| 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, write to the Free Software
|
| 16 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
|
| 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.h>
|
| 36 | #include <linux/interrupt.h>
|
| 37 | #include <linux/sched.h>
|
| 38 | #include <linux/nmi.h>
|
| 39 | #include <linux/atomic.h>
|
| 40 | #include <linux/bitops.h>
|
| 41 | #include <linux/export.h>
|
| 42 | #include <linux/completion.h>
|
| 43 | #include <linux/moduleparam.h>
|
| 44 | #include <linux/module.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 <linux/prefetch.h>
|
| 54 | #include <linux/delay.h>
|
| 55 | #include <linux/stop_machine.h>
|
| 56 | #include <linux/random.h>
|
| 57 | #include <linux/ftrace_event.h>
|
| 58 | #include <linux/suspend.h>
|
| 59 |
|
| 60 | #include "tree.h"
|
| 61 | #include <trace/events/rcu.h>
|
| 62 |
|
| 63 | #include "rcu.h"
|
| 64 |
|
| 65 | MODULE_ALIAS("rcutree");
|
| 66 | #ifdef MODULE_PARAM_PREFIX
|
| 67 | #undef MODULE_PARAM_PREFIX
|
| 68 | #endif
|
| 69 | #define MODULE_PARAM_PREFIX "rcutree."
|
| 70 |
|
| 71 |
|
| 72 |
|
| 73 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
|
| 74 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
|
| 75 |
|
| 76 | |
| 77 | * In order to export the rcu_state name to the tracing tools, it
|
| 78 | * needs to be added in the __tracepoint_string section.
|
| 79 | * This requires defining a separate variable tp_<sname>_varname
|
| 80 | * that points to the string being used, and this will allow
|
| 81 | * the tracing userspace tools to be able to decipher the string
|
| 82 | * address to the matching string.
|
| 83 |
|
| 84 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
|
| 85 | static char sname##_varname[] = #sname; \
|
| 86 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
|
| 87 | struct rcu_state sname##_state = { \
|
| 88 | .level = { &sname##_state.node[0] }, \
|
| 89 | .call = cr, \
|
| 90 | .fqs_state = RCU_GP_IDLE, \
|
| 91 | .gpnum = 0UL - 300UL, \
|
| 92 | .completed = 0UL - 300UL, \
|
| 93 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
|
| 94 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \
|
| 95 | .orphan_donetail = &sname##_state.orphan_donelist, \
|
| 96 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
|
| 97 | .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
|
| 98 | .name = sname##_varname, \
|
| 99 | .abbr = sabbr, \
|
| 100 | }; \
|
| 101 | DEFINE_PER_CPU(struct rcu_data, sname##_data)
|
| 102 |
|
| 103 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
|
| 104 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
|
| 105 |
|
| 106 | static struct rcu_state *rcu_state;
|
| 107 | LIST_HEAD(rcu_struct_flavors);
|
| 108 |
|
| 109 |
|
| 110 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
|
| 111 | module_param(rcu_fanout_leaf, int, 0444);
|
| 112 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
|
| 113 | static int num_rcu_lvl[] = {
|
| 114 | NUM_RCU_LVL_0,
|
| 115 | NUM_RCU_LVL_1,
|
| 116 | NUM_RCU_LVL_2,
|
| 117 | NUM_RCU_LVL_3,
|
| 118 | NUM_RCU_LVL_4,
|
| 119 | };
|
| 120 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES;
|
| 121 |
|
| 122 | |
| 123 | * The rcu_scheduler_active variable transitions from zero to one just
|
| 124 | * before the first task is spawned. So when this variable is zero, RCU
|
| 125 | * can assume that there is but one task, allowing RCU to (for example)
|
| 126 | * optimize synchronize_sched() to a simple barrier(). When this variable
|
| 127 | * is one, RCU must actually do all the hard work required to detect real
|
| 128 | * grace periods. This variable is also used to suppress boot-time false
|
| 129 | * positives from lockdep-RCU error checking.
|
| 130 |
|
| 131 | int rcu_scheduler_active __read_mostly;
|
| 132 | EXPORT_SYMBOL_GPL(rcu_scheduler_active);
|
| 133 |
|
| 134 | |
| 135 | * The rcu_scheduler_fully_active variable transitions from zero to one
|
| 136 | * during the early_initcall() processing, which is after the scheduler
|
| 137 | * is capable of creating new tasks. So RCU processing (for example,
|
| 138 | * creating tasks for RCU priority boosting) must be delayed until after
|
| 139 | * rcu_scheduler_fully_active transitions from zero to one. We also
|
| 140 | * currently delay invocation of any RCU callbacks until after this point.
|
| 141 | *
|
| 142 | * It might later prove better for people registering RCU callbacks during
|
| 143 | * early boot to take responsibility for these callbacks, but one step at
|
| 144 | * a time.
|
| 145 |
|
| 146 | static int rcu_scheduler_fully_active __read_mostly;
|
| 147 |
|
| 148 | #ifdef CONFIG_RCU_BOOST
|
| 149 |
|
| 150 | |
| 151 | * Control variables for per-CPU and per-rcu_node kthreads. These
|
| 152 | * handle all flavors of RCU.
|
| 153 |
|
| 154 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
|
| 155 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
|
| 156 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
|
| 157 | DEFINE_PER_CPU(char, rcu_cpu_has_work);
|
| 158 |
|
| 159 | #endif
|
| 160 |
|
| 161 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
|
| 162 | static void invoke_rcu_core(void);
|
| 163 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
|
| 164 |
|
| 165 | |
| 166 | * Track the rcutorture test sequence number and the update version
|
| 167 | * number within a given test. The rcutorture_testseq is incremented
|
| 168 | * on every rcutorture module load and unload, so has an odd value
|
| 169 | * when a test is running. The rcutorture_vernum is set to zero
|
| 170 | * when rcutorture starts and is incremented on each rcutorture update.
|
| 171 | * These variables enable correlating rcutorture output with the
|
| 172 | * RCU tracing information.
|
| 173 |
|
| 174 | unsigned long rcutorture_testseq;
|
| 175 | unsigned long rcutorture_vernum;
|
| 176 |
|
| 177 | |
| 178 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
|
| 179 | * permit this function to be invoked without holding the root rcu_node
|
| 180 | * structure's ->lock, but of course results can be subject to change.
|
| 181 |
|
| 182 | static int rcu_gp_in_progress(struct rcu_state *rsp)
|
| 183 | {
|
| 184 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
|
| 185 | }
|
| 186 |
|
| 187 | |
| 188 | * Note a quiescent state. Because we do not need to know
|
| 189 | * how many quiescent states passed, just if there was at least
|
| 190 | * one since the start of the grace period, this just sets a flag.
|
| 191 | * The caller must have disabled preemption.
|
| 192 |
|
| 193 | void rcu_sched_qs(int cpu)
|
| 194 | {
|
| 195 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
|
| 196 |
|
| 197 | if (rdp->passed_quiesce == 0)
|
| 198 | trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
|
| 199 | rdp->passed_quiesce = 1;
|
| 200 | }
|
| 201 |
|
| 202 | void rcu_bh_qs(int cpu)
|
| 203 | {
|
| 204 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
|
| 205 |
|
| 206 | if (rdp->passed_quiesce == 0)
|
| 207 | trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
|
| 208 | rdp->passed_quiesce = 1;
|
| 209 | }
|
| 210 |
|
| 211 | |
| 212 | * Note a context switch. This is a quiescent state for RCU-sched,
|
| 213 | * and requires special handling for preemptible RCU.
|
| 214 | * The caller must have disabled preemption.
|
| 215 |
|
| 216 | void rcu_note_context_switch(int cpu)
|
| 217 | {
|
| 218 | trace_rcu_utilization(TPS("Start context switch"));
|
| 219 | rcu_sched_qs(cpu);
|
| 220 | rcu_preempt_note_context_switch(cpu);
|
| 221 | trace_rcu_utilization(TPS("End context switch"));
|
| 222 | }
|
| 223 | EXPORT_SYMBOL_GPL(rcu_note_context_switch);
|
| 224 |
|
| 225 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
|
| 226 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
|
| 227 | .dynticks = ATOMIC_INIT(1),
|
| 228 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
|
| 229 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
|
| 230 | .dynticks_idle = ATOMIC_INIT(1),
|
| 231 | #endif
|
| 232 | };
|
| 233 |
|
| 234 | static long blimit = 10;
|
| 235 | static long qhimark = 10000;
|
| 236 | static long qlowmark = 100;
|
| 237 |
|
| 238 | module_param(blimit, long, 0444);
|
| 239 | module_param(qhimark, long, 0444);
|
| 240 | module_param(qlowmark, long, 0444);
|
| 241 |
|
| 242 | static ulong jiffies_till_first_fqs = ULONG_MAX;
|
| 243 | static ulong jiffies_till_next_fqs = ULONG_MAX;
|
| 244 |
|
| 245 | module_param(jiffies_till_first_fqs, ulong, 0644);
|
| 246 | module_param(jiffies_till_next_fqs, ulong, 0644);
|
| 247 |
|
| 248 | static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 249 | struct rcu_data *rdp);
|
| 250 | static void force_qs_rnp(struct rcu_state *rsp,
|
| 251 | int (*f)(struct rcu_data *rsp, bool *isidle,
|
| 252 | unsigned long *maxj),
|
| 253 | bool *isidle, unsigned long *maxj);
|
| 254 | static void force_quiescent_state(struct rcu_state *rsp);
|
| 255 | static int rcu_pending(int cpu);
|
| 256 |
|
| 257 | |
| 258 | * Return the number of RCU-sched batches processed thus far for debug & stats.
|
| 259 |
|
| 260 | long rcu_batches_completed_sched(void)
|
| 261 | {
|
| 262 | return rcu_sched_state.completed;
|
| 263 | }
|
| 264 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
|
| 265 |
|
| 266 | |
| 267 | * Return the number of RCU BH batches processed thus far for debug & stats.
|
| 268 |
|
| 269 | long rcu_batches_completed_bh(void)
|
| 270 | {
|
| 271 | return rcu_bh_state.completed;
|
| 272 | }
|
| 273 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
|
| 274 |
|
| 275 | |
| 276 | * Force a quiescent state for RCU BH.
|
| 277 |
|
| 278 | void rcu_bh_force_quiescent_state(void)
|
| 279 | {
|
| 280 | force_quiescent_state(&rcu_bh_state);
|
| 281 | }
|
| 282 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
|
| 283 |
|
| 284 | |
| 285 | * Record the number of times rcutorture tests have been initiated and
|
| 286 | * terminated. This information allows the debugfs tracing stats to be
|
| 287 | * correlated to the rcutorture messages, even when the rcutorture module
|
| 288 | * is being repeatedly loaded and unloaded. In other words, we cannot
|
| 289 | * store this state in rcutorture itself.
|
| 290 |
|
| 291 | void rcutorture_record_test_transition(void)
|
| 292 | {
|
| 293 | rcutorture_testseq++;
|
| 294 | rcutorture_vernum = 0;
|
| 295 | }
|
| 296 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
|
| 297 |
|
| 298 | |
| 299 | * Record the number of writer passes through the current rcutorture test.
|
| 300 | * This is also used to correlate debugfs tracing stats with the rcutorture
|
| 301 | * messages.
|
| 302 |
|
| 303 | void rcutorture_record_progress(unsigned long vernum)
|
| 304 | {
|
| 305 | rcutorture_vernum++;
|
| 306 | }
|
| 307 | EXPORT_SYMBOL_GPL(rcutorture_record_progress);
|
| 308 |
|
| 309 | |
| 310 | * Force a quiescent state for RCU-sched.
|
| 311 |
|
| 312 | void rcu_sched_force_quiescent_state(void)
|
| 313 | {
|
| 314 | force_quiescent_state(&rcu_sched_state);
|
| 315 | }
|
| 316 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
|
| 317 |
|
| 318 | |
| 319 | * Does the CPU have callbacks ready to be invoked?
|
| 320 |
|
| 321 | static int
|
| 322 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
|
| 323 | {
|
| 324 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
|
| 325 | rdp->nxttail[RCU_DONE_TAIL] != NULL;
|
| 326 | }
|
| 327 |
|
| 328 | |
| 329 | * Does the current CPU require a not-yet-started grace period?
|
| 330 | * The caller must have disabled interrupts to prevent races with
|
| 331 | * normal callback registry.
|
| 332 |
|
| 333 | static int
|
| 334 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 335 | {
|
| 336 | int i;
|
| 337 |
|
| 338 | if (rcu_gp_in_progress(rsp))
|
| 339 | return 0;
|
| 340 | if (rcu_nocb_needs_gp(rsp))
|
| 341 | return 1;
|
| 342 | if (!rdp->nxttail[RCU_NEXT_TAIL])
|
| 343 | return 0;
|
| 344 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
|
| 345 | return 1;
|
| 346 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
|
| 347 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
|
| 348 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
|
| 349 | rdp->nxtcompleted[i]))
|
| 350 | return 1;
|
| 351 | return 0;
|
| 352 | }
|
| 353 |
|
| 354 | |
| 355 | * Return the root node of the specified rcu_state structure.
|
| 356 |
|
| 357 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
|
| 358 | {
|
| 359 | return &rsp->node[0];
|
| 360 | }
|
| 361 |
|
| 362 | |
| 363 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
|
| 364 | *
|
| 365 | * If the new value of the ->dynticks_nesting counter now is zero,
|
| 366 | * we really have entered idle, and must do the appropriate accounting.
|
| 367 | * The caller must have disabled interrupts.
|
| 368 |
|
| 369 | static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
|
| 370 | bool user)
|
| 371 | {
|
| 372 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
|
| 373 | if (!user && !is_idle_task(current)) {
|
| 374 | struct task_struct *idle __maybe_unused =
|
| 375 | idle_task(smp_processor_id());
|
| 376 |
|
| 377 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
|
| 378 | ftrace_dump(DUMP_ORIG);
|
| 379 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
|
| 380 | current->pid, current->comm,
|
| 381 | idle->pid, idle->comm);
|
| 382 | }
|
| 383 | rcu_prepare_for_idle(smp_processor_id());
|
| 384 |
|
| 385 | smp_mb__before_atomic_inc();
|
| 386 | atomic_inc(&rdtp->dynticks);
|
| 387 | smp_mb__after_atomic_inc();
|
| 388 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
|
| 389 |
|
| 390 | |
| 391 | * It is illegal to enter an extended quiescent state while
|
| 392 | * in an RCU read-side critical section.
|
| 393 |
|
| 394 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
|
| 395 | "Illegal idle entry in RCU read-side critical section.");
|
| 396 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
|
| 397 | "Illegal idle entry in RCU-bh read-side critical section.");
|
| 398 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
|
| 399 | "Illegal idle entry in RCU-sched read-side critical section.");
|
| 400 | }
|
| 401 |
|
| 402 | |
| 403 | * Enter an RCU extended quiescent state, which can be either the
|
| 404 | * idle loop or adaptive-tickless usermode execution.
|
| 405 |
|
| 406 | static void rcu_eqs_enter(bool user)
|
| 407 | {
|
| 408 | long long oldval;
|
| 409 | struct rcu_dynticks *rdtp;
|
| 410 |
|
| 411 | rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 412 | oldval = rdtp->dynticks_nesting;
|
| 413 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
|
| 414 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
|
| 415 | rdtp->dynticks_nesting = 0;
|
| 416 | else
|
| 417 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
|
| 418 | rcu_eqs_enter_common(rdtp, oldval, user);
|
| 419 | }
|
| 420 |
|
| 421 | |
| 422 | * rcu_idle_enter - inform RCU that current CPU is entering idle
|
| 423 | *
|
| 424 | * Enter idle mode, in other words, -leave- the mode in which RCU
|
| 425 | * read-side critical sections can occur. (Though RCU read-side
|
| 426 | * critical sections can occur in irq handlers in idle, a possibility
|
| 427 | * handled by irq_enter() and irq_exit().)
|
| 428 | *
|
| 429 | * We crowbar the ->dynticks_nesting field to zero to allow for
|
| 430 | * the possibility of usermode upcalls having messed up our count
|
| 431 | * of interrupt nesting level during the prior busy period.
|
| 432 |
|
| 433 | void rcu_idle_enter(void)
|
| 434 | {
|
| 435 | unsigned long flags;
|
| 436 |
|
| 437 | local_irq_save(flags);
|
| 438 | rcu_eqs_enter(false);
|
| 439 | rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
|
| 440 | local_irq_restore(flags);
|
| 441 | }
|
| 442 | EXPORT_SYMBOL_GPL(rcu_idle_enter);
|
| 443 |
|
| 444 | #ifdef CONFIG_RCU_USER_QS
|
| 445 | |
| 446 | * rcu_user_enter - inform RCU that we are resuming userspace.
|
| 447 | *
|
| 448 | * Enter RCU idle mode right before resuming userspace. No use of RCU
|
| 449 | * is permitted between this call and rcu_user_exit(). This way the
|
| 450 | * CPU doesn't need to maintain the tick for RCU maintenance purposes
|
| 451 | * when the CPU runs in userspace.
|
| 452 |
|
| 453 | void rcu_user_enter(void)
|
| 454 | {
|
| 455 | rcu_eqs_enter(1);
|
| 456 | }
|
| 457 | #endif
|
| 458 |
|
| 459 | |
| 460 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
|
| 461 | *
|
| 462 | * Exit from an interrupt handler, which might possibly result in entering
|
| 463 | * idle mode, in other words, leaving the mode in which read-side critical
|
| 464 | * sections can occur.
|
| 465 | *
|
| 466 | * This code assumes that the idle loop never does anything that might
|
| 467 | * result in unbalanced calls to irq_enter() and irq_exit(). If your
|
| 468 | * architecture violates this assumption, RCU will give you what you
|
| 469 | * deserve, good and hard. But very infrequently and irreproducibly.
|
| 470 | *
|
| 471 | * Use things like work queues to work around this limitation.
|
| 472 | *
|
| 473 | * You have been warned.
|
| 474 |
|
| 475 | void rcu_irq_exit(void)
|
| 476 | {
|
| 477 | unsigned long flags;
|
| 478 | long long oldval;
|
| 479 | struct rcu_dynticks *rdtp;
|
| 480 |
|
| 481 | local_irq_save(flags);
|
| 482 | rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 483 | oldval = rdtp->dynticks_nesting;
|
| 484 | rdtp->dynticks_nesting--;
|
| 485 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
|
| 486 | if (rdtp->dynticks_nesting)
|
| 487 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
|
| 488 | else
|
| 489 | rcu_eqs_enter_common(rdtp, oldval, true);
|
| 490 | rcu_sysidle_enter(rdtp, 1);
|
| 491 | local_irq_restore(flags);
|
| 492 | }
|
| 493 |
|
| 494 | |
| 495 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
|
| 496 | *
|
| 497 | * If the new value of the ->dynticks_nesting counter was previously zero,
|
| 498 | * we really have exited idle, and must do the appropriate accounting.
|
| 499 | * The caller must have disabled interrupts.
|
| 500 |
|
| 501 | static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
|
| 502 | int user)
|
| 503 | {
|
| 504 | smp_mb__before_atomic_inc();
|
| 505 | atomic_inc(&rdtp->dynticks);
|
| 506 |
|
| 507 | smp_mb__after_atomic_inc();
|
| 508 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
|
| 509 | rcu_cleanup_after_idle(smp_processor_id());
|
| 510 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
|
| 511 | if (!user && !is_idle_task(current)) {
|
| 512 | struct task_struct *idle __maybe_unused =
|
| 513 | idle_task(smp_processor_id());
|
| 514 |
|
| 515 | trace_rcu_dyntick(TPS("Error on exit: not idle task"),
|
| 516 | oldval, rdtp->dynticks_nesting);
|
| 517 | ftrace_dump(DUMP_ORIG);
|
| 518 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
|
| 519 | current->pid, current->comm,
|
| 520 | idle->pid, idle->comm);
|
| 521 | }
|
| 522 | }
|
| 523 |
|
| 524 | |
| 525 | * Exit an RCU extended quiescent state, which can be either the
|
| 526 | * idle loop or adaptive-tickless usermode execution.
|
| 527 |
|
| 528 | static void rcu_eqs_exit(bool user)
|
| 529 | {
|
| 530 | struct rcu_dynticks *rdtp;
|
| 531 | long long oldval;
|
| 532 |
|
| 533 | rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 534 | oldval = rdtp->dynticks_nesting;
|
| 535 | WARN_ON_ONCE(oldval < 0);
|
| 536 | if (oldval & DYNTICK_TASK_NEST_MASK)
|
| 537 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
|
| 538 | else
|
| 539 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
|
| 540 | rcu_eqs_exit_common(rdtp, oldval, user);
|
| 541 | }
|
| 542 |
|
| 543 | |
| 544 | * rcu_idle_exit - inform RCU that current CPU is leaving idle
|
| 545 | *
|
| 546 | * Exit idle mode, in other words, -enter- the mode in which RCU
|
| 547 | * read-side critical sections can occur.
|
| 548 | *
|
| 549 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
|
| 550 | * allow for the possibility of usermode upcalls messing up our count
|
| 551 | * of interrupt nesting level during the busy period that is just
|
| 552 | * now starting.
|
| 553 |
|
| 554 | void rcu_idle_exit(void)
|
| 555 | {
|
| 556 | unsigned long flags;
|
| 557 |
|
| 558 | local_irq_save(flags);
|
| 559 | rcu_eqs_exit(false);
|
| 560 | rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
|
| 561 | local_irq_restore(flags);
|
| 562 | }
|
| 563 | EXPORT_SYMBOL_GPL(rcu_idle_exit);
|
| 564 |
|
| 565 | #ifdef CONFIG_RCU_USER_QS
|
| 566 | |
| 567 | * rcu_user_exit - inform RCU that we are exiting userspace.
|
| 568 | *
|
| 569 | * Exit RCU idle mode while entering the kernel because it can
|
| 570 | * run a RCU read side critical section anytime.
|
| 571 |
|
| 572 | void rcu_user_exit(void)
|
| 573 | {
|
| 574 | rcu_eqs_exit(1);
|
| 575 | }
|
| 576 | #endif
|
| 577 |
|
| 578 | |
| 579 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
|
| 580 | *
|
| 581 | * Enter an interrupt handler, which might possibly result in exiting
|
| 582 | * idle mode, in other words, entering the mode in which read-side critical
|
| 583 | * sections can occur.
|
| 584 | *
|
| 585 | * Note that the Linux kernel is fully capable of entering an interrupt
|
| 586 | * handler that it never exits, for example when doing upcalls to
|
| 587 | * user mode! This code assumes that the idle loop never does upcalls to
|
| 588 | * user mode. If your architecture does do upcalls from the idle loop (or
|
| 589 | * does anything else that results in unbalanced calls to the irq_enter()
|
| 590 | * and irq_exit() functions), RCU will give you what you deserve, good
|
| 591 | * and hard. But very infrequently and irreproducibly.
|
| 592 | *
|
| 593 | * Use things like work queues to work around this limitation.
|
| 594 | *
|
| 595 | * You have been warned.
|
| 596 |
|
| 597 | void rcu_irq_enter(void)
|
| 598 | {
|
| 599 | unsigned long flags;
|
| 600 | struct rcu_dynticks *rdtp;
|
| 601 | long long oldval;
|
| 602 |
|
| 603 | local_irq_save(flags);
|
| 604 | rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 605 | oldval = rdtp->dynticks_nesting;
|
| 606 | rdtp->dynticks_nesting++;
|
| 607 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
|
| 608 | if (oldval)
|
| 609 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
|
| 610 | else
|
| 611 | rcu_eqs_exit_common(rdtp, oldval, true);
|
| 612 | rcu_sysidle_exit(rdtp, 1);
|
| 613 | local_irq_restore(flags);
|
| 614 | }
|
| 615 |
|
| 616 | |
| 617 | * rcu_nmi_enter - inform RCU of entry to NMI context
|
| 618 | *
|
| 619 | * If the CPU was idle with dynamic ticks active, and there is no
|
| 620 | * irq handler running, this updates rdtp->dynticks_nmi to let the
|
| 621 | * RCU grace-period handling know that the CPU is active.
|
| 622 |
|
| 623 | void rcu_nmi_enter(void)
|
| 624 | {
|
| 625 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 626 |
|
| 627 | if (rdtp->dynticks_nmi_nesting == 0 &&
|
| 628 | (atomic_read(&rdtp->dynticks) & 0x1))
|
| 629 | return;
|
| 630 | rdtp->dynticks_nmi_nesting++;
|
| 631 | smp_mb__before_atomic_inc();
|
| 632 | atomic_inc(&rdtp->dynticks);
|
| 633 |
|
| 634 | smp_mb__after_atomic_inc();
|
| 635 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
|
| 636 | }
|
| 637 |
|
| 638 | |
| 639 | * rcu_nmi_exit - inform RCU of exit from NMI context
|
| 640 | *
|
| 641 | * If the CPU was idle with dynamic ticks active, and there is no
|
| 642 | * irq handler running, this updates rdtp->dynticks_nmi to let the
|
| 643 | * RCU grace-period handling know that the CPU is no longer active.
|
| 644 |
|
| 645 | void rcu_nmi_exit(void)
|
| 646 | {
|
| 647 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
|
| 648 |
|
| 649 | if (rdtp->dynticks_nmi_nesting == 0 ||
|
| 650 | --rdtp->dynticks_nmi_nesting != 0)
|
| 651 | return;
|
| 652 |
|
| 653 | smp_mb__before_atomic_inc();
|
| 654 | atomic_inc(&rdtp->dynticks);
|
| 655 | smp_mb__after_atomic_inc();
|
| 656 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
|
| 657 | }
|
| 658 |
|
| 659 | |
| 660 | * __rcu_is_watching - are RCU read-side critical sections safe?
|
| 661 | *
|
| 662 | * Return true if RCU is watching the running CPU, which means that
|
| 663 | * this CPU can safely enter RCU read-side critical sections. Unlike
|
| 664 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at
|
| 665 | * least disabled preemption.
|
| 666 |
|
| 667 | bool notrace __rcu_is_watching(void)
|
| 668 | {
|
| 669 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
|
| 670 | }
|
| 671 |
|
| 672 | |
| 673 | * rcu_is_watching - see if RCU thinks that the current CPU is idle
|
| 674 | *
|
| 675 | * If the current CPU is in its idle loop and is neither in an interrupt
|
| 676 | * or NMI handler, return true.
|
| 677 |
|
| 678 | bool notrace rcu_is_watching(void)
|
| 679 | {
|
| 680 | int ret;
|
| 681 |
|
| 682 | preempt_disable();
|
| 683 | ret = __rcu_is_watching();
|
| 684 | preempt_enable();
|
| 685 | return ret;
|
| 686 | }
|
| 687 | EXPORT_SYMBOL_GPL(rcu_is_watching);
|
| 688 |
|
| 689 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
|
| 690 |
|
| 691 | |
| 692 | * Is the current CPU online? Disable preemption to avoid false positives
|
| 693 | * that could otherwise happen due to the current CPU number being sampled,
|
| 694 | * this task being preempted, its old CPU being taken offline, resuming
|
| 695 | * on some other CPU, then determining that its old CPU is now offline.
|
| 696 | * It is OK to use RCU on an offline processor during initial boot, hence
|
| 697 | * the check for rcu_scheduler_fully_active. Note also that it is OK
|
| 698 | * for a CPU coming online to use RCU for one jiffy prior to marking itself
|
| 699 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going
|
| 700 | * offline to continue to use RCU for one jiffy after marking itself
|
| 701 | * offline in the cpu_online_mask. This leniency is necessary given the
|
| 702 | * non-atomic nature of the online and offline processing, for example,
|
| 703 | * the fact that a CPU enters the scheduler after completing the CPU_DYING
|
| 704 | * notifiers.
|
| 705 | *
|
| 706 | * This is also why RCU internally marks CPUs online during the
|
| 707 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
|
| 708 | *
|
| 709 | * Disable checking if in an NMI handler because we cannot safely report
|
| 710 | * errors from NMI handlers anyway.
|
| 711 |
|
| 712 | bool rcu_lockdep_current_cpu_online(void)
|
| 713 | {
|
| 714 | struct rcu_data *rdp;
|
| 715 | struct rcu_node *rnp;
|
| 716 | bool ret;
|
| 717 |
|
| 718 | if (in_nmi())
|
| 719 | return 1;
|
| 720 | preempt_disable();
|
| 721 | rdp = this_cpu_ptr(&rcu_sched_data);
|
| 722 | rnp = rdp->mynode;
|
| 723 | ret = (rdp->grpmask & rnp->qsmaskinit) ||
|
| 724 | !rcu_scheduler_fully_active;
|
| 725 | preempt_enable();
|
| 726 | return ret;
|
| 727 | }
|
| 728 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
|
| 729 |
|
| 730 | #endif
|
| 731 |
|
| 732 | |
| 733 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
|
| 734 | *
|
| 735 | * If the current CPU is idle or running at a first-level (not nested)
|
| 736 | * interrupt from idle, return true. The caller must have at least
|
| 737 | * disabled preemption.
|
| 738 |
|
| 739 | static int rcu_is_cpu_rrupt_from_idle(void)
|
| 740 | {
|
| 741 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
|
| 742 | }
|
| 743 |
|
| 744 | |
| 745 | * Snapshot the specified CPU's dynticks counter so that we can later
|
| 746 | * credit them with an implicit quiescent state. Return 1 if this CPU
|
| 747 | * is in dynticks idle mode, which is an extended quiescent state.
|
| 748 |
|
| 749 | static int dyntick_save_progress_counter(struct rcu_data *rdp,
|
| 750 | bool *isidle, unsigned long *maxj)
|
| 751 | {
|
| 752 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
|
| 753 | rcu_sysidle_check_cpu(rdp, isidle, maxj);
|
| 754 | return (rdp->dynticks_snap & 0x1) == 0;
|
| 755 | }
|
| 756 |
|
| 757 | |
| 758 | * Return true if the specified CPU has passed through a quiescent
|
| 759 | * state by virtue of being in or having passed through an dynticks
|
| 760 | * idle state since the last call to dyntick_save_progress_counter()
|
| 761 | * for this same CPU, or by virtue of having been offline.
|
| 762 |
|
| 763 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
|
| 764 | bool *isidle, unsigned long *maxj)
|
| 765 | {
|
| 766 | unsigned int curr;
|
| 767 | unsigned int snap;
|
| 768 |
|
| 769 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
|
| 770 | snap = (unsigned int)rdp->dynticks_snap;
|
| 771 |
|
| 772 | |
| 773 | * If the CPU passed through or entered a dynticks idle phase with
|
| 774 | * no active irq/NMI handlers, then we can safely pretend that the CPU
|
| 775 | * already acknowledged the request to pass through a quiescent
|
| 776 | * state. Either way, that CPU cannot possibly be in an RCU
|
| 777 | * read-side critical section that started before the beginning
|
| 778 | * of the current RCU grace period.
|
| 779 |
|
| 780 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
|
| 781 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
|
| 782 | rdp->dynticks_fqs++;
|
| 783 | return 1;
|
| 784 | }
|
| 785 |
|
| 786 | |
| 787 | * Check for the CPU being offline, but only if the grace period
|
| 788 | * is old enough. We don't need to worry about the CPU changing
|
| 789 | * state: If we see it offline even once, it has been through a
|
| 790 | * quiescent state.
|
| 791 | *
|
| 792 | * The reason for insisting that the grace period be at least
|
| 793 | * one jiffy old is that CPUs that are not quite online and that
|
| 794 | * have just gone offline can still execute RCU read-side critical
|
| 795 | * sections.
|
| 796 |
|
| 797 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
|
| 798 | return 0;
|
| 799 | barrier();
|
| 800 | if (cpu_is_offline(rdp->cpu)) {
|
| 801 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
|
| 802 | rdp->offline_fqs++;
|
| 803 | return 1;
|
| 804 | }
|
| 805 |
|
| 806 | |
| 807 | * There is a possibility that a CPU in adaptive-ticks state
|
| 808 | * might run in the kernel with the scheduling-clock tick disabled
|
| 809 | * for an extended time period. Invoke rcu_kick_nohz_cpu() to
|
| 810 | * force the CPU to restart the scheduling-clock tick in this
|
| 811 | * CPU is in this state.
|
| 812 |
|
| 813 | rcu_kick_nohz_cpu(rdp->cpu);
|
| 814 |
|
| 815 | return 0;
|
| 816 | }
|
| 817 |
|
| 818 | static void record_gp_stall_check_time(struct rcu_state *rsp)
|
| 819 | {
|
| 820 | unsigned long j = ACCESS_ONCE(jiffies);
|
| 821 |
|
| 822 | rsp->gp_start = j;
|
| 823 | smp_wmb();
|
| 824 | rsp->jiffies_stall = j + rcu_jiffies_till_stall_check();
|
| 825 | }
|
| 826 |
|
| 827 | |
| 828 | * Dump stacks of all tasks running on stalled CPUs. This is a fallback
|
| 829 | * for architectures that do not implement trigger_all_cpu_backtrace().
|
| 830 | * The NMI-triggered stack traces are more accurate because they are
|
| 831 | * printed by the target CPU.
|
| 832 |
|
| 833 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
|
| 834 | {
|
| 835 | int cpu;
|
| 836 | unsigned long flags;
|
| 837 | struct rcu_node *rnp;
|
| 838 |
|
| 839 | rcu_for_each_leaf_node(rsp, rnp) {
|
| 840 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 841 | if (rnp->qsmask != 0) {
|
| 842 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
|
| 843 | if (rnp->qsmask & (1UL << cpu))
|
| 844 | dump_cpu_task(rnp->grplo + cpu);
|
| 845 | }
|
| 846 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 847 | }
|
| 848 | }
|
| 849 |
|
| 850 | static void print_other_cpu_stall(struct rcu_state *rsp)
|
| 851 | {
|
| 852 | int cpu;
|
| 853 | long delta;
|
| 854 | unsigned long flags;
|
| 855 | int ndetected = 0;
|
| 856 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 857 | long totqlen = 0;
|
| 858 |
|
| 859 |
|
| 860 |
|
| 861 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 862 | delta = jiffies - rsp->jiffies_stall;
|
| 863 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
|
| 864 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 865 | return;
|
| 866 | }
|
| 867 | rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
|
| 868 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 869 |
|
| 870 | |
| 871 | * OK, time to rat on our buddy...
|
| 872 | * See Documentation/RCU/stallwarn.txt for info on how to debug
|
| 873 | * RCU CPU stall warnings.
|
| 874 |
|
| 875 | pr_err("INFO: %s detected stalls on CPUs/tasks:",
|
| 876 | rsp->name);
|
| 877 | print_cpu_stall_info_begin();
|
| 878 | rcu_for_each_leaf_node(rsp, rnp) {
|
| 879 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 880 | ndetected += rcu_print_task_stall(rnp);
|
| 881 | if (rnp->qsmask != 0) {
|
| 882 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
|
| 883 | if (rnp->qsmask & (1UL << cpu)) {
|
| 884 | print_cpu_stall_info(rsp,
|
| 885 | rnp->grplo + cpu);
|
| 886 | ndetected++;
|
| 887 | }
|
| 888 | }
|
| 889 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 890 | }
|
| 891 |
|
| 892 | |
| 893 | * Now rat on any tasks that got kicked up to the root rcu_node
|
| 894 | * due to CPU offlining.
|
| 895 |
|
| 896 | rnp = rcu_get_root(rsp);
|
| 897 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 898 | ndetected += rcu_print_task_stall(rnp);
|
| 899 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 900 |
|
| 901 | print_cpu_stall_info_end();
|
| 902 | for_each_possible_cpu(cpu)
|
| 903 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
|
| 904 | pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
|
| 905 | smp_processor_id(), (long)(jiffies - rsp->gp_start),
|
| 906 | rsp->gpnum, rsp->completed, totqlen);
|
| 907 | if (ndetected == 0)
|
| 908 | pr_err("INFO: Stall ended before state dump start\n");
|
| 909 | else if (!trigger_all_cpu_backtrace())
|
| 910 | rcu_dump_cpu_stacks(rsp);
|
| 911 |
|
| 912 |
|
| 913 |
|
| 914 | rcu_print_detail_task_stall(rsp);
|
| 915 |
|
| 916 | force_quiescent_state(rsp);
|
| 917 | }
|
| 918 |
|
| 919 | |
| 920 | * This function really isn't for public consumption, but RCU is special in
|
| 921 | * that context switches can allow the state machine to make progress.
|
| 922 |
|
| 923 | extern void resched_cpu(int cpu);
|
| 924 |
|
| 925 | static void print_cpu_stall(struct rcu_state *rsp)
|
| 926 | {
|
| 927 | int cpu;
|
| 928 | unsigned long flags;
|
| 929 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 930 | long totqlen = 0;
|
| 931 |
|
| 932 | |
| 933 | * OK, time to rat on ourselves...
|
| 934 | * See Documentation/RCU/stallwarn.txt for info on how to debug
|
| 935 | * RCU CPU stall warnings.
|
| 936 |
|
| 937 | pr_err("INFO: %s self-detected stall on CPU", rsp->name);
|
| 938 | print_cpu_stall_info_begin();
|
| 939 | print_cpu_stall_info(rsp, smp_processor_id());
|
| 940 | print_cpu_stall_info_end();
|
| 941 | for_each_possible_cpu(cpu)
|
| 942 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
|
| 943 | pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
|
| 944 | jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
|
| 945 | if (!trigger_all_cpu_backtrace())
|
| 946 | dump_stack();
|
| 947 |
|
| 948 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 949 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
|
| 950 | rsp->jiffies_stall = jiffies +
|
| 951 | 3 * rcu_jiffies_till_stall_check() + 3;
|
| 952 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 953 |
|
| 954 | |
| 955 | * Attempt to revive the RCU machinery by forcing a context switch.
|
| 956 | *
|
| 957 | * A context switch would normally allow the RCU state machine to make
|
| 958 | * progress and it could be we're stuck in kernel space without context
|
| 959 | * switches for an entirely unreasonable amount of time.
|
| 960 |
|
| 961 | resched_cpu(smp_processor_id());
|
| 962 | }
|
| 963 |
|
| 964 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 965 | {
|
| 966 | unsigned long completed;
|
| 967 | unsigned long gpnum;
|
| 968 | unsigned long gps;
|
| 969 | unsigned long j;
|
| 970 | unsigned long js;
|
| 971 | struct rcu_node *rnp;
|
| 972 |
|
| 973 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
|
| 974 | return;
|
| 975 | j = ACCESS_ONCE(jiffies);
|
| 976 |
|
| 977 | |
| 978 | * Lots of memory barriers to reject false positives.
|
| 979 | *
|
| 980 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
|
| 981 | * then rsp->gp_start, and finally rsp->completed. These values
|
| 982 | * are updated in the opposite order with memory barriers (or
|
| 983 | * equivalent) during grace-period initialization and cleanup.
|
| 984 | * Now, a false positive can occur if we get an new value of
|
| 985 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given
|
| 986 | * the memory barriers, the only way that this can happen is if one
|
| 987 | * grace period ends and another starts between these two fetches.
|
| 988 | * Detect this by comparing rsp->completed with the previous fetch
|
| 989 | * from rsp->gpnum.
|
| 990 | *
|
| 991 | * Given this check, comparisons of jiffies, rsp->jiffies_stall,
|
| 992 | * and rsp->gp_start suffice to forestall false positives.
|
| 993 |
|
| 994 | gpnum = ACCESS_ONCE(rsp->gpnum);
|
| 995 | smp_rmb();
|
| 996 | js = ACCESS_ONCE(rsp->jiffies_stall);
|
| 997 | smp_rmb();
|
| 998 | gps = ACCESS_ONCE(rsp->gp_start);
|
| 999 | smp_rmb();
|
| 1000 | completed = ACCESS_ONCE(rsp->completed);
|
| 1001 | if (ULONG_CMP_GE(completed, gpnum) ||
|
| 1002 | ULONG_CMP_LT(j, js) ||
|
| 1003 | ULONG_CMP_GE(gps, js))
|
| 1004 | return;
|
| 1005 | rnp = rdp->mynode;
|
| 1006 | if (rcu_gp_in_progress(rsp) &&
|
| 1007 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
|
| 1008 |
|
| 1009 |
|
| 1010 | print_cpu_stall(rsp);
|
| 1011 |
|
| 1012 | } else if (rcu_gp_in_progress(rsp) &&
|
| 1013 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
|
| 1014 |
|
| 1015 |
|
| 1016 | print_other_cpu_stall(rsp);
|
| 1017 | }
|
| 1018 | }
|
| 1019 |
|
| 1020 | |
| 1021 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
|
| 1022 | *
|
| 1023 | * Set the stall-warning timeout way off into the future, thus preventing
|
| 1024 | * any RCU CPU stall-warning messages from appearing in the current set of
|
| 1025 | * RCU grace periods.
|
| 1026 | *
|
| 1027 | * The caller must disable hard irqs.
|
| 1028 |
|
| 1029 | void rcu_cpu_stall_reset(void)
|
| 1030 | {
|
| 1031 | struct rcu_state *rsp;
|
| 1032 |
|
| 1033 | for_each_rcu_flavor(rsp)
|
| 1034 | rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
|
| 1035 | }
|
| 1036 |
|
| 1037 | |
| 1038 | * Initialize the specified rcu_data structure's callback list to empty.
|
| 1039 |
|
| 1040 | static void init_callback_list(struct rcu_data *rdp)
|
| 1041 | {
|
| 1042 | int i;
|
| 1043 |
|
| 1044 | if (init_nocb_callback_list(rdp))
|
| 1045 | return;
|
| 1046 | rdp->nxtlist = NULL;
|
| 1047 | for (i = 0; i < RCU_NEXT_SIZE; i++)
|
| 1048 | rdp->nxttail[i] = &rdp->nxtlist;
|
| 1049 | }
|
| 1050 |
|
| 1051 | |
| 1052 | * Determine the value that ->completed will have at the end of the
|
| 1053 | * next subsequent grace period. This is used to tag callbacks so that
|
| 1054 | * a CPU can invoke callbacks in a timely fashion even if that CPU has
|
| 1055 | * been dyntick-idle for an extended period with callbacks under the
|
| 1056 | * influence of RCU_FAST_NO_HZ.
|
| 1057 | *
|
| 1058 | * The caller must hold rnp->lock with interrupts disabled.
|
| 1059 |
|
| 1060 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
|
| 1061 | struct rcu_node *rnp)
|
| 1062 | {
|
| 1063 | |
| 1064 | * If RCU is idle, we just wait for the next grace period.
|
| 1065 | * But we can only be sure that RCU is idle if we are looking
|
| 1066 | * at the root rcu_node structure -- otherwise, a new grace
|
| 1067 | * period might have started, but just not yet gotten around
|
| 1068 | * to initializing the current non-root rcu_node structure.
|
| 1069 |
|
| 1070 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
|
| 1071 | return rnp->completed + 1;
|
| 1072 |
|
| 1073 | |
| 1074 | * Otherwise, wait for a possible partial grace period and
|
| 1075 | * then the subsequent full grace period.
|
| 1076 |
|
| 1077 | return rnp->completed + 2;
|
| 1078 | }
|
| 1079 |
|
| 1080 | |
| 1081 | * Trace-event helper function for rcu_start_future_gp() and
|
| 1082 | * rcu_nocb_wait_gp().
|
| 1083 |
|
| 1084 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
|
| 1085 | unsigned long c, const char *s)
|
| 1086 | {
|
| 1087 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
|
| 1088 | rnp->completed, c, rnp->level,
|
| 1089 | rnp->grplo, rnp->grphi, s);
|
| 1090 | }
|
| 1091 |
|
| 1092 | |
| 1093 | * Start some future grace period, as needed to handle newly arrived
|
| 1094 | * callbacks. The required future grace periods are recorded in each
|
| 1095 | * rcu_node structure's ->need_future_gp field.
|
| 1096 | *
|
| 1097 | * The caller must hold the specified rcu_node structure's ->lock.
|
| 1098 |
|
| 1099 | static unsigned long __maybe_unused
|
| 1100 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
|
| 1101 | {
|
| 1102 | unsigned long c;
|
| 1103 | int i;
|
| 1104 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
|
| 1105 |
|
| 1106 | |
| 1107 | * Pick up grace-period number for new callbacks. If this
|
| 1108 | * grace period is already marked as needed, return to the caller.
|
| 1109 |
|
| 1110 | c = rcu_cbs_completed(rdp->rsp, rnp);
|
| 1111 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
|
| 1112 | if (rnp->need_future_gp[c & 0x1]) {
|
| 1113 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
|
| 1114 | return c;
|
| 1115 | }
|
| 1116 |
|
| 1117 | |
| 1118 | * If either this rcu_node structure or the root rcu_node structure
|
| 1119 | * believe that a grace period is in progress, then we must wait
|
| 1120 | * for the one following, which is in "c". Because our request
|
| 1121 | * will be noticed at the end of the current grace period, we don't
|
| 1122 | * need to explicitly start one.
|
| 1123 |
|
| 1124 | if (rnp->gpnum != rnp->completed ||
|
| 1125 | ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
|
| 1126 | rnp->need_future_gp[c & 0x1]++;
|
| 1127 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
|
| 1128 | return c;
|
| 1129 | }
|
| 1130 |
|
| 1131 | |
| 1132 | * There might be no grace period in progress. If we don't already
|
| 1133 | * hold it, acquire the root rcu_node structure's lock in order to
|
| 1134 | * start one (if needed).
|
| 1135 |
|
| 1136 | if (rnp != rnp_root)
|
| 1137 | raw_spin_lock(&rnp_root->lock);
|
| 1138 |
|
| 1139 | |
| 1140 | * Get a new grace-period number. If there really is no grace
|
| 1141 | * period in progress, it will be smaller than the one we obtained
|
| 1142 | * earlier. Adjust callbacks as needed. Note that even no-CBs
|
| 1143 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
|
| 1144 |
|
| 1145 | c = rcu_cbs_completed(rdp->rsp, rnp_root);
|
| 1146 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
|
| 1147 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
|
| 1148 | rdp->nxtcompleted[i] = c;
|
| 1149 |
|
| 1150 | |
| 1151 | * If the needed for the required grace period is already
|
| 1152 | * recorded, trace and leave.
|
| 1153 |
|
| 1154 | if (rnp_root->need_future_gp[c & 0x1]) {
|
| 1155 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
|
| 1156 | goto unlock_out;
|
| 1157 | }
|
| 1158 |
|
| 1159 |
|
| 1160 | rnp_root->need_future_gp[c & 0x1]++;
|
| 1161 |
|
| 1162 |
|
| 1163 | if (rnp_root->gpnum != rnp_root->completed) {
|
| 1164 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
|
| 1165 | } else {
|
| 1166 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
|
| 1167 | rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
|
| 1168 | }
|
| 1169 | unlock_out:
|
| 1170 | if (rnp != rnp_root)
|
| 1171 | raw_spin_unlock(&rnp_root->lock);
|
| 1172 | return c;
|
| 1173 | }
|
| 1174 |
|
| 1175 | |
| 1176 | * Clean up any old requests for the just-ended grace period. Also return
|
| 1177 | * whether any additional grace periods have been requested. Also invoke
|
| 1178 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
|
| 1179 | * waiting for this grace period to complete.
|
| 1180 |
|
| 1181 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
|
| 1182 | {
|
| 1183 | int c = rnp->completed;
|
| 1184 | int needmore;
|
| 1185 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
|
| 1186 |
|
| 1187 | rcu_nocb_gp_cleanup(rsp, rnp);
|
| 1188 | rnp->need_future_gp[c & 0x1] = 0;
|
| 1189 | needmore = rnp->need_future_gp[(c + 1) & 0x1];
|
| 1190 | trace_rcu_future_gp(rnp, rdp, c,
|
| 1191 | needmore ? TPS("CleanupMore") : TPS("Cleanup"));
|
| 1192 | return needmore;
|
| 1193 | }
|
| 1194 |
|
| 1195 | |
| 1196 | * If there is room, assign a ->completed number to any callbacks on
|
| 1197 | * this CPU that have not already been assigned. Also accelerate any
|
| 1198 | * callbacks that were previously assigned a ->completed number that has
|
| 1199 | * since proven to be too conservative, which can happen if callbacks get
|
| 1200 | * assigned a ->completed number while RCU is idle, but with reference to
|
| 1201 | * a non-root rcu_node structure. This function is idempotent, so it does
|
| 1202 | * not hurt to call it repeatedly.
|
| 1203 | *
|
| 1204 | * The caller must hold rnp->lock with interrupts disabled.
|
| 1205 |
|
| 1206 | static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 1207 | struct rcu_data *rdp)
|
| 1208 | {
|
| 1209 | unsigned long c;
|
| 1210 | int i;
|
| 1211 |
|
| 1212 |
|
| 1213 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
|
| 1214 | return;
|
| 1215 |
|
| 1216 | |
| 1217 | * Starting from the sublist containing the callbacks most
|
| 1218 | * recently assigned a ->completed number and working down, find the
|
| 1219 | * first sublist that is not assignable to an upcoming grace period.
|
| 1220 | * Such a sublist has something in it (first two tests) and has
|
| 1221 | * a ->completed number assigned that will complete sooner than
|
| 1222 | * the ->completed number for newly arrived callbacks (last test).
|
| 1223 | *
|
| 1224 | * The key point is that any later sublist can be assigned the
|
| 1225 | * same ->completed number as the newly arrived callbacks, which
|
| 1226 | * means that the callbacks in any of these later sublist can be
|
| 1227 | * grouped into a single sublist, whether or not they have already
|
| 1228 | * been assigned a ->completed number.
|
| 1229 |
|
| 1230 | c = rcu_cbs_completed(rsp, rnp);
|
| 1231 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
|
| 1232 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
|
| 1233 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
|
| 1234 | break;
|
| 1235 |
|
| 1236 | |
| 1237 | * If there are no sublist for unassigned callbacks, leave.
|
| 1238 | * At the same time, advance "i" one sublist, so that "i" will
|
| 1239 | * index into the sublist where all the remaining callbacks should
|
| 1240 | * be grouped into.
|
| 1241 |
|
| 1242 | if (++i >= RCU_NEXT_TAIL)
|
| 1243 | return;
|
| 1244 |
|
| 1245 | |
| 1246 | * Assign all subsequent callbacks' ->completed number to the next
|
| 1247 | * full grace period and group them all in the sublist initially
|
| 1248 | * indexed by "i".
|
| 1249 |
|
| 1250 | for (; i <= RCU_NEXT_TAIL; i++) {
|
| 1251 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
|
| 1252 | rdp->nxtcompleted[i] = c;
|
| 1253 | }
|
| 1254 |
|
| 1255 | rcu_start_future_gp(rnp, rdp);
|
| 1256 |
|
| 1257 |
|
| 1258 | if (!*rdp->nxttail[RCU_WAIT_TAIL])
|
| 1259 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
|
| 1260 | else
|
| 1261 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
|
| 1262 | }
|
| 1263 |
|
| 1264 | |
| 1265 | * Move any callbacks whose grace period has completed to the
|
| 1266 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and
|
| 1267 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
|
| 1268 | * sublist. This function is idempotent, so it does not hurt to
|
| 1269 | * invoke it repeatedly. As long as it is not invoked -too- often...
|
| 1270 | *
|
| 1271 | * The caller must hold rnp->lock with interrupts disabled.
|
| 1272 |
|
| 1273 | static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 1274 | struct rcu_data *rdp)
|
| 1275 | {
|
| 1276 | int i, j;
|
| 1277 |
|
| 1278 |
|
| 1279 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
|
| 1280 | return;
|
| 1281 |
|
| 1282 | |
| 1283 | * Find all callbacks whose ->completed numbers indicate that they
|
| 1284 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
|
| 1285 |
|
| 1286 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
|
| 1287 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
|
| 1288 | break;
|
| 1289 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
|
| 1290 | }
|
| 1291 |
|
| 1292 | for (j = RCU_WAIT_TAIL; j < i; j++)
|
| 1293 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
|
| 1294 |
|
| 1295 |
|
| 1296 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
|
| 1297 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
|
| 1298 | break;
|
| 1299 | rdp->nxttail[j] = rdp->nxttail[i];
|
| 1300 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
|
| 1301 | }
|
| 1302 |
|
| 1303 |
|
| 1304 | rcu_accelerate_cbs(rsp, rnp, rdp);
|
| 1305 | }
|
| 1306 |
|
| 1307 | |
| 1308 | * Update CPU-local rcu_data state to record the beginnings and ends of
|
| 1309 | * grace periods. The caller must hold the ->lock of the leaf rcu_node
|
| 1310 | * structure corresponding to the current CPU, and must have irqs disabled.
|
| 1311 |
|
| 1312 | static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
|
| 1313 | {
|
| 1314 |
|
| 1315 | if (rdp->completed == rnp->completed) {
|
| 1316 |
|
| 1317 |
|
| 1318 | rcu_accelerate_cbs(rsp, rnp, rdp);
|
| 1319 |
|
| 1320 | } else {
|
| 1321 |
|
| 1322 |
|
| 1323 | rcu_advance_cbs(rsp, rnp, rdp);
|
| 1324 |
|
| 1325 |
|
| 1326 | rdp->completed = rnp->completed;
|
| 1327 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
|
| 1328 | }
|
| 1329 |
|
| 1330 | if (rdp->gpnum != rnp->gpnum) {
|
| 1331 | |
| 1332 | * If the current grace period is waiting for this CPU,
|
| 1333 | * set up to detect a quiescent state, otherwise don't
|
| 1334 | * go looking for one.
|
| 1335 |
|
| 1336 | rdp->gpnum = rnp->gpnum;
|
| 1337 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
|
| 1338 | rdp->passed_quiesce = 0;
|
| 1339 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
|
| 1340 | zero_cpu_stall_ticks(rdp);
|
| 1341 | }
|
| 1342 | }
|
| 1343 |
|
| 1344 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 1345 | {
|
| 1346 | unsigned long flags;
|
| 1347 | struct rcu_node *rnp;
|
| 1348 |
|
| 1349 | local_irq_save(flags);
|
| 1350 | rnp = rdp->mynode;
|
| 1351 | if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
|
| 1352 | rdp->completed == ACCESS_ONCE(rnp->completed)) ||
|
| 1353 | !raw_spin_trylock(&rnp->lock)) {
|
| 1354 | local_irq_restore(flags);
|
| 1355 | return;
|
| 1356 | }
|
| 1357 | __note_gp_changes(rsp, rnp, rdp);
|
| 1358 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 1359 | }
|
| 1360 |
|
| 1361 | |
| 1362 | * Initialize a new grace period. Return 0 if no grace period required.
|
| 1363 |
|
| 1364 | static int rcu_gp_init(struct rcu_state *rsp)
|
| 1365 | {
|
| 1366 | struct rcu_data *rdp;
|
| 1367 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 1368 |
|
| 1369 | rcu_bind_gp_kthread();
|
| 1370 | raw_spin_lock_irq(&rnp->lock);
|
| 1371 | if (rsp->gp_flags == 0) {
|
| 1372 |
|
| 1373 | raw_spin_unlock_irq(&rnp->lock);
|
| 1374 | return 0;
|
| 1375 | }
|
| 1376 | rsp->gp_flags = 0;
|
| 1377 |
|
| 1378 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
|
| 1379 | |
| 1380 | * Grace period already in progress, don't start another.
|
| 1381 | * Not supposed to be able to happen.
|
| 1382 |
|
| 1383 | raw_spin_unlock_irq(&rnp->lock);
|
| 1384 | return 0;
|
| 1385 | }
|
| 1386 |
|
| 1387 |
|
| 1388 | record_gp_stall_check_time(rsp);
|
| 1389 | smp_wmb();
|
| 1390 | rsp->gpnum++;
|
| 1391 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
|
| 1392 | raw_spin_unlock_irq(&rnp->lock);
|
| 1393 |
|
| 1394 |
|
| 1395 | mutex_lock(&rsp->onoff_mutex);
|
| 1396 |
|
| 1397 | |
| 1398 | * Set the quiescent-state-needed bits in all the rcu_node
|
| 1399 | * structures for all currently online CPUs in breadth-first order,
|
| 1400 | * starting from the root rcu_node structure, relying on the layout
|
| 1401 | * of the tree within the rsp->node[] array. Note that other CPUs
|
| 1402 | * will access only the leaves of the hierarchy, thus seeing that no
|
| 1403 | * grace period is in progress, at least until the corresponding
|
| 1404 | * leaf node has been initialized. In addition, we have excluded
|
| 1405 | * CPU-hotplug operations.
|
| 1406 | *
|
| 1407 | * The grace period cannot complete until the initialization
|
| 1408 | * process finishes, because this kthread handles both.
|
| 1409 |
|
| 1410 | rcu_for_each_node_breadth_first(rsp, rnp) {
|
| 1411 | raw_spin_lock_irq(&rnp->lock);
|
| 1412 | rdp = this_cpu_ptr(rsp->rda);
|
| 1413 | rcu_preempt_check_blocked_tasks(rnp);
|
| 1414 | rnp->qsmask = rnp->qsmaskinit;
|
| 1415 | ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
|
| 1416 | WARN_ON_ONCE(rnp->completed != rsp->completed);
|
| 1417 | ACCESS_ONCE(rnp->completed) = rsp->completed;
|
| 1418 | if (rnp == rdp->mynode)
|
| 1419 | __note_gp_changes(rsp, rnp, rdp);
|
| 1420 | rcu_preempt_boost_start_gp(rnp);
|
| 1421 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
|
| 1422 | rnp->level, rnp->grplo,
|
| 1423 | rnp->grphi, rnp->qsmask);
|
| 1424 | raw_spin_unlock_irq(&rnp->lock);
|
| 1425 | #ifdef CONFIG_PROVE_RCU_DELAY
|
| 1426 | if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
|
| 1427 | system_state == SYSTEM_RUNNING)
|
| 1428 | udelay(200);
|
| 1429 | #endif
|
| 1430 | cond_resched();
|
| 1431 | }
|
| 1432 |
|
| 1433 | mutex_unlock(&rsp->onoff_mutex);
|
| 1434 | return 1;
|
| 1435 | }
|
| 1436 |
|
| 1437 | |
| 1438 | * Do one round of quiescent-state forcing.
|
| 1439 |
|
| 1440 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
|
| 1441 | {
|
| 1442 | int fqs_state = fqs_state_in;
|
| 1443 | bool isidle = false;
|
| 1444 | unsigned long maxj;
|
| 1445 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 1446 |
|
| 1447 | rsp->n_force_qs++;
|
| 1448 | if (fqs_state == RCU_SAVE_DYNTICK) {
|
| 1449 |
|
| 1450 | if (is_sysidle_rcu_state(rsp)) {
|
| 1451 | isidle = 1;
|
| 1452 | maxj = jiffies - ULONG_MAX / 4;
|
| 1453 | }
|
| 1454 | force_qs_rnp(rsp, dyntick_save_progress_counter,
|
| 1455 | &isidle, &maxj);
|
| 1456 | rcu_sysidle_report_gp(rsp, isidle, maxj);
|
| 1457 | fqs_state = RCU_FORCE_QS;
|
| 1458 | } else {
|
| 1459 |
|
| 1460 | isidle = 0;
|
| 1461 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
|
| 1462 | }
|
| 1463 |
|
| 1464 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
|
| 1465 | raw_spin_lock_irq(&rnp->lock);
|
| 1466 | rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
|
| 1467 | raw_spin_unlock_irq(&rnp->lock);
|
| 1468 | }
|
| 1469 | return fqs_state;
|
| 1470 | }
|
| 1471 |
|
| 1472 | |
| 1473 | * Clean up after the old grace period.
|
| 1474 |
|
| 1475 | static void rcu_gp_cleanup(struct rcu_state *rsp)
|
| 1476 | {
|
| 1477 | unsigned long gp_duration;
|
| 1478 | int nocb = 0;
|
| 1479 | struct rcu_data *rdp;
|
| 1480 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 1481 |
|
| 1482 | raw_spin_lock_irq(&rnp->lock);
|
| 1483 | gp_duration = jiffies - rsp->gp_start;
|
| 1484 | if (gp_duration > rsp->gp_max)
|
| 1485 | rsp->gp_max = gp_duration;
|
| 1486 |
|
| 1487 | |
| 1488 | * We know the grace period is complete, but to everyone else
|
| 1489 | * it appears to still be ongoing. But it is also the case
|
| 1490 | * that to everyone else it looks like there is nothing that
|
| 1491 | * they can do to advance the grace period. It is therefore
|
| 1492 | * safe for us to drop the lock in order to mark the grace
|
| 1493 | * period as completed in all of the rcu_node structures.
|
| 1494 |
|
| 1495 | raw_spin_unlock_irq(&rnp->lock);
|
| 1496 |
|
| 1497 | |
| 1498 | * Propagate new ->completed value to rcu_node structures so
|
| 1499 | * that other CPUs don't have to wait until the start of the next
|
| 1500 | * grace period to process their callbacks. This also avoids
|
| 1501 | * some nasty RCU grace-period initialization races by forcing
|
| 1502 | * the end of the current grace period to be completely recorded in
|
| 1503 | * all of the rcu_node structures before the beginning of the next
|
| 1504 | * grace period is recorded in any of the rcu_node structures.
|
| 1505 |
|
| 1506 | rcu_for_each_node_breadth_first(rsp, rnp) {
|
| 1507 | raw_spin_lock_irq(&rnp->lock);
|
| 1508 | ACCESS_ONCE(rnp->completed) = rsp->gpnum;
|
| 1509 | rdp = this_cpu_ptr(rsp->rda);
|
| 1510 | if (rnp == rdp->mynode)
|
| 1511 | __note_gp_changes(rsp, rnp, rdp);
|
| 1512 | nocb += rcu_future_gp_cleanup(rsp, rnp);
|
| 1513 | raw_spin_unlock_irq(&rnp->lock);
|
| 1514 | cond_resched();
|
| 1515 | }
|
| 1516 | rnp = rcu_get_root(rsp);
|
| 1517 | raw_spin_lock_irq(&rnp->lock);
|
| 1518 | rcu_nocb_gp_set(rnp, nocb);
|
| 1519 |
|
| 1520 | rsp->completed = rsp->gpnum;
|
| 1521 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
|
| 1522 | rsp->fqs_state = RCU_GP_IDLE;
|
| 1523 | rdp = this_cpu_ptr(rsp->rda);
|
| 1524 | rcu_advance_cbs(rsp, rnp, rdp);
|
| 1525 | if (cpu_needs_another_gp(rsp, rdp)) {
|
| 1526 | rsp->gp_flags = RCU_GP_FLAG_INIT;
|
| 1527 | trace_rcu_grace_period(rsp->name,
|
| 1528 | ACCESS_ONCE(rsp->gpnum),
|
| 1529 | TPS("newreq"));
|
| 1530 | }
|
| 1531 | raw_spin_unlock_irq(&rnp->lock);
|
| 1532 | }
|
| 1533 |
|
| 1534 | |
| 1535 | * Body of kthread that handles grace periods.
|
| 1536 |
|
| 1537 | static int __noreturn rcu_gp_kthread(void *arg)
|
| 1538 | {
|
| 1539 | int fqs_state;
|
| 1540 | int gf;
|
| 1541 | unsigned long j;
|
| 1542 | int ret;
|
| 1543 | struct rcu_state *rsp = arg;
|
| 1544 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 1545 |
|
| 1546 | for (;;) {
|
| 1547 |
|
| 1548 |
|
| 1549 | for (;;) {
|
| 1550 | trace_rcu_grace_period(rsp->name,
|
| 1551 | ACCESS_ONCE(rsp->gpnum),
|
| 1552 | TPS("reqwait"));
|
| 1553 | wait_event_interruptible(rsp->gp_wq,
|
| 1554 | ACCESS_ONCE(rsp->gp_flags) &
|
| 1555 | RCU_GP_FLAG_INIT);
|
| 1556 | if (rcu_gp_init(rsp))
|
| 1557 | break;
|
| 1558 | cond_resched();
|
| 1559 | flush_signals(current);
|
| 1560 | trace_rcu_grace_period(rsp->name,
|
| 1561 | ACCESS_ONCE(rsp->gpnum),
|
| 1562 | TPS("reqwaitsig"));
|
| 1563 | }
|
| 1564 |
|
| 1565 |
|
| 1566 | fqs_state = RCU_SAVE_DYNTICK;
|
| 1567 | j = jiffies_till_first_fqs;
|
| 1568 | if (j > HZ) {
|
| 1569 | j = HZ;
|
| 1570 | jiffies_till_first_fqs = HZ;
|
| 1571 | }
|
| 1572 | ret = 0;
|
| 1573 | for (;;) {
|
| 1574 | if (!ret)
|
| 1575 | rsp->jiffies_force_qs = jiffies + j;
|
| 1576 | trace_rcu_grace_period(rsp->name,
|
| 1577 | ACCESS_ONCE(rsp->gpnum),
|
| 1578 | TPS("fqswait"));
|
| 1579 | ret = wait_event_interruptible_timeout(rsp->gp_wq,
|
| 1580 | ((gf = ACCESS_ONCE(rsp->gp_flags)) &
|
| 1581 | RCU_GP_FLAG_FQS) ||
|
| 1582 | (!ACCESS_ONCE(rnp->qsmask) &&
|
| 1583 | !rcu_preempt_blocked_readers_cgp(rnp)),
|
| 1584 | j);
|
| 1585 |
|
| 1586 | if (!ACCESS_ONCE(rnp->qsmask) &&
|
| 1587 | !rcu_preempt_blocked_readers_cgp(rnp))
|
| 1588 | break;
|
| 1589 |
|
| 1590 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
|
| 1591 | (gf & RCU_GP_FLAG_FQS)) {
|
| 1592 | trace_rcu_grace_period(rsp->name,
|
| 1593 | ACCESS_ONCE(rsp->gpnum),
|
| 1594 | TPS("fqsstart"));
|
| 1595 | fqs_state = rcu_gp_fqs(rsp, fqs_state);
|
| 1596 | trace_rcu_grace_period(rsp->name,
|
| 1597 | ACCESS_ONCE(rsp->gpnum),
|
| 1598 | TPS("fqsend"));
|
| 1599 | cond_resched();
|
| 1600 | } else {
|
| 1601 |
|
| 1602 | cond_resched();
|
| 1603 | flush_signals(current);
|
| 1604 | trace_rcu_grace_period(rsp->name,
|
| 1605 | ACCESS_ONCE(rsp->gpnum),
|
| 1606 | TPS("fqswaitsig"));
|
| 1607 | }
|
| 1608 | j = jiffies_till_next_fqs;
|
| 1609 | if (j > HZ) {
|
| 1610 | j = HZ;
|
| 1611 | jiffies_till_next_fqs = HZ;
|
| 1612 | } else if (j < 1) {
|
| 1613 | j = 1;
|
| 1614 | jiffies_till_next_fqs = 1;
|
| 1615 | }
|
| 1616 | }
|
| 1617 |
|
| 1618 |
|
| 1619 | rcu_gp_cleanup(rsp);
|
| 1620 | }
|
| 1621 | }
|
| 1622 |
|
| 1623 | static void rsp_wakeup(struct irq_work *work)
|
| 1624 | {
|
| 1625 | struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
|
| 1626 |
|
| 1627 |
|
| 1628 | wake_up(&rsp->gp_wq);
|
| 1629 | }
|
| 1630 |
|
| 1631 | |
| 1632 | * Start a new RCU grace period if warranted, re-initializing the hierarchy
|
| 1633 | * in preparation for detecting the next grace period. The caller must hold
|
| 1634 | * the root node's ->lock and hard irqs must be disabled.
|
| 1635 | *
|
| 1636 | * Note that it is legal for a dying CPU (which is marked as offline) to
|
| 1637 | * invoke this function. This can happen when the dying CPU reports its
|
| 1638 | * quiescent state.
|
| 1639 |
|
| 1640 | static void
|
| 1641 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
|
| 1642 | struct rcu_data *rdp)
|
| 1643 | {
|
| 1644 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
|
| 1645 | |
| 1646 | * Either we have not yet spawned the grace-period
|
| 1647 | * task, this CPU does not need another grace period,
|
| 1648 | * or a grace period is already in progress.
|
| 1649 | * Either way, don't start a new grace period.
|
| 1650 |
|
| 1651 | return;
|
| 1652 | }
|
| 1653 | rsp->gp_flags = RCU_GP_FLAG_INIT;
|
| 1654 | trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
|
| 1655 | TPS("newreq"));
|
| 1656 |
|
| 1657 | |
| 1658 | * We can't do wakeups while holding the rnp->lock, as that
|
| 1659 | * could cause possible deadlocks with the rq->lock. Defer
|
| 1660 | * the wakeup to interrupt context. And don't bother waking
|
| 1661 | * up the running kthread.
|
| 1662 |
|
| 1663 | if (current != rsp->gp_kthread)
|
| 1664 | irq_work_queue(&rsp->wakeup_work);
|
| 1665 | }
|
| 1666 |
|
| 1667 | |
| 1668 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
|
| 1669 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
|
| 1670 | * is invoked indirectly from rcu_advance_cbs(), which would result in
|
| 1671 | * endless recursion -- or would do so if it wasn't for the self-deadlock
|
| 1672 | * that is encountered beforehand.
|
| 1673 |
|
| 1674 | static void
|
| 1675 | rcu_start_gp(struct rcu_state *rsp)
|
| 1676 | {
|
| 1677 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
|
| 1678 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 1679 |
|
| 1680 | |
| 1681 | * If there is no grace period in progress right now, any
|
| 1682 | * callbacks we have up to this point will be satisfied by the
|
| 1683 | * next grace period. Also, advancing the callbacks reduces the
|
| 1684 | * probability of false positives from cpu_needs_another_gp()
|
| 1685 | * resulting in pointless grace periods. So, advance callbacks
|
| 1686 | * then start the grace period!
|
| 1687 |
|
| 1688 | rcu_advance_cbs(rsp, rnp, rdp);
|
| 1689 | rcu_start_gp_advanced(rsp, rnp, rdp);
|
| 1690 | }
|
| 1691 |
|
| 1692 | |
| 1693 | * Report a full set of quiescent states to the specified rcu_state
|
| 1694 | * data structure. This involves cleaning up after the prior grace
|
| 1695 | * period and letting rcu_start_gp() start up the next grace period
|
| 1696 | * if one is needed. Note that the caller must hold rnp->lock, which
|
| 1697 | * is released before return.
|
| 1698 |
|
| 1699 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
|
| 1700 | __releases(rcu_get_root(rsp)->lock)
|
| 1701 | {
|
| 1702 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
|
| 1703 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
|
| 1704 | wake_up(&rsp->gp_wq);
|
| 1705 | }
|
| 1706 |
|
| 1707 | |
| 1708 | * Similar to rcu_report_qs_rdp(), for which it is a helper function.
|
| 1709 | * Allows quiescent states for a group of CPUs to be reported at one go
|
| 1710 | * to the specified rcu_node structure, though all the CPUs in the group
|
| 1711 | * must be represented by the same rcu_node structure (which need not be
|
| 1712 | * a leaf rcu_node structure, though it often will be). That structure's
|
| 1713 | * lock must be held upon entry, and it is released before return.
|
| 1714 |
|
| 1715 | static void
|
| 1716 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
|
| 1717 | struct rcu_node *rnp, unsigned long flags)
|
| 1718 | __releases(rnp->lock)
|
| 1719 | {
|
| 1720 | struct rcu_node *rnp_c;
|
| 1721 |
|
| 1722 |
|
| 1723 | for (;;) {
|
| 1724 | if (!(rnp->qsmask & mask)) {
|
| 1725 |
|
| 1726 |
|
| 1727 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 1728 | return;
|
| 1729 | }
|
| 1730 | rnp->qsmask &= ~mask;
|
| 1731 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
|
| 1732 | mask, rnp->qsmask, rnp->level,
|
| 1733 | rnp->grplo, rnp->grphi,
|
| 1734 | !!rnp->gp_tasks);
|
| 1735 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
|
| 1736 |
|
| 1737 |
|
| 1738 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 1739 | return;
|
| 1740 | }
|
| 1741 | mask = rnp->grpmask;
|
| 1742 | if (rnp->parent == NULL) {
|
| 1743 |
|
| 1744 |
|
| 1745 |
|
| 1746 | break;
|
| 1747 | }
|
| 1748 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 1749 | rnp_c = rnp;
|
| 1750 | rnp = rnp->parent;
|
| 1751 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 1752 | WARN_ON_ONCE(rnp_c->qsmask);
|
| 1753 | }
|
| 1754 |
|
| 1755 | |
| 1756 | * Get here if we are the last CPU to pass through a quiescent
|
| 1757 | * state for this grace period. Invoke rcu_report_qs_rsp()
|
| 1758 | * to clean up and start the next grace period if one is needed.
|
| 1759 |
|
| 1760 | rcu_report_qs_rsp(rsp, flags);
|
| 1761 | }
|
| 1762 |
|
| 1763 | |
| 1764 | * Record a quiescent state for the specified CPU to that CPU's rcu_data
|
| 1765 | * structure. This must be either called from the specified CPU, or
|
| 1766 | * called when the specified CPU is known to be offline (and when it is
|
| 1767 | * also known that no other CPU is concurrently trying to help the offline
|
| 1768 | * CPU). The lastcomp argument is used to make sure we are still in the
|
| 1769 | * grace period of interest. We don't want to end the current grace period
|
| 1770 | * based on quiescent states detected in an earlier grace period!
|
| 1771 |
|
| 1772 | static void
|
| 1773 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
|
| 1774 | {
|
| 1775 | unsigned long flags;
|
| 1776 | unsigned long mask;
|
| 1777 | struct rcu_node *rnp;
|
| 1778 |
|
| 1779 | rnp = rdp->mynode;
|
| 1780 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 1781 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
|
| 1782 | rnp->completed == rnp->gpnum) {
|
| 1783 |
|
| 1784 | |
| 1785 | * The grace period in which this quiescent state was
|
| 1786 | * recorded has ended, so don't report it upwards.
|
| 1787 | * We will instead need a new quiescent state that lies
|
| 1788 | * within the current grace period.
|
| 1789 |
|
| 1790 | rdp->passed_quiesce = 0;
|
| 1791 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 1792 | return;
|
| 1793 | }
|
| 1794 | mask = rdp->grpmask;
|
| 1795 | if ((rnp->qsmask & mask) == 0) {
|
| 1796 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 1797 | } else {
|
| 1798 | rdp->qs_pending = 0;
|
| 1799 |
|
| 1800 | |
| 1801 | * This GP can't end until cpu checks in, so all of our
|
| 1802 | * callbacks can be processed during the next GP.
|
| 1803 |
|
| 1804 | rcu_accelerate_cbs(rsp, rnp, rdp);
|
| 1805 |
|
| 1806 | rcu_report_qs_rnp(mask, rsp, rnp, flags);
|
| 1807 | }
|
| 1808 | }
|
| 1809 |
|
| 1810 | |
| 1811 | * Check to see if there is a new grace period of which this CPU
|
| 1812 | * is not yet aware, and if so, set up local rcu_data state for it.
|
| 1813 | * Otherwise, see if this CPU has just passed through its first
|
| 1814 | * quiescent state for this grace period, and record that fact if so.
|
| 1815 |
|
| 1816 | static void
|
| 1817 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 1818 | {
|
| 1819 |
|
| 1820 | note_gp_changes(rsp, rdp);
|
| 1821 |
|
| 1822 | |
| 1823 | * Does this CPU still need to do its part for current grace period?
|
| 1824 | * If no, return and let the other CPUs do their part as well.
|
| 1825 |
|
| 1826 | if (!rdp->qs_pending)
|
| 1827 | return;
|
| 1828 |
|
| 1829 | |
| 1830 | * Was there a quiescent state since the beginning of the grace
|
| 1831 | * period? If no, then exit and wait for the next call.
|
| 1832 |
|
| 1833 | if (!rdp->passed_quiesce)
|
| 1834 | return;
|
| 1835 |
|
| 1836 | |
| 1837 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the
|
| 1838 | * judge of that).
|
| 1839 |
|
| 1840 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
|
| 1841 | }
|
| 1842 |
|
| 1843 | #ifdef CONFIG_HOTPLUG_CPU
|
| 1844 |
|
| 1845 | |
| 1846 | * Send the specified CPU's RCU callbacks to the orphanage. The
|
| 1847 | * specified CPU must be offline, and the caller must hold the
|
| 1848 | * ->orphan_lock.
|
| 1849 |
|
| 1850 | static void
|
| 1851 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
|
| 1852 | struct rcu_node *rnp, struct rcu_data *rdp)
|
| 1853 | {
|
| 1854 |
|
| 1855 | if (rcu_is_nocb_cpu(rdp->cpu))
|
| 1856 | return;
|
| 1857 |
|
| 1858 | |
| 1859 | * Orphan the callbacks. First adjust the counts. This is safe
|
| 1860 | * because _rcu_barrier() excludes CPU-hotplug operations, so it
|
| 1861 | * cannot be running now. Thus no memory barrier is required.
|
| 1862 |
|
| 1863 | if (rdp->nxtlist != NULL) {
|
| 1864 | rsp->qlen_lazy += rdp->qlen_lazy;
|
| 1865 | rsp->qlen += rdp->qlen;
|
| 1866 | rdp->n_cbs_orphaned += rdp->qlen;
|
| 1867 | rdp->qlen_lazy = 0;
|
| 1868 | ACCESS_ONCE(rdp->qlen) = 0;
|
| 1869 | }
|
| 1870 |
|
| 1871 | |
| 1872 | * Next, move those callbacks still needing a grace period to
|
| 1873 | * the orphanage, where some other CPU will pick them up.
|
| 1874 | * Some of the callbacks might have gone partway through a grace
|
| 1875 | * period, but that is too bad. They get to start over because we
|
| 1876 | * cannot assume that grace periods are synchronized across CPUs.
|
| 1877 | * We don't bother updating the ->nxttail[] array yet, instead
|
| 1878 | * we just reset the whole thing later on.
|
| 1879 |
|
| 1880 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
|
| 1881 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
|
| 1882 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
|
| 1883 | *rdp->nxttail[RCU_DONE_TAIL] = NULL;
|
| 1884 | }
|
| 1885 |
|
| 1886 | |
| 1887 | * Then move the ready-to-invoke callbacks to the orphanage,
|
| 1888 | * where some other CPU will pick them up. These will not be
|
| 1889 | * required to pass though another grace period: They are done.
|
| 1890 |
|
| 1891 | if (rdp->nxtlist != NULL) {
|
| 1892 | *rsp->orphan_donetail = rdp->nxtlist;
|
| 1893 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
|
| 1894 | }
|
| 1895 |
|
| 1896 |
|
| 1897 | init_callback_list(rdp);
|
| 1898 | }
|
| 1899 |
|
| 1900 | |
| 1901 | * Adopt the RCU callbacks from the specified rcu_state structure's
|
| 1902 | * orphanage. The caller must hold the ->orphan_lock.
|
| 1903 |
|
| 1904 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
|
| 1905 | {
|
| 1906 | int i;
|
| 1907 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
|
| 1908 |
|
| 1909 |
|
| 1910 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
|
| 1911 | return;
|
| 1912 |
|
| 1913 |
|
| 1914 | rdp->qlen_lazy += rsp->qlen_lazy;
|
| 1915 | rdp->qlen += rsp->qlen;
|
| 1916 | rdp->n_cbs_adopted += rsp->qlen;
|
| 1917 | if (rsp->qlen_lazy != rsp->qlen)
|
| 1918 | rcu_idle_count_callbacks_posted();
|
| 1919 | rsp->qlen_lazy = 0;
|
| 1920 | rsp->qlen = 0;
|
| 1921 |
|
| 1922 | |
| 1923 | * We do not need a memory barrier here because the only way we
|
| 1924 | * can get here if there is an rcu_barrier() in flight is if
|
| 1925 | * we are the task doing the rcu_barrier().
|
| 1926 |
|
| 1927 |
|
| 1928 |
|
| 1929 | if (rsp->orphan_donelist != NULL) {
|
| 1930 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
|
| 1931 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
|
| 1932 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
|
| 1933 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
|
| 1934 | rdp->nxttail[i] = rsp->orphan_donetail;
|
| 1935 | rsp->orphan_donelist = NULL;
|
| 1936 | rsp->orphan_donetail = &rsp->orphan_donelist;
|
| 1937 | }
|
| 1938 |
|
| 1939 |
|
| 1940 | if (rsp->orphan_nxtlist != NULL) {
|
| 1941 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
|
| 1942 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
|
| 1943 | rsp->orphan_nxtlist = NULL;
|
| 1944 | rsp->orphan_nxttail = &rsp->orphan_nxtlist;
|
| 1945 | }
|
| 1946 | }
|
| 1947 |
|
| 1948 | |
| 1949 | * Trace the fact that this CPU is going offline.
|
| 1950 |
|
| 1951 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
|
| 1952 | {
|
| 1953 | RCU_TRACE(unsigned long mask);
|
| 1954 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
|
| 1955 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
|
| 1956 |
|
| 1957 | RCU_TRACE(mask = rdp->grpmask);
|
| 1958 | trace_rcu_grace_period(rsp->name,
|
| 1959 | rnp->gpnum + 1 - !!(rnp->qsmask & mask),
|
| 1960 | TPS("cpuofl"));
|
| 1961 | }
|
| 1962 |
|
| 1963 | |
| 1964 | * The CPU has been completely removed, and some other CPU is reporting
|
| 1965 | * this fact from process context. Do the remainder of the cleanup,
|
| 1966 | * including orphaning the outgoing CPU's RCU callbacks, and also
|
| 1967 | * adopting them. There can only be one CPU hotplug operation at a time,
|
| 1968 | * so no other CPU can be attempting to update rcu_cpu_kthread_task.
|
| 1969 |
|
| 1970 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
|
| 1971 | {
|
| 1972 | unsigned long flags;
|
| 1973 | unsigned long mask;
|
| 1974 | int need_report = 0;
|
| 1975 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 1976 | struct rcu_node *rnp = rdp->mynode;
|
| 1977 |
|
| 1978 |
|
| 1979 | rcu_boost_kthread_setaffinity(rnp, -1);
|
| 1980 |
|
| 1981 |
|
| 1982 |
|
| 1983 |
|
| 1984 | mutex_lock(&rsp->onoff_mutex);
|
| 1985 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
|
| 1986 |
|
| 1987 |
|
| 1988 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
|
| 1989 | rcu_adopt_orphan_cbs(rsp);
|
| 1990 |
|
| 1991 |
|
| 1992 | mask = rdp->grpmask;
|
| 1993 | do {
|
| 1994 | raw_spin_lock(&rnp->lock);
|
| 1995 | rnp->qsmaskinit &= ~mask;
|
| 1996 | if (rnp->qsmaskinit != 0) {
|
| 1997 | if (rnp != rdp->mynode)
|
| 1998 | raw_spin_unlock(&rnp->lock);
|
| 1999 | break;
|
| 2000 | }
|
| 2001 | if (rnp == rdp->mynode)
|
| 2002 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
|
| 2003 | else
|
| 2004 | raw_spin_unlock(&rnp->lock);
|
| 2005 | mask = rnp->grpmask;
|
| 2006 | rnp = rnp->parent;
|
| 2007 | } while (rnp != NULL);
|
| 2008 |
|
| 2009 | |
| 2010 | * We still hold the leaf rcu_node structure lock here, and
|
| 2011 | * irqs are still disabled. The reason for this subterfuge is
|
| 2012 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
|
| 2013 | * held leads to deadlock.
|
| 2014 |
|
| 2015 | raw_spin_unlock(&rsp->orphan_lock);
|
| 2016 | rnp = rdp->mynode;
|
| 2017 | if (need_report & RCU_OFL_TASKS_NORM_GP)
|
| 2018 | rcu_report_unblock_qs_rnp(rnp, flags);
|
| 2019 | else
|
| 2020 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 2021 | if (need_report & RCU_OFL_TASKS_EXP_GP)
|
| 2022 | rcu_report_exp_rnp(rsp, rnp, true);
|
| 2023 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
|
| 2024 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
|
| 2025 | cpu, rdp->qlen, rdp->nxtlist);
|
| 2026 | init_callback_list(rdp);
|
| 2027 |
|
| 2028 | rdp->nxttail[RCU_NEXT_TAIL] = NULL;
|
| 2029 | mutex_unlock(&rsp->onoff_mutex);
|
| 2030 | }
|
| 2031 |
|
| 2032 | #else
|
| 2033 |
|
| 2034 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
|
| 2035 | {
|
| 2036 | }
|
| 2037 |
|
| 2038 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
|
| 2039 | {
|
| 2040 | }
|
| 2041 |
|
| 2042 | #endif
|
| 2043 |
|
| 2044 | |
| 2045 | * Invoke any RCU callbacks that have made it to the end of their grace
|
| 2046 | * period. Thottle as specified by rdp->blimit.
|
| 2047 |
|
| 2048 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 2049 | {
|
| 2050 | unsigned long flags;
|
| 2051 | struct rcu_head *next, *list, **tail;
|
| 2052 | long bl, count, count_lazy;
|
| 2053 | int i;
|
| 2054 |
|
| 2055 |
|
| 2056 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
|
| 2057 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
|
| 2058 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
|
| 2059 | need_resched(), is_idle_task(current),
|
| 2060 | rcu_is_callbacks_kthread());
|
| 2061 | return;
|
| 2062 | }
|
| 2063 |
|
| 2064 | |
| 2065 | * Extract the list of ready callbacks, disabling to prevent
|
| 2066 | * races with call_rcu() from interrupt handlers.
|
| 2067 |
|
| 2068 | local_irq_save(flags);
|
| 2069 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
|
| 2070 | bl = rdp->blimit;
|
| 2071 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
|
| 2072 | list = rdp->nxtlist;
|
| 2073 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
|
| 2074 | *rdp->nxttail[RCU_DONE_TAIL] = NULL;
|
| 2075 | tail = rdp->nxttail[RCU_DONE_TAIL];
|
| 2076 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
|
| 2077 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
|
| 2078 | rdp->nxttail[i] = &rdp->nxtlist;
|
| 2079 | local_irq_restore(flags);
|
| 2080 |
|
| 2081 |
|
| 2082 | count = count_lazy = 0;
|
| 2083 | while (list) {
|
| 2084 | next = list->next;
|
| 2085 | prefetch(next);
|
| 2086 | debug_rcu_head_unqueue(list);
|
| 2087 | if (__rcu_reclaim(rsp->name, list))
|
| 2088 | count_lazy++;
|
| 2089 | list = next;
|
| 2090 |
|
| 2091 | if (++count >= bl &&
|
| 2092 | (need_resched() ||
|
| 2093 | (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
|
| 2094 | break;
|
| 2095 | }
|
| 2096 |
|
| 2097 | local_irq_save(flags);
|
| 2098 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
|
| 2099 | is_idle_task(current),
|
| 2100 | rcu_is_callbacks_kthread());
|
| 2101 |
|
| 2102 |
|
| 2103 | if (list != NULL) {
|
| 2104 | *tail = rdp->nxtlist;
|
| 2105 | rdp->nxtlist = list;
|
| 2106 | for (i = 0; i < RCU_NEXT_SIZE; i++)
|
| 2107 | if (&rdp->nxtlist == rdp->nxttail[i])
|
| 2108 | rdp->nxttail[i] = tail;
|
| 2109 | else
|
| 2110 | break;
|
| 2111 | }
|
| 2112 | smp_mb();
|
| 2113 | rdp->qlen_lazy -= count_lazy;
|
| 2114 | ACCESS_ONCE(rdp->qlen) -= count;
|
| 2115 | rdp->n_cbs_invoked += count;
|
| 2116 |
|
| 2117 |
|
| 2118 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
|
| 2119 | rdp->blimit = blimit;
|
| 2120 |
|
| 2121 |
|
| 2122 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
|
| 2123 | rdp->qlen_last_fqs_check = 0;
|
| 2124 | rdp->n_force_qs_snap = rsp->n_force_qs;
|
| 2125 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
|
| 2126 | rdp->qlen_last_fqs_check = rdp->qlen;
|
| 2127 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
|
| 2128 |
|
| 2129 | local_irq_restore(flags);
|
| 2130 |
|
| 2131 |
|
| 2132 | if (cpu_has_callbacks_ready_to_invoke(rdp))
|
| 2133 | invoke_rcu_core();
|
| 2134 | }
|
| 2135 |
|
| 2136 | |
| 2137 | * Check to see if this CPU is in a non-context-switch quiescent state
|
| 2138 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
|
| 2139 | * Also schedule RCU core processing.
|
| 2140 | *
|
| 2141 | * This function must be called from hardirq context. It is normally
|
| 2142 | * invoked from the scheduling-clock interrupt. If rcu_pending returns
|
| 2143 | * false, there is no point in invoking rcu_check_callbacks().
|
| 2144 |
|
| 2145 | void rcu_check_callbacks(int cpu, int user)
|
| 2146 | {
|
| 2147 | trace_rcu_utilization(TPS("Start scheduler-tick"));
|
| 2148 | increment_cpu_stall_ticks();
|
| 2149 | if (user || rcu_is_cpu_rrupt_from_idle()) {
|
| 2150 |
|
| 2151 | |
| 2152 | * Get here if this CPU took its interrupt from user
|
| 2153 | * mode or from the idle loop, and if this is not a
|
| 2154 | * nested interrupt. In this case, the CPU is in
|
| 2155 | * a quiescent state, so note it.
|
| 2156 | *
|
| 2157 | * No memory barrier is required here because both
|
| 2158 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
|
| 2159 | * variables that other CPUs neither access nor modify,
|
| 2160 | * at least not while the corresponding CPU is online.
|
| 2161 |
|
| 2162 |
|
| 2163 | rcu_sched_qs(cpu);
|
| 2164 | rcu_bh_qs(cpu);
|
| 2165 |
|
| 2166 | } else if (!in_softirq()) {
|
| 2167 |
|
| 2168 | |
| 2169 | * Get here if this CPU did not take its interrupt from
|
| 2170 | * softirq, in other words, if it is not interrupting
|
| 2171 | * a rcu_bh read-side critical section. This is an _bh
|
| 2172 | * critical section, so note it.
|
| 2173 |
|
| 2174 |
|
| 2175 | rcu_bh_qs(cpu);
|
| 2176 | }
|
| 2177 | rcu_preempt_check_callbacks(cpu);
|
| 2178 | if (rcu_pending(cpu))
|
| 2179 | invoke_rcu_core();
|
| 2180 | trace_rcu_utilization(TPS("End scheduler-tick"));
|
| 2181 | }
|
| 2182 |
|
| 2183 | |
| 2184 | * Scan the leaf rcu_node structures, processing dyntick state for any that
|
| 2185 | * have not yet encountered a quiescent state, using the function specified.
|
| 2186 | * Also initiate boosting for any threads blocked on the root rcu_node.
|
| 2187 | *
|
| 2188 | * The caller must have suppressed start of new grace periods.
|
| 2189 |
|
| 2190 | static void force_qs_rnp(struct rcu_state *rsp,
|
| 2191 | int (*f)(struct rcu_data *rsp, bool *isidle,
|
| 2192 | unsigned long *maxj),
|
| 2193 | bool *isidle, unsigned long *maxj)
|
| 2194 | {
|
| 2195 | unsigned long bit;
|
| 2196 | int cpu;
|
| 2197 | unsigned long flags;
|
| 2198 | unsigned long mask;
|
| 2199 | struct rcu_node *rnp;
|
| 2200 |
|
| 2201 | rcu_for_each_leaf_node(rsp, rnp) {
|
| 2202 | cond_resched();
|
| 2203 | mask = 0;
|
| 2204 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 2205 | if (!rcu_gp_in_progress(rsp)) {
|
| 2206 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 2207 | return;
|
| 2208 | }
|
| 2209 | if (rnp->qsmask == 0) {
|
| 2210 | rcu_initiate_boost(rnp, flags);
|
| 2211 | continue;
|
| 2212 | }
|
| 2213 | cpu = rnp->grplo;
|
| 2214 | bit = 1;
|
| 2215 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
|
| 2216 | if ((rnp->qsmask & bit) != 0) {
|
| 2217 | if ((rnp->qsmaskinit & bit) != 0)
|
| 2218 | *isidle = 0;
|
| 2219 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
|
| 2220 | mask |= bit;
|
| 2221 | }
|
| 2222 | }
|
| 2223 | if (mask != 0) {
|
| 2224 |
|
| 2225 |
|
| 2226 | rcu_report_qs_rnp(mask, rsp, rnp, flags);
|
| 2227 | continue;
|
| 2228 | }
|
| 2229 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 2230 | }
|
| 2231 | rnp = rcu_get_root(rsp);
|
| 2232 | if (rnp->qsmask == 0) {
|
| 2233 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 2234 | rcu_initiate_boost(rnp, flags);
|
| 2235 | }
|
| 2236 | }
|
| 2237 |
|
| 2238 | |
| 2239 | * Force quiescent states on reluctant CPUs, and also detect which
|
| 2240 | * CPUs are in dyntick-idle mode.
|
| 2241 |
|
| 2242 | static void force_quiescent_state(struct rcu_state *rsp)
|
| 2243 | {
|
| 2244 | unsigned long flags;
|
| 2245 | bool ret;
|
| 2246 | struct rcu_node *rnp;
|
| 2247 | struct rcu_node *rnp_old = NULL;
|
| 2248 |
|
| 2249 |
|
| 2250 | rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
|
| 2251 | for (; rnp != NULL; rnp = rnp->parent) {
|
| 2252 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
|
| 2253 | !raw_spin_trylock(&rnp->fqslock);
|
| 2254 | if (rnp_old != NULL)
|
| 2255 | raw_spin_unlock(&rnp_old->fqslock);
|
| 2256 | if (ret) {
|
| 2257 | rsp->n_force_qs_lh++;
|
| 2258 | return;
|
| 2259 | }
|
| 2260 | rnp_old = rnp;
|
| 2261 | }
|
| 2262 |
|
| 2263 |
|
| 2264 |
|
| 2265 | raw_spin_lock_irqsave(&rnp_old->lock, flags);
|
| 2266 | raw_spin_unlock(&rnp_old->fqslock);
|
| 2267 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
|
| 2268 | rsp->n_force_qs_lh++;
|
| 2269 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
|
| 2270 | return;
|
| 2271 | }
|
| 2272 | rsp->gp_flags |= RCU_GP_FLAG_FQS;
|
| 2273 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
|
| 2274 | wake_up(&rsp->gp_wq);
|
| 2275 | }
|
| 2276 |
|
| 2277 | |
| 2278 | * This does the RCU core processing work for the specified rcu_state
|
| 2279 | * and rcu_data structures. This may be called only from the CPU to
|
| 2280 | * whom the rdp belongs.
|
| 2281 |
|
| 2282 | static void
|
| 2283 | __rcu_process_callbacks(struct rcu_state *rsp)
|
| 2284 | {
|
| 2285 | unsigned long flags;
|
| 2286 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
|
| 2287 |
|
| 2288 | WARN_ON_ONCE(rdp->beenonline == 0);
|
| 2289 |
|
| 2290 |
|
| 2291 | rcu_check_quiescent_state(rsp, rdp);
|
| 2292 |
|
| 2293 |
|
| 2294 | local_irq_save(flags);
|
| 2295 | if (cpu_needs_another_gp(rsp, rdp)) {
|
| 2296 | raw_spin_lock(&rcu_get_root(rsp)->lock);
|
| 2297 | rcu_start_gp(rsp);
|
| 2298 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
|
| 2299 | } else {
|
| 2300 | local_irq_restore(flags);
|
| 2301 | }
|
| 2302 |
|
| 2303 |
|
| 2304 | if (cpu_has_callbacks_ready_to_invoke(rdp))
|
| 2305 | invoke_rcu_callbacks(rsp, rdp);
|
| 2306 | }
|
| 2307 |
|
| 2308 | |
| 2309 | * Do RCU core processing for the current CPU.
|
| 2310 |
|
| 2311 | static void rcu_process_callbacks(struct softirq_action *unused)
|
| 2312 | {
|
| 2313 | struct rcu_state *rsp;
|
| 2314 |
|
| 2315 | if (cpu_is_offline(smp_processor_id()))
|
| 2316 | return;
|
| 2317 | trace_rcu_utilization(TPS("Start RCU core"));
|
| 2318 | for_each_rcu_flavor(rsp)
|
| 2319 | __rcu_process_callbacks(rsp);
|
| 2320 | trace_rcu_utilization(TPS("End RCU core"));
|
| 2321 | }
|
| 2322 |
|
| 2323 | |
| 2324 | * Schedule RCU callback invocation. If the specified type of RCU
|
| 2325 | * does not support RCU priority boosting, just do a direct call,
|
| 2326 | * otherwise wake up the per-CPU kernel kthread. Note that because we
|
| 2327 | * are running on the current CPU with interrupts disabled, the
|
| 2328 | * rcu_cpu_kthread_task cannot disappear out from under us.
|
| 2329 |
|
| 2330 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 2331 | {
|
| 2332 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
|
| 2333 | return;
|
| 2334 | if (likely(!rsp->boost)) {
|
| 2335 | rcu_do_batch(rsp, rdp);
|
| 2336 | return;
|
| 2337 | }
|
| 2338 | invoke_rcu_callbacks_kthread();
|
| 2339 | }
|
| 2340 |
|
| 2341 | static void invoke_rcu_core(void)
|
| 2342 | {
|
| 2343 | if (cpu_online(smp_processor_id()))
|
| 2344 | raise_softirq(RCU_SOFTIRQ);
|
| 2345 | }
|
| 2346 |
|
| 2347 | |
| 2348 | * Handle any core-RCU processing required by a call_rcu() invocation.
|
| 2349 |
|
| 2350 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
|
| 2351 | struct rcu_head *head, unsigned long flags)
|
| 2352 | {
|
| 2353 | |
| 2354 | * If called from an extended quiescent state, invoke the RCU
|
| 2355 | * core in order to force a re-evaluation of RCU's idleness.
|
| 2356 |
|
| 2357 | if (!rcu_is_watching() && cpu_online(smp_processor_id()))
|
| 2358 | invoke_rcu_core();
|
| 2359 |
|
| 2360 |
|
| 2361 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
|
| 2362 | return;
|
| 2363 |
|
| 2364 | |
| 2365 | * Force the grace period if too many callbacks or too long waiting.
|
| 2366 | * Enforce hysteresis, and don't invoke force_quiescent_state()
|
| 2367 | * if some other CPU has recently done so. Also, don't bother
|
| 2368 | * invoking force_quiescent_state() if the newly enqueued callback
|
| 2369 | * is the only one waiting for a grace period to complete.
|
| 2370 |
|
| 2371 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
|
| 2372 |
|
| 2373 |
|
| 2374 | note_gp_changes(rsp, rdp);
|
| 2375 |
|
| 2376 |
|
| 2377 | if (!rcu_gp_in_progress(rsp)) {
|
| 2378 | struct rcu_node *rnp_root = rcu_get_root(rsp);
|
| 2379 |
|
| 2380 | raw_spin_lock(&rnp_root->lock);
|
| 2381 | rcu_start_gp(rsp);
|
| 2382 | raw_spin_unlock(&rnp_root->lock);
|
| 2383 | } else {
|
| 2384 |
|
| 2385 | rdp->blimit = LONG_MAX;
|
| 2386 | if (rsp->n_force_qs == rdp->n_force_qs_snap &&
|
| 2387 | *rdp->nxttail[RCU_DONE_TAIL] != head)
|
| 2388 | force_quiescent_state(rsp);
|
| 2389 | rdp->n_force_qs_snap = rsp->n_force_qs;
|
| 2390 | rdp->qlen_last_fqs_check = rdp->qlen;
|
| 2391 | }
|
| 2392 | }
|
| 2393 | }
|
| 2394 |
|
| 2395 | |
| 2396 | * RCU callback function to leak a callback.
|
| 2397 |
|
| 2398 | static void rcu_leak_callback(struct rcu_head *rhp)
|
| 2399 | {
|
| 2400 | }
|
| 2401 |
|
| 2402 | |
| 2403 | * Helper function for call_rcu() and friends. The cpu argument will
|
| 2404 | * normally be -1, indicating "currently running CPU". It may specify
|
| 2405 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
|
| 2406 | * is expected to specify a CPU.
|
| 2407 |
|
| 2408 | static void
|
| 2409 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
|
| 2410 | struct rcu_state *rsp, int cpu, bool lazy)
|
| 2411 | {
|
| 2412 | unsigned long flags;
|
| 2413 | struct rcu_data *rdp;
|
| 2414 |
|
| 2415 | WARN_ON_ONCE((unsigned long)head & 0x3);
|
| 2416 | if (debug_rcu_head_queue(head)) {
|
| 2417 |
|
| 2418 | ACCESS_ONCE(head->func) = rcu_leak_callback;
|
| 2419 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
|
| 2420 | return;
|
| 2421 | }
|
| 2422 | head->func = func;
|
| 2423 | head->next = NULL;
|
| 2424 |
|
| 2425 | |
| 2426 | * Opportunistically note grace-period endings and beginnings.
|
| 2427 | * Note that we might see a beginning right after we see an
|
| 2428 | * end, but never vice versa, since this CPU has to pass through
|
| 2429 | * a quiescent state betweentimes.
|
| 2430 |
|
| 2431 | local_irq_save(flags);
|
| 2432 | rdp = this_cpu_ptr(rsp->rda);
|
| 2433 |
|
| 2434 |
|
| 2435 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
|
| 2436 | int offline;
|
| 2437 |
|
| 2438 | if (cpu != -1)
|
| 2439 | rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 2440 | offline = !__call_rcu_nocb(rdp, head, lazy);
|
| 2441 | WARN_ON_ONCE(offline);
|
| 2442 |
|
| 2443 | local_irq_restore(flags);
|
| 2444 | return;
|
| 2445 | }
|
| 2446 | ACCESS_ONCE(rdp->qlen)++;
|
| 2447 | if (lazy)
|
| 2448 | rdp->qlen_lazy++;
|
| 2449 | else
|
| 2450 | rcu_idle_count_callbacks_posted();
|
| 2451 | smp_mb();
|
| 2452 | *rdp->nxttail[RCU_NEXT_TAIL] = head;
|
| 2453 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
|
| 2454 |
|
| 2455 | if (__is_kfree_rcu_offset((unsigned long)func))
|
| 2456 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
|
| 2457 | rdp->qlen_lazy, rdp->qlen);
|
| 2458 | else
|
| 2459 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
|
| 2460 |
|
| 2461 |
|
| 2462 | __call_rcu_core(rsp, rdp, head, flags);
|
| 2463 | local_irq_restore(flags);
|
| 2464 | }
|
| 2465 |
|
| 2466 | |
| 2467 | * Queue an RCU-sched callback for invocation after a grace period.
|
| 2468 |
|
| 2469 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
|
| 2470 | {
|
| 2471 | __call_rcu(head, func, &rcu_sched_state, -1, 0);
|
| 2472 | }
|
| 2473 | EXPORT_SYMBOL_GPL(call_rcu_sched);
|
| 2474 |
|
| 2475 | |
| 2476 | * Queue an RCU callback for invocation after a quicker grace period.
|
| 2477 |
|
| 2478 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
|
| 2479 | {
|
| 2480 | __call_rcu(head, func, &rcu_bh_state, -1, 0);
|
| 2481 | }
|
| 2482 | EXPORT_SYMBOL_GPL(call_rcu_bh);
|
| 2483 |
|
| 2484 | |
| 2485 | * Because a context switch is a grace period for RCU-sched and RCU-bh,
|
| 2486 | * any blocking grace-period wait automatically implies a grace period
|
| 2487 | * if there is only one CPU online at any point time during execution
|
| 2488 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
|
| 2489 | * occasionally incorrectly indicate that there are multiple CPUs online
|
| 2490 | * when there was in fact only one the whole time, as this just adds
|
| 2491 | * some overhead: RCU still operates correctly.
|
| 2492 |
|
| 2493 | static inline int rcu_blocking_is_gp(void)
|
| 2494 | {
|
| 2495 | int ret;
|
| 2496 |
|
| 2497 | might_sleep();
|
| 2498 | preempt_disable();
|
| 2499 | ret = num_online_cpus() <= 1;
|
| 2500 | preempt_enable();
|
| 2501 | return ret;
|
| 2502 | }
|
| 2503 |
|
| 2504 | |
| 2505 | * synchronize_sched - wait until an rcu-sched grace period has elapsed.
|
| 2506 | *
|
| 2507 | * Control will return to the caller some time after a full rcu-sched
|
| 2508 | * grace period has elapsed, in other words after all currently executing
|
| 2509 | * rcu-sched read-side critical sections have completed. These read-side
|
| 2510 | * critical sections are delimited by rcu_read_lock_sched() and
|
| 2511 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
|
| 2512 | * local_irq_disable(), and so on may be used in place of
|
| 2513 | * rcu_read_lock_sched().
|
| 2514 | *
|
| 2515 | * This means that all preempt_disable code sequences, including NMI and
|
| 2516 | * non-threaded hardware-interrupt handlers, in progress on entry will
|
| 2517 | * have completed before this primitive returns. However, this does not
|
| 2518 | * guarantee that softirq handlers will have completed, since in some
|
| 2519 | * kernels, these handlers can run in process context, and can block.
|
| 2520 | *
|
| 2521 | * Note that this guarantee implies further memory-ordering guarantees.
|
| 2522 | * On systems with more than one CPU, when synchronize_sched() returns,
|
| 2523 | * each CPU is guaranteed to have executed a full memory barrier since the
|
| 2524 | * end of its last RCU-sched read-side critical section whose beginning
|
| 2525 | * preceded the call to synchronize_sched(). In addition, each CPU having
|
| 2526 | * an RCU read-side critical section that extends beyond the return from
|
| 2527 | * synchronize_sched() is guaranteed to have executed a full memory barrier
|
| 2528 | * after the beginning of synchronize_sched() and before the beginning of
|
| 2529 | * that RCU read-side critical section. Note that these guarantees include
|
| 2530 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs
|
| 2531 | * that are executing in the kernel.
|
| 2532 | *
|
| 2533 | * Furthermore, if CPU A invoked synchronize_sched(), which returned
|
| 2534 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed
|
| 2535 | * to have executed a full memory barrier during the execution of
|
| 2536 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
|
| 2537 | * again only if the system has more than one CPU).
|
| 2538 | *
|
| 2539 | * This primitive provides the guarantees made by the (now removed)
|
| 2540 | * synchronize_kernel() API. In contrast, synchronize_rcu() only
|
| 2541 | * guarantees that rcu_read_lock() sections will have completed.
|
| 2542 | * In "classic RCU", these two guarantees happen to be one and
|
| 2543 | * the same, but can differ in realtime RCU implementations.
|
| 2544 |
|
| 2545 | void synchronize_sched(void)
|
| 2546 | {
|
| 2547 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
|
| 2548 | !lock_is_held(&rcu_lock_map) &&
|
| 2549 | !lock_is_held(&rcu_sched_lock_map),
|
| 2550 | "Illegal synchronize_sched() in RCU-sched read-side critical section");
|
| 2551 | if (rcu_blocking_is_gp())
|
| 2552 | return;
|
| 2553 | if (rcu_expedited)
|
| 2554 | synchronize_sched_expedited();
|
| 2555 | else
|
| 2556 | wait_rcu_gp(call_rcu_sched);
|
| 2557 | }
|
| 2558 | EXPORT_SYMBOL_GPL(synchronize_sched);
|
| 2559 |
|
| 2560 | |
| 2561 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
|
| 2562 | *
|
| 2563 | * Control will return to the caller some time after a full rcu_bh grace
|
| 2564 | * period has elapsed, in other words after all currently executing rcu_bh
|
| 2565 | * read-side critical sections have completed. RCU read-side critical
|
| 2566 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
|
| 2567 | * and may be nested.
|
| 2568 | *
|
| 2569 | * See the description of synchronize_sched() for more detailed information
|
| 2570 | * on memory ordering guarantees.
|
| 2571 |
|
| 2572 | void synchronize_rcu_bh(void)
|
| 2573 | {
|
| 2574 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
|
| 2575 | !lock_is_held(&rcu_lock_map) &&
|
| 2576 | !lock_is_held(&rcu_sched_lock_map),
|
| 2577 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
|
| 2578 | if (rcu_blocking_is_gp())
|
| 2579 | return;
|
| 2580 | if (rcu_expedited)
|
| 2581 | synchronize_rcu_bh_expedited();
|
| 2582 | else
|
| 2583 | wait_rcu_gp(call_rcu_bh);
|
| 2584 | }
|
| 2585 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
|
| 2586 |
|
| 2587 | static int synchronize_sched_expedited_cpu_stop(void *data)
|
| 2588 | {
|
| 2589 | |
| 2590 | * There must be a full memory barrier on each affected CPU
|
| 2591 | * between the time that try_stop_cpus() is called and the
|
| 2592 | * time that it returns.
|
| 2593 | *
|
| 2594 | * In the current initial implementation of cpu_stop, the
|
| 2595 | * above condition is already met when the control reaches
|
| 2596 | * this point and the following smp_mb() is not strictly
|
| 2597 | * necessary. Do smp_mb() anyway for documentation and
|
| 2598 | * robustness against future implementation changes.
|
| 2599 |
|
| 2600 | smp_mb();
|
| 2601 | return 0;
|
| 2602 | }
|
| 2603 |
|
| 2604 | |
| 2605 | * synchronize_sched_expedited - Brute-force RCU-sched grace period
|
| 2606 | *
|
| 2607 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
|
| 2608 | * approach to force the grace period to end quickly. This consumes
|
| 2609 | * significant time on all CPUs and is unfriendly to real-time workloads,
|
| 2610 | * so is thus not recommended for any sort of common-case code. In fact,
|
| 2611 | * if you are using synchronize_sched_expedited() in a loop, please
|
| 2612 | * restructure your code to batch your updates, and then use a single
|
| 2613 | * synchronize_sched() instead.
|
| 2614 | *
|
| 2615 | * Note that it is illegal to call this function while holding any lock
|
| 2616 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
|
| 2617 | * to call this function from a CPU-hotplug notifier. Failing to observe
|
| 2618 | * these restriction will result in deadlock.
|
| 2619 | *
|
| 2620 | * This implementation can be thought of as an application of ticket
|
| 2621 | * locking to RCU, with sync_sched_expedited_started and
|
| 2622 | * sync_sched_expedited_done taking on the roles of the halves
|
| 2623 | * of the ticket-lock word. Each task atomically increments
|
| 2624 | * sync_sched_expedited_started upon entry, snapshotting the old value,
|
| 2625 | * then attempts to stop all the CPUs. If this succeeds, then each
|
| 2626 | * CPU will have executed a context switch, resulting in an RCU-sched
|
| 2627 | * grace period. We are then done, so we use atomic_cmpxchg() to
|
| 2628 | * update sync_sched_expedited_done to match our snapshot -- but
|
| 2629 | * only if someone else has not already advanced past our snapshot.
|
| 2630 | *
|
| 2631 | * On the other hand, if try_stop_cpus() fails, we check the value
|
| 2632 | * of sync_sched_expedited_done. If it has advanced past our
|
| 2633 | * initial snapshot, then someone else must have forced a grace period
|
| 2634 | * some time after we took our snapshot. In this case, our work is
|
| 2635 | * done for us, and we can simply return. Otherwise, we try again,
|
| 2636 | * but keep our initial snapshot for purposes of checking for someone
|
| 2637 | * doing our work for us.
|
| 2638 | *
|
| 2639 | * If we fail too many times in a row, we fall back to synchronize_sched().
|
| 2640 |
|
| 2641 | void synchronize_sched_expedited(void)
|
| 2642 | {
|
| 2643 | long firstsnap, s, snap;
|
| 2644 | int trycount = 0;
|
| 2645 | struct rcu_state *rsp = &rcu_sched_state;
|
| 2646 |
|
| 2647 | |
| 2648 | * If we are in danger of counter wrap, just do synchronize_sched().
|
| 2649 | * By allowing sync_sched_expedited_started to advance no more than
|
| 2650 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
|
| 2651 | * that more than 3.5 billion CPUs would be required to force a
|
| 2652 | * counter wrap on a 32-bit system. Quite a few more CPUs would of
|
| 2653 | * course be required on a 64-bit system.
|
| 2654 |
|
| 2655 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
|
| 2656 | (ulong)atomic_long_read(&rsp->expedited_done) +
|
| 2657 | ULONG_MAX / 8)) {
|
| 2658 | synchronize_sched();
|
| 2659 | atomic_long_inc(&rsp->expedited_wrap);
|
| 2660 | return;
|
| 2661 | }
|
| 2662 |
|
| 2663 | |
| 2664 | * Take a ticket. Note that atomic_inc_return() implies a
|
| 2665 | * full memory barrier.
|
| 2666 |
|
| 2667 | snap = atomic_long_inc_return(&rsp->expedited_start);
|
| 2668 | firstsnap = snap;
|
| 2669 | get_online_cpus();
|
| 2670 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
|
| 2671 |
|
| 2672 | |
| 2673 | * Each pass through the following loop attempts to force a
|
| 2674 | * context switch on each CPU.
|
| 2675 |
|
| 2676 | while (try_stop_cpus(cpu_online_mask,
|
| 2677 | synchronize_sched_expedited_cpu_stop,
|
| 2678 | NULL) == -EAGAIN) {
|
| 2679 | put_online_cpus();
|
| 2680 | atomic_long_inc(&rsp->expedited_tryfail);
|
| 2681 |
|
| 2682 |
|
| 2683 | s = atomic_long_read(&rsp->expedited_done);
|
| 2684 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
|
| 2685 |
|
| 2686 | smp_mb__before_atomic_inc();
|
| 2687 | atomic_long_inc(&rsp->expedited_workdone1);
|
| 2688 | return;
|
| 2689 | }
|
| 2690 |
|
| 2691 |
|
| 2692 | if (trycount++ < 10) {
|
| 2693 | udelay(trycount * num_online_cpus());
|
| 2694 | } else {
|
| 2695 | wait_rcu_gp(call_rcu_sched);
|
| 2696 | atomic_long_inc(&rsp->expedited_normal);
|
| 2697 | return;
|
| 2698 | }
|
| 2699 |
|
| 2700 |
|
| 2701 | s = atomic_long_read(&rsp->expedited_done);
|
| 2702 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
|
| 2703 |
|
| 2704 | smp_mb__before_atomic_inc();
|
| 2705 | atomic_long_inc(&rsp->expedited_workdone2);
|
| 2706 | return;
|
| 2707 | }
|
| 2708 |
|
| 2709 | |
| 2710 | * Refetching sync_sched_expedited_started allows later
|
| 2711 | * callers to piggyback on our grace period. We retry
|
| 2712 | * after they started, so our grace period works for them,
|
| 2713 | * and they started after our first try, so their grace
|
| 2714 | * period works for us.
|
| 2715 |
|
| 2716 | get_online_cpus();
|
| 2717 | snap = atomic_long_read(&rsp->expedited_start);
|
| 2718 | smp_mb();
|
| 2719 | }
|
| 2720 | atomic_long_inc(&rsp->expedited_stoppedcpus);
|
| 2721 |
|
| 2722 | |
| 2723 | * Everyone up to our most recent fetch is covered by our grace
|
| 2724 | * period. Update the counter, but only if our work is still
|
| 2725 | * relevant -- which it won't be if someone who started later
|
| 2726 | * than we did already did their update.
|
| 2727 |
|
| 2728 | do {
|
| 2729 | atomic_long_inc(&rsp->expedited_done_tries);
|
| 2730 | s = atomic_long_read(&rsp->expedited_done);
|
| 2731 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
|
| 2732 |
|
| 2733 | smp_mb__before_atomic_inc();
|
| 2734 | atomic_long_inc(&rsp->expedited_done_lost);
|
| 2735 | break;
|
| 2736 | }
|
| 2737 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
|
| 2738 | atomic_long_inc(&rsp->expedited_done_exit);
|
| 2739 |
|
| 2740 | put_online_cpus();
|
| 2741 | }
|
| 2742 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
|
| 2743 |
|
| 2744 | |
| 2745 | * Check to see if there is any immediate RCU-related work to be done
|
| 2746 | * by the current CPU, for the specified type of RCU, returning 1 if so.
|
| 2747 | * The checks are in order of increasing expense: checks that can be
|
| 2748 | * carried out against CPU-local state are performed first. However,
|
| 2749 | * we must check for CPU stalls first, else we might not get a chance.
|
| 2750 |
|
| 2751 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
|
| 2752 | {
|
| 2753 | struct rcu_node *rnp = rdp->mynode;
|
| 2754 |
|
| 2755 | rdp->n_rcu_pending++;
|
| 2756 |
|
| 2757 |
|
| 2758 | check_cpu_stall(rsp, rdp);
|
| 2759 |
|
| 2760 |
|
| 2761 | if (rcu_scheduler_fully_active &&
|
| 2762 | rdp->qs_pending && !rdp->passed_quiesce) {
|
| 2763 | rdp->n_rp_qs_pending++;
|
| 2764 | } else if (rdp->qs_pending && rdp->passed_quiesce) {
|
| 2765 | rdp->n_rp_report_qs++;
|
| 2766 | return 1;
|
| 2767 | }
|
| 2768 |
|
| 2769 |
|
| 2770 | if (cpu_has_callbacks_ready_to_invoke(rdp)) {
|
| 2771 | rdp->n_rp_cb_ready++;
|
| 2772 | return 1;
|
| 2773 | }
|
| 2774 |
|
| 2775 |
|
| 2776 | if (cpu_needs_another_gp(rsp, rdp)) {
|
| 2777 | rdp->n_rp_cpu_needs_gp++;
|
| 2778 | return 1;
|
| 2779 | }
|
| 2780 |
|
| 2781 |
|
| 2782 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) {
|
| 2783 | rdp->n_rp_gp_completed++;
|
| 2784 | return 1;
|
| 2785 | }
|
| 2786 |
|
| 2787 |
|
| 2788 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) {
|
| 2789 | rdp->n_rp_gp_started++;
|
| 2790 | return 1;
|
| 2791 | }
|
| 2792 |
|
| 2793 |
|
| 2794 | rdp->n_rp_need_nothing++;
|
| 2795 | return 0;
|
| 2796 | }
|
| 2797 |
|
| 2798 | |
| 2799 | * Check to see if there is any immediate RCU-related work to be done
|
| 2800 | * by the current CPU, returning 1 if so. This function is part of the
|
| 2801 | * RCU implementation; it is -not- an exported member of the RCU API.
|
| 2802 |
|
| 2803 | static int rcu_pending(int cpu)
|
| 2804 | {
|
| 2805 | struct rcu_state *rsp;
|
| 2806 |
|
| 2807 | for_each_rcu_flavor(rsp)
|
| 2808 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
|
| 2809 | return 1;
|
| 2810 | return 0;
|
| 2811 | }
|
| 2812 |
|
| 2813 | |
| 2814 | * Return true if the specified CPU has any callback. If all_lazy is
|
| 2815 | * non-NULL, store an indication of whether all callbacks are lazy.
|
| 2816 | * (If there are no callbacks, all of them are deemed to be lazy.)
|
| 2817 |
|
| 2818 | static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
|
| 2819 | {
|
| 2820 | bool al = true;
|
| 2821 | bool hc = false;
|
| 2822 | struct rcu_data *rdp;
|
| 2823 | struct rcu_state *rsp;
|
| 2824 |
|
| 2825 | for_each_rcu_flavor(rsp) {
|
| 2826 | rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 2827 | if (!rdp->nxtlist)
|
| 2828 | continue;
|
| 2829 | hc = true;
|
| 2830 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
|
| 2831 | al = false;
|
| 2832 | break;
|
| 2833 | }
|
| 2834 | }
|
| 2835 | if (all_lazy)
|
| 2836 | *all_lazy = al;
|
| 2837 | return hc;
|
| 2838 | }
|
| 2839 |
|
| 2840 | |
| 2841 | * Helper function for _rcu_barrier() tracing. If tracing is disabled,
|
| 2842 | * the compiler is expected to optimize this away.
|
| 2843 |
|
| 2844 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
|
| 2845 | int cpu, unsigned long done)
|
| 2846 | {
|
| 2847 | trace_rcu_barrier(rsp->name, s, cpu,
|
| 2848 | atomic_read(&rsp->barrier_cpu_count), done);
|
| 2849 | }
|
| 2850 |
|
| 2851 | |
| 2852 | * RCU callback function for _rcu_barrier(). If we are last, wake
|
| 2853 | * up the task executing _rcu_barrier().
|
| 2854 |
|
| 2855 | static void rcu_barrier_callback(struct rcu_head *rhp)
|
| 2856 | {
|
| 2857 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
|
| 2858 | struct rcu_state *rsp = rdp->rsp;
|
| 2859 |
|
| 2860 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
|
| 2861 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
|
| 2862 | complete(&rsp->barrier_completion);
|
| 2863 | } else {
|
| 2864 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
|
| 2865 | }
|
| 2866 | }
|
| 2867 |
|
| 2868 | |
| 2869 | * Called with preemption disabled, and from cross-cpu IRQ context.
|
| 2870 |
|
| 2871 | static void rcu_barrier_func(void *type)
|
| 2872 | {
|
| 2873 | struct rcu_state *rsp = type;
|
| 2874 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
|
| 2875 |
|
| 2876 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
|
| 2877 | atomic_inc(&rsp->barrier_cpu_count);
|
| 2878 | rsp->call(&rdp->barrier_head, rcu_barrier_callback);
|
| 2879 | }
|
| 2880 |
|
| 2881 | |
| 2882 | * Orchestrate the specified type of RCU barrier, waiting for all
|
| 2883 | * RCU callbacks of the specified type to complete.
|
| 2884 |
|
| 2885 | static void _rcu_barrier(struct rcu_state *rsp)
|
| 2886 | {
|
| 2887 | int cpu;
|
| 2888 | struct rcu_data *rdp;
|
| 2889 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
|
| 2890 | unsigned long snap_done;
|
| 2891 |
|
| 2892 | _rcu_barrier_trace(rsp, "Begin", -1, snap);
|
| 2893 |
|
| 2894 |
|
| 2895 | mutex_lock(&rsp->barrier_mutex);
|
| 2896 |
|
| 2897 | |
| 2898 | * Ensure that all prior references, including to ->n_barrier_done,
|
| 2899 | * are ordered before the _rcu_barrier() machinery.
|
| 2900 |
|
| 2901 | smp_mb();
|
| 2902 |
|
| 2903 | |
| 2904 | * Recheck ->n_barrier_done to see if others did our work for us.
|
| 2905 | * This means checking ->n_barrier_done for an even-to-odd-to-even
|
| 2906 | * transition. The "if" expression below therefore rounds the old
|
| 2907 | * value up to the next even number and adds two before comparing.
|
| 2908 |
|
| 2909 | snap_done = rsp->n_barrier_done;
|
| 2910 | _rcu_barrier_trace(rsp, "Check", -1, snap_done);
|
| 2911 |
|
| 2912 | |
| 2913 | * If the value in snap is odd, we needed to wait for the current
|
| 2914 | * rcu_barrier() to complete, then wait for the next one, in other
|
| 2915 | * words, we need the value of snap_done to be three larger than
|
| 2916 | * the value of snap. On the other hand, if the value in snap is
|
| 2917 | * even, we only had to wait for the next rcu_barrier() to complete,
|
| 2918 | * in other words, we need the value of snap_done to be only two
|
| 2919 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes
|
| 2920 | * this for us (thank you, Linus!).
|
| 2921 |
|
| 2922 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
|
| 2923 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
|
| 2924 | smp_mb();
|
| 2925 | mutex_unlock(&rsp->barrier_mutex);
|
| 2926 | return;
|
| 2927 | }
|
| 2928 |
|
| 2929 | |
| 2930 | * Increment ->n_barrier_done to avoid duplicate work. Use
|
| 2931 | * ACCESS_ONCE() to prevent the compiler from speculating
|
| 2932 | * the increment to precede the early-exit check.
|
| 2933 |
|
| 2934 | ACCESS_ONCE(rsp->n_barrier_done)++;
|
| 2935 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
|
| 2936 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
|
| 2937 | smp_mb();
|
| 2938 |
|
| 2939 | |
| 2940 | * Initialize the count to one rather than to zero in order to
|
| 2941 | * avoid a too-soon return to zero in case of a short grace period
|
| 2942 | * (or preemption of this task). Exclude CPU-hotplug operations
|
| 2943 | * to ensure that no offline CPU has callbacks queued.
|
| 2944 |
|
| 2945 | init_completion(&rsp->barrier_completion);
|
| 2946 | atomic_set(&rsp->barrier_cpu_count, 1);
|
| 2947 | get_online_cpus();
|
| 2948 |
|
| 2949 | |
| 2950 | * Force each CPU with callbacks to register a new callback.
|
| 2951 | * When that callback is invoked, we will know that all of the
|
| 2952 | * corresponding CPU's preceding callbacks have been invoked.
|
| 2953 |
|
| 2954 | for_each_possible_cpu(cpu) {
|
| 2955 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
|
| 2956 | continue;
|
| 2957 | rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 2958 | if (rcu_is_nocb_cpu(cpu)) {
|
| 2959 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
|
| 2960 | rsp->n_barrier_done);
|
| 2961 | atomic_inc(&rsp->barrier_cpu_count);
|
| 2962 | __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
|
| 2963 | rsp, cpu, 0);
|
| 2964 | } else if (ACCESS_ONCE(rdp->qlen)) {
|
| 2965 | _rcu_barrier_trace(rsp, "OnlineQ", cpu,
|
| 2966 | rsp->n_barrier_done);
|
| 2967 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
|
| 2968 | } else {
|
| 2969 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
|
| 2970 | rsp->n_barrier_done);
|
| 2971 | }
|
| 2972 | }
|
| 2973 | put_online_cpus();
|
| 2974 |
|
| 2975 | |
| 2976 | * Now that we have an rcu_barrier_callback() callback on each
|
| 2977 | * CPU, and thus each counted, remove the initial count.
|
| 2978 |
|
| 2979 | if (atomic_dec_and_test(&rsp->barrier_cpu_count))
|
| 2980 | complete(&rsp->barrier_completion);
|
| 2981 |
|
| 2982 |
|
| 2983 | smp_mb();
|
| 2984 | ACCESS_ONCE(rsp->n_barrier_done)++;
|
| 2985 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
|
| 2986 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
|
| 2987 | smp_mb();
|
| 2988 |
|
| 2989 |
|
| 2990 | wait_for_completion(&rsp->barrier_completion);
|
| 2991 |
|
| 2992 |
|
| 2993 | mutex_unlock(&rsp->barrier_mutex);
|
| 2994 | }
|
| 2995 |
|
| 2996 | |
| 2997 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
|
| 2998 |
|
| 2999 | void rcu_barrier_bh(void)
|
| 3000 | {
|
| 3001 | _rcu_barrier(&rcu_bh_state);
|
| 3002 | }
|
| 3003 | EXPORT_SYMBOL_GPL(rcu_barrier_bh);
|
| 3004 |
|
| 3005 | |
| 3006 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
|
| 3007 |
|
| 3008 | void rcu_barrier_sched(void)
|
| 3009 | {
|
| 3010 | _rcu_barrier(&rcu_sched_state);
|
| 3011 | }
|
| 3012 | EXPORT_SYMBOL_GPL(rcu_barrier_sched);
|
| 3013 |
|
| 3014 | |
| 3015 | * Do boot-time initialization of a CPU's per-CPU RCU data.
|
| 3016 |
|
| 3017 | static void __init
|
| 3018 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
|
| 3019 | {
|
| 3020 | unsigned long flags;
|
| 3021 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 3022 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 3023 |
|
| 3024 |
|
| 3025 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 3026 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
|
| 3027 | init_callback_list(rdp);
|
| 3028 | rdp->qlen_lazy = 0;
|
| 3029 | ACCESS_ONCE(rdp->qlen) = 0;
|
| 3030 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
|
| 3031 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
|
| 3032 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
|
| 3033 | rdp->cpu = cpu;
|
| 3034 | rdp->rsp = rsp;
|
| 3035 | rcu_boot_init_nocb_percpu_data(rdp);
|
| 3036 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 3037 | }
|
| 3038 |
|
| 3039 | |
| 3040 | * Initialize a CPU's per-CPU RCU data. Note that only one online or
|
| 3041 | * offline event can be happening at a given time. Note also that we
|
| 3042 | * can accept some slop in the rsp->completed access due to the fact
|
| 3043 | * that this CPU cannot possibly have any RCU callbacks in flight yet.
|
| 3044 |
|
| 3045 | static void
|
| 3046 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
|
| 3047 | {
|
| 3048 | unsigned long flags;
|
| 3049 | unsigned long mask;
|
| 3050 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
|
| 3051 | struct rcu_node *rnp = rcu_get_root(rsp);
|
| 3052 |
|
| 3053 |
|
| 3054 | mutex_lock(&rsp->onoff_mutex);
|
| 3055 |
|
| 3056 |
|
| 3057 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 3058 | rdp->beenonline = 1;
|
| 3059 | rdp->preemptible = preemptible;
|
| 3060 | rdp->qlen_last_fqs_check = 0;
|
| 3061 | rdp->n_force_qs_snap = rsp->n_force_qs;
|
| 3062 | rdp->blimit = blimit;
|
| 3063 | init_callback_list(rdp);
|
| 3064 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
|
| 3065 | rcu_sysidle_init_percpu_data(rdp->dynticks);
|
| 3066 | atomic_set(&rdp->dynticks->dynticks,
|
| 3067 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
|
| 3068 | raw_spin_unlock(&rnp->lock);
|
| 3069 |
|
| 3070 |
|
| 3071 | rnp = rdp->mynode;
|
| 3072 | mask = rdp->grpmask;
|
| 3073 | do {
|
| 3074 |
|
| 3075 | raw_spin_lock(&rnp->lock);
|
| 3076 | rnp->qsmaskinit |= mask;
|
| 3077 | mask = rnp->grpmask;
|
| 3078 | if (rnp == rdp->mynode) {
|
| 3079 | |
| 3080 | * If there is a grace period in progress, we will
|
| 3081 | * set up to wait for it next time we run the
|
| 3082 | * RCU core code.
|
| 3083 |
|
| 3084 | rdp->gpnum = rnp->completed;
|
| 3085 | rdp->completed = rnp->completed;
|
| 3086 | rdp->passed_quiesce = 0;
|
| 3087 | rdp->qs_pending = 0;
|
| 3088 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
|
| 3089 | }
|
| 3090 | raw_spin_unlock(&rnp->lock);
|
| 3091 | rnp = rnp->parent;
|
| 3092 | } while (rnp != NULL && !(rnp->qsmaskinit & mask));
|
| 3093 | local_irq_restore(flags);
|
| 3094 |
|
| 3095 | mutex_unlock(&rsp->onoff_mutex);
|
| 3096 | }
|
| 3097 |
|
| 3098 | static void rcu_prepare_cpu(int cpu)
|
| 3099 | {
|
| 3100 | struct rcu_state *rsp;
|
| 3101 |
|
| 3102 | for_each_rcu_flavor(rsp)
|
| 3103 | rcu_init_percpu_data(cpu, rsp,
|
| 3104 | strcmp(rsp->name, "rcu_preempt") == 0);
|
| 3105 | }
|
| 3106 |
|
| 3107 | |
| 3108 | * Handle CPU online/offline notification events.
|
| 3109 |
|
| 3110 | static int rcu_cpu_notify(struct notifier_block *self,
|
| 3111 | unsigned long action, void *hcpu)
|
| 3112 | {
|
| 3113 | long cpu = (long)hcpu;
|
| 3114 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
|
| 3115 | struct rcu_node *rnp = rdp->mynode;
|
| 3116 | struct rcu_state *rsp;
|
| 3117 |
|
| 3118 | trace_rcu_utilization(TPS("Start CPU hotplug"));
|
| 3119 | switch (action) {
|
| 3120 | case CPU_UP_PREPARE:
|
| 3121 | case CPU_UP_PREPARE_FROZEN:
|
| 3122 | rcu_prepare_cpu(cpu);
|
| 3123 | rcu_prepare_kthreads(cpu);
|
| 3124 | break;
|
| 3125 | case CPU_ONLINE:
|
| 3126 | case CPU_DOWN_FAILED:
|
| 3127 | rcu_boost_kthread_setaffinity(rnp, -1);
|
| 3128 | break;
|
| 3129 | case CPU_DOWN_PREPARE:
|
| 3130 | rcu_boost_kthread_setaffinity(rnp, cpu);
|
| 3131 | break;
|
| 3132 | case CPU_DYING:
|
| 3133 | case CPU_DYING_FROZEN:
|
| 3134 | for_each_rcu_flavor(rsp)
|
| 3135 | rcu_cleanup_dying_cpu(rsp);
|
| 3136 | break;
|
| 3137 | case CPU_DEAD:
|
| 3138 | case CPU_DEAD_FROZEN:
|
| 3139 | case CPU_UP_CANCELED:
|
| 3140 | case CPU_UP_CANCELED_FROZEN:
|
| 3141 | for_each_rcu_flavor(rsp)
|
| 3142 | rcu_cleanup_dead_cpu(cpu, rsp);
|
| 3143 | break;
|
| 3144 | default:
|
| 3145 | break;
|
| 3146 | }
|
| 3147 | trace_rcu_utilization(TPS("End CPU hotplug"));
|
| 3148 | return NOTIFY_OK;
|
| 3149 | }
|
| 3150 |
|
| 3151 | static int rcu_pm_notify(struct notifier_block *self,
|
| 3152 | unsigned long action, void *hcpu)
|
| 3153 | {
|
| 3154 | switch (action) {
|
| 3155 | case PM_HIBERNATION_PREPARE:
|
| 3156 | case PM_SUSPEND_PREPARE:
|
| 3157 | if (nr_cpu_ids <= 256)
|
| 3158 | rcu_expedited = 1;
|
| 3159 | break;
|
| 3160 | case PM_POST_HIBERNATION:
|
| 3161 | case PM_POST_SUSPEND:
|
| 3162 | rcu_expedited = 0;
|
| 3163 | break;
|
| 3164 | default:
|
| 3165 | break;
|
| 3166 | }
|
| 3167 | return NOTIFY_OK;
|
| 3168 | }
|
| 3169 |
|
| 3170 | |
| 3171 | * Spawn the kthread that handles this RCU flavor's grace periods.
|
| 3172 |
|
| 3173 | static int __init rcu_spawn_gp_kthread(void)
|
| 3174 | {
|
| 3175 | unsigned long flags;
|
| 3176 | struct rcu_node *rnp;
|
| 3177 | struct rcu_state *rsp;
|
| 3178 | struct task_struct *t;
|
| 3179 |
|
| 3180 | for_each_rcu_flavor(rsp) {
|
| 3181 | t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
|
| 3182 | BUG_ON(IS_ERR(t));
|
| 3183 | rnp = rcu_get_root(rsp);
|
| 3184 | raw_spin_lock_irqsave(&rnp->lock, flags);
|
| 3185 | rsp->gp_kthread = t;
|
| 3186 | raw_spin_unlock_irqrestore(&rnp->lock, flags);
|
| 3187 | rcu_spawn_nocb_kthreads(rsp);
|
| 3188 | }
|
| 3189 | return 0;
|
| 3190 | }
|
| 3191 | early_initcall(rcu_spawn_gp_kthread);
|
| 3192 |
|
| 3193 | |
| 3194 | * This function is invoked towards the end of the scheduler's initialization
|
| 3195 | * process. Before this is called, the idle task might contain
|
| 3196 | * RCU read-side critical sections (during which time, this idle
|
| 3197 | * task is booting the system). After this function is called, the
|
| 3198 | * idle tasks are prohibited from containing RCU read-side critical
|
| 3199 | * sections. This function also enables RCU lockdep checking.
|
| 3200 |
|
| 3201 | void rcu_scheduler_starting(void)
|
| 3202 | {
|
| 3203 | WARN_ON(num_online_cpus() != 1);
|
| 3204 | WARN_ON(nr_context_switches() > 0);
|
| 3205 | rcu_scheduler_active = 1;
|
| 3206 | }
|
| 3207 |
|
| 3208 | |
| 3209 | * Compute the per-level fanout, either using the exact fanout specified
|
| 3210 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
|
| 3211 |
|
| 3212 | #ifdef CONFIG_RCU_FANOUT_EXACT
|
| 3213 | static void __init rcu_init_levelspread(struct rcu_state *rsp)
|
| 3214 | {
|
| 3215 | int i;
|
| 3216 |
|
| 3217 | for (i = rcu_num_lvls - 1; i > 0; i--)
|
| 3218 | rsp->levelspread[i] = CONFIG_RCU_FANOUT;
|
| 3219 | rsp->levelspread[0] = rcu_fanout_leaf;
|
| 3220 | }
|
| 3221 | #else
|
| 3222 | static void __init rcu_init_levelspread(struct rcu_state *rsp)
|
| 3223 | {
|
| 3224 | int ccur;
|
| 3225 | int cprv;
|
| 3226 | int i;
|
| 3227 |
|
| 3228 | cprv = nr_cpu_ids;
|
| 3229 | for (i = rcu_num_lvls - 1; i >= 0; i--) {
|
| 3230 | ccur = rsp->levelcnt[i];
|
| 3231 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
|
| 3232 | cprv = ccur;
|
| 3233 | }
|
| 3234 | }
|
| 3235 | #endif
|
| 3236 |
|
| 3237 | |
| 3238 | * Helper function for rcu_init() that initializes one rcu_state structure.
|
| 3239 |
|
| 3240 | static void __init rcu_init_one(struct rcu_state *rsp,
|
| 3241 | struct rcu_data __percpu *rda)
|
| 3242 | {
|
| 3243 | static char *buf[] = { "rcu_node_0",
|
| 3244 | "rcu_node_1",
|
| 3245 | "rcu_node_2",
|
| 3246 | "rcu_node_3" };
|
| 3247 | static char *fqs[] = { "rcu_node_fqs_0",
|
| 3248 | "rcu_node_fqs_1",
|
| 3249 | "rcu_node_fqs_2",
|
| 3250 | "rcu_node_fqs_3" };
|
| 3251 | int cpustride = 1;
|
| 3252 | int i;
|
| 3253 | int j;
|
| 3254 | struct rcu_node *rnp;
|
| 3255 |
|
| 3256 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));
|
| 3257 |
|
| 3258 |
|
| 3259 | if (rcu_num_lvls > RCU_NUM_LVLS)
|
| 3260 | panic("rcu_init_one: rcu_num_lvls overflow");
|
| 3261 |
|
| 3262 |
|
| 3263 |
|
| 3264 | for (i = 0; i < rcu_num_lvls; i++)
|
| 3265 | rsp->levelcnt[i] = num_rcu_lvl[i];
|
| 3266 | for (i = 1; i < rcu_num_lvls; i++)
|
| 3267 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
|
| 3268 | rcu_init_levelspread(rsp);
|
| 3269 |
|
| 3270 |
|
| 3271 |
|
| 3272 | for (i = rcu_num_lvls - 1; i >= 0; i--) {
|
| 3273 | cpustride *= rsp->levelspread[i];
|
| 3274 | rnp = rsp->level[i];
|
| 3275 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
|
| 3276 | raw_spin_lock_init(&rnp->lock);
|
| 3277 | lockdep_set_class_and_name(&rnp->lock,
|
| 3278 | &rcu_node_class[i], buf[i]);
|
| 3279 | raw_spin_lock_init(&rnp->fqslock);
|
| 3280 | lockdep_set_class_and_name(&rnp->fqslock,
|
| 3281 | &rcu_fqs_class[i], fqs[i]);
|
| 3282 | rnp->gpnum = rsp->gpnum;
|
| 3283 | rnp->completed = rsp->completed;
|
| 3284 | rnp->qsmask = 0;
|
| 3285 | rnp->qsmaskinit = 0;
|
| 3286 | rnp->grplo = j * cpustride;
|
| 3287 | rnp->grphi = (j + 1) * cpustride - 1;
|
| 3288 | if (rnp->grphi >= NR_CPUS)
|
| 3289 | rnp->grphi = NR_CPUS - 1;
|
| 3290 | if (i == 0) {
|
| 3291 | rnp->grpnum = 0;
|
| 3292 | rnp->grpmask = 0;
|
| 3293 | rnp->parent = NULL;
|
| 3294 | } else {
|
| 3295 | rnp->grpnum = j % rsp->levelspread[i - 1];
|
| 3296 | rnp->grpmask = 1UL << rnp->grpnum;
|
| 3297 | rnp->parent = rsp->level[i - 1] +
|
| 3298 | j / rsp->levelspread[i - 1];
|
| 3299 | }
|
| 3300 | rnp->level = i;
|
| 3301 | INIT_LIST_HEAD(&rnp->blkd_tasks);
|
| 3302 | rcu_init_one_nocb(rnp);
|
| 3303 | }
|
| 3304 | }
|
| 3305 |
|
| 3306 | rsp->rda = rda;
|
| 3307 | init_waitqueue_head(&rsp->gp_wq);
|
| 3308 | init_irq_work(&rsp->wakeup_work, rsp_wakeup);
|
| 3309 | rnp = rsp->level[rcu_num_lvls - 1];
|
| 3310 | for_each_possible_cpu(i) {
|
| 3311 | while (i > rnp->grphi)
|
| 3312 | rnp++;
|
| 3313 | per_cpu_ptr(rsp->rda, i)->mynode = rnp;
|
| 3314 | rcu_boot_init_percpu_data(i, rsp);
|
| 3315 | }
|
| 3316 | list_add(&rsp->flavors, &rcu_struct_flavors);
|
| 3317 | }
|
| 3318 |
|
| 3319 | |
| 3320 | * Compute the rcu_node tree geometry from kernel parameters. This cannot
|
| 3321 | * replace the definitions in tree.h because those are needed to size
|
| 3322 | * the ->node array in the rcu_state structure.
|
| 3323 |
|
| 3324 | static void __init rcu_init_geometry(void)
|
| 3325 | {
|
| 3326 | ulong d;
|
| 3327 | int i;
|
| 3328 | int j;
|
| 3329 | int n = nr_cpu_ids;
|
| 3330 | int rcu_capacity[MAX_RCU_LVLS + 1];
|
| 3331 |
|
| 3332 | |
| 3333 | * Initialize any unspecified boot parameters.
|
| 3334 | * The default values of jiffies_till_first_fqs and
|
| 3335 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
|
| 3336 | * value, which is a function of HZ, then adding one for each
|
| 3337 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
|
| 3338 |
|
| 3339 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
|
| 3340 | if (jiffies_till_first_fqs == ULONG_MAX)
|
| 3341 | jiffies_till_first_fqs = d;
|
| 3342 | if (jiffies_till_next_fqs == ULONG_MAX)
|
| 3343 | jiffies_till_next_fqs = d;
|
| 3344 |
|
| 3345 |
|
| 3346 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
|
| 3347 | nr_cpu_ids == NR_CPUS)
|
| 3348 | return;
|
| 3349 |
|
| 3350 | |
| 3351 | * Compute number of nodes that can be handled an rcu_node tree
|
| 3352 | * with the given number of levels. Setting rcu_capacity[0] makes
|
| 3353 | * some of the arithmetic easier.
|
| 3354 |
|
| 3355 | rcu_capacity[0] = 1;
|
| 3356 | rcu_capacity[1] = rcu_fanout_leaf;
|
| 3357 | for (i = 2; i <= MAX_RCU_LVLS; i++)
|
| 3358 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
|
| 3359 |
|
| 3360 | |
| 3361 | * The boot-time rcu_fanout_leaf parameter is only permitted
|
| 3362 | * to increase the leaf-level fanout, not decrease it. Of course,
|
| 3363 | * the leaf-level fanout cannot exceed the number of bits in
|
| 3364 | * the rcu_node masks. Finally, the tree must be able to accommodate
|
| 3365 | * the configured number of CPUs. Complain and fall back to the
|
| 3366 | * compile-time values if these limits are exceeded.
|
| 3367 |
|
| 3368 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
|
| 3369 | rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
|
| 3370 | n > rcu_capacity[MAX_RCU_LVLS]) {
|
| 3371 | WARN_ON(1);
|
| 3372 | return;
|
| 3373 | }
|
| 3374 |
|
| 3375 |
|
| 3376 | for (i = 1; i <= MAX_RCU_LVLS; i++)
|
| 3377 | if (n <= rcu_capacity[i]) {
|
| 3378 | for (j = 0; j <= i; j++)
|
| 3379 | num_rcu_lvl[j] =
|
| 3380 | DIV_ROUND_UP(n, rcu_capacity[i - j]);
|
| 3381 | rcu_num_lvls = i;
|
| 3382 | for (j = i + 1; j <= MAX_RCU_LVLS; j++)
|
| 3383 | num_rcu_lvl[j] = 0;
|
| 3384 | break;
|
| 3385 | }
|
| 3386 |
|
| 3387 |
|
| 3388 | rcu_num_nodes = 0;
|
| 3389 | for (i = 0; i <= MAX_RCU_LVLS; i++)
|
| 3390 | rcu_num_nodes += num_rcu_lvl[i];
|
| 3391 | rcu_num_nodes -= n;
|
| 3392 | }
|
| 3393 |
|
| 3394 | void __init rcu_init(void)
|
| 3395 | {
|
| 3396 | int cpu;
|
| 3397 |
|
| 3398 | rcu_bootup_announce();
|
| 3399 | rcu_init_geometry();
|
| 3400 | rcu_init_one(&rcu_bh_state, &rcu_bh_data);
|
| 3401 | rcu_init_one(&rcu_sched_state, &rcu_sched_data);
|
| 3402 | __rcu_init_preempt();
|
| 3403 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
|
| 3404 |
|
| 3405 | |
| 3406 | * We don't need protection against CPU-hotplug here because
|
| 3407 | * this is called early in boot, before either interrupts
|
| 3408 | * or the scheduler are operational.
|
| 3409 |
|
| 3410 | cpu_notifier(rcu_cpu_notify, 0);
|
| 3411 | pm_notifier(rcu_pm_notify, 0);
|
| 3412 | for_each_online_cpu(cpu)
|
| 3413 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
|
| 3414 | }
|
| 3415 |
|
| 3416 | #include "tree_plugin.h"
|