // handler and that the rcu_node lock is an irq-disabled lock | |
657 | // prevents self-deadlock. So we can safely recheck under the lock. | |
658 | // Note that the nohz_full state currently cannot change. | |
659 | raw_spin_lock_rcu_node(rdp->mynode); | |
343640cb | 660 | if (READ_ONCE(rdp->rcu_urgent_qs) && !rdp->rcu_forced_tick) { |
aaf2bc50 PM |
661 | // A nohz_full CPU is in the kernel and RCU needs a |
662 | // quiescent state. Turn on the tick! | |
663 | WRITE_ONCE(rdp->rcu_forced_tick, true); | |
664 | tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU); | |
665 | } | |
666 | raw_spin_unlock_rcu_node(rdp->mynode); | |
667 | } | |
7a29fb4a | 668 | NOKPROBE_SYMBOL(__rcu_irq_enter_check_tick); |
d1ec4c34 | 669 | #endif /* CONFIG_NO_HZ_FULL */ |
19dd1591 | 670 | |
bc849e91 PM |
671 | /* |
672 | * Check to see if any future non-offloaded RCU-related work will need | |
673 | * to be done by the current CPU, even if none need be done immediately, | |
674 | * returning 1 if so. This function is part of the RCU implementation; | |
675 | * it is -not- an exported member of the RCU API. This is used by | |
676 | * the idle-entry code to figure out whether it is safe to disable the | |
677 | * scheduler-clock interrupt. | |
678 | * | |
679 | * Just check whether or not this CPU has non-offloaded RCU callbacks | |
680 | * queued. | |
681 | */ | |
29845399 | 682 | int rcu_needs_cpu(void) |
bc849e91 | 683 | { |
bc849e91 PM |
684 | return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) && |
685 | !rcu_rdp_is_offloaded(this_cpu_ptr(&rcu_data)); | |
686 | } | |
687 | ||
66e4c33b | 688 | /* |
516e5ae0 JFG |
689 | * If any sort of urgency was applied to the current CPU (for example, |
690 | * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order | |
691 | * to get to a quiescent state, disable it. | |
66e4c33b | 692 | */ |
516e5ae0 | 693 | static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp) |
66e4c33b | 694 | { |
5b14557b | 695 | raw_lockdep_assert_held_rcu_node(rdp->mynode); |
516e5ae0 JFG |
696 | WRITE_ONCE(rdp->rcu_urgent_qs, false); |
697 | WRITE_ONCE(rdp->rcu_need_heavy_qs, false); | |
66e4c33b PM |
698 | if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) { |
699 | tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU); | |
2a2ae872 | 700 | WRITE_ONCE(rdp->rcu_forced_tick, false); |
66e4c33b PM |
701 | } |
702 | } | |
703 | ||
5c173eb8 | 704 | /** |
c924bf5a | 705 | * rcu_is_watching - RCU read-side critical sections permitted on current CPU? |
64db4cff | 706 | * |
c924bf5a PM |
707 | * Return @true if RCU is watching the running CPU and @false otherwise. |
708 | * An @true return means that this CPU can safely enter RCU read-side | |
709 | * critical sections. | |
710 | * | |
711 | * Although calls to rcu_is_watching() from most parts of the kernel | |
712 | * will return @true, there are important exceptions. For example, if the | |
713 | * current CPU is deep within its idle loop, in kernel entry/exit code, | |
714 | * or offline, rcu_is_watching() will return @false. | |
d2098b44 PZ |
715 | * |
716 | * Make notrace because it can be called by the internal functions of | |
717 | * ftrace, and making this notrace removes unnecessary recursion calls. | |
64db4cff | 718 | */ |
d2098b44 | 719 | notrace bool rcu_is_watching(void) |
64db4cff | 720 | { |
f534ed1f | 721 | bool ret; |
34240697 | 722 | |
46f00d18 | 723 | preempt_disable_notrace(); |
791875d1 | 724 | ret = !rcu_dynticks_curr_cpu_in_eqs(); |
46f00d18 | 725 | preempt_enable_notrace(); |
34240697 | 726 | return ret; |
64db4cff | 727 | } |
5c173eb8 | 728 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
64db4cff | 729 | |
bcbfdd01 PM |
730 | /* |
731 | * If a holdout task is actually running, request an urgent quiescent | |
732 | * state from its CPU. This is unsynchronized, so migrations can cause | |
733 | * the request to go to the wrong CPU. Which is OK, all that will happen | |
734 | * is that the CPU's next context switch will be a bit slower and next | |
735 | * time around this task will generate another request. | |
736 | */ | |
737 | void rcu_request_urgent_qs_task(struct task_struct *t) | |
738 | { | |
739 | int cpu; | |
740 | ||
741 | barrier(); | |
742 | cpu = task_cpu(t); | |
743 | if (!task_curr(t)) | |
744 | return; /* This task is not running on that CPU. */ | |
2dba13f0 | 745 | smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true); |
bcbfdd01 PM |
746 | } |
747 | ||
9b9500da | 748 | /* |
277ffe1b | 749 | * When trying to report a quiescent state on behalf of some other CPU, |
9b9500da | 750 | * it is our responsibility to check for and handle potential overflow |
a66ae8ae | 751 | * of the rcu_node ->gp_seq counter with respect to the rcu_data counters. |
9b9500da PM |
752 | * After all, the CPU might be in deep idle state, and thus executing no |
753 | * code whatsoever. | |
754 | */ | |
755 | static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp) | |
756 | { | |
a32e01ee | 757 | raw_lockdep_assert_held_rcu_node(rnp); |
a66ae8ae PM |
758 | if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4, |
759 | rnp->gp_seq)) | |
9b9500da | 760 | WRITE_ONCE(rdp->gpwrap, true); |
8aa670cd PM |
761 | if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq)) |
762 | rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4; | |
9b9500da PM |
763 | } |
764 | ||
64db4cff PM |
765 | /* |
766 | * Snapshot the specified CPU's dynticks counter so that we can later | |
767 | * credit them with an implicit quiescent state. Return 1 if this CPU | |
1eba8f84 | 768 | * is in dynticks idle mode, which is an extended quiescent state. |
64db4cff | 769 | */ |
fe5ac724 | 770 | static int dyntick_save_progress_counter(struct rcu_data *rdp) |
64db4cff | 771 | { |
62e2412d | 772 | rdp->dynticks_snap = rcu_dynticks_snap(rdp->cpu); |
02a5c550 | 773 | if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) { |
88d1bead | 774 | trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti")); |
9b9500da | 775 | rcu_gpnum_ovf(rdp->mynode, rdp); |
23a9bacd | 776 | return 1; |
7941dbde | 777 | } |
23a9bacd | 778 | return 0; |
64db4cff PM |
779 | } |
780 | ||
781 | /* | |
85d68222 PZ |
782 | * Returns positive if the specified CPU has passed through a quiescent state |
783 | * by virtue of being in or having passed through an dynticks idle state since | |
784 | * the last call to dyntick_save_progress_counter() for this same CPU, or by | |
785 | * virtue of having been offline. | |
786 | * | |
787 | * Returns negative if the specified CPU needs a force resched. | |
788 | * | |
789 | * Returns zero otherwise. | |
64db4cff | 790 | */ |
fe5ac724 | 791 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp) |
64db4cff | 792 | { |
3a19b46a | 793 | unsigned long jtsq; |
85d68222 | 794 | int ret = 0; |
9b9500da | 795 | struct rcu_node *rnp = rdp->mynode; |
64db4cff PM |
796 | |
797 | /* | |
798 | * If the CPU passed through or entered a dynticks idle phase with | |
799 | * no active irq/NMI handlers, then we can safely pretend that the CPU | |
800 | * already acknowledged the request to pass through a quiescent | |
801 | * state. Either way, that CPU cannot possibly be in an RCU | |
802 | * read-side critical section that started before the beginning | |
803 | * of the current RCU grace period. | |
804 | */ | |
dc5a4f29 | 805 | if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) { |
88d1bead | 806 | trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti")); |
9b9500da | 807 | rcu_gpnum_ovf(rnp, rdp); |
3a19b46a PM |
808 | return 1; |
809 | } | |
810 | ||
666ca290 JFG |
811 | /* |
812 | * Complain if a CPU that is considered to be offline from RCU's | |
813 | * perspective has not yet reported a quiescent state. After all, | |
814 | * the offline CPU should have reported a quiescent state during | |
815 | * the CPU-offline process, or, failing that, by rcu_gp_init() | |
816 | * if it ran concurrently with either the CPU going offline or the | |
817 | * last task on a leaf rcu_node structure exiting its RCU read-side | |
818 | * critical section while all CPUs corresponding to that structure | |
819 | * are offline. This added warning detects bugs in any of these | |
820 | * code paths. | |
821 | * | |
822 | * The rcu_node structure's ->lock is held here, which excludes | |
823 | * the relevant portions the CPU-hotplug code, the grace-period | |
824 | * initialization code, and the rcu_read_unlock() code paths. | |
825 | * | |
826 | * For more detail, please refer to the "Hotplug CPU" section | |
827 | * of RCU's Requirements documentation. | |
828 | */ | |
5ae0f1b5 | 829 | if (WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp))) { |
f2e2df59 PM |
830 | struct rcu_node *rnp1; |
831 | ||
f2e2df59 PM |
832 | pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n", |
833 | __func__, rnp->grplo, rnp->grphi, rnp->level, | |
834 | (long)rnp->gp_seq, (long)rnp->completedqs); | |
835 | for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent) | |
836 | pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n", | |
837 | __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask); | |
f2e2df59 | 838 | pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n", |
5ae0f1b5 | 839 | __func__, rdp->cpu, ".o"[rcu_rdp_cpu_online(rdp)], |
ae2b217a PM |
840 | (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_state, |
841 | (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_state); | |
f2e2df59 PM |
842 | return 1; /* Break things loose after complaining. */ |
843 | } | |
844 | ||
65d798f0 | 845 | /* |
4a81e832 | 846 | * A CPU running for an extended time within the kernel can |
c06aed0e PM |
847 | * delay RCU grace periods: (1) At age jiffies_to_sched_qs, |
848 | * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set | |
7e28c5af PM |
849 | * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the |
850 | * unsynchronized assignments to the per-CPU rcu_need_heavy_qs | |
851 | * variable are safe because the assignments are repeated if this | |
852 | * CPU failed to pass through a quiescent state. This code | |
c06aed0e | 853 | * also checks .jiffies_resched in case jiffies_to_sched_qs |
7e28c5af | 854 | * is set way high. |
6193c76a | 855 | */ |
c06aed0e | 856 | jtsq = READ_ONCE(jiffies_to_sched_qs); |
88ee23ef | 857 | if (!READ_ONCE(rdp->rcu_need_heavy_qs) && |
7e28c5af | 858 | (time_after(jiffies, rcu_state.gp_start + jtsq * 2) || |
b2b00ddf PM |
859 | time_after(jiffies, rcu_state.jiffies_resched) || |
860 | rcu_state.cbovld)) { | |
88ee23ef | 861 | WRITE_ONCE(rdp->rcu_need_heavy_qs, true); |
9226b10d | 862 | /* Store rcu_need_heavy_qs before rcu_urgent_qs. */ |
9424b867 | 863 | smp_store_release(&rdp->rcu_urgent_qs, true); |
7e28c5af | 864 | } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) { |
9424b867 | 865 | WRITE_ONCE(rdp->rcu_urgent_qs, true); |
6193c76a PM |
866 | } |
867 | ||
28053bc7 | 868 | /* |
c98cac60 | 869 | * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq! |
d3052109 PM |
870 | * The above code handles this, but only for straight cond_resched(). |
871 | * And some in-kernel loops check need_resched() before calling | |
872 | * cond_resched(), which defeats the above code for CPUs that are | |
873 | * running in-kernel with scheduling-clock interrupts disabled. | |
874 | * So hit them over the head with the resched_cpu() hammer! | |
28053bc7 | 875 | */ |
d3052109 | 876 | if (tick_nohz_full_cpu(rdp->cpu) && |
b2b00ddf PM |
877 | (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) || |
878 | rcu_state.cbovld)) { | |
9424b867 | 879 | WRITE_ONCE(rdp->rcu_urgent_qs, true); |
d3052109 | 880 | WRITE_ONCE(rdp->last_fqs_resched, jiffies); |
85d68222 | 881 | ret = -1; |
d3052109 PM |
882 | } |
883 | ||
884 | /* | |
885 | * If more than halfway to RCU CPU stall-warning time, invoke | |
886 | * resched_cpu() more frequently to try to loosen things up a bit. | |
887 | * Also check to see if the CPU is getting hammered with interrupts, | |
888 | * but only once per grace period, just to keep the IPIs down to | |
889 | * a dull roar. | |
890 | */ | |
891 | if (time_after(jiffies, rcu_state.jiffies_resched)) { | |
892 | if (time_after(jiffies, | |
893 | READ_ONCE(rdp->last_fqs_resched) + jtsq)) { | |
d3052109 | 894 | WRITE_ONCE(rdp->last_fqs_resched, jiffies); |
85d68222 | 895 | ret = -1; |
d3052109 | 896 | } |
9b9500da | 897 | if (IS_ENABLED(CONFIG_IRQ_WORK) && |
8aa670cd | 898 | !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq && |
9b9500da | 899 | (rnp->ffmask & rdp->grpmask)) { |
9b9500da | 900 | rdp->rcu_iw_pending = true; |
8aa670cd | 901 | rdp->rcu_iw_gp_seq = rnp->gp_seq; |
9b9500da PM |
902 | irq_work_queue_on(&rdp->rcu_iw, rdp->cpu); |
903 | } | |
be42f00b ZL |
904 | |
905 | if (rcu_cpu_stall_cputime && rdp->snap_record.gp_seq != rdp->gp_seq) { | |
906 | int cpu = rdp->cpu; | |
907 | struct rcu_snap_record *rsrp; | |
908 | struct kernel_cpustat *kcsp; | |
909 | ||
910 | kcsp = &kcpustat_cpu(cpu); | |
911 | ||
912 | rsrp = &rdp->snap_record; | |
913 | rsrp->cputime_irq = kcpustat_field(kcsp, CPUTIME_IRQ, cpu); | |
914 | rsrp->cputime_softirq = kcpustat_field(kcsp, CPUTIME_SOFTIRQ, cpu); | |
915 | rsrp->cputime_system = kcpustat_field(kcsp, CPUTIME_SYSTEM, cpu); | |
916 | rsrp->nr_hardirqs = kstat_cpu_irqs_sum(rdp->cpu); | |
917 | rsrp->nr_softirqs = kstat_cpu_softirqs_sum(rdp->cpu); | |
918 | rsrp->nr_csw = nr_context_switches_cpu(rdp->cpu); | |
919 | rsrp->jiffies = jiffies; | |
920 | rsrp->gp_seq = rdp->gp_seq; | |
921 | } | |
9b9500da | 922 | } |
4914950a | 923 | |
85d68222 | 924 | return ret; |
64db4cff PM |
925 | } |
926 | ||
41e80595 PM |
927 | /* Trace-event wrapper function for trace_rcu_future_grace_period. */ |
928 | static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp, | |
b73de91d | 929 | unsigned long gp_seq_req, const char *s) |
0446be48 | 930 | { |
0937d045 PM |
931 | trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq), |
932 | gp_seq_req, rnp->level, | |
933 | rnp->grplo, rnp->grphi, s); | |
0446be48 PM |
934 | } |
935 | ||
936 | /* | |
b73de91d | 937 | * rcu_start_this_gp - Request the start of a particular grace period |
df2bf8f7 | 938 | * @rnp_start: The leaf node of the CPU from which to start. |
b73de91d JF |
939 | * @rdp: The rcu_data corresponding to the CPU from which to start. |
940 | * @gp_seq_req: The gp_seq of the grace period to start. | |
941 | * | |
41e80595 | 942 | * Start the specified grace period, as needed to handle newly arrived |
0446be48 | 943 | * callbacks. The required future grace periods are recorded in each |
7a1d0f23 | 944 | * rcu_node structure's ->gp_seq_needed field. Returns true if there |
48a7639c | 945 | * is reason to awaken the grace-period kthread. |
0446be48 | 946 | * |
d5cd9685 PM |
947 | * The caller must hold the specified rcu_node structure's ->lock, which |
948 | * is why the caller is responsible for waking the grace-period kthread. | |
b73de91d JF |
949 | * |
950 | * Returns true if the GP thread needs to be awakened else false. | |
0446be48 | 951 | */ |
df2bf8f7 | 952 | static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp, |
b73de91d | 953 | unsigned long gp_seq_req) |
0446be48 | 954 | { |
48a7639c | 955 | bool ret = false; |
df2bf8f7 | 956 | struct rcu_node *rnp; |
0446be48 PM |
957 | |
958 | /* | |
360e0da6 PM |
959 | * Use funnel locking to either acquire the root rcu_node |
960 | * structure's lock or bail out if the need for this grace period | |
df2bf8f7 JFG |
961 | * has already been recorded -- or if that grace period has in |
962 | * fact already started. If there is already a grace period in | |
963 | * progress in a non-leaf node, no recording is needed because the | |
964 | * end of the grace period will scan the leaf rcu_node structures. | |
965 | * Note that rnp_start->lock must not be released. | |
0446be48 | 966 | */ |
df2bf8f7 JFG |
967 | raw_lockdep_assert_held_rcu_node(rnp_start); |
968 | trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf")); | |
969 | for (rnp = rnp_start; 1; rnp = rnp->parent) { | |
970 | if (rnp != rnp_start) | |
971 | raw_spin_lock_rcu_node(rnp); | |
972 | if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) || | |
973 | rcu_seq_started(&rnp->gp_seq, gp_seq_req) || | |
974 | (rnp != rnp_start && | |
975 | rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) { | |
976 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, | |
b73de91d | 977 | TPS("Prestarted")); |
360e0da6 PM |
978 | goto unlock_out; |
979 | } | |
8ff37290 | 980 | WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req); |
226ca5e7 | 981 | if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) { |
a2165e41 | 982 | /* |
226ca5e7 JFG |
983 | * We just marked the leaf or internal node, and a |
984 | * grace period is in progress, which means that | |
985 | * rcu_gp_cleanup() will see the marking. Bail to | |
986 | * reduce contention. | |
a2165e41 | 987 | */ |
df2bf8f7 | 988 | trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, |
b73de91d | 989 | TPS("Startedleaf")); |
a2165e41 PM |
990 | goto unlock_out; |
991 | } | |
df2bf8f7 JFG |
992 | if (rnp != rnp_start && rnp->parent != NULL) |
993 | raw_spin_unlock_rcu_node(rnp); | |
994 | if (!rnp->parent) | |
360e0da6 | 995 | break; /* At root, and perhaps also leaf. */ |
0446be48 PM |
996 | } |
997 | ||
360e0da6 | 998 | /* If GP already in progress, just leave, otherwise start one. */ |
de8e8730 | 999 | if (rcu_gp_in_progress()) { |
df2bf8f7 | 1000 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot")); |
0446be48 PM |
1001 | goto unlock_out; |
1002 | } | |
df2bf8f7 | 1003 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot")); |
9cbc5b97 | 1004 | WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT); |
2906d215 | 1005 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); |
5648d659 | 1006 | if (!READ_ONCE(rcu_state.gp_kthread)) { |
df2bf8f7 | 1007 | trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread")); |
360e0da6 | 1008 | goto unlock_out; |
0446be48 | 1009 | } |
62ae1951 | 1010 | trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq")); |
360e0da6 | 1011 | ret = true; /* Caller must wake GP kthread. */ |
0446be48 | 1012 | unlock_out: |
ab5e869c | 1013 | /* Push furthest requested GP to leaf node and rcu_data structure. */ |
df2bf8f7 | 1014 | if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) { |
8ff37290 PM |
1015 | WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed); |
1016 | WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed); | |
ab5e869c | 1017 | } |
df2bf8f7 JFG |
1018 | if (rnp != rnp_start) |
1019 | raw_spin_unlock_rcu_node(rnp); | |
48a7639c | 1020 | return ret; |
0446be48 PM |
1021 | } |
1022 | ||
1023 | /* | |
1024 | * Clean up any old requests for the just-ended grace period. Also return | |
d1e4f01d | 1025 | * whether any additional grace periods have been requested. |
0446be48 | 1026 | */ |
3481f2ea | 1027 | static bool rcu_future_gp_cleanup(struct rcu_node *rnp) |
0446be48 | 1028 | { |
fb31340f | 1029 | bool needmore; |
da1df50d | 1030 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
0446be48 | 1031 | |
7a1d0f23 PM |
1032 | needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed); |
1033 | if (!needmore) | |
1034 | rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */ | |
b73de91d | 1035 | trace_rcu_this_gp(rnp, rdp, rnp->gp_seq, |
41e80595 | 1036 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); |
0446be48 PM |
1037 | return needmore; |
1038 | } | |
1039 | ||
e787644c FW |
1040 | static void swake_up_one_online_ipi(void *arg) |
1041 | { | |
1042 | struct swait_queue_head *wqh = arg; | |
1043 | ||
1044 | swake_up_one(wqh); | |
1045 | } | |
1046 | ||
1047 | static void swake_up_one_online(struct swait_queue_head *wqh) | |
1048 | { | |
1049 | int cpu = get_cpu(); | |
1050 | ||
1051 | /* | |
1052 | * If called from rcutree_report_cpu_starting(), wake up | |
1053 | * is dangerous that late in the CPU-down hotplug process. The | |
1054 | * scheduler might queue an ignored hrtimer. Defer the wake up | |
1055 | * to an online CPU instead. | |
1056 | */ | |
1057 | if (unlikely(cpu_is_offline(cpu))) { | |
1058 | int target; | |
1059 | ||
1060 | target = cpumask_any_and(housekeeping_cpumask(HK_TYPE_RCU), | |
1061 | cpu_online_mask); | |
1062 | ||
1063 | smp_call_function_single(target, swake_up_one_online_ipi, | |
1064 | wqh, 0); | |
1065 | put_cpu(); | |
1066 | } else { | |
1067 | put_cpu(); | |
1068 | swake_up_one(wqh); | |
1069 | } | |
1070 | } | |
1071 | ||
48a7639c | 1072 | /* |
5648d659 PM |
1073 | * Awaken the grace-period kthread. Don't do a self-awaken (unless in an |
1074 | * interrupt or softirq handler, in which case we just might immediately | |
1075 | * sleep upon return, resulting in a grace-period hang), and don't bother | |
1076 | * awakening when there is nothing for the grace-period kthread to do | |
1077 | * (as in several CPUs raced to awaken, we lost), and finally don't try | |
1078 | * to awaken a kthread that has not yet been created. If all those checks | |
1079 | * are passed, track some debug information and awaken. | |
1d1f898d ZJ |
1080 | * |
1081 | * So why do the self-wakeup when in an interrupt or softirq handler | |
1082 | * in the grace-period kthread's context? Because the kthread might have | |
1083 | * been interrupted just as it was going to sleep, and just after the final | |
1084 | * pre-sleep check of the awaken condition. In this case, a wakeup really | |
1085 | * is required, and is therefore supplied. | |
48a7639c | 1086 | */ |
532c00c9 | 1087 | static void rcu_gp_kthread_wake(void) |
48a7639c | 1088 | { |
5648d659 PM |
1089 | struct task_struct *t = READ_ONCE(rcu_state.gp_kthread); |
1090 | ||
2407a64f | 1091 | if ((current == t && !in_hardirq() && !in_serving_softirq()) || |
5648d659 | 1092 | !READ_ONCE(rcu_state.gp_flags) || !t) |
48a7639c | 1093 | return; |
fd897573 PM |
1094 | WRITE_ONCE(rcu_state.gp_wake_time, jiffies); |
1095 | WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq)); | |
e787644c | 1096 | swake_up_one_online(&rcu_state.gp_wq); |
48a7639c PM |
1097 | } |
1098 | ||
dc35c893 | 1099 | /* |
29365e56 PM |
1100 | * If there is room, assign a ->gp_seq number to any callbacks on this |
1101 | * CPU that have not already been assigned. Also accelerate any callbacks | |
1102 | * that were previously assigned a ->gp_seq number that has since proven | |
1103 | * to be too conservative, which can happen if callbacks get assigned a | |
1104 | * ->gp_seq number while RCU is idle, but with reference to a non-root | |
1105 | * rcu_node structure. This function is idempotent, so it does not hurt | |
1106 | * to call it repeatedly. Returns an flag saying that we should awaken | |
1107 | * the RCU grace-period kthread. | |
dc35c893 PM |
1108 | * |
1109 | * The caller must hold rnp->lock with interrupts disabled. | |
1110 | */ | |
02f50142 | 1111 | static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp) |
dc35c893 | 1112 | { |
b73de91d | 1113 | unsigned long gp_seq_req; |
15fecf89 | 1114 | bool ret = false; |
dc35c893 | 1115 | |
d1b222c6 | 1116 | rcu_lockdep_assert_cblist_protected(rdp); |
a32e01ee | 1117 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1118 | |
15fecf89 PM |
1119 | /* If no pending (not yet ready to invoke) callbacks, nothing to do. */ |
1120 | if (!rcu_segcblist_pend_cbs(&rdp->cblist)) | |
48a7639c | 1121 | return false; |
dc35c893 | 1122 | |
3afe7fa5 JFG |
1123 | trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc")); |
1124 | ||
dc35c893 | 1125 | /* |
15fecf89 PM |
1126 | * Callbacks are often registered with incomplete grace-period |
1127 | * information. Something about the fact that getting exact | |
1128 | * information requires acquiring a global lock... RCU therefore | |
1129 | * makes a conservative estimate of the grace period number at which | |
1130 | * a given callback will become ready to invoke. The following | |
1131 | * code checks this estimate and improves it when possible, thus | |
1132 | * accelerating callback invocation to an earlier grace-period | |
1133 | * number. | |
dc35c893 | 1134 | */ |
9cbc5b97 | 1135 | gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq); |
b73de91d JF |
1136 | if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req)) |
1137 | ret = rcu_start_this_gp(rnp, rdp, gp_seq_req); | |
6d4b418c PM |
1138 | |
1139 | /* Trace depending on how much we were able to accelerate. */ | |
15fecf89 | 1140 | if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL)) |
a7886e89 | 1141 | trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB")); |
6d4b418c | 1142 | else |
a7886e89 JFG |
1143 | trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB")); |
1144 | ||
3afe7fa5 JFG |
1145 | trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc")); |
1146 | ||
48a7639c | 1147 | return ret; |
dc35c893 PM |
1148 | } |
1149 | ||
e44e73ca PM |
1150 | /* |
1151 | * Similar to rcu_accelerate_cbs(), but does not require that the leaf | |
1152 | * rcu_node structure's ->lock be held. It consults the cached value | |
1153 | * of ->gp_seq_needed in the rcu_data structure, and if that indicates | |
1154 | * that a new grace-period request be made, invokes rcu_accelerate_cbs() | |
1155 | * while holding the leaf rcu_node structure's ->lock. | |
1156 | */ | |
c6e09b97 | 1157 | static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp, |
e44e73ca PM |
1158 | struct rcu_data *rdp) |
1159 | { | |
1160 | unsigned long c; | |
1161 | bool needwake; | |
1162 | ||
d1b222c6 | 1163 | rcu_lockdep_assert_cblist_protected(rdp); |
c6e09b97 | 1164 | c = rcu_seq_snap(&rcu_state.gp_seq); |
a5b89501 | 1165 | if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) { |
e44e73ca PM |
1166 | /* Old request still live, so mark recent callbacks. */ |
1167 | (void)rcu_segcblist_accelerate(&rdp->cblist, c); | |
1168 | return; | |
1169 | } | |
1170 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ | |
02f50142 | 1171 | needwake = rcu_accelerate_cbs(rnp, rdp); |
e44e73ca PM |
1172 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
1173 | if (needwake) | |
532c00c9 | 1174 | rcu_gp_kthread_wake(); |
e44e73ca PM |
1175 | } |
1176 | ||
dc35c893 PM |
1177 | /* |
1178 | * Move any callbacks whose grace period has completed to the | |
1179 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and | |
29365e56 | 1180 | * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL |
dc35c893 PM |
1181 | * sublist. This function is idempotent, so it does not hurt to |
1182 | * invoke it repeatedly. As long as it is not invoked -too- often... | |
48a7639c | 1183 | * Returns true if the RCU grace-period kthread needs to be awakened. |
dc35c893 PM |
1184 | * |
1185 | * The caller must hold rnp->lock with interrupts disabled. | |
1186 | */ | |
834f56bf | 1187 | static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp) |
dc35c893 | 1188 | { |
d1b222c6 | 1189 | rcu_lockdep_assert_cblist_protected(rdp); |
a32e01ee | 1190 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1191 | |
15fecf89 PM |
1192 | /* If no pending (not yet ready to invoke) callbacks, nothing to do. */ |
1193 | if (!rcu_segcblist_pend_cbs(&rdp->cblist)) | |
48a7639c | 1194 | return false; |
dc35c893 PM |
1195 | |
1196 | /* | |
29365e56 | 1197 | * Find all callbacks whose ->gp_seq numbers indicate that they |
dc35c893 PM |
1198 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. |
1199 | */ | |
29365e56 | 1200 | rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq); |
dc35c893 PM |
1201 | |
1202 | /* Classify any remaining callbacks. */ | |
02f50142 | 1203 | return rcu_accelerate_cbs(rnp, rdp); |
dc35c893 PM |
1204 | } |
1205 | ||
7f36ef82 PM |
1206 | /* |
1207 | * Move and classify callbacks, but only if doing so won't require | |
1208 | * that the RCU grace-period kthread be awakened. | |
1209 | */ | |
1210 | static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp, | |
1211 | struct rcu_data *rdp) | |
1212 | { | |
d1b222c6 | 1213 | rcu_lockdep_assert_cblist_protected(rdp); |
614ddad1 | 1214 | if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) || !raw_spin_trylock_rcu_node(rnp)) |
7f36ef82 | 1215 | return; |
614ddad1 PM |
1216 | // The grace period cannot end while we hold the rcu_node lock. |
1217 | if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) | |
1218 | WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp)); | |
6608c3a0 | 1219 | raw_spin_unlock_rcu_node(rnp); |
7f36ef82 PM |
1220 | } |
1221 | ||
1a2f5d57 PM |
1222 | /* |
1223 | * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a | |
1224 | * quiescent state. This is intended to be invoked when the CPU notices | |
1225 | * a new grace period. | |
1226 | */ | |
1227 | static void rcu_strict_gp_check_qs(void) | |
1228 | { | |
1229 | if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) { | |
1230 | rcu_read_lock(); | |
1231 | rcu_read_unlock(); | |
1232 | } | |
1233 | } | |
1234 | ||
d09b62df | 1235 | /* |
ba9fbe95 PM |
1236 | * Update CPU-local rcu_data state to record the beginnings and ends of |
1237 | * grace periods. The caller must hold the ->lock of the leaf rcu_node | |
1238 | * structure corresponding to the current CPU, and must have irqs disabled. | |
48a7639c | 1239 | * Returns true if the grace-period kthread needs to be awakened. |
d09b62df | 1240 | */ |
c7e48f7b | 1241 | static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp) |
d09b62df | 1242 | { |
5d6742b3 | 1243 | bool ret = false; |
b5ea0370 | 1244 | bool need_qs; |
3820b513 | 1245 | const bool offloaded = rcu_rdp_is_offloaded(rdp); |
48a7639c | 1246 | |
a32e01ee | 1247 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 1248 | |
67e14c1e PM |
1249 | if (rdp->gp_seq == rnp->gp_seq) |
1250 | return false; /* Nothing to do. */ | |
d09b62df | 1251 | |
67e14c1e PM |
1252 | /* Handle the ends of any preceding grace periods first. */ |
1253 | if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) || | |
1254 | unlikely(READ_ONCE(rdp->gpwrap))) { | |
5d6742b3 PM |
1255 | if (!offloaded) |
1256 | ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */ | |
b5ea0370 | 1257 | rdp->core_needs_qs = false; |
9cbc5b97 | 1258 | trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend")); |
67e14c1e | 1259 | } else { |
5d6742b3 PM |
1260 | if (!offloaded) |
1261 | ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */ | |
b5ea0370 PM |
1262 | if (rdp->core_needs_qs) |
1263 | rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask); | |
d09b62df | 1264 | } |
398ebe60 | 1265 | |
67e14c1e PM |
1266 | /* Now handle the beginnings of any new-to-this-CPU grace periods. */ |
1267 | if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) || | |
1268 | unlikely(READ_ONCE(rdp->gpwrap))) { | |
6eaef633 PM |
1269 | /* |
1270 | * If the current grace period is waiting for this CPU, | |
1271 | * set up to detect a quiescent state, otherwise don't | |
1272 | * go looking for one. | |
1273 | */ | |
9cbc5b97 | 1274 | trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart")); |
b5ea0370 PM |
1275 | need_qs = !!(rnp->qsmask & rdp->grpmask); |
1276 | rdp->cpu_no_qs.b.norm = need_qs; | |
1277 | rdp->core_needs_qs = need_qs; | |
6eaef633 PM |
1278 | zero_cpu_stall_ticks(rdp); |
1279 | } | |
67e14c1e | 1280 | rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */ |
13dc7d0c | 1281 | if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap) |
8ff37290 | 1282 | WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed); |
c708b08c PM |
1283 | if (IS_ENABLED(CONFIG_PROVE_RCU) && READ_ONCE(rdp->gpwrap)) |
1284 | WRITE_ONCE(rdp->last_sched_clock, jiffies); | |
3d18469a PM |
1285 | WRITE_ONCE(rdp->gpwrap, false); |
1286 | rcu_gpnum_ovf(rnp, rdp); | |
48a7639c | 1287 | return ret; |
6eaef633 PM |
1288 | } |
1289 | ||
15cabdff | 1290 | static void note_gp_changes(struct rcu_data *rdp) |
6eaef633 PM |
1291 | { |
1292 | unsigned long flags; | |
48a7639c | 1293 | bool needwake; |
6eaef633 PM |
1294 | struct rcu_node *rnp; |
1295 | ||
1296 | local_irq_save(flags); | |
1297 | rnp = rdp->mynode; | |
67e14c1e | 1298 | if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) && |
7d0ae808 | 1299 | !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */ |
2a67e741 | 1300 | !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */ |
6eaef633 PM |
1301 | local_irq_restore(flags); |
1302 | return; | |
1303 | } | |
c7e48f7b | 1304 | needwake = __note_gp_changes(rnp, rdp); |
67c583a7 | 1305 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1a2f5d57 | 1306 | rcu_strict_gp_check_qs(); |
48a7639c | 1307 | if (needwake) |
532c00c9 | 1308 | rcu_gp_kthread_wake(); |
6eaef633 PM |
1309 | } |
1310 | ||
99d6a2ac PM |
1311 | static atomic_t *rcu_gp_slow_suppress; |
1312 | ||
1313 | /* Register a counter to suppress debugging grace-period delays. */ | |
1314 | void rcu_gp_slow_register(atomic_t *rgssp) | |
1315 | { | |
1316 | WARN_ON_ONCE(rcu_gp_slow_suppress); | |
1317 | ||
1318 | WRITE_ONCE(rcu_gp_slow_suppress, rgssp); | |
1319 | } | |
1320 | EXPORT_SYMBOL_GPL(rcu_gp_slow_register); | |
1321 | ||
1322 | /* Unregister a counter, with NULL for not caring which. */ | |
1323 | void rcu_gp_slow_unregister(atomic_t *rgssp) | |
1324 | { | |
0ae9942f | 1325 | WARN_ON_ONCE(rgssp && rgssp != rcu_gp_slow_suppress && rcu_gp_slow_suppress != NULL); |
99d6a2ac PM |
1326 | |
1327 | WRITE_ONCE(rcu_gp_slow_suppress, NULL); | |
1328 | } | |
1329 | EXPORT_SYMBOL_GPL(rcu_gp_slow_unregister); | |
1330 | ||
1331 | static bool rcu_gp_slow_is_suppressed(void) | |
1332 | { | |
1333 | atomic_t *rgssp = READ_ONCE(rcu_gp_slow_suppress); | |
1334 | ||
1335 | return rgssp && atomic_read(rgssp); | |
1336 | } | |
1337 | ||
22212332 | 1338 | static void rcu_gp_slow(int delay) |
0f41c0dd | 1339 | { |
99d6a2ac PM |
1340 | if (!rcu_gp_slow_is_suppressed() && delay > 0 && |
1341 | !(rcu_seq_ctr(rcu_state.gp_seq) % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay))) | |
77865dea | 1342 | schedule_timeout_idle(delay); |
0f41c0dd PM |
1343 | } |
1344 | ||
55b2dcf5 PM |
1345 | static unsigned long sleep_duration; |
1346 | ||
1347 | /* Allow rcutorture to stall the grace-period kthread. */ | |
1348 | void rcu_gp_set_torture_wait(int duration) | |
1349 | { | |
1350 | if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0) | |
1351 | WRITE_ONCE(sleep_duration, duration); | |
1352 | } | |
1353 | EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait); | |
1354 | ||
1355 | /* Actually implement the aforementioned wait. */ | |
1356 | static void rcu_gp_torture_wait(void) | |
1357 | { | |
1358 | unsigned long duration; | |
1359 | ||
1360 | if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST)) | |
1361 | return; | |
1362 | duration = xchg(&sleep_duration, 0UL); | |
1363 | if (duration > 0) { | |
1364 | pr_alert("%s: Waiting %lu jiffies\n", __func__, duration); | |
77865dea | 1365 | schedule_timeout_idle(duration); |
55b2dcf5 PM |
1366 | pr_alert("%s: Wait complete\n", __func__); |
1367 | } | |
1368 | } | |
1369 | ||
933ada2c PM |
1370 | /* |
1371 | * Handler for on_each_cpu() to invoke the target CPU's RCU core | |
1372 | * processing. | |
1373 | */ | |
1374 | static void rcu_strict_gp_boundary(void *unused) | |
1375 | { | |
1376 | invoke_rcu_core(); | |
1377 | } | |
1378 | ||
bf95b2bc PM |
1379 | // Make the polled API aware of the beginning of a grace period. |
1380 | static void rcu_poll_gp_seq_start(unsigned long *snap) | |
1381 | { | |
1382 | struct rcu_node *rnp = rcu_get_root(); | |
1383 | ||
3f6c3d29 | 1384 | if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) |
bf95b2bc PM |
1385 | raw_lockdep_assert_held_rcu_node(rnp); |
1386 | ||
1387 | // If RCU was idle, note beginning of GP. | |
1388 | if (!rcu_seq_state(rcu_state.gp_seq_polled)) | |
1389 | rcu_seq_start(&rcu_state.gp_seq_polled); | |
1390 | ||
1391 | // Either way, record current state. | |
1392 | *snap = rcu_state.gp_seq_polled; | |
1393 | } | |
1394 | ||
1395 | // Make the polled API aware of the end of a grace period. | |
1396 | static void rcu_poll_gp_seq_end(unsigned long *snap) | |
1397 | { | |
1398 | struct rcu_node *rnp = rcu_get_root(); | |
1399 | ||
3f6c3d29 | 1400 | if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) |
bf95b2bc PM |
1401 | raw_lockdep_assert_held_rcu_node(rnp); |
1402 | ||
1403 | // If the previously noted GP is still in effect, record the | |
1404 | // end of that GP. Either way, zero counter to avoid counter-wrap | |
1405 | // problems. | |
1406 | if (*snap && *snap == rcu_state.gp_seq_polled) { | |
1407 | rcu_seq_end(&rcu_state.gp_seq_polled); | |
1408 | rcu_state.gp_seq_polled_snap = 0; | |
dd041405 | 1409 | rcu_state.gp_seq_polled_exp_snap = 0; |
bf95b2bc PM |
1410 | } else { |
1411 | *snap = 0; | |
1412 | } | |
1413 | } | |
1414 | ||
1415 | // Make the polled API aware of the beginning of a grace period, but | |
1416 | // where caller does not hold the root rcu_node structure's lock. | |
1417 | static void rcu_poll_gp_seq_start_unlocked(unsigned long *snap) | |
1418 | { | |
31d8aaa8 | 1419 | unsigned long flags; |
bf95b2bc PM |
1420 | struct rcu_node *rnp = rcu_get_root(); |
1421 | ||
1422 | if (rcu_init_invoked()) { | |
3f6c3d29 PM |
1423 | if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) |
1424 | lockdep_assert_irqs_enabled(); | |
31d8aaa8 | 1425 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
bf95b2bc PM |
1426 | } |
1427 | rcu_poll_gp_seq_start(snap); | |
1428 | if (rcu_init_invoked()) | |
31d8aaa8 | 1429 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
bf95b2bc PM |
1430 | } |
1431 | ||
1432 | // Make the polled API aware of the end of a grace period, but where | |
1433 | // caller does not hold the root rcu_node structure's lock. | |
1434 | static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap) | |
1435 | { | |
31d8aaa8 | 1436 | unsigned long flags; |
bf95b2bc PM |
1437 | struct rcu_node *rnp = rcu_get_root(); |
1438 | ||
1439 | if (rcu_init_invoked()) { | |
3f6c3d29 PM |
1440 | if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) |
1441 | lockdep_assert_irqs_enabled(); | |
31d8aaa8 | 1442 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
bf95b2bc PM |
1443 | } |
1444 | rcu_poll_gp_seq_end(snap); | |
1445 | if (rcu_init_invoked()) | |
31d8aaa8 | 1446 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
bf95b2bc PM |
1447 | } |
1448 | ||
988f569a URS |
1449 | /* |
1450 | * There is a single llist, which is used for handling | |
1451 | * synchronize_rcu() users' enqueued rcu_synchronize nodes. | |
1452 | * Within this llist, there are two tail pointers: | |
1453 | * | |
1454 | * wait tail: Tracks the set of nodes, which need to | |
1455 | * wait for the current GP to complete. | |
1456 | * done tail: Tracks the set of nodes, for which grace | |
1457 | * period has elapsed. These nodes processing | |
1458 | * will be done as part of the cleanup work | |
1459 | * execution by a kworker. | |
1460 | * | |
1461 | * At every grace period init, a new wait node is added | |
1462 | * to the llist. This wait node is used as wait tail | |
1463 | * for this new grace period. Given that there are a fixed | |
1464 | * number of wait nodes, if all wait nodes are in use | |
1465 | * (which can happen when kworker callback processing | |
1466 | * is delayed) and additional grace period is requested. | |
1467 | * This means, a system is slow in processing callbacks. | |
1468 | * | |
1469 | * TODO: If a slow processing is detected, a first node | |
1470 | * in the llist should be used as a wait-tail for this | |
1471 | * grace period, therefore users which should wait due | |
1472 | * to a slow process are handled by _this_ grace period | |
1473 | * and not next. | |
1474 | * | |
1475 | * Below is an illustration of how the done and wait | |
1476 | * tail pointers move from one set of rcu_synchronize nodes | |
1477 | * to the other, as grace periods start and finish and | |
1478 | * nodes are processed by kworker. | |
1479 | * | |
1480 | * | |
1481 | * a. Initial llist callbacks list: | |
1482 | * | |
1483 | * +----------+ +--------+ +-------+ | |
1484 | * | | | | | | | |
1485 | * | head |---------> | cb2 |--------->| cb1 | | |
1486 | * | | | | | | | |
1487 | * +----------+ +--------+ +-------+ | |
1488 | * | |
1489 | * | |
1490 | * | |
1491 | * b. New GP1 Start: | |
1492 | * | |
1493 | * WAIT TAIL | |
1494 | * | | |
1495 | * | | |
1496 | * v | |
1497 | * +----------+ +--------+ +--------+ +-------+ | |
1498 | * | | | | | | | | | |
1499 | * | head ------> wait |------> cb2 |------> | cb1 | | |
1500 | * | | | head1 | | | | | | |
1501 | * +----------+ +--------+ +--------+ +-------+ | |
1502 | * | |
1503 | * | |
1504 | * | |
1505 | * c. GP completion: | |
1506 | * | |
1507 | * WAIT_TAIL == DONE_TAIL | |
1508 | * | |
1509 | * DONE TAIL | |
1510 | * | | |
1511 | * | | |
1512 | * v | |
1513 | * +----------+ +--------+ +--------+ +-------+ | |
1514 | * | | | | | | | | | |
1515 | * | head ------> wait |------> cb2 |------> | cb1 | | |
1516 | * | | | head1 | | | | | | |
1517 | * +----------+ +--------+ +--------+ +-------+ | |
1518 | * | |
1519 | * | |
1520 | * | |
1521 | * d. New callbacks and GP2 start: | |
1522 | * | |
1523 | * WAIT TAIL DONE TAIL | |
1524 | * | | | |
1525 | * | | | |
1526 | * v v | |
1527 | * +----------+ +------+ +------+ +------+ +-----+ +-----+ +-----+ | |
1528 | * | | | | | | | | | | | | | | | |
1529 | * | head ------> wait |--->| cb4 |--->| cb3 |--->|wait |--->| cb2 |--->| cb1 | | |
1530 | * | | | head2| | | | | |head1| | | | | | |
1531 | * +----------+ +------+ +------+ +------+ +-----+ +-----+ +-----+ | |
1532 | * | |
1533 | * | |
1534 | * | |
1535 | * e. GP2 completion: | |
1536 | * | |
1537 | * WAIT_TAIL == DONE_TAIL | |
1538 | * DONE TAIL | |
1539 | * | | |
1540 | * | | |
1541 | * v | |
1542 | * +----------+ +------+ +------+ +------+ +-----+ +-----+ +-----+ | |
1543 | * | | | | | | | | | | | | | | | |
1544 | * | head ------> wait |--->| cb4 |--->| cb3 |--->|wait |--->| cb2 |--->| cb1 | | |
1545 | * | | | head2| | | | | |head1| | | | | | |
1546 | * +----------+ +------+ +------+ +------+ +-----+ +-----+ +-----+ | |
1547 | * | |
1548 | * | |
1549 | * While the llist state transitions from d to e, a kworker | |
1550 | * can start executing rcu_sr_normal_gp_cleanup_work() and | |
1551 | * can observe either the old done tail (@c) or the new | |
1552 | * done tail (@e). So, done tail updates and reads need | |
1553 | * to use the rel-acq semantics. If the concurrent kworker | |
1554 | * observes the old done tail, the newly queued work | |
1555 | * execution will process the updated done tail. If the | |
1556 | * concurrent kworker observes the new done tail, then | |
1557 | * the newly queued work will skip processing the done | |
1558 | * tail, as workqueue semantics guarantees that the new | |
1559 | * work is executed only after the previous one completes. | |
1560 | * | |
1561 | * f. kworker callbacks processing complete: | |
1562 | * | |
1563 | * | |
1564 | * DONE TAIL | |
1565 | * | | |
1566 | * | | |
1567 | * v | |
1568 | * +----------+ +--------+ | |
1569 | * | | | | | |
1570 | * | head ------> wait | | |
1571 | * | | | head2 | | |
1572 | * +----------+ +--------+ | |
1573 | * | |
1574 | */ | |
1575 | static bool rcu_sr_is_wait_head(struct llist_node *node) | |
1576 | { | |
1577 | return &(rcu_state.srs_wait_nodes)[0].node <= node && | |
1578 | node <= &(rcu_state.srs_wait_nodes)[SR_NORMAL_GP_WAIT_HEAD_MAX - 1].node; | |
1579 | } | |
1580 | ||
1581 | static struct llist_node *rcu_sr_get_wait_head(void) | |
1582 | { | |
1583 | struct sr_wait_node *sr_wn; | |
1584 | int i; | |
1585 | ||
1586 | for (i = 0; i < SR_NORMAL_GP_WAIT_HEAD_MAX; i++) { | |
1587 | sr_wn = &(rcu_state.srs_wait_nodes)[i]; | |
1588 | ||
1589 | if (!atomic_cmpxchg_acquire(&sr_wn->inuse, 0, 1)) | |
1590 | return &sr_wn->node; | |
1591 | } | |
1592 | ||
1593 | return NULL; | |
1594 | } | |
1595 | ||
1596 | static void rcu_sr_put_wait_head(struct llist_node *node) | |
1597 | { | |
1598 | struct sr_wait_node *sr_wn = container_of(node, struct sr_wait_node, node); | |
1599 | ||
1600 | atomic_set_release(&sr_wn->inuse, 0); | |
1601 | } | |
1602 | ||
1603 | /* Disabled by default. */ | |
1604 | static int rcu_normal_wake_from_gp; | |
1605 | module_param(rcu_normal_wake_from_gp, int, 0644); | |
0fd210ba | 1606 | static struct workqueue_struct *sync_wq; |
988f569a URS |
1607 | |
1608 | static void rcu_sr_normal_complete(struct llist_node *node) | |
1609 | { | |
1610 | struct rcu_synchronize *rs = container_of( | |
1611 | (struct rcu_head *) node, struct rcu_synchronize, head); | |
1612 | unsigned long oldstate = (unsigned long) rs->head.func; | |
1613 | ||
1614 | WARN_ONCE(IS_ENABLED(CONFIG_PROVE_RCU) && | |
1615 | !poll_state_synchronize_rcu(oldstate), | |
1616 | "A full grace period is not passed yet: %lu", | |
1617 | rcu_seq_diff(get_state_synchronize_rcu(), oldstate)); | |
1618 | ||
1619 | /* Finally. */ | |
1620 | complete(&rs->completion); | |
1621 | } | |
1622 | ||
1623 | static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work) | |
1624 | { | |
1625 | struct llist_node *done, *rcu, *next, *head; | |
1626 | ||
1627 | /* | |
1628 | * This work execution can potentially execute | |
1629 | * while a new done tail is being updated by | |
1630 | * grace period kthread in rcu_sr_normal_gp_cleanup(). | |
1631 | * So, read and updates of done tail need to | |
1632 | * follow acq-rel semantics. | |
1633 | * | |
1634 | * Given that wq semantics guarantees that a single work | |
1635 | * cannot execute concurrently by multiple kworkers, | |
1636 | * the done tail list manipulations are protected here. | |
1637 | */ | |
1638 | done = smp_load_acquire(&rcu_state.srs_done_tail); | |
1639 | if (!done) | |
1640 | return; | |
1641 | ||
1642 | WARN_ON_ONCE(!rcu_sr_is_wait_head(done)); | |
1643 | head = done->next; | |
1644 | done->next = NULL; | |
1645 | ||
1646 | /* | |
1647 | * The dummy node, which is pointed to by the | |
1648 | * done tail which is acq-read above is not removed | |
1649 | * here. This allows lockless additions of new | |
1650 | * rcu_synchronize nodes in rcu_sr_normal_add_req(), | |
1651 | * while the cleanup work executes. The dummy | |
1652 | * nodes is removed, in next round of cleanup | |
1653 | * work execution. | |
1654 | */ | |
1655 | llist_for_each_safe(rcu, next, head) { | |
1656 | if (!rcu_sr_is_wait_head(rcu)) { | |
1657 | rcu_sr_normal_complete(rcu); | |
1658 | continue; | |
1659 | } | |
1660 | ||
1661 | rcu_sr_put_wait_head(rcu); | |
1662 | } | |
1663 | } | |
1664 | ||
1665 | /* | |
1666 | * Helper function for rcu_gp_cleanup(). | |
1667 | */ | |
1668 | static void rcu_sr_normal_gp_cleanup(void) | |
1669 | { | |
462df2f5 URS |
1670 | struct llist_node *wait_tail, *next, *rcu; |
1671 | int done = 0; | |
988f569a URS |
1672 | |
1673 | wait_tail = rcu_state.srs_wait_tail; | |
1674 | if (wait_tail == NULL) | |
1675 | return; | |
1676 | ||
1677 | rcu_state.srs_wait_tail = NULL; | |
1678 | ASSERT_EXCLUSIVE_WRITER(rcu_state.srs_wait_tail); | |
462df2f5 URS |
1679 | WARN_ON_ONCE(!rcu_sr_is_wait_head(wait_tail)); |
1680 | ||
1681 | /* | |
1682 | * Process (a) and (d) cases. See an illustration. | |
1683 | */ | |
1684 | llist_for_each_safe(rcu, next, wait_tail->next) { | |
1685 | if (rcu_sr_is_wait_head(rcu)) | |
1686 | break; | |
1687 | ||
1688 | rcu_sr_normal_complete(rcu); | |
1689 | // It can be last, update a next on this step. | |
1690 | wait_tail->next = next; | |
1691 | ||
1692 | if (++done == SR_MAX_USERS_WAKE_FROM_GP) | |
1693 | break; | |
1694 | } | |
988f569a URS |
1695 | |
1696 | // concurrent sr_normal_gp_cleanup work might observe this update. | |
1697 | smp_store_release(&rcu_state.srs_done_tail, wait_tail); | |
1698 | ASSERT_EXCLUSIVE_WRITER(rcu_state.srs_done_tail); | |
1699 | ||
462df2f5 URS |
1700 | /* |
1701 | * We schedule a work in order to perform a final processing | |
1702 | * of outstanding users(if still left) and releasing wait-heads | |
1703 | * added by rcu_sr_normal_gp_init() call. | |
1704 | */ | |
0fd210ba | 1705 | queue_work(sync_wq, &rcu_state.srs_cleanup_work); |
988f569a URS |
1706 | } |
1707 | ||
1708 | /* | |
1709 | * Helper function for rcu_gp_init(). | |
1710 | */ | |
1711 | static bool rcu_sr_normal_gp_init(void) | |
1712 | { | |
1713 | struct llist_node *first; | |
1714 | struct llist_node *wait_head; | |
1715 | bool start_new_poll = false; | |
1716 | ||
1717 | first = READ_ONCE(rcu_state.srs_next.first); | |
1718 | if (!first || rcu_sr_is_wait_head(first)) | |
1719 | return start_new_poll; | |
1720 | ||
1721 | wait_head = rcu_sr_get_wait_head(); | |
1722 | if (!wait_head) { | |
1723 | // Kick another GP to retry. | |
1724 | start_new_poll = true; | |
1725 | return start_new_poll; | |
1726 | } | |
1727 | ||
1728 | /* Inject a wait-dummy-node. */ | |
1729 | llist_add(wait_head, &rcu_state.srs_next); | |
1730 | ||
1731 | /* | |
1732 | * A waiting list of rcu_synchronize nodes should be empty on | |
1733 | * this step, since a GP-kthread, rcu_gp_init() -> gp_cleanup(), | |
1734 | * rolls it over. If not, it is a BUG, warn a user. | |
1735 | */ | |
1736 | WARN_ON_ONCE(rcu_state.srs_wait_tail != NULL); | |
1737 | rcu_state.srs_wait_tail = wait_head; | |
1738 | ASSERT_EXCLUSIVE_WRITER(rcu_state.srs_wait_tail); | |
1739 | ||
1740 | return start_new_poll; | |
1741 | } | |
1742 | ||
1743 | static void rcu_sr_normal_add_req(struct rcu_synchronize *rs) | |
1744 | { | |
1745 | llist_add((struct llist_node *) &rs->head, &rcu_state.srs_next); | |
1746 | } | |
1747 | ||
b3dbec76 | 1748 | /* |
45fed3e7 | 1749 | * Initialize a new grace period. Return false if no grace period required. |
b3dbec76 | 1750 | */ |
f74126dc | 1751 | static noinline_for_stack bool rcu_gp_init(void) |
b3dbec76 | 1752 | { |
ec2c2976 | 1753 | unsigned long flags; |
0aa04b05 | 1754 | unsigned long oldmask; |
ec2c2976 | 1755 | unsigned long mask; |
b3dbec76 | 1756 | struct rcu_data *rdp; |
336a4f6c | 1757 | struct rcu_node *rnp = rcu_get_root(); |
988f569a | 1758 | bool start_new_poll; |
b3dbec76 | 1759 | |
9cbc5b97 | 1760 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
2a67e741 | 1761 | raw_spin_lock_irq_rcu_node(rnp); |
62bb24c4 | 1762 | if (!rcu_state.gp_flags) { |
f7be8209 | 1763 | /* Spurious wakeup, tell caller to go back to sleep. */ |
67c583a7 | 1764 | raw_spin_unlock_irq_rcu_node(rnp); |
45fed3e7 | 1765 | return false; |
f7be8209 | 1766 | } |
9cbc5b97 | 1767 | WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */ |
b3dbec76 | 1768 | |
de8e8730 | 1769 | if (WARN_ON_ONCE(rcu_gp_in_progress())) { |
f7be8209 PM |
1770 | /* |
1771 | * Grace period already in progress, don't start another. | |
1772 | * Not supposed to be able to happen. | |
1773 | */ | |
67c583a7 | 1774 | raw_spin_unlock_irq_rcu_node(rnp); |
45fed3e7 | 1775 | return false; |
7fdefc10 PM |
1776 | } |
1777 | ||
7fdefc10 | 1778 | /* Advance to a new grace period and initialize state. */ |
ad3832e9 | 1779 | record_gp_stall_check_time(); |
ff3bb6f4 | 1780 | /* Record GP times before starting GP, hence rcu_seq_start(). */ |
9cbc5b97 | 1781 | rcu_seq_start(&rcu_state.gp_seq); |
62ae1951 | 1782 | ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); |
988f569a | 1783 | start_new_poll = rcu_sr_normal_gp_init(); |
9cbc5b97 | 1784 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); |
bf95b2bc | 1785 | rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap); |
67c583a7 | 1786 | raw_spin_unlock_irq_rcu_node(rnp); |
7fdefc10 | 1787 | |
988f569a URS |
1788 | /* |
1789 | * The "start_new_poll" is set to true, only when this GP is not able | |
1790 | * to handle anything and there are outstanding users. It happens when | |
1791 | * the rcu_sr_normal_gp_init() function was not able to insert a dummy | |
1792 | * separator to the llist, because there were no left any dummy-nodes. | |
1793 | * | |
1794 | * Number of dummy-nodes is fixed, it could be that we are run out of | |
1795 | * them, if so we start a new pool request to repeat a try. It is rare | |
1796 | * and it means that a system is doing a slow processing of callbacks. | |
1797 | */ | |
1798 | if (start_new_poll) | |
1799 | (void) start_poll_synchronize_rcu(); | |
1800 | ||
0aa04b05 | 1801 | /* |
f37599e6 JFG |
1802 | * Apply per-leaf buffered online and offline operations to |
1803 | * the rcu_node tree. Note that this new grace period need not | |
1804 | * wait for subsequent online CPUs, and that RCU hooks in the CPU | |
1805 | * offlining path, when combined with checks in this function, | |
1806 | * will handle CPUs that are currently going offline or that will | |
1807 | * go offline later. Please also refer to "Hotplug CPU" section | |
1808 | * of RCU's Requirements documentation. | |
0aa04b05 | 1809 | */ |
683954e5 | 1810 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF); |
82980b16 | 1811 | /* Exclude CPU hotplug operations. */ |
aedf4ba9 | 1812 | rcu_for_each_leaf_node(rnp) { |
82980b16 DW |
1813 | local_irq_save(flags); |
1814 | arch_spin_lock(&rcu_state.ofl_lock); | |
1815 | raw_spin_lock_rcu_node(rnp); | |
0aa04b05 PM |
1816 | if (rnp->qsmaskinit == rnp->qsmaskinitnext && |
1817 | !rnp->wait_blkd_tasks) { | |
1818 | /* Nothing to do on this leaf rcu_node structure. */ | |
82980b16 DW |
1819 | raw_spin_unlock_rcu_node(rnp); |
1820 | arch_spin_unlock(&rcu_state.ofl_lock); | |
1821 | local_irq_restore(flags); | |
0aa04b05 PM |
1822 | continue; |
1823 | } | |
1824 | ||
1825 | /* Record old state, apply changes to ->qsmaskinit field. */ | |
1826 | oldmask = rnp->qsmaskinit; | |
1827 | rnp->qsmaskinit = rnp->qsmaskinitnext; | |
1828 | ||
1829 | /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ | |
1830 | if (!oldmask != !rnp->qsmaskinit) { | |
962aff03 PM |
1831 | if (!oldmask) { /* First online CPU for rcu_node. */ |
1832 | if (!rnp->wait_blkd_tasks) /* Ever offline? */ | |
1833 | rcu_init_new_rnp(rnp); | |
1834 | } else if (rcu_preempt_has_tasks(rnp)) { | |
1835 | rnp->wait_blkd_tasks = true; /* blocked tasks */ | |
1836 | } else { /* Last offline CPU and can propagate. */ | |
0aa04b05 | 1837 | rcu_cleanup_dead_rnp(rnp); |
962aff03 | 1838 | } |
0aa04b05 PM |
1839 | } |
1840 | ||
1841 | /* | |
1842 | * If all waited-on tasks from prior grace period are | |
1843 | * done, and if all this rcu_node structure's CPUs are | |
1844 | * still offline, propagate up the rcu_node tree and | |
1845 | * clear ->wait_blkd_tasks. Otherwise, if one of this | |
1846 | * rcu_node structure's CPUs has since come back online, | |
962aff03 | 1847 | * simply clear ->wait_blkd_tasks. |
0aa04b05 PM |
1848 | */ |
1849 | if (rnp->wait_blkd_tasks && | |
962aff03 | 1850 | (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) { |
0aa04b05 | 1851 | rnp->wait_blkd_tasks = false; |
962aff03 PM |
1852 | if (!rnp->qsmaskinit) |
1853 | rcu_cleanup_dead_rnp(rnp); | |
0aa04b05 PM |
1854 | } |
1855 | ||
82980b16 DW |
1856 | raw_spin_unlock_rcu_node(rnp); |
1857 | arch_spin_unlock(&rcu_state.ofl_lock); | |
1858 | local_irq_restore(flags); | |
0aa04b05 | 1859 | } |
22212332 | 1860 | rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */ |
7fdefc10 PM |
1861 | |
1862 | /* | |
1863 | * Set the quiescent-state-needed bits in all the rcu_node | |
9cbc5b97 PM |
1864 | * structures for all currently online CPUs in breadth-first |
1865 | * order, starting from the root rcu_node structure, relying on the | |
1866 | * layout of the tree within the rcu_state.node[] array. Note that | |
1867 | * other CPUs will access only the leaves of the hierarchy, thus | |
1868 | * seeing that no grace period is in progress, at least until the | |
1869 | * corresponding leaf node has been initialized. | |
7fdefc10 PM |
1870 | * |
1871 | * The grace period cannot complete until the initialization | |
1872 | * process finishes, because this kthread handles both. | |
1873 | */ | |
683954e5 | 1874 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT); |
aedf4ba9 | 1875 | rcu_for_each_node_breadth_first(rnp) { |
22212332 | 1876 | rcu_gp_slow(gp_init_delay); |
ec2c2976 | 1877 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
da1df50d | 1878 | rdp = this_cpu_ptr(&rcu_data); |
81ab59a3 | 1879 | rcu_preempt_check_blocked_tasks(rnp); |
7fdefc10 | 1880 | rnp->qsmask = rnp->qsmaskinit; |
9cbc5b97 | 1881 | WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq); |
7fdefc10 | 1882 | if (rnp == rdp->mynode) |
c7e48f7b | 1883 | (void)__note_gp_changes(rnp, rdp); |
7fdefc10 | 1884 | rcu_preempt_boost_start_gp(rnp); |
9cbc5b97 | 1885 | trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq, |
7fdefc10 PM |
1886 | rnp->level, rnp->grplo, |
1887 | rnp->grphi, rnp->qsmask); | |
ec2c2976 PM |
1888 | /* Quiescent states for tasks on any now-offline CPUs. */ |
1889 | mask = rnp->qsmask & ~rnp->qsmaskinitnext; | |
f2e2df59 | 1890 | rnp->rcu_gp_init_mask = mask; |
ec2c2976 | 1891 | if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp)) |
b50912d0 | 1892 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
ec2c2976 PM |
1893 | else |
1894 | raw_spin_unlock_irq_rcu_node(rnp); | |
cee43939 | 1895 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 1896 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
7fdefc10 | 1897 | } |
b3dbec76 | 1898 | |
933ada2c PM |
1899 | // If strict, make all CPUs aware of new grace period. |
1900 | if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) | |
1901 | on_each_cpu(rcu_strict_gp_boundary, NULL, 0); | |
1902 | ||
45fed3e7 | 1903 | return true; |
7fdefc10 | 1904 | } |
b3dbec76 | 1905 | |
b9a425cf | 1906 | /* |
b3dae109 | 1907 | * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state |
d5374226 | 1908 | * time. |
b9a425cf | 1909 | */ |
0854a05c | 1910 | static bool rcu_gp_fqs_check_wake(int *gfp) |
b9a425cf | 1911 | { |
336a4f6c | 1912 | struct rcu_node *rnp = rcu_get_root(); |
b9a425cf | 1913 | |
1fca4d12 PM |
1914 | // If under overload conditions, force an immediate FQS scan. |
1915 | if (*gfp & RCU_GP_FLAG_OVLD) | |
1916 | return true; | |
1917 | ||
1918 | // Someone like call_rcu() requested a force-quiescent-state scan. | |
0854a05c | 1919 | *gfp = READ_ONCE(rcu_state.gp_flags); |
b9a425cf PM |
1920 | if (*gfp & RCU_GP_FLAG_FQS) |
1921 | return true; | |
1922 | ||
1fca4d12 | 1923 | // The current grace period has completed. |
b9a425cf PM |
1924 | if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp)) |
1925 | return true; | |
1926 | ||
1927 | return false; | |
1928 | } | |
1929 | ||
4cdfc175 PM |
1930 | /* |
1931 | * Do one round of quiescent-state forcing. | |
1932 | */ | |
0854a05c | 1933 | static void rcu_gp_fqs(bool first_time) |
4cdfc175 | 1934 | { |
b96e7a5f | 1935 | int nr_fqs = READ_ONCE(rcu_state.nr_fqs_jiffies_stall); |
336a4f6c | 1936 | struct rcu_node *rnp = rcu_get_root(); |
4cdfc175 | 1937 | |
9cbc5b97 | 1938 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
2431774f | 1939 | WRITE_ONCE(rcu_state.n_force_qs, rcu_state.n_force_qs + 1); |
b96e7a5f JFG |
1940 | |
1941 | WARN_ON_ONCE(nr_fqs > 3); | |
1942 | /* Only countdown nr_fqs for stall purposes if jiffies moves. */ | |
1943 | if (nr_fqs) { | |
1944 | if (nr_fqs == 1) { | |
1945 | WRITE_ONCE(rcu_state.jiffies_stall, | |
1946 | jiffies + rcu_jiffies_till_stall_check()); | |
1947 | } | |
1948 | WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, --nr_fqs); | |
1949 | } | |
1950 | ||
77f81fe0 | 1951 | if (first_time) { |
4cdfc175 | 1952 | /* Collect dyntick-idle snapshots. */ |
e9ecb780 | 1953 | force_qs_rnp(dyntick_save_progress_counter); |
4cdfc175 PM |
1954 | } else { |
1955 | /* Handle dyntick-idle and offline CPUs. */ | |
e9ecb780 | 1956 | force_qs_rnp(rcu_implicit_dynticks_qs); |
4cdfc175 PM |
1957 | } |
1958 | /* Clear flag to prevent immediate re-entry. */ | |
9cbc5b97 | 1959 | if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) { |
2a67e741 | 1960 | raw_spin_lock_irq_rcu_node(rnp); |
62bb24c4 | 1961 | WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & ~RCU_GP_FLAG_FQS); |
67c583a7 | 1962 | raw_spin_unlock_irq_rcu_node(rnp); |
4cdfc175 | 1963 | } |
4cdfc175 PM |
1964 | } |
1965 | ||
c3854a05 PM |
1966 | /* |
1967 | * Loop doing repeated quiescent-state forcing until the grace period ends. | |
1968 | */ | |
f74126dc | 1969 | static noinline_for_stack void rcu_gp_fqs_loop(void) |
c3854a05 | 1970 | { |
9bdb5b3a | 1971 | bool first_gp_fqs = true; |
1fca4d12 | 1972 | int gf = 0; |
c3854a05 PM |
1973 | unsigned long j; |
1974 | int ret; | |
1975 | struct rcu_node *rnp = rcu_get_root(); | |
1976 | ||
c06aed0e | 1977 | j = READ_ONCE(jiffies_till_first_fqs); |
1fca4d12 PM |
1978 | if (rcu_state.cbovld) |
1979 | gf = RCU_GP_FLAG_OVLD; | |
c3854a05 PM |
1980 | ret = 0; |
1981 | for (;;) { | |
fb77dccf PM |
1982 | if (rcu_state.cbovld) { |
1983 | j = (j + 2) / 3; | |
1984 | if (j <= 0) | |
1985 | j = 1; | |
1986 | } | |
1987 | if (!ret || time_before(jiffies + j, rcu_state.jiffies_force_qs)) { | |
683954e5 NU |
1988 | WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j); |
1989 | /* | |
1990 | * jiffies_force_qs before RCU_GP_WAIT_FQS state | |
1991 | * update; required for stall checks. | |
1992 | */ | |
1993 | smp_wmb(); | |
c3854a05 | 1994 | WRITE_ONCE(rcu_state.jiffies_kick_kthreads, |
9cf422a8 | 1995 | jiffies + (j ? 3 * j : 2)); |
c3854a05 | 1996 | } |
0f11ad32 | 1997 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 | 1998 | TPS("fqswait")); |
683954e5 | 1999 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS); |
eb880949 LS |
2000 | (void)swait_event_idle_timeout_exclusive(rcu_state.gp_wq, |
2001 | rcu_gp_fqs_check_wake(&gf), j); | |
55b2dcf5 | 2002 | rcu_gp_torture_wait(); |
683954e5 | 2003 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS); |
c3854a05 | 2004 | /* Locking provides needed memory barriers. */ |
a03ae49c NU |
2005 | /* |
2006 | * Exit the loop if the root rcu_node structure indicates that the grace period | |
2007 | * has ended, leave the loop. The rcu_preempt_blocked_readers_cgp(rnp) check | |
2008 | * is required only for single-node rcu_node trees because readers blocking | |
2009 | * the current grace period are queued only on leaf rcu_node structures. | |
2010 | * For multi-node trees, checking the root node's ->qsmask suffices, because a | |
2011 | * given root node's ->qsmask bit is cleared only when all CPUs and tasks from | |
2012 | * the corresponding leaf nodes have passed through their quiescent state. | |
2013 | */ | |
c3854a05 PM |
2014 | if (!READ_ONCE(rnp->qsmask) && |
2015 | !rcu_preempt_blocked_readers_cgp(rnp)) | |
2016 | break; | |
2017 | /* If time for quiescent-state forcing, do it. */ | |
29ffebc5 | 2018 | if (!time_after(rcu_state.jiffies_force_qs, jiffies) || |
9c392453 | 2019 | (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) { |
0f11ad32 | 2020 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
2021 | TPS("fqsstart")); |
2022 | rcu_gp_fqs(first_gp_fqs); | |
1fca4d12 PM |
2023 | gf = 0; |
2024 | if (first_gp_fqs) { | |
2025 | first_gp_fqs = false; | |
2026 | gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0; | |
2027 | } | |
0f11ad32 | 2028 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
2029 | TPS("fqsend")); |
2030 | cond_resched_tasks_rcu_qs(); | |
2031 | WRITE_ONCE(rcu_state.gp_activity, jiffies); | |
2032 | ret = 0; /* Force full wait till next FQS. */ | |
c06aed0e | 2033 | j = READ_ONCE(jiffies_till_next_fqs); |
c3854a05 PM |
2034 | } else { |
2035 | /* Deal with stray signal. */ | |
2036 | cond_resched_tasks_rcu_qs(); | |
2037 | WRITE_ONCE(rcu_state.gp_activity, jiffies); | |
2038 | WARN_ON(signal_pending(current)); | |
0f11ad32 | 2039 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
c3854a05 PM |
2040 | TPS("fqswaitsig")); |
2041 | ret = 1; /* Keep old FQS timing. */ | |
2042 | j = jiffies; | |
2043 | if (time_after(jiffies, rcu_state.jiffies_force_qs)) | |
2044 | j = 1; | |
2045 | else | |
2046 | j = rcu_state.jiffies_force_qs - j; | |
1fca4d12 | 2047 | gf = 0; |
c3854a05 PM |
2048 | } |
2049 | } | |
2050 | } | |
2051 | ||
7fdefc10 PM |
2052 | /* |
2053 | * Clean up after the old grace period. | |
2054 | */ | |
2f20de99 | 2055 | static noinline void rcu_gp_cleanup(void) |
7fdefc10 | 2056 | { |
b2b00ddf | 2057 | int cpu; |
48a7639c | 2058 | bool needgp = false; |
b2b00ddf | 2059 | unsigned long gp_duration; |
de30ad51 | 2060 | unsigned long new_gp_seq; |
5d6742b3 | 2061 | bool offloaded; |
7fdefc10 | 2062 | struct rcu_data *rdp; |
336a4f6c | 2063 | struct rcu_node *rnp = rcu_get_root(); |
abedf8e2 | 2064 | struct swait_queue_head *sq; |
b3dbec76 | 2065 | |
9cbc5b97 | 2066 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
2a67e741 | 2067 | raw_spin_lock_irq_rcu_node(rnp); |
c51d7b5e PM |
2068 | rcu_state.gp_end = jiffies; |
2069 | gp_duration = rcu_state.gp_end - rcu_state.gp_start; | |
9cbc5b97 PM |
2070 | if (gp_duration > rcu_state.gp_max) |
2071 | rcu_state.gp_max = gp_duration; | |
b3dbec76 | 2072 | |
7fdefc10 PM |
2073 | /* |
2074 | * We know the grace period is complete, but to everyone else | |
2075 | * it appears to still be ongoing. But it is also the case | |
2076 | * that to everyone else it looks like there is nothing that | |
2077 | * they can do to advance the grace period. It is therefore | |
2078 | * safe for us to drop the lock in order to mark the grace | |
2079 | * period as completed in all of the rcu_node structures. | |
7fdefc10 | 2080 | */ |
bf95b2bc | 2081 | rcu_poll_gp_seq_end(&rcu_state.gp_seq_polled_snap); |
67c583a7 | 2082 | raw_spin_unlock_irq_rcu_node(rnp); |
b3dbec76 | 2083 | |
5d4b8659 | 2084 | /* |
ff3bb6f4 PM |
2085 | * Propagate new ->gp_seq value to rcu_node structures so that |
2086 | * other CPUs don't have to wait until the start of the next grace | |
2087 | * period to process their callbacks. This also avoids some nasty | |
2088 | * RCU grace-period initialization races by forcing the end of | |
2089 | * the current grace period to be completely recorded in all of | |
2090 | * the rcu_node structures before the beginning of the next grace | |
2091 | * period is recorded in any of the rcu_node structures. | |
5d4b8659 | 2092 | */ |
9cbc5b97 | 2093 | new_gp_seq = rcu_state.gp_seq; |
de30ad51 | 2094 | rcu_seq_end(&new_gp_seq); |
aedf4ba9 | 2095 | rcu_for_each_node_breadth_first(rnp) { |
2a67e741 | 2096 | raw_spin_lock_irq_rcu_node(rnp); |
4bc8d555 | 2097 | if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp))) |
81ab59a3 | 2098 | dump_blkd_tasks(rnp, 10); |
5c60d25f | 2099 | WARN_ON_ONCE(rnp->qsmask); |
de30ad51 | 2100 | WRITE_ONCE(rnp->gp_seq, new_gp_seq); |
3fdefca9 PM |
2101 | if (!rnp->parent) |
2102 | smp_mb(); // Order against failing poll_state_synchronize_rcu_full(). | |
da1df50d | 2103 | rdp = this_cpu_ptr(&rcu_data); |
b11cc576 | 2104 | if (rnp == rdp->mynode) |
c7e48f7b | 2105 | needgp = __note_gp_changes(rnp, rdp) || needgp; |
78e4bc34 | 2106 | /* smp_mb() provided by prior unlock-lock pair. */ |
3481f2ea | 2107 | needgp = rcu_future_gp_cleanup(rnp) || needgp; |
b2b00ddf PM |
2108 | // Reset overload indication for CPUs no longer overloaded |
2109 | if (rcu_is_leaf_node(rnp)) | |
2110 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) { | |
2111 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
2112 | check_cb_ovld_locked(rdp, rnp); | |
2113 | } | |
065bb78c | 2114 | sq = rcu_nocb_gp_get(rnp); |
67c583a7 | 2115 | raw_spin_unlock_irq_rcu_node(rnp); |
065bb78c | 2116 | rcu_nocb_gp_cleanup(sq); |
cee43939 | 2117 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 2118 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
22212332 | 2119 | rcu_gp_slow(gp_cleanup_delay); |
7fdefc10 | 2120 | } |
336a4f6c | 2121 | rnp = rcu_get_root(); |
9cbc5b97 | 2122 | raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */ |
7fdefc10 | 2123 | |
0a89e5a4 | 2124 | /* Declare grace period done, trace first to use old GP number. */ |
9cbc5b97 | 2125 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end")); |
0a89e5a4 | 2126 | rcu_seq_end(&rcu_state.gp_seq); |
62ae1951 | 2127 | ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); |
683954e5 | 2128 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE); |
fb31340f | 2129 | /* Check for GP requests since above loop. */ |
da1df50d | 2130 | rdp = this_cpu_ptr(&rcu_data); |
5b55072f | 2131 | if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) { |
abd13fdd | 2132 | trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed, |
41e80595 | 2133 | TPS("CleanupMore")); |
fb31340f PM |
2134 | needgp = true; |
2135 | } | |
48a7639c | 2136 | /* Advance CBs to reduce false positives below. */ |
3820b513 | 2137 | offloaded = rcu_rdp_is_offloaded(rdp); |
5d6742b3 | 2138 | if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) { |
75182a4e PM |
2139 | |
2140 | // We get here if a grace period was needed (“needgp”) | |
2141 | // and the above call to rcu_accelerate_cbs() did not set | |
2142 | // the RCU_GP_FLAG_INIT bit in ->gp_state (which records | |
2143 | // the need for another grace period). The purpose | |
2144 | // of the “offloaded” check is to avoid invoking | |
2145 | // rcu_accelerate_cbs() on an offloaded CPU because we do not | |
2146 | // hold the ->nocb_lock needed to safely access an offloaded | |
2147 | // ->cblist. We do not want to acquire that lock because | |
2148 | // it can be heavily contended during callback floods. | |
2149 | ||
9cbc5b97 | 2150 | WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT); |
2906d215 | 2151 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); |
75182a4e | 2152 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("newreq")); |
18390aea | 2153 | } else { |
75182a4e PM |
2154 | |
2155 | // We get here either if there is no need for an | |
2156 | // additional grace period or if rcu_accelerate_cbs() has | |
2157 | // already set the RCU_GP_FLAG_INIT bit in ->gp_flags. | |
2158 | // So all we need to do is to clear all of the other | |
2159 | // ->gp_flags bits. | |
2160 | ||
2161 | WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags & RCU_GP_FLAG_INIT); | |
bb311ecc | 2162 | } |
67c583a7 | 2163 | raw_spin_unlock_irq_rcu_node(rnp); |
4e025f52 | 2164 | |
988f569a URS |
2165 | // Make synchronize_rcu() users aware of the end of old grace period. |
2166 | rcu_sr_normal_gp_cleanup(); | |
2167 | ||
4e025f52 PM |
2168 | // If strict, make all CPUs aware of the end of the old grace period. |
2169 | if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) | |
2170 | on_each_cpu(rcu_strict_gp_boundary, NULL, 0); | |
7fdefc10 PM |
2171 | } |
2172 | ||
2173 | /* | |
2174 | * Body of kthread that handles grace periods. | |
2175 | */ | |
0854a05c | 2176 | static int __noreturn rcu_gp_kthread(void *unused) |
7fdefc10 | 2177 | { |
5871968d | 2178 | rcu_bind_gp_kthread(); |
7fdefc10 PM |
2179 | for (;;) { |
2180 | ||
2181 | /* Handle grace-period start. */ | |
2182 | for (;;) { | |
0f11ad32 | 2183 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
63c4db78 | 2184 | TPS("reqwait")); |
683954e5 | 2185 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS); |
9cbc5b97 PM |
2186 | swait_event_idle_exclusive(rcu_state.gp_wq, |
2187 | READ_ONCE(rcu_state.gp_flags) & | |
2188 | RCU_GP_FLAG_INIT); | |
55b2dcf5 | 2189 | rcu_gp_torture_wait(); |
683954e5 | 2190 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS); |
78e4bc34 | 2191 | /* Locking provides needed memory barrier. */ |
0854a05c | 2192 | if (rcu_gp_init()) |
7fdefc10 | 2193 | break; |
cee43939 | 2194 | cond_resched_tasks_rcu_qs(); |
9cbc5b97 | 2195 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
73a860cd | 2196 | WARN_ON(signal_pending(current)); |
0f11ad32 | 2197 | trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, |
63c4db78 | 2198 | TPS("reqwaitsig")); |
7fdefc10 | 2199 | } |
cabc49c1 | 2200 | |
4cdfc175 | 2201 | /* Handle quiescent-state forcing. */ |
c3854a05 | 2202 | rcu_gp_fqs_loop(); |
4cdfc175 PM |
2203 | |
2204 | /* Handle grace-period end. */ | |
683954e5 | 2205 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP); |
0854a05c | 2206 | rcu_gp_cleanup(); |
683954e5 | 2207 | WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED); |
b3dbec76 | 2208 | } |
b3dbec76 PM |
2209 | } |
2210 | ||
f41d911f | 2211 | /* |
49918a54 PM |
2212 | * Report a full set of quiescent states to the rcu_state data structure. |
2213 | * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if | |
2214 | * another grace period is required. Whether we wake the grace-period | |
2215 | * kthread or it awakens itself for the next round of quiescent-state | |
2216 | * forcing, that kthread will clean up after the just-completed grace | |
2217 | * period. Note that the caller must hold rnp->lock, which is released | |
2218 | * before return. | |
f41d911f | 2219 | */ |
aff4e9ed | 2220 | static void rcu_report_qs_rsp(unsigned long flags) |
336a4f6c | 2221 | __releases(rcu_get_root()->lock) |
f41d911f | 2222 | { |
336a4f6c | 2223 | raw_lockdep_assert_held_rcu_node(rcu_get_root()); |
de8e8730 | 2224 | WARN_ON_ONCE(!rcu_gp_in_progress()); |
62bb24c4 | 2225 | WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_FQS); |
336a4f6c | 2226 | raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags); |
532c00c9 | 2227 | rcu_gp_kthread_wake(); |
f41d911f PM |
2228 | } |
2229 | ||
64db4cff | 2230 | /* |
d3f6bad3 PM |
2231 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
2232 | * Allows quiescent states for a group of CPUs to be reported at one go | |
2233 | * to the specified rcu_node structure, though all the CPUs in the group | |
654e9533 PM |
2234 | * must be represented by the same rcu_node structure (which need not be a |
2235 | * leaf rcu_node structure, though it often will be). The gps parameter | |
2236 | * is the grace-period snapshot, which means that the quiescent states | |
c9a24e2d | 2237 | * are valid only if rnp->gp_seq is equal to gps. That structure's lock |
654e9533 | 2238 | * must be held upon entry, and it is released before return. |
ec2c2976 PM |
2239 | * |
2240 | * As a special case, if mask is zero, the bit-already-cleared check is | |
2241 | * disabled. This allows propagating quiescent state due to resumed tasks | |
2242 | * during grace-period initialization. | |
64db4cff | 2243 | */ |
b50912d0 PM |
2244 | static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp, |
2245 | unsigned long gps, unsigned long flags) | |
64db4cff PM |
2246 | __releases(rnp->lock) |
2247 | { | |
654e9533 | 2248 | unsigned long oldmask = 0; |
28ecd580 PM |
2249 | struct rcu_node *rnp_c; |
2250 | ||
a32e01ee | 2251 | raw_lockdep_assert_held_rcu_node(rnp); |
c0b334c5 | 2252 | |
64db4cff PM |
2253 | /* Walk up the rcu_node hierarchy. */ |
2254 | for (;;) { | |
ec2c2976 | 2255 | if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) { |
64db4cff | 2256 | |
654e9533 PM |
2257 | /* |
2258 | * Our bit has already been cleared, or the | |
2259 | * relevant grace period is already over, so done. | |
2260 | */ | |
67c583a7 | 2261 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2262 | return; |
2263 | } | |
654e9533 | 2264 | WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ |
5b4c11d5 | 2265 | WARN_ON_ONCE(!rcu_is_leaf_node(rnp) && |
2dee9404 | 2266 | rcu_preempt_blocked_readers_cgp(rnp)); |
7672d647 | 2267 | WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask); |
67a0edbf | 2268 | trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq, |
d4c08f2a PM |
2269 | mask, rnp->qsmask, rnp->level, |
2270 | rnp->grplo, rnp->grphi, | |
2271 | !!rnp->gp_tasks); | |
27f4d280 | 2272 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
64db4cff PM |
2273 | |
2274 | /* Other bits still set at this level, so done. */ | |
67c583a7 | 2275 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2276 | return; |
2277 | } | |
d43a5d32 | 2278 | rnp->completedqs = rnp->gp_seq; |
64db4cff PM |
2279 | mask = rnp->grpmask; |
2280 | if (rnp->parent == NULL) { | |
2281 | ||
2282 | /* No more levels. Exit loop holding root lock. */ | |
2283 | ||
2284 | break; | |
2285 | } | |
67c583a7 | 2286 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
28ecd580 | 2287 | rnp_c = rnp; |
64db4cff | 2288 | rnp = rnp->parent; |
2a67e741 | 2289 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
0937d045 | 2290 | oldmask = READ_ONCE(rnp_c->qsmask); |
64db4cff PM |
2291 | } |
2292 | ||
2293 | /* | |
2294 | * Get here if we are the last CPU to pass through a quiescent | |
d3f6bad3 | 2295 | * state for this grace period. Invoke rcu_report_qs_rsp() |
f41d911f | 2296 | * to clean up and start the next grace period if one is needed. |
64db4cff | 2297 | */ |
aff4e9ed | 2298 | rcu_report_qs_rsp(flags); /* releases rnp->lock. */ |
64db4cff PM |
2299 | } |
2300 | ||
cc99a310 PM |
2301 | /* |
2302 | * Record a quiescent state for all tasks that were previously queued | |
2303 | * on the specified rcu_node structure and that were blocking the current | |
49918a54 | 2304 | * RCU grace period. The caller must hold the corresponding rnp->lock with |
cc99a310 PM |
2305 | * irqs disabled, and this lock is released upon return, but irqs remain |
2306 | * disabled. | |
2307 | */ | |
17a8212b | 2308 | static void __maybe_unused |
139ad4da | 2309 | rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
cc99a310 PM |
2310 | __releases(rnp->lock) |
2311 | { | |
654e9533 | 2312 | unsigned long gps; |
cc99a310 PM |
2313 | unsigned long mask; |
2314 | struct rcu_node *rnp_p; | |
2315 | ||
a32e01ee | 2316 | raw_lockdep_assert_held_rcu_node(rnp); |
c130d2dc | 2317 | if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) || |
c74859d1 PM |
2318 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) || |
2319 | rnp->qsmask != 0) { | |
67c583a7 | 2320 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
cc99a310 PM |
2321 | return; /* Still need more quiescent states! */ |
2322 | } | |
2323 | ||
77cfc7bf | 2324 | rnp->completedqs = rnp->gp_seq; |
cc99a310 PM |
2325 | rnp_p = rnp->parent; |
2326 | if (rnp_p == NULL) { | |
2327 | /* | |
a77da14c PM |
2328 | * Only one rcu_node structure in the tree, so don't |
2329 | * try to report up to its nonexistent parent! | |
cc99a310 | 2330 | */ |
aff4e9ed | 2331 | rcu_report_qs_rsp(flags); |
cc99a310 PM |
2332 | return; |
2333 | } | |
2334 | ||
c9a24e2d PM |
2335 | /* Report up the rest of the hierarchy, tracking current ->gp_seq. */ |
2336 | gps = rnp->gp_seq; | |
cc99a310 | 2337 | mask = rnp->grpmask; |
67c583a7 | 2338 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
2a67e741 | 2339 | raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */ |
b50912d0 | 2340 | rcu_report_qs_rnp(mask, rnp_p, gps, flags); |
cc99a310 PM |
2341 | } |
2342 | ||
64db4cff | 2343 | /* |
d3f6bad3 | 2344 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
4b455dc3 | 2345 | * structure. This must be called from the specified CPU. |
64db4cff PM |
2346 | */ |
2347 | static void | |
cfeac397 | 2348 | rcu_report_qs_rdp(struct rcu_data *rdp) |
64db4cff PM |
2349 | { |
2350 | unsigned long flags; | |
2351 | unsigned long mask; | |
b3bb02fe | 2352 | bool needacc = false; |
64db4cff PM |
2353 | struct rcu_node *rnp; |
2354 | ||
cfeac397 | 2355 | WARN_ON_ONCE(rdp->cpu != smp_processor_id()); |
64db4cff | 2356 | rnp = rdp->mynode; |
2a67e741 | 2357 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
c9a24e2d PM |
2358 | if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq || |
2359 | rdp->gpwrap) { | |
64db4cff PM |
2360 | |
2361 | /* | |
e4cc1f22 PM |
2362 | * The grace period in which this quiescent state was |
2363 | * recorded has ended, so don't report it upwards. | |
2364 | * We will instead need a new quiescent state that lies | |
2365 | * within the current grace period. | |
64db4cff | 2366 | */ |
5b74c458 | 2367 | rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */ |
67c583a7 | 2368 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff PM |
2369 | return; |
2370 | } | |
2371 | mask = rdp->grpmask; | |
cfeac397 | 2372 | rdp->core_needs_qs = false; |
64db4cff | 2373 | if ((rnp->qsmask & mask) == 0) { |
67c583a7 | 2374 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff | 2375 | } else { |
64db4cff PM |
2376 | /* |
2377 | * This GP can't end until cpu checks in, so all of our | |
2378 | * callbacks can be processed during the next GP. | |
24ee940d | 2379 | * |
b3bb02fe | 2380 | * NOCB kthreads have their own way to deal with that... |
64db4cff | 2381 | */ |
b3bb02fe | 2382 | if (!rcu_rdp_is_offloaded(rdp)) { |
46103fe0 Z |
2383 | /* |
2384 | * The current GP has not yet ended, so it | |
2385 | * should not be possible for rcu_accelerate_cbs() | |
2386 | * to return true. So complain, but don't awaken. | |
2387 | */ | |
2388 | WARN_ON_ONCE(rcu_accelerate_cbs(rnp, rdp)); | |
b3bb02fe FW |
2389 | } else if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) { |
2390 | /* | |
2391 | * ...but NOCB kthreads may miss or delay callbacks acceleration | |
2392 | * if in the middle of a (de-)offloading process. | |
2393 | */ | |
2394 | needacc = true; | |
2395 | } | |
64db4cff | 2396 | |
516e5ae0 | 2397 | rcu_disable_urgency_upon_qs(rdp); |
b50912d0 | 2398 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
654e9533 | 2399 | /* ^^^ Released rnp->lock */ |
b3bb02fe FW |
2400 | |
2401 | if (needacc) { | |
2402 | rcu_nocb_lock_irqsave(rdp, flags); | |
2403 | rcu_accelerate_cbs_unlocked(rnp, rdp); | |
2404 | rcu_nocb_unlock_irqrestore(rdp, flags); | |
2405 | } | |
64db4cff PM |
2406 | } |
2407 | } | |
2408 | ||
2409 | /* | |
2410 | * Check to see if there is a new grace period of which this CPU | |
2411 | * is not yet aware, and if so, set up local rcu_data state for it. | |
2412 | * Otherwise, see if this CPU has just passed through its first | |
2413 | * quiescent state for this grace period, and record that fact if so. | |
2414 | */ | |
2415 | static void | |
8087d3e3 | 2416 | rcu_check_quiescent_state(struct rcu_data *rdp) |
64db4cff | 2417 | { |
05eb552b | 2418 | /* Check for grace-period ends and beginnings. */ |
15cabdff | 2419 | note_gp_changes(rdp); |
64db4cff PM |
2420 | |
2421 | /* | |
2422 | * Does this CPU still need to do its part for current grace period? | |
2423 | * If no, return and let the other CPUs do their part as well. | |
2424 | */ | |
97c668b8 | 2425 | if (!rdp->core_needs_qs) |
64db4cff PM |
2426 | return; |
2427 | ||
2428 | /* | |
2429 | * Was there a quiescent state since the beginning of the grace | |
2430 | * period? If no, then exit and wait for the next call. | |
2431 | */ | |
3a19b46a | 2432 | if (rdp->cpu_no_qs.b.norm) |
64db4cff PM |
2433 | return; |
2434 | ||
d3f6bad3 PM |
2435 | /* |
2436 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the | |
2437 | * judge of that). | |
2438 | */ | |
cfeac397 | 2439 | rcu_report_qs_rdp(rdp); |
64db4cff PM |
2440 | } |
2441 | ||
fea1c1f0 | 2442 | /* Return true if callback-invocation time limit exceeded. */ |
f51164a8 PM |
2443 | static bool rcu_do_batch_check_time(long count, long tlimit, |
2444 | bool jlimit_check, unsigned long jlimit) | |
fea1c1f0 PM |
2445 | { |
2446 | // Invoke local_clock() only once per 32 consecutive callbacks. | |
f51164a8 PM |
2447 | return unlikely(tlimit) && |
2448 | (!likely(count & 31) || | |
2449 | (IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) && | |
2450 | jlimit_check && time_after(jiffies, jlimit))) && | |
2451 | local_clock() >= tlimit; | |
fea1c1f0 PM |
2452 | } |
2453 | ||
64db4cff PM |
2454 | /* |
2455 | * Invoke any RCU callbacks that have made it to the end of their grace | |
a616aec9 | 2456 | * period. Throttle as specified by rdp->blimit. |
64db4cff | 2457 | */ |
5bb5d09c | 2458 | static void rcu_do_batch(struct rcu_data *rdp) |
64db4cff | 2459 | { |
f51164a8 PM |
2460 | long bl; |
2461 | long count = 0; | |
b5374b2d | 2462 | int div; |
b4e6039e | 2463 | bool __maybe_unused empty; |
64db4cff | 2464 | unsigned long flags; |
f51164a8 PM |
2465 | unsigned long jlimit; |
2466 | bool jlimit_check = false; | |
2467 | long pending; | |
15fecf89 | 2468 | struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl); |
f51164a8 PM |
2469 | struct rcu_head *rhp; |
2470 | long tlimit = 0; | |
64db4cff | 2471 | |
dc35c893 | 2472 | /* If no callbacks are ready, just return. */ |
15fecf89 | 2473 | if (!rcu_segcblist_ready_cbs(&rdp->cblist)) { |
3c779dfe | 2474 | trace_rcu_batch_start(rcu_state.name, |
15fecf89 | 2475 | rcu_segcblist_n_cbs(&rdp->cblist), 0); |
3c779dfe | 2476 | trace_rcu_batch_end(rcu_state.name, 0, |
15fecf89 | 2477 | !rcu_segcblist_empty(&rdp->cblist), |
4968c300 | 2478 | need_resched(), is_idle_task(current), |
51038506 | 2479 | rcu_is_callbacks_kthread(rdp)); |
64db4cff | 2480 | return; |
29c00b4a | 2481 | } |
64db4cff PM |
2482 | |
2483 | /* | |
7b65dfa3 | 2484 | * Extract the list of ready callbacks, disabling IRQs to prevent |
15fecf89 PM |
2485 | * races with call_rcu() from interrupt handlers. Leave the |
2486 | * callback counts, as rcu_barrier() needs to be conservative. | |
1e8e6951 FW |
2487 | * |
2488 | * Callbacks execution is fully ordered against preceding grace period | |
2489 | * completion (materialized by rnp->gp_seq update) thanks to the | |
2490 | * smp_mb__after_unlock_lock() upon node locking required for callbacks | |
2491 | * advancing. In NOCB mode this ordering is then further relayed through | |
2492 | * the nocb locking that protects both callbacks advancing and extraction. | |
64db4cff | 2493 | */ |
7b65dfa3 | 2494 | rcu_nocb_lock_irqsave(rdp, flags); |
8146c4e2 | 2495 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
253cbbff | 2496 | pending = rcu_segcblist_get_seglen(&rdp->cblist, RCU_DONE_TAIL); |
b5374b2d PM |
2497 | div = READ_ONCE(rcu_divisor); |
2498 | div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div; | |
2499 | bl = max(rdp->blimit, pending >> div); | |
fea1c1f0 | 2500 | if ((in_serving_softirq() || rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING) && |
f51164a8 PM |
2501 | (IS_ENABLED(CONFIG_RCU_DOUBLE_CHECK_CB_TIME) || unlikely(bl > 100))) { |
2502 | const long npj = NSEC_PER_SEC / HZ; | |
a2b354b9 PM |
2503 | long rrn = READ_ONCE(rcu_resched_ns); |
2504 | ||
2505 | rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn; | |
2506 | tlimit = local_clock() + rrn; | |
f51164a8 PM |
2507 | jlimit = jiffies + (rrn + npj + 1) / npj; |
2508 | jlimit_check = true; | |
a2b354b9 | 2509 | } |
3c779dfe | 2510 | trace_rcu_batch_start(rcu_state.name, |
15fecf89 PM |
2511 | rcu_segcblist_n_cbs(&rdp->cblist), bl); |
2512 | rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl); | |
344e219d | 2513 | if (rcu_rdp_is_offloaded(rdp)) |
7f36ef82 | 2514 | rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist); |
3afe7fa5 JFG |
2515 | |
2516 | trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued")); | |
5d6742b3 | 2517 | rcu_nocb_unlock_irqrestore(rdp, flags); |
64db4cff PM |
2518 | |
2519 | /* Invoke callbacks. */ | |
6a949b7a | 2520 | tick_dep_set_task(current, TICK_DEP_BIT_RCU); |
15fecf89 | 2521 | rhp = rcu_cblist_dequeue(&rcl); |
3afe7fa5 | 2522 | |
15fecf89 | 2523 | for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) { |
77a40f97 JFG |
2524 | rcu_callback_t f; |
2525 | ||
6bc33582 | 2526 | count++; |
15fecf89 | 2527 | debug_rcu_head_unqueue(rhp); |
77a40f97 JFG |
2528 | |
2529 | rcu_lock_acquire(&rcu_callback_map); | |
2530 | trace_rcu_invoke_callback(rcu_state.name, rhp); | |
2531 | ||
2532 | f = rhp->func; | |
2cbc482d | 2533 | debug_rcu_head_callback(rhp); |
77a40f97 JFG |
2534 | WRITE_ONCE(rhp->func, (rcu_callback_t)0L); |
2535 | f(rhp); | |
2536 | ||
2537 | rcu_lock_release(&rcu_callback_map); | |
2538 | ||
15fecf89 PM |
2539 | /* |
2540 | * Stop only if limit reached and CPU has something to do. | |
15fecf89 | 2541 | */ |
3e61e95e FW |
2542 | if (in_serving_softirq()) { |
2543 | if (count >= bl && (need_resched() || !is_idle_task(current))) | |
2544 | break; | |
a554ba28 FW |
2545 | /* |
2546 | * Make sure we don't spend too much time here and deprive other | |
2547 | * softirq vectors of CPU cycles. | |
2548 | */ | |
f51164a8 | 2549 | if (rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit)) |
a554ba28 | 2550 | break; |
3e61e95e | 2551 | } else { |
fea1c1f0 PM |
2552 | // In rcuc/rcuoc context, so no worries about |
2553 | // depriving other softirq vectors of CPU cycles. | |
5d6742b3 PM |
2554 | local_bh_enable(); |
2555 | lockdep_assert_irqs_enabled(); | |
2556 | cond_resched_tasks_rcu_qs(); | |
2557 | lockdep_assert_irqs_enabled(); | |
2558 | local_bh_disable(); | |
fea1c1f0 PM |
2559 | // But rcuc kthreads can delay quiescent-state |
2560 | // reporting, so check time limits for them. | |
2561 | if (rdp->rcu_cpu_kthread_status == RCU_KTHREAD_RUNNING && | |
f51164a8 | 2562 | rcu_do_batch_check_time(count, tlimit, jlimit_check, jlimit)) { |
fea1c1f0 PM |
2563 | rdp->rcu_cpu_has_work = 1; |
2564 | break; | |
2565 | } | |
5d6742b3 | 2566 | } |
64db4cff PM |
2567 | } |
2568 | ||
7b65dfa3 | 2569 | rcu_nocb_lock_irqsave(rdp, flags); |
e816d56f | 2570 | rdp->n_cbs_invoked += count; |
3c779dfe | 2571 | trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(), |
51038506 | 2572 | is_idle_task(current), rcu_is_callbacks_kthread(rdp)); |
64db4cff | 2573 | |
15fecf89 PM |
2574 | /* Update counts and requeue any remaining callbacks. */ |
2575 | rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl); | |
6bc33582 | 2576 | rcu_segcblist_add_len(&rdp->cblist, -count); |
64db4cff PM |
2577 | |
2578 | /* Reinstate batch limit if we have worked down the excess. */ | |
15fecf89 | 2579 | count = rcu_segcblist_n_cbs(&rdp->cblist); |
d5a9a8c3 | 2580 | if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark) |
64db4cff PM |
2581 | rdp->blimit = blimit; |
2582 | ||
37c72e56 | 2583 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
15fecf89 | 2584 | if (count == 0 && rdp->qlen_last_fqs_check != 0) { |
37c72e56 | 2585 | rdp->qlen_last_fqs_check = 0; |
2431774f | 2586 | rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs); |
15fecf89 PM |
2587 | } else if (count < rdp->qlen_last_fqs_check - qhimark) |
2588 | rdp->qlen_last_fqs_check = count; | |
efd88b02 PM |
2589 | |
2590 | /* | |
2591 | * The following usually indicates a double call_rcu(). To track | |
2592 | * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y. | |
2593 | */ | |
b4e6039e JFG |
2594 | empty = rcu_segcblist_empty(&rdp->cblist); |
2595 | WARN_ON_ONCE(count == 0 && !empty); | |
d1b222c6 | 2596 | WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) && |
b4e6039e JFG |
2597 | count != 0 && empty); |
2598 | WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0); | |
2599 | WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0); | |
37c72e56 | 2600 | |
5d6742b3 | 2601 | rcu_nocb_unlock_irqrestore(rdp, flags); |
64db4cff | 2602 | |
6a949b7a | 2603 | tick_dep_clear_task(current, TICK_DEP_BIT_RCU); |
64db4cff PM |
2604 | } |
2605 | ||
2606 | /* | |
c98cac60 PM |
2607 | * This function is invoked from each scheduling-clock interrupt, |
2608 | * and checks to see if this CPU is in a non-context-switch quiescent | |
2609 | * state, for example, user mode or idle loop. It also schedules RCU | |
2610 | * core processing. If the current grace period has gone on too long, | |
2611 | * it will ask the scheduler to manufacture a context switch for the sole | |
277ffe1b | 2612 | * purpose of providing the needed quiescent state. |
64db4cff | 2613 | */ |
c98cac60 | 2614 | void rcu_sched_clock_irq(int user) |
64db4cff | 2615 | { |
c708b08c PM |
2616 | unsigned long j; |
2617 | ||
2618 | if (IS_ENABLED(CONFIG_PROVE_RCU)) { | |
2619 | j = jiffies; | |
2620 | WARN_ON_ONCE(time_before(j, __this_cpu_read(rcu_data.last_sched_clock))); | |
2621 | __this_cpu_write(rcu_data.last_sched_clock, j); | |
2622 | } | |
f7f7bac9 | 2623 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
a649d25d | 2624 | lockdep_assert_irqs_disabled(); |
4e95020c | 2625 | raw_cpu_inc(rcu_data.ticks_this_gp); |
92aa39e9 | 2626 | /* The load-acquire pairs with the store-release setting to true. */ |
2dba13f0 | 2627 | if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) { |
92aa39e9 | 2628 | /* Idle and userspace execution already are quiescent states. */ |
a0ef9ec2 | 2629 | if (!rcu_is_cpu_rrupt_from_idle() && !user) { |
92aa39e9 PM |
2630 | set_tsk_need_resched(current); |
2631 | set_preempt_need_resched(); | |
2632 | } | |
2dba13f0 | 2633 | __this_cpu_write(rcu_data.rcu_urgent_qs, false); |
64db4cff | 2634 | } |
c98cac60 | 2635 | rcu_flavor_sched_clock_irq(user); |
dd7dafd1 | 2636 | if (rcu_pending(user)) |
a46e0899 | 2637 | invoke_rcu_core(); |
528262f5 Z |
2638 | if (user || rcu_is_cpu_rrupt_from_idle()) |
2639 | rcu_note_voluntary_context_switch(current); | |
a649d25d | 2640 | lockdep_assert_irqs_disabled(); |
07f27570 | 2641 | |
f7f7bac9 | 2642 | trace_rcu_utilization(TPS("End scheduler-tick")); |
64db4cff PM |
2643 | } |
2644 | ||
64db4cff | 2645 | /* |
5d8a752e ZZ |
2646 | * Scan the leaf rcu_node structures. For each structure on which all |
2647 | * CPUs have reported a quiescent state and on which there are tasks | |
2648 | * blocking the current grace period, initiate RCU priority boosting. | |
2649 | * Otherwise, invoke the specified function to check dyntick state for | |
2650 | * each CPU that has not yet reported a quiescent state. | |
64db4cff | 2651 | */ |
8ff0b907 | 2652 | static void force_qs_rnp(int (*f)(struct rcu_data *rdp)) |
64db4cff | 2653 | { |
64db4cff PM |
2654 | int cpu; |
2655 | unsigned long flags; | |
a0b6c9a7 | 2656 | struct rcu_node *rnp; |
64db4cff | 2657 | |
b2b00ddf PM |
2658 | rcu_state.cbovld = rcu_state.cbovldnext; |
2659 | rcu_state.cbovldnext = false; | |
aedf4ba9 | 2660 | rcu_for_each_leaf_node(rnp) { |
85d68222 PZ |
2661 | unsigned long mask = 0; |
2662 | unsigned long rsmask = 0; | |
2663 | ||
cee43939 | 2664 | cond_resched_tasks_rcu_qs(); |
2a67e741 | 2665 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
b2b00ddf | 2666 | rcu_state.cbovldnext |= !!rnp->cbovldmask; |
a0b6c9a7 | 2667 | if (rnp->qsmask == 0) { |
9b1ce0ac | 2668 | if (rcu_preempt_blocked_readers_cgp(rnp)) { |
a77da14c PM |
2669 | /* |
2670 | * No point in scanning bits because they | |
2671 | * are all zero. But we might need to | |
2672 | * priority-boost blocked readers. | |
2673 | */ | |
2674 | rcu_initiate_boost(rnp, flags); | |
2675 | /* rcu_initiate_boost() releases rnp->lock */ | |
2676 | continue; | |
2677 | } | |
92816435 PM |
2678 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
2679 | continue; | |
64db4cff | 2680 | } |
7441e766 | 2681 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) { |
85d68222 PZ |
2682 | struct rcu_data *rdp; |
2683 | int ret; | |
2684 | ||
7441e766 | 2685 | rdp = per_cpu_ptr(&rcu_data, cpu); |
85d68222 PZ |
2686 | ret = f(rdp); |
2687 | if (ret > 0) { | |
7441e766 PM |
2688 | mask |= rdp->grpmask; |
2689 | rcu_disable_urgency_upon_qs(rdp); | |
0edd1b17 | 2690 | } |
85d68222 PZ |
2691 | if (ret < 0) |
2692 | rsmask |= rdp->grpmask; | |
64db4cff | 2693 | } |
45f014c5 | 2694 | if (mask != 0) { |
c9a24e2d | 2695 | /* Idle/offline CPUs, report (releases rnp->lock). */ |
b50912d0 | 2696 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
0aa04b05 PM |
2697 | } else { |
2698 | /* Nothing to do here, so just drop the lock. */ | |
67c583a7 | 2699 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
64db4cff | 2700 | } |
85d68222 PZ |
2701 | |
2702 | for_each_leaf_node_cpu_mask(rnp, cpu, rsmask) | |
2703 | resched_cpu(cpu); | |
64db4cff | 2704 | } |
64db4cff PM |
2705 | } |
2706 | ||
2707 | /* | |
2708 | * Force quiescent states on reluctant CPUs, and also detect which | |
2709 | * CPUs are in dyntick-idle mode. | |
2710 | */ | |
cd920e5a | 2711 | void rcu_force_quiescent_state(void) |
64db4cff PM |
2712 | { |
2713 | unsigned long flags; | |
394f2769 PM |
2714 | bool ret; |
2715 | struct rcu_node *rnp; | |
2716 | struct rcu_node *rnp_old = NULL; | |
2717 | ||
dee39c0c Z |
2718 | if (!rcu_gp_in_progress()) |
2719 | return; | |
394f2769 | 2720 | /* Funnel through hierarchy to reduce memory contention. */ |
ceb1c8c9 | 2721 | rnp = raw_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 | } |
62bb24c4 | 2740 | WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_FQS); |
67c583a7 | 2741 | raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags); |
532c00c9 | 2742 | rcu_gp_kthread_wake(); |
64db4cff | 2743 | } |
cd920e5a | 2744 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); |
64db4cff | 2745 | |
a657f261 PM |
2746 | // Workqueue handler for an RCU reader for kernels enforcing struct RCU |
2747 | // grace periods. | |
2748 | static void strict_work_handler(struct work_struct *work) | |
2749 | { | |
2750 | rcu_read_lock(); | |
2751 | rcu_read_unlock(); | |
2752 | } | |
2753 | ||
fb60e533 | 2754 | /* Perform RCU core processing work for the current CPU. */ |
48d07c04 | 2755 | static __latent_entropy void rcu_core(void) |
64db4cff PM |
2756 | { |
2757 | unsigned long flags; | |
da1df50d | 2758 | struct rcu_data *rdp = raw_cpu_ptr(&rcu_data); |
26d950a9 | 2759 | struct rcu_node *rnp = rdp->mynode; |
fbb94cbd FW |
2760 | /* |
2761 | * On RT rcu_core() can be preempted when IRQs aren't disabled. | |
2762 | * Therefore this function can race with concurrent NOCB (de-)offloading | |
2763 | * on this CPU and the below condition must be considered volatile. | |
2764 | * However if we race with: | |
2765 | * | |
2766 | * _ Offloading: In the worst case we accelerate or process callbacks | |
2767 | * concurrently with NOCB kthreads. We are guaranteed to | |
2768 | * call rcu_nocb_lock() if that happens. | |
2769 | * | |
2770 | * _ Deoffloading: In the worst case we miss callbacks acceleration or | |
2771 | * processing. This is fine because the early stage | |
2772 | * of deoffloading invokes rcu_core() after setting | |
2773 | * SEGCBLIST_RCU_CORE. So we guarantee that we'll process | |
2774 | * what could have been dismissed without the need to wait | |
2775 | * for the next rcu_pending() check in the next jiffy. | |
2776 | */ | |
32aa2f41 | 2777 | const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist); |
64db4cff | 2778 | |
b049fdf8 PM |
2779 | if (cpu_is_offline(smp_processor_id())) |
2780 | return; | |
2781 | trace_rcu_utilization(TPS("Start RCU core")); | |
50dc7def | 2782 | WARN_ON_ONCE(!rdp->beenonline); |
2e597558 | 2783 | |
3e310098 | 2784 | /* Report any deferred quiescent states if preemption enabled. */ |
790da248 | 2785 | if (IS_ENABLED(CONFIG_PREEMPT_COUNT) && (!(preempt_count() & PREEMPT_MASK))) { |
3e310098 | 2786 | rcu_preempt_deferred_qs(current); |
fced9c8c PM |
2787 | } else if (rcu_preempt_need_deferred_qs(current)) { |
2788 | set_tsk_need_resched(current); | |
2789 | set_preempt_need_resched(); | |
2790 | } | |
3e310098 | 2791 | |
64db4cff | 2792 | /* Update RCU state based on any recent quiescent states. */ |
8087d3e3 | 2793 | rcu_check_quiescent_state(rdp); |
64db4cff | 2794 | |
bd7af846 | 2795 | /* No grace period and unregistered callbacks? */ |
de8e8730 | 2796 | if (!rcu_gp_in_progress() && |
634954c2 FW |
2797 | rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) { |
2798 | rcu_nocb_lock_irqsave(rdp, flags); | |
e44e73ca | 2799 | if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL)) |
c6e09b97 | 2800 | rcu_accelerate_cbs_unlocked(rnp, rdp); |
634954c2 | 2801 | rcu_nocb_unlock_irqrestore(rdp, flags); |
64db4cff PM |
2802 | } |
2803 | ||
791416c4 | 2804 | rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check()); |
26d950a9 | 2805 | |
64db4cff | 2806 | /* If there are callbacks ready, invoke them. */ |
32aa2f41 | 2807 | if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) && |
0598a4d4 | 2808 | likely(READ_ONCE(rcu_scheduler_fully_active))) { |
43e903ad | 2809 | rcu_do_batch(rdp); |
0598a4d4 FW |
2810 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
2811 | if (rcu_segcblist_ready_cbs(&rdp->cblist)) | |
2812 | invoke_rcu_core(); | |
2813 | } | |
96d3fd0d PM |
2814 | |
2815 | /* Do any needed deferred wakeups of rcuo kthreads. */ | |
2816 | do_nocb_deferred_wakeup(rdp); | |
f7f7bac9 | 2817 | trace_rcu_utilization(TPS("End RCU core")); |
a657f261 PM |
2818 | |
2819 | // If strict GPs, schedule an RCU reader in a clean environment. | |
2820 | if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) | |
2821 | queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work); | |
64db4cff PM |
2822 | } |
2823 | ||
48d07c04 SAS |
2824 | static void rcu_core_si(struct softirq_action *h) |
2825 | { | |
2826 | rcu_core(); | |
2827 | } | |
2828 | ||
2829 | static void rcu_wake_cond(struct task_struct *t, int status) | |
2830 | { | |
2831 | /* | |
2832 | * If the thread is yielding, only wake it when this | |
2833 | * is invoked from idle | |
2834 | */ | |
2835 | if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current))) | |
2836 | wake_up_process(t); | |
2837 | } | |
2838 | ||
2839 | static void invoke_rcu_core_kthread(void) | |
2840 | { | |
2841 | struct task_struct *t; | |
2842 | unsigned long flags; | |
2843 | ||
2844 | local_irq_save(flags); | |
2845 | __this_cpu_write(rcu_data.rcu_cpu_has_work, 1); | |
2846 | t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task); | |
2847 | if (t != NULL && t != current) | |
2848 | rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status)); | |
2849 | local_irq_restore(flags); | |
2850 | } | |
2851 | ||
48d07c04 SAS |
2852 | /* |
2853 | * Wake up this CPU's rcuc kthread to do RCU core processing. | |
2854 | */ | |
a46e0899 | 2855 | static void invoke_rcu_core(void) |
09223371 | 2856 | { |
48d07c04 SAS |
2857 | if (!cpu_online(smp_processor_id())) |
2858 | return; | |
2859 | if (use_softirq) | |
b0f74036 | 2860 | raise_softirq(RCU_SOFTIRQ); |
48d07c04 SAS |
2861 | else |
2862 | invoke_rcu_core_kthread(); | |
2863 | } | |
2864 | ||
2865 | static void rcu_cpu_kthread_park(unsigned int cpu) | |
2866 | { | |
2867 | per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; | |
2868 | } | |
2869 | ||
2870 | static int rcu_cpu_kthread_should_run(unsigned int cpu) | |
2871 | { | |
2872 | return __this_cpu_read(rcu_data.rcu_cpu_has_work); | |
2873 | } | |
2874 | ||
2875 | /* | |
2876 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces | |
2877 | * the RCU softirq used in configurations of RCU that do not support RCU | |
2878 | * priority boosting. | |
2879 | */ | |
2880 | static void rcu_cpu_kthread(unsigned int cpu) | |
2881 | { | |
2882 | unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status); | |
2883 | char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work); | |
c9515875 | 2884 | unsigned long *j = this_cpu_ptr(&rcu_data.rcuc_activity); |
48d07c04 SAS |
2885 | int spincnt; |
2886 | ||
2488a5e6 | 2887 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_run")); |
48d07c04 | 2888 | for (spincnt = 0; spincnt < 10; spincnt++) { |
c9515875 | 2889 | WRITE_ONCE(*j, jiffies); |
48d07c04 SAS |
2890 | local_bh_disable(); |
2891 | *statusp = RCU_KTHREAD_RUNNING; | |
2892 | local_irq_disable(); | |
2893 | work = *workp; | |
a24c1aab | 2894 | WRITE_ONCE(*workp, 0); |
48d07c04 SAS |
2895 | local_irq_enable(); |
2896 | if (work) | |
2897 | rcu_core(); | |
2898 | local_bh_enable(); | |
a24c1aab | 2899 | if (!READ_ONCE(*workp)) { |
48d07c04 SAS |
2900 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); |
2901 | *statusp = RCU_KTHREAD_WAITING; | |
2902 | return; | |
2903 | } | |
2904 | } | |
2905 | *statusp = RCU_KTHREAD_YIELDING; | |
2906 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); | |
77865dea | 2907 | schedule_timeout_idle(2); |
48d07c04 SAS |
2908 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); |
2909 | *statusp = RCU_KTHREAD_WAITING; | |
c9515875 | 2910 | WRITE_ONCE(*j, jiffies); |
48d07c04 SAS |
2911 | } |
2912 | ||
2913 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { | |
2914 | .store = &rcu_data.rcu_cpu_kthread_task, | |
2915 | .thread_should_run = rcu_cpu_kthread_should_run, | |
2916 | .thread_fn = rcu_cpu_kthread, | |
2917 | .thread_comm = "rcuc/%u", | |
2918 | .setup = rcu_cpu_kthread_setup, | |
2919 | .park = rcu_cpu_kthread_park, | |
2920 | }; | |
2921 | ||
2922 | /* | |
2923 | * Spawn per-CPU RCU core processing kthreads. | |
2924 | */ | |
2925 | static int __init rcu_spawn_core_kthreads(void) | |
2926 | { | |
2927 | int cpu; | |
2928 | ||
2929 | for_each_possible_cpu(cpu) | |
2930 | per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0; | |
4b4399b2 | 2931 | if (use_softirq) |
48d07c04 SAS |
2932 | return 0; |
2933 | WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec), | |
2934 | "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__); | |
2935 | return 0; | |
09223371 SL |
2936 | } |
2937 | ||
afd4e696 FW |
2938 | static void rcutree_enqueue(struct rcu_data *rdp, struct rcu_head *head, rcu_callback_t func) |
2939 | { | |
2940 | rcu_segcblist_enqueue(&rdp->cblist, head); | |
2941 | if (__is_kvfree_rcu_offset((unsigned long)func)) | |
2942 | trace_rcu_kvfree_callback(rcu_state.name, head, | |
2943 | (unsigned long)func, | |
2944 | rcu_segcblist_n_cbs(&rdp->cblist)); | |
2945 | else | |
2946 | trace_rcu_callback(rcu_state.name, head, | |
2947 | rcu_segcblist_n_cbs(&rdp->cblist)); | |
2948 | trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued")); | |
2949 | } | |
2950 | ||
29154c57 PM |
2951 | /* |
2952 | * Handle any core-RCU processing required by a call_rcu() invocation. | |
2953 | */ | |
afd4e696 FW |
2954 | static void call_rcu_core(struct rcu_data *rdp, struct rcu_head *head, |
2955 | rcu_callback_t func, unsigned long flags) | |
64db4cff | 2956 | { |
afd4e696 | 2957 | rcutree_enqueue(rdp, head, func); |
62fde6ed PM |
2958 | /* |
2959 | * If called from an extended quiescent state, invoke the RCU | |
2960 | * core in order to force a re-evaluation of RCU's idleness. | |
2961 | */ | |
9910affa | 2962 | if (!rcu_is_watching()) |
62fde6ed PM |
2963 | invoke_rcu_core(); |
2964 | ||
a16b7a69 | 2965 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
29154c57 | 2966 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
2655d57e | 2967 | return; |
64db4cff | 2968 | |
37c72e56 PM |
2969 | /* |
2970 | * Force the grace period if too many callbacks or too long waiting. | |
cd920e5a | 2971 | * Enforce hysteresis, and don't invoke rcu_force_quiescent_state() |
37c72e56 | 2972 | * if some other CPU has recently done so. Also, don't bother |
cd920e5a | 2973 | * invoking rcu_force_quiescent_state() if the newly enqueued callback |
37c72e56 PM |
2974 | * is the only one waiting for a grace period to complete. |
2975 | */ | |
15fecf89 PM |
2976 | if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) > |
2977 | rdp->qlen_last_fqs_check + qhimark)) { | |
b52573d2 PM |
2978 | |
2979 | /* Are we ignoring a completed grace period? */ | |
15cabdff | 2980 | note_gp_changes(rdp); |
b52573d2 PM |
2981 | |
2982 | /* Start a new grace period if one not already started. */ | |
de8e8730 | 2983 | if (!rcu_gp_in_progress()) { |
c6e09b97 | 2984 | rcu_accelerate_cbs_unlocked(rdp->mynode, rdp); |
b52573d2 PM |
2985 | } else { |
2986 | /* Give the grace period a kick. */ | |
d5a9a8c3 | 2987 | rdp->blimit = DEFAULT_MAX_RCU_BLIMIT; |
2431774f | 2988 | if (READ_ONCE(rcu_state.n_force_qs) == rdp->n_force_qs_snap && |
15fecf89 | 2989 | rcu_segcblist_first_pend_cb(&rdp->cblist) != head) |
cd920e5a | 2990 | rcu_force_quiescent_state(); |
2431774f | 2991 | rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs); |
15fecf89 | 2992 | rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist); |
b52573d2 | 2993 | } |
4cdfc175 | 2994 | } |
29154c57 PM |
2995 | } |
2996 | ||
ae150184 PM |
2997 | /* |
2998 | * RCU callback function to leak a callback. | |
2999 | */ | |
3000 | static void rcu_leak_callback(struct rcu_head *rhp) | |
3001 | { | |
3002 | } | |
3003 | ||
3fbfbf7a | 3004 | /* |
b2b00ddf PM |
3005 | * Check and if necessary update the leaf rcu_node structure's |
3006 | * ->cbovldmask bit corresponding to the current CPU based on that CPU's | |
3007 | * number of queued RCU callbacks. The caller must hold the leaf rcu_node | |
3008 | * structure's ->lock. | |
3fbfbf7a | 3009 | */ |
b2b00ddf PM |
3010 | static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp) |
3011 | { | |
3012 | raw_lockdep_assert_held_rcu_node(rnp); | |
3013 | if (qovld_calc <= 0) | |
3014 | return; // Early boot and wildcard value set. | |
3015 | if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) | |
3016 | WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask); | |
3017 | else | |
3018 | WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask); | |
3019 | } | |
3020 | ||
3021 | /* | |
3022 | * Check and if necessary update the leaf rcu_node structure's | |
3023 | * ->cbovldmask bit corresponding to the current CPU based on that CPU's | |
3024 | * number of queued RCU callbacks. No locks need be held, but the | |
3025 | * caller must have disabled interrupts. | |
3026 | * | |
3027 | * Note that this function ignores the possibility that there are a lot | |
3028 | * of callbacks all of which have already seen the end of their respective | |
3029 | * grace periods. This omission is due to the need for no-CBs CPUs to | |
3030 | * be holding ->nocb_lock to do this check, which is too heavy for a | |
3031 | * common-case operation. | |
3fbfbf7a | 3032 | */ |
b2b00ddf PM |
3033 | static void check_cb_ovld(struct rcu_data *rdp) |
3034 | { | |
3035 | struct rcu_node *const rnp = rdp->mynode; | |
3036 | ||
3037 | if (qovld_calc <= 0 || | |
3038 | ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) == | |
3039 | !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask))) | |
3040 | return; // Early boot wildcard value or already set correctly. | |
3041 | raw_spin_lock_rcu_node(rnp); | |
3042 | check_cb_ovld_locked(rdp, rnp); | |
3043 | raw_spin_unlock_rcu_node(rnp); | |
3044 | } | |
3045 | ||
3cb278e7 | 3046 | static void |
cf7066b9 | 3047 | __call_rcu_common(struct rcu_head *head, rcu_callback_t func, bool lazy_in) |
64db4cff | 3048 | { |
b4b7914a | 3049 | static atomic_t doublefrees; |
64db4cff | 3050 | unsigned long flags; |
cf7066b9 | 3051 | bool lazy; |
64db4cff PM |
3052 | struct rcu_data *rdp; |
3053 | ||
b8f2ed53 PM |
3054 | /* Misaligned rcu_head! */ |
3055 | WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1)); | |
3056 | ||
ae150184 | 3057 | if (debug_rcu_head_queue(head)) { |
fa3c6647 PM |
3058 | /* |
3059 | * Probable double call_rcu(), so leak the callback. | |
3060 | * Use rcu:rcu_callback trace event to find the previous | |
1fe09ebe | 3061 | * time callback was passed to call_rcu(). |
fa3c6647 | 3062 | */ |
b4b7914a PM |
3063 | if (atomic_inc_return(&doublefrees) < 4) { |
3064 | pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func); | |
3065 | mem_dump_obj(head); | |
3066 | } | |
7d0ae808 | 3067 | WRITE_ONCE(head->func, rcu_leak_callback); |
ae150184 PM |
3068 | return; |
3069 | } | |
64db4cff PM |
3070 | head->func = func; |
3071 | head->next = NULL; | |
300c0c5e | 3072 | kasan_record_aux_stack_noalloc(head); |
d818cc76 | 3073 | local_irq_save(flags); |
da1df50d | 3074 | rdp = this_cpu_ptr(&rcu_data); |
cf7066b9 | 3075 | lazy = lazy_in && !rcu_async_should_hurry(); |
64db4cff PM |
3076 | |
3077 | /* Add the callback to our list. */ | |
5d6742b3 PM |
3078 | if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) { |
3079 | // This can trigger due to call_rcu() from offline CPU: | |
3080 | WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE); | |
34404ca8 | 3081 | WARN_ON_ONCE(!rcu_is_watching()); |
5d6742b3 PM |
3082 | // Very early boot, before rcu_init(). Initialize if needed |
3083 | // and then drop through to queue the callback. | |
15fecf89 PM |
3084 | if (rcu_segcblist_empty(&rdp->cblist)) |
3085 | rcu_segcblist_init(&rdp->cblist); | |
0d8ee37e | 3086 | } |
77a40f97 | 3087 | |
b2b00ddf | 3088 | check_cb_ovld(rdp); |
d4c08f2a | 3089 | |
afd4e696 FW |
3090 | if (unlikely(rcu_rdp_is_offloaded(rdp))) |
3091 | call_rcu_nocb(rdp, head, func, flags, lazy); | |
3092 | else | |
3093 | call_rcu_core(rdp, head, func, flags); | |
b913c3fe | 3094 | local_irq_restore(flags); |
64db4cff | 3095 | } |
64db4cff | 3096 | |
3cb278e7 | 3097 | #ifdef CONFIG_RCU_LAZY |
7f66f099 QY |
3098 | static bool enable_rcu_lazy __read_mostly = !IS_ENABLED(CONFIG_RCU_LAZY_DEFAULT_OFF); |
3099 | module_param(enable_rcu_lazy, bool, 0444); | |
3100 | ||
3cb278e7 JFG |
3101 | /** |
3102 | * call_rcu_hurry() - Queue RCU callback for invocation after grace period, and | |
3103 | * flush all lazy callbacks (including the new one) to the main ->cblist while | |
3104 | * doing so. | |
3105 | * | |
3106 | * @head: structure to be used for queueing the RCU updates. | |
3107 | * @func: actual callback function to be invoked after the grace period | |
3108 | * | |
3109 | * The callback function will be invoked some time after a full grace | |
3110 | * period elapses, in other words after all pre-existing RCU read-side | |
3111 | * critical sections have completed. | |
3112 | * | |
3113 | * Use this API instead of call_rcu() if you don't want the callback to be | |
3114 | * invoked after very long periods of time, which can happen on systems without | |
3115 | * memory pressure and on systems which are lightly loaded or mostly idle. | |
3116 | * This function will cause callbacks to be invoked sooner than later at the | |
3117 | * expense of extra power. Other than that, this function is identical to, and | |
3118 | * reuses call_rcu()'s logic. Refer to call_rcu() for more details about memory | |
3119 | * ordering and other functionality. | |
3120 | */ | |
3121 | void call_rcu_hurry(struct rcu_head *head, rcu_callback_t func) | |
3122 | { | |
4502138a | 3123 | __call_rcu_common(head, func, false); |
3cb278e7 JFG |
3124 | } |
3125 | EXPORT_SYMBOL_GPL(call_rcu_hurry); | |
7f66f099 QY |
3126 | #else |
3127 | #define enable_rcu_lazy false | |
3cb278e7 JFG |
3128 | #endif |
3129 | ||
3130 | /** | |
3131 | * call_rcu() - Queue an RCU callback for invocation after a grace period. | |
3132 | * By default the callbacks are 'lazy' and are kept hidden from the main | |
3133 | * ->cblist to prevent starting of grace periods too soon. | |
3134 | * If you desire grace periods to start very soon, use call_rcu_hurry(). | |
3135 | * | |
3136 | * @head: structure to be used for queueing the RCU updates. | |
3137 | * @func: actual callback function to be invoked after the grace period | |
3138 | * | |
3139 | * The callback function will be invoked some time after a full grace | |
3140 | * period elapses, in other words after all pre-existing RCU read-side | |
3141 | * critical sections have completed. However, the callback function | |
3142 | * might well execute concurrently with RCU read-side critical sections | |
3143 | * that started after call_rcu() was invoked. | |
3144 | * | |
3145 | * RCU read-side critical sections are delimited by rcu_read_lock() | |
3146 | * and rcu_read_unlock(), and may be nested. In addition, but only in | |
3147 | * v5.0 and later, regions of code across which interrupts, preemption, | |
3148 | * or softirqs have been disabled also serve as RCU read-side critical | |
3149 | * sections. This includes hardware interrupt handlers, softirq handlers, | |
3150 | * and NMI handlers. | |
3151 | * | |
3152 | * Note that all CPUs must agree that the grace period extended beyond | |
3153 | * all pre-existing RCU read-side critical section. On systems with more | |
3154 | * than one CPU, this means that when "func()" is invoked, each CPU is | |
3155 | * guaranteed to have executed a full memory barrier since the end of its | |
3156 | * last RCU read-side critical section whose beginning preceded the call | |
3157 | * to call_rcu(). It also means that each CPU executing an RCU read-side | |
3158 | * critical section that continues beyond the start of "func()" must have | |
3159 | * executed a memory barrier after the call_rcu() but before the beginning | |
3160 | * of that RCU read-side critical section. Note that these guarantees | |
3161 | * include CPUs that are offline, idle, or executing in user mode, as | |
3162 | * well as CPUs that are executing in the kernel. | |
3163 | * | |
3164 | * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the | |
3165 | * resulting RCU callback function "func()", then both CPU A and CPU B are | |
3166 | * guaranteed to execute a full memory barrier during the time interval | |
3167 | * between the call to call_rcu() and the invocation of "func()" -- even | |
3168 | * if CPU A and CPU B are the same CPU (but again only if the system has | |
3169 | * more than one CPU). | |
3170 | * | |
3171 | * Implementation of these memory-ordering guarantees is described here: | |
3172 | * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst. | |
3173 | */ | |
3174 | void call_rcu(struct rcu_head *head, rcu_callback_t func) | |
3175 | { | |
7f66f099 | 3176 | __call_rcu_common(head, func, enable_rcu_lazy); |
3cb278e7 JFG |
3177 | } |
3178 | EXPORT_SYMBOL_GPL(call_rcu); | |
a35d1690 BP |
3179 | |
3180 | /* Maximum number of jiffies to wait before draining a batch. */ | |
51824b78 | 3181 | #define KFREE_DRAIN_JIFFIES (5 * HZ) |
0392bebe | 3182 | #define KFREE_N_BATCHES 2 |
5f3c8d62 | 3183 | #define FREE_N_CHANNELS 2 |
34c88174 URS |
3184 | |
3185 | /** | |
5f3c8d62 | 3186 | * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers |
27538e18 | 3187 | * @list: List node. All blocks are linked between each other |
cc37d520 | 3188 | * @gp_snap: Snapshot of RCU state for objects placed to this bulk |
34c88174 | 3189 | * @nr_records: Number of active pointers in the array |
5f3c8d62 | 3190 | * @records: Array of the kvfree_rcu() pointers |
34c88174 | 3191 | */ |
5f3c8d62 | 3192 | struct kvfree_rcu_bulk_data { |
27538e18 | 3193 | struct list_head list; |
cdfa0f6f | 3194 | struct rcu_gp_oldstate gp_snap; |
34c88174 | 3195 | unsigned long nr_records; |
3af84862 | 3196 | void *records[]; |
34c88174 URS |
3197 | }; |
3198 | ||
3af84862 URS |
3199 | /* |
3200 | * This macro defines how many entries the "records" array | |
3201 | * will contain. It is based on the fact that the size of | |
5f3c8d62 | 3202 | * kvfree_rcu_bulk_data structure becomes exactly one page. |
3af84862 | 3203 | */ |
5f3c8d62 URS |
3204 | #define KVFREE_BULK_MAX_ENTR \ |
3205 | ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *)) | |
3af84862 | 3206 | |
a35d1690 | 3207 | /** |
0392bebe | 3208 | * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests |
a35d1690 | 3209 | * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period |
0392bebe | 3210 | * @head_free: List of kfree_rcu() objects waiting for a grace period |
f32276a3 | 3211 | * @head_free_gp_snap: Grace-period snapshot to check for attempted premature frees. |
27538e18 | 3212 | * @bulk_head_free: Bulk-List of kvfree_rcu() objects waiting for a grace period |
0392bebe JFG |
3213 | * @krcp: Pointer to @kfree_rcu_cpu structure |
3214 | */ | |
3215 | ||
3216 | struct kfree_rcu_cpu_work { | |
3217 | struct rcu_work rcu_work; | |
3218 | struct rcu_head *head_free; | |
f32276a3 | 3219 | struct rcu_gp_oldstate head_free_gp_snap; |
27538e18 | 3220 | struct list_head bulk_head_free[FREE_N_CHANNELS]; |
0392bebe JFG |
3221 | struct kfree_rcu_cpu *krcp; |
3222 | }; | |
3223 | ||
3224 | /** | |
3225 | * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period | |
a35d1690 | 3226 | * @head: List of kfree_rcu() objects not yet waiting for a grace period |
2ca836b1 | 3227 | * @head_gp_snap: Snapshot of RCU state for objects placed to "@head" |
27538e18 | 3228 | * @bulk_head: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period |
0392bebe | 3229 | * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period |
a35d1690 BP |
3230 | * @lock: Synchronize access to this structure |
3231 | * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES | |
69f08d39 | 3232 | * @initialized: The @rcu_work fields have been initialized |
4c33464a URS |
3233 | * @head_count: Number of objects in rcu_head singular list |
3234 | * @bulk_count: Number of objects in bulk-list | |
72a2fbda MCC |
3235 | * @bkvcache: |
3236 | * A simple cache list that contains objects for reuse purpose. | |
3237 | * In order to save some per-cpu space the list is singular. | |
3238 | * Even though it is lockless an access has to be protected by the | |
3239 | * per-cpu lock. | |
56292e86 | 3240 | * @page_cache_work: A work to refill the cache when it is empty |
d0bfa8b3 | 3241 | * @backoff_page_cache_fill: Delay cache refills |
56292e86 URS |
3242 | * @work_in_progress: Indicates that page_cache_work is running |
3243 | * @hrtimer: A hrtimer for scheduling a page_cache_work | |
72a2fbda | 3244 | * @nr_bkv_objs: number of allocated objects at @bkvcache. |
a35d1690 BP |
3245 | * |
3246 | * This is a per-CPU structure. The reason that it is not included in | |
3247 | * the rcu_data structure is to permit this code to be extracted from | |
3248 | * the RCU files. Such extraction could allow further optimization of | |
3249 | * the interactions with the slab allocators. | |
3250 | */ | |
3251 | struct kfree_rcu_cpu { | |
4c33464a URS |
3252 | // Objects queued on a linked list |
3253 | // through their rcu_head structures. | |
a35d1690 | 3254 | struct rcu_head *head; |
2ca836b1 | 3255 | unsigned long head_gp_snap; |
4c33464a URS |
3256 | atomic_t head_count; |
3257 | ||
3258 | // Objects queued on a bulk-list. | |
27538e18 | 3259 | struct list_head bulk_head[FREE_N_CHANNELS]; |
4c33464a URS |
3260 | atomic_t bulk_count[FREE_N_CHANNELS]; |
3261 | ||
0392bebe | 3262 | struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES]; |
8ac88f71 | 3263 | raw_spinlock_t lock; |
a35d1690 | 3264 | struct delayed_work monitor_work; |
a35d1690 | 3265 | bool initialized; |
56292e86 | 3266 | |
d0bfa8b3 ZQ |
3267 | struct delayed_work page_cache_work; |
3268 | atomic_t backoff_page_cache_fill; | |
56292e86 URS |
3269 | atomic_t work_in_progress; |
3270 | struct hrtimer hrtimer; | |
3271 | ||
53c72b59 URS |
3272 | struct llist_head bkvcache; |
3273 | int nr_bkv_objs; | |
a35d1690 BP |
3274 | }; |
3275 | ||
69f08d39 SAS |
3276 | static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = { |
3277 | .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock), | |
3278 | }; | |
a35d1690 | 3279 | |
34c88174 | 3280 | static __always_inline void |
5f3c8d62 | 3281 | debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead) |
34c88174 URS |
3282 | { |
3283 | #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD | |
446044eb JFG |
3284 | int i; |
3285 | ||
3286 | for (i = 0; i < bhead->nr_records; i++) | |
3287 | debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i])); | |
34c88174 URS |
3288 | #endif |
3289 | } | |
3290 | ||
952371d6 URS |
3291 | static inline struct kfree_rcu_cpu * |
3292 | krc_this_cpu_lock(unsigned long *flags) | |
3293 | { | |
3294 | struct kfree_rcu_cpu *krcp; | |
3295 | ||
3296 | local_irq_save(*flags); // For safely calling this_cpu_ptr(). | |
3297 | krcp = this_cpu_ptr(&krc); | |
69f08d39 | 3298 | raw_spin_lock(&krcp->lock); |
952371d6 URS |
3299 | |
3300 | return krcp; | |
3301 | } | |
3302 | ||
3303 | static inline void | |
3304 | krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags) | |
3305 | { | |
7ffc9ec8 | 3306 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
952371d6 URS |
3307 | } |
3308 | ||
5f3c8d62 | 3309 | static inline struct kvfree_rcu_bulk_data * |
53c72b59 URS |
3310 | get_cached_bnode(struct kfree_rcu_cpu *krcp) |
3311 | { | |
3312 | if (!krcp->nr_bkv_objs) | |
3313 | return NULL; | |
3314 | ||
ac7625eb | 3315 | WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1); |
5f3c8d62 | 3316 | return (struct kvfree_rcu_bulk_data *) |
53c72b59 URS |
3317 | llist_del_first(&krcp->bkvcache); |
3318 | } | |
3319 | ||
3320 | static inline bool | |
3321 | put_cached_bnode(struct kfree_rcu_cpu *krcp, | |
5f3c8d62 | 3322 | struct kvfree_rcu_bulk_data *bnode) |
53c72b59 URS |
3323 | { |
3324 | // Check the limit. | |
3325 | if (krcp->nr_bkv_objs >= rcu_min_cached_objs) | |
3326 | return false; | |
3327 | ||
3328 | llist_add((struct llist_node *) bnode, &krcp->bkvcache); | |
ac7625eb | 3329 | WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1); |
53c72b59 | 3330 | return true; |
53c72b59 URS |
3331 | } |
3332 | ||
d0bfa8b3 ZQ |
3333 | static int |
3334 | drain_page_cache(struct kfree_rcu_cpu *krcp) | |
3335 | { | |
3336 | unsigned long flags; | |
3337 | struct llist_node *page_list, *pos, *n; | |
3338 | int freed = 0; | |
53c72b59 | 3339 | |
6b706e56 Z |
3340 | if (!rcu_min_cached_objs) |
3341 | return 0; | |
3342 | ||
d0bfa8b3 ZQ |
3343 | raw_spin_lock_irqsave(&krcp->lock, flags); |
3344 | page_list = llist_del_all(&krcp->bkvcache); | |
ac7625eb | 3345 | WRITE_ONCE(krcp->nr_bkv_objs, 0); |
d0bfa8b3 | 3346 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
53c72b59 | 3347 | |
d0bfa8b3 ZQ |
3348 | llist_for_each_safe(pos, n, page_list) { |
3349 | free_page((unsigned long)pos); | |
3350 | freed++; | |
3351 | } | |
3352 | ||
3353 | return freed; | |
53c72b59 URS |
3354 | } |
3355 | ||
8c15a9e8 URS |
3356 | static void |
3357 | kvfree_rcu_bulk(struct kfree_rcu_cpu *krcp, | |
3358 | struct kvfree_rcu_bulk_data *bnode, int idx) | |
3359 | { | |
3360 | unsigned long flags; | |
3361 | int i; | |
3362 | ||
cdfa0f6f | 3363 | if (!WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&bnode->gp_snap))) { |
1e237994 | 3364 | debug_rcu_bhead_unqueue(bnode); |
cdfa0f6f PM |
3365 | rcu_lock_acquire(&rcu_callback_map); |
3366 | if (idx == 0) { // kmalloc() / kfree(). | |
3367 | trace_rcu_invoke_kfree_bulk_callback( | |
3368 | rcu_state.name, bnode->nr_records, | |
3369 | bnode->records); | |
3370 | ||
3371 | kfree_bulk(bnode->nr_records, bnode->records); | |
3372 | } else { // vmalloc() / vfree(). | |
3373 | for (i = 0; i < bnode->nr_records; i++) { | |
3374 | trace_rcu_invoke_kvfree_callback( | |
3375 | rcu_state.name, bnode->records[i], 0); | |
3376 | ||
3377 | vfree(bnode->records[i]); | |
3378 | } | |
8c15a9e8 | 3379 | } |
cdfa0f6f | 3380 | rcu_lock_release(&rcu_callback_map); |
8c15a9e8 | 3381 | } |
8c15a9e8 URS |
3382 | |
3383 | raw_spin_lock_irqsave(&krcp->lock, flags); | |
3384 | if (put_cached_bnode(krcp, bnode)) | |
3385 | bnode = NULL; | |
3386 | raw_spin_unlock_irqrestore(&krcp->lock, flags); | |
3387 | ||
3388 | if (bnode) | |
3389 | free_page((unsigned long) bnode); | |
3390 | ||
3391 | cond_resched_tasks_rcu_qs(); | |
3392 | } | |
3393 | ||
3394 | static void | |
3395 | kvfree_rcu_list(struct rcu_head *head) | |
3396 | { | |
3397 | struct rcu_head *next; | |
3398 | ||
3399 | for (; head; head = next) { | |
3400 | void *ptr = (void *) head->func; | |
3401 | unsigned long offset = (void *) head - ptr; | |
3402 | ||
3403 | next = head->next; | |
3404 | debug_rcu_head_unqueue((struct rcu_head *)ptr); | |
3405 | rcu_lock_acquire(&rcu_callback_map); | |
3406 | trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset); | |
3407 | ||
3408 | if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset))) | |
3409 | kvfree(ptr); | |
3410 | ||
3411 | rcu_lock_release(&rcu_callback_map); | |
3412 | cond_resched_tasks_rcu_qs(); | |
3413 | } | |
3414 | } | |
3415 | ||
495aa969 | 3416 | /* |
a35d1690 | 3417 | * This function is invoked in workqueue context after a grace period. |
27538e18 | 3418 | * It frees all the objects queued on ->bulk_head_free or ->head_free. |
495aa969 | 3419 | */ |
a35d1690 BP |
3420 | static void kfree_rcu_work(struct work_struct *work) |
3421 | { | |
3422 | unsigned long flags; | |
27538e18 URS |
3423 | struct kvfree_rcu_bulk_data *bnode, *n; |
3424 | struct list_head bulk_head[FREE_N_CHANNELS]; | |
8c15a9e8 | 3425 | struct rcu_head *head; |
a35d1690 | 3426 | struct kfree_rcu_cpu *krcp; |
0392bebe | 3427 | struct kfree_rcu_cpu_work *krwp; |
f32276a3 | 3428 | struct rcu_gp_oldstate head_gp_snap; |
8c15a9e8 | 3429 | int i; |
a35d1690 | 3430 | |
0392bebe | 3431 | krwp = container_of(to_rcu_work(work), |
cc37d520 | 3432 | struct kfree_rcu_cpu_work, rcu_work); |
0392bebe | 3433 | krcp = krwp->krcp; |
34c88174 | 3434 | |
8ac88f71 | 3435 | raw_spin_lock_irqsave(&krcp->lock, flags); |
5f3c8d62 | 3436 | // Channels 1 and 2. |
27538e18 URS |
3437 | for (i = 0; i < FREE_N_CHANNELS; i++) |
3438 | list_replace_init(&krwp->bulk_head_free[i], &bulk_head[i]); | |
34c88174 | 3439 | |
5f3c8d62 | 3440 | // Channel 3. |
0392bebe JFG |
3441 | head = krwp->head_free; |
3442 | krwp->head_free = NULL; | |
f32276a3 | 3443 | head_gp_snap = krwp->head_free_gp_snap; |
8ac88f71 | 3444 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
a35d1690 | 3445 | |
277ffe1b | 3446 | // Handle the first two channels. |
5f3c8d62 | 3447 | for (i = 0; i < FREE_N_CHANNELS; i++) { |
cc37d520 | 3448 | // Start from the tail page, so a GP is likely passed for it. |
8c15a9e8 URS |
3449 | list_for_each_entry_safe(bnode, n, &bulk_head[i], list) |
3450 | kvfree_rcu_bulk(krcp, bnode, i); | |
34c88174 URS |
3451 | } |
3452 | ||
3453 | /* | |
d8628f35 URS |
3454 | * This is used when the "bulk" path can not be used for the |
3455 | * double-argument of kvfree_rcu(). This happens when the | |
3456 | * page-cache is empty, which means that objects are instead | |
3457 | * queued on a linked list through their rcu_head structures. | |
3458 | * This list is named "Channel 3". | |
34c88174 | 3459 | */ |
f32276a3 URS |
3460 | if (head && !WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&head_gp_snap))) |
3461 | kvfree_rcu_list(head); | |
a35d1690 BP |
3462 | } |
3463 | ||
82d26c36 JFG |
3464 | static bool |
3465 | need_offload_krc(struct kfree_rcu_cpu *krcp) | |
3466 | { | |
3467 | int i; | |
3468 | ||
3469 | for (i = 0; i < FREE_N_CHANNELS; i++) | |
27538e18 | 3470 | if (!list_empty(&krcp->bulk_head[i])) |
82d26c36 JFG |
3471 | return true; |
3472 | ||
96274561 | 3473 | return !!READ_ONCE(krcp->head); |
82d26c36 JFG |
3474 | } |
3475 | ||
5da7cb19 ZD |
3476 | static bool |
3477 | need_wait_for_krwp_work(struct kfree_rcu_cpu_work *krwp) | |
3478 | { | |
3479 | int i; | |
3480 | ||
3481 | for (i = 0; i < FREE_N_CHANNELS; i++) | |
3482 | if (!list_empty(&krwp->bulk_head_free[i])) | |
3483 | return true; | |
3484 | ||
3485 | return !!krwp->head_free; | |
3486 | } | |
3487 | ||
4c33464a URS |
3488 | static int krc_count(struct kfree_rcu_cpu *krcp) |
3489 | { | |
3490 | int sum = atomic_read(&krcp->head_count); | |
3491 | int i; | |
3492 | ||
3493 | for (i = 0; i < FREE_N_CHANNELS; i++) | |
3494 | sum += atomic_read(&krcp->bulk_count[i]); | |
3495 | ||
3496 | return sum; | |
82d26c36 JFG |
3497 | } |
3498 | ||
51824b78 URS |
3499 | static void |
3500 | schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp) | |
3501 | { | |
3502 | long delay, delay_left; | |
3503 | ||
4c33464a | 3504 | delay = krc_count(krcp) >= KVFREE_BULK_MAX_ENTR ? 1:KFREE_DRAIN_JIFFIES; |
51824b78 URS |
3505 | if (delayed_work_pending(&krcp->monitor_work)) { |
3506 | delay_left = krcp->monitor_work.timer.expires - jiffies; | |
3507 | if (delay < delay_left) | |
3508 | mod_delayed_work(system_wq, &krcp->monitor_work, delay); | |
3509 | return; | |
3510 | } | |
3511 | queue_delayed_work(system_wq, &krcp->monitor_work, delay); | |
3512 | } | |
3513 | ||
2ca836b1 URS |
3514 | static void |
3515 | kvfree_rcu_drain_ready(struct kfree_rcu_cpu *krcp) | |
3516 | { | |
3517 | struct list_head bulk_ready[FREE_N_CHANNELS]; | |
3518 | struct kvfree_rcu_bulk_data *bnode, *n; | |
3519 | struct rcu_head *head_ready = NULL; | |
3520 | unsigned long flags; | |
3521 | int i; | |
3522 | ||
3523 | raw_spin_lock_irqsave(&krcp->lock, flags); | |
3524 | for (i = 0; i < FREE_N_CHANNELS; i++) { | |
3525 | INIT_LIST_HEAD(&bulk_ready[i]); | |
3526 | ||
3527 | list_for_each_entry_safe_reverse(bnode, n, &krcp->bulk_head[i], list) { | |
cdfa0f6f | 3528 | if (!poll_state_synchronize_rcu_full(&bnode->gp_snap)) |
2ca836b1 URS |
3529 | break; |
3530 | ||
3531 | atomic_sub(bnode->nr_records, &krcp->bulk_count[i]); | |
3532 | list_move(&bnode->list, &bulk_ready[i]); | |
3533 | } | |
3534 | } | |
3535 | ||
3536 | if (krcp->head && poll_state_synchronize_rcu(krcp->head_gp_snap)) { | |
3537 | head_ready = krcp->head; | |
3538 | atomic_set(&krcp->head_count, 0); | |
3539 | WRITE_ONCE(krcp->head, NULL); | |
3540 | } | |
3541 | raw_spin_unlock_irqrestore(&krcp->lock, flags); | |
3542 | ||
3543 | for (i = 0; i < FREE_N_CHANNELS; i++) { | |
3544 | list_for_each_entry_safe(bnode, n, &bulk_ready[i], list) | |
3545 | kvfree_rcu_bulk(krcp, bnode, i); | |
3546 | } | |
3547 | ||
3548 | if (head_ready) | |
3549 | kvfree_rcu_list(head_ready); | |
3550 | } | |
3551 | ||
495aa969 | 3552 | /* |
a78d4a2a | 3553 | * This function is invoked after the KFREE_DRAIN_JIFFIES timeout. |
a35d1690 | 3554 | */ |
a78d4a2a | 3555 | static void kfree_rcu_monitor(struct work_struct *work) |
a35d1690 | 3556 | { |
a78d4a2a URS |
3557 | struct kfree_rcu_cpu *krcp = container_of(work, |
3558 | struct kfree_rcu_cpu, monitor_work.work); | |
3559 | unsigned long flags; | |
5f3c8d62 | 3560 | int i, j; |
0392bebe | 3561 | |
2ca836b1 URS |
3562 | // Drain ready for reclaim. |
3563 | kvfree_rcu_drain_ready(krcp); | |
3564 | ||
a78d4a2a | 3565 | raw_spin_lock_irqsave(&krcp->lock, flags); |
a35d1690 | 3566 | |
a78d4a2a | 3567 | // Attempt to start a new batch. |
34c88174 | 3568 | for (i = 0; i < KFREE_N_BATCHES; i++) { |
a78d4a2a | 3569 | struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]); |
a35d1690 | 3570 | |
5da7cb19 ZD |
3571 | // Try to detach bulk_head or head and attach it, only when |
3572 | // all channels are free. Any channel is not free means at krwp | |
3573 | // there is on-going rcu work to handle krwp's free business. | |
3574 | if (need_wait_for_krwp_work(krwp)) | |
3575 | continue; | |
27538e18 | 3576 | |
5da7cb19 ZD |
3577 | // kvfree_rcu_drain_ready() might handle this krcp, if so give up. |
3578 | if (need_offload_krc(krcp)) { | |
d8628f35 URS |
3579 | // Channel 1 corresponds to the SLAB-pointer bulk path. |
3580 | // Channel 2 corresponds to vmalloc-pointer bulk path. | |
5f3c8d62 | 3581 | for (j = 0; j < FREE_N_CHANNELS; j++) { |
4c33464a | 3582 | if (list_empty(&krwp->bulk_head_free[j])) { |
4c33464a | 3583 | atomic_set(&krcp->bulk_count[j], 0); |
2ca836b1 URS |
3584 | list_replace_init(&krcp->bulk_head[j], |
3585 | &krwp->bulk_head_free[j]); | |
5f3c8d62 | 3586 | } |
34c88174 URS |
3587 | } |
3588 | ||
d8628f35 URS |
3589 | // Channel 3 corresponds to both SLAB and vmalloc |
3590 | // objects queued on the linked list. | |
34c88174 URS |
3591 | if (!krwp->head_free) { |
3592 | krwp->head_free = krcp->head; | |
f32276a3 | 3593 | get_state_synchronize_rcu_full(&krwp->head_free_gp_snap); |
4c33464a | 3594 | atomic_set(&krcp->head_count, 0); |
2ca836b1 | 3595 | WRITE_ONCE(krcp->head, NULL); |
34c88174 URS |
3596 | } |
3597 | ||
a78d4a2a URS |
3598 | // One work is per one batch, so there are three |
3599 | // "free channels", the batch can handle. It can | |
3600 | // be that the work is in the pending state when | |
3601 | // channels have been detached following by each | |
3602 | // other. | |
34c88174 | 3603 | queue_rcu_work(system_wq, &krwp->rcu_work); |
34c88174 URS |
3604 | } |
3605 | } | |
3606 | ||
8fc5494a URS |
3607 | raw_spin_unlock_irqrestore(&krcp->lock, flags); |
3608 | ||
a78d4a2a URS |
3609 | // If there is nothing to detach, it means that our job is |
3610 | // successfully done here. In case of having at least one | |
3611 | // of the channels that is still busy we should rearm the | |
3612 | // work to repeat an attempt. Because previous batches are | |
3613 | // still in progress. | |
82d26c36 | 3614 | if (need_offload_krc(krcp)) |
51824b78 | 3615 | schedule_delayed_monitor_work(krcp); |
a35d1690 BP |
3616 | } |
3617 | ||
56292e86 URS |
3618 | static enum hrtimer_restart |
3619 | schedule_page_work_fn(struct hrtimer *t) | |
3620 | { | |
3621 | struct kfree_rcu_cpu *krcp = | |
3622 | container_of(t, struct kfree_rcu_cpu, hrtimer); | |
3623 | ||
d0bfa8b3 | 3624 | queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0); |
56292e86 URS |
3625 | return HRTIMER_NORESTART; |
3626 | } | |
3627 | ||
3628 | static void fill_page_cache_func(struct work_struct *work) | |
3629 | { | |
3630 | struct kvfree_rcu_bulk_data *bnode; | |
3631 | struct kfree_rcu_cpu *krcp = | |
3632 | container_of(work, struct kfree_rcu_cpu, | |
d0bfa8b3 | 3633 | page_cache_work.work); |
56292e86 | 3634 | unsigned long flags; |
d0bfa8b3 | 3635 | int nr_pages; |
56292e86 URS |
3636 | bool pushed; |
3637 | int i; | |
3638 | ||
d0bfa8b3 ZQ |
3639 | nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ? |
3640 | 1 : rcu_min_cached_objs; | |
3641 | ||
60888b77 | 3642 | for (i = READ_ONCE(krcp->nr_bkv_objs); i < nr_pages; i++) { |
56292e86 | 3643 | bnode = (struct kvfree_rcu_bulk_data *) |
ee6ddf58 | 3644 | __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); |
56292e86 | 3645 | |
093590c1 MH |
3646 | if (!bnode) |
3647 | break; | |
56292e86 | 3648 | |
093590c1 MH |
3649 | raw_spin_lock_irqsave(&krcp->lock, flags); |
3650 | pushed = put_cached_bnode(krcp, bnode); | |
3651 | raw_spin_unlock_irqrestore(&krcp->lock, flags); | |
3652 | ||
3653 | if (!pushed) { | |
3654 | free_page((unsigned long) bnode); | |
3655 | break; | |
56292e86 URS |
3656 | } |
3657 | } | |
3658 | ||
3659 | atomic_set(&krcp->work_in_progress, 0); | |
d0bfa8b3 | 3660 | atomic_set(&krcp->backoff_page_cache_fill, 0); |
56292e86 URS |
3661 | } |
3662 | ||
3663 | static void | |
3664 | run_page_cache_worker(struct kfree_rcu_cpu *krcp) | |
3665 | { | |
021a5ff8 URS |
3666 | // If cache disabled, bail out. |
3667 | if (!rcu_min_cached_objs) | |
3668 | return; | |
3669 | ||
56292e86 URS |
3670 | if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING && |
3671 | !atomic_xchg(&krcp->work_in_progress, 1)) { | |
d0bfa8b3 ZQ |
3672 | if (atomic_read(&krcp->backoff_page_cache_fill)) { |
3673 | queue_delayed_work(system_wq, | |
3674 | &krcp->page_cache_work, | |
3675 | msecs_to_jiffies(rcu_delay_page_cache_fill_msec)); | |
3676 | } else { | |
3677 | hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
3678 | krcp->hrtimer.function = schedule_page_work_fn; | |
3679 | hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL); | |
3680 | } | |
56292e86 URS |
3681 | } |
3682 | } | |
3683 | ||
148e3731 URS |
3684 | // Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock() |
3685 | // state specified by flags. If can_alloc is true, the caller must | |
3686 | // be schedulable and not be holding any locks or mutexes that might be | |
3687 | // acquired by the memory allocator or anything that it might invoke. | |
3688 | // Returns true if ptr was successfully recorded, else the caller must | |
3689 | // use a fallback. | |
34c88174 | 3690 | static inline bool |
148e3731 URS |
3691 | add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp, |
3692 | unsigned long *flags, void *ptr, bool can_alloc) | |
34c88174 | 3693 | { |
5f3c8d62 URS |
3694 | struct kvfree_rcu_bulk_data *bnode; |
3695 | int idx; | |
34c88174 | 3696 | |
148e3731 URS |
3697 | *krcp = krc_this_cpu_lock(flags); |
3698 | if (unlikely(!(*krcp)->initialized)) | |
34c88174 URS |
3699 | return false; |
3700 | ||
5f3c8d62 | 3701 | idx = !!is_vmalloc_addr(ptr); |
27538e18 URS |
3702 | bnode = list_first_entry_or_null(&(*krcp)->bulk_head[idx], |
3703 | struct kvfree_rcu_bulk_data, list); | |
34c88174 URS |
3704 | |
3705 | /* Check if a new block is required. */ | |
27538e18 | 3706 | if (!bnode || bnode->nr_records == KVFREE_BULK_MAX_ENTR) { |
148e3731 URS |
3707 | bnode = get_cached_bnode(*krcp); |
3708 | if (!bnode && can_alloc) { | |
3709 | krc_this_cpu_unlock(*krcp, *flags); | |
3e7ce7a1 URS |
3710 | |
3711 | // __GFP_NORETRY - allows a light-weight direct reclaim | |
3712 | // what is OK from minimizing of fallback hitting point of | |
3713 | // view. Apart of that it forbids any OOM invoking what is | |
3714 | // also beneficial since we are about to release memory soon. | |
3715 | // | |
3716 | // __GFP_NOMEMALLOC - prevents from consuming of all the | |
3717 | // memory reserves. Please note we have a fallback path. | |
3718 | // | |
3719 | // __GFP_NOWARN - it is supposed that an allocation can | |
3720 | // be failed under low memory or high memory pressure | |
3721 | // scenarios. | |
148e3731 | 3722 | bnode = (struct kvfree_rcu_bulk_data *) |
3e7ce7a1 | 3723 | __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN); |
309a4316 | 3724 | raw_spin_lock_irqsave(&(*krcp)->lock, *flags); |
148e3731 URS |
3725 | } |
3726 | ||
56292e86 | 3727 | if (!bnode) |
34c88174 URS |
3728 | return false; |
3729 | ||
27538e18 | 3730 | // Initialize the new block and attach it. |
34c88174 | 3731 | bnode->nr_records = 0; |
27538e18 | 3732 | list_add(&bnode->list, &(*krcp)->bulk_head[idx]); |
34c88174 URS |
3733 | } |
3734 | ||
cc37d520 | 3735 | // Finally insert and update the GP for this page. |
27538e18 | 3736 | bnode->records[bnode->nr_records++] = ptr; |
cdfa0f6f | 3737 | get_state_synchronize_rcu_full(&bnode->gp_snap); |
4c33464a | 3738 | atomic_inc(&(*krcp)->bulk_count[idx]); |
34c88174 URS |
3739 | |
3740 | return true; | |
3741 | } | |
3742 | ||
a35d1690 | 3743 | /* |
277ffe1b ZZ |
3744 | * Queue a request for lazy invocation of the appropriate free routine |
3745 | * after a grace period. Please note that three paths are maintained, | |
3746 | * two for the common case using arrays of pointers and a third one that | |
3747 | * is used only when the main paths cannot be used, for example, due to | |
3748 | * memory pressure. | |
a35d1690 | 3749 | * |
c408b215 | 3750 | * Each kvfree_call_rcu() request is added to a batch. The batch will be drained |
34c88174 URS |
3751 | * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will |
3752 | * be free'd in workqueue context. This allows us to: batch requests together to | |
5f3c8d62 | 3753 | * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load. |
495aa969 | 3754 | */ |
04a522b7 | 3755 | void kvfree_call_rcu(struct rcu_head *head, void *ptr) |
495aa969 | 3756 | { |
a35d1690 BP |
3757 | unsigned long flags; |
3758 | struct kfree_rcu_cpu *krcp; | |
3042f83f | 3759 | bool success; |
a35d1690 | 3760 | |
04a522b7 URS |
3761 | /* |
3762 | * Please note there is a limitation for the head-less | |
3763 | * variant, that is why there is a clear rule for such | |
3764 | * objects: it can be used from might_sleep() context | |
3765 | * only. For other places please embed an rcu_head to | |
3766 | * your data. | |
3767 | */ | |
3768 | if (!head) | |
3042f83f | 3769 | might_sleep(); |
3042f83f | 3770 | |
a35d1690 | 3771 | // Queue the object but don't yet schedule the batch. |
446044eb | 3772 | if (debug_rcu_head_queue(ptr)) { |
e99637be JFG |
3773 | // Probable double kfree_rcu(), just leak. |
3774 | WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n", | |
3775 | __func__, head); | |
3042f83f URS |
3776 | |
3777 | // Mark as success and leave. | |
148e3731 | 3778 | return; |
e99637be | 3779 | } |
34c88174 | 3780 | |
300c0c5e | 3781 | kasan_record_aux_stack_noalloc(ptr); |
148e3731 | 3782 | success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head); |
3042f83f | 3783 | if (!success) { |
56292e86 URS |
3784 | run_page_cache_worker(krcp); |
3785 | ||
3042f83f URS |
3786 | if (head == NULL) |
3787 | // Inline if kvfree_rcu(one_arg) call. | |
3788 | goto unlock_return; | |
3789 | ||
04a522b7 | 3790 | head->func = ptr; |
34c88174 | 3791 | head->next = krcp->head; |
8fc5494a | 3792 | WRITE_ONCE(krcp->head, head); |
4c33464a | 3793 | atomic_inc(&krcp->head_count); |
2ca836b1 URS |
3794 | |
3795 | // Take a snapshot for this krcp. | |
3796 | krcp->head_gp_snap = get_state_synchronize_rcu(); | |
3042f83f | 3797 | success = true; |
34c88174 | 3798 | } |
a35d1690 | 3799 | |
5f98fd03 CM |
3800 | /* |
3801 | * The kvfree_rcu() caller considers the pointer freed at this point | |
3802 | * and likely removes any references to it. Since the actual slab | |
3803 | * freeing (and kmemleak_free()) is deferred, tell kmemleak to ignore | |
3804 | * this object (no scanning or false positives reporting). | |
3805 | */ | |
3806 | kmemleak_ignore(ptr); | |
3807 | ||
a35d1690 | 3808 | // Set timer to drain after KFREE_DRAIN_JIFFIES. |
82d26c36 | 3809 | if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING) |
51824b78 | 3810 | schedule_delayed_monitor_work(krcp); |
a35d1690 | 3811 | |
e99637be | 3812 | unlock_return: |
952371d6 | 3813 | krc_this_cpu_unlock(krcp, flags); |
3042f83f URS |
3814 | |
3815 | /* | |
3816 | * Inline kvfree() after synchronize_rcu(). We can do | |
3817 | * it from might_sleep() context only, so the current | |
3818 | * CPU can pass the QS state. | |
3819 | */ | |
3820 | if (!success) { | |
3821 | debug_rcu_head_unqueue((struct rcu_head *) ptr); | |
3822 | synchronize_rcu(); | |
3823 | kvfree(ptr); | |
3824 | } | |
495aa969 | 3825 | } |
c408b215 | 3826 | EXPORT_SYMBOL_GPL(kvfree_call_rcu); |
495aa969 | 3827 | |
9154244c JFG |
3828 | static unsigned long |
3829 | kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc) | |
3830 | { | |
3831 | int cpu; | |
a6a82ce1 | 3832 | unsigned long count = 0; |
9154244c JFG |
3833 | |
3834 | /* Snapshot count of all CPUs */ | |
70060b87 | 3835 | for_each_possible_cpu(cpu) { |
9154244c JFG |
3836 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); |
3837 | ||
4c33464a | 3838 | count += krc_count(krcp); |
ac7625eb | 3839 | count += READ_ONCE(krcp->nr_bkv_objs); |
d0bfa8b3 | 3840 | atomic_set(&krcp->backoff_page_cache_fill, 1); |
9154244c JFG |
3841 | } |
3842 | ||
38269096 | 3843 | return count == 0 ? SHRINK_EMPTY : count; |
9154244c JFG |
3844 | } |
3845 | ||
3846 | static unsigned long | |
3847 | kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc) | |
3848 | { | |
3849 | int cpu, freed = 0; | |
9154244c | 3850 | |
70060b87 | 3851 | for_each_possible_cpu(cpu) { |
9154244c JFG |
3852 | int count; |
3853 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); | |
3854 | ||
4c33464a | 3855 | count = krc_count(krcp); |
d0bfa8b3 | 3856 | count += drain_page_cache(krcp); |
7fe1da33 | 3857 | kfree_rcu_monitor(&krcp->monitor_work.work); |
9154244c JFG |
3858 | |
3859 | sc->nr_to_scan -= count; | |
3860 | freed += count; | |
3861 | ||
3862 | if (sc->nr_to_scan <= 0) | |
3863 | break; | |
3864 | } | |
3865 | ||
c6dfd72b | 3866 | return freed == 0 ? SHRINK_STOP : freed; |
9154244c JFG |
3867 | } |
3868 | ||
a35d1690 BP |
3869 | void __init kfree_rcu_scheduler_running(void) |
3870 | { | |
3871 | int cpu; | |
a35d1690 | 3872 | |
70060b87 | 3873 | for_each_possible_cpu(cpu) { |
a35d1690 BP |
3874 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); |
3875 | ||
82d26c36 | 3876 | if (need_offload_krc(krcp)) |
51824b78 | 3877 | schedule_delayed_monitor_work(krcp); |
a35d1690 BP |
3878 | } |
3879 | } | |
3880 | ||
e5bc3af7 PM |
3881 | /* |
3882 | * During early boot, any blocking grace-period wait automatically | |
258f887a | 3883 | * implies a grace period. |
e5bc3af7 | 3884 | * |
258f887a PM |
3885 | * Later on, this could in theory be the case for kernels built with |
3886 | * CONFIG_SMP=y && CONFIG_PREEMPTION=y running on a single CPU, but this | |
3887 | * is not a common case. Furthermore, this optimization would cause | |
3888 | * the rcu_gp_oldstate structure to expand by 50%, so this potential | |
3889 | * grace-period optimization is ignored once the scheduler is running. | |
e5bc3af7 PM |
3890 | */ |
3891 | static int rcu_blocking_is_gp(void) | |
3892 | { | |
3d1adf7a Z |
3893 | if (rcu_scheduler_active != RCU_SCHEDULER_INACTIVE) { |
3894 | might_sleep(); | |
258f887a | 3895 | return false; |
3d1adf7a | 3896 | } |
258f887a | 3897 | return true; |
e5bc3af7 PM |
3898 | } |
3899 | ||
988f569a URS |
3900 | /* |
3901 | * Helper function for the synchronize_rcu() API. | |
3902 | */ | |
3903 | static void synchronize_rcu_normal(void) | |
3904 | { | |
3905 | struct rcu_synchronize rs; | |
3906 | ||
2053937a URS |
3907 | trace_rcu_sr_normal(rcu_state.name, &rs.head, TPS("request")); |
3908 | ||
988f569a URS |
3909 | if (!READ_ONCE(rcu_normal_wake_from_gp)) { |
3910 | wait_rcu_gp(call_rcu_hurry); | |
2053937a | 3911 | goto trace_complete_out; |
988f569a URS |
3912 | } |
3913 | ||
3914 | init_rcu_head_on_stack(&rs.head); | |
3915 | init_completion(&rs.completion); | |
3916 | ||
3917 | /* | |
3918 | * This code might be preempted, therefore take a GP | |
3919 | * snapshot before adding a request. | |
3920 | */ | |
3921 | if (IS_ENABLED(CONFIG_PROVE_RCU)) | |
3922 | rs.head.func = (void *) get_state_synchronize_rcu(); | |
3923 | ||
3924 | rcu_sr_normal_add_req(&rs); | |
3925 | ||
3926 | /* Kick a GP and start waiting. */ | |
3927 | (void) start_poll_synchronize_rcu(); | |
3928 | ||
3929 | /* Now we can wait. */ | |
3930 | wait_for_completion(&rs.completion); | |
3931 | destroy_rcu_head_on_stack(&rs.head); | |
2053937a URS |
3932 | |
3933 | trace_complete_out: | |
3934 | trace_rcu_sr_normal(rcu_state.name, &rs.head, TPS("complete")); | |
988f569a URS |
3935 | } |
3936 | ||
e5bc3af7 PM |
3937 | /** |
3938 | * synchronize_rcu - wait until a grace period has elapsed. | |
3939 | * | |
3940 | * Control will return to the caller some time after a full grace | |
3941 | * period has elapsed, in other words after all currently executing RCU | |
3942 | * read-side critical sections have completed. Note, however, that | |
3943 | * upon return from synchronize_rcu(), the caller might well be executing | |
3944 | * concurrently with new RCU read-side critical sections that began while | |
1893afd6 PM |
3945 | * synchronize_rcu() was waiting. |
3946 | * | |
3947 | * RCU read-side critical sections are delimited by rcu_read_lock() | |
3948 | * and rcu_read_unlock(), and may be nested. In addition, but only in | |
3949 | * v5.0 and later, regions of code across which interrupts, preemption, | |
3950 | * or softirqs have been disabled also serve as RCU read-side critical | |
e5bc3af7 PM |
3951 | * sections. This includes hardware interrupt handlers, softirq handlers, |
3952 | * and NMI handlers. | |
3953 | * | |
3954 | * Note that this guarantee implies further memory-ordering guarantees. | |
3955 | * On systems with more than one CPU, when synchronize_rcu() returns, | |
3956 | * each CPU is guaranteed to have executed a full memory barrier since | |
3957 | * the end of its last RCU read-side critical section whose beginning | |
3958 | * preceded the call to synchronize_rcu(). In addition, each CPU having | |
3959 | * an RCU read-side critical section that extends beyond the return from | |
3960 | * synchronize_rcu() is guaranteed to have executed a full memory barrier | |
3961 | * after the beginning of synchronize_rcu() and before the beginning of | |
3962 | * that RCU read-side critical section. Note that these guarantees include | |
3963 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs | |
3964 | * that are executing in the kernel. | |
3965 | * | |
3966 | * Furthermore, if CPU A invoked synchronize_rcu(), which returned | |
3967 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
3968 | * to have executed a full memory barrier during the execution of | |
3969 | * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but | |
3970 | * again only if the system has more than one CPU). | |
3d3a0d1b PM |
3971 | * |
3972 | * Implementation of these memory-ordering guarantees is described here: | |
3973 | * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst. | |
e5bc3af7 PM |
3974 | */ |
3975 | void synchronize_rcu(void) | |
3976 | { | |
910e1209 PM |
3977 | unsigned long flags; |
3978 | struct rcu_node *rnp; | |
3979 | ||
e5bc3af7 PM |
3980 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
3981 | lock_is_held(&rcu_lock_map) || | |
3982 | lock_is_held(&rcu_sched_lock_map), | |
3983 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
910e1209 PM |
3984 | if (!rcu_blocking_is_gp()) { |
3985 | if (rcu_gp_is_expedited()) | |
3986 | synchronize_rcu_expedited(); | |
3987 | else | |
988f569a | 3988 | synchronize_rcu_normal(); |
910e1209 | 3989 | return; |
bf95b2bc | 3990 | } |
910e1209 PM |
3991 | |
3992 | // Context allows vacuous grace periods. | |
3993 | // Note well that this code runs with !PREEMPT && !SMP. | |
3994 | // In addition, all code that advances grace periods runs at | |
3995 | // process level. Therefore, this normal GP overlaps with other | |
3996 | // normal GPs only by being fully nested within them, which allows | |
3997 | // reuse of ->gp_seq_polled_snap. | |
3998 | rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_snap); | |
3999 | rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_snap); | |
4000 | ||
d761de8a PM |
4001 | // Update the normal grace-period counters to record |
4002 | // this grace period, but only those used by the boot CPU. | |
4003 | // The rcu_scheduler_starting() will take care of the rest of | |
4004 | // these counters. | |
910e1209 PM |
4005 | local_irq_save(flags); |
4006 | WARN_ON_ONCE(num_online_cpus() > 1); | |
4007 | rcu_state.gp_seq += (1 << RCU_SEQ_CTR_SHIFT); | |
d761de8a | 4008 | for (rnp = this_cpu_ptr(&rcu_data)->mynode; rnp; rnp = rnp->parent) |
910e1209 PM |
4009 | rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq; |
4010 | local_irq_restore(flags); | |
e5bc3af7 PM |
4011 | } |
4012 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
4013 | ||
91a967fd PM |
4014 | /** |
4015 | * get_completed_synchronize_rcu_full - Return a full pre-completed polled state cookie | |
4016 | * @rgosp: Place to put state cookie | |
4017 | * | |
4018 | * Stores into @rgosp a value that will always be treated by functions | |
4019 | * like poll_state_synchronize_rcu_full() as a cookie whose grace period | |
4020 | * has already completed. | |
4021 | */ | |
4022 | void get_completed_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp) | |
4023 | { | |
4024 | rgosp->rgos_norm = RCU_GET_STATE_COMPLETED; | |
4025 | rgosp->rgos_exp = RCU_GET_STATE_COMPLETED; | |
91a967fd PM |
4026 | } |
4027 | EXPORT_SYMBOL_GPL(get_completed_synchronize_rcu_full); | |
4028 | ||
765a3f4f PM |
4029 | /** |
4030 | * get_state_synchronize_rcu - Snapshot current RCU state | |
4031 | * | |
4032 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
7abb18bd PM |
4033 | * or poll_state_synchronize_rcu() to determine whether or not a full |
4034 | * grace period has elapsed in the meantime. | |
765a3f4f PM |
4035 | */ |
4036 | unsigned long get_state_synchronize_rcu(void) | |
4037 | { | |
4038 | /* | |
4039 | * Any prior manipulation of RCU-protected data must happen | |
e4be81a2 | 4040 | * before the load from ->gp_seq. |
765a3f4f PM |
4041 | */ |
4042 | smp_mb(); /* ^^^ */ | |
bf95b2bc | 4043 | return rcu_seq_snap(&rcu_state.gp_seq_polled); |
765a3f4f PM |
4044 | } |
4045 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); | |
4046 | ||
7abb18bd | 4047 | /** |
3fdefca9 PM |
4048 | * get_state_synchronize_rcu_full - Snapshot RCU state, both normal and expedited |
4049 | * @rgosp: location to place combined normal/expedited grace-period state | |
7abb18bd | 4050 | * |
3fdefca9 PM |
4051 | * Places the normal and expedited grace-period states in @rgosp. This |
4052 | * state value can be passed to a later call to cond_synchronize_rcu_full() | |
4053 | * or poll_state_synchronize_rcu_full() to determine whether or not a | |
4054 | * grace period (whether normal or expedited) has elapsed in the meantime. | |
4055 | * The rcu_gp_oldstate structure takes up twice the memory of an unsigned | |
4056 | * long, but is guaranteed to see all grace periods. In contrast, the | |
4057 | * combined state occupies less memory, but can sometimes fail to take | |
4058 | * grace periods into account. | |
7abb18bd | 4059 | * |
3fdefca9 PM |
4060 | * This does not guarantee that the needed grace period will actually |
4061 | * start. | |
7abb18bd | 4062 | */ |
3fdefca9 PM |
4063 | void get_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp) |
4064 | { | |
4065 | struct rcu_node *rnp = rcu_get_root(); | |
4066 | ||
4067 | /* | |
4068 | * Any prior manipulation of RCU-protected data must happen | |
4069 | * before the loads from ->gp_seq and ->expedited_sequence. | |
4070 | */ | |
4071 | smp_mb(); /* ^^^ */ | |
4072 | rgosp->rgos_norm = rcu_seq_snap(&rnp->gp_seq); | |
4073 | rgosp->rgos_exp = rcu_seq_snap(&rcu_state.expedited_sequence); | |
3fdefca9 PM |
4074 | } |
4075 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu_full); | |
4076 | ||
76ea3641 PM |
4077 | /* |
4078 | * Helper function for start_poll_synchronize_rcu() and | |
4079 | * start_poll_synchronize_rcu_full(). | |
7abb18bd | 4080 | */ |
76ea3641 | 4081 | static void start_poll_synchronize_rcu_common(void) |
7abb18bd PM |
4082 | { |
4083 | unsigned long flags; | |
7abb18bd PM |
4084 | bool needwake; |
4085 | struct rcu_data *rdp; | |
4086 | struct rcu_node *rnp; | |
4087 | ||
4088 | lockdep_assert_irqs_enabled(); | |
4089 | local_irq_save(flags); | |
4090 | rdp = this_cpu_ptr(&rcu_data); | |
4091 | rnp = rdp->mynode; | |
4092 | raw_spin_lock_rcu_node(rnp); // irqs already disabled. | |
bf95b2bc PM |
4093 | // Note it is possible for a grace period to have elapsed between |
4094 | // the above call to get_state_synchronize_rcu() and the below call | |
4095 | // to rcu_seq_snap. This is OK, the worst that happens is that we | |
4096 | // get a grace period that no one needed. These accesses are ordered | |
4097 | // by smp_mb(), and we are accessing them in the opposite order | |
4098 | // from which they are updated at grace-period start, as required. | |
4099 | needwake = rcu_start_this_gp(rnp, rdp, rcu_seq_snap(&rcu_state.gp_seq)); | |
7abb18bd PM |
4100 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
4101 | if (needwake) | |
4102 | rcu_gp_kthread_wake(); | |
76ea3641 PM |
4103 | } |
4104 | ||
4105 | /** | |
4106 | * start_poll_synchronize_rcu - Snapshot and start RCU grace period | |
4107 | * | |
4108 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() | |
4109 | * or poll_state_synchronize_rcu() to determine whether or not a full | |
4110 | * grace period has elapsed in the meantime. If the needed grace period | |
4111 | * is not already slated to start, notifies RCU core of the need for that | |
4112 | * grace period. | |
4113 | * | |
4114 | * Interrupts must be enabled for the case where it is necessary to awaken | |
4115 | * the grace-period kthread. | |
4116 | */ | |
4117 | unsigned long start_poll_synchronize_rcu(void) | |
4118 | { | |
4119 | unsigned long gp_seq = get_state_synchronize_rcu(); | |
4120 | ||
4121 | start_poll_synchronize_rcu_common(); | |
7abb18bd PM |
4122 | return gp_seq; |
4123 | } | |
4124 | EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu); | |
4125 | ||
4126 | /** | |
76ea3641 PM |
4127 | * start_poll_synchronize_rcu_full - Take a full snapshot and start RCU grace period |
4128 | * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full() | |
7abb18bd | 4129 | * |
76ea3641 PM |
4130 | * Places the normal and expedited grace-period states in *@rgos. This |
4131 | * state value can be passed to a later call to cond_synchronize_rcu_full() | |
4132 | * or poll_state_synchronize_rcu_full() to determine whether or not a | |
4133 | * grace period (whether normal or expedited) has elapsed in the meantime. | |
4134 | * If the needed grace period is not already slated to start, notifies | |
4135 | * RCU core of the need for that grace period. | |
4136 | * | |
4137 | * Interrupts must be enabled for the case where it is necessary to awaken | |
4138 | * the grace-period kthread. | |
4139 | */ | |
4140 | void start_poll_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp) | |
4141 | { | |
4142 | get_state_synchronize_rcu_full(rgosp); | |
4143 | ||
4144 | start_poll_synchronize_rcu_common(); | |
4145 | } | |
4146 | EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_full); | |
4147 | ||
7abb18bd | 4148 | /** |
91a967fd | 4149 | * poll_state_synchronize_rcu - Has the specified RCU grace period completed? |
3d3a0d1b | 4150 | * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu() |
7abb18bd PM |
4151 | * |
4152 | * If a full RCU grace period has elapsed since the earlier call from | |
f21e0143 | 4153 | * which @oldstate was obtained, return @true, otherwise return @false. |
a616aec9 | 4154 | * If @false is returned, it is the caller's responsibility to invoke this |
7abb18bd PM |
4155 | * function later on until it does return @true. Alternatively, the caller |
4156 | * can explicitly wait for a grace period, for example, by passing @oldstate | |
95ff24ee PM |
4157 | * to either cond_synchronize_rcu() or cond_synchronize_rcu_expedited() |
4158 | * on the one hand or by directly invoking either synchronize_rcu() or | |
4159 | * synchronize_rcu_expedited() on the other. | |
7abb18bd PM |
4160 | * |
4161 | * Yes, this function does not take counter wrap into account. | |
4162 | * But counter wrap is harmless. If the counter wraps, we have waited for | |
2403e804 | 4163 | * more than a billion grace periods (and way more on a 64-bit system!). |
f21e0143 | 4164 | * Those needing to keep old state values for very long time periods |
91a967fd PM |
4165 | * (many hours even on 32-bit systems) should check them occasionally and |
4166 | * either refresh them or set a flag indicating that the grace period has | |
4167 | * completed. Alternatively, they can use get_completed_synchronize_rcu() | |
4168 | * to get a guaranteed-completed grace-period state. | |
3d3a0d1b | 4169 | * |
95ff24ee PM |
4170 | * In addition, because oldstate compresses the grace-period state for |
4171 | * both normal and expedited grace periods into a single unsigned long, | |
4172 | * it can miss a grace period when synchronize_rcu() runs concurrently | |
4173 | * with synchronize_rcu_expedited(). If this is unacceptable, please | |
4174 | * instead use the _full() variant of these polling APIs. | |
4175 | * | |
3d3a0d1b PM |
4176 | * This function provides the same memory-ordering guarantees that |
4177 | * would be provided by a synchronize_rcu() that was invoked at the call | |
4178 | * to the function that provided @oldstate, and that returned at the end | |
4179 | * of this function. | |
7abb18bd PM |
4180 | */ |
4181 | bool poll_state_synchronize_rcu(unsigned long oldstate) | |
4182 | { | |
414c1238 | 4183 | if (oldstate == RCU_GET_STATE_COMPLETED || |
bf95b2bc | 4184 | rcu_seq_done_exact(&rcu_state.gp_seq_polled, oldstate)) { |
7abb18bd PM |
4185 | smp_mb(); /* Ensure GP ends before subsequent accesses. */ |
4186 | return true; | |
4187 | } | |
4188 | return false; | |
4189 | } | |
4190 | EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu); | |
4191 | ||
765a3f4f | 4192 | /** |
91a967fd PM |
4193 | * poll_state_synchronize_rcu_full - Has the specified RCU grace period completed? |
4194 | * @rgosp: value from get_state_synchronize_rcu_full() or start_poll_synchronize_rcu_full() | |
765a3f4f | 4195 | * |
91a967fd PM |
4196 | * If a full RCU grace period has elapsed since the earlier call from |
4197 | * which *rgosp was obtained, return @true, otherwise return @false. | |
4198 | * If @false is returned, it is the caller's responsibility to invoke this | |
4199 | * function later on until it does return @true. Alternatively, the caller | |
4200 | * can explicitly wait for a grace period, for example, by passing @rgosp | |
4201 | * to cond_synchronize_rcu() or by directly invoking synchronize_rcu(). | |
4202 | * | |
4203 | * Yes, this function does not take counter wrap into account. | |
4204 | * But counter wrap is harmless. If the counter wraps, we have waited | |
4205 | * for more than a billion grace periods (and way more on a 64-bit | |
4206 | * system!). Those needing to keep rcu_gp_oldstate values for very | |
4207 | * long time periods (many hours even on 32-bit systems) should check | |
4208 | * them occasionally and either refresh them or set a flag indicating | |
4209 | * that the grace period has completed. Alternatively, they can use | |
4210 | * get_completed_synchronize_rcu_full() to get a guaranteed-completed | |
4211 | * grace-period state. | |
765a3f4f | 4212 | * |
91a967fd PM |
4213 | * This function provides the same memory-ordering guarantees that would |
4214 | * be provided by a synchronize_rcu() that was invoked at the call to | |
4215 | * the function that provided @rgosp, and that returned at the end of this | |
4216 | * function. And this guarantee requires that the root rcu_node structure's | |
4217 | * ->gp_seq field be checked instead of that of the rcu_state structure. | |
4218 | * The problem is that the just-ending grace-period's callbacks can be | |
4219 | * invoked between the time that the root rcu_node structure's ->gp_seq | |
4220 | * field is updated and the time that the rcu_state structure's ->gp_seq | |
4221 | * field is updated. Therefore, if a single synchronize_rcu() is to | |
4222 | * cause a subsequent poll_state_synchronize_rcu_full() to return @true, | |
4223 | * then the root rcu_node structure is the one that needs to be polled. | |
4224 | */ | |
4225 | bool poll_state_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp) | |
4226 | { | |
4227 | struct rcu_node *rnp = rcu_get_root(); | |
4228 | ||
4229 | smp_mb(); // Order against root rcu_node structure grace-period cleanup. | |
4230 | if (rgosp->rgos_norm == RCU_GET_STATE_COMPLETED || | |
4231 | rcu_seq_done_exact(&rnp->gp_seq, rgosp->rgos_norm) || | |
4232 | rgosp->rgos_exp == RCU_GET_STATE_COMPLETED || | |
7ecef087 | 4233 | rcu_seq_done_exact(&rcu_state.expedited_sequence, rgosp->rgos_exp)) { |
91a967fd PM |
4234 | smp_mb(); /* Ensure GP ends before subsequent accesses. */ |
4235 | return true; | |
4236 | } | |
4237 | return false; | |
4238 | } | |
4239 | EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu_full); | |
4240 | ||
765a3f4f PM |
4241 | /** |
4242 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period | |
d96c52fe | 4243 | * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited() |
765a3f4f PM |
4244 | * |
4245 | * If a full RCU grace period has elapsed since the earlier call to | |
7abb18bd PM |
4246 | * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return. |
4247 | * Otherwise, invoke synchronize_rcu() to wait for a full grace period. | |
765a3f4f | 4248 | * |
d96c52fe PM |
4249 | * Yes, this function does not take counter wrap into account. |
4250 | * But counter wrap is harmless. If the counter wraps, we have waited for | |
765a3f4f | 4251 | * more than 2 billion grace periods (and way more on a 64-bit system!), |
d96c52fe | 4252 | * so waiting for a couple of additional grace periods should be just fine. |
3d3a0d1b PM |
4253 | * |
4254 | * This function provides the same memory-ordering guarantees that | |
4255 | * would be provided by a synchronize_rcu() that was invoked at the call | |
d96c52fe | 4256 | * to the function that provided @oldstate and that returned at the end |
3d3a0d1b | 4257 | * of this function. |
765a3f4f PM |
4258 | */ |
4259 | void cond_synchronize_rcu(unsigned long oldstate) | |
4260 | { | |
7abb18bd | 4261 | if (!poll_state_synchronize_rcu(oldstate)) |
765a3f4f PM |
4262 | synchronize_rcu(); |
4263 | } | |
4264 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); | |
4265 | ||
b6fe4917 PM |
4266 | /** |
4267 | * cond_synchronize_rcu_full - Conditionally wait for an RCU grace period | |
4268 | * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full() | |
4269 | * | |
4270 | * If a full RCU grace period has elapsed since the call to | |
4271 | * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), | |
4272 | * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was | |
4273 | * obtained, just return. Otherwise, invoke synchronize_rcu() to wait | |
4274 | * for a full grace period. | |
4275 | * | |
4276 | * Yes, this function does not take counter wrap into account. | |
4277 | * But counter wrap is harmless. If the counter wraps, we have waited for | |
4278 | * more than 2 billion grace periods (and way more on a 64-bit system!), | |
4279 | * so waiting for a couple of additional grace periods should be just fine. | |
4280 | * | |
4281 | * This function provides the same memory-ordering guarantees that | |
4282 | * would be provided by a synchronize_rcu() that was invoked at the call | |
4283 | * to the function that provided @rgosp and that returned at the end of | |
4284 | * this function. | |
4285 | */ | |
4286 | void cond_synchronize_rcu_full(struct rcu_gp_oldstate *rgosp) | |
4287 | { | |
4288 | if (!poll_state_synchronize_rcu_full(rgosp)) | |
4289 | synchronize_rcu(); | |
4290 | } | |
4291 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu_full); | |
4292 | ||
64db4cff | 4293 | /* |
98ece508 | 4294 | * Check to see if there is any immediate RCU-related work to be done by |
49918a54 PM |
4295 | * the current CPU, returning 1 if so and zero otherwise. The checks are |
4296 | * in order of increasing expense: checks that can be carried out against | |
4297 | * CPU-local state are performed first. However, we must check for CPU | |
4298 | * stalls first, else we might not get a chance. | |
64db4cff | 4299 | */ |
dd7dafd1 | 4300 | static int rcu_pending(int user) |
64db4cff | 4301 | { |
ed93dfc6 | 4302 | bool gp_in_progress; |
98ece508 | 4303 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
2f51f988 PM |
4304 | struct rcu_node *rnp = rdp->mynode; |
4305 | ||
a649d25d PM |
4306 | lockdep_assert_irqs_disabled(); |
4307 | ||
64db4cff | 4308 | /* Check for CPU stalls, if enabled. */ |
ea12ff2b | 4309 | check_cpu_stall(rdp); |
64db4cff | 4310 | |
85f69b32 | 4311 | /* Does this CPU need a deferred NOCB wakeup? */ |
87090516 | 4312 | if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE)) |
85f69b32 PM |
4313 | return 1; |
4314 | ||
dd7dafd1 PM |
4315 | /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */ |
4316 | if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu()) | |
a096932f PM |
4317 | return 0; |
4318 | ||
64db4cff | 4319 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
ed93dfc6 PM |
4320 | gp_in_progress = rcu_gp_in_progress(); |
4321 | if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress) | |
64db4cff PM |
4322 | return 1; |
4323 | ||
4324 | /* Does this CPU have callbacks ready to invoke? */ | |
3820b513 | 4325 | if (!rcu_rdp_is_offloaded(rdp) && |
bd56e0a4 | 4326 | rcu_segcblist_ready_cbs(&rdp->cblist)) |
64db4cff PM |
4327 | return 1; |
4328 | ||
4329 | /* Has RCU gone idle with this CPU needing another grace period? */ | |
ed93dfc6 | 4330 | if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) && |
3820b513 | 4331 | !rcu_rdp_is_offloaded(rdp) && |
c1935209 | 4332 | !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL)) |
64db4cff PM |
4333 | return 1; |
4334 | ||
67e14c1e PM |
4335 | /* Have RCU grace period completed or started? */ |
4336 | if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq || | |
01c495f7 | 4337 | unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */ |
64db4cff PM |
4338 | return 1; |
4339 | ||
64db4cff PM |
4340 | /* nothing to do */ |
4341 | return 0; | |
4342 | } | |
4343 | ||
a83eff0a | 4344 | /* |
dd46a788 | 4345 | * Helper function for rcu_barrier() tracing. If tracing is disabled, |
a83eff0a PM |
4346 | * the compiler is expected to optimize this away. |
4347 | */ | |
dd46a788 | 4348 | static void rcu_barrier_trace(const char *s, int cpu, unsigned long done) |
a83eff0a | 4349 | { |
8344b871 PM |
4350 | trace_rcu_barrier(rcu_state.name, s, cpu, |
4351 | atomic_read(&rcu_state.barrier_cpu_count), done); | |
a83eff0a PM |
4352 | } |
4353 | ||
b1420f1c | 4354 | /* |
dd46a788 PM |
4355 | * RCU callback function for rcu_barrier(). If we are last, wake |
4356 | * up the task executing rcu_barrier(). | |
aa24f937 PM |
4357 | * |
4358 | * Note that the value of rcu_state.barrier_sequence must be captured | |
4359 | * before the atomic_dec_and_test(). Otherwise, if this CPU is not last, | |
4360 | * other CPUs might count the value down to zero before this CPU gets | |
4361 | * around to invoking rcu_barrier_trace(), which might result in bogus | |
4362 | * data from the next instance of rcu_barrier(). | |
b1420f1c | 4363 | */ |
24ebbca8 | 4364 | static void rcu_barrier_callback(struct rcu_head *rhp) |
d0ec774c | 4365 | { |
aa24f937 PM |
4366 | unsigned long __maybe_unused s = rcu_state.barrier_sequence; |
4367 | ||
ec9f5835 | 4368 | if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) { |
aa24f937 | 4369 | rcu_barrier_trace(TPS("LastCB"), -1, s); |
ec9f5835 | 4370 | complete(&rcu_state.barrier_completion); |
a83eff0a | 4371 | } else { |
aa24f937 | 4372 | rcu_barrier_trace(TPS("CB"), -1, s); |
a83eff0a | 4373 | } |
d0ec774c PM |
4374 | } |
4375 | ||
4376 | /* | |
a16578dd | 4377 | * If needed, entrain an rcu_barrier() callback on rdp->cblist. |
d0ec774c | 4378 | */ |
a16578dd | 4379 | static void rcu_barrier_entrain(struct rcu_data *rdp) |
d0ec774c | 4380 | { |
a16578dd PM |
4381 | unsigned long gseq = READ_ONCE(rcu_state.barrier_sequence); |
4382 | unsigned long lseq = READ_ONCE(rdp->barrier_seq_snap); | |
b8f7aca3 FW |
4383 | bool wake_nocb = false; |
4384 | bool was_alldone = false; | |
d0ec774c | 4385 | |
80b3fd47 | 4386 | lockdep_assert_held(&rcu_state.barrier_lock); |
a16578dd PM |
4387 | if (rcu_seq_state(lseq) || !rcu_seq_state(gseq) || rcu_seq_ctr(lseq) != rcu_seq_ctr(gseq)) |
4388 | return; | |
dd46a788 | 4389 | rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence); |
f92c734f PM |
4390 | rdp->barrier_head.func = rcu_barrier_callback; |
4391 | debug_rcu_head_queue(&rdp->barrier_head); | |
5d6742b3 | 4392 | rcu_nocb_lock(rdp); |
b8f7aca3 FW |
4393 | /* |
4394 | * Flush bypass and wakeup rcuog if we add callbacks to an empty regular | |
4395 | * queue. This way we don't wait for bypass timer that can reach seconds | |
4396 | * if it's fully lazy. | |
4397 | */ | |
4398 | was_alldone = rcu_rdp_is_offloaded(rdp) && !rcu_segcblist_pend_cbs(&rdp->cblist); | |
3cb278e7 | 4399 | WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false)); |
b8f7aca3 | 4400 | wake_nocb = was_alldone && rcu_segcblist_pend_cbs(&rdp->cblist); |
77a40f97 | 4401 | if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) { |
ec9f5835 | 4402 | atomic_inc(&rcu_state.barrier_cpu_count); |
f92c734f PM |
4403 | } else { |
4404 | debug_rcu_head_unqueue(&rdp->barrier_head); | |
a16578dd | 4405 | rcu_barrier_trace(TPS("IRQNQ"), -1, rcu_state.barrier_sequence); |
f92c734f | 4406 | } |
5d6742b3 | 4407 | rcu_nocb_unlock(rdp); |
b8f7aca3 FW |
4408 | if (wake_nocb) |
4409 | wake_nocb_gp(rdp, false); | |
a16578dd PM |
4410 | smp_store_release(&rdp->barrier_seq_snap, gseq); |
4411 | } | |
4412 | ||
4413 | /* | |
4414 | * Called with preemption disabled, and from cross-cpu IRQ context. | |
4415 | */ | |
4416 | static void rcu_barrier_handler(void *cpu_in) | |
4417 | { | |
4418 | uintptr_t cpu = (uintptr_t)cpu_in; | |
4419 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
4420 | ||
4421 | lockdep_assert_irqs_disabled(); | |
4422 | WARN_ON_ONCE(cpu != rdp->cpu); | |
4423 | WARN_ON_ONCE(cpu != smp_processor_id()); | |
80b3fd47 | 4424 | raw_spin_lock(&rcu_state.barrier_lock); |
a16578dd | 4425 | rcu_barrier_entrain(rdp); |
80b3fd47 | 4426 | raw_spin_unlock(&rcu_state.barrier_lock); |
d0ec774c PM |
4427 | } |
4428 | ||
dd46a788 PM |
4429 | /** |
4430 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
4431 | * | |
4432 | * Note that this primitive does not necessarily wait for an RCU grace period | |
4433 | * to complete. For example, if there are no RCU callbacks queued anywhere | |
4434 | * in the system, then rcu_barrier() is within its rights to return | |
4435 | * immediately, without waiting for anything, much less an RCU grace period. | |
d0ec774c | 4436 | */ |
dd46a788 | 4437 | void rcu_barrier(void) |
d0ec774c | 4438 | { |
127e2981 | 4439 | uintptr_t cpu; |
a16578dd PM |
4440 | unsigned long flags; |
4441 | unsigned long gseq; | |
b1420f1c | 4442 | struct rcu_data *rdp; |
ec9f5835 | 4443 | unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence); |
b1420f1c | 4444 | |
dd46a788 | 4445 | rcu_barrier_trace(TPS("Begin"), -1, s); |
b1420f1c | 4446 | |
e74f4c45 | 4447 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
ec9f5835 | 4448 | mutex_lock(&rcu_state.barrier_mutex); |
b1420f1c | 4449 | |
4f525a52 | 4450 | /* Did someone else do our work for us? */ |
ec9f5835 | 4451 | if (rcu_seq_done(&rcu_state.barrier_sequence, s)) { |
0cabb47a | 4452 | rcu_barrier_trace(TPS("EarlyExit"), -1, rcu_state.barrier_sequence); |
cf3a9c48 | 4453 | smp_mb(); /* caller's subsequent code after above check. */ |
ec9f5835 | 4454 | mutex_unlock(&rcu_state.barrier_mutex); |
cf3a9c48 PM |
4455 | return; |
4456 | } | |
4457 | ||
4f525a52 | 4458 | /* Mark the start of the barrier operation. */ |
80b3fd47 | 4459 | raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags); |
ec9f5835 | 4460 | rcu_seq_start(&rcu_state.barrier_sequence); |
a16578dd | 4461 | gseq = rcu_state.barrier_sequence; |
dd46a788 | 4462 | rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence); |
b1420f1c | 4463 | |
d0ec774c | 4464 | /* |
127e2981 PM |
4465 | * Initialize the count to two rather than to zero in order |
4466 | * to avoid a too-soon return to zero in case of an immediate | |
4467 | * invocation of the just-enqueued callback (or preemption of | |
4468 | * this task). Exclude CPU-hotplug operations to ensure that no | |
4469 | * offline non-offloaded CPU has callbacks queued. | |
d0ec774c | 4470 | */ |
ec9f5835 | 4471 | init_completion(&rcu_state.barrier_completion); |
127e2981 | 4472 | atomic_set(&rcu_state.barrier_cpu_count, 2); |
80b3fd47 | 4473 | raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags); |
b1420f1c PM |
4474 | |
4475 | /* | |
1331e7a1 PM |
4476 | * Force each CPU with callbacks to register a new callback. |
4477 | * When that callback is invoked, we will know that all of the | |
4478 | * corresponding CPU's preceding callbacks have been invoked. | |
b1420f1c | 4479 | */ |
3fbfbf7a | 4480 | for_each_possible_cpu(cpu) { |
da1df50d | 4481 | rdp = per_cpu_ptr(&rcu_data, cpu); |
a16578dd PM |
4482 | retry: |
4483 | if (smp_load_acquire(&rdp->barrier_seq_snap) == gseq) | |
ce5215c1 | 4484 | continue; |
80b3fd47 | 4485 | raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags); |
0cabb47a | 4486 | if (!rcu_segcblist_n_cbs(&rdp->cblist)) { |
a16578dd | 4487 | WRITE_ONCE(rdp->barrier_seq_snap, gseq); |
80b3fd47 | 4488 | raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags); |
0cabb47a | 4489 | rcu_barrier_trace(TPS("NQ"), cpu, rcu_state.barrier_sequence); |
ce5215c1 | 4490 | continue; |
0cabb47a | 4491 | } |
a16578dd PM |
4492 | if (!rcu_rdp_cpu_online(rdp)) { |
4493 | rcu_barrier_entrain(rdp); | |
4494 | WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq); | |
80b3fd47 | 4495 | raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags); |
0cabb47a | 4496 | rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu, rcu_state.barrier_sequence); |
a16578dd | 4497 | continue; |
b1420f1c | 4498 | } |
80b3fd47 | 4499 | raw_spin_unlock_irqrestore(&rcu_state.barrier_lock, flags); |
a16578dd PM |
4500 | if (smp_call_function_single(cpu, rcu_barrier_handler, (void *)cpu, 1)) { |
4501 | schedule_timeout_uninterruptible(1); | |
4502 | goto retry; | |
b1420f1c | 4503 | } |
a16578dd PM |
4504 | WARN_ON_ONCE(READ_ONCE(rdp->barrier_seq_snap) != gseq); |
4505 | rcu_barrier_trace(TPS("OnlineQ"), cpu, rcu_state.barrier_sequence); | |
b1420f1c | 4506 | } |
b1420f1c PM |
4507 | |
4508 | /* | |
4509 | * Now that we have an rcu_barrier_callback() callback on each | |
4510 | * CPU, and thus each counted, remove the initial count. | |
4511 | */ | |
127e2981 | 4512 | if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count)) |
ec9f5835 | 4513 | complete(&rcu_state.barrier_completion); |
b1420f1c PM |
4514 | |
4515 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ | |
ec9f5835 | 4516 | wait_for_completion(&rcu_state.barrier_completion); |
b1420f1c | 4517 | |
4f525a52 | 4518 | /* Mark the end of the barrier operation. */ |
dd46a788 | 4519 | rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence); |
ec9f5835 | 4520 | rcu_seq_end(&rcu_state.barrier_sequence); |
a16578dd PM |
4521 | gseq = rcu_state.barrier_sequence; |
4522 | for_each_possible_cpu(cpu) { | |
4523 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
4524 | ||
4525 | WRITE_ONCE(rdp->barrier_seq_snap, gseq); | |
4526 | } | |
4f525a52 | 4527 | |
b1420f1c | 4528 | /* Other rcu_barrier() invocations can now safely proceed. */ |
ec9f5835 | 4529 | mutex_unlock(&rcu_state.barrier_mutex); |
d0ec774c | 4530 | } |
45975c7d | 4531 | EXPORT_SYMBOL_GPL(rcu_barrier); |
d0ec774c | 4532 | |
16128b1f PM |
4533 | static unsigned long rcu_barrier_last_throttle; |
4534 | ||
4535 | /** | |
4536 | * rcu_barrier_throttled - Do rcu_barrier(), but limit to one per second | |
4537 | * | |
4538 | * This can be thought of as guard rails around rcu_barrier() that | |
4539 | * permits unrestricted userspace use, at least assuming the hardware's | |
4540 | * try_cmpxchg() is robust. There will be at most one call per second to | |
4541 | * rcu_barrier() system-wide from use of this function, which means that | |
4542 | * callers might needlessly wait a second or three. | |
4543 | * | |
4544 | * This is intended for use by test suites to avoid OOM by flushing RCU | |
4545 | * callbacks from the previous test before starting the next. See the | |
4546 | * rcutree.do_rcu_barrier module parameter for more information. | |
4547 | * | |
4548 | * Why not simply make rcu_barrier() more scalable? That might be | |
4549 | * the eventual endpoint, but let's keep it simple for the time being. | |
4550 | * Note that the module parameter infrastructure serializes calls to a | |
4551 | * given .set() function, but should concurrent .set() invocation ever be | |
4552 | * possible, we are ready! | |
4553 | */ | |
4554 | static void rcu_barrier_throttled(void) | |
4555 | { | |
4556 | unsigned long j = jiffies; | |
4557 | unsigned long old = READ_ONCE(rcu_barrier_last_throttle); | |
4558 | unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence); | |
4559 | ||
4560 | while (time_in_range(j, old, old + HZ / 16) || | |
4561 | !try_cmpxchg(&rcu_barrier_last_throttle, &old, j)) { | |
4562 | schedule_timeout_idle(HZ / 16); | |
4563 | if (rcu_seq_done(&rcu_state.barrier_sequence, s)) { | |
4564 | smp_mb(); /* caller's subsequent code after above check. */ | |
4565 | return; | |
4566 | } | |
4567 | j = jiffies; | |
4568 | old = READ_ONCE(rcu_barrier_last_throttle); | |
4569 | } | |
4570 | rcu_barrier(); | |
4571 | } | |
4572 | ||
4573 | /* | |
4574 | * Invoke rcu_barrier_throttled() when a rcutree.do_rcu_barrier | |
4575 | * request arrives. We insist on a true value to allow for possible | |
4576 | * future expansion. | |
4577 | */ | |
4578 | static int param_set_do_rcu_barrier(const char *val, const struct kernel_param *kp) | |
4579 | { | |
4580 | bool b; | |
4581 | int ret; | |
4582 | ||
4583 | if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) | |
4584 | return -EAGAIN; | |
4585 | ret = kstrtobool(val, &b); | |
4586 | if (!ret && b) { | |
4587 | atomic_inc((atomic_t *)kp->arg); | |
4588 | rcu_barrier_throttled(); | |
4589 | atomic_dec((atomic_t *)kp->arg); | |
4590 | } | |
4591 | return ret; | |
4592 | } | |
4593 | ||
4594 | /* | |
4595 | * Output the number of outstanding rcutree.do_rcu_barrier requests. | |
4596 | */ | |
4597 | static int param_get_do_rcu_barrier(char *buffer, const struct kernel_param *kp) | |
4598 | { | |
4599 | return sprintf(buffer, "%d\n", atomic_read((atomic_t *)kp->arg)); | |
4600 | } | |
4601 | ||
4602 | static const struct kernel_param_ops do_rcu_barrier_ops = { | |
4603 | .set = param_set_do_rcu_barrier, | |
4604 | .get = param_get_do_rcu_barrier, | |
4605 | }; | |
4606 | static atomic_t do_rcu_barrier; | |
4607 | module_param_cb(do_rcu_barrier, &do_rcu_barrier_ops, &do_rcu_barrier, 0644); | |
4608 | ||
5a04848d PM |
4609 | /* |
4610 | * Compute the mask of online CPUs for the specified rcu_node structure. | |
4611 | * This will not be stable unless the rcu_node structure's ->lock is | |
4612 | * held, but the bit corresponding to the current CPU will be stable | |
4613 | * in most contexts. | |
4614 | */ | |
4615 | static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) | |
4616 | { | |
4617 | return READ_ONCE(rnp->qsmaskinitnext); | |
4618 | } | |
4619 | ||
4620 | /* | |
4621 | * Is the CPU corresponding to the specified rcu_data structure online | |
4622 | * from RCU's perspective? This perspective is given by that structure's | |
4623 | * ->qsmaskinitnext field rather than by the global cpu_online_mask. | |
4624 | */ | |
4625 | static bool rcu_rdp_cpu_online(struct rcu_data *rdp) | |
4626 | { | |
4627 | return !!(rdp->grpmask & rcu_rnp_online_cpus(rdp->mynode)); | |
4628 | } | |
4629 | ||
2be4686d FW |
4630 | bool rcu_cpu_online(int cpu) |
4631 | { | |
4632 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
4633 | ||
4634 | return rcu_rdp_cpu_online(rdp); | |
4635 | } | |
4636 | ||
5a04848d PM |
4637 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
4638 | ||
4639 | /* | |
4640 | * Is the current CPU online as far as RCU is concerned? | |
4641 | * | |
4642 | * Disable preemption to avoid false positives that could otherwise | |
4643 | * happen due to the current CPU number being sampled, this task being | |
4644 | * preempted, its old CPU being taken offline, resuming on some other CPU, | |
4645 | * then determining that its old CPU is now offline. | |
4646 | * | |
4647 | * Disable checking if in an NMI handler because we cannot safely | |
4648 | * report errors from NMI handlers anyway. In addition, it is OK to use | |
4649 | * RCU on an offline processor during initial boot, hence the check for | |
4650 | * rcu_scheduler_fully_active. | |
4651 | */ | |
4652 | bool rcu_lockdep_current_cpu_online(void) | |
4653 | { | |
4654 | struct rcu_data *rdp; | |
4655 | bool ret = false; | |
4656 | ||
4657 | if (in_nmi() || !rcu_scheduler_fully_active) | |
4658 | return true; | |
4659 | preempt_disable_notrace(); | |
4660 | rdp = this_cpu_ptr(&rcu_data); | |
4661 | /* | |
4662 | * Strictly, we care here about the case where the current CPU is | |
448e9f34 | 4663 | * in rcutree_report_cpu_starting() and thus has an excuse for rdp->grpmask |
5a04848d PM |
4664 | * not being up to date. So arch_spin_is_locked() might have a |
4665 | * false positive if it's held by some *other* CPU, but that's | |
4666 | * OK because that just means a false *negative* on the warning. | |
4667 | */ | |
4668 | if (rcu_rdp_cpu_online(rdp) || arch_spin_is_locked(&rcu_state.ofl_lock)) | |
4669 | ret = true; | |
4670 | preempt_enable_notrace(); | |
4671 | return ret; | |
4672 | } | |
4673 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); | |
4674 | ||
4675 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ | |
4676 | ||
4677 | // Has rcu_init() been invoked? This is used (for example) to determine | |
4678 | // whether spinlocks may be acquired safely. | |
4679 | static bool rcu_init_invoked(void) | |
4680 | { | |
09e077cf | 4681 | return !!READ_ONCE(rcu_state.n_online_cpus); |
5a04848d PM |
4682 | } |
4683 | ||
5a04848d PM |
4684 | /* |
4685 | * All CPUs for the specified rcu_node structure have gone offline, | |
4686 | * and all tasks that were preempted within an RCU read-side critical | |
4687 | * section while running on one of those CPUs have since exited their RCU | |
4688 | * read-side critical section. Some other CPU is reporting this fact with | |
4689 | * the specified rcu_node structure's ->lock held and interrupts disabled. | |
4690 | * This function therefore goes up the tree of rcu_node structures, | |
4691 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that | |
4692 | * the leaf rcu_node structure's ->qsmaskinit field has already been | |
4693 | * updated. | |
4694 | * | |
4695 | * This function does check that the specified rcu_node structure has | |
4696 | * all CPUs offline and no blocked tasks, so it is OK to invoke it | |
4697 | * prematurely. That said, invoking it after the fact will cost you | |
4698 | * a needless lock acquisition. So once it has done its work, don't | |
4699 | * invoke it again. | |
4700 | */ | |
4701 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) | |
4702 | { | |
4703 | long mask; | |
4704 | struct rcu_node *rnp = rnp_leaf; | |
4705 | ||
4706 | raw_lockdep_assert_held_rcu_node(rnp_leaf); | |
4707 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || | |
4708 | WARN_ON_ONCE(rnp_leaf->qsmaskinit) || | |
4709 | WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf))) | |
4710 | return; | |
4711 | for (;;) { | |
4712 | mask = rnp->grpmask; | |
4713 | rnp = rnp->parent; | |
4714 | if (!rnp) | |
4715 | break; | |
4716 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ | |
4717 | rnp->qsmaskinit &= ~mask; | |
4718 | /* Between grace periods, so better already be zero! */ | |
4719 | WARN_ON_ONCE(rnp->qsmask); | |
4720 | if (rnp->qsmaskinit) { | |
4721 | raw_spin_unlock_rcu_node(rnp); | |
4722 | /* irqs remain disabled. */ | |
4723 | return; | |
4724 | } | |
4725 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ | |
4726 | } | |
4727 | } | |
4728 | ||
0aa04b05 PM |
4729 | /* |
4730 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the | |
4731 | * first CPU in a given leaf rcu_node structure coming online. The caller | |
a616aec9 | 4732 | * must hold the corresponding leaf rcu_node ->lock with interrupts |
0aa04b05 PM |
4733 | * disabled. |
4734 | */ | |
4735 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) | |
4736 | { | |
4737 | long mask; | |
8d672fa6 | 4738 | long oldmask; |
0aa04b05 PM |
4739 | struct rcu_node *rnp = rnp_leaf; |
4740 | ||
8d672fa6 | 4741 | raw_lockdep_assert_held_rcu_node(rnp_leaf); |
962aff03 | 4742 | WARN_ON_ONCE(rnp->wait_blkd_tasks); |
0aa04b05 PM |
4743 | for (;;) { |
4744 | mask = rnp->grpmask; | |
4745 | rnp = rnp->parent; | |
4746 | if (rnp == NULL) | |
4747 | return; | |
6cf10081 | 4748 | raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */ |
8d672fa6 | 4749 | oldmask = rnp->qsmaskinit; |
0aa04b05 | 4750 | rnp->qsmaskinit |= mask; |
67c583a7 | 4751 | raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */ |
8d672fa6 PM |
4752 | if (oldmask) |
4753 | return; | |
0aa04b05 PM |
4754 | } |
4755 | } | |
4756 | ||
64db4cff | 4757 | /* |
27569620 | 4758 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
64db4cff | 4759 | */ |
27569620 | 4760 | static void __init |
53b46303 | 4761 | rcu_boot_init_percpu_data(int cpu) |
64db4cff | 4762 | { |
904e600e | 4763 | struct context_tracking *ct = this_cpu_ptr(&context_tracking); |
da1df50d | 4764 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
27569620 PM |
4765 | |
4766 | /* Set up local state, ensuring consistent view of global state. */ | |
bc75e999 | 4767 | rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu); |
a657f261 | 4768 | INIT_WORK(&rdp->strict_work, strict_work_handler); |
904e600e | 4769 | WARN_ON_ONCE(ct->dynticks_nesting != 1); |
62e2412d | 4770 | WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(cpu))); |
a16578dd | 4771 | rdp->barrier_seq_snap = rcu_state.barrier_sequence; |
53b46303 | 4772 | rdp->rcu_ofl_gp_seq = rcu_state.gp_seq; |
ae2b217a | 4773 | rdp->rcu_ofl_gp_state = RCU_GP_CLEANED; |
53b46303 | 4774 | rdp->rcu_onl_gp_seq = rcu_state.gp_seq; |
ae2b217a | 4775 | rdp->rcu_onl_gp_state = RCU_GP_CLEANED; |
c708b08c | 4776 | rdp->last_sched_clock = jiffies; |
27569620 | 4777 | rdp->cpu = cpu; |
3fbfbf7a | 4778 | rcu_boot_init_nocb_percpu_data(rdp); |
27569620 PM |
4779 | } |
4780 | ||
c19e5d3b | 4781 | struct kthread_worker *rcu_exp_gp_kworker; |
c19e5d3b | 4782 | |
8e5e6215 | 4783 | static void rcu_spawn_exp_par_gp_kworker(struct rcu_node *rnp) |
c19e5d3b | 4784 | { |
8e5e6215 FW |
4785 | struct kthread_worker *kworker; |
4786 | const char *name = "rcu_exp_par_gp_kthread_worker/%d"; | |
c19e5d3b | 4787 | struct sched_param param = { .sched_priority = kthread_prio }; |
8e5e6215 | 4788 | int rnp_index = rnp - rcu_get_root(); |
c19e5d3b | 4789 | |
8e5e6215 FW |
4790 | if (rnp->exp_kworker) |
4791 | return; | |
4792 | ||
4793 | kworker = kthread_create_worker(0, name, rnp_index); | |
4794 | if (IS_ERR_OR_NULL(kworker)) { | |
4795 | pr_err("Failed to create par gp kworker on %d/%d\n", | |
4796 | rnp->grplo, rnp->grphi); | |
c19e5d3b FW |
4797 | return; |
4798 | } | |
8e5e6215 | 4799 | WRITE_ONCE(rnp->exp_kworker, kworker); |
23da2ad6 FW |
4800 | |
4801 | if (IS_ENABLED(CONFIG_RCU_EXP_KTHREAD)) | |
4802 | sched_setscheduler_nocheck(kworker->task, SCHED_FIFO, ¶m); | |
8e5e6215 | 4803 | } |
c19e5d3b | 4804 | |
b67cffcb FW |
4805 | static struct task_struct *rcu_exp_par_gp_task(struct rcu_node *rnp) |
4806 | { | |
4807 | struct kthread_worker *kworker = READ_ONCE(rnp->exp_kworker); | |
4808 | ||
4809 | if (!kworker) | |
4810 | return NULL; | |
4811 | ||
4812 | return kworker->task; | |
4813 | } | |
4814 | ||
8e5e6215 FW |
4815 | static void __init rcu_start_exp_gp_kworker(void) |
4816 | { | |
4817 | const char *name = "rcu_exp_gp_kthread_worker"; | |
4818 | struct sched_param param = { .sched_priority = kthread_prio }; | |
4819 | ||
4820 | rcu_exp_gp_kworker = kthread_create_worker(0, name); | |
4821 | if (IS_ERR_OR_NULL(rcu_exp_gp_kworker)) { | |
4822 | pr_err("Failed to create %s!\n", name); | |
c19e5d3b FW |
4823 | rcu_exp_gp_kworker = NULL; |
4824 | return; | |
4825 | } | |
c19e5d3b | 4826 | |
23da2ad6 FW |
4827 | if (IS_ENABLED(CONFIG_RCU_EXP_KTHREAD)) |
4828 | sched_setscheduler_nocheck(rcu_exp_gp_kworker->task, SCHED_FIFO, ¶m); | |
c19e5d3b | 4829 | } |
c19e5d3b | 4830 | |
8e5e6215 FW |
4831 | static void rcu_spawn_rnp_kthreads(struct rcu_node *rnp) |
4832 | { | |
23da2ad6 | 4833 | if (rcu_scheduler_fully_active) { |
8e5e6215 FW |
4834 | mutex_lock(&rnp->kthread_mutex); |
4835 | rcu_spawn_one_boost_kthread(rnp); | |
4836 | rcu_spawn_exp_par_gp_kworker(rnp); | |
4837 | mutex_unlock(&rnp->kthread_mutex); | |
4838 | } | |
4839 | } | |
4840 | ||
27569620 | 4841 | /* |
53b46303 PM |
4842 | * Invoked early in the CPU-online process, when pretty much all services |
4843 | * are available. The incoming CPU is not present. | |
4844 | * | |
4845 | * Initializes a CPU's per-CPU RCU data. Note that only one online or | |
ff3bb6f4 PM |
4846 | * offline event can be happening at a given time. Note also that we can |
4847 | * accept some slop in the rsp->gp_seq access due to the fact that this | |
e83e73f5 PM |
4848 | * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet. |
4849 | * And any offloaded callbacks are being numbered elsewhere. | |
64db4cff | 4850 | */ |
53b46303 | 4851 | int rcutree_prepare_cpu(unsigned int cpu) |
64db4cff PM |
4852 | { |
4853 | unsigned long flags; | |
904e600e | 4854 | struct context_tracking *ct = per_cpu_ptr(&context_tracking, cpu); |
da1df50d | 4855 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
336a4f6c | 4856 | struct rcu_node *rnp = rcu_get_root(); |
64db4cff PM |
4857 | |
4858 | /* Set up local state, ensuring consistent view of global state. */ | |
6cf10081 | 4859 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
37c72e56 | 4860 | rdp->qlen_last_fqs_check = 0; |
2431774f | 4861 | rdp->n_force_qs_snap = READ_ONCE(rcu_state.n_force_qs); |
64db4cff | 4862 | rdp->blimit = blimit; |
904e600e | 4863 | ct->dynticks_nesting = 1; /* CPU not up, no tearing. */ |
67c583a7 | 4864 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
ec711bc1 | 4865 | |
126d9d49 | 4866 | /* |
ec711bc1 FW |
4867 | * Only non-NOCB CPUs that didn't have early-boot callbacks need to be |
4868 | * (re-)initialized. | |
126d9d49 | 4869 | */ |
ec711bc1 | 4870 | if (!rcu_segcblist_is_enabled(&rdp->cblist)) |
126d9d49 | 4871 | rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */ |
64db4cff | 4872 | |
0aa04b05 PM |
4873 | /* |
4874 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed | |
4875 | * propagation up the rcu_node tree will happen at the beginning | |
4876 | * of the next grace period. | |
4877 | */ | |
64db4cff | 4878 | rnp = rdp->mynode; |
2a67e741 | 4879 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
8ff37290 PM |
4880 | rdp->gp_seq = READ_ONCE(rnp->gp_seq); |
4881 | rdp->gp_seq_needed = rdp->gp_seq; | |
5b74c458 | 4882 | rdp->cpu_no_qs.b.norm = true; |
97c668b8 | 4883 | rdp->core_needs_qs = false; |
9b9500da | 4884 | rdp->rcu_iw_pending = false; |
7a9f50a0 | 4885 | rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler); |
8ff37290 | 4886 | rdp->rcu_iw_gp_seq = rdp->gp_seq - 1; |
53b46303 | 4887 | trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl")); |
67c583a7 | 4888 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
8e5e6215 | 4889 | rcu_spawn_rnp_kthreads(rnp); |
ad368d15 | 4890 | rcu_spawn_cpu_nocb_kthread(cpu); |
09e077cf | 4891 | ASSERT_EXCLUSIVE_WRITER(rcu_state.n_online_cpus); |
ed73860c | 4892 | WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1); |
4df83742 TG |
4893 | |
4894 | return 0; | |
4895 | } | |
4896 | ||
deb34f36 | 4897 | /* |
b67cffcb FW |
4898 | * Update kthreads affinity during CPU-hotplug changes. |
4899 | * | |
4900 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are | |
4901 | * served by the rcu_node in question. The CPU hotplug lock is still | |
4902 | * held, so the value of rnp->qsmaskinit will be stable. | |
4903 | * | |
4904 | * We don't include outgoingcpu in the affinity set, use -1 if there is | |
4905 | * no outgoing CPU. If there are no CPUs left in the affinity set, | |
4906 | * this function allows the kthread to execute on any CPU. | |
4907 | * | |
4908 | * Any future concurrent calls are serialized via ->kthread_mutex. | |
deb34f36 | 4909 | */ |
b67cffcb | 4910 | static void rcutree_affinity_setting(unsigned int cpu, int outgoingcpu) |
4df83742 | 4911 | { |
b67cffcb FW |
4912 | cpumask_var_t cm; |
4913 | unsigned long mask; | |
4914 | struct rcu_data *rdp; | |
4915 | struct rcu_node *rnp; | |
4916 | struct task_struct *task_boost, *task_exp; | |
4df83742 | 4917 | |
b67cffcb FW |
4918 | rdp = per_cpu_ptr(&rcu_data, cpu); |
4919 | rnp = rdp->mynode; | |
4920 | ||
4921 | task_boost = rcu_boost_task(rnp); | |
4922 | task_exp = rcu_exp_par_gp_task(rnp); | |
4923 | ||
4924 | /* | |
4925 | * If CPU is the boot one, those tasks are created later from early | |
4926 | * initcall since kthreadd must be created first. | |
4927 | */ | |
4928 | if (!task_boost && !task_exp) | |
4929 | return; | |
4930 | ||
4931 | if (!zalloc_cpumask_var(&cm, GFP_KERNEL)) | |
4932 | return; | |
4933 | ||
4934 | mutex_lock(&rnp->kthread_mutex); | |
4935 | mask = rcu_rnp_online_cpus(rnp); | |
4936 | for_each_leaf_node_possible_cpu(rnp, cpu) | |
4937 | if ((mask & leaf_node_cpu_bit(rnp, cpu)) && | |
4938 | cpu != outgoingcpu) | |
4939 | cpumask_set_cpu(cpu, cm); | |
4940 | cpumask_and(cm, cm, housekeeping_cpumask(HK_TYPE_RCU)); | |
4941 | if (cpumask_empty(cm)) { | |
4942 | cpumask_copy(cm, housekeeping_cpumask(HK_TYPE_RCU)); | |
4943 | if (outgoingcpu >= 0) | |
4944 | cpumask_clear_cpu(outgoingcpu, cm); | |
4945 | } | |
4946 | ||
4947 | if (task_exp) | |
4948 | set_cpus_allowed_ptr(task_exp, cm); | |
4949 | ||
4950 | if (task_boost) | |
4951 | set_cpus_allowed_ptr(task_boost, cm); | |
4df83742 | 4952 | |
b67cffcb | 4953 | mutex_unlock(&rnp->kthread_mutex); |
4df83742 | 4954 | |
b67cffcb | 4955 | free_cpumask_var(cm); |
4df83742 TG |
4956 | } |
4957 | ||
401b0de3 PM |
4958 | /* |
4959 | * Has the specified (known valid) CPU ever been fully online? | |
4960 | */ | |
4961 | bool rcu_cpu_beenfullyonline(int cpu) | |
4962 | { | |
4963 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
4964 | ||
4965 | return smp_load_acquire(&rdp->beenonline); | |
4966 | } | |
4967 | ||
deb34f36 PM |
4968 | /* |
4969 | * Near the end of the CPU-online process. Pretty much all services | |
4970 | * enabled, and the CPU is now very much alive. | |
4971 | */ | |
4df83742 TG |
4972 | int rcutree_online_cpu(unsigned int cpu) |
4973 | { | |
9b9500da PM |
4974 | unsigned long flags; |
4975 | struct rcu_data *rdp; | |
4976 | struct rcu_node *rnp; | |
9b9500da | 4977 | |
b97d23c5 PM |
4978 | rdp = per_cpu_ptr(&rcu_data, cpu); |
4979 | rnp = rdp->mynode; | |
4980 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
4981 | rnp->ffmask |= rdp->grpmask; | |
4982 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
9b9500da PM |
4983 | if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE) |
4984 | return 0; /* Too early in boot for scheduler work. */ | |
4985 | sync_sched_exp_online_cleanup(cpu); | |
4986 | rcutree_affinity_setting(cpu, -1); | |
96926686 PM |
4987 | |
4988 | // Stop-machine done, so allow nohz_full to disable tick. | |
4989 | tick_dep_clear(TICK_DEP_BIT_RCU); | |
4df83742 TG |
4990 | return 0; |
4991 | } | |
4992 | ||
7ec99de3 PM |
4993 | /* |
4994 | * Mark the specified CPU as being online so that subsequent grace periods | |
4995 | * (both expedited and normal) will wait on it. Note that this means that | |
4996 | * incoming CPUs are not allowed to use RCU read-side critical sections | |
4997 | * until this function is called. Failing to observe this restriction | |
4998 | * will result in lockdep splats. | |
deb34f36 PM |
4999 | * |
5000 | * Note that this function is special in that it is invoked directly | |
5001 | * from the incoming CPU rather than from the cpuhp_step mechanism. | |
5002 | * This is because this function must be invoked at a precise location. | |
15d44dfa | 5003 | * This incoming CPU must not have enabled interrupts yet. |
448e9f34 FW |
5004 | * |
5005 | * This mirrors the effects of rcutree_report_cpu_dead(). | |
7ec99de3 | 5006 | */ |
448e9f34 | 5007 | void rcutree_report_cpu_starting(unsigned int cpu) |
7ec99de3 | 5008 | { |
7ec99de3 PM |
5009 | unsigned long mask; |
5010 | struct rcu_data *rdp; | |
5011 | struct rcu_node *rnp; | |
abfce041 | 5012 | bool newcpu; |
7ec99de3 | 5013 | |
15d44dfa | 5014 | lockdep_assert_irqs_disabled(); |
c0f97f20 PM |
5015 | rdp = per_cpu_ptr(&rcu_data, cpu); |
5016 | if (rdp->cpu_started) | |
f64c6013 | 5017 | return; |
c0f97f20 | 5018 | rdp->cpu_started = true; |
f64c6013 | 5019 | |
b97d23c5 PM |
5020 | rnp = rdp->mynode; |
5021 | mask = rdp->grpmask; | |
82980b16 | 5022 | arch_spin_lock(&rcu_state.ofl_lock); |
2caebefb | 5023 | rcu_dynticks_eqs_online(); |
80b3fd47 | 5024 | raw_spin_lock(&rcu_state.barrier_lock); |
82980b16 | 5025 | raw_spin_lock_rcu_node(rnp); |
105abf82 | 5026 | WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask); |
80b3fd47 | 5027 | raw_spin_unlock(&rcu_state.barrier_lock); |
abfce041 | 5028 | newcpu = !(rnp->expmaskinitnext & mask); |
b97d23c5 | 5029 | rnp->expmaskinitnext |= mask; |
b97d23c5 | 5030 | /* Allow lockless access for expedited grace periods. */ |
abfce041 | 5031 | smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */ |
2f084695 | 5032 | ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus); |
b97d23c5 | 5033 | rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */ |
eb7a6653 | 5034 | rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq); |
ae2b217a | 5035 | rdp->rcu_onl_gp_state = READ_ONCE(rcu_state.gp_state); |
9f866dac JFG |
5036 | |
5037 | /* An incoming CPU should never be blocking a grace period. */ | |
5038 | if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */ | |
82980b16 | 5039 | /* rcu_report_qs_rnp() *really* wants some flags to restore */ |
15d44dfa | 5040 | unsigned long flags; |
82980b16 | 5041 | |
15d44dfa | 5042 | local_irq_save(flags); |
516e5ae0 | 5043 | rcu_disable_urgency_upon_qs(rdp); |
b97d23c5 | 5044 | /* Report QS -after- changing ->qsmaskinitnext! */ |
15d44dfa | 5045 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
b97d23c5 | 5046 | } else { |
82980b16 | 5047 | raw_spin_unlock_rcu_node(rnp); |
7ec99de3 | 5048 | } |
82980b16 | 5049 | arch_spin_unlock(&rcu_state.ofl_lock); |
401b0de3 | 5050 | smp_store_release(&rdp->beenonline, true); |
313517fc | 5051 | smp_mb(); /* Ensure RCU read-side usage follows above initialization. */ |
7ec99de3 PM |
5052 | } |
5053 | ||
27d50c7e | 5054 | /* |
53b46303 PM |
5055 | * The outgoing function has no further need of RCU, so remove it from |
5056 | * the rcu_node tree's ->qsmaskinitnext bit masks. | |
5057 | * | |
5058 | * Note that this function is special in that it is invoked directly | |
5059 | * from the outgoing CPU rather than from the cpuhp_step mechanism. | |
5060 | * This is because this function must be invoked at a precise location. | |
448e9f34 FW |
5061 | * |
5062 | * This mirrors the effect of rcutree_report_cpu_starting(). | |
27d50c7e | 5063 | */ |
448e9f34 | 5064 | void rcutree_report_cpu_dead(void) |
27d50c7e | 5065 | { |
358662a9 | 5066 | unsigned long flags; |
27d50c7e | 5067 | unsigned long mask; |
c964c1f5 | 5068 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
27d50c7e TG |
5069 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
5070 | ||
358662a9 FW |
5071 | /* |
5072 | * IRQS must be disabled from now on and until the CPU dies, or an interrupt | |
5073 | * may introduce a new READ-side while it is actually off the QS masks. | |
5074 | */ | |
5075 | lockdep_assert_irqs_disabled(); | |
147c6852 PM |
5076 | // Do any dangling deferred wakeups. |
5077 | do_nocb_deferred_wakeup(rdp); | |
5078 | ||
53b46303 PM |
5079 | rcu_preempt_deferred_qs(current); |
5080 | ||
27d50c7e TG |
5081 | /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ |
5082 | mask = rdp->grpmask; | |
82980b16 | 5083 | arch_spin_lock(&rcu_state.ofl_lock); |
27d50c7e | 5084 | raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */ |
53b46303 | 5085 | rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq); |
ae2b217a | 5086 | rdp->rcu_ofl_gp_state = READ_ONCE(rcu_state.gp_state); |
fece2776 PM |
5087 | if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */ |
5088 | /* Report quiescent state -before- changing ->qsmaskinitnext! */ | |
e2bb1288 | 5089 | rcu_disable_urgency_upon_qs(rdp); |
b50912d0 | 5090 | rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags); |
fece2776 PM |
5091 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
5092 | } | |
105abf82 | 5093 | WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask); |
710d60cb | 5094 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
82980b16 | 5095 | arch_spin_unlock(&rcu_state.ofl_lock); |
c0f97f20 | 5096 | rdp->cpu_started = false; |
27d50c7e | 5097 | } |
a58163d8 | 5098 | |
04e613de | 5099 | #ifdef CONFIG_HOTPLUG_CPU |
53b46303 PM |
5100 | /* |
5101 | * The outgoing CPU has just passed through the dying-idle state, and we | |
5102 | * are being invoked from the CPU that was IPIed to continue the offline | |
5103 | * operation. Migrate the outgoing CPU's callbacks to the current CPU. | |
5104 | */ | |
5105 | void rcutree_migrate_callbacks(int cpu) | |
a58163d8 PM |
5106 | { |
5107 | unsigned long flags; | |
b1a2d79f | 5108 | struct rcu_data *my_rdp; |
c00045be | 5109 | struct rcu_node *my_rnp; |
da1df50d | 5110 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
ec4eacce | 5111 | bool needwake; |
a58163d8 | 5112 | |
3820b513 | 5113 | if (rcu_rdp_is_offloaded(rdp) || |
ce5215c1 | 5114 | rcu_segcblist_empty(&rdp->cblist)) |
95335c03 PM |
5115 | return; /* No callbacks to migrate. */ |
5116 | ||
80b3fd47 | 5117 | raw_spin_lock_irqsave(&rcu_state.barrier_lock, flags); |
a16578dd PM |
5118 | WARN_ON_ONCE(rcu_rdp_cpu_online(rdp)); |
5119 | rcu_barrier_entrain(rdp); | |
da1df50d | 5120 | my_rdp = this_cpu_ptr(&rcu_data); |
c00045be | 5121 | my_rnp = my_rdp->mynode; |
5d6742b3 | 5122 | rcu_nocb_lock(my_rdp); /* irqs already disabled. */ |
3cb278e7 | 5123 | WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies, false)); |
c00045be | 5124 | raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */ |
ec4eacce | 5125 | /* Leverage recent GPs and set GP for new callbacks. */ |
c00045be PM |
5126 | needwake = rcu_advance_cbs(my_rnp, rdp) || |
5127 | rcu_advance_cbs(my_rnp, my_rdp); | |
f2dbe4a5 | 5128 | rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist); |
80b3fd47 | 5129 | raw_spin_unlock(&rcu_state.barrier_lock); /* irqs remain disabled. */ |
23651d9b | 5130 | needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp); |
c035280f | 5131 | rcu_segcblist_disable(&rdp->cblist); |
a16578dd | 5132 | WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) != !rcu_segcblist_n_cbs(&my_rdp->cblist)); |
52c1d81e | 5133 | check_cb_ovld_locked(my_rdp, my_rnp); |
3820b513 | 5134 | if (rcu_rdp_is_offloaded(my_rdp)) { |
5d6742b3 PM |
5135 | raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */ |
5136 | __call_rcu_nocb_wake(my_rdp, true, flags); | |
5137 | } else { | |
5138 | rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */ | |
b913c3fe | 5139 | raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */ |
5d6742b3 | 5140 | } |
b913c3fe | 5141 | local_irq_restore(flags); |
ec4eacce | 5142 | if (needwake) |
532c00c9 | 5143 | rcu_gp_kthread_wake(); |
5d6742b3 | 5144 | lockdep_assert_irqs_enabled(); |
a58163d8 PM |
5145 | WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 || |
5146 | !rcu_segcblist_empty(&rdp->cblist), | |
5147 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n", | |
5148 | cpu, rcu_segcblist_n_cbs(&rdp->cblist), | |
5149 | rcu_segcblist_first_cb(&rdp->cblist)); | |
5150 | } | |
2cb1f6e9 FW |
5151 | |
5152 | /* | |
5153 | * The CPU has been completely removed, and some other CPU is reporting | |
5154 | * this fact from process context. Do the remainder of the cleanup. | |
5155 | * There can only be one CPU hotplug operation at a time, so no need for | |
5156 | * explicit locking. | |
5157 | */ | |
5158 | int rcutree_dead_cpu(unsigned int cpu) | |
5159 | { | |
09e077cf | 5160 | ASSERT_EXCLUSIVE_WRITER(rcu_state.n_online_cpus); |
2cb1f6e9 FW |
5161 | WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1); |
5162 | // Stop-machine done, so allow nohz_full to disable tick. | |
5163 | tick_dep_clear(TICK_DEP_BIT_RCU); | |
5164 | return 0; | |
5165 | } | |
5166 | ||
5167 | /* | |
5168 | * Near the end of the offline process. Trace the fact that this CPU | |
5169 | * is going offline. | |
5170 | */ | |
5171 | int rcutree_dying_cpu(unsigned int cpu) | |
5172 | { | |
5173 | bool blkd; | |
5174 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); | |
5175 | struct rcu_node *rnp = rdp->mynode; | |
5176 | ||
5177 | blkd = !!(READ_ONCE(rnp->qsmask) & rdp->grpmask); | |
5178 | trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq), | |
5179 | blkd ? TPS("cpuofl-bgp") : TPS("cpuofl")); | |
5180 | return 0; | |
5181 | } | |
5182 | ||
5183 | /* | |
5184 | * Near the beginning of the process. The CPU is still very much alive | |
5185 | * with pretty much all services enabled. | |
5186 | */ | |
5187 | int rcutree_offline_cpu(unsigned int cpu) | |
5188 | { | |
5189 | unsigned long flags; | |
5190 | struct rcu_data *rdp; | |
5191 | struct rcu_node *rnp; | |
5192 | ||
5193 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
5194 | rnp = rdp->mynode; | |
5195 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
5196 | rnp->ffmask &= ~rdp->grpmask; | |
5197 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
5198 | ||
5199 | rcutree_affinity_setting(cpu, cpu); | |
5200 | ||
5201 | // nohz_full CPUs need the tick for stop-machine to work quickly | |
5202 | tick_dep_set(TICK_DEP_BIT_RCU); | |
5203 | return 0; | |
5204 | } | |
5205 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
27d50c7e | 5206 | |
deb34f36 PM |
5207 | /* |
5208 | * On non-huge systems, use expedited RCU grace periods to make suspend | |
5209 | * and hibernation run faster. | |
5210 | */ | |
d1d74d14 BP |
5211 | static int rcu_pm_notify(struct notifier_block *self, |
5212 | unsigned long action, void *hcpu) | |
5213 | { | |
5214 | switch (action) { | |
5215 | case PM_HIBERNATION_PREPARE: | |
5216 | case PM_SUSPEND_PREPARE: | |
6efdda8b | 5217 | rcu_async_hurry(); |
e85e6a21 | 5218 | rcu_expedite_gp(); |
d1d74d14 BP |
5219 | break; |
5220 | case PM_POST_HIBERNATION: | |
5221 | case PM_POST_SUSPEND: | |
e85e6a21 | 5222 | rcu_unexpedite_gp(); |
6efdda8b | 5223 | rcu_async_relax(); |
d1d74d14 BP |
5224 | break; |
5225 | default: | |
5226 | break; | |
5227 | } | |
5228 | return NOTIFY_OK; | |
5229 | } | |
5230 | ||
b3dbec76 | 5231 | /* |
49918a54 | 5232 | * Spawn the kthreads that handle RCU's grace periods. |
b3dbec76 PM |
5233 | */ |
5234 | static int __init rcu_spawn_gp_kthread(void) | |
5235 | { | |
5236 | unsigned long flags; | |
5237 | struct rcu_node *rnp; | |
a94844b2 | 5238 | struct sched_param sp; |
b3dbec76 | 5239 | struct task_struct *t; |
3352911f | 5240 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
b3dbec76 | 5241 | |
9386c0b7 | 5242 | rcu_scheduler_fully_active = 1; |
b97d23c5 | 5243 | t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name); |
08543bda PM |
5244 | if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__)) |
5245 | return 0; | |
b97d23c5 PM |
5246 | if (kthread_prio) { |
5247 | sp.sched_priority = kthread_prio; | |
5248 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
b3dbec76 | 5249 | } |
b97d23c5 PM |
5250 | rnp = rcu_get_root(); |
5251 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
5648d659 PM |
5252 | WRITE_ONCE(rcu_state.gp_activity, jiffies); |
5253 | WRITE_ONCE(rcu_state.gp_req_activity, jiffies); | |
5254 | // Reset .gp_activity and .gp_req_activity before setting .gp_kthread. | |
5255 | smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */ | |
b97d23c5 PM |
5256 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
5257 | wake_up_process(t); | |
3352911f FW |
5258 | /* This is a pre-SMP initcall, we expect a single CPU */ |
5259 | WARN_ON(num_online_cpus() > 1); | |
87c5adf0 FW |
5260 | /* |
5261 | * Those kthreads couldn't be created on rcu_init() -> rcutree_prepare_cpu() | |
5262 | * due to rcu_scheduler_fully_active. | |
5263 | */ | |
5264 | rcu_spawn_cpu_nocb_kthread(smp_processor_id()); | |
8e5e6215 | 5265 | rcu_spawn_rnp_kthreads(rdp->mynode); |
8e4b1d2b | 5266 | rcu_spawn_core_kthreads(); |
9621fbee | 5267 | /* Create kthread worker for expedited GPs */ |
8e5e6215 | 5268 | rcu_start_exp_gp_kworker(); |
b3dbec76 PM |
5269 | return 0; |
5270 | } | |
5271 | early_initcall(rcu_spawn_gp_kthread); | |
5272 | ||
bbad9379 | 5273 | /* |
52d7e48b PM |
5274 | * This function is invoked towards the end of the scheduler's |
5275 | * initialization process. Before this is called, the idle task might | |
5276 | * contain synchronous grace-period primitives (during which time, this idle | |
5277 | * task is booting the system, and such primitives are no-ops). After this | |
5278 | * function is called, any synchronous grace-period primitives are run as | |
5279 | * expedited, with the requesting task driving the grace period forward. | |
900b1028 | 5280 | * A later core_initcall() rcu_set_runtime_mode() will switch to full |
52d7e48b | 5281 | * runtime RCU functionality. |
bbad9379 PM |
5282 | */ |
5283 | void rcu_scheduler_starting(void) | |
5284 | { | |
d761de8a PM |
5285 | unsigned long flags; |
5286 | struct rcu_node *rnp; | |
5287 | ||
bbad9379 PM |
5288 | WARN_ON(num_online_cpus() != 1); |
5289 | WARN_ON(nr_context_switches() > 0); | |
52d7e48b | 5290 | rcu_test_sync_prims(); |
d761de8a PM |
5291 | |
5292 | // Fix up the ->gp_seq counters. | |
5293 | local_irq_save(flags); | |
5294 | rcu_for_each_node_breadth_first(rnp) | |
5295 | rnp->gp_seq_needed = rnp->gp_seq = rcu_state.gp_seq; | |
5296 | local_irq_restore(flags); | |
5297 | ||
5298 | // Switch out of early boot mode. | |
52d7e48b PM |
5299 | rcu_scheduler_active = RCU_SCHEDULER_INIT; |
5300 | rcu_test_sync_prims(); | |
bbad9379 PM |
5301 | } |
5302 | ||
64db4cff | 5303 | /* |
49918a54 | 5304 | * Helper function for rcu_init() that initializes the rcu_state structure. |
64db4cff | 5305 | */ |
b8bb1f63 | 5306 | static void __init rcu_init_one(void) |
64db4cff | 5307 | { |
cb007102 AG |
5308 | static const char * const buf[] = RCU_NODE_NAME_INIT; |
5309 | static const char * const fqs[] = RCU_FQS_NAME_INIT; | |
3dc5dbe9 PM |
5310 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
5311 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; | |
199977bf | 5312 | |
199977bf | 5313 | int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */ |
64db4cff PM |
5314 | int cpustride = 1; |
5315 | int i; | |
5316 | int j; | |
5317 | struct rcu_node *rnp; | |
5318 | ||
05b84aec | 5319 | BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
b6407e86 | 5320 | |
3eaaaf6c PM |
5321 | /* Silence gcc 4.8 false positive about array index out of range. */ |
5322 | if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS) | |
5323 | panic("rcu_init_one: rcu_num_lvls out of range"); | |
4930521a | 5324 | |
64db4cff PM |
5325 | /* Initialize the level-tracking arrays. */ |
5326 | ||
f885b7f2 | 5327 | for (i = 1; i < rcu_num_lvls; i++) |
eb7a6653 PM |
5328 | rcu_state.level[i] = |
5329 | rcu_state.level[i - 1] + num_rcu_lvl[i - 1]; | |
41f5c631 | 5330 | rcu_init_levelspread(levelspread, num_rcu_lvl); |
64db4cff PM |
5331 | |
5332 | /* Initialize the elements themselves, starting from the leaves. */ | |
5333 | ||
f885b7f2 | 5334 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
199977bf | 5335 | cpustride *= levelspread[i]; |
eb7a6653 | 5336 | rnp = rcu_state.level[i]; |
41f5c631 | 5337 | for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) { |
67c583a7 BF |
5338 | raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock)); |
5339 | lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock), | |
b6407e86 | 5340 | &rcu_node_class[i], buf[i]); |
394f2769 PM |
5341 | raw_spin_lock_init(&rnp->fqslock); |
5342 | lockdep_set_class_and_name(&rnp->fqslock, | |
5343 | &rcu_fqs_class[i], fqs[i]); | |
eb7a6653 PM |
5344 | rnp->gp_seq = rcu_state.gp_seq; |
5345 | rnp->gp_seq_needed = rcu_state.gp_seq; | |
5346 | rnp->completedqs = rcu_state.gp_seq; | |
64db4cff PM |
5347 | rnp->qsmask = 0; |
5348 | rnp->qsmaskinit = 0; | |
5349 | rnp->grplo = j * cpustride; | |
5350 | rnp->grphi = (j + 1) * cpustride - 1; | |
595f3900 HS |
5351 | if (rnp->grphi >= nr_cpu_ids) |
5352 | rnp->grphi = nr_cpu_ids - 1; | |
64db4cff PM |
5353 | if (i == 0) { |
5354 | rnp->grpnum = 0; | |
5355 | rnp->grpmask = 0; | |
5356 | rnp->parent = NULL; | |
5357 | } else { | |
199977bf | 5358 | rnp->grpnum = j % levelspread[i - 1]; |
df63fa5b | 5359 | rnp->grpmask = BIT(rnp->grpnum); |
eb7a6653 | 5360 | rnp->parent = rcu_state.level[i - 1] + |
199977bf | 5361 | j / levelspread[i - 1]; |
64db4cff PM |
5362 | } |
5363 | rnp->level = i; | |
12f5f524 | 5364 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
dae6e64d | 5365 | rcu_init_one_nocb(rnp); |
f6a12f34 PM |
5366 | init_waitqueue_head(&rnp->exp_wq[0]); |
5367 | init_waitqueue_head(&rnp->exp_wq[1]); | |
3b5f668e PM |
5368 | init_waitqueue_head(&rnp->exp_wq[2]); |
5369 | init_waitqueue_head(&rnp->exp_wq[3]); | |
f6a12f34 | 5370 | spin_lock_init(&rnp->exp_lock); |
7836b270 | 5371 | mutex_init(&rnp->kthread_mutex); |
d96c52fe PM |
5372 | raw_spin_lock_init(&rnp->exp_poll_lock); |
5373 | rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED; | |
5374 | INIT_WORK(&rnp->exp_poll_wq, sync_rcu_do_polled_gp); | |
64db4cff PM |
5375 | } |
5376 | } | |
0c34029a | 5377 | |
eb7a6653 PM |
5378 | init_swait_queue_head(&rcu_state.gp_wq); |
5379 | init_swait_queue_head(&rcu_state.expedited_wq); | |
aedf4ba9 | 5380 | rnp = rcu_first_leaf_node(); |
0c34029a | 5381 | for_each_possible_cpu(i) { |
4a90a068 | 5382 | while (i > rnp->grphi) |
0c34029a | 5383 | rnp++; |
da1df50d | 5384 | per_cpu_ptr(&rcu_data, i)->mynode = rnp; |
53b46303 | 5385 | rcu_boot_init_percpu_data(i); |
0c34029a | 5386 | } |
64db4cff PM |
5387 | } |
5388 | ||
c8db27dd AC |
5389 | /* |
5390 | * Force priority from the kernel command-line into range. | |
5391 | */ | |
5392 | static void __init sanitize_kthread_prio(void) | |
5393 | { | |
5394 | int kthread_prio_in = kthread_prio; | |
5395 | ||
5396 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2 | |
5397 | && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST)) | |
5398 | kthread_prio = 2; | |
5399 | else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) | |
5400 | kthread_prio = 1; | |
5401 | else if (kthread_prio < 0) | |
5402 | kthread_prio = 0; | |
5403 | else if (kthread_prio > 99) | |
5404 | kthread_prio = 99; | |
5405 | ||
5406 | if (kthread_prio != kthread_prio_in) | |
5407 | pr_alert("%s: Limited prio to %d from %d\n", | |
5408 | __func__, kthread_prio, kthread_prio_in); | |
5409 | } | |
5410 | ||
f885b7f2 PM |
5411 | /* |
5412 | * Compute the rcu_node tree geometry from kernel parameters. This cannot | |
4102adab | 5413 | * replace the definitions in tree.h because those are needed to size |
f885b7f2 PM |
5414 | * the ->node array in the rcu_state structure. |
5415 | */ | |
b5befe84 | 5416 | void rcu_init_geometry(void) |
f885b7f2 | 5417 | { |
026ad283 | 5418 | ulong d; |
f885b7f2 | 5419 | int i; |
b5befe84 | 5420 | static unsigned long old_nr_cpu_ids; |
05b84aec | 5421 | int rcu_capacity[RCU_NUM_LVLS]; |
b5befe84 FW |
5422 | static bool initialized; |
5423 | ||
5424 | if (initialized) { | |
5425 | /* | |
5426 | * Warn if setup_nr_cpu_ids() had not yet been invoked, | |
5427 | * unless nr_cpus_ids == NR_CPUS, in which case who cares? | |
5428 | */ | |
5429 | WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids); | |
5430 | return; | |
5431 | } | |
5432 | ||
5433 | old_nr_cpu_ids = nr_cpu_ids; | |
5434 | initialized = true; | |
f885b7f2 | 5435 | |
026ad283 PM |
5436 | /* |
5437 | * Initialize any unspecified boot parameters. | |
5438 | * The default values of jiffies_till_first_fqs and | |
5439 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS | |
5440 | * value, which is a function of HZ, then adding one for each | |
5441 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. | |
5442 | */ | |
5443 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; | |
5444 | if (jiffies_till_first_fqs == ULONG_MAX) | |
5445 | jiffies_till_first_fqs = d; | |
5446 | if (jiffies_till_next_fqs == ULONG_MAX) | |
5447 | jiffies_till_next_fqs = d; | |
6973032a | 5448 | adjust_jiffies_till_sched_qs(); |
026ad283 | 5449 | |
f885b7f2 | 5450 | /* If the compile-time values are accurate, just leave. */ |
47d631af | 5451 | if (rcu_fanout_leaf == RCU_FANOUT_LEAF && |
b17c7035 | 5452 | nr_cpu_ids == NR_CPUS) |
f885b7f2 | 5453 | return; |
a7538352 | 5454 | pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n", |
39479098 | 5455 | rcu_fanout_leaf, nr_cpu_ids); |
f885b7f2 | 5456 | |
f885b7f2 | 5457 | /* |
ee968ac6 PM |
5458 | * The boot-time rcu_fanout_leaf parameter must be at least two |
5459 | * and cannot exceed the number of bits in the rcu_node masks. | |
5460 | * Complain and fall back to the compile-time values if this | |
5461 | * limit is exceeded. | |
f885b7f2 | 5462 | */ |
ee968ac6 | 5463 | if (rcu_fanout_leaf < 2 || |
75cf15a4 | 5464 | rcu_fanout_leaf > sizeof(unsigned long) * 8) { |
13bd6494 | 5465 | rcu_fanout_leaf = RCU_FANOUT_LEAF; |
f885b7f2 PM |
5466 | WARN_ON(1); |
5467 | return; | |
5468 | } | |
5469 | ||
f885b7f2 PM |
5470 | /* |
5471 | * Compute number of nodes that can be handled an rcu_node tree | |
9618138b | 5472 | * with the given number of levels. |
f885b7f2 | 5473 | */ |
9618138b | 5474 | rcu_capacity[0] = rcu_fanout_leaf; |
05b84aec | 5475 | for (i = 1; i < RCU_NUM_LVLS; i++) |
05c5df31 | 5476 | rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT; |
f885b7f2 PM |
5477 | |
5478 | /* | |
75cf15a4 | 5479 | * The tree must be able to accommodate the configured number of CPUs. |
ee968ac6 | 5480 | * If this limit is exceeded, fall back to the compile-time values. |
f885b7f2 | 5481 | */ |
ee968ac6 PM |
5482 | if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) { |
5483 | rcu_fanout_leaf = RCU_FANOUT_LEAF; | |
5484 | WARN_ON(1); | |
5485 | return; | |
5486 | } | |
f885b7f2 | 5487 | |
679f9858 | 5488 | /* Calculate the number of levels in the tree. */ |
9618138b | 5489 | for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) { |
679f9858 | 5490 | } |
9618138b | 5491 | rcu_num_lvls = i + 1; |
679f9858 | 5492 | |
f885b7f2 | 5493 | /* Calculate the number of rcu_nodes at each level of the tree. */ |
679f9858 | 5494 | for (i = 0; i < rcu_num_lvls; i++) { |
9618138b | 5495 | int cap = rcu_capacity[(rcu_num_lvls - 1) - i]; |
679f9858 AG |
5496 | num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap); |
5497 | } | |
f885b7f2 PM |
5498 | |
5499 | /* Calculate the total number of rcu_node structures. */ | |
5500 | rcu_num_nodes = 0; | |
679f9858 | 5501 | for (i = 0; i < rcu_num_lvls; i++) |
f885b7f2 | 5502 | rcu_num_nodes += num_rcu_lvl[i]; |
f885b7f2 PM |
5503 | } |
5504 | ||
a3dc2948 PM |
5505 | /* |
5506 | * Dump out the structure of the rcu_node combining tree associated | |
49918a54 | 5507 | * with the rcu_state structure. |
a3dc2948 | 5508 | */ |
b8bb1f63 | 5509 | static void __init rcu_dump_rcu_node_tree(void) |
a3dc2948 PM |
5510 | { |
5511 | int level = 0; | |
5512 | struct rcu_node *rnp; | |
5513 | ||
5514 | pr_info("rcu_node tree layout dump\n"); | |
5515 | pr_info(" "); | |
aedf4ba9 | 5516 | rcu_for_each_node_breadth_first(rnp) { |
a3dc2948 PM |
5517 | if (rnp->level != level) { |
5518 | pr_cont("\n"); | |
5519 | pr_info(" "); | |
5520 | level = rnp->level; | |
5521 | } | |
5522 | pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum); | |
5523 | } | |
5524 | pr_cont("\n"); | |
5525 | } | |
5526 | ||
ad7c946b PM |
5527 | struct workqueue_struct *rcu_gp_wq; |
5528 | ||
a35d1690 BP |
5529 | static void __init kfree_rcu_batch_init(void) |
5530 | { | |
5531 | int cpu; | |
27538e18 | 5532 | int i, j; |
21e0b932 | 5533 | struct shrinker *kfree_rcu_shrinker; |
a35d1690 | 5534 | |
d0bfa8b3 ZQ |
5535 | /* Clamp it to [0:100] seconds interval. */ |
5536 | if (rcu_delay_page_cache_fill_msec < 0 || | |
5537 | rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) { | |
5538 | ||
5539 | rcu_delay_page_cache_fill_msec = | |
5540 | clamp(rcu_delay_page_cache_fill_msec, 0, | |
5541 | (int) (100 * MSEC_PER_SEC)); | |
5542 | ||
5543 | pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n", | |
5544 | rcu_delay_page_cache_fill_msec); | |
5545 | } | |
5546 | ||
a35d1690 BP |
5547 | for_each_possible_cpu(cpu) { |
5548 | struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu); | |
5549 | ||
34c88174 URS |
5550 | for (i = 0; i < KFREE_N_BATCHES; i++) { |
5551 | INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work); | |
0392bebe | 5552 | krcp->krw_arr[i].krcp = krcp; |
27538e18 URS |
5553 | |
5554 | for (j = 0; j < FREE_N_CHANNELS; j++) | |
5555 | INIT_LIST_HEAD(&krcp->krw_arr[i].bulk_head_free[j]); | |
34c88174 URS |
5556 | } |
5557 | ||
27538e18 URS |
5558 | for (i = 0; i < FREE_N_CHANNELS; i++) |
5559 | INIT_LIST_HEAD(&krcp->bulk_head[i]); | |
5560 | ||
a35d1690 | 5561 | INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor); |
d0bfa8b3 | 5562 | INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func); |
a35d1690 BP |
5563 | krcp->initialized = true; |
5564 | } | |
21e0b932 QZ |
5565 | |
5566 | kfree_rcu_shrinker = shrinker_alloc(0, "rcu-kfree"); | |
5567 | if (!kfree_rcu_shrinker) { | |
5568 | pr_err("Failed to allocate kfree_rcu() shrinker!\n"); | |
5569 | return; | |
5570 | } | |
5571 | ||
5572 | kfree_rcu_shrinker->count_objects = kfree_rcu_shrink_count; | |
5573 | kfree_rcu_shrinker->scan_objects = kfree_rcu_shrink_scan; | |
5574 | ||
5575 | shrinker_register(kfree_rcu_shrinker); | |
a35d1690 BP |
5576 | } |
5577 | ||
9f680ab4 | 5578 | void __init rcu_init(void) |
64db4cff | 5579 | { |
2eed973a | 5580 | int cpu = smp_processor_id(); |
9f680ab4 | 5581 | |
47627678 PM |
5582 | rcu_early_boot_tests(); |
5583 | ||
a35d1690 | 5584 | kfree_rcu_batch_init(); |
f41d911f | 5585 | rcu_bootup_announce(); |
c8db27dd | 5586 | sanitize_kthread_prio(); |
f885b7f2 | 5587 | rcu_init_geometry(); |
b8bb1f63 | 5588 | rcu_init_one(); |
a3dc2948 | 5589 | if (dump_tree) |
b8bb1f63 | 5590 | rcu_dump_rcu_node_tree(); |
48d07c04 SAS |
5591 | if (use_softirq) |
5592 | open_softirq(RCU_SOFTIRQ, rcu_core_si); | |
9f680ab4 PM |
5593 | |
5594 | /* | |
5595 | * We don't need protection against CPU-hotplug here because | |
5596 | * this is called early in boot, before either interrupts | |
5597 | * or the scheduler are operational. | |
5598 | */ | |
d1d74d14 | 5599 | pm_notifier(rcu_pm_notify, 0); |
2eed973a FW |
5600 | WARN_ON(num_online_cpus() > 1); // Only one CPU this early in boot. |
5601 | rcutree_prepare_cpu(cpu); | |
448e9f34 | 5602 | rcutree_report_cpu_starting(cpu); |
2eed973a | 5603 | rcutree_online_cpu(cpu); |
ad7c946b | 5604 | |
277ffe1b | 5605 | /* Create workqueue for Tree SRCU and for expedited GPs. */ |
ad7c946b PM |
5606 | rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0); |
5607 | WARN_ON(!rcu_gp_wq); | |
b2b00ddf | 5608 | |
0fd210ba URS |
5609 | sync_wq = alloc_workqueue("sync_wq", WQ_MEM_RECLAIM, 0); |
5610 | WARN_ON(!sync_wq); | |
5611 | ||
b2b00ddf PM |
5612 | /* Fill in default value for rcutree.qovld boot parameter. */ |
5613 | /* -After- the rcu_node ->lock fields are initialized! */ | |
5614 | if (qovld < 0) | |
5615 | qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark; | |
5616 | else | |
5617 | qovld_calc = qovld; | |
d96c52fe | 5618 | |
7ea91307 Z |
5619 | // Kick-start in case any polled grace periods started early. |
5620 | (void)start_poll_synchronize_rcu_expedited(); | |
748bf47a PM |
5621 | |
5622 | rcu_test_sync_prims(); | |
30ef0963 PM |
5623 | |
5624 | tasks_cblist_init_generic(); | |
64db4cff PM |
5625 | } |
5626 | ||
10462d6f | 5627 | #include "tree_stall.h" |
3549c2bc | 5628 | #include "tree_exp.h" |
dfcb2754 | 5629 | #include "tree_nocb.h" |
4102adab | 5630 | #include "tree_plugin.h" |