| 1 | |
| 2 | * Read-Copy Update mechanism for mutual exclusion
|
| 3 | *
|
| 4 | * This program is free software; you can redistribute it and/or modify
|
| 5 | * it under the terms of the GNU General Public License as published by
|
| 6 | * the Free Software Foundation; either version 2 of the License, or
|
| 7 | * (at your option) any later version.
|
| 8 | *
|
| 9 | * This program is distributed in the hope that it will be useful,
|
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
| 12 | * GNU General Public License for more details.
|
| 13 | *
|
| 14 | * You should have received a copy of the GNU General Public License
|
| 15 | * along with this program; if not, you can access it online at
|
| 16 | * http://www.gnu.org/licenses/gpl-2.0.html.
|
| 17 | *
|
| 18 | * Copyright IBM Corporation, 2008
|
| 19 | *
|
| 20 | * Authors: Dipankar Sarma <[email protected]>
|
| 21 | * Manfred Spraul <[email protected]>
|
| 22 | * Paul E. McKenney <[email protected]> Hierarchical version
|
| 23 | *
|
| 24 | * Based on the original work by Paul McKenney <[email protected]>
|
| 25 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
|
| 26 | *
|
| 27 | * For detailed explanation of Read-Copy Update mechanism see -
|
| 28 | * Documentation/RCU
|
| 29 |
|
| 30 | #include <linux/types.h>
|
| 31 | #include <linux/kernel.h>
|
| 32 | #include <linux/init.h>
|
| 33 | #include <linux/spinlock.h>
|
| 34 | #include <linux/smp.h>
|
| 35 | #include <linux/rcupdate_wait.h>
|
| 36 | #include <linux/interrupt.h>
|
| 37 | #include <linux/sched.h>
|
| 38 | #include <linux/sched/debug.h>
|
| 39 | #include <linux/nmi.h>
|
| 40 | #include <linux/atomic.h>
|
| 41 | #include <linux/bitops.h>
|
| 42 | #include <linux/export.h>
|
| 43 | #include <linux/completion.h>
|
| 44 | #include <linux/moduleparam.h>
|
| 45 | #include <linux/percpu.h>
|
| 46 | #include <linux/notifier.h>
|
| 47 | #include <linux/cpu.h>
|
| 48 | #include <linux/mutex.h>
|
| 49 | #include <linux/time.h>
|
| 50 | #include <linux/kernel_stat.h>
|
| 51 | #include <linux/wait.h>
|
| 52 | #include <linux/kthread.h>
|
| 53 | #include <uapi/linux/sched/types.h>
|
| 54 | #include <linux/prefetch.h>
|
| 55 | #include <linux/delay.h>
|
| 56 | #include <linux/stop_machine.h>
|
| 57 | #include <linux/random.h>
|
| 58 | #include <linux/trace_events.h>
|
| 59 | #include <linux/suspend.h>
|
| 60 | #include <linux/ftrace.h>
|
| 61 |
|
| 62 | #include "tree.h"
|
| 63 | #include "rcu.h"
|
| 64 |
|
| 65 | #ifdef MODULE_PARAM_PREFIX
|
| 66 | #undef MODULE_PARAM_PREFIX
|
| 67 | #endif
|
| 68 | #define MODULE_PARAM_PREFIX "rcutree."
|
| 69 |
|
| 70 |
|
| 71 |
|
| 72 | |
| 73 | * In order to export the rcu_state name to the tracing tools, it
|
| 74 | * needs to be added in the __tracepoint_string section.
|
| 75 | * This requires defining a separate variable tp_<sname>_varname
|
| 76 | * that points to the string being used, and this will allow
|
| 77 | * the tracing userspace tools to be able to decipher the string
|
| 78 | * address to the matching string.
|
| 79 |
|
| 80 | #ifdef CONFIG_TRACING
|
| 81 | # define DEFINE_RCU_TPS(sname) \
|
| 82 | static char sname##_varname[] = #sname; \
|
| 83 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
|
| 84 | # define RCU_STATE_NAME(sname) sname##_varname
|
| 85 | #else
|
| 86 | # define DEFINE_RCU_TPS(sname)
|
| 87 | # define RCU_STATE_NAME(sname) __stringify(sname)
|
| 88 | #endif
|
| 89 |
|
| 90 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
|
| 91 | DEFINE_RCU_TPS(sname) \
|
| 92 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
|
| 93 | struct rcu_state sname##_state = { \
|
| 94 | .level = { &sname##_state.node[0] }, \
|
| 95 | .rda = &sname##_data, \
|
| 96 | .call = cr, \
|
| 97 | .gp_state = RCU_GP_IDLE, \
|
| 98 | .gpnum = 0UL - 300UL, \
|
| 99 | .completed = 0UL - 300UL, \
|
| 100 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
|
| 101 | .orphan_pend = RCU_CBLIST_INITIALIZER(sname##_state.orphan_pend), \
|
| 102 | .orphan_done = RCU_CBLIST_INITIALIZER(sname##_state.orphan_done), \
|
| 103 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
|
| 104 | .name = RCU_STATE_NAME(sname), \
|
| 105 | .abbr = sabbr, \
|
| 106 | .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
|
| 107 | .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
|
| 108 | }
|
| 109 |
|
| 110 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
|
| 111 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
|
| 112 |
|
| 113 | static struct rcu_state *const rcu_state_p;
|
| 114 | LIST_HEAD(rcu_struct_flavors);
|
| 115 |
|
| 116 |
|
| 117 | static bool dump_tree;
|
| 118 | module_param(dump_tree, bool, 0444);
|
| 119 |
|
| 120 | static bool rcu_fanout_exact;
|
| 121 | module_param(rcu_fanout_exact, bool, 0444);
|
| 122 |
|
| 123 | static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
|
| 124 | module_param(rcu_fanout_leaf, int, 0444);
|
| 125 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
|
| 126 |
|
| 127 | int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
|
| 128 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES;
|
| 129 |
|
| 130 | int sysctl_panic_on_rcu_stall __read_mostly;
|
| 131 |
|
| 132 | |
| 133 | * The rcu_scheduler_active variable is initialized to the value
|
| 134 | * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
|
| 135 | * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
|
| 136 | * RCU can assume that there is but one task, allowing RCU to (for example)
|
| 137 | * optimize synchronize_rcu() to a simple barrier(). When this variable
|
| 138 | * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
|
| 139 | * to detect real grace periods. This variable is also used to suppress
|
| 140 | * boot-time false positives from lockdep-RCU error checking. Finally, it
|
| 141 | * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
|
| 142 | * is fully initialized, including all of its kthreads having been spawned.
|
| 143 |
|
| 144 | int rcu_scheduler_active __read_mostly;
|
| 145 | EXPORT_SYMBOL_GPL(rcu_scheduler_active);
|
| 146 |
|
| 147 | |
| 148 | * The rcu_scheduler_fully_active variable transitions from zero to one
|
| 149 | * during the early_initcall() processing, which is after the scheduler
|
| 150 | * is capable of creating new tasks. So RCU processing (for example,
|
| 151 | * creating tasks for RCU priority boosting) must be delayed until after
|
| 152 | * rcu_scheduler_fully_active transitions from zero to one. We also
|
| 153 | * currently delay invocation of any RCU callbacks until after this point.
|
| 154 | *
|
| 155 | * It might later prove better for people registering RCU callbacks during
|
| 156 | * early boot to take responsibility for these callbacks, but one step at
|
| 157 | * a time.
|
| 158 |
|
| 159 | static int rcu_scheduler_fully_active __read_mostly;
|
| 160 |
|
| 161 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
|
| 162 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
|
| 163 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
|
| 164 | static void invoke_rcu_core(void);
|
| 165 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
|
| 166 | static void rcu_report_exp_rdp(struct rcu_state *rsp,
|
| 167 | struct rcu_data *rdp, bool wake);
|
| 168 | static void sync_sched_exp_online_cleanup(int cpu);
|
| 169 |
|
| 170 |
|
| 171 | static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
|
| 172 | module_param(kthread_prio, int, 0644);
|
| 173 |
|
| 174 |
|
| 175 |
|
| 176 | static int gp_preinit_delay;
|
| 177 | module_param(gp_preinit_delay, int, 0444);
|
| 178 | static int gp_init_delay;
|
| 179 | module_param(gp_init_delay, int, 0444);
|
| 180 | static int gp_cleanup_delay;
|
| 181 | module_param(gp_cleanup_delay, int, 0444);
|
| 182 |
|
| 183 | |
| 184 | * Number of grace periods between delays, normalized by the duration of
|
| 185 | * the delay. The longer the delay, the more the grace periods between
|
| 186 | * each delay. The reason for this normalization is that it means that,
|
| 187 | * for non-zero delays, the overall slowdown of grace periods is constant
|
| 188 | * regardless of the duration of the delay. This arrangement balances
|
| 189 | * the need for long delays to increase some race probabilities with the
|
| 190 | * need for fast grace periods to increase other race probabilities.
|
| 191 |
|
| 192 | #define PER_RCU_NODE_PERIOD 3
|
| 193 |
|
| 194 | |
| 195 | * Track the rcutorture test sequence number and the update version
|
| 196 | * number within a given test. The rcutorture_testseq is incremented
|
| 197 | * on every rcutorture module load and unload, so has an odd value
|
| 198 | * when a test is running. The rcutorture_vernum is set to zero
|
| 199 | * when rcutorture starts and is incremented on each rcutorture update.
|
| 200 | * These variables enable correlating rcutorture output with the
|
| 201 | * RCU tracing information.
|
| 202 |
|
| 203 | unsigned long rcutorture_testseq;
|
| 204 | unsigned long rcutorture_vernum;
|
| 205 |
|
| 206 | |
| 207 | * Compute the mask of online CPUs for the specified rcu_node structure.
|
| 208 | * This will not be stable unless the rcu_node structure's ->lock is
|
| 209 | * held, but the bit corresponding to the current CPU will be stable
|
| 210 | * in most contexts.
|
| 211 |
|
| 212 | unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
|
| 213 | {
|
| 214 | return READ_ONCE(rnp->qsmaskinitnext);
|
| 215 | }
|
| 216 |
|
| 217 | |
| 218 | * Return true if an RCU grace period is in progress. The READ_ONCE()s
|
| 219 | * permit this function to be invoked without holding the root rcu_node
|
| 220 | * structure's ->lock, but of course results can be subject to change.
|
| 221 |
|
| 222 | static int rcu_gp_in_progress(struct rcu_state *rsp)
|
| 223 | {
|
| 224 | return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
|
| 225 | }
|
| 226 |
|
| 227 | |
| 228 | * Note a quiescent state. Because we do not need to know
|
| 229 | * how many quiescent states passed, just if there was at least
|
| 230 | * one since the start of the grace period, this just sets a flag.
|
| 231 | * The caller must have disabled preemption.
|
| 232 |
|
| 233 | void rcu_sched_qs(void)
|
| 234 | {
|
| 235 | RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
|
| 236 | if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
|
| 237 | return;
|
| 238 | trace_rcu_grace_period(TPS("rcu_sched"),
|
| 239 | __this_cpu_read(rcu_sched_data.gpnum),
|
| 240 | TPS("cpuqs"));
|
| 241 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
|
| 242 | if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
|
| 243 | return;
|
| 244 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
|
| 245 | rcu_report_exp_rdp(&rcu_sched_state,
|
| 246 | this_cpu_ptr(&rcu_sched_data), true);
|
| 247 | }
|
| 248 |
|
| 249 | void rcu_bh_qs(void)
|
| 250 | {
|
| 251 | RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
|
| 252 | if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
|
| 253 | trace_rcu_grace_period(TPS("rcu_bh"),
|
| 254 | __this_cpu_read(rcu_bh_data.gpnum),
|
| 255 | TPS("cpuqs"));
|
| 256 | __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
|
| 257 | }
|
| 258 | }
|
| 259 |
|
| 260 | |
| 261 | * Steal a bit from the bottom of ->dynticks for idle entry/exit
|
| 262 | * control. Initially this is for TLB flushing.
|
| 263 |
|
| 264 | #define RCU_DYNTICK_CTRL_MASK 0x1
|
| 265 | #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
|
| 266 | #ifndef rcu_eqs_special_exit
|
| 267 | #define rcu_eqs_special_exit() do { } while (0)
|
| 268 | #endif
|
| 269 |
|
| 270 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
|
| 271 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
|
| 272 | .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
|
| 273 | };
|
| 274 |
|
| 275 | |
| 276 | * There's a few places, currently just in the tracing infrastructure,
|
| 277 | * that uses rcu_irq_enter() to make sure RCU is watching. But there's
|
| 278 | * a small location where that will not even work. In those cases
|
| 279 | * rcu_irq_enter_disabled() needs to be checked to make sure rcu_irq_enter()
|
| 280 | * can be called.
|
| 281 |
|
| 282 | static DEFINE_PER_CPU(bool, disable_rcu_irq_enter);
|
| 283 |
|
| 284 | bool rcu_irq_enter_disabled(void)
|
| 285 | {
|
| 286 | return this_cpu_read(disable_rcu_irq_enter);
|
| 287 | }
|
| 288 |
|
| 289 | |
| 290 | * Record entry into an extended quiescent state. This is only to be
|
| 291 | * called when not already in an extended quiescent state.
|
| 292 |
|
| 293 | static void rcu_dynticks_eqs_enter(void)
|
| 294 | {
|
| 295 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 296 | int seq;
|
| 297 |
|
| 298 | |
| 299 | * CPUs seeing atomic_add_return() must see prior RCU read-side
|
| 300 | * critical sections, and we also must force ordering with the
|
| 301 | * next idle sojourn.
|
| 302 |
|
| 303 | seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
|
| 304 |
|
| 305 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
|
| 306 | (seq & RCU_DYNTICK_CTRL_CTR));
|
| 307 |
|
| 308 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
|
| 309 | (seq & RCU_DYNTICK_CTRL_MASK));
|
| 310 | }
|
| 311 |
|
| 312 | |
| 313 | * Record exit from an extended quiescent state. This is only to be
|
| 314 | * called from an extended quiescent state.
|
| 315 |
|
| 316 | static void rcu_dynticks_eqs_exit(void)
|
| 317 | {
|
| 318 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 319 | int seq;
|
| 320 |
|
| 321 | |
| 322 | * CPUs seeing atomic_add_return() must see prior idle sojourns,
|
| 323 | * and we also must force ordering with the next RCU read-side
|
| 324 | * critical section.
|
| 325 |
|
| 326 | seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
|
| 327 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
|
| 328 | !(seq & RCU_DYNTICK_CTRL_CTR));
|
| 329 | if (seq & RCU_DYNTICK_CTRL_MASK) {
|
| 330 | atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
|
| 331 | smp_mb__after_atomic();
|
| 332 |
|
| 333 | rcu_eqs_special_exit();
|
| 334 | }
|
| 335 | }
|
| 336 |
|
| 337 | |
| 338 | * Reset the current CPU's ->dynticks counter to indicate that the
|
| 339 | * newly onlined CPU is no longer in an extended quiescent state.
|
| 340 | * This will either leave the counter unchanged, or increment it
|
| 341 | * to the next non-quiescent value.
|
| 342 | *
|
| 343 | * The non-atomic test/increment sequence works because the upper bits
|
| 344 | * of the ->dynticks counter are manipulated only by the corresponding CPU,
|
| 345 | * or when the corresponding CPU is offline.
|
| 346 |
|
| 347 | static void rcu_dynticks_eqs_online(void)
|
| 348 | {
|
| 349 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 350 |
|
| 351 | if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
|
| 352 | return;
|
| 353 | atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
|
| 354 | }
|
| 355 |
|
| 356 | |
| 357 | * Is the current CPU in an extended quiescent state?
|
| 358 | *
|
| 359 | * No ordering, as we are sampling CPU-local information.
|
| 360 |
|
| 361 | bool rcu_dynticks_curr_cpu_in_eqs(void)
|
| 362 | {
|
| 363 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 364 |
|
| 365 | return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
|
| 366 | }
|
| 367 |
|
| 368 | |
| 369 | * Snapshot the ->dynticks counter with full ordering so as to allow
|
| 370 | * stable comparison of this counter with past and future snapshots.
|
| 371 |
|
| 372 | int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
|
| 373 | {
|
| 374 | int snap = atomic_add_return(0, &rdtp->dynticks);
|
| 375 |
|
| 376 | return snap & ~RCU_DYNTICK_CTRL_MASK;
|
| 377 | }
|
| 378 |
|
| 379 | |
| 380 | * Return true if the snapshot returned from rcu_dynticks_snap()
|
| 381 | * indicates that RCU is in an extended quiescent state.
|
| 382 |
|
| 383 | static bool rcu_dynticks_in_eqs(int snap)
|
| 384 | {
|
| 385 | return !(snap & RCU_DYNTICK_CTRL_CTR);
|
| 386 | }
|
| 387 |
|
| 388 | |
| 389 | * Return true if the CPU corresponding to the specified rcu_dynticks
|
| 390 | * structure has spent some time in an extended quiescent state since
|
| 391 | * rcu_dynticks_snap() returned the specified snapshot.
|
| 392 |
|
| 393 | static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
|
| 394 | {
|
| 395 | return snap != rcu_dynticks_snap(rdtp);
|
| 396 | }
|
| 397 |
|
| 398 | |
| 399 | * Do a double-increment of the ->dynticks counter to emulate a
|
| 400 | * momentary idle-CPU quiescent state.
|
| 401 |
|
| 402 | static void rcu_dynticks_momentary_idle(void)
|
| 403 | {
|
| 404 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 405 | int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
|
| 406 | &rdtp->dynticks);
|
| 407 |
|
| 408 |
|
| 409 | WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
|
| 410 | }
|
| 411 |
|
| 412 | |
| 413 | * Set the special (bottom) bit of the specified CPU so that it
|
| 414 | * will take special action (such as flushing its TLB) on the
|
| 415 | * next exit from an extended quiescent state. Returns true if
|
| 416 | * the bit was successfully set, or false if the CPU was not in
|
| 417 | * an extended quiescent state.
|
| 418 |
|
| 419 | bool rcu_eqs_special_set(int cpu)
|
| 420 | {
|
| 421 | int old;
|
| 422 | int new;
|
| 423 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
|
| 424 |
|
| 425 | do {
|
| 426 | old = atomic_read(&rdtp->dynticks);
|
| 427 | if (old & RCU_DYNTICK_CTRL_CTR)
|
| 428 | return false;
|
| 429 | new = old | RCU_DYNTICK_CTRL_MASK;
|
| 430 | } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
|
| 431 | return true;
|
| 432 | }
|
| 433 |
|
| 434 | |
| 435 | * Let the RCU core know that this CPU has gone through the scheduler,
|
| 436 | * which is a quiescent state. This is called when the need for a
|
| 437 | * quiescent state is urgent, so we burn an atomic operation and full
|
| 438 | * memory barriers to let the RCU core know about it, regardless of what
|
| 439 | * this CPU might (or might not) do in the near future.
|
| 440 | *
|
| 441 | * We inform the RCU core by emulating a zero-duration dyntick-idle period.
|
| 442 | *
|
| 443 | * The caller must have disabled interrupts.
|
| 444 |
|
| 445 | static void rcu_momentary_dyntick_idle(void)
|
| 446 | {
|
| 447 | raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
|
| 448 | rcu_dynticks_momentary_idle();
|
| 449 | }
|
| 450 |
|
| 451 | |
| 452 | * Note a context switch. This is a quiescent state for RCU-sched,
|
| 453 | * and requires special handling for preemptible RCU.
|
| 454 | * The caller must have disabled interrupts.
|
| 455 |
|
| 456 | void rcu_note_context_switch(bool preempt)
|
| 457 | {
|
| 458 | barrier();
|
| 459 | trace_rcu_utilization(TPS("Start context switch"));
|
| 460 | rcu_sched_qs();
|
| 461 | rcu_preempt_note_context_switch(preempt);
|
| 462 |
|
| 463 | if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
|
| 464 | goto out;
|
| 465 | this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
|
| 466 | if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
|
| 467 | rcu_momentary_dyntick_idle();
|
| 468 | this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
|
| 469 | if (!preempt)
|
| 470 | rcu_note_voluntary_context_switch_lite(current);
|
| 471 | out:
|
| 472 | trace_rcu_utilization(TPS("End context switch"));
|
| 473 | barrier();
|
| 474 | }
|
| 475 | EXPORT_SYMBOL_GPL(rcu_note_context_switch);
|
| 476 |
|
| 477 | |
| 478 | * Register a quiescent state for all RCU flavors. If there is an
|
| 479 | * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
|
| 480 | * dyntick-idle quiescent state visible to other CPUs (but only for those
|
| 481 | * RCU flavors in desperate need of a quiescent state, which will normally
|
| 482 | * be none of them). Either way, do a lightweight quiescent state for
|
| 483 | * all RCU flavors.
|
| 484 | *
|
| 485 | * The barrier() calls are redundant in the common case when this is
|
| 486 | * called externally, but just in case this is called from within this
|
| 487 | * file.
|
| 488 | *
|
| 489 |
|
| 490 | void rcu_all_qs(void)
|
| 491 | {
|
| 492 | unsigned long flags;
|
| 493 |
|
| 494 | if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
|
| 495 | return;
|
| 496 | preempt_disable();
|
| 497 |
|
| 498 | if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
|
| 499 | preempt_enable();
|
| 500 | return;
|
| 501 | }
|
| 502 | this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
|
| 503 | barrier();
|
| 504 | if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
|
| 505 | local_irq_save(flags);
|
| 506 | rcu_momentary_dyntick_idle();
|
| 507 | local_irq_restore(flags);
|
| 508 | }
|
| 509 | if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
|
| 510 | rcu_sched_qs();
|
| 511 | this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
|
| 512 | barrier();
|
| 513 | preempt_enable();
|
| 514 | }
|
| 515 | EXPORT_SYMBOL_GPL(rcu_all_qs);
|
| 516 |
|
| 517 | #define DEFAULT_RCU_BLIMIT 10
|
| 518 | static long blimit = DEFAULT_RCU_BLIMIT;
|
| 519 | #define DEFAULT_RCU_QHIMARK 10000
|
| 520 | static long qhimark = DEFAULT_RCU_QHIMARK;
|
| 521 | #define DEFAULT_RCU_QLOMARK 100
|
| 522 | static long qlowmark = DEFAULT_RCU_QLOMARK;
|
| 523 |
|
| 524 | module_param(blimit, long, 0444);
|
| 525 | module_param(qhimark, long, 0444);
|
| 526 | module_param(qlowmark, long, 0444);
|
| 527 |
|
| 528 | static ulong jiffies_till_first_fqs = ULONG_MAX;
|
| 529 | static ulong jiffies_till_next_fqs = ULONG_MAX;
|
| 530 | static bool rcu_kick_kthreads;
|
| 531 |
|
| 532 | module_param(jiffies_till_first_fqs, ulong, 0644);
|
| 533 | module_param(jiffies_till_next_fqs, ulong, 0644);
|
| 534 | module_param(rcu_kick_kthreads, bool, 0644);
|
| 535 |
|
| 536 | |
| 537 | * How long the grace period must be before we start recruiting
|
| 538 | * quiescent-state help from rcu_note_context_switch().
|
| 539 |
|
| 540 | static ulong jiffies_till_sched_qs = HZ / 20;
|
| 541 | module_param(jiffies_till_sched_qs, ulong, 0644);
|
| 542 |
|
| 543 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 544 | struct rcu_data *rdp);
|
| 545 | static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
|
| 546 | static void force_quiescent_state(struct rcu_state *rsp);
|
| 547 | static int rcu_pending(void);
|
| 548 |
|
| 549 | |
| 550 | * Return the number of RCU batches started thus far for debug & stats.
|
| 551 |
|
| 552 | unsigned long rcu_batches_started(void)
|
| 553 | {
|
| 554 | return rcu_state_p->gpnum;
|
| 555 | }
|
| 556 | EXPORT_SYMBOL_GPL(rcu_batches_started);
|
| 557 |
|
| 558 | |
| 559 | * Return the number of RCU-sched batches started thus far for debug & stats.
|
| 560 |
|
| 561 | unsigned long rcu_batches_started_sched(void)
|
| 562 | {
|
| 563 | return rcu_sched_state.gpnum;
|
| 564 | }
|
| 565 | EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
|
| 566 |
|
| 567 | |
| 568 | * Return the number of RCU BH batches started thus far for debug & stats.
|
| 569 |
|
| 570 | unsigned long rcu_batches_started_bh(void)
|
| 571 | {
|
| 572 | return rcu_bh_state.gpnum;
|
| 573 | }
|
| 574 | EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
|
| 575 |
|
| 576 | |
| 577 | * Return the number of RCU batches completed thus far for debug & stats.
|
| 578 |
|
| 579 | unsigned long rcu_batches_completed(void)
|
| 580 | {
|
| 581 | return rcu_state_p->completed;
|
| 582 | }
|
| 583 | EXPORT_SYMBOL_GPL(rcu_batches_completed);
|
| 584 |
|
| 585 | |
| 586 | * Return the number of RCU-sched batches completed thus far for debug & stats.
|
| 587 |
|
| 588 | unsigned long rcu_batches_completed_sched(void)
|
| 589 | {
|
| 590 | return rcu_sched_state.completed;
|
| 591 | }
|
| 592 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
|
| 593 |
|
| 594 | |
| 595 | * Return the number of RCU BH batches completed thus far for debug & stats.
|
| 596 |
|
| 597 | unsigned long rcu_batches_completed_bh(void)
|
| 598 | {
|
| 599 | return rcu_bh_state.completed;
|
| 600 | }
|
| 601 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
|
| 602 |
|
| 603 | |
| 604 | * Return the number of RCU expedited batches completed thus far for
|
| 605 | * debug & stats. Odd numbers mean that a batch is in progress, even
|
| 606 | * numbers mean idle. The value returned will thus be roughly double
|
| 607 | * the cumulative batches since boot.
|
| 608 |
|
| 609 | unsigned long rcu_exp_batches_completed(void)
|
| 610 | {
|
| 611 | return rcu_state_p->expedited_sequence;
|
| 612 | }
|
| 613 | EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
|
| 614 |
|
| 615 | |
| 616 | * Return the number of RCU-sched expedited batches completed thus far
|
| 617 | * for debug & stats. Similar to rcu_exp_batches_completed().
|
| 618 |
|
| 619 | unsigned long rcu_exp_batches_completed_sched(void)
|
| 620 | {
|
| 621 | return rcu_sched_state.expedited_sequence;
|
| 622 | }
|
| 623 | EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
|
| 624 |
|
| 625 | |
| 626 | * Force a quiescent state.
|
| 627 |
|
| 628 | void rcu_force_quiescent_state(void)
|
| 629 | {
|
| 630 | force_quiescent_state(rcu_state_p);
|
| 631 | }
|
| 632 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
|
| 633 |
|
| 634 | |
| 635 | * Force a quiescent state for RCU BH.
|
| 636 |
|
| 637 | void rcu_bh_force_quiescent_state(void)
|
| 638 | {
|
| 639 | force_quiescent_state(&rcu_bh_state);
|
| 640 | }
|
| 641 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
|
| 642 |
|
| 643 | |
| 644 | * Force a quiescent state for RCU-sched.
|
| 645 |
|
| 646 | void rcu_sched_force_quiescent_state(void)
|
| 647 | {
|
| 648 | force_quiescent_state(&rcu_sched_state);
|
| 649 | }
|
| 650 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
|
| 651 |
|
| 652 | |
| 653 | * Show the state of the grace-period kthreads.
|
| 654 |
|
| 655 | void show_rcu_gp_kthreads(void)
|
| 656 | {
|
| 657 | struct rcu_state *rsp;
|
| 658 |
|
| 659 | for_each_rcu_flavor(rsp) {
|
| 660 | pr_info("%s: wait state: %d ->state: %#lx\n",
|
| 661 | rsp->name, rsp->gp_state, rsp->gp_kthread->state);
|
| 662 |
|
| 663 | }
|
| 664 | }
|
| 665 | EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
|
| 666 |
|
| 667 | |
| 668 | * Record the number of times rcutorture tests have been initiated and
|
| 669 | * terminated. This information allows the debugfs tracing stats to be
|
| 670 | * correlated to the rcutorture messages, even when the rcutorture module
|
| 671 | * is being repeatedly loaded and unloaded. In other words, we cannot
|
| 672 | * store this state in rcutorture itself.
|
| 673 |
|
| 674 | void rcutorture_record_test_transition(void)
|
| 675 | {
|
| 676 | rcutorture_testseq++;
|
| 677 | rcutorture_vernum = 0;
|
| 678 | }
|
| 679 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
|
| 680 |
|
| 681 | |
| 682 | * Send along grace-period-related data for rcutorture diagnostics.
|
| 683 |
|
| 684 | void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
|
| 685 | unsigned long *gpnum, unsigned long *completed)
|
| 686 | {
|
| 687 | struct rcu_state *rsp = NULL;
|
| 688 |
|
| 689 | switch (test_type) {
|
| 690 | case RCU_FLAVOR:
|
| 691 | rsp = rcu_state_p;
|
| 692 | break;
|
| 693 | case RCU_BH_FLAVOR:
|
| 694 | rsp = &rcu_bh_state;
|
| 695 | break;
|
| 696 | case RCU_SCHED_FLAVOR:
|
| 697 | rsp = &rcu_sched_state;
|
| 698 | break;
|
| 699 | default:
|
| 700 | break;
|
| 701 | }
|
| 702 | if (rsp == NULL)
|
| 703 | return;
|
| 704 | *flags = READ_ONCE(rsp->gp_flags);
|
| 705 | *gpnum = READ_ONCE(rsp->gpnum);
|
| 706 | *completed = READ_ONCE(rsp->completed);
|
| 707 | }
|
| 708 | EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
|
| 709 |
|
| 710 | |
| 711 | * Record the number of writer passes through the current rcutorture test.
|
| 712 | * This is also used to correlate debugfs tracing stats with the rcutorture
|
| 713 | * messages.
|
| 714 |
|
| 715 | void rcutorture_record_progress(unsigned long vernum)
|
| 716 | {
|
| 717 | rcutorture_vernum++;
|
| 718 | }
|
| 719 | EXPORT_SYMBOL_GPL(rcutorture_record_progress);
|
| 720 |
|
| 721 | |
| 722 | * Return the root node of the specified rcu_state structure.
|
| 723 |
|
| 724 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
|
| 725 | {
|
| 726 | return &rsp->node[0];
|
| 727 | }
|
| 728 |
|
| 729 | |
| 730 | * Is there any need for future grace periods?
|
| 731 | * Interrupts must be disabled. If the caller does not hold the root
|
| 732 | * rnp_node structure's ->lock, the results are advisory only.
|
| 733 |
|
| 734 | static int rcu_future_needs_gp(struct rcu_state *rsp)
|
| 735 | {
|
| 736 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 737 | int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
|
| 738 | int *fp = &rnp->need_future_gp[idx];
|
| 739 |
|
| 740 | RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_future_needs_gp() invoked with irqs enabled!!!");
|
| 741 | return READ_ONCE(*fp);
|
| 742 | }
|
| 743 |
|
| 744 | |
| 745 | * Does the current CPU require a not-yet-started grace period?
|
| 746 | * The caller must have disabled interrupts to prevent races with
|
| 747 | * normal callback registry.
|
| 748 |
|
| 749 | static bool
|
| 750 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 751 | {
|
| 752 | RCU_LOCKDEP_WARN(!irqs_disabled(), "cpu_needs_another_gp() invoked with irqs enabled!!!");
|
| 753 | if (rcu_gp_in_progress(rsp))
|
| 754 | return false;
|
| 755 | if (rcu_future_needs_gp(rsp))
|
| 756 | return true;
|
| 757 | if (!rcu_segcblist_is_enabled(&rdp->cblist))
|
| 758 | return false;
|
| 759 | if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
|
| 760 | return true;
|
| 761 | if (rcu_segcblist_future_gp_needed(&rdp->cblist,
|
| 762 | READ_ONCE(rsp->completed)))
|
| 763 | return true;
|
| 764 | return false;
|
| 765 | }
|
| 766 |
|
| 767 | |
| 768 | * rcu_eqs_enter_common - current CPU is entering an extended quiescent state
|
| 769 | *
|
| 770 | * Enter idle, doing appropriate accounting. The caller must have
|
| 771 | * disabled interrupts.
|
| 772 |
|
| 773 | static void rcu_eqs_enter_common(bool user)
|
| 774 | {
|
| 775 | struct rcu_state *rsp;
|
| 776 | struct rcu_data *rdp;
|
| 777 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 778 |
|
| 779 | RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_eqs_enter_common() invoked with irqs enabled!!!");
|
| 780 | trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0);
|
| 781 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
|
| 782 | !user && !is_idle_task(current)) {
|
| 783 | struct task_struct *idle __maybe_unused =
|
| 784 | idle_task(smp_processor_id());
|
| 785 |
|
| 786 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), rdtp->dynticks_nesting, 0);
|
| 787 | rcu_ftrace_dump(DUMP_ORIG);
|
| 788 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
|
| 789 | current->pid, current->comm,
|
| 790 | idle->pid, idle->comm);
|
| 791 | }
|
| 792 | for_each_rcu_flavor(rsp) {
|
| 793 | rdp = this_cpu_ptr(rsp->rda);
|
| 794 | do_nocb_deferred_wakeup(rdp);
|
| 795 | }
|
| 796 | rcu_prepare_for_idle();
|
| 797 | __this_cpu_inc(disable_rcu_irq_enter);
|
| 798 | rdtp->dynticks_nesting = 0;
|
| 799 | rcu_dynticks_eqs_enter();
|
| 800 | __this_cpu_dec(disable_rcu_irq_enter);
|
| 801 | rcu_dynticks_task_enter();
|
| 802 |
|
| 803 | |
| 804 | * It is illegal to enter an extended quiescent state while
|
| 805 | * in an RCU read-side critical section.
|
| 806 |
|
| 807 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
|
| 808 | "Illegal idle entry in RCU read-side critical section.");
|
| 809 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
|
| 810 | "Illegal idle entry in RCU-bh read-side critical section.");
|
| 811 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
|
| 812 | "Illegal idle entry in RCU-sched read-side critical section.");
|
| 813 | }
|
| 814 |
|
| 815 | |
| 816 | * Enter an RCU extended quiescent state, which can be either the
|
| 817 | * idle loop or adaptive-tickless usermode execution.
|
| 818 |
|
| 819 | static void rcu_eqs_enter(bool user)
|
| 820 | {
|
| 821 | struct rcu_dynticks *rdtp;
|
| 822 |
|
| 823 | rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 824 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
|
| 825 | (rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == 0);
|
| 826 | if ((rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
|
| 827 | rcu_eqs_enter_common(user);
|
| 828 | else
|
| 829 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
|
| 830 | }
|
| 831 |
|
| 832 | |
| 833 | * rcu_idle_enter - inform RCU that current CPU is entering idle
|
| 834 | *
|
| 835 | * Enter idle mode, in other words, -leave- the mode in which RCU
|
| 836 | * read-side critical sections can occur. (Though RCU read-side
|
| 837 | * critical sections can occur in irq handlers in idle, a possibility
|
| 838 | * handled by irq_enter() and irq_exit().)
|
| 839 | *
|
| 840 | * We crowbar the ->dynticks_nesting field to zero to allow for
|
| 841 | * the possibility of usermode upcalls having messed up our count
|
| 842 | * of interrupt nesting level during the prior busy period.
|
| 843 |
|
| 844 | void rcu_idle_enter(void)
|
| 845 | {
|
| 846 | unsigned long flags;
|
| 847 |
|
| 848 | local_irq_save(flags);
|
| 849 | rcu_eqs_enter(false);
|
| 850 | local_irq_restore(flags);
|
| 851 | }
|
| 852 | EXPORT_SYMBOL_GPL(rcu_idle_enter);
|
| 853 |
|
| 854 | #ifdef CONFIG_NO_HZ_FULL
|
| 855 | |
| 856 | * rcu_user_enter - inform RCU that we are resuming userspace.
|
| 857 | *
|
| 858 | * Enter RCU idle mode right before resuming userspace. No use of RCU
|
| 859 | * is permitted between this call and rcu_user_exit(). This way the
|
| 860 | * CPU doesn't need to maintain the tick for RCU maintenance purposes
|
| 861 | * when the CPU runs in userspace.
|
| 862 |
|
| 863 | void rcu_user_enter(void)
|
| 864 | {
|
| 865 | rcu_eqs_enter(1);
|
| 866 | }
|
| 867 | #endif
|
| 868 |
|
| 869 | |
| 870 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
|
| 871 | *
|
| 872 | * Exit from an interrupt handler, which might possibly result in entering
|
| 873 | * idle mode, in other words, leaving the mode in which read-side critical
|
| 874 | * sections can occur. The caller must have disabled interrupts.
|
| 875 | *
|
| 876 | * This code assumes that the idle loop never does anything that might
|
| 877 | * result in unbalanced calls to irq_enter() and irq_exit(). If your
|
| 878 | * architecture violates this assumption, RCU will give you what you
|
| 879 | * deserve, good and hard. But very infrequently and irreproducibly.
|
| 880 | *
|
| 881 | * Use things like work queues to work around this limitation.
|
| 882 | *
|
| 883 | * You have been warned.
|
| 884 |
|
| 885 | void rcu_irq_exit(void)
|
| 886 | {
|
| 887 | struct rcu_dynticks *rdtp;
|
| 888 |
|
| 889 | RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
|
| 890 | rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 891 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
|
| 892 | rdtp->dynticks_nesting < 1);
|
| 893 | if (rdtp->dynticks_nesting <= 1) {
|
| 894 | rcu_eqs_enter_common(true);
|
| 895 | } else {
|
| 896 | trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nesting, rdtp->dynticks_nesting - 1);
|
| 897 | rdtp->dynticks_nesting--;
|
| 898 | }
|
| 899 | }
|
| 900 |
|
| 901 | |
| 902 | * Wrapper for rcu_irq_exit() where interrupts are enabled.
|
| 903 |
|
| 904 | void rcu_irq_exit_irqson(void)
|
| 905 | {
|
| 906 | unsigned long flags;
|
| 907 |
|
| 908 | local_irq_save(flags);
|
| 909 | rcu_irq_exit();
|
| 910 | local_irq_restore(flags);
|
| 911 | }
|
| 912 |
|
| 913 | |
| 914 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
|
| 915 | *
|
| 916 | * If the new value of the ->dynticks_nesting counter was previously zero,
|
| 917 | * we really have exited idle, and must do the appropriate accounting.
|
| 918 | * The caller must have disabled interrupts.
|
| 919 |
|
| 920 | static void rcu_eqs_exit_common(long long oldval, int user)
|
| 921 | {
|
| 922 | RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
|
| 923 |
|
| 924 | rcu_dynticks_task_exit();
|
| 925 | rcu_dynticks_eqs_exit();
|
| 926 | rcu_cleanup_after_idle();
|
| 927 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
|
| 928 | if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
|
| 929 | !user && !is_idle_task(current)) {
|
| 930 | struct task_struct *idle __maybe_unused =
|
| 931 | idle_task(smp_processor_id());
|
| 932 |
|
| 933 | trace_rcu_dyntick(TPS("Error on exit: not idle task"),
|
| 934 | oldval, rdtp->dynticks_nesting);
|
| 935 | rcu_ftrace_dump(DUMP_ORIG);
|
| 936 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
|
| 937 | current->pid, current->comm,
|
| 938 | idle->pid, idle->comm);
|
| 939 | }
|
| 940 | }
|
| 941 |
|
| 942 | |
| 943 | * Exit an RCU extended quiescent state, which can be either the
|
| 944 | * idle loop or adaptive-tickless usermode execution.
|
| 945 |
|
| 946 | static void rcu_eqs_exit(bool user)
|
| 947 | {
|
| 948 | struct rcu_dynticks *rdtp;
|
| 949 | long long oldval;
|
| 950 |
|
| 951 | RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_eqs_exit() invoked with irqs enabled!!!");
|
| 952 | rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 953 | oldval = rdtp->dynticks_nesting;
|
| 954 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
|
| 955 | if (oldval & DYNTICK_TASK_NEST_MASK) {
|
| 956 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
|
| 957 | } else {
|
| 958 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
|
| 959 | rcu_eqs_exit_common(oldval, user);
|
| 960 | }
|
| 961 | }
|
| 962 |
|
| 963 | |
| 964 | * rcu_idle_exit - inform RCU that current CPU is leaving idle
|
| 965 | *
|
| 966 | * Exit idle mode, in other words, -enter- the mode in which RCU
|
| 967 | * read-side critical sections can occur.
|
| 968 | *
|
| 969 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
|
| 970 | * allow for the possibility of usermode upcalls messing up our count
|
| 971 | * of interrupt nesting level during the busy period that is just
|
| 972 | * now starting.
|
| 973 |
|
| 974 | void rcu_idle_exit(void)
|
| 975 | {
|
| 976 | unsigned long flags;
|
| 977 |
|
| 978 | local_irq_save(flags);
|
| 979 | rcu_eqs_exit(false);
|
| 980 | local_irq_restore(flags);
|
| 981 | }
|
| 982 | EXPORT_SYMBOL_GPL(rcu_idle_exit);
|
| 983 |
|
| 984 | #ifdef CONFIG_NO_HZ_FULL
|
| 985 | |
| 986 | * rcu_user_exit - inform RCU that we are exiting userspace.
|
| 987 | *
|
| 988 | * Exit RCU idle mode while entering the kernel because it can
|
| 989 | * run a RCU read side critical section anytime.
|
| 990 |
|
| 991 | void rcu_user_exit(void)
|
| 992 | {
|
| 993 | rcu_eqs_exit(1);
|
| 994 | }
|
| 995 | #endif
|
| 996 |
|
| 997 | |
| 998 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
|
| 999 | *
|
| 1000 | * Enter an interrupt handler, which might possibly result in exiting
|
| 1001 | * idle mode, in other words, entering the mode in which read-side critical
|
| 1002 | * sections can occur. The caller must have disabled interrupts.
|
| 1003 | *
|
| 1004 | * Note that the Linux kernel is fully capable of entering an interrupt
|
| 1005 | * handler that it never exits, for example when doing upcalls to
|
| 1006 | * user mode! This code assumes that the idle loop never does upcalls to
|
| 1007 | * user mode. If your architecture does do upcalls from the idle loop (or
|
| 1008 | * does anything else that results in unbalanced calls to the irq_enter()
|
| 1009 | * and irq_exit() functions), RCU will give you what you deserve, good
|
| 1010 | * and hard. But very infrequently and irreproducibly.
|
| 1011 | *
|
| 1012 | * Use things like work queues to work around this limitation.
|
| 1013 | *
|
| 1014 | * You have been warned.
|
| 1015 |
|
| 1016 | void rcu_irq_enter(void)
|
| 1017 | {
|
| 1018 | struct rcu_dynticks *rdtp;
|
| 1019 | long long oldval;
|
| 1020 |
|
| 1021 | RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
|
| 1022 | rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 1023 | oldval = rdtp->dynticks_nesting;
|
| 1024 | rdtp->dynticks_nesting++;
|
| 1025 | WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
|
| 1026 | rdtp->dynticks_nesting == 0);
|
| 1027 | if (oldval)
|
| 1028 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
|
| 1029 | else
|
| 1030 | rcu_eqs_exit_common(oldval, true);
|
| 1031 | }
|
| 1032 |
|
| 1033 | |
| 1034 | * Wrapper for rcu_irq_enter() where interrupts are enabled.
|
| 1035 |
|
| 1036 | void rcu_irq_enter_irqson(void)
|
| 1037 | {
|
| 1038 | unsigned long flags;
|
| 1039 |
|
| 1040 | local_irq_save(flags);
|
| 1041 | rcu_irq_enter();
|
| 1042 | local_irq_restore(flags);
|
| 1043 | }
|
| 1044 |
|
| 1045 | |
| 1046 | * rcu_nmi_enter - inform RCU of entry to NMI context
|
| 1047 | *
|
| 1048 | * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
|
| 1049 | * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
|
| 1050 | * that the CPU is active. This implementation permits nested NMIs, as
|
| 1051 | * long as the nesting level does not overflow an int. (You will probably
|
| 1052 | * run out of stack space first.)
|
| 1053 |
|
| 1054 | void rcu_nmi_enter(void)
|
| 1055 | {
|
| 1056 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 1057 | int incby = 2;
|
| 1058 |
|
| 1059 |
|
| 1060 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
|
| 1061 |
|
| 1062 | |
| 1063 | * If idle from RCU viewpoint, atomically increment ->dynticks
|
| 1064 | * to mark non-idle and increment ->dynticks_nmi_nesting by one.
|
| 1065 | * Otherwise, increment ->dynticks_nmi_nesting by two. This means
|
| 1066 | * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
|
| 1067 | * to be in the outermost NMI handler that interrupted an RCU-idle
|
| 1068 | * period (observation due to Andy Lutomirski).
|
| 1069 |
|
| 1070 | if (rcu_dynticks_curr_cpu_in_eqs()) {
|
| 1071 | rcu_dynticks_eqs_exit();
|
| 1072 | incby = 1;
|
| 1073 | }
|
| 1074 | rdtp->dynticks_nmi_nesting += incby;
|
| 1075 | barrier();
|
| 1076 | }
|
| 1077 |
|
| 1078 | |
| 1079 | * rcu_nmi_exit - inform RCU of exit from NMI context
|
| 1080 | *
|
| 1081 | * If we are returning from the outermost NMI handler that interrupted an
|
| 1082 | * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
|
| 1083 | * to let the RCU grace-period handling know that the CPU is back to
|
| 1084 | * being RCU-idle.
|
| 1085 |
|
| 1086 | void rcu_nmi_exit(void)
|
| 1087 | {
|
| 1088 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 1089 |
|
| 1090 | |
| 1091 | * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
|
| 1092 | * (We are exiting an NMI handler, so RCU better be paying attention
|
| 1093 | * to us!)
|
| 1094 |
|
| 1095 | WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
|
| 1096 | WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
|
| 1097 |
|
| 1098 | |
| 1099 | * If the nesting level is not 1, the CPU wasn't RCU-idle, so
|
| 1100 | * leave it in non-RCU-idle state.
|
| 1101 |
|
| 1102 | if (rdtp->dynticks_nmi_nesting != 1) {
|
| 1103 | rdtp->dynticks_nmi_nesting -= 2;
|
| 1104 | return;
|
| 1105 | }
|
| 1106 |
|
| 1107 |
|
| 1108 | rdtp->dynticks_nmi_nesting = 0;
|
| 1109 | rcu_dynticks_eqs_enter();
|
| 1110 | }
|
| 1111 |
|
| 1112 | |
| 1113 | * rcu_is_watching - see if RCU thinks that the current CPU is idle
|
| 1114 | *
|
| 1115 | * Return true if RCU is watching the running CPU, which means that this
|
| 1116 | * CPU can safely enter RCU read-side critical sections. In other words,
|
| 1117 | * if the current CPU is in its idle loop and is neither in an interrupt
|
| 1118 | * or NMI handler, return true.
|
| 1119 |
|
| 1120 | bool notrace rcu_is_watching(void)
|
| 1121 | {
|
| 1122 | bool ret;
|
| 1123 |
|
| 1124 | preempt_disable_notrace();
|
| 1125 | ret = !rcu_dynticks_curr_cpu_in_eqs();
|
| 1126 | preempt_enable_notrace();
|
| 1127 | return ret;
|
| 1128 | }
|
| 1129 | EXPORT_SYMBOL_GPL(rcu_is_watching);
|
| 1130 |
|
| 1131 | |
| 1132 | * If a holdout task is actually running, request an urgent quiescent
|
| 1133 | * state from its CPU. This is unsynchronized, so migrations can cause
|
| 1134 | * the request to go to the wrong CPU. Which is OK, all that will happen
|
| 1135 | * is that the CPU's next context switch will be a bit slower and next
|
| 1136 | * time around this task will generate another request.
|
| 1137 |
|
| 1138 | void rcu_request_urgent_qs_task(struct task_struct *t)
|
| 1139 | {
|
| 1140 | int cpu;
|
| 1141 |
|
| 1142 | barrier();
|
| 1143 | cpu = task_cpu(t);
|
| 1144 | if (!task_curr(t))
|
| 1145 | return;
|
| 1146 | smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
|
| 1147 | }
|
| 1148 |
|
| 1149 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
|
| 1150 |
|
| 1151 | |
| 1152 | * Is the current CPU online? Disable preemption to avoid false positives
|
| 1153 | * that could otherwise happen due to the current CPU number being sampled,
|
| 1154 | * this task being preempted, its old CPU being taken offline, resuming
|
| 1155 | * on some other CPU, then determining that its old CPU is now offline.
|
| 1156 | * It is OK to use RCU on an offline processor during initial boot, hence
|
| 1157 | * the check for rcu_scheduler_fully_active. Note also that it is OK
|
| 1158 | * for a CPU coming online to use RCU for one jiffy prior to marking itself
|
| 1159 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going
|
| 1160 | * offline to continue to use RCU for one jiffy after marking itself
|
| 1161 | * offline in the cpu_online_mask. This leniency is necessary given the
|
| 1162 | * non-atomic nature of the online and offline processing, for example,
|
| 1163 | * the fact that a CPU enters the scheduler after completing the teardown
|
| 1164 | * of the CPU.
|
| 1165 | *
|
| 1166 | * This is also why RCU internally marks CPUs online during in the
|
| 1167 | * preparation phase and offline after the CPU has been taken down.
|
| 1168 | *
|
| 1169 | * Disable checking if in an NMI handler because we cannot safely report
|
| 1170 | * errors from NMI handlers anyway.
|
| 1171 |
|
| 1172 | bool rcu_lockdep_current_cpu_online(void)
|
| 1173 | {
|
| 1174 | struct rcu_data *rdp;
|
| 1175 | struct rcu_node *rnp;
|
| 1176 | bool ret;
|
| 1177 |
|
| 1178 | if (in_nmi())
|
| 1179 | return true;
|
| 1180 | preempt_disable();
|
| 1181 | rdp = this_cpu_ptr(&rcu_sched_data);
|
| 1182 | rnp = rdp->mynode;
|
| 1183 | ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
|
| 1184 | !rcu_scheduler_fully_active;
|
| 1185 | preempt_enable();
|
| 1186 | return ret;
|
| 1187 | }
|
| 1188 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
|
| 1189 |
|
| 1190 | #endif
|
| 1191 |
|
| 1192 | |
| 1193 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
|
| 1194 | *
|
| 1195 | * If the current CPU is idle or running at a first-level (not nested)
|
| 1196 | * interrupt from idle, return true. The caller must have at least
|
| 1197 | * disabled preemption.
|
| 1198 |
|
| 1199 | static int rcu_is_cpu_rrupt_from_idle(void)
|
| 1200 | {
|
| 1201 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
|
| 1202 | }
|
| 1203 |
|
| 1204 | |
| 1205 | * Snapshot the specified CPU's dynticks counter so that we can later
|
| 1206 | * credit them with an implicit quiescent state. Return 1 if this CPU
|
| 1207 | * is in dynticks idle mode, which is an extended quiescent state.
|
| 1208 |
|
| 1209 | static int dyntick_save_progress_counter(struct rcu_data *rdp)
|
| 1210 | {
|
| 1211 | rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
|
| 1212 | if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
|
| 1213 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
|
| 1214 | if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
|
| 1215 | rdp->mynode->gpnum))
|
| 1216 | WRITE_ONCE(rdp->gpwrap, true);
|
| 1217 | return 1;
|
| 1218 | }
|
| 1219 | return 0;
|
| 1220 | }
|
| 1221 |
|
| 1222 | |
| 1223 | * Return true if the specified CPU has passed through a quiescent
|
| 1224 | * state by virtue of being in or having passed through an dynticks
|
| 1225 | * idle state since the last call to dyntick_save_progress_counter()
|
| 1226 | * for this same CPU, or by virtue of having been offline.
|
| 1227 |
|
| 1228 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
|
| 1229 | {
|
| 1230 | unsigned long jtsq;
|
| 1231 | bool *rnhqp;
|
| 1232 | bool *ruqp;
|
| 1233 | unsigned long rjtsc;
|
| 1234 | struct rcu_node *rnp;
|
| 1235 |
|
| 1236 | |
| 1237 | * If the CPU passed through or entered a dynticks idle phase with
|
| 1238 | * no active irq/NMI handlers, then we can safely pretend that the CPU
|
| 1239 | * already acknowledged the request to pass through a quiescent
|
| 1240 | * state. Either way, that CPU cannot possibly be in an RCU
|
| 1241 | * read-side critical section that started before the beginning
|
| 1242 | * of the current RCU grace period.
|
| 1243 |
|
| 1244 | if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
|
| 1245 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
|
| 1246 | rdp->dynticks_fqs++;
|
| 1247 | return 1;
|
| 1248 | }
|
| 1249 |
|
| 1250 |
|
| 1251 | jtsq = jiffies_till_sched_qs;
|
| 1252 | rjtsc = rcu_jiffies_till_stall_check();
|
| 1253 | if (jtsq > rjtsc / 2) {
|
| 1254 | WRITE_ONCE(jiffies_till_sched_qs, rjtsc);
|
| 1255 | jtsq = rjtsc / 2;
|
| 1256 | } else if (jtsq < 1) {
|
| 1257 | WRITE_ONCE(jiffies_till_sched_qs, 1);
|
| 1258 | jtsq = 1;
|
| 1259 | }
|
| 1260 |
|
| 1261 | |
| 1262 | * Has this CPU encountered a cond_resched_rcu_qs() since the
|
| 1263 | * beginning of the grace period? For this to be the case,
|
| 1264 | * the CPU has to have noticed the current grace period. This
|
| 1265 | * might not be the case for nohz_full CPUs looping in the kernel.
|
| 1266 |
|
| 1267 | rnp = rdp->mynode;
|
| 1268 | ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
|
| 1269 | if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
|
| 1270 | READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
|
| 1271 | READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
|
| 1272 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
|
| 1273 | return 1;
|
| 1274 | } else {
|
| 1275 |
|
| 1276 | smp_store_release(ruqp, true);
|
| 1277 | }
|
| 1278 |
|
| 1279 |
|
| 1280 | if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
|
| 1281 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
|
| 1282 | rdp->offline_fqs++;
|
| 1283 | return 1;
|
| 1284 | }
|
| 1285 |
|
| 1286 | |
| 1287 | * A CPU running for an extended time within the kernel can
|
| 1288 | * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
|
| 1289 | * even context-switching back and forth between a pair of
|
| 1290 | * in-kernel CPU-bound tasks cannot advance grace periods.
|
| 1291 | * So if the grace period is old enough, make the CPU pay attention.
|
| 1292 | * Note that the unsynchronized assignments to the per-CPU
|
| 1293 | * rcu_need_heavy_qs variable are safe. Yes, setting of
|
| 1294 | * bits can be lost, but they will be set again on the next
|
| 1295 | * force-quiescent-state pass. So lost bit sets do not result
|
| 1296 | * in incorrect behavior, merely in a grace period lasting
|
| 1297 | * a few jiffies longer than it might otherwise. Because
|
| 1298 | * there are at most four threads involved, and because the
|
| 1299 | * updates are only once every few jiffies, the probability of
|
| 1300 | * lossage (and thus of slight grace-period extension) is
|
| 1301 | * quite low.
|
| 1302 | *
|
| 1303 | * Note that if the jiffies_till_sched_qs boot/sysfs parameter
|
| 1304 | * is set too high, we override with half of the RCU CPU stall
|
| 1305 | * warning delay.
|
| 1306 |
|
| 1307 | rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
|
| 1308 | if (!READ_ONCE(*rnhqp) &&
|
| 1309 | (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
|
| 1310 | time_after(jiffies, rdp->rsp->jiffies_resched))) {
|
| 1311 | WRITE_ONCE(*rnhqp, true);
|
| 1312 |
|
| 1313 | smp_store_release(ruqp, true);
|
| 1314 | rdp->rsp->jiffies_resched += 5;
|
| 1315 | }
|
| 1316 |
|
| 1317 | |
| 1318 | * If more than halfway to RCU CPU stall-warning time, do
|
| 1319 | * a resched_cpu() to try to loosen things up a bit.
|
| 1320 |
|
| 1321 | if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2)
|
| 1322 | resched_cpu(rdp->cpu);
|
| 1323 |
|
| 1324 | return 0;
|
| 1325 | }
|
| 1326 |
|
| 1327 | static void record_gp_stall_check_time(struct rcu_state *rsp)
|
| 1328 | {
|
| 1329 | unsigned long j = jiffies;
|
| 1330 | unsigned long j1;
|
| 1331 |
|
| 1332 | rsp->gp_start = j;
|
| 1333 | smp_wmb();
|
| 1334 | j1 = rcu_jiffies_till_stall_check();
|
| 1335 | WRITE_ONCE(rsp->jiffies_stall, j + j1);
|
| 1336 | rsp->jiffies_resched = j + j1 / 2;
|
| 1337 | rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
|
| 1338 | }
|
| 1339 |
|
| 1340 | |
| 1341 | * Convert a ->gp_state value to a character string.
|
| 1342 |
|
| 1343 | static const char *gp_state_getname(short gs)
|
| 1344 | {
|
| 1345 | if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
|
| 1346 | return "???";
|
| 1347 | return gp_state_names[gs];
|
| 1348 | }
|
| 1349 |
|
| 1350 | |
| 1351 | * Complain about starvation of grace-period kthread.
|
| 1352 |
|
| 1353 | static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
|
| 1354 | {
|
| 1355 | unsigned long gpa;
|
| 1356 | unsigned long j;
|
| 1357 |
|
| 1358 | j = jiffies;
|
| 1359 | gpa = READ_ONCE(rsp->gp_activity);
|
| 1360 | if (j - gpa > 2 * HZ) {
|
| 1361 | pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
|
| 1362 | rsp->name, j - gpa,
|
| 1363 | rsp->gpnum, rsp->completed,
|
| 1364 | rsp->gp_flags,
|
| 1365 | gp_state_getname(rsp->gp_state), rsp->gp_state,
|
| 1366 | rsp->gp_kthread ? rsp->gp_kthread->state : ~0);
|
| 1367 | if (rsp->gp_kthread) {
|
| 1368 | sched_show_task(rsp->gp_kthread);
|
| 1369 | wake_up_process(rsp->gp_kthread);
|
| 1370 | }
|
| 1371 | }
|
| 1372 | }
|
| 1373 |
|
| 1374 | |
| 1375 | * Dump stacks of all tasks running on stalled CPUs. First try using
|
| 1376 | * NMIs, but fall back to manual remote stack tracing on architectures
|
| 1377 | * that don't support NMI-based stack dumps. The NMI-triggered stack
|
| 1378 | * traces are more accurate because they are printed by the target CPU.
|
| 1379 |
|
| 1380 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
|
| 1381 | {
|
| 1382 | int cpu;
|
| 1383 | unsigned long flags;
|
| 1384 | struct rcu_node *rnp;
|
| 1385 |
|
| 1386 | rcu_for_each_leaf_node(rsp, rnp) {
|
| 1387 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 1388 | for_each_leaf_node_possible_cpu(rnp, cpu)
|
| 1389 | if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
|
| 1390 | if (!trigger_single_cpu_backtrace(cpu))
|
| 1391 | dump_cpu_task(cpu);
|
| 1392 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 1393 | }
|
| 1394 | }
|
| 1395 |
|
| 1396 | |
| 1397 | * If too much time has passed in the current grace period, and if
|
| 1398 | * so configured, go kick the relevant kthreads.
|
| 1399 |
|
| 1400 | static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
|
| 1401 | {
|
| 1402 | unsigned long j;
|
| 1403 |
|
| 1404 | if (!rcu_kick_kthreads)
|
| 1405 | return;
|
| 1406 | j = READ_ONCE(rsp->jiffies_kick_kthreads);
|
| 1407 | if (time_after(jiffies, j) && rsp->gp_kthread &&
|
| 1408 | (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
|
| 1409 | WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
|
| 1410 | rcu_ftrace_dump(DUMP_ALL);
|
| 1411 | wake_up_process(rsp->gp_kthread);
|
| 1412 | WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
|
| 1413 | }
|
| 1414 | }
|
| 1415 |
|
| 1416 | static inline void panic_on_rcu_stall(void)
|
| 1417 | {
|
| 1418 | if (sysctl_panic_on_rcu_stall)
|
| 1419 | panic("RCU Stall\n");
|
| 1420 | }
|
| 1421 |
|
| 1422 | static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
|
| 1423 | {
|
| 1424 | int cpu;
|
| 1425 | long delta;
|
| 1426 | unsigned long flags;
|
| 1427 | unsigned long gpa;
|
| 1428 | unsigned long j;
|
| 1429 | int ndetected = 0;
|
| 1430 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 1431 | long totqlen = 0;
|
| 1432 |
|
| 1433 |
|
| 1434 | rcu_stall_kick_kthreads(rsp);
|
| 1435 | if (rcu_cpu_stall_suppress)
|
| 1436 | return;
|
| 1437 |
|
| 1438 |
|
| 1439 |
|
| 1440 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 1441 | delta = jiffies - READ_ONCE(rsp->jiffies_stall);
|
| 1442 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
|
| 1443 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 1444 | return;
|
| 1445 | }
|
| 1446 | WRITE_ONCE(rsp->jiffies_stall,
|
| 1447 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
|
| 1448 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 1449 |
|
| 1450 | |
| 1451 | * OK, time to rat on our buddy...
|
| 1452 | * See Documentation/RCU/stallwarn.txt for info on how to debug
|
| 1453 | * RCU CPU stall warnings.
|
| 1454 |
|
| 1455 | pr_err("INFO: %s detected stalls on CPUs/tasks:",
|
| 1456 | rsp->name);
|
| 1457 | print_cpu_stall_info_begin();
|
| 1458 | rcu_for_each_leaf_node(rsp, rnp) {
|
| 1459 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 1460 | ndetected += rcu_print_task_stall(rnp);
|
| 1461 | if (rnp->qsmask != 0) {
|
| 1462 | for_each_leaf_node_possible_cpu(rnp, cpu)
|
| 1463 | if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
|
| 1464 | print_cpu_stall_info(rsp, cpu);
|
| 1465 | ndetected++;
|
| 1466 | }
|
| 1467 | }
|
| 1468 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 1469 | }
|
| 1470 |
|
| 1471 | print_cpu_stall_info_end();
|
| 1472 | for_each_possible_cpu(cpu)
|
| 1473 | totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
|
| 1474 | cpu)->cblist);
|
| 1475 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
|
| 1476 | smp_processor_id(), (long)(jiffies - rsp->gp_start),
|
| 1477 | (long)rsp->gpnum, (long)rsp->completed, totqlen);
|
| 1478 | if (ndetected) {
|
| 1479 | rcu_dump_cpu_stacks(rsp);
|
| 1480 |
|
| 1481 |
|
| 1482 | rcu_print_detail_task_stall(rsp);
|
| 1483 | } else {
|
| 1484 | if (READ_ONCE(rsp->gpnum) != gpnum ||
|
| 1485 | READ_ONCE(rsp->completed) == gpnum) {
|
| 1486 | pr_err("INFO: Stall ended before state dump start\n");
|
| 1487 | } else {
|
| 1488 | j = jiffies;
|
| 1489 | gpa = READ_ONCE(rsp->gp_activity);
|
| 1490 | pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
|
| 1491 | rsp->name, j - gpa, j, gpa,
|
| 1492 | jiffies_till_next_fqs,
|
| 1493 | rcu_get_root(rsp)->qsmask);
|
| 1494 |
|
| 1495 | sched_show_task(current);
|
| 1496 | }
|
| 1497 | }
|
| 1498 |
|
| 1499 | rcu_check_gp_kthread_starvation(rsp);
|
| 1500 |
|
| 1501 | panic_on_rcu_stall();
|
| 1502 |
|
| 1503 | force_quiescent_state(rsp);
|
| 1504 | }
|
| 1505 |
|
| 1506 | static void print_cpu_stall(struct rcu_state *rsp)
|
| 1507 | {
|
| 1508 | int cpu;
|
| 1509 | unsigned long flags;
|
| 1510 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 1511 | long totqlen = 0;
|
| 1512 |
|
| 1513 |
|
| 1514 | rcu_stall_kick_kthreads(rsp);
|
| 1515 | if (rcu_cpu_stall_suppress)
|
| 1516 | return;
|
| 1517 |
|
| 1518 | |
| 1519 | * OK, time to rat on ourselves...
|
| 1520 | * See Documentation/RCU/stallwarn.txt for info on how to debug
|
| 1521 | * RCU CPU stall warnings.
|
| 1522 |
|
| 1523 | pr_err("INFO: %s self-detected stall on CPU", rsp->name);
|
| 1524 | print_cpu_stall_info_begin();
|
| 1525 | print_cpu_stall_info(rsp, smp_processor_id());
|
| 1526 | print_cpu_stall_info_end();
|
| 1527 | for_each_possible_cpu(cpu)
|
| 1528 | totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
|
| 1529 | cpu)->cblist);
|
| 1530 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
|
| 1531 | jiffies - rsp->gp_start,
|
| 1532 | (long)rsp->gpnum, (long)rsp->completed, totqlen);
|
| 1533 |
|
| 1534 | rcu_check_gp_kthread_starvation(rsp);
|
| 1535 |
|
| 1536 | rcu_dump_cpu_stacks(rsp);
|
| 1537 |
|
| 1538 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 1539 | if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
|
| 1540 | WRITE_ONCE(rsp->jiffies_stall,
|
| 1541 | jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
|
| 1542 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 1543 |
|
| 1544 | panic_on_rcu_stall();
|
| 1545 |
|
| 1546 | |
| 1547 | * Attempt to revive the RCU machinery by forcing a context switch.
|
| 1548 | *
|
| 1549 | * A context switch would normally allow the RCU state machine to make
|
| 1550 | * progress and it could be we're stuck in kernel space without context
|
| 1551 | * switches for an entirely unreasonable amount of time.
|
| 1552 |
|
| 1553 | resched_cpu(smp_processor_id());
|
| 1554 | }
|
| 1555 |
|
| 1556 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 1557 | {
|
| 1558 | unsigned long completed;
|
| 1559 | unsigned long gpnum;
|
| 1560 | unsigned long gps;
|
| 1561 | unsigned long j;
|
| 1562 | unsigned long js;
|
| 1563 | struct rcu_node *rnp;
|
| 1564 |
|
| 1565 | if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
|
| 1566 | !rcu_gp_in_progress(rsp))
|
| 1567 | return;
|
| 1568 | rcu_stall_kick_kthreads(rsp);
|
| 1569 | j = jiffies;
|
| 1570 |
|
| 1571 | |
| 1572 | * Lots of memory barriers to reject false positives.
|
| 1573 | *
|
| 1574 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
|
| 1575 | * then rsp->gp_start, and finally rsp->completed. These values
|
| 1576 | * are updated in the opposite order with memory barriers (or
|
| 1577 | * equivalent) during grace-period initialization and cleanup.
|
| 1578 | * Now, a false positive can occur if we get an new value of
|
| 1579 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given
|
| 1580 | * the memory barriers, the only way that this can happen is if one
|
| 1581 | * grace period ends and another starts between these two fetches.
|
| 1582 | * Detect this by comparing rsp->completed with the previous fetch
|
| 1583 | * from rsp->gpnum.
|
| 1584 | *
|
| 1585 | * Given this check, comparisons of jiffies, rsp->jiffies_stall,
|
| 1586 | * and rsp->gp_start suffice to forestall false positives.
|
| 1587 |
|
| 1588 | gpnum = READ_ONCE(rsp->gpnum);
|
| 1589 | smp_rmb();
|
| 1590 | js = READ_ONCE(rsp->jiffies_stall);
|
| 1591 | smp_rmb();
|
| 1592 | gps = READ_ONCE(rsp->gp_start);
|
| 1593 | smp_rmb();
|
| 1594 | completed = READ_ONCE(rsp->completed);
|
| 1595 | if (ULONG_CMP_GE(completed, gpnum) ||
|
| 1596 | ULONG_CMP_LT(j, js) ||
|
| 1597 | ULONG_CMP_GE(gps, js))
|
| 1598 | return;
|
| 1599 | rnp = rdp->mynode;
|
| 1600 | if (rcu_gp_in_progress(rsp) &&
|
| 1601 | (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
|
| 1602 |
|
| 1603 |
|
| 1604 | print_cpu_stall(rsp);
|
| 1605 |
|
| 1606 | } else if (rcu_gp_in_progress(rsp) &&
|
| 1607 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
|
| 1608 |
|
| 1609 |
|
| 1610 | print_other_cpu_stall(rsp, gpnum);
|
| 1611 | }
|
| 1612 | }
|
| 1613 |
|
| 1614 | |
| 1615 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
|
| 1616 | *
|
| 1617 | * Set the stall-warning timeout way off into the future, thus preventing
|
| 1618 | * any RCU CPU stall-warning messages from appearing in the current set of
|
| 1619 | * RCU grace periods.
|
| 1620 | *
|
| 1621 | * The caller must disable hard irqs.
|
| 1622 |
|
| 1623 | void rcu_cpu_stall_reset(void)
|
| 1624 | {
|
| 1625 | struct rcu_state *rsp;
|
| 1626 |
|
| 1627 | for_each_rcu_flavor(rsp)
|
| 1628 | WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
|
| 1629 | }
|
| 1630 |
|
| 1631 | |
| 1632 | * Determine the value that ->completed will have at the end of the
|
| 1633 | * next subsequent grace period. This is used to tag callbacks so that
|
| 1634 | * a CPU can invoke callbacks in a timely fashion even if that CPU has
|
| 1635 | * been dyntick-idle for an extended period with callbacks under the
|
| 1636 | * influence of RCU_FAST_NO_HZ.
|
| 1637 | *
|
| 1638 | * The caller must hold rnp->lock with interrupts disabled.
|
| 1639 |
|
| 1640 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
|
| 1641 | struct rcu_node *rnp)
|
| 1642 | {
|
| 1643 | lockdep_assert_held(&rnp->lock);
|
| 1644 |
|
| 1645 | |
| 1646 | * If RCU is idle, we just wait for the next grace period.
|
| 1647 | * But we can only be sure that RCU is idle if we are looking
|
| 1648 | * at the root rcu_node structure -- otherwise, a new grace
|
| 1649 | * period might have started, but just not yet gotten around
|
| 1650 | * to initializing the current non-root rcu_node structure.
|
| 1651 |
|
| 1652 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
|
| 1653 | return rnp->completed + 1;
|
| 1654 |
|
| 1655 | |
| 1656 | * Otherwise, wait for a possible partial grace period and
|
| 1657 | * then the subsequent full grace period.
|
| 1658 |
|
| 1659 | return rnp->completed + 2;
|
| 1660 | }
|
| 1661 |
|
| 1662 | |
| 1663 | * Trace-event helper function for rcu_start_future_gp() and
|
| 1664 | * rcu_nocb_wait_gp().
|
| 1665 |
|
| 1666 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
|
| 1667 | unsigned long c, const char *s)
|
| 1668 | {
|
| 1669 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
|
| 1670 | rnp->completed, c, rnp->level,
|
| 1671 | rnp->grplo, rnp->grphi, s);
|
| 1672 | }
|
| 1673 |
|
| 1674 | |
| 1675 | * Start some future grace period, as needed to handle newly arrived
|
| 1676 | * callbacks. The required future grace periods are recorded in each
|
| 1677 | * rcu_node structure's ->need_future_gp field. Returns true if there
|
| 1678 | * is reason to awaken the grace-period kthread.
|
| 1679 | *
|
| 1680 | * The caller must hold the specified rcu_node structure's ->lock.
|
| 1681 |
|
| 1682 | static bool __maybe_unused
|
| 1683 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
|
| 1684 | unsigned long *c_out)
|
| 1685 | {
|
| 1686 | unsigned long c;
|
| 1687 | bool ret = false;
|
| 1688 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
|
| 1689 |
|
| 1690 | lockdep_assert_held(&rnp->lock);
|
| 1691 |
|
| 1692 | |
| 1693 | * Pick up grace-period number for new callbacks. If this
|
| 1694 | * grace period is already marked as needed, return to the caller.
|
| 1695 |
|
| 1696 | c = rcu_cbs_completed(rdp->rsp, rnp);
|
| 1697 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
|
| 1698 | if (rnp->need_future_gp[c & 0x1]) {
|
| 1699 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
|
| 1700 | goto out;
|
| 1701 | }
|
| 1702 |
|
| 1703 | |
| 1704 | * If either this rcu_node structure or the root rcu_node structure
|
| 1705 | * believe that a grace period is in progress, then we must wait
|
| 1706 | * for the one following, which is in "c". Because our request
|
| 1707 | * will be noticed at the end of the current grace period, we don't
|
| 1708 | * need to explicitly start one. We only do the lockless check
|
| 1709 | * of rnp_root's fields if the current rcu_node structure thinks
|
| 1710 | * there is no grace period in flight, and because we hold rnp->lock,
|
| 1711 | * the only possible change is when rnp_root's two fields are
|
| 1712 | * equal, in which case rnp_root->gpnum might be concurrently
|
| 1713 | * incremented. But that is OK, as it will just result in our
|
| 1714 | * doing some extra useless work.
|
| 1715 |
|
| 1716 | if (rnp->gpnum != rnp->completed ||
|
| 1717 | READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
|
| 1718 | rnp->need_future_gp[c & 0x1]++;
|
| 1719 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
|
| 1720 | goto out;
|
| 1721 | }
|
| 1722 |
|
| 1723 | |
| 1724 | * There might be no grace period in progress. If we don't already
|
| 1725 | * hold it, acquire the root rcu_node structure's lock in order to
|
| 1726 | * start one (if needed).
|
| 1727 |
|
| 1728 | if (rnp != rnp_root)
|
| 1729 | raw_spin_lock_rcu_node(rnp_root);
|
| 1730 |
|
| 1731 | |
| 1732 | * Get a new grace-period number. If there really is no grace
|
| 1733 | * period in progress, it will be smaller than the one we obtained
|
| 1734 | * earlier. Adjust callbacks as needed.
|
| 1735 |
|
| 1736 | c = rcu_cbs_completed(rdp->rsp, rnp_root);
|
| 1737 | if (!rcu_is_nocb_cpu(rdp->cpu))
|
| 1738 | (void)rcu_segcblist_accelerate(&rdp->cblist, c);
|
| 1739 |
|
| 1740 | |
| 1741 | * If the needed for the required grace period is already
|
| 1742 | * recorded, trace and leave.
|
| 1743 |
|
| 1744 | if (rnp_root->need_future_gp[c & 0x1]) {
|
| 1745 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
|
| 1746 | goto unlock_out;
|
| 1747 | }
|
| 1748 |
|
| 1749 |
|
| 1750 | rnp_root->need_future_gp[c & 0x1]++;
|
| 1751 |
|
| 1752 |
|
| 1753 | if (rnp_root->gpnum != rnp_root->completed) {
|
| 1754 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
|
| 1755 | } else {
|
| 1756 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
|
| 1757 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
|
| 1758 | }
|
| 1759 | unlock_out:
|
| 1760 | if (rnp != rnp_root)
|
| 1761 | raw_spin_unlock_rcu_node(rnp_root);
|
| 1762 | out:
|
| 1763 | if (c_out != NULL)
|
| 1764 | *c_out = c;
|
| 1765 | return ret;
|
| 1766 | }
|
| 1767 |
|
| 1768 | |
| 1769 | * Clean up any old requests for the just-ended grace period. Also return
|
| 1770 | * whether any additional grace periods have been requested.
|
| 1771 |
|
| 1772 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
|
| 1773 | {
|
| 1774 | int c = rnp->completed;
|
| 1775 | int needmore;
|
| 1776 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
|
| 1777 |
|
| 1778 | rnp->need_future_gp[c & 0x1] = 0;
|
| 1779 | needmore = rnp->need_future_gp[(c + 1) & 0x1];
|
| 1780 | trace_rcu_future_gp(rnp, rdp, c,
|
| 1781 | needmore ? TPS("CleanupMore") : TPS("Cleanup"));
|
| 1782 | return needmore;
|
| 1783 | }
|
| 1784 |
|
| 1785 | |
| 1786 | * Awaken the grace-period kthread for the specified flavor of RCU.
|
| 1787 | * Don't do a self-awaken, and don't bother awakening when there is
|
| 1788 | * nothing for the grace-period kthread to do (as in several CPUs
|
| 1789 | * raced to awaken, and we lost), and finally don't try to awaken
|
| 1790 | * a kthread that has not yet been created.
|
| 1791 |
|
| 1792 | static void rcu_gp_kthread_wake(struct rcu_state *rsp)
|
| 1793 | {
|
| 1794 | if (current == rsp->gp_kthread ||
|
| 1795 | !READ_ONCE(rsp->gp_flags) ||
|
| 1796 | !rsp->gp_kthread)
|
| 1797 | return;
|
| 1798 | swake_up(&rsp->gp_wq);
|
| 1799 | }
|
| 1800 |
|
| 1801 | |
| 1802 | * If there is room, assign a ->completed number to any callbacks on
|
| 1803 | * this CPU that have not already been assigned. Also accelerate any
|
| 1804 | * callbacks that were previously assigned a ->completed number that has
|
| 1805 | * since proven to be too conservative, which can happen if callbacks get
|
| 1806 | * assigned a ->completed number while RCU is idle, but with reference to
|
| 1807 | * a non-root rcu_node structure. This function is idempotent, so it does
|
| 1808 | * not hurt to call it repeatedly. Returns an flag saying that we should
|
| 1809 | * awaken the RCU grace-period kthread.
|
| 1810 | *
|
| 1811 | * The caller must hold rnp->lock with interrupts disabled.
|
| 1812 |
|
| 1813 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 1814 | struct rcu_data *rdp)
|
| 1815 | {
|
| 1816 | bool ret = false;
|
| 1817 |
|
| 1818 | lockdep_assert_held(&rnp->lock);
|
| 1819 |
|
| 1820 |
|
| 1821 | if (!rcu_segcblist_pend_cbs(&rdp->cblist))
|
| 1822 | return false;
|
| 1823 |
|
| 1824 | |
| 1825 | * Callbacks are often registered with incomplete grace-period
|
| 1826 | * information. Something about the fact that getting exact
|
| 1827 | * information requires acquiring a global lock... RCU therefore
|
| 1828 | * makes a conservative estimate of the grace period number at which
|
| 1829 | * a given callback will become ready to invoke. The following
|
| 1830 | * code checks this estimate and improves it when possible, thus
|
| 1831 | * accelerating callback invocation to an earlier grace-period
|
| 1832 | * number.
|
| 1833 |
|
| 1834 | if (rcu_segcblist_accelerate(&rdp->cblist, rcu_cbs_completed(rsp, rnp)))
|
| 1835 | ret = rcu_start_future_gp(rnp, rdp, NULL);
|
| 1836 |
|
| 1837 |
|
| 1838 | if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
|
| 1839 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
|
| 1840 | else
|
| 1841 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
|
| 1842 | return ret;
|
| 1843 | }
|
| 1844 |
|
| 1845 | |
| 1846 | * Move any callbacks whose grace period has completed to the
|
| 1847 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and
|
| 1848 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
|
| 1849 | * sublist. This function is idempotent, so it does not hurt to
|
| 1850 | * invoke it repeatedly. As long as it is not invoked -too- often...
|
| 1851 | * Returns true if the RCU grace-period kthread needs to be awakened.
|
| 1852 | *
|
| 1853 | * The caller must hold rnp->lock with interrupts disabled.
|
| 1854 |
|
| 1855 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 1856 | struct rcu_data *rdp)
|
| 1857 | {
|
| 1858 | lockdep_assert_held(&rnp->lock);
|
| 1859 |
|
| 1860 |
|
| 1861 | if (!rcu_segcblist_pend_cbs(&rdp->cblist))
|
| 1862 | return false;
|
| 1863 |
|
| 1864 | |
| 1865 | * Find all callbacks whose ->completed numbers indicate that they
|
| 1866 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
|
| 1867 |
|
| 1868 | rcu_segcblist_advance(&rdp->cblist, rnp->completed);
|
| 1869 |
|
| 1870 |
|
| 1871 | return rcu_accelerate_cbs(rsp, rnp, rdp);
|
| 1872 | }
|
| 1873 |
|
| 1874 | |
| 1875 | * Update CPU-local rcu_data state to record the beginnings and ends of
|
| 1876 | * grace periods. The caller must hold the ->lock of the leaf rcu_node
|
| 1877 | * structure corresponding to the current CPU, and must have irqs disabled.
|
| 1878 | * Returns true if the grace-period kthread needs to be awakened.
|
| 1879 |
|
| 1880 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 1881 | struct rcu_data *rdp)
|
| 1882 | {
|
| 1883 | bool ret;
|
| 1884 | bool need_gp;
|
| 1885 |
|
| 1886 | lockdep_assert_held(&rnp->lock);
|
| 1887 |
|
| 1888 |
|
| 1889 | if (rdp->completed == rnp->completed &&
|
| 1890 | !unlikely(READ_ONCE(rdp->gpwrap))) {
|
| 1891 |
|
| 1892 |
|
| 1893 | ret = rcu_accelerate_cbs(rsp, rnp, rdp);
|
| 1894 |
|
| 1895 | } else {
|
| 1896 |
|
| 1897 |
|
| 1898 | ret = rcu_advance_cbs(rsp, rnp, rdp);
|
| 1899 |
|
| 1900 |
|
| 1901 | rdp->completed = rnp->completed;
|
| 1902 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
|
| 1903 | }
|
| 1904 |
|
| 1905 | if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
|
| 1906 | |
| 1907 | * If the current grace period is waiting for this CPU,
|
| 1908 | * set up to detect a quiescent state, otherwise don't
|
| 1909 | * go looking for one.
|
| 1910 |
|
| 1911 | rdp->gpnum = rnp->gpnum;
|
| 1912 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
|
| 1913 | need_gp = !!(rnp->qsmask & rdp->grpmask);
|
| 1914 | rdp->cpu_no_qs.b.norm = need_gp;
|
| 1915 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
|
| 1916 | rdp->core_needs_qs = need_gp;
|
| 1917 | zero_cpu_stall_ticks(rdp);
|
| 1918 | WRITE_ONCE(rdp->gpwrap, false);
|
| 1919 | }
|
| 1920 | return ret;
|
| 1921 | }
|
| 1922 |
|
| 1923 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 1924 | {
|
| 1925 | unsigned long flags;
|
| 1926 | bool needwake;
|
| 1927 | struct rcu_node *rnp;
|
| 1928 |
|
| 1929 | local_irq_save(flags);
|
| 1930 | rnp = rdp->mynode;
|
| 1931 | if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
|
| 1932 | rdp->completed == READ_ONCE(rnp->completed) &&
|
| 1933 | !unlikely(READ_ONCE(rdp->gpwrap))) ||
|
| 1934 | !raw_spin_trylock_rcu_node(rnp)) {
|
| 1935 | local_irq_restore(flags);
|
| 1936 | return;
|
| 1937 | }
|
| 1938 | needwake = __note_gp_changes(rsp, rnp, rdp);
|
| 1939 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 1940 | if (needwake)
|
| 1941 | rcu_gp_kthread_wake(rsp);
|
| 1942 | }
|
| 1943 |
|
| 1944 | static void rcu_gp_slow(struct rcu_state *rsp, int delay)
|
| 1945 | {
|
| 1946 | if (delay > 0 &&
|
| 1947 | !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
|
| 1948 | schedule_timeout_uninterruptible(delay);
|
| 1949 | }
|
| 1950 |
|
| 1951 | |
| 1952 | * Initialize a new grace period. Return false if no grace period required.
|
| 1953 |
|
| 1954 | static bool rcu_gp_init(struct rcu_state *rsp)
|
| 1955 | {
|
| 1956 | unsigned long oldmask;
|
| 1957 | struct rcu_data *rdp;
|
| 1958 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 1959 |
|
| 1960 | WRITE_ONCE(rsp->gp_activity, jiffies);
|
| 1961 | raw_spin_lock_irq_rcu_node(rnp);
|
| 1962 | if (!READ_ONCE(rsp->gp_flags)) {
|
| 1963 |
|
| 1964 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 1965 | return false;
|
| 1966 | }
|
| 1967 | WRITE_ONCE(rsp->gp_flags, 0);
|
| 1968 |
|
| 1969 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
|
| 1970 | |
| 1971 | * Grace period already in progress, don't start another.
|
| 1972 | * Not supposed to be able to happen.
|
| 1973 |
|
| 1974 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 1975 | return false;
|
| 1976 | }
|
| 1977 |
|
| 1978 |
|
| 1979 | record_gp_stall_check_time(rsp);
|
| 1980 |
|
| 1981 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
|
| 1982 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
|
| 1983 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 1984 |
|
| 1985 | |
| 1986 | * Apply per-leaf buffered online and offline operations to the
|
| 1987 | * rcu_node tree. Note that this new grace period need not wait
|
| 1988 | * for subsequent online CPUs, and that quiescent-state forcing
|
| 1989 | * will handle subsequent offline CPUs.
|
| 1990 |
|
| 1991 | rcu_for_each_leaf_node(rsp, rnp) {
|
| 1992 | rcu_gp_slow(rsp, gp_preinit_delay);
|
| 1993 | raw_spin_lock_irq_rcu_node(rnp);
|
| 1994 | if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
|
| 1995 | !rnp->wait_blkd_tasks) {
|
| 1996 |
|
| 1997 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 1998 | continue;
|
| 1999 | }
|
| 2000 |
|
| 2001 |
|
| 2002 | oldmask = rnp->qsmaskinit;
|
| 2003 | rnp->qsmaskinit = rnp->qsmaskinitnext;
|
| 2004 |
|
| 2005 |
|
| 2006 | if (!oldmask != !rnp->qsmaskinit) {
|
| 2007 | if (!oldmask)
|
| 2008 | rcu_init_new_rnp(rnp);
|
| 2009 | else if (rcu_preempt_has_tasks(rnp))
|
| 2010 | rnp->wait_blkd_tasks = true;
|
| 2011 | else
|
| 2012 | rcu_cleanup_dead_rnp(rnp);
|
| 2013 | }
|
| 2014 |
|
| 2015 | |
| 2016 | * If all waited-on tasks from prior grace period are
|
| 2017 | * done, and if all this rcu_node structure's CPUs are
|
| 2018 | * still offline, propagate up the rcu_node tree and
|
| 2019 | * clear ->wait_blkd_tasks. Otherwise, if one of this
|
| 2020 | * rcu_node structure's CPUs has since come back online,
|
| 2021 | * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
|
| 2022 | * checks for this, so just call it unconditionally).
|
| 2023 |
|
| 2024 | if (rnp->wait_blkd_tasks &&
|
| 2025 | (!rcu_preempt_has_tasks(rnp) ||
|
| 2026 | rnp->qsmaskinit)) {
|
| 2027 | rnp->wait_blkd_tasks = false;
|
| 2028 | rcu_cleanup_dead_rnp(rnp);
|
| 2029 | }
|
| 2030 |
|
| 2031 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 2032 | }
|
| 2033 |
|
| 2034 | |
| 2035 | * Set the quiescent-state-needed bits in all the rcu_node
|
| 2036 | * structures for all currently online CPUs in breadth-first order,
|
| 2037 | * starting from the root rcu_node structure, relying on the layout
|
| 2038 | * of the tree within the rsp->node[] array. Note that other CPUs
|
| 2039 | * will access only the leaves of the hierarchy, thus seeing that no
|
| 2040 | * grace period is in progress, at least until the corresponding
|
| 2041 | * leaf node has been initialized.
|
| 2042 | *
|
| 2043 | * The grace period cannot complete until the initialization
|
| 2044 | * process finishes, because this kthread handles both.
|
| 2045 |
|
| 2046 | rcu_for_each_node_breadth_first(rsp, rnp) {
|
| 2047 | rcu_gp_slow(rsp, gp_init_delay);
|
| 2048 | raw_spin_lock_irq_rcu_node(rnp);
|
| 2049 | rdp = this_cpu_ptr(rsp->rda);
|
| 2050 | rcu_preempt_check_blocked_tasks(rnp);
|
| 2051 | rnp->qsmask = rnp->qsmaskinit;
|
| 2052 | WRITE_ONCE(rnp->gpnum, rsp->gpnum);
|
| 2053 | if (WARN_ON_ONCE(rnp->completed != rsp->completed))
|
| 2054 | WRITE_ONCE(rnp->completed, rsp->completed);
|
| 2055 | if (rnp == rdp->mynode)
|
| 2056 | (void)__note_gp_changes(rsp, rnp, rdp);
|
| 2057 | rcu_preempt_boost_start_gp(rnp);
|
| 2058 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
|
| 2059 | rnp->level, rnp->grplo,
|
| 2060 | rnp->grphi, rnp->qsmask);
|
| 2061 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 2062 | cond_resched_rcu_qs();
|
| 2063 | WRITE_ONCE(rsp->gp_activity, jiffies);
|
| 2064 | }
|
| 2065 |
|
| 2066 | return true;
|
| 2067 | }
|
| 2068 |
|
| 2069 | |
| 2070 | * Helper function for wait_event_interruptible_timeout() wakeup
|
| 2071 | * at force-quiescent-state time.
|
| 2072 |
|
| 2073 | static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
|
| 2074 | {
|
| 2075 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 2076 |
|
| 2077 |
|
| 2078 | *gfp = READ_ONCE(rsp->gp_flags);
|
| 2079 | if (*gfp & RCU_GP_FLAG_FQS)
|
| 2080 | return true;
|
| 2081 |
|
| 2082 |
|
| 2083 | if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
|
| 2084 | return true;
|
| 2085 |
|
| 2086 | return false;
|
| 2087 | }
|
| 2088 |
|
| 2089 | |
| 2090 | * Do one round of quiescent-state forcing.
|
| 2091 |
|
| 2092 | static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
|
| 2093 | {
|
| 2094 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 2095 |
|
| 2096 | WRITE_ONCE(rsp->gp_activity, jiffies);
|
| 2097 | rsp->n_force_qs++;
|
| 2098 | if (first_time) {
|
| 2099 |
|
| 2100 | force_qs_rnp(rsp, dyntick_save_progress_counter);
|
| 2101 | } else {
|
| 2102 |
|
| 2103 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
|
| 2104 | }
|
| 2105 |
|
| 2106 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
|
| 2107 | raw_spin_lock_irq_rcu_node(rnp);
|
| 2108 | WRITE_ONCE(rsp->gp_flags,
|
| 2109 | READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
|
| 2110 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 2111 | }
|
| 2112 | }
|
| 2113 |
|
| 2114 | |
| 2115 | * Clean up after the old grace period.
|
| 2116 |
|
| 2117 | static void rcu_gp_cleanup(struct rcu_state *rsp)
|
| 2118 | {
|
| 2119 | unsigned long gp_duration;
|
| 2120 | bool needgp = false;
|
| 2121 | int nocb = 0;
|
| 2122 | struct rcu_data *rdp;
|
| 2123 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 2124 | struct swait_queue_head *sq;
|
| 2125 |
|
| 2126 | WRITE_ONCE(rsp->gp_activity, jiffies);
|
| 2127 | raw_spin_lock_irq_rcu_node(rnp);
|
| 2128 | gp_duration = jiffies - rsp->gp_start;
|
| 2129 | if (gp_duration > rsp->gp_max)
|
| 2130 | rsp->gp_max = gp_duration;
|
| 2131 |
|
| 2132 | |
| 2133 | * We know the grace period is complete, but to everyone else
|
| 2134 | * it appears to still be ongoing. But it is also the case
|
| 2135 | * that to everyone else it looks like there is nothing that
|
| 2136 | * they can do to advance the grace period. It is therefore
|
| 2137 | * safe for us to drop the lock in order to mark the grace
|
| 2138 | * period as completed in all of the rcu_node structures.
|
| 2139 |
|
| 2140 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 2141 |
|
| 2142 | |
| 2143 | * Propagate new ->completed value to rcu_node structures so
|
| 2144 | * that other CPUs don't have to wait until the start of the next
|
| 2145 | * grace period to process their callbacks. This also avoids
|
| 2146 | * some nasty RCU grace-period initialization races by forcing
|
| 2147 | * the end of the current grace period to be completely recorded in
|
| 2148 | * all of the rcu_node structures before the beginning of the next
|
| 2149 | * grace period is recorded in any of the rcu_node structures.
|
| 2150 |
|
| 2151 | rcu_for_each_node_breadth_first(rsp, rnp) {
|
| 2152 | raw_spin_lock_irq_rcu_node(rnp);
|
| 2153 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
|
| 2154 | WARN_ON_ONCE(rnp->qsmask);
|
| 2155 | WRITE_ONCE(rnp->completed, rsp->gpnum);
|
| 2156 | rdp = this_cpu_ptr(rsp->rda);
|
| 2157 | if (rnp == rdp->mynode)
|
| 2158 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
|
| 2159 |
|
| 2160 | nocb += rcu_future_gp_cleanup(rsp, rnp);
|
| 2161 | sq = rcu_nocb_gp_get(rnp);
|
| 2162 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 2163 | rcu_nocb_gp_cleanup(sq);
|
| 2164 | cond_resched_rcu_qs();
|
| 2165 | WRITE_ONCE(rsp->gp_activity, jiffies);
|
| 2166 | rcu_gp_slow(rsp, gp_cleanup_delay);
|
| 2167 | }
|
| 2168 | rnp = rcu_get_root(rsp);
|
| 2169 | raw_spin_lock_irq_rcu_node(rnp);
|
| 2170 | rcu_nocb_gp_set(rnp, nocb);
|
| 2171 |
|
| 2172 |
|
| 2173 | WRITE_ONCE(rsp->completed, rsp->gpnum);
|
| 2174 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
|
| 2175 | rsp->gp_state = RCU_GP_IDLE;
|
| 2176 | rdp = this_cpu_ptr(rsp->rda);
|
| 2177 |
|
| 2178 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
|
| 2179 | if (needgp || cpu_needs_another_gp(rsp, rdp)) {
|
| 2180 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
|
| 2181 | trace_rcu_grace_period(rsp->name,
|
| 2182 | READ_ONCE(rsp->gpnum),
|
| 2183 | TPS("newreq"));
|
| 2184 | }
|
| 2185 | raw_spin_unlock_irq_rcu_node(rnp);
|
| 2186 | }
|
| 2187 |
|
| 2188 | |
| 2189 | * Body of kthread that handles grace periods.
|
| 2190 |
|
| 2191 | static int __noreturn rcu_gp_kthread(void *arg)
|
| 2192 | {
|
| 2193 | bool first_gp_fqs;
|
| 2194 | int gf;
|
| 2195 | unsigned long j;
|
| 2196 | int ret;
|
| 2197 | struct rcu_state *rsp = arg;
|
| 2198 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 2199 |
|
| 2200 | rcu_bind_gp_kthread();
|
| 2201 | for (;;) {
|
| 2202 |
|
| 2203 |
|
| 2204 | for (;;) {
|
| 2205 | trace_rcu_grace_period(rsp->name,
|
| 2206 | READ_ONCE(rsp->gpnum),
|
| 2207 | TPS("reqwait"));
|
| 2208 | rsp->gp_state = RCU_GP_WAIT_GPS;
|
| 2209 | swait_event_interruptible(rsp->gp_wq,
|
| 2210 | READ_ONCE(rsp->gp_flags) &
|
| 2211 | RCU_GP_FLAG_INIT);
|
| 2212 | rsp->gp_state = RCU_GP_DONE_GPS;
|
| 2213 |
|
| 2214 | if (rcu_gp_init(rsp))
|
| 2215 | break;
|
| 2216 | cond_resched_rcu_qs();
|
| 2217 | WRITE_ONCE(rsp->gp_activity, jiffies);
|
| 2218 | WARN_ON(signal_pending(current));
|
| 2219 | trace_rcu_grace_period(rsp->name,
|
| 2220 | READ_ONCE(rsp->gpnum),
|
| 2221 | TPS("reqwaitsig"));
|
| 2222 | }
|
| 2223 |
|
| 2224 |
|
| 2225 | first_gp_fqs = true;
|
| 2226 | j = jiffies_till_first_fqs;
|
| 2227 | if (j > HZ) {
|
| 2228 | j = HZ;
|
| 2229 | jiffies_till_first_fqs = HZ;
|
| 2230 | }
|
| 2231 | ret = 0;
|
| 2232 | for (;;) {
|
| 2233 | if (!ret) {
|
| 2234 | rsp->jiffies_force_qs = jiffies + j;
|
| 2235 | WRITE_ONCE(rsp->jiffies_kick_kthreads,
|
| 2236 | jiffies + 3 * j);
|
| 2237 | }
|
| 2238 | trace_rcu_grace_period(rsp->name,
|
| 2239 | READ_ONCE(rsp->gpnum),
|
| 2240 | TPS("fqswait"));
|
| 2241 | rsp->gp_state = RCU_GP_WAIT_FQS;
|
| 2242 | ret = swait_event_interruptible_timeout(rsp->gp_wq,
|
| 2243 | rcu_gp_fqs_check_wake(rsp, &gf), j);
|
| 2244 | rsp->gp_state = RCU_GP_DOING_FQS;
|
| 2245 |
|
| 2246 |
|
| 2247 | if (!READ_ONCE(rnp->qsmask) &&
|
| 2248 | !rcu_preempt_blocked_readers_cgp(rnp))
|
| 2249 | break;
|
| 2250 |
|
| 2251 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
|
| 2252 | (gf & RCU_GP_FLAG_FQS)) {
|
| 2253 | trace_rcu_grace_period(rsp->name,
|
| 2254 | READ_ONCE(rsp->gpnum),
|
| 2255 | TPS("fqsstart"));
|
| 2256 | rcu_gp_fqs(rsp, first_gp_fqs);
|
| 2257 | first_gp_fqs = false;
|
| 2258 | trace_rcu_grace_period(rsp->name,
|
| 2259 | READ_ONCE(rsp->gpnum),
|
| 2260 | TPS("fqsend"));
|
| 2261 | cond_resched_rcu_qs();
|
| 2262 | WRITE_ONCE(rsp->gp_activity, jiffies);
|
| 2263 | ret = 0;
|
| 2264 | j = jiffies_till_next_fqs;
|
| 2265 | if (j > HZ) {
|
| 2266 | j = HZ;
|
| 2267 | jiffies_till_next_fqs = HZ;
|
| 2268 | } else if (j < 1) {
|
| 2269 | j = 1;
|
| 2270 | jiffies_till_next_fqs = 1;
|
| 2271 | }
|
| 2272 | } else {
|
| 2273 |
|
| 2274 | cond_resched_rcu_qs();
|
| 2275 | WRITE_ONCE(rsp->gp_activity, jiffies);
|
| 2276 | WARN_ON(signal_pending(current));
|
| 2277 | trace_rcu_grace_period(rsp->name,
|
| 2278 | READ_ONCE(rsp->gpnum),
|
| 2279 | TPS("fqswaitsig"));
|
| 2280 | ret = 1;
|
| 2281 | j = jiffies;
|
| 2282 | if (time_after(jiffies, rsp->jiffies_force_qs))
|
| 2283 | j = 1;
|
| 2284 | else
|
| 2285 | j = rsp->jiffies_force_qs - j;
|
| 2286 | }
|
| 2287 | }
|
| 2288 |
|
| 2289 |
|
| 2290 | rsp->gp_state = RCU_GP_CLEANUP;
|
| 2291 | rcu_gp_cleanup(rsp);
|
| 2292 | rsp->gp_state = RCU_GP_CLEANED;
|
| 2293 | }
|
| 2294 | }
|
| 2295 |
|
| 2296 | |
| 2297 | * Start a new RCU grace period if warranted, re-initializing the hierarchy
|
| 2298 | * in preparation for detecting the next grace period. The caller must hold
|
| 2299 | * the root node's ->lock and hard irqs must be disabled.
|
| 2300 | *
|
| 2301 | * Note that it is legal for a dying CPU (which is marked as offline) to
|
| 2302 | * invoke this function. This can happen when the dying CPU reports its
|
| 2303 | * quiescent state.
|
| 2304 | *
|
| 2305 | * Returns true if the grace-period kthread must be awakened.
|
| 2306 |
|
| 2307 | static bool
|
| 2308 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 2309 | struct rcu_data *rdp)
|
| 2310 | {
|
| 2311 | lockdep_assert_held(&rnp->lock);
|
| 2312 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
|
| 2313 | |
| 2314 | * Either we have not yet spawned the grace-period
|
| 2315 | * task, this CPU does not need another grace period,
|
| 2316 | * or a grace period is already in progress.
|
| 2317 | * Either way, don't start a new grace period.
|
| 2318 |
|
| 2319 | return false;
|
| 2320 | }
|
| 2321 | WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
|
| 2322 | trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
|
| 2323 | TPS("newreq"));
|
| 2324 |
|
| 2325 | |
| 2326 | * We can't do wakeups while holding the rnp->lock, as that
|
| 2327 | * could cause possible deadlocks with the rq->lock. Defer
|
| 2328 | * the wakeup to our caller.
|
| 2329 |
|
| 2330 | return true;
|
| 2331 | }
|
| 2332 |
|
| 2333 | |
| 2334 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
|
| 2335 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
|
| 2336 | * is invoked indirectly from rcu_advance_cbs(), which would result in
|
| 2337 | * endless recursion -- or would do so if it wasn't for the self-deadlock
|
| 2338 | * that is encountered beforehand.
|
| 2339 | *
|
| 2340 | * Returns true if the grace-period kthread needs to be awakened.
|
| 2341 |
|
| 2342 | static bool rcu_start_gp(struct rcu_state *rsp)
|
| 2343 | {
|
| 2344 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
|
| 2345 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 2346 | bool ret = false;
|
| 2347 |
|
| 2348 | |
| 2349 | * If there is no grace period in progress right now, any
|
| 2350 | * callbacks we have up to this point will be satisfied by the
|
| 2351 | * next grace period. Also, advancing the callbacks reduces the
|
| 2352 | * probability of false positives from cpu_needs_another_gp()
|
| 2353 | * resulting in pointless grace periods. So, advance callbacks
|
| 2354 | * then start the grace period!
|
| 2355 |
|
| 2356 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
|
| 2357 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
|
| 2358 | return ret;
|
| 2359 | }
|
| 2360 |
|
| 2361 | |
| 2362 | * Report a full set of quiescent states to the specified rcu_state data
|
| 2363 | * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
|
| 2364 | * kthread if another grace period is required. Whether we wake
|
| 2365 | * the grace-period kthread or it awakens itself for the next round
|
| 2366 | * of quiescent-state forcing, that kthread will clean up after the
|
| 2367 | * just-completed grace period. Note that the caller must hold rnp->lock,
|
| 2368 | * which is released before return.
|
| 2369 |
|
| 2370 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
|
| 2371 | __releases(rcu_get_root(rsp)->lock)
|
| 2372 | {
|
| 2373 | lockdep_assert_held(&rcu_get_root(rsp)->lock);
|
| 2374 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
|
| 2375 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
|
| 2376 | raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
|
| 2377 | rcu_gp_kthread_wake(rsp);
|
| 2378 | }
|
| 2379 |
|
| 2380 | |
| 2381 | * Similar to rcu_report_qs_rdp(), for which it is a helper function.
|
| 2382 | * Allows quiescent states for a group of CPUs to be reported at one go
|
| 2383 | * to the specified rcu_node structure, though all the CPUs in the group
|
| 2384 | * must be represented by the same rcu_node structure (which need not be a
|
| 2385 | * leaf rcu_node structure, though it often will be). The gps parameter
|
| 2386 | * is the grace-period snapshot, which means that the quiescent states
|
| 2387 | * are valid only if rnp->gpnum is equal to gps. That structure's lock
|
| 2388 | * must be held upon entry, and it is released before return.
|
| 2389 |
|
| 2390 | static void
|
| 2391 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
|
| 2392 | struct rcu_node *rnp, unsigned long gps, unsigned long flags)
|
| 2393 | __releases(rnp->lock)
|
| 2394 | {
|
| 2395 | unsigned long oldmask = 0;
|
| 2396 | struct rcu_node *rnp_c;
|
| 2397 |
|
| 2398 | lockdep_assert_held(&rnp->lock);
|
| 2399 |
|
| 2400 |
|
| 2401 | for (;;) {
|
| 2402 | if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
|
| 2403 |
|
| 2404 | |
| 2405 | * Our bit has already been cleared, or the
|
| 2406 | * relevant grace period is already over, so done.
|
| 2407 |
|
| 2408 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 2409 | return;
|
| 2410 | }
|
| 2411 | WARN_ON_ONCE(oldmask);
|
| 2412 | rnp->qsmask &= ~mask;
|
| 2413 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
|
| 2414 | mask, rnp->qsmask, rnp->level,
|
| 2415 | rnp->grplo, rnp->grphi,
|
| 2416 | !!rnp->gp_tasks);
|
| 2417 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
|
| 2418 |
|
| 2419 |
|
| 2420 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 2421 | return;
|
| 2422 | }
|
| 2423 | mask = rnp->grpmask;
|
| 2424 | if (rnp->parent == NULL) {
|
| 2425 |
|
| 2426 |
|
| 2427 |
|
| 2428 | break;
|
| 2429 | }
|
| 2430 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 2431 | rnp_c = rnp;
|
| 2432 | rnp = rnp->parent;
|
| 2433 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 2434 | oldmask = rnp_c->qsmask;
|
| 2435 | }
|
| 2436 |
|
| 2437 | |
| 2438 | * Get here if we are the last CPU to pass through a quiescent
|
| 2439 | * state for this grace period. Invoke rcu_report_qs_rsp()
|
| 2440 | * to clean up and start the next grace period if one is needed.
|
| 2441 |
|
| 2442 | rcu_report_qs_rsp(rsp, flags);
|
| 2443 | }
|
| 2444 |
|
| 2445 | |
| 2446 | * Record a quiescent state for all tasks that were previously queued
|
| 2447 | * on the specified rcu_node structure and that were blocking the current
|
| 2448 | * RCU grace period. The caller must hold the specified rnp->lock with
|
| 2449 | * irqs disabled, and this lock is released upon return, but irqs remain
|
| 2450 | * disabled.
|
| 2451 |
|
| 2452 | static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
|
| 2453 | struct rcu_node *rnp, unsigned long flags)
|
| 2454 | __releases(rnp->lock)
|
| 2455 | {
|
| 2456 | unsigned long gps;
|
| 2457 | unsigned long mask;
|
| 2458 | struct rcu_node *rnp_p;
|
| 2459 |
|
| 2460 | lockdep_assert_held(&rnp->lock);
|
| 2461 | if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
|
| 2462 | rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
|
| 2463 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 2464 | return;
|
| 2465 | }
|
| 2466 |
|
| 2467 | rnp_p = rnp->parent;
|
| 2468 | if (rnp_p == NULL) {
|
| 2469 | |
| 2470 | * Only one rcu_node structure in the tree, so don't
|
| 2471 | * try to report up to its nonexistent parent!
|
| 2472 |
|
| 2473 | rcu_report_qs_rsp(rsp, flags);
|
| 2474 | return;
|
| 2475 | }
|
| 2476 |
|
| 2477 |
|
| 2478 | gps = rnp->gpnum;
|
| 2479 | mask = rnp->grpmask;
|
| 2480 | raw_spin_unlock_rcu_node(rnp);
|
| 2481 | raw_spin_lock_rcu_node(rnp_p);
|
| 2482 | rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
|
| 2483 | }
|
| 2484 |
|
| 2485 | |
| 2486 | * Record a quiescent state for the specified CPU to that CPU's rcu_data
|
| 2487 | * structure. This must be called from the specified CPU.
|
| 2488 |
|
| 2489 | static void
|
| 2490 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
|
| 2491 | {
|
| 2492 | unsigned long flags;
|
| 2493 | unsigned long mask;
|
| 2494 | bool needwake;
|
| 2495 | struct rcu_node *rnp;
|
| 2496 |
|
| 2497 | rnp = rdp->mynode;
|
| 2498 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 2499 | if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
|
| 2500 | rnp->completed == rnp->gpnum || rdp->gpwrap) {
|
| 2501 |
|
| 2502 | |
| 2503 | * The grace period in which this quiescent state was
|
| 2504 | * recorded has ended, so don't report it upwards.
|
| 2505 | * We will instead need a new quiescent state that lies
|
| 2506 | * within the current grace period.
|
| 2507 |
|
| 2508 | rdp->cpu_no_qs.b.norm = true;
|
| 2509 | rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
|
| 2510 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 2511 | return;
|
| 2512 | }
|
| 2513 | mask = rdp->grpmask;
|
| 2514 | if ((rnp->qsmask & mask) == 0) {
|
| 2515 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 2516 | } else {
|
| 2517 | rdp->core_needs_qs = false;
|
| 2518 |
|
| 2519 | |
| 2520 | * This GP can't end until cpu checks in, so all of our
|
| 2521 | * callbacks can be processed during the next GP.
|
| 2522 |
|
| 2523 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
|
| 2524 |
|
| 2525 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
|
| 2526 |
|
| 2527 | if (needwake)
|
| 2528 | rcu_gp_kthread_wake(rsp);
|
| 2529 | }
|
| 2530 | }
|
| 2531 |
|
| 2532 | |
| 2533 | * Check to see if there is a new grace period of which this CPU
|
| 2534 | * is not yet aware, and if so, set up local rcu_data state for it.
|
| 2535 | * Otherwise, see if this CPU has just passed through its first
|
| 2536 | * quiescent state for this grace period, and record that fact if so.
|
| 2537 |
|
| 2538 | static void
|
| 2539 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 2540 | {
|
| 2541 |
|
| 2542 | note_gp_changes(rsp, rdp);
|
| 2543 |
|
| 2544 | |
| 2545 | * Does this CPU still need to do its part for current grace period?
|
| 2546 | * If no, return and let the other CPUs do their part as well.
|
| 2547 |
|
| 2548 | if (!rdp->core_needs_qs)
|
| 2549 | return;
|
| 2550 |
|
| 2551 | |
| 2552 | * Was there a quiescent state since the beginning of the grace
|
| 2553 | * period? If no, then exit and wait for the next call.
|
| 2554 |
|
| 2555 | if (rdp->cpu_no_qs.b.norm)
|
| 2556 | return;
|
| 2557 |
|
| 2558 | |
| 2559 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the
|
| 2560 | * judge of that).
|
| 2561 |
|
| 2562 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
|
| 2563 | }
|
| 2564 |
|
| 2565 | |
| 2566 | * Send the specified CPU's RCU callbacks to the orphanage. The
|
| 2567 | * specified CPU must be offline, and the caller must hold the
|
| 2568 | * ->orphan_lock.
|
| 2569 |
|
| 2570 | static void
|
| 2571 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
|
| 2572 | struct rcu_node *rnp, struct rcu_data *rdp)
|
| 2573 | {
|
| 2574 | lockdep_assert_held(&rsp->orphan_lock);
|
| 2575 |
|
| 2576 |
|
| 2577 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
|
| 2578 | return;
|
| 2579 |
|
| 2580 | |
| 2581 | * Orphan the callbacks. First adjust the counts. This is safe
|
| 2582 | * because _rcu_barrier() excludes CPU-hotplug operations, so it
|
| 2583 | * cannot be running now. Thus no memory barrier is required.
|
| 2584 |
|
| 2585 | rdp->n_cbs_orphaned += rcu_segcblist_n_cbs(&rdp->cblist);
|
| 2586 | rcu_segcblist_extract_count(&rdp->cblist, &rsp->orphan_done);
|
| 2587 |
|
| 2588 | |
| 2589 | * Next, move those callbacks still needing a grace period to
|
| 2590 | * the orphanage, where some other CPU will pick them up.
|
| 2591 | * Some of the callbacks might have gone partway through a grace
|
| 2592 | * period, but that is too bad. They get to start over because we
|
| 2593 | * cannot assume that grace periods are synchronized across CPUs.
|
| 2594 |
|
| 2595 | rcu_segcblist_extract_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
|
| 2596 |
|
| 2597 | |
| 2598 | * Then move the ready-to-invoke callbacks to the orphanage,
|
| 2599 | * where some other CPU will pick them up. These will not be
|
| 2600 | * required to pass though another grace period: They are done.
|
| 2601 |
|
| 2602 | rcu_segcblist_extract_done_cbs(&rdp->cblist, &rsp->orphan_done);
|
| 2603 |
|
| 2604 |
|
| 2605 | rcu_segcblist_disable(&rdp->cblist);
|
| 2606 | }
|
| 2607 |
|
| 2608 | |
| 2609 | * Adopt the RCU callbacks from the specified rcu_state structure's
|
| 2610 | * orphanage. The caller must hold the ->orphan_lock.
|
| 2611 |
|
| 2612 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
|
| 2613 | {
|
| 2614 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
|
| 2615 |
|
| 2616 | lockdep_assert_held(&rsp->orphan_lock);
|
| 2617 |
|
| 2618 |
|
| 2619 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
|
| 2620 | rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
|
| 2621 | return;
|
| 2622 |
|
| 2623 |
|
| 2624 | rdp->n_cbs_adopted += rsp->orphan_done.len;
|
| 2625 | if (rsp->orphan_done.len_lazy != rsp->orphan_done.len)
|
| 2626 | rcu_idle_count_callbacks_posted();
|
| 2627 | rcu_segcblist_insert_count(&rdp->cblist, &rsp->orphan_done);
|
| 2628 |
|
| 2629 | |
| 2630 | * We do not need a memory barrier here because the only way we
|
| 2631 | * can get here if there is an rcu_barrier() in flight is if
|
| 2632 | * we are the task doing the rcu_barrier().
|
| 2633 |
|
| 2634 |
|
| 2635 |
|
| 2636 | rcu_segcblist_insert_done_cbs(&rdp->cblist, &rsp->orphan_done);
|
| 2637 | WARN_ON_ONCE(rsp->orphan_done.head);
|
| 2638 | rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rsp->orphan_pend);
|
| 2639 | WARN_ON_ONCE(rsp->orphan_pend.head);
|
| 2640 | WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) !=
|
| 2641 | !rcu_segcblist_n_cbs(&rdp->cblist));
|
| 2642 | }
|
| 2643 |
|
| 2644 | |
| 2645 | * Trace the fact that this CPU is going offline.
|
| 2646 |
|
| 2647 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
|
| 2648 | {
|
| 2649 | RCU_TRACE(unsigned long mask;)
|
| 2650 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
|
| 2651 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
|
| 2652 |
|
| 2653 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
|
| 2654 | return;
|
| 2655 |
|
| 2656 | RCU_TRACE(mask = rdp->grpmask;)
|
| 2657 | trace_rcu_grace_period(rsp->name,
|
| 2658 | rnp->gpnum + 1 - !!(rnp->qsmask & mask),
|
| 2659 | TPS("cpuofl"));
|
| 2660 | }
|
| 2661 |
|
| 2662 | |
| 2663 | * All CPUs for the specified rcu_node structure have gone offline,
|
| 2664 | * and all tasks that were preempted within an RCU read-side critical
|
| 2665 | * section while running on one of those CPUs have since exited their RCU
|
| 2666 | * read-side critical section. Some other CPU is reporting this fact with
|
| 2667 | * the specified rcu_node structure's ->lock held and interrupts disabled.
|
| 2668 | * This function therefore goes up the tree of rcu_node structures,
|
| 2669 | * clearing the corresponding bits in the ->qsmaskinit fields. Note that
|
| 2670 | * the leaf rcu_node structure's ->qsmaskinit field has already been
|
| 2671 | * updated
|
| 2672 | *
|
| 2673 | * This function does check that the specified rcu_node structure has
|
| 2674 | * all CPUs offline and no blocked tasks, so it is OK to invoke it
|
| 2675 | * prematurely. That said, invoking it after the fact will cost you
|
| 2676 | * a needless lock acquisition. So once it has done its work, don't
|
| 2677 | * invoke it again.
|
| 2678 |
|
| 2679 | static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
|
| 2680 | {
|
| 2681 | long mask;
|
| 2682 | struct rcu_node *rnp = rnp_leaf;
|
| 2683 |
|
| 2684 | lockdep_assert_held(&rnp->lock);
|
| 2685 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
|
| 2686 | rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
|
| 2687 | return;
|
| 2688 | for (;;) {
|
| 2689 | mask = rnp->grpmask;
|
| 2690 | rnp = rnp->parent;
|
| 2691 | if (!rnp)
|
| 2692 | break;
|
| 2693 | raw_spin_lock_rcu_node(rnp);
|
| 2694 | rnp->qsmaskinit &= ~mask;
|
| 2695 | rnp->qsmask &= ~mask;
|
| 2696 | if (rnp->qsmaskinit) {
|
| 2697 | raw_spin_unlock_rcu_node(rnp);
|
| 2698 |
|
| 2699 | return;
|
| 2700 | }
|
| 2701 | raw_spin_unlock_rcu_node(rnp);
|
| 2702 | }
|
| 2703 | }
|
| 2704 |
|
| 2705 | |
| 2706 | * The CPU has been completely removed, and some other CPU is reporting
|
| 2707 | * this fact from process context. Do the remainder of the cleanup,
|
| 2708 | * including orphaning the outgoing CPU's RCU callbacks, and also
|
| 2709 | * adopting them. There can only be one CPU hotplug operation at a time,
|
| 2710 | * so no other CPU can be attempting to update rcu_cpu_kthread_task.
|
| 2711 |
|
| 2712 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
|
| 2713 | {
|
| 2714 | unsigned long flags;
|
| 2715 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 2716 | struct rcu_node *rnp = rdp->mynode;
|
| 2717 |
|
| 2718 | if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
|
| 2719 | return;
|
| 2720 |
|
| 2721 |
|
| 2722 | rcu_boost_kthread_setaffinity(rnp, -1);
|
| 2723 |
|
| 2724 |
|
| 2725 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
|
| 2726 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
|
| 2727 | rcu_adopt_orphan_cbs(rsp, flags);
|
| 2728 | raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
|
| 2729 |
|
| 2730 | WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
|
| 2731 | !rcu_segcblist_empty(&rdp->cblist),
|
| 2732 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
|
| 2733 | cpu, rcu_segcblist_n_cbs(&rdp->cblist),
|
| 2734 | rcu_segcblist_first_cb(&rdp->cblist));
|
| 2735 | }
|
| 2736 |
|
| 2737 | |
| 2738 | * Invoke any RCU callbacks that have made it to the end of their grace
|
| 2739 | * period. Thottle as specified by rdp->blimit.
|
| 2740 |
|
| 2741 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 2742 | {
|
| 2743 | unsigned long flags;
|
| 2744 | struct rcu_head *rhp;
|
| 2745 | struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
|
| 2746 | long bl, count;
|
| 2747 |
|
| 2748 |
|
| 2749 | if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
|
| 2750 | trace_rcu_batch_start(rsp->name,
|
| 2751 | rcu_segcblist_n_lazy_cbs(&rdp->cblist),
|
| 2752 | rcu_segcblist_n_cbs(&rdp->cblist), 0);
|
| 2753 | trace_rcu_batch_end(rsp->name, 0,
|
| 2754 | !rcu_segcblist_empty(&rdp->cblist),
|
| 2755 | need_resched(), is_idle_task(current),
|
| 2756 | rcu_is_callbacks_kthread());
|
| 2757 | return;
|
| 2758 | }
|
| 2759 |
|
| 2760 | |
| 2761 | * Extract the list of ready callbacks, disabling to prevent
|
| 2762 | * races with call_rcu() from interrupt handlers. Leave the
|
| 2763 | * callback counts, as rcu_barrier() needs to be conservative.
|
| 2764 |
|
| 2765 | local_irq_save(flags);
|
| 2766 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
|
| 2767 | bl = rdp->blimit;
|
| 2768 | trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
|
| 2769 | rcu_segcblist_n_cbs(&rdp->cblist), bl);
|
| 2770 | rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
|
| 2771 | local_irq_restore(flags);
|
| 2772 |
|
| 2773 |
|
| 2774 | rhp = rcu_cblist_dequeue(&rcl);
|
| 2775 | for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
|
| 2776 | debug_rcu_head_unqueue(rhp);
|
| 2777 | if (__rcu_reclaim(rsp->name, rhp))
|
| 2778 | rcu_cblist_dequeued_lazy(&rcl);
|
| 2779 | |
| 2780 | * Stop only if limit reached and CPU has something to do.
|
| 2781 | * Note: The rcl structure counts down from zero.
|
| 2782 |
|
| 2783 | if (-rcl.len >= bl &&
|
| 2784 | (need_resched() ||
|
| 2785 | (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
|
| 2786 | break;
|
| 2787 | }
|
| 2788 |
|
| 2789 | local_irq_save(flags);
|
| 2790 | count = -rcl.len;
|
| 2791 | trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
|
| 2792 | is_idle_task(current), rcu_is_callbacks_kthread());
|
| 2793 |
|
| 2794 |
|
| 2795 | rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
|
| 2796 | smp_mb();
|
| 2797 | rdp->n_cbs_invoked += count;
|
| 2798 | rcu_segcblist_insert_count(&rdp->cblist, &rcl);
|
| 2799 |
|
| 2800 |
|
| 2801 | count = rcu_segcblist_n_cbs(&rdp->cblist);
|
| 2802 | if (rdp->blimit == LONG_MAX && count <= qlowmark)
|
| 2803 | rdp->blimit = blimit;
|
| 2804 |
|
| 2805 |
|
| 2806 | if (count == 0 && rdp->qlen_last_fqs_check != 0) {
|
| 2807 | rdp->qlen_last_fqs_check = 0;
|
| 2808 | rdp->n_force_qs_snap = rsp->n_force_qs;
|
| 2809 | } else if (count < rdp->qlen_last_fqs_check - qhimark)
|
| 2810 | rdp->qlen_last_fqs_check = count;
|
| 2811 | WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
|
| 2812 |
|
| 2813 | local_irq_restore(flags);
|
| 2814 |
|
| 2815 |
|
| 2816 | if (rcu_segcblist_ready_cbs(&rdp->cblist))
|
| 2817 | invoke_rcu_core();
|
| 2818 | }
|
| 2819 |
|
| 2820 | |
| 2821 | * Check to see if this CPU is in a non-context-switch quiescent state
|
| 2822 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
|
| 2823 | * Also schedule RCU core processing.
|
| 2824 | *
|
| 2825 | * This function must be called from hardirq context. It is normally
|
| 2826 | * invoked from the scheduling-clock interrupt.
|
| 2827 |
|
| 2828 | void rcu_check_callbacks(int user)
|
| 2829 | {
|
| 2830 | trace_rcu_utilization(TPS("Start scheduler-tick"));
|
| 2831 | increment_cpu_stall_ticks();
|
| 2832 | if (user || rcu_is_cpu_rrupt_from_idle()) {
|
| 2833 |
|
| 2834 | |
| 2835 | * Get here if this CPU took its interrupt from user
|
| 2836 | * mode or from the idle loop, and if this is not a
|
| 2837 | * nested interrupt. In this case, the CPU is in
|
| 2838 | * a quiescent state, so note it.
|
| 2839 | *
|
| 2840 | * No memory barrier is required here because both
|
| 2841 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
|
| 2842 | * variables that other CPUs neither access nor modify,
|
| 2843 | * at least not while the corresponding CPU is online.
|
| 2844 |
|
| 2845 |
|
| 2846 | rcu_sched_qs();
|
| 2847 | rcu_bh_qs();
|
| 2848 |
|
| 2849 | } else if (!in_softirq()) {
|
| 2850 |
|
| 2851 | |
| 2852 | * Get here if this CPU did not take its interrupt from
|
| 2853 | * softirq, in other words, if it is not interrupting
|
| 2854 | * a rcu_bh read-side critical section. This is an _bh
|
| 2855 | * critical section, so note it.
|
| 2856 |
|
| 2857 |
|
| 2858 | rcu_bh_qs();
|
| 2859 | }
|
| 2860 | rcu_preempt_check_callbacks();
|
| 2861 | if (rcu_pending())
|
| 2862 | invoke_rcu_core();
|
| 2863 | if (user)
|
| 2864 | rcu_note_voluntary_context_switch(current);
|
| 2865 | trace_rcu_utilization(TPS("End scheduler-tick"));
|
| 2866 | }
|
| 2867 |
|
| 2868 | |
| 2869 | * Scan the leaf rcu_node structures, processing dyntick state for any that
|
| 2870 | * have not yet encountered a quiescent state, using the function specified.
|
| 2871 | * Also initiate boosting for any threads blocked on the root rcu_node.
|
| 2872 | *
|
| 2873 | * The caller must have suppressed start of new grace periods.
|
| 2874 |
|
| 2875 | static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
|
| 2876 | {
|
| 2877 | int cpu;
|
| 2878 | unsigned long flags;
|
| 2879 | unsigned long mask;
|
| 2880 | struct rcu_node *rnp;
|
| 2881 |
|
| 2882 | rcu_for_each_leaf_node(rsp, rnp) {
|
| 2883 | cond_resched_rcu_qs();
|
| 2884 | mask = 0;
|
| 2885 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 2886 | if (rnp->qsmask == 0) {
|
| 2887 | if (rcu_state_p == &rcu_sched_state ||
|
| 2888 | rsp != rcu_state_p ||
|
| 2889 | rcu_preempt_blocked_readers_cgp(rnp)) {
|
| 2890 | |
| 2891 | * No point in scanning bits because they
|
| 2892 | * are all zero. But we might need to
|
| 2893 | * priority-boost blocked readers.
|
| 2894 |
|
| 2895 | rcu_initiate_boost(rnp, flags);
|
| 2896 |
|
| 2897 | continue;
|
| 2898 | }
|
| 2899 | if (rnp->parent &&
|
| 2900 | (rnp->parent->qsmask & rnp->grpmask)) {
|
| 2901 | |
| 2902 | * Race between grace-period
|
| 2903 | * initialization and task exiting RCU
|
| 2904 | * read-side critical section: Report.
|
| 2905 |
|
| 2906 | rcu_report_unblock_qs_rnp(rsp, rnp, flags);
|
| 2907 |
|
| 2908 | continue;
|
| 2909 | }
|
| 2910 | }
|
| 2911 | for_each_leaf_node_possible_cpu(rnp, cpu) {
|
| 2912 | unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
|
| 2913 | if ((rnp->qsmask & bit) != 0) {
|
| 2914 | if (f(per_cpu_ptr(rsp->rda, cpu)))
|
| 2915 | mask |= bit;
|
| 2916 | }
|
| 2917 | }
|
| 2918 | if (mask != 0) {
|
| 2919 |
|
| 2920 | rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
|
| 2921 | } else {
|
| 2922 |
|
| 2923 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 2924 | }
|
| 2925 | }
|
| 2926 | }
|
| 2927 |
|
| 2928 | |
| 2929 | * Force quiescent states on reluctant CPUs, and also detect which
|
| 2930 | * CPUs are in dyntick-idle mode.
|
| 2931 |
|
| 2932 | static void force_quiescent_state(struct rcu_state *rsp)
|
| 2933 | {
|
| 2934 | unsigned long flags;
|
| 2935 | bool ret;
|
| 2936 | struct rcu_node *rnp;
|
| 2937 | struct rcu_node *rnp_old = NULL;
|
| 2938 |
|
| 2939 |
|
| 2940 | rnp = __this_cpu_read(rsp->rda->mynode);
|
| 2941 | for (; rnp != NULL; rnp = rnp->parent) {
|
| 2942 | ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
|
| 2943 | !raw_spin_trylock(&rnp->fqslock);
|
| 2944 | if (rnp_old != NULL)
|
| 2945 | raw_spin_unlock(&rnp_old->fqslock);
|
| 2946 | if (ret) {
|
| 2947 | rsp->n_force_qs_lh++;
|
| 2948 | return;
|
| 2949 | }
|
| 2950 | rnp_old = rnp;
|
| 2951 | }
|
| 2952 |
|
| 2953 |
|
| 2954 |
|
| 2955 | raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
|
| 2956 | raw_spin_unlock(&rnp_old->fqslock);
|
| 2957 | if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
|
| 2958 | rsp->n_force_qs_lh++;
|
| 2959 | raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
|
| 2960 | return;
|
| 2961 | }
|
| 2962 | WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
|
| 2963 | raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
|
| 2964 | rcu_gp_kthread_wake(rsp);
|
| 2965 | }
|
| 2966 |
|
| 2967 | |
| 2968 | * This does the RCU core processing work for the specified rcu_state
|
| 2969 | * and rcu_data structures. This may be called only from the CPU to
|
| 2970 | * whom the rdp belongs.
|
| 2971 |
|
| 2972 | static void
|
| 2973 | __rcu_process_callbacks(struct rcu_state *rsp)
|
| 2974 | {
|
| 2975 | unsigned long flags;
|
| 2976 | bool needwake;
|
| 2977 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
|
| 2978 |
|
| 2979 | WARN_ON_ONCE(!rdp->beenonline);
|
| 2980 |
|
| 2981 |
|
| 2982 | rcu_check_quiescent_state(rsp, rdp);
|
| 2983 |
|
| 2984 |
|
| 2985 | local_irq_save(flags);
|
| 2986 | if (cpu_needs_another_gp(rsp, rdp)) {
|
| 2987 | raw_spin_lock_rcu_node(rcu_get_root(rsp));
|
| 2988 | needwake = rcu_start_gp(rsp);
|
| 2989 | raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
|
| 2990 | if (needwake)
|
| 2991 | rcu_gp_kthread_wake(rsp);
|
| 2992 | } else {
|
| 2993 | local_irq_restore(flags);
|
| 2994 | }
|
| 2995 |
|
| 2996 |
|
| 2997 | if (rcu_segcblist_ready_cbs(&rdp->cblist))
|
| 2998 | invoke_rcu_callbacks(rsp, rdp);
|
| 2999 |
|
| 3000 |
|
| 3001 | do_nocb_deferred_wakeup(rdp);
|
| 3002 | }
|
| 3003 |
|
| 3004 | |
| 3005 | * Do RCU core processing for the current CPU.
|
| 3006 |
|
| 3007 | static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
|
| 3008 | {
|
| 3009 | struct rcu_state *rsp;
|
| 3010 |
|
| 3011 | if (cpu_is_offline(smp_processor_id()))
|
| 3012 | return;
|
| 3013 | trace_rcu_utilization(TPS("Start RCU core"));
|
| 3014 | for_each_rcu_flavor(rsp)
|
| 3015 | __rcu_process_callbacks(rsp);
|
| 3016 | trace_rcu_utilization(TPS("End RCU core"));
|
| 3017 | }
|
| 3018 |
|
| 3019 | |
| 3020 | * Schedule RCU callback invocation. If the specified type of RCU
|
| 3021 | * does not support RCU priority boosting, just do a direct call,
|
| 3022 | * otherwise wake up the per-CPU kernel kthread. Note that because we
|
| 3023 | * are running on the current CPU with softirqs disabled, the
|
| 3024 | * rcu_cpu_kthread_task cannot disappear out from under us.
|
| 3025 |
|
| 3026 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 3027 | {
|
| 3028 | if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
|
| 3029 | return;
|
| 3030 | if (likely(!rsp->boost)) {
|
| 3031 | rcu_do_batch(rsp, rdp);
|
| 3032 | return;
|
| 3033 | }
|
| 3034 | invoke_rcu_callbacks_kthread();
|
| 3035 | }
|
| 3036 |
|
| 3037 | static void invoke_rcu_core(void)
|
| 3038 | {
|
| 3039 | if (cpu_online(smp_processor_id()))
|
| 3040 | raise_softirq(RCU_SOFTIRQ);
|
| 3041 | }
|
| 3042 |
|
| 3043 | |
| 3044 | * Handle any core-RCU processing required by a call_rcu() invocation.
|
| 3045 |
|
| 3046 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
|
| 3047 | struct rcu_head *head, unsigned long flags)
|
| 3048 | {
|
| 3049 | bool needwake;
|
| 3050 |
|
| 3051 | |
| 3052 | * If called from an extended quiescent state, invoke the RCU
|
| 3053 | * core in order to force a re-evaluation of RCU's idleness.
|
| 3054 |
|
| 3055 | if (!rcu_is_watching())
|
| 3056 | invoke_rcu_core();
|
| 3057 |
|
| 3058 |
|
| 3059 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
|
| 3060 | return;
|
| 3061 |
|
| 3062 | |
| 3063 | * Force the grace period if too many callbacks or too long waiting.
|
| 3064 | * Enforce hysteresis, and don't invoke force_quiescent_state()
|
| 3065 | * if some other CPU has recently done so. Also, don't bother
|
| 3066 | * invoking force_quiescent_state() if the newly enqueued callback
|
| 3067 | * is the only one waiting for a grace period to complete.
|
| 3068 |
|
| 3069 | if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
|
| 3070 | rdp->qlen_last_fqs_check + qhimark)) {
|
| 3071 |
|
| 3072 |
|
| 3073 | note_gp_changes(rsp, rdp);
|
| 3074 |
|
| 3075 |
|
| 3076 | if (!rcu_gp_in_progress(rsp)) {
|
| 3077 | struct rcu_node *rnp_root = rcu_get_root(rsp);
|
| 3078 |
|
| 3079 | raw_spin_lock_rcu_node(rnp_root);
|
| 3080 | needwake = rcu_start_gp(rsp);
|
| 3081 | raw_spin_unlock_rcu_node(rnp_root);
|
| 3082 | if (needwake)
|
| 3083 | rcu_gp_kthread_wake(rsp);
|
| 3084 | } else {
|
| 3085 |
|
| 3086 | rdp->blimit = LONG_MAX;
|
| 3087 | if (rsp->n_force_qs == rdp->n_force_qs_snap &&
|
| 3088 | rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
|
| 3089 | force_quiescent_state(rsp);
|
| 3090 | rdp->n_force_qs_snap = rsp->n_force_qs;
|
| 3091 | rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
|
| 3092 | }
|
| 3093 | }
|
| 3094 | }
|
| 3095 |
|
| 3096 | |
| 3097 | * RCU callback function to leak a callback.
|
| 3098 |
|
| 3099 | static void rcu_leak_callback(struct rcu_head *rhp)
|
| 3100 | {
|
| 3101 | }
|
| 3102 |
|
| 3103 | |
| 3104 | * Helper function for call_rcu() and friends. The cpu argument will
|
| 3105 | * normally be -1, indicating "currently running CPU". It may specify
|
| 3106 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
|
| 3107 | * is expected to specify a CPU.
|
| 3108 |
|
| 3109 | static void
|
| 3110 | __call_rcu(struct rcu_head *head, rcu_callback_t func,
|
| 3111 | struct rcu_state *rsp, int cpu, bool lazy)
|
| 3112 | {
|
| 3113 | unsigned long flags;
|
| 3114 | struct rcu_data *rdp;
|
| 3115 |
|
| 3116 |
|
| 3117 | WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
|
| 3118 |
|
| 3119 | if (debug_rcu_head_queue(head)) {
|
| 3120 | |
| 3121 | * Probable double call_rcu(), so leak the callback.
|
| 3122 | * Use rcu:rcu_callback trace event to find the previous
|
| 3123 | * time callback was passed to __call_rcu().
|
| 3124 |
|
| 3125 | WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
|
| 3126 | head, head->func);
|
| 3127 | WRITE_ONCE(head->func, rcu_leak_callback);
|
| 3128 | return;
|
| 3129 | }
|
| 3130 | head->func = func;
|
| 3131 | head->next = NULL;
|
| 3132 | local_irq_save(flags);
|
| 3133 | rdp = this_cpu_ptr(rsp->rda);
|
| 3134 |
|
| 3135 |
|
| 3136 | if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
|
| 3137 | int offline;
|
| 3138 |
|
| 3139 | if (cpu != -1)
|
| 3140 | rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 3141 | if (likely(rdp->mynode)) {
|
| 3142 |
|
| 3143 | offline = !__call_rcu_nocb(rdp, head, lazy, flags);
|
| 3144 | WARN_ON_ONCE(offline);
|
| 3145 |
|
| 3146 | local_irq_restore(flags);
|
| 3147 | return;
|
| 3148 | }
|
| 3149 | |
| 3150 | * Very early boot, before rcu_init(). Initialize if needed
|
| 3151 | * and then drop through to queue the callback.
|
| 3152 |
|
| 3153 | BUG_ON(cpu != -1);
|
| 3154 | WARN_ON_ONCE(!rcu_is_watching());
|
| 3155 | if (rcu_segcblist_empty(&rdp->cblist))
|
| 3156 | rcu_segcblist_init(&rdp->cblist);
|
| 3157 | }
|
| 3158 | rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
|
| 3159 | if (!lazy)
|
| 3160 | rcu_idle_count_callbacks_posted();
|
| 3161 |
|
| 3162 | if (__is_kfree_rcu_offset((unsigned long)func))
|
| 3163 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
|
| 3164 | rcu_segcblist_n_lazy_cbs(&rdp->cblist),
|
| 3165 | rcu_segcblist_n_cbs(&rdp->cblist));
|
| 3166 | else
|
| 3167 | trace_rcu_callback(rsp->name, head,
|
| 3168 | rcu_segcblist_n_lazy_cbs(&rdp->cblist),
|
| 3169 | rcu_segcblist_n_cbs(&rdp->cblist));
|
| 3170 |
|
| 3171 |
|
| 3172 | __call_rcu_core(rsp, rdp, head, flags);
|
| 3173 | local_irq_restore(flags);
|
| 3174 | }
|
| 3175 |
|
| 3176 | |
| 3177 | * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
|
| 3178 | * @head: structure to be used for queueing the RCU updates.
|
| 3179 | * @func: actual callback function to be invoked after the grace period
|
| 3180 | *
|
| 3181 | * The callback function will be invoked some time after a full grace
|
| 3182 | * period elapses, in other words after all currently executing RCU
|
| 3183 | * read-side critical sections have completed. call_rcu_sched() assumes
|
| 3184 | * that the read-side critical sections end on enabling of preemption
|
| 3185 | * or on voluntary preemption.
|
| 3186 | * RCU read-side critical sections are delimited by :
|
| 3187 | * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
|
| 3188 | * - anything that disables preemption.
|
| 3189 | *
|
| 3190 | * These may be nested.
|
| 3191 | *
|
| 3192 | * See the description of call_rcu() for more detailed information on
|
| 3193 | * memory ordering guarantees.
|
| 3194 |
|
| 3195 | void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
|
| 3196 | {
|
| 3197 | __call_rcu(head, func, &rcu_sched_state, -1, 0);
|
| 3198 | }
|
| 3199 | EXPORT_SYMBOL_GPL(call_rcu_sched);
|
| 3200 |
|
| 3201 | |
| 3202 | * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
|
| 3203 | * @head: structure to be used for queueing the RCU updates.
|
| 3204 | * @func: actual callback function to be invoked after the grace period
|
| 3205 | *
|
| 3206 | * The callback function will be invoked some time after a full grace
|
| 3207 | * period elapses, in other words after all currently executing RCU
|
| 3208 | * read-side critical sections have completed. call_rcu_bh() assumes
|
| 3209 | * that the read-side critical sections end on completion of a softirq
|
| 3210 | * handler. This means that read-side critical sections in process
|
| 3211 | * context must not be interrupted by softirqs. This interface is to be
|
| 3212 | * used when most of the read-side critical sections are in softirq context.
|
| 3213 | * RCU read-side critical sections are delimited by :
|
| 3214 | * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context.
|
| 3215 | * OR
|
| 3216 | * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
|
| 3217 | * These may be nested.
|
| 3218 | *
|
| 3219 | * See the description of call_rcu() for more detailed information on
|
| 3220 | * memory ordering guarantees.
|
| 3221 |
|
| 3222 | void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
|
| 3223 | {
|
| 3224 | __call_rcu(head, func, &rcu_bh_state, -1, 0);
|
| 3225 | }
|
| 3226 | EXPORT_SYMBOL_GPL(call_rcu_bh);
|
| 3227 |
|
| 3228 | |
| 3229 | * Queue an RCU callback for lazy invocation after a grace period.
|
| 3230 | * This will likely be later named something like "call_rcu_lazy()",
|
| 3231 | * but this change will require some way of tagging the lazy RCU
|
| 3232 | * callbacks in the list of pending callbacks. Until then, this
|
| 3233 | * function may only be called from __kfree_rcu().
|
| 3234 |
|
| 3235 | void kfree_call_rcu(struct rcu_head *head,
|
| 3236 | rcu_callback_t func)
|
| 3237 | {
|
| 3238 | __call_rcu(head, func, rcu_state_p, -1, 1);
|
| 3239 | }
|
| 3240 | EXPORT_SYMBOL_GPL(kfree_call_rcu);
|
| 3241 |
|
| 3242 | |
| 3243 | * Because a context switch is a grace period for RCU-sched and RCU-bh,
|
| 3244 | * any blocking grace-period wait automatically implies a grace period
|
| 3245 | * if there is only one CPU online at any point time during execution
|
| 3246 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
|
| 3247 | * occasionally incorrectly indicate that there are multiple CPUs online
|
| 3248 | * when there was in fact only one the whole time, as this just adds
|
| 3249 | * some overhead: RCU still operates correctly.
|
| 3250 |
|
| 3251 | static inline int rcu_blocking_is_gp(void)
|
| 3252 | {
|
| 3253 | int ret;
|
| 3254 |
|
| 3255 | might_sleep();
|
| 3256 | preempt_disable();
|
| 3257 | ret = num_online_cpus() <= 1;
|
| 3258 | preempt_enable();
|
| 3259 | return ret;
|
| 3260 | }
|
| 3261 |
|
| 3262 | |
| 3263 | * synchronize_sched - wait until an rcu-sched grace period has elapsed.
|
| 3264 | *
|
| 3265 | * Control will return to the caller some time after a full rcu-sched
|
| 3266 | * grace period has elapsed, in other words after all currently executing
|
| 3267 | * rcu-sched read-side critical sections have completed. These read-side
|
| 3268 | * critical sections are delimited by rcu_read_lock_sched() and
|
| 3269 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
|
| 3270 | * local_irq_disable(), and so on may be used in place of
|
| 3271 | * rcu_read_lock_sched().
|
| 3272 | *
|
| 3273 | * This means that all preempt_disable code sequences, including NMI and
|
| 3274 | * non-threaded hardware-interrupt handlers, in progress on entry will
|
| 3275 | * have completed before this primitive returns. However, this does not
|
| 3276 | * guarantee that softirq handlers will have completed, since in some
|
| 3277 | * kernels, these handlers can run in process context, and can block.
|
| 3278 | *
|
| 3279 | * Note that this guarantee implies further memory-ordering guarantees.
|
| 3280 | * On systems with more than one CPU, when synchronize_sched() returns,
|
| 3281 | * each CPU is guaranteed to have executed a full memory barrier since the
|
| 3282 | * end of its last RCU-sched read-side critical section whose beginning
|
| 3283 | * preceded the call to synchronize_sched(). In addition, each CPU having
|
| 3284 | * an RCU read-side critical section that extends beyond the return from
|
| 3285 | * synchronize_sched() is guaranteed to have executed a full memory barrier
|
| 3286 | * after the beginning of synchronize_sched() and before the beginning of
|
| 3287 | * that RCU read-side critical section. Note that these guarantees include
|
| 3288 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs
|
| 3289 | * that are executing in the kernel.
|
| 3290 | *
|
| 3291 | * Furthermore, if CPU A invoked synchronize_sched(), which returned
|
| 3292 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed
|
| 3293 | * to have executed a full memory barrier during the execution of
|
| 3294 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
|
| 3295 | * again only if the system has more than one CPU).
|
| 3296 |
|
| 3297 | void synchronize_sched(void)
|
| 3298 | {
|
| 3299 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
|
| 3300 | lock_is_held(&rcu_lock_map) ||
|
| 3301 | lock_is_held(&rcu_sched_lock_map),
|
| 3302 | "Illegal synchronize_sched() in RCU-sched read-side critical section");
|
| 3303 | if (rcu_blocking_is_gp())
|
| 3304 | return;
|
| 3305 | if (rcu_gp_is_expedited())
|
| 3306 | synchronize_sched_expedited();
|
| 3307 | else
|
| 3308 | wait_rcu_gp(call_rcu_sched);
|
| 3309 | }
|
| 3310 | EXPORT_SYMBOL_GPL(synchronize_sched);
|
| 3311 |
|
| 3312 | |
| 3313 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
|
| 3314 | *
|
| 3315 | * Control will return to the caller some time after a full rcu_bh grace
|
| 3316 | * period has elapsed, in other words after all currently executing rcu_bh
|
| 3317 | * read-side critical sections have completed. RCU read-side critical
|
| 3318 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
|
| 3319 | * and may be nested.
|
| 3320 | *
|
| 3321 | * See the description of synchronize_sched() for more detailed information
|
| 3322 | * on memory ordering guarantees.
|
| 3323 |
|
| 3324 | void synchronize_rcu_bh(void)
|
| 3325 | {
|
| 3326 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
|
| 3327 | lock_is_held(&rcu_lock_map) ||
|
| 3328 | lock_is_held(&rcu_sched_lock_map),
|
| 3329 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
|
| 3330 | if (rcu_blocking_is_gp())
|
| 3331 | return;
|
| 3332 | if (rcu_gp_is_expedited())
|
| 3333 | synchronize_rcu_bh_expedited();
|
| 3334 | else
|
| 3335 | wait_rcu_gp(call_rcu_bh);
|
| 3336 | }
|
| 3337 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
|
| 3338 |
|
| 3339 | |
| 3340 | * get_state_synchronize_rcu - Snapshot current RCU state
|
| 3341 | *
|
| 3342 | * Returns a cookie that is used by a later call to cond_synchronize_rcu()
|
| 3343 | * to determine whether or not a full grace period has elapsed in the
|
| 3344 | * meantime.
|
| 3345 |
|
| 3346 | unsigned long get_state_synchronize_rcu(void)
|
| 3347 | {
|
| 3348 | |
| 3349 | * Any prior manipulation of RCU-protected data must happen
|
| 3350 | * before the load from ->gpnum.
|
| 3351 |
|
| 3352 | smp_mb();
|
| 3353 |
|
| 3354 | |
| 3355 | * Make sure this load happens before the purportedly
|
| 3356 | * time-consuming work between get_state_synchronize_rcu()
|
| 3357 | * and cond_synchronize_rcu().
|
| 3358 |
|
| 3359 | return smp_load_acquire(&rcu_state_p->gpnum);
|
| 3360 | }
|
| 3361 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
|
| 3362 |
|
| 3363 | |
| 3364 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period
|
| 3365 | *
|
| 3366 | * @oldstate: return value from earlier call to get_state_synchronize_rcu()
|
| 3367 | *
|
| 3368 | * If a full RCU grace period has elapsed since the earlier call to
|
| 3369 | * get_state_synchronize_rcu(), just return. Otherwise, invoke
|
| 3370 | * synchronize_rcu() to wait for a full grace period.
|
| 3371 | *
|
| 3372 | * Yes, this function does not take counter wrap into account. But
|
| 3373 | * counter wrap is harmless. If the counter wraps, we have waited for
|
| 3374 | * more than 2 billion grace periods (and way more on a 64-bit system!),
|
| 3375 | * so waiting for one additional grace period should be just fine.
|
| 3376 |
|
| 3377 | void cond_synchronize_rcu(unsigned long oldstate)
|
| 3378 | {
|
| 3379 | unsigned long newstate;
|
| 3380 |
|
| 3381 | |
| 3382 | * Ensure that this load happens before any RCU-destructive
|
| 3383 | * actions the caller might carry out after we return.
|
| 3384 |
|
| 3385 | newstate = smp_load_acquire(&rcu_state_p->completed);
|
| 3386 | if (ULONG_CMP_GE(oldstate, newstate))
|
| 3387 | synchronize_rcu();
|
| 3388 | }
|
| 3389 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
|
| 3390 |
|
| 3391 | |
| 3392 | * get_state_synchronize_sched - Snapshot current RCU-sched state
|
| 3393 | *
|
| 3394 | * Returns a cookie that is used by a later call to cond_synchronize_sched()
|
| 3395 | * to determine whether or not a full grace period has elapsed in the
|
| 3396 | * meantime.
|
| 3397 |
|
| 3398 | unsigned long get_state_synchronize_sched(void)
|
| 3399 | {
|
| 3400 | |
| 3401 | * Any prior manipulation of RCU-protected data must happen
|
| 3402 | * before the load from ->gpnum.
|
| 3403 |
|
| 3404 | smp_mb();
|
| 3405 |
|
| 3406 | |
| 3407 | * Make sure this load happens before the purportedly
|
| 3408 | * time-consuming work between get_state_synchronize_sched()
|
| 3409 | * and cond_synchronize_sched().
|
| 3410 |
|
| 3411 | return smp_load_acquire(&rcu_sched_state.gpnum);
|
| 3412 | }
|
| 3413 | EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
|
| 3414 |
|
| 3415 | |
| 3416 | * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
|
| 3417 | *
|
| 3418 | * @oldstate: return value from earlier call to get_state_synchronize_sched()
|
| 3419 | *
|
| 3420 | * If a full RCU-sched grace period has elapsed since the earlier call to
|
| 3421 | * get_state_synchronize_sched(), just return. Otherwise, invoke
|
| 3422 | * synchronize_sched() to wait for a full grace period.
|
| 3423 | *
|
| 3424 | * Yes, this function does not take counter wrap into account. But
|
| 3425 | * counter wrap is harmless. If the counter wraps, we have waited for
|
| 3426 | * more than 2 billion grace periods (and way more on a 64-bit system!),
|
| 3427 | * so waiting for one additional grace period should be just fine.
|
| 3428 |
|
| 3429 | void cond_synchronize_sched(unsigned long oldstate)
|
| 3430 | {
|
| 3431 | unsigned long newstate;
|
| 3432 |
|
| 3433 | |
| 3434 | * Ensure that this load happens before any RCU-destructive
|
| 3435 | * actions the caller might carry out after we return.
|
| 3436 |
|
| 3437 | newstate = smp_load_acquire(&rcu_sched_state.completed);
|
| 3438 | if (ULONG_CMP_GE(oldstate, newstate))
|
| 3439 | synchronize_sched();
|
| 3440 | }
|
| 3441 | EXPORT_SYMBOL_GPL(cond_synchronize_sched);
|
| 3442 |
|
| 3443 | |
| 3444 | * Check to see if there is any immediate RCU-related work to be done
|
| 3445 | * by the current CPU, for the specified type of RCU, returning 1 if so.
|
| 3446 | * The checks are in order of increasing expense: checks that can be
|
| 3447 | * carried out against CPU-local state are performed first. However,
|
| 3448 | * we must check for CPU stalls first, else we might not get a chance.
|
| 3449 |
|
| 3450 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 3451 | {
|
| 3452 | struct rcu_node *rnp = rdp->mynode;
|
| 3453 |
|
| 3454 | rdp->n_rcu_pending++;
|
| 3455 |
|
| 3456 |
|
| 3457 | check_cpu_stall(rsp, rdp);
|
| 3458 |
|
| 3459 |
|
| 3460 | if (rcu_nohz_full_cpu(rsp))
|
| 3461 | return 0;
|
| 3462 |
|
| 3463 |
|
| 3464 | if (rcu_scheduler_fully_active &&
|
| 3465 | rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
|
| 3466 | rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_dynticks.rcu_qs_ctr)) {
|
| 3467 | rdp->n_rp_core_needs_qs++;
|
| 3468 | } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
|
| 3469 | rdp->n_rp_report_qs++;
|
| 3470 | return 1;
|
| 3471 | }
|
| 3472 |
|
| 3473 |
|
| 3474 | if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
|
| 3475 | rdp->n_rp_cb_ready++;
|
| 3476 | return 1;
|
| 3477 | }
|
| 3478 |
|
| 3479 |
|
| 3480 | if (cpu_needs_another_gp(rsp, rdp)) {
|
| 3481 | rdp->n_rp_cpu_needs_gp++;
|
| 3482 | return 1;
|
| 3483 | }
|
| 3484 |
|
| 3485 |
|
| 3486 | if (READ_ONCE(rnp->completed) != rdp->completed) {
|
| 3487 | rdp->n_rp_gp_completed++;
|
| 3488 | return 1;
|
| 3489 | }
|
| 3490 |
|
| 3491 |
|
| 3492 | if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
|
| 3493 | unlikely(READ_ONCE(rdp->gpwrap))) {
|
| 3494 | rdp->n_rp_gp_started++;
|
| 3495 | return 1;
|
| 3496 | }
|
| 3497 |
|
| 3498 |
|
| 3499 | if (rcu_nocb_need_deferred_wakeup(rdp)) {
|
| 3500 | rdp->n_rp_nocb_defer_wakeup++;
|
| 3501 | return 1;
|
| 3502 | }
|
| 3503 |
|
| 3504 |
|
| 3505 | rdp->n_rp_need_nothing++;
|
| 3506 | return 0;
|
| 3507 | }
|
| 3508 |
|
| 3509 | |
| 3510 | * Check to see if there is any immediate RCU-related work to be done
|
| 3511 | * by the current CPU, returning 1 if so. This function is part of the
|
| 3512 | * RCU implementation; it is -not- an exported member of the RCU API.
|
| 3513 |
|
| 3514 | static int rcu_pending(void)
|
| 3515 | {
|
| 3516 | struct rcu_state *rsp;
|
| 3517 |
|
| 3518 | for_each_rcu_flavor(rsp)
|
| 3519 | if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
|
| 3520 | return 1;
|
| 3521 | return 0;
|
| 3522 | }
|
| 3523 |
|
| 3524 | |
| 3525 | * Return true if the specified CPU has any callback. If all_lazy is
|
| 3526 | * non-NULL, store an indication of whether all callbacks are lazy.
|
| 3527 | * (If there are no callbacks, all of them are deemed to be lazy.)
|
| 3528 |
|
| 3529 | static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
|
| 3530 | {
|
| 3531 | bool al = true;
|
| 3532 | bool hc = false;
|
| 3533 | struct rcu_data *rdp;
|
| 3534 | struct rcu_state *rsp;
|
| 3535 |
|
| 3536 | for_each_rcu_flavor(rsp) {
|
| 3537 | rdp = this_cpu_ptr(rsp->rda);
|
| 3538 | if (rcu_segcblist_empty(&rdp->cblist))
|
| 3539 | continue;
|
| 3540 | hc = true;
|
| 3541 | if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
|
| 3542 | al = false;
|
| 3543 | break;
|
| 3544 | }
|
| 3545 | }
|
| 3546 | if (all_lazy)
|
| 3547 | *all_lazy = al;
|
| 3548 | return hc;
|
| 3549 | }
|
| 3550 |
|
| 3551 | |
| 3552 | * Helper function for _rcu_barrier() tracing. If tracing is disabled,
|
| 3553 | * the compiler is expected to optimize this away.
|
| 3554 |
|
| 3555 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
|
| 3556 | int cpu, unsigned long done)
|
| 3557 | {
|
| 3558 | trace_rcu_barrier(rsp->name, s, cpu,
|
| 3559 | atomic_read(&rsp->barrier_cpu_count), done);
|
| 3560 | }
|
| 3561 |
|
| 3562 | |
| 3563 | * RCU callback function for _rcu_barrier(). If we are last, wake
|
| 3564 | * up the task executing _rcu_barrier().
|
| 3565 |
|
| 3566 | static void rcu_barrier_callback(struct rcu_head *rhp)
|
| 3567 | {
|
| 3568 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
|
| 3569 | struct rcu_state *rsp = rdp->rsp;
|
| 3570 |
|
| 3571 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
|
| 3572 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
|
| 3573 | complete(&rsp->barrier_completion);
|
| 3574 | } else {
|
| 3575 | _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
|
| 3576 | }
|
| 3577 | }
|
| 3578 |
|
| 3579 | |
| 3580 | * Called with preemption disabled, and from cross-cpu IRQ context.
|
| 3581 |
|
| 3582 | static void rcu_barrier_func(void *type)
|
| 3583 | {
|
| 3584 | struct rcu_state *rsp = type;
|
| 3585 | struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
|
| 3586 |
|
| 3587 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
|
| 3588 | rdp->barrier_head.func = rcu_barrier_callback;
|
| 3589 | debug_rcu_head_queue(&rdp->barrier_head);
|
| 3590 | if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
|
| 3591 | atomic_inc(&rsp->barrier_cpu_count);
|
| 3592 | } else {
|
| 3593 | debug_rcu_head_unqueue(&rdp->barrier_head);
|
| 3594 | _rcu_barrier_trace(rsp, "IRQNQ", -1, rsp->barrier_sequence);
|
| 3595 | }
|
| 3596 | }
|
| 3597 |
|
| 3598 | |
| 3599 | * Orchestrate the specified type of RCU barrier, waiting for all
|
| 3600 | * RCU callbacks of the specified type to complete.
|
| 3601 |
|
| 3602 | static void _rcu_barrier(struct rcu_state *rsp)
|
| 3603 | {
|
| 3604 | int cpu;
|
| 3605 | struct rcu_data *rdp;
|
| 3606 | unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
|
| 3607 |
|
| 3608 | _rcu_barrier_trace(rsp, "Begin", -1, s);
|
| 3609 |
|
| 3610 |
|
| 3611 | mutex_lock(&rsp->barrier_mutex);
|
| 3612 |
|
| 3613 |
|
| 3614 | if (rcu_seq_done(&rsp->barrier_sequence, s)) {
|
| 3615 | _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
|
| 3616 | smp_mb();
|
| 3617 | mutex_unlock(&rsp->barrier_mutex);
|
| 3618 | return;
|
| 3619 | }
|
| 3620 |
|
| 3621 |
|
| 3622 | rcu_seq_start(&rsp->barrier_sequence);
|
| 3623 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
|
| 3624 |
|
| 3625 | |
| 3626 | * Initialize the count to one rather than to zero in order to
|
| 3627 | * avoid a too-soon return to zero in case of a short grace period
|
| 3628 | * (or preemption of this task). Exclude CPU-hotplug operations
|
| 3629 | * to ensure that no offline CPU has callbacks queued.
|
| 3630 |
|
| 3631 | init_completion(&rsp->barrier_completion);
|
| 3632 | atomic_set(&rsp->barrier_cpu_count, 1);
|
| 3633 | get_online_cpus();
|
| 3634 |
|
| 3635 | |
| 3636 | * Force each CPU with callbacks to register a new callback.
|
| 3637 | * When that callback is invoked, we will know that all of the
|
| 3638 | * corresponding CPU's preceding callbacks have been invoked.
|
| 3639 |
|
| 3640 | for_each_possible_cpu(cpu) {
|
| 3641 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
|
| 3642 | continue;
|
| 3643 | rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 3644 | if (rcu_is_nocb_cpu(cpu)) {
|
| 3645 | if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
|
| 3646 | _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
|
| 3647 | rsp->barrier_sequence);
|
| 3648 | } else {
|
| 3649 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
|
| 3650 | rsp->barrier_sequence);
|
| 3651 | smp_mb__before_atomic();
|
| 3652 | atomic_inc(&rsp->barrier_cpu_count);
|
| 3653 | __call_rcu(&rdp->barrier_head,
|
| 3654 | rcu_barrier_callback, rsp, cpu, 0);
|
| 3655 | }
|
| 3656 | } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
|
| 3657 | _rcu_barrier_trace(rsp, "OnlineQ", cpu,
|
| 3658 | rsp->barrier_sequence);
|
| 3659 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
|
| 3660 | } else {
|
| 3661 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
|
| 3662 | rsp->barrier_sequence);
|
| 3663 | }
|
| 3664 | }
|
| 3665 | put_online_cpus();
|
| 3666 |
|
| 3667 | |
| 3668 | * Now that we have an rcu_barrier_callback() callback on each
|
| 3669 | * CPU, and thus each counted, remove the initial count.
|
| 3670 |
|
| 3671 | if (atomic_dec_and_test(&rsp->barrier_cpu_count))
|
| 3672 | complete(&rsp->barrier_completion);
|
| 3673 |
|
| 3674 |
|
| 3675 | wait_for_completion(&rsp->barrier_completion);
|
| 3676 |
|
| 3677 |
|
| 3678 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
|
| 3679 | rcu_seq_end(&rsp->barrier_sequence);
|
| 3680 |
|
| 3681 |
|
| 3682 | mutex_unlock(&rsp->barrier_mutex);
|
| 3683 | }
|
| 3684 |
|
| 3685 | |
| 3686 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
|
| 3687 |
|
| 3688 | void rcu_barrier_bh(void)
|
| 3689 | {
|
| 3690 | _rcu_barrier(&rcu_bh_state);
|
| 3691 | }
|
| 3692 | EXPORT_SYMBOL_GPL(rcu_barrier_bh);
|
| 3693 |
|
| 3694 | |
| 3695 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
|
| 3696 |
|
| 3697 | void rcu_barrier_sched(void)
|
| 3698 | {
|
| 3699 | _rcu_barrier(&rcu_sched_state);
|
| 3700 | }
|
| 3701 | EXPORT_SYMBOL_GPL(rcu_barrier_sched);
|
| 3702 |
|
| 3703 | |
| 3704 | * Propagate ->qsinitmask bits up the rcu_node tree to account for the
|
| 3705 | * first CPU in a given leaf rcu_node structure coming online. The caller
|
| 3706 | * must hold the corresponding leaf rcu_node ->lock with interrrupts
|
| 3707 | * disabled.
|
| 3708 |
|
| 3709 | static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
|
| 3710 | {
|
| 3711 | long mask;
|
| 3712 | struct rcu_node *rnp = rnp_leaf;
|
| 3713 |
|
| 3714 | lockdep_assert_held(&rnp->lock);
|
| 3715 | for (;;) {
|
| 3716 | mask = rnp->grpmask;
|
| 3717 | rnp = rnp->parent;
|
| 3718 | if (rnp == NULL)
|
| 3719 | return;
|
| 3720 | raw_spin_lock_rcu_node(rnp);
|
| 3721 | rnp->qsmaskinit |= mask;
|
| 3722 | raw_spin_unlock_rcu_node(rnp);
|
| 3723 | }
|
| 3724 | }
|
| 3725 |
|
| 3726 | |
| 3727 | * Do boot-time initialization of a CPU's per-CPU RCU data.
|
| 3728 |
|
| 3729 | static void __init
|
| 3730 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
|
| 3731 | {
|
| 3732 | unsigned long flags;
|
| 3733 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 3734 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 3735 |
|
| 3736 |
|
| 3737 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 3738 | rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
|
| 3739 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
|
| 3740 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
|
| 3741 | WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
|
| 3742 | rdp->cpu = cpu;
|
| 3743 | rdp->rsp = rsp;
|
| 3744 | rcu_boot_init_nocb_percpu_data(rdp);
|
| 3745 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 3746 | }
|
| 3747 |
|
| 3748 | |
| 3749 | * Initialize a CPU's per-CPU RCU data. Note that only one online or
|
| 3750 | * offline event can be happening at a given time. Note also that we
|
| 3751 | * can accept some slop in the rsp->completed access due to the fact
|
| 3752 | * that this CPU cannot possibly have any RCU callbacks in flight yet.
|
| 3753 |
|
| 3754 | static void
|
| 3755 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
|
| 3756 | {
|
| 3757 | unsigned long flags;
|
| 3758 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 3759 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 3760 |
|
| 3761 |
|
| 3762 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 3763 | rdp->qlen_last_fqs_check = 0;
|
| 3764 | rdp->n_force_qs_snap = rsp->n_force_qs;
|
| 3765 | rdp->blimit = blimit;
|
| 3766 | if (rcu_segcblist_empty(&rdp->cblist) &&
|
| 3767 | !init_nocb_callback_list(rdp))
|
| 3768 | rcu_segcblist_init(&rdp->cblist);
|
| 3769 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
|
| 3770 | rcu_dynticks_eqs_online();
|
| 3771 | raw_spin_unlock_rcu_node(rnp);
|
| 3772 |
|
| 3773 | |
| 3774 | * Add CPU to leaf rcu_node pending-online bitmask. Any needed
|
| 3775 | * propagation up the rcu_node tree will happen at the beginning
|
| 3776 | * of the next grace period.
|
| 3777 |
|
| 3778 | rnp = rdp->mynode;
|
| 3779 | raw_spin_lock_rcu_node(rnp);
|
| 3780 | if (!rdp->beenonline)
|
| 3781 | WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
|
| 3782 | rdp->beenonline = true;
|
| 3783 | rdp->gpnum = rnp->completed;
|
| 3784 | rdp->completed = rnp->completed;
|
| 3785 | rdp->cpu_no_qs.b.norm = true;
|
| 3786 | rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
|
| 3787 | rdp->core_needs_qs = false;
|
| 3788 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
|
| 3789 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 3790 | }
|
| 3791 |
|
| 3792 | |
| 3793 | * Invoked early in the CPU-online process, when pretty much all
|
| 3794 | * services are available. The incoming CPU is not present.
|
| 3795 |
|
| 3796 | int rcutree_prepare_cpu(unsigned int cpu)
|
| 3797 | {
|
| 3798 | struct rcu_state *rsp;
|
| 3799 |
|
| 3800 | for_each_rcu_flavor(rsp)
|
| 3801 | rcu_init_percpu_data(cpu, rsp);
|
| 3802 |
|
| 3803 | rcu_prepare_kthreads(cpu);
|
| 3804 | rcu_spawn_all_nocb_kthreads(cpu);
|
| 3805 |
|
| 3806 | return 0;
|
| 3807 | }
|
| 3808 |
|
| 3809 | |
| 3810 | * Update RCU priority boot kthread affinity for CPU-hotplug changes.
|
| 3811 |
|
| 3812 | static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
|
| 3813 | {
|
| 3814 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
|
| 3815 |
|
| 3816 | rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
|
| 3817 | }
|
| 3818 |
|
| 3819 | |
| 3820 | * Near the end of the CPU-online process. Pretty much all services
|
| 3821 | * enabled, and the CPU is now very much alive.
|
| 3822 |
|
| 3823 | int rcutree_online_cpu(unsigned int cpu)
|
| 3824 | {
|
| 3825 | sync_sched_exp_online_cleanup(cpu);
|
| 3826 | rcutree_affinity_setting(cpu, -1);
|
| 3827 | if (IS_ENABLED(CONFIG_TREE_SRCU))
|
| 3828 | srcu_online_cpu(cpu);
|
| 3829 | return 0;
|
| 3830 | }
|
| 3831 |
|
| 3832 | |
| 3833 | * Near the beginning of the process. The CPU is still very much alive
|
| 3834 | * with pretty much all services enabled.
|
| 3835 |
|
| 3836 | int rcutree_offline_cpu(unsigned int cpu)
|
| 3837 | {
|
| 3838 | rcutree_affinity_setting(cpu, cpu);
|
| 3839 | if (IS_ENABLED(CONFIG_TREE_SRCU))
|
| 3840 | srcu_offline_cpu(cpu);
|
| 3841 | return 0;
|
| 3842 | }
|
| 3843 |
|
| 3844 | |
| 3845 | * Near the end of the offline process. We do only tracing here.
|
| 3846 |
|
| 3847 | int rcutree_dying_cpu(unsigned int cpu)
|
| 3848 | {
|
| 3849 | struct rcu_state *rsp;
|
| 3850 |
|
| 3851 | for_each_rcu_flavor(rsp)
|
| 3852 | rcu_cleanup_dying_cpu(rsp);
|
| 3853 | return 0;
|
| 3854 | }
|
| 3855 |
|
| 3856 | |
| 3857 | * The outgoing CPU is gone and we are running elsewhere.
|
| 3858 |
|
| 3859 | int rcutree_dead_cpu(unsigned int cpu)
|
| 3860 | {
|
| 3861 | struct rcu_state *rsp;
|
| 3862 |
|
| 3863 | for_each_rcu_flavor(rsp) {
|
| 3864 | rcu_cleanup_dead_cpu(cpu, rsp);
|
| 3865 | do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
|
| 3866 | }
|
| 3867 | return 0;
|
| 3868 | }
|
| 3869 |
|
| 3870 | |
| 3871 | * Mark the specified CPU as being online so that subsequent grace periods
|
| 3872 | * (both expedited and normal) will wait on it. Note that this means that
|
| 3873 | * incoming CPUs are not allowed to use RCU read-side critical sections
|
| 3874 | * until this function is called. Failing to observe this restriction
|
| 3875 | * will result in lockdep splats.
|
| 3876 | *
|
| 3877 | * Note that this function is special in that it is invoked directly
|
| 3878 | * from the incoming CPU rather than from the cpuhp_step mechanism.
|
| 3879 | * This is because this function must be invoked at a precise location.
|
| 3880 |
|
| 3881 | void rcu_cpu_starting(unsigned int cpu)
|
| 3882 | {
|
| 3883 | unsigned long flags;
|
| 3884 | unsigned long mask;
|
| 3885 | struct rcu_data *rdp;
|
| 3886 | struct rcu_node *rnp;
|
| 3887 | struct rcu_state *rsp;
|
| 3888 |
|
| 3889 | for_each_rcu_flavor(rsp) {
|
| 3890 | rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 3891 | rnp = rdp->mynode;
|
| 3892 | mask = rdp->grpmask;
|
| 3893 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 3894 | rnp->qsmaskinitnext |= mask;
|
| 3895 | rnp->expmaskinitnext |= mask;
|
| 3896 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 3897 | }
|
| 3898 | }
|
| 3899 |
|
| 3900 | #ifdef CONFIG_HOTPLUG_CPU
|
| 3901 | |
| 3902 | * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
|
| 3903 | * function. We now remove it from the rcu_node tree's ->qsmaskinit
|
| 3904 | * bit masks.
|
| 3905 |
|
| 3906 | static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
|
| 3907 | {
|
| 3908 | unsigned long flags;
|
| 3909 | unsigned long mask;
|
| 3910 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 3911 | struct rcu_node *rnp = rdp->mynode;
|
| 3912 |
|
| 3913 |
|
| 3914 | mask = rdp->grpmask;
|
| 3915 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 3916 | rnp->qsmaskinitnext &= ~mask;
|
| 3917 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 3918 | }
|
| 3919 |
|
| 3920 | |
| 3921 | * The outgoing function has no further need of RCU, so remove it from
|
| 3922 | * the list of CPUs that RCU must track.
|
| 3923 | *
|
| 3924 | * Note that this function is special in that it is invoked directly
|
| 3925 | * from the outgoing CPU rather than from the cpuhp_step mechanism.
|
| 3926 | * This is because this function must be invoked at a precise location.
|
| 3927 |
|
| 3928 | void rcu_report_dead(unsigned int cpu)
|
| 3929 | {
|
| 3930 | struct rcu_state *rsp;
|
| 3931 |
|
| 3932 |
|
| 3933 | preempt_disable();
|
| 3934 | rcu_report_exp_rdp(&rcu_sched_state,
|
| 3935 | this_cpu_ptr(rcu_sched_state.rda), true);
|
| 3936 | preempt_enable();
|
| 3937 | for_each_rcu_flavor(rsp)
|
| 3938 | rcu_cleanup_dying_idle_cpu(cpu, rsp);
|
| 3939 | }
|
| 3940 | #endif
|
| 3941 |
|
| 3942 | |
| 3943 | * On non-huge systems, use expedited RCU grace periods to make suspend
|
| 3944 | * and hibernation run faster.
|
| 3945 |
|
| 3946 | static int rcu_pm_notify(struct notifier_block *self,
|
| 3947 | unsigned long action, void *hcpu)
|
| 3948 | {
|
| 3949 | switch (action) {
|
| 3950 | case PM_HIBERNATION_PREPARE:
|
| 3951 | case PM_SUSPEND_PREPARE:
|
| 3952 | if (nr_cpu_ids <= 256)
|
| 3953 | rcu_expedite_gp();
|
| 3954 | break;
|
| 3955 | case PM_POST_HIBERNATION:
|
| 3956 | case PM_POST_SUSPEND:
|
| 3957 | if (nr_cpu_ids <= 256)
|
| 3958 | rcu_unexpedite_gp();
|
| 3959 | break;
|
| 3960 | default:
|
| 3961 | break;
|
| 3962 | }
|
| 3963 | return NOTIFY_OK;
|
| 3964 | }
|
| 3965 |
|
| 3966 | |
| 3967 | * Spawn the kthreads that handle each RCU flavor's grace periods.
|
| 3968 |
|
| 3969 | static int __init rcu_spawn_gp_kthread(void)
|
| 3970 | {
|
| 3971 | unsigned long flags;
|
| 3972 | int kthread_prio_in = kthread_prio;
|
| 3973 | struct rcu_node *rnp;
|
| 3974 | struct rcu_state *rsp;
|
| 3975 | struct sched_param sp;
|
| 3976 | struct task_struct *t;
|
| 3977 |
|
| 3978 |
|
| 3979 | if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
|
| 3980 | kthread_prio = 1;
|
| 3981 | else if (kthread_prio < 0)
|
| 3982 | kthread_prio = 0;
|
| 3983 | else if (kthread_prio > 99)
|
| 3984 | kthread_prio = 99;
|
| 3985 | if (kthread_prio != kthread_prio_in)
|
| 3986 | pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
|
| 3987 | kthread_prio, kthread_prio_in);
|
| 3988 |
|
| 3989 | rcu_scheduler_fully_active = 1;
|
| 3990 | for_each_rcu_flavor(rsp) {
|
| 3991 | t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
|
| 3992 | BUG_ON(IS_ERR(t));
|
| 3993 | rnp = rcu_get_root(rsp);
|
| 3994 | raw_spin_lock_irqsave_rcu_node(rnp, flags);
|
| 3995 | rsp->gp_kthread = t;
|
| 3996 | if (kthread_prio) {
|
| 3997 | sp.sched_priority = kthread_prio;
|
| 3998 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
|
| 3999 | }
|
| 4000 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
|
| 4001 | wake_up_process(t);
|
| 4002 | }
|
| 4003 | rcu_spawn_nocb_kthreads();
|
| 4004 | rcu_spawn_boost_kthreads();
|
| 4005 | return 0;
|
| 4006 | }
|
| 4007 | early_initcall(rcu_spawn_gp_kthread);
|
| 4008 |
|
| 4009 | |
| 4010 | * This function is invoked towards the end of the scheduler's
|
| 4011 | * initialization process. Before this is called, the idle task might
|
| 4012 | * contain synchronous grace-period primitives (during which time, this idle
|
| 4013 | * task is booting the system, and such primitives are no-ops). After this
|
| 4014 | * function is called, any synchronous grace-period primitives are run as
|
| 4015 | * expedited, with the requesting task driving the grace period forward.
|
| 4016 | * A later core_initcall() rcu_set_runtime_mode() will switch to full
|
| 4017 | * runtime RCU functionality.
|
| 4018 |
|
| 4019 | void rcu_scheduler_starting(void)
|
| 4020 | {
|
| 4021 | WARN_ON(num_online_cpus() != 1);
|
| 4022 | WARN_ON(nr_context_switches() > 0);
|
| 4023 | rcu_test_sync_prims();
|
| 4024 | rcu_scheduler_active = RCU_SCHEDULER_INIT;
|
| 4025 | rcu_test_sync_prims();
|
| 4026 | }
|
| 4027 |
|
| 4028 | |
| 4029 | * Helper function for rcu_init() that initializes one rcu_state structure.
|
| 4030 |
|
| 4031 | static void __init rcu_init_one(struct rcu_state *rsp)
|
| 4032 | {
|
| 4033 | static const char * const buf[] = RCU_NODE_NAME_INIT;
|
| 4034 | static const char * const fqs[] = RCU_FQS_NAME_INIT;
|
| 4035 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
|
| 4036 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
|
| 4037 |
|
| 4038 | int levelspread[RCU_NUM_LVLS];
|
| 4039 | int cpustride = 1;
|
| 4040 | int i;
|
| 4041 | int j;
|
| 4042 | struct rcu_node *rnp;
|
| 4043 |
|
| 4044 | BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf));
|
| 4045 |
|
| 4046 |
|
| 4047 | if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
|
| 4048 | panic("rcu_init_one: rcu_num_lvls out of range");
|
| 4049 |
|
| 4050 |
|
| 4051 |
|
| 4052 | for (i = 1; i < rcu_num_lvls; i++)
|
| 4053 | rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
|
| 4054 | rcu_init_levelspread(levelspread, num_rcu_lvl);
|
| 4055 |
|
| 4056 |
|
| 4057 |
|
| 4058 | for (i = rcu_num_lvls - 1; i >= 0; i--) {
|
| 4059 | cpustride *= levelspread[i];
|
| 4060 | rnp = rsp->level[i];
|
| 4061 | for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
|
| 4062 | raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
|
| 4063 | lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
|
| 4064 | &rcu_node_class[i], buf[i]);
|
| 4065 | raw_spin_lock_init(&rnp->fqslock);
|
| 4066 | lockdep_set_class_and_name(&rnp->fqslock,
|
| 4067 | &rcu_fqs_class[i], fqs[i]);
|
| 4068 | rnp->gpnum = rsp->gpnum;
|
| 4069 | rnp->completed = rsp->completed;
|
| 4070 | rnp->qsmask = 0;
|
| 4071 | rnp->qsmaskinit = 0;
|
| 4072 | rnp->grplo = j * cpustride;
|
| 4073 | rnp->grphi = (j + 1) * cpustride - 1;
|
| 4074 | if (rnp->grphi >= nr_cpu_ids)
|
| 4075 | rnp->grphi = nr_cpu_ids - 1;
|
| 4076 | if (i == 0) {
|
| 4077 | rnp->grpnum = 0;
|
| 4078 | rnp->grpmask = 0;
|
| 4079 | rnp->parent = NULL;
|
| 4080 | } else {
|
| 4081 | rnp->grpnum = j % levelspread[i - 1];
|
| 4082 | rnp->grpmask = 1UL << rnp->grpnum;
|
| 4083 | rnp->parent = rsp->level[i - 1] +
|
| 4084 | j / levelspread[i - 1];
|
| 4085 | }
|
| 4086 | rnp->level = i;
|
| 4087 | INIT_LIST_HEAD(&rnp->blkd_tasks);
|
| 4088 | rcu_init_one_nocb(rnp);
|
| 4089 | init_waitqueue_head(&rnp->exp_wq[0]);
|
| 4090 | init_waitqueue_head(&rnp->exp_wq[1]);
|
| 4091 | init_waitqueue_head(&rnp->exp_wq[2]);
|
| 4092 | init_waitqueue_head(&rnp->exp_wq[3]);
|
| 4093 | spin_lock_init(&rnp->exp_lock);
|
| 4094 | }
|
| 4095 | }
|
| 4096 |
|
| 4097 | init_swait_queue_head(&rsp->gp_wq);
|
| 4098 | init_swait_queue_head(&rsp->expedited_wq);
|
| 4099 | rnp = rsp->level[rcu_num_lvls - 1];
|
| 4100 | for_each_possible_cpu(i) {
|
| 4101 | while (i > rnp->grphi)
|
| 4102 | rnp++;
|
| 4103 | per_cpu_ptr(rsp->rda, i)->mynode = rnp;
|
| 4104 | rcu_boot_init_percpu_data(i, rsp);
|
| 4105 | }
|
| 4106 | list_add(&rsp->flavors, &rcu_struct_flavors);
|
| 4107 | }
|
| 4108 |
|
| 4109 | |
| 4110 | * Compute the rcu_node tree geometry from kernel parameters. This cannot
|
| 4111 | * replace the definitions in tree.h because those are needed to size
|
| 4112 | * the ->node array in the rcu_state structure.
|
| 4113 |
|
| 4114 | static void __init rcu_init_geometry(void)
|
| 4115 | {
|
| 4116 | ulong d;
|
| 4117 | int i;
|
| 4118 | int rcu_capacity[RCU_NUM_LVLS];
|
| 4119 |
|
| 4120 | |
| 4121 | * Initialize any unspecified boot parameters.
|
| 4122 | * The default values of jiffies_till_first_fqs and
|
| 4123 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
|
| 4124 | * value, which is a function of HZ, then adding one for each
|
| 4125 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
|
| 4126 |
|
| 4127 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
|
| 4128 | if (jiffies_till_first_fqs == ULONG_MAX)
|
| 4129 | jiffies_till_first_fqs = d;
|
| 4130 | if (jiffies_till_next_fqs == ULONG_MAX)
|
| 4131 | jiffies_till_next_fqs = d;
|
| 4132 |
|
| 4133 |
|
| 4134 | if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
|
| 4135 | nr_cpu_ids == NR_CPUS)
|
| 4136 | return;
|
| 4137 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
|
| 4138 | rcu_fanout_leaf, nr_cpu_ids);
|
| 4139 |
|
| 4140 | |
| 4141 | * The boot-time rcu_fanout_leaf parameter must be at least two
|
| 4142 | * and cannot exceed the number of bits in the rcu_node masks.
|
| 4143 | * Complain and fall back to the compile-time values if this
|
| 4144 | * limit is exceeded.
|
| 4145 |
|
| 4146 | if (rcu_fanout_leaf < 2 ||
|
| 4147 | rcu_fanout_leaf > sizeof(unsigned long) * 8) {
|
| 4148 | rcu_fanout_leaf = RCU_FANOUT_LEAF;
|
| 4149 | WARN_ON(1);
|
| 4150 | return;
|
| 4151 | }
|
| 4152 |
|
| 4153 | |
| 4154 | * Compute number of nodes that can be handled an rcu_node tree
|
| 4155 | * with the given number of levels.
|
| 4156 |
|
| 4157 | rcu_capacity[0] = rcu_fanout_leaf;
|
| 4158 | for (i = 1; i < RCU_NUM_LVLS; i++)
|
| 4159 | rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
|
| 4160 |
|
| 4161 | |
| 4162 | * The tree must be able to accommodate the configured number of CPUs.
|
| 4163 | * If this limit is exceeded, fall back to the compile-time values.
|
| 4164 |
|
| 4165 | if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
|
| 4166 | rcu_fanout_leaf = RCU_FANOUT_LEAF;
|
| 4167 | WARN_ON(1);
|
| 4168 | return;
|
| 4169 | }
|
| 4170 |
|
| 4171 |
|
| 4172 | for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
|
| 4173 | }
|
| 4174 | rcu_num_lvls = i + 1;
|
| 4175 |
|
| 4176 |
|
| 4177 | for (i = 0; i < rcu_num_lvls; i++) {
|
| 4178 | int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
|
| 4179 | num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
|
| 4180 | }
|
| 4181 |
|
| 4182 |
|
| 4183 | rcu_num_nodes = 0;
|
| 4184 | for (i = 0; i < rcu_num_lvls; i++)
|
| 4185 | rcu_num_nodes += num_rcu_lvl[i];
|
| 4186 | }
|
| 4187 |
|
| 4188 | |
| 4189 | * Dump out the structure of the rcu_node combining tree associated
|
| 4190 | * with the rcu_state structure referenced by rsp.
|
| 4191 |
|
| 4192 | static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
|
| 4193 | {
|
| 4194 | int level = 0;
|
| 4195 | struct rcu_node *rnp;
|
| 4196 |
|
| 4197 | pr_info("rcu_node tree layout dump\n");
|
| 4198 | pr_info(" ");
|
| 4199 | rcu_for_each_node_breadth_first(rsp, rnp) {
|
| 4200 | if (rnp->level != level) {
|
| 4201 | pr_cont("\n");
|
| 4202 | pr_info(" ");
|
| 4203 | level = rnp->level;
|
| 4204 | }
|
| 4205 | pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
|
| 4206 | }
|
| 4207 | pr_cont("\n");
|
| 4208 | }
|
| 4209 |
|
| 4210 | void __init rcu_init(void)
|
| 4211 | {
|
| 4212 | int cpu;
|
| 4213 |
|
| 4214 | rcu_early_boot_tests();
|
| 4215 |
|
| 4216 | rcu_bootup_announce();
|
| 4217 | rcu_init_geometry();
|
| 4218 | rcu_init_one(&rcu_bh_state);
|
| 4219 | rcu_init_one(&rcu_sched_state);
|
| 4220 | if (dump_tree)
|
| 4221 | rcu_dump_rcu_node_tree(&rcu_sched_state);
|
| 4222 | __rcu_init_preempt();
|
| 4223 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
|
| 4224 |
|
| 4225 | |
| 4226 | * We don't need protection against CPU-hotplug here because
|
| 4227 | * this is called early in boot, before either interrupts
|
| 4228 | * or the scheduler are operational.
|
| 4229 |
|
| 4230 | pm_notifier(rcu_pm_notify, 0);
|
| 4231 | for_each_online_cpu(cpu) {
|
| 4232 | rcutree_prepare_cpu(cpu);
|
| 4233 | rcu_cpu_starting(cpu);
|
| 4234 | if (IS_ENABLED(CONFIG_TREE_SRCU))
|
| 4235 | srcu_online_cpu(cpu);
|
| 4236 | }
|
| 4237 | }
|
| 4238 |
|
| 4239 | #include "tree_exp.h"
|
| 4240 | #include "tree_plugin.h"
|