| 1 | /* |
| 2 | * Read-Copy Update mechanism for mutual exclusion |
| 3 | * |
| 4 | * This program is free software; you can redistribute it and/or modify |
| 5 | * it under the terms of the GNU General Public License as published by |
| 6 | * the Free Software Foundation; either version 2 of the License, or |
| 7 | * (at your option) any later version. |
| 8 | * |
| 9 | * This program is distributed in the hope that it will be useful, |
| 10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 12 | * GNU General Public License for more details. |
| 13 | * |
| 14 | * You should have received a copy of the GNU General Public License |
| 15 | * along with this program; if not, you can access it online at |
| 16 | * http://www.gnu.org/licenses/gpl-2.0.html. |
| 17 | * |
| 18 | * Copyright IBM Corporation, 2008 |
| 19 | * |
| 20 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> |
| 21 | * Manfred Spraul <manfred@colorfullife.com> |
| 22 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version |
| 23 | * |
| 24 | * Based on the original work by Paul McKenney <paulmck@us.ibm.com> |
| 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 "rcu.h" |
| 62 | |
| 63 | MODULE_ALIAS("rcutree"); |
| 64 | #ifdef MODULE_PARAM_PREFIX |
| 65 | #undef MODULE_PARAM_PREFIX |
| 66 | #endif |
| 67 | #define MODULE_PARAM_PREFIX "rcutree." |
| 68 | |
| 69 | /* Data structures. */ |
| 70 | |
| 71 | static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; |
| 72 | static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; |
| 73 | |
| 74 | /* |
| 75 | * In order to export the rcu_state name to the tracing tools, it |
| 76 | * needs to be added in the __tracepoint_string section. |
| 77 | * This requires defining a separate variable tp_<sname>_varname |
| 78 | * that points to the string being used, and this will allow |
| 79 | * the tracing userspace tools to be able to decipher the string |
| 80 | * address to the matching string. |
| 81 | */ |
| 82 | #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ |
| 83 | static char sname##_varname[] = #sname; \ |
| 84 | static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \ |
| 85 | struct rcu_state sname##_state = { \ |
| 86 | .level = { &sname##_state.node[0] }, \ |
| 87 | .call = cr, \ |
| 88 | .fqs_state = RCU_GP_IDLE, \ |
| 89 | .gpnum = 0UL - 300UL, \ |
| 90 | .completed = 0UL - 300UL, \ |
| 91 | .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ |
| 92 | .orphan_nxttail = &sname##_state.orphan_nxtlist, \ |
| 93 | .orphan_donetail = &sname##_state.orphan_donelist, \ |
| 94 | .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ |
| 95 | .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \ |
| 96 | .name = sname##_varname, \ |
| 97 | .abbr = sabbr, \ |
| 98 | }; \ |
| 99 | DEFINE_PER_CPU(struct rcu_data, sname##_data) |
| 100 | |
| 101 | RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); |
| 102 | RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); |
| 103 | |
| 104 | static struct rcu_state *rcu_state; |
| 105 | LIST_HEAD(rcu_struct_flavors); |
| 106 | |
| 107 | /* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */ |
| 108 | static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF; |
| 109 | module_param(rcu_fanout_leaf, int, 0444); |
| 110 | int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; |
| 111 | static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */ |
| 112 | NUM_RCU_LVL_0, |
| 113 | NUM_RCU_LVL_1, |
| 114 | NUM_RCU_LVL_2, |
| 115 | NUM_RCU_LVL_3, |
| 116 | NUM_RCU_LVL_4, |
| 117 | }; |
| 118 | int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ |
| 119 | |
| 120 | /* |
| 121 | * The rcu_scheduler_active variable transitions from zero to one just |
| 122 | * before the first task is spawned. So when this variable is zero, RCU |
| 123 | * can assume that there is but one task, allowing RCU to (for example) |
| 124 | * optimize synchronize_sched() to a simple barrier(). When this variable |
| 125 | * is one, RCU must actually do all the hard work required to detect real |
| 126 | * grace periods. This variable is also used to suppress boot-time false |
| 127 | * positives from lockdep-RCU error checking. |
| 128 | */ |
| 129 | int rcu_scheduler_active __read_mostly; |
| 130 | EXPORT_SYMBOL_GPL(rcu_scheduler_active); |
| 131 | |
| 132 | /* |
| 133 | * The rcu_scheduler_fully_active variable transitions from zero to one |
| 134 | * during the early_initcall() processing, which is after the scheduler |
| 135 | * is capable of creating new tasks. So RCU processing (for example, |
| 136 | * creating tasks for RCU priority boosting) must be delayed until after |
| 137 | * rcu_scheduler_fully_active transitions from zero to one. We also |
| 138 | * currently delay invocation of any RCU callbacks until after this point. |
| 139 | * |
| 140 | * It might later prove better for people registering RCU callbacks during |
| 141 | * early boot to take responsibility for these callbacks, but one step at |
| 142 | * a time. |
| 143 | */ |
| 144 | static int rcu_scheduler_fully_active __read_mostly; |
| 145 | |
| 146 | #ifdef CONFIG_RCU_BOOST |
| 147 | |
| 148 | /* |
| 149 | * Control variables for per-CPU and per-rcu_node kthreads. These |
| 150 | * handle all flavors of RCU. |
| 151 | */ |
| 152 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); |
| 153 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); |
| 154 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); |
| 155 | DEFINE_PER_CPU(char, rcu_cpu_has_work); |
| 156 | |
| 157 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 158 | |
| 159 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); |
| 160 | static void invoke_rcu_core(void); |
| 161 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); |
| 162 | |
| 163 | /* |
| 164 | * Track the rcutorture test sequence number and the update version |
| 165 | * number within a given test. The rcutorture_testseq is incremented |
| 166 | * on every rcutorture module load and unload, so has an odd value |
| 167 | * when a test is running. The rcutorture_vernum is set to zero |
| 168 | * when rcutorture starts and is incremented on each rcutorture update. |
| 169 | * These variables enable correlating rcutorture output with the |
| 170 | * RCU tracing information. |
| 171 | */ |
| 172 | unsigned long rcutorture_testseq; |
| 173 | unsigned long rcutorture_vernum; |
| 174 | |
| 175 | /* |
| 176 | * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s |
| 177 | * permit this function to be invoked without holding the root rcu_node |
| 178 | * structure's ->lock, but of course results can be subject to change. |
| 179 | */ |
| 180 | static int rcu_gp_in_progress(struct rcu_state *rsp) |
| 181 | { |
| 182 | return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum); |
| 183 | } |
| 184 | |
| 185 | /* |
| 186 | * Note a quiescent state. Because we do not need to know |
| 187 | * how many quiescent states passed, just if there was at least |
| 188 | * one since the start of the grace period, this just sets a flag. |
| 189 | * The caller must have disabled preemption. |
| 190 | */ |
| 191 | void rcu_sched_qs(int cpu) |
| 192 | { |
| 193 | struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu); |
| 194 | |
| 195 | if (rdp->passed_quiesce == 0) |
| 196 | trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs")); |
| 197 | rdp->passed_quiesce = 1; |
| 198 | } |
| 199 | |
| 200 | void rcu_bh_qs(int cpu) |
| 201 | { |
| 202 | struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu); |
| 203 | |
| 204 | if (rdp->passed_quiesce == 0) |
| 205 | trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs")); |
| 206 | rdp->passed_quiesce = 1; |
| 207 | } |
| 208 | |
| 209 | /* |
| 210 | * Note a context switch. This is a quiescent state for RCU-sched, |
| 211 | * and requires special handling for preemptible RCU. |
| 212 | * The caller must have disabled preemption. |
| 213 | */ |
| 214 | void rcu_note_context_switch(int cpu) |
| 215 | { |
| 216 | trace_rcu_utilization(TPS("Start context switch")); |
| 217 | rcu_sched_qs(cpu); |
| 218 | rcu_preempt_note_context_switch(cpu); |
| 219 | trace_rcu_utilization(TPS("End context switch")); |
| 220 | } |
| 221 | EXPORT_SYMBOL_GPL(rcu_note_context_switch); |
| 222 | |
| 223 | static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { |
| 224 | .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, |
| 225 | .dynticks = ATOMIC_INIT(1), |
| 226 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
| 227 | .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, |
| 228 | .dynticks_idle = ATOMIC_INIT(1), |
| 229 | #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
| 230 | }; |
| 231 | |
| 232 | static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ |
| 233 | static long qhimark = 10000; /* If this many pending, ignore blimit. */ |
| 234 | static long qlowmark = 100; /* Once only this many pending, use blimit. */ |
| 235 | |
| 236 | module_param(blimit, long, 0444); |
| 237 | module_param(qhimark, long, 0444); |
| 238 | module_param(qlowmark, long, 0444); |
| 239 | |
| 240 | static ulong jiffies_till_first_fqs = ULONG_MAX; |
| 241 | static ulong jiffies_till_next_fqs = ULONG_MAX; |
| 242 | |
| 243 | module_param(jiffies_till_first_fqs, ulong, 0644); |
| 244 | module_param(jiffies_till_next_fqs, ulong, 0644); |
| 245 | |
| 246 | static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
| 247 | struct rcu_data *rdp); |
| 248 | static void force_qs_rnp(struct rcu_state *rsp, |
| 249 | int (*f)(struct rcu_data *rsp, bool *isidle, |
| 250 | unsigned long *maxj), |
| 251 | bool *isidle, unsigned long *maxj); |
| 252 | static void force_quiescent_state(struct rcu_state *rsp); |
| 253 | static int rcu_pending(int cpu); |
| 254 | |
| 255 | /* |
| 256 | * Return the number of RCU-sched batches processed thus far for debug & stats. |
| 257 | */ |
| 258 | long rcu_batches_completed_sched(void) |
| 259 | { |
| 260 | return rcu_sched_state.completed; |
| 261 | } |
| 262 | EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); |
| 263 | |
| 264 | /* |
| 265 | * Return the number of RCU BH batches processed thus far for debug & stats. |
| 266 | */ |
| 267 | long rcu_batches_completed_bh(void) |
| 268 | { |
| 269 | return rcu_bh_state.completed; |
| 270 | } |
| 271 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); |
| 272 | |
| 273 | /* |
| 274 | * Force a quiescent state for RCU BH. |
| 275 | */ |
| 276 | void rcu_bh_force_quiescent_state(void) |
| 277 | { |
| 278 | force_quiescent_state(&rcu_bh_state); |
| 279 | } |
| 280 | EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); |
| 281 | |
| 282 | /* |
| 283 | * Record the number of times rcutorture tests have been initiated and |
| 284 | * terminated. This information allows the debugfs tracing stats to be |
| 285 | * correlated to the rcutorture messages, even when the rcutorture module |
| 286 | * is being repeatedly loaded and unloaded. In other words, we cannot |
| 287 | * store this state in rcutorture itself. |
| 288 | */ |
| 289 | void rcutorture_record_test_transition(void) |
| 290 | { |
| 291 | rcutorture_testseq++; |
| 292 | rcutorture_vernum = 0; |
| 293 | } |
| 294 | EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); |
| 295 | |
| 296 | /* |
| 297 | * Record the number of writer passes through the current rcutorture test. |
| 298 | * This is also used to correlate debugfs tracing stats with the rcutorture |
| 299 | * messages. |
| 300 | */ |
| 301 | void rcutorture_record_progress(unsigned long vernum) |
| 302 | { |
| 303 | rcutorture_vernum++; |
| 304 | } |
| 305 | EXPORT_SYMBOL_GPL(rcutorture_record_progress); |
| 306 | |
| 307 | /* |
| 308 | * Force a quiescent state for RCU-sched. |
| 309 | */ |
| 310 | void rcu_sched_force_quiescent_state(void) |
| 311 | { |
| 312 | force_quiescent_state(&rcu_sched_state); |
| 313 | } |
| 314 | EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); |
| 315 | |
| 316 | /* |
| 317 | * Does the CPU have callbacks ready to be invoked? |
| 318 | */ |
| 319 | static int |
| 320 | cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) |
| 321 | { |
| 322 | return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && |
| 323 | rdp->nxttail[RCU_DONE_TAIL] != NULL; |
| 324 | } |
| 325 | |
| 326 | /* |
| 327 | * Return the root node of the specified rcu_state structure. |
| 328 | */ |
| 329 | static struct rcu_node *rcu_get_root(struct rcu_state *rsp) |
| 330 | { |
| 331 | return &rsp->node[0]; |
| 332 | } |
| 333 | |
| 334 | /* |
| 335 | * Is there any need for future grace periods? |
| 336 | * Interrupts must be disabled. If the caller does not hold the root |
| 337 | * rnp_node structure's ->lock, the results are advisory only. |
| 338 | */ |
| 339 | static int rcu_future_needs_gp(struct rcu_state *rsp) |
| 340 | { |
| 341 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 342 | int idx = (ACCESS_ONCE(rnp->completed) + 1) & 0x1; |
| 343 | int *fp = &rnp->need_future_gp[idx]; |
| 344 | |
| 345 | return ACCESS_ONCE(*fp); |
| 346 | } |
| 347 | |
| 348 | /* |
| 349 | * Does the current CPU require a not-yet-started grace period? |
| 350 | * The caller must have disabled interrupts to prevent races with |
| 351 | * normal callback registry. |
| 352 | */ |
| 353 | static int |
| 354 | cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) |
| 355 | { |
| 356 | int i; |
| 357 | |
| 358 | if (rcu_gp_in_progress(rsp)) |
| 359 | return 0; /* No, a grace period is already in progress. */ |
| 360 | if (rcu_future_needs_gp(rsp)) |
| 361 | return 1; /* Yes, a no-CBs CPU needs one. */ |
| 362 | if (!rdp->nxttail[RCU_NEXT_TAIL]) |
| 363 | return 0; /* No, this is a no-CBs (or offline) CPU. */ |
| 364 | if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) |
| 365 | return 1; /* Yes, this CPU has newly registered callbacks. */ |
| 366 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) |
| 367 | if (rdp->nxttail[i - 1] != rdp->nxttail[i] && |
| 368 | ULONG_CMP_LT(ACCESS_ONCE(rsp->completed), |
| 369 | rdp->nxtcompleted[i])) |
| 370 | return 1; /* Yes, CBs for future grace period. */ |
| 371 | return 0; /* No grace period needed. */ |
| 372 | } |
| 373 | |
| 374 | /* |
| 375 | * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state |
| 376 | * |
| 377 | * If the new value of the ->dynticks_nesting counter now is zero, |
| 378 | * we really have entered idle, and must do the appropriate accounting. |
| 379 | * The caller must have disabled interrupts. |
| 380 | */ |
| 381 | static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval, |
| 382 | bool user) |
| 383 | { |
| 384 | struct rcu_state *rsp; |
| 385 | struct rcu_data *rdp; |
| 386 | |
| 387 | trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); |
| 388 | if (!user && !is_idle_task(current)) { |
| 389 | struct task_struct *idle __maybe_unused = |
| 390 | idle_task(smp_processor_id()); |
| 391 | |
| 392 | trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); |
| 393 | ftrace_dump(DUMP_ORIG); |
| 394 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
| 395 | current->pid, current->comm, |
| 396 | idle->pid, idle->comm); /* must be idle task! */ |
| 397 | } |
| 398 | for_each_rcu_flavor(rsp) { |
| 399 | rdp = this_cpu_ptr(rsp->rda); |
| 400 | do_nocb_deferred_wakeup(rdp); |
| 401 | } |
| 402 | rcu_prepare_for_idle(smp_processor_id()); |
| 403 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
| 404 | smp_mb__before_atomic_inc(); /* See above. */ |
| 405 | atomic_inc(&rdtp->dynticks); |
| 406 | smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */ |
| 407 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
| 408 | |
| 409 | /* |
| 410 | * It is illegal to enter an extended quiescent state while |
| 411 | * in an RCU read-side critical section. |
| 412 | */ |
| 413 | rcu_lockdep_assert(!lock_is_held(&rcu_lock_map), |
| 414 | "Illegal idle entry in RCU read-side critical section."); |
| 415 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map), |
| 416 | "Illegal idle entry in RCU-bh read-side critical section."); |
| 417 | rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map), |
| 418 | "Illegal idle entry in RCU-sched read-side critical section."); |
| 419 | } |
| 420 | |
| 421 | /* |
| 422 | * Enter an RCU extended quiescent state, which can be either the |
| 423 | * idle loop or adaptive-tickless usermode execution. |
| 424 | */ |
| 425 | static void rcu_eqs_enter(bool user) |
| 426 | { |
| 427 | long long oldval; |
| 428 | struct rcu_dynticks *rdtp; |
| 429 | |
| 430 | rdtp = this_cpu_ptr(&rcu_dynticks); |
| 431 | oldval = rdtp->dynticks_nesting; |
| 432 | WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0); |
| 433 | if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { |
| 434 | rdtp->dynticks_nesting = 0; |
| 435 | rcu_eqs_enter_common(rdtp, oldval, user); |
| 436 | } else { |
| 437 | rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; |
| 438 | } |
| 439 | } |
| 440 | |
| 441 | /** |
| 442 | * rcu_idle_enter - inform RCU that current CPU is entering idle |
| 443 | * |
| 444 | * Enter idle mode, in other words, -leave- the mode in which RCU |
| 445 | * read-side critical sections can occur. (Though RCU read-side |
| 446 | * critical sections can occur in irq handlers in idle, a possibility |
| 447 | * handled by irq_enter() and irq_exit().) |
| 448 | * |
| 449 | * We crowbar the ->dynticks_nesting field to zero to allow for |
| 450 | * the possibility of usermode upcalls having messed up our count |
| 451 | * of interrupt nesting level during the prior busy period. |
| 452 | */ |
| 453 | void rcu_idle_enter(void) |
| 454 | { |
| 455 | unsigned long flags; |
| 456 | |
| 457 | local_irq_save(flags); |
| 458 | rcu_eqs_enter(false); |
| 459 | rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0); |
| 460 | local_irq_restore(flags); |
| 461 | } |
| 462 | EXPORT_SYMBOL_GPL(rcu_idle_enter); |
| 463 | |
| 464 | #ifdef CONFIG_RCU_USER_QS |
| 465 | /** |
| 466 | * rcu_user_enter - inform RCU that we are resuming userspace. |
| 467 | * |
| 468 | * Enter RCU idle mode right before resuming userspace. No use of RCU |
| 469 | * is permitted between this call and rcu_user_exit(). This way the |
| 470 | * CPU doesn't need to maintain the tick for RCU maintenance purposes |
| 471 | * when the CPU runs in userspace. |
| 472 | */ |
| 473 | void rcu_user_enter(void) |
| 474 | { |
| 475 | rcu_eqs_enter(1); |
| 476 | } |
| 477 | #endif /* CONFIG_RCU_USER_QS */ |
| 478 | |
| 479 | /** |
| 480 | * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle |
| 481 | * |
| 482 | * Exit from an interrupt handler, which might possibly result in entering |
| 483 | * idle mode, in other words, leaving the mode in which read-side critical |
| 484 | * sections can occur. |
| 485 | * |
| 486 | * This code assumes that the idle loop never does anything that might |
| 487 | * result in unbalanced calls to irq_enter() and irq_exit(). If your |
| 488 | * architecture violates this assumption, RCU will give you what you |
| 489 | * deserve, good and hard. But very infrequently and irreproducibly. |
| 490 | * |
| 491 | * Use things like work queues to work around this limitation. |
| 492 | * |
| 493 | * You have been warned. |
| 494 | */ |
| 495 | void rcu_irq_exit(void) |
| 496 | { |
| 497 | unsigned long flags; |
| 498 | long long oldval; |
| 499 | struct rcu_dynticks *rdtp; |
| 500 | |
| 501 | local_irq_save(flags); |
| 502 | rdtp = this_cpu_ptr(&rcu_dynticks); |
| 503 | oldval = rdtp->dynticks_nesting; |
| 504 | rdtp->dynticks_nesting--; |
| 505 | WARN_ON_ONCE(rdtp->dynticks_nesting < 0); |
| 506 | if (rdtp->dynticks_nesting) |
| 507 | trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); |
| 508 | else |
| 509 | rcu_eqs_enter_common(rdtp, oldval, true); |
| 510 | rcu_sysidle_enter(rdtp, 1); |
| 511 | local_irq_restore(flags); |
| 512 | } |
| 513 | |
| 514 | /* |
| 515 | * rcu_eqs_exit_common - current CPU moving away from extended quiescent state |
| 516 | * |
| 517 | * If the new value of the ->dynticks_nesting counter was previously zero, |
| 518 | * we really have exited idle, and must do the appropriate accounting. |
| 519 | * The caller must have disabled interrupts. |
| 520 | */ |
| 521 | static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval, |
| 522 | int user) |
| 523 | { |
| 524 | smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */ |
| 525 | atomic_inc(&rdtp->dynticks); |
| 526 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ |
| 527 | smp_mb__after_atomic_inc(); /* See above. */ |
| 528 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); |
| 529 | rcu_cleanup_after_idle(smp_processor_id()); |
| 530 | trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); |
| 531 | if (!user && !is_idle_task(current)) { |
| 532 | struct task_struct *idle __maybe_unused = |
| 533 | idle_task(smp_processor_id()); |
| 534 | |
| 535 | trace_rcu_dyntick(TPS("Error on exit: not idle task"), |
| 536 | oldval, rdtp->dynticks_nesting); |
| 537 | ftrace_dump(DUMP_ORIG); |
| 538 | WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", |
| 539 | current->pid, current->comm, |
| 540 | idle->pid, idle->comm); /* must be idle task! */ |
| 541 | } |
| 542 | } |
| 543 | |
| 544 | /* |
| 545 | * Exit an RCU extended quiescent state, which can be either the |
| 546 | * idle loop or adaptive-tickless usermode execution. |
| 547 | */ |
| 548 | static void rcu_eqs_exit(bool user) |
| 549 | { |
| 550 | struct rcu_dynticks *rdtp; |
| 551 | long long oldval; |
| 552 | |
| 553 | rdtp = this_cpu_ptr(&rcu_dynticks); |
| 554 | oldval = rdtp->dynticks_nesting; |
| 555 | WARN_ON_ONCE(oldval < 0); |
| 556 | if (oldval & DYNTICK_TASK_NEST_MASK) { |
| 557 | rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; |
| 558 | } else { |
| 559 | rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
| 560 | rcu_eqs_exit_common(rdtp, oldval, user); |
| 561 | } |
| 562 | } |
| 563 | |
| 564 | /** |
| 565 | * rcu_idle_exit - inform RCU that current CPU is leaving idle |
| 566 | * |
| 567 | * Exit idle mode, in other words, -enter- the mode in which RCU |
| 568 | * read-side critical sections can occur. |
| 569 | * |
| 570 | * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to |
| 571 | * allow for the possibility of usermode upcalls messing up our count |
| 572 | * of interrupt nesting level during the busy period that is just |
| 573 | * now starting. |
| 574 | */ |
| 575 | void rcu_idle_exit(void) |
| 576 | { |
| 577 | unsigned long flags; |
| 578 | |
| 579 | local_irq_save(flags); |
| 580 | rcu_eqs_exit(false); |
| 581 | rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0); |
| 582 | local_irq_restore(flags); |
| 583 | } |
| 584 | EXPORT_SYMBOL_GPL(rcu_idle_exit); |
| 585 | |
| 586 | #ifdef CONFIG_RCU_USER_QS |
| 587 | /** |
| 588 | * rcu_user_exit - inform RCU that we are exiting userspace. |
| 589 | * |
| 590 | * Exit RCU idle mode while entering the kernel because it can |
| 591 | * run a RCU read side critical section anytime. |
| 592 | */ |
| 593 | void rcu_user_exit(void) |
| 594 | { |
| 595 | rcu_eqs_exit(1); |
| 596 | } |
| 597 | #endif /* CONFIG_RCU_USER_QS */ |
| 598 | |
| 599 | /** |
| 600 | * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle |
| 601 | * |
| 602 | * Enter an interrupt handler, which might possibly result in exiting |
| 603 | * idle mode, in other words, entering the mode in which read-side critical |
| 604 | * sections can occur. |
| 605 | * |
| 606 | * Note that the Linux kernel is fully capable of entering an interrupt |
| 607 | * handler that it never exits, for example when doing upcalls to |
| 608 | * user mode! This code assumes that the idle loop never does upcalls to |
| 609 | * user mode. If your architecture does do upcalls from the idle loop (or |
| 610 | * does anything else that results in unbalanced calls to the irq_enter() |
| 611 | * and irq_exit() functions), RCU will give you what you deserve, good |
| 612 | * and hard. But very infrequently and irreproducibly. |
| 613 | * |
| 614 | * Use things like work queues to work around this limitation. |
| 615 | * |
| 616 | * You have been warned. |
| 617 | */ |
| 618 | void rcu_irq_enter(void) |
| 619 | { |
| 620 | unsigned long flags; |
| 621 | struct rcu_dynticks *rdtp; |
| 622 | long long oldval; |
| 623 | |
| 624 | local_irq_save(flags); |
| 625 | rdtp = this_cpu_ptr(&rcu_dynticks); |
| 626 | oldval = rdtp->dynticks_nesting; |
| 627 | rdtp->dynticks_nesting++; |
| 628 | WARN_ON_ONCE(rdtp->dynticks_nesting == 0); |
| 629 | if (oldval) |
| 630 | trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); |
| 631 | else |
| 632 | rcu_eqs_exit_common(rdtp, oldval, true); |
| 633 | rcu_sysidle_exit(rdtp, 1); |
| 634 | local_irq_restore(flags); |
| 635 | } |
| 636 | |
| 637 | /** |
| 638 | * rcu_nmi_enter - inform RCU of entry to NMI context |
| 639 | * |
| 640 | * If the CPU was idle with dynamic ticks active, and there is no |
| 641 | * irq handler running, this updates rdtp->dynticks_nmi to let the |
| 642 | * RCU grace-period handling know that the CPU is active. |
| 643 | */ |
| 644 | void rcu_nmi_enter(void) |
| 645 | { |
| 646 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
| 647 | |
| 648 | if (rdtp->dynticks_nmi_nesting == 0 && |
| 649 | (atomic_read(&rdtp->dynticks) & 0x1)) |
| 650 | return; |
| 651 | rdtp->dynticks_nmi_nesting++; |
| 652 | smp_mb__before_atomic_inc(); /* Force delay from prior write. */ |
| 653 | atomic_inc(&rdtp->dynticks); |
| 654 | /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */ |
| 655 | smp_mb__after_atomic_inc(); /* See above. */ |
| 656 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1)); |
| 657 | } |
| 658 | |
| 659 | /** |
| 660 | * rcu_nmi_exit - inform RCU of exit from NMI context |
| 661 | * |
| 662 | * If the CPU was idle with dynamic ticks active, and there is no |
| 663 | * irq handler running, this updates rdtp->dynticks_nmi to let the |
| 664 | * RCU grace-period handling know that the CPU is no longer active. |
| 665 | */ |
| 666 | void rcu_nmi_exit(void) |
| 667 | { |
| 668 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
| 669 | |
| 670 | if (rdtp->dynticks_nmi_nesting == 0 || |
| 671 | --rdtp->dynticks_nmi_nesting != 0) |
| 672 | return; |
| 673 | /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */ |
| 674 | smp_mb__before_atomic_inc(); /* See above. */ |
| 675 | atomic_inc(&rdtp->dynticks); |
| 676 | smp_mb__after_atomic_inc(); /* Force delay to next write. */ |
| 677 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1); |
| 678 | } |
| 679 | |
| 680 | /** |
| 681 | * __rcu_is_watching - are RCU read-side critical sections safe? |
| 682 | * |
| 683 | * Return true if RCU is watching the running CPU, which means that |
| 684 | * this CPU can safely enter RCU read-side critical sections. Unlike |
| 685 | * rcu_is_watching(), the caller of __rcu_is_watching() must have at |
| 686 | * least disabled preemption. |
| 687 | */ |
| 688 | bool notrace __rcu_is_watching(void) |
| 689 | { |
| 690 | return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1; |
| 691 | } |
| 692 | |
| 693 | /** |
| 694 | * rcu_is_watching - see if RCU thinks that the current CPU is idle |
| 695 | * |
| 696 | * If the current CPU is in its idle loop and is neither in an interrupt |
| 697 | * or NMI handler, return true. |
| 698 | */ |
| 699 | bool notrace rcu_is_watching(void) |
| 700 | { |
| 701 | int ret; |
| 702 | |
| 703 | preempt_disable(); |
| 704 | ret = __rcu_is_watching(); |
| 705 | preempt_enable(); |
| 706 | return ret; |
| 707 | } |
| 708 | EXPORT_SYMBOL_GPL(rcu_is_watching); |
| 709 | |
| 710 | #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) |
| 711 | |
| 712 | /* |
| 713 | * Is the current CPU online? Disable preemption to avoid false positives |
| 714 | * that could otherwise happen due to the current CPU number being sampled, |
| 715 | * this task being preempted, its old CPU being taken offline, resuming |
| 716 | * on some other CPU, then determining that its old CPU is now offline. |
| 717 | * It is OK to use RCU on an offline processor during initial boot, hence |
| 718 | * the check for rcu_scheduler_fully_active. Note also that it is OK |
| 719 | * for a CPU coming online to use RCU for one jiffy prior to marking itself |
| 720 | * online in the cpu_online_mask. Similarly, it is OK for a CPU going |
| 721 | * offline to continue to use RCU for one jiffy after marking itself |
| 722 | * offline in the cpu_online_mask. This leniency is necessary given the |
| 723 | * non-atomic nature of the online and offline processing, for example, |
| 724 | * the fact that a CPU enters the scheduler after completing the CPU_DYING |
| 725 | * notifiers. |
| 726 | * |
| 727 | * This is also why RCU internally marks CPUs online during the |
| 728 | * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase. |
| 729 | * |
| 730 | * Disable checking if in an NMI handler because we cannot safely report |
| 731 | * errors from NMI handlers anyway. |
| 732 | */ |
| 733 | bool rcu_lockdep_current_cpu_online(void) |
| 734 | { |
| 735 | struct rcu_data *rdp; |
| 736 | struct rcu_node *rnp; |
| 737 | bool ret; |
| 738 | |
| 739 | if (in_nmi()) |
| 740 | return true; |
| 741 | preempt_disable(); |
| 742 | rdp = this_cpu_ptr(&rcu_sched_data); |
| 743 | rnp = rdp->mynode; |
| 744 | ret = (rdp->grpmask & rnp->qsmaskinit) || |
| 745 | !rcu_scheduler_fully_active; |
| 746 | preempt_enable(); |
| 747 | return ret; |
| 748 | } |
| 749 | EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); |
| 750 | |
| 751 | #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ |
| 752 | |
| 753 | /** |
| 754 | * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle |
| 755 | * |
| 756 | * If the current CPU is idle or running at a first-level (not nested) |
| 757 | * interrupt from idle, return true. The caller must have at least |
| 758 | * disabled preemption. |
| 759 | */ |
| 760 | static int rcu_is_cpu_rrupt_from_idle(void) |
| 761 | { |
| 762 | return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; |
| 763 | } |
| 764 | |
| 765 | /* |
| 766 | * Snapshot the specified CPU's dynticks counter so that we can later |
| 767 | * credit them with an implicit quiescent state. Return 1 if this CPU |
| 768 | * is in dynticks idle mode, which is an extended quiescent state. |
| 769 | */ |
| 770 | static int dyntick_save_progress_counter(struct rcu_data *rdp, |
| 771 | bool *isidle, unsigned long *maxj) |
| 772 | { |
| 773 | rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks); |
| 774 | rcu_sysidle_check_cpu(rdp, isidle, maxj); |
| 775 | if ((rdp->dynticks_snap & 0x1) == 0) { |
| 776 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
| 777 | return 1; |
| 778 | } else { |
| 779 | return 0; |
| 780 | } |
| 781 | } |
| 782 | |
| 783 | /* |
| 784 | * This function really isn't for public consumption, but RCU is special in |
| 785 | * that context switches can allow the state machine to make progress. |
| 786 | */ |
| 787 | extern void resched_cpu(int cpu); |
| 788 | |
| 789 | /* |
| 790 | * Return true if the specified CPU has passed through a quiescent |
| 791 | * state by virtue of being in or having passed through an dynticks |
| 792 | * idle state since the last call to dyntick_save_progress_counter() |
| 793 | * for this same CPU, or by virtue of having been offline. |
| 794 | */ |
| 795 | static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, |
| 796 | bool *isidle, unsigned long *maxj) |
| 797 | { |
| 798 | unsigned int curr; |
| 799 | unsigned int snap; |
| 800 | |
| 801 | curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks); |
| 802 | snap = (unsigned int)rdp->dynticks_snap; |
| 803 | |
| 804 | /* |
| 805 | * If the CPU passed through or entered a dynticks idle phase with |
| 806 | * no active irq/NMI handlers, then we can safely pretend that the CPU |
| 807 | * already acknowledged the request to pass through a quiescent |
| 808 | * state. Either way, that CPU cannot possibly be in an RCU |
| 809 | * read-side critical section that started before the beginning |
| 810 | * of the current RCU grace period. |
| 811 | */ |
| 812 | if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) { |
| 813 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); |
| 814 | rdp->dynticks_fqs++; |
| 815 | return 1; |
| 816 | } |
| 817 | |
| 818 | /* |
| 819 | * Check for the CPU being offline, but only if the grace period |
| 820 | * is old enough. We don't need to worry about the CPU changing |
| 821 | * state: If we see it offline even once, it has been through a |
| 822 | * quiescent state. |
| 823 | * |
| 824 | * The reason for insisting that the grace period be at least |
| 825 | * one jiffy old is that CPUs that are not quite online and that |
| 826 | * have just gone offline can still execute RCU read-side critical |
| 827 | * sections. |
| 828 | */ |
| 829 | if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies)) |
| 830 | return 0; /* Grace period is not old enough. */ |
| 831 | barrier(); |
| 832 | if (cpu_is_offline(rdp->cpu)) { |
| 833 | trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); |
| 834 | rdp->offline_fqs++; |
| 835 | return 1; |
| 836 | } |
| 837 | |
| 838 | /* |
| 839 | * There is a possibility that a CPU in adaptive-ticks state |
| 840 | * might run in the kernel with the scheduling-clock tick disabled |
| 841 | * for an extended time period. Invoke rcu_kick_nohz_cpu() to |
| 842 | * force the CPU to restart the scheduling-clock tick in this |
| 843 | * CPU is in this state. |
| 844 | */ |
| 845 | rcu_kick_nohz_cpu(rdp->cpu); |
| 846 | |
| 847 | /* |
| 848 | * Alternatively, the CPU might be running in the kernel |
| 849 | * for an extended period of time without a quiescent state. |
| 850 | * Attempt to force the CPU through the scheduler to gain the |
| 851 | * needed quiescent state, but only if the grace period has gone |
| 852 | * on for an uncommonly long time. If there are many stuck CPUs, |
| 853 | * we will beat on the first one until it gets unstuck, then move |
| 854 | * to the next. Only do this for the primary flavor of RCU. |
| 855 | */ |
| 856 | if (rdp->rsp == rcu_state && |
| 857 | ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) { |
| 858 | rdp->rsp->jiffies_resched += 5; |
| 859 | resched_cpu(rdp->cpu); |
| 860 | } |
| 861 | |
| 862 | return 0; |
| 863 | } |
| 864 | |
| 865 | static void record_gp_stall_check_time(struct rcu_state *rsp) |
| 866 | { |
| 867 | unsigned long j = jiffies; |
| 868 | unsigned long j1; |
| 869 | |
| 870 | rsp->gp_start = j; |
| 871 | smp_wmb(); /* Record start time before stall time. */ |
| 872 | j1 = rcu_jiffies_till_stall_check(); |
| 873 | ACCESS_ONCE(rsp->jiffies_stall) = j + j1; |
| 874 | rsp->jiffies_resched = j + j1 / 2; |
| 875 | } |
| 876 | |
| 877 | /* |
| 878 | * Dump stacks of all tasks running on stalled CPUs. This is a fallback |
| 879 | * for architectures that do not implement trigger_all_cpu_backtrace(). |
| 880 | * The NMI-triggered stack traces are more accurate because they are |
| 881 | * printed by the target CPU. |
| 882 | */ |
| 883 | static void rcu_dump_cpu_stacks(struct rcu_state *rsp) |
| 884 | { |
| 885 | int cpu; |
| 886 | unsigned long flags; |
| 887 | struct rcu_node *rnp; |
| 888 | |
| 889 | rcu_for_each_leaf_node(rsp, rnp) { |
| 890 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 891 | if (rnp->qsmask != 0) { |
| 892 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) |
| 893 | if (rnp->qsmask & (1UL << cpu)) |
| 894 | dump_cpu_task(rnp->grplo + cpu); |
| 895 | } |
| 896 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 897 | } |
| 898 | } |
| 899 | |
| 900 | static void print_other_cpu_stall(struct rcu_state *rsp) |
| 901 | { |
| 902 | int cpu; |
| 903 | long delta; |
| 904 | unsigned long flags; |
| 905 | int ndetected = 0; |
| 906 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 907 | long totqlen = 0; |
| 908 | |
| 909 | /* Only let one CPU complain about others per time interval. */ |
| 910 | |
| 911 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 912 | delta = jiffies - ACCESS_ONCE(rsp->jiffies_stall); |
| 913 | if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { |
| 914 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 915 | return; |
| 916 | } |
| 917 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + 3 * rcu_jiffies_till_stall_check() + 3; |
| 918 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 919 | |
| 920 | /* |
| 921 | * OK, time to rat on our buddy... |
| 922 | * See Documentation/RCU/stallwarn.txt for info on how to debug |
| 923 | * RCU CPU stall warnings. |
| 924 | */ |
| 925 | pr_err("INFO: %s detected stalls on CPUs/tasks:", |
| 926 | rsp->name); |
| 927 | print_cpu_stall_info_begin(); |
| 928 | rcu_for_each_leaf_node(rsp, rnp) { |
| 929 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 930 | ndetected += rcu_print_task_stall(rnp); |
| 931 | if (rnp->qsmask != 0) { |
| 932 | for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++) |
| 933 | if (rnp->qsmask & (1UL << cpu)) { |
| 934 | print_cpu_stall_info(rsp, |
| 935 | rnp->grplo + cpu); |
| 936 | ndetected++; |
| 937 | } |
| 938 | } |
| 939 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 940 | } |
| 941 | |
| 942 | /* |
| 943 | * Now rat on any tasks that got kicked up to the root rcu_node |
| 944 | * due to CPU offlining. |
| 945 | */ |
| 946 | rnp = rcu_get_root(rsp); |
| 947 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 948 | ndetected += rcu_print_task_stall(rnp); |
| 949 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 950 | |
| 951 | print_cpu_stall_info_end(); |
| 952 | for_each_possible_cpu(cpu) |
| 953 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; |
| 954 | pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", |
| 955 | smp_processor_id(), (long)(jiffies - rsp->gp_start), |
| 956 | (long)rsp->gpnum, (long)rsp->completed, totqlen); |
| 957 | if (ndetected == 0) |
| 958 | pr_err("INFO: Stall ended before state dump start\n"); |
| 959 | else if (!trigger_all_cpu_backtrace()) |
| 960 | rcu_dump_cpu_stacks(rsp); |
| 961 | |
| 962 | /* Complain about tasks blocking the grace period. */ |
| 963 | |
| 964 | rcu_print_detail_task_stall(rsp); |
| 965 | |
| 966 | force_quiescent_state(rsp); /* Kick them all. */ |
| 967 | } |
| 968 | |
| 969 | /* |
| 970 | * This function really isn't for public consumption, but RCU is special in |
| 971 | * that context switches can allow the state machine to make progress. |
| 972 | */ |
| 973 | extern void resched_cpu(int cpu); |
| 974 | |
| 975 | static void print_cpu_stall(struct rcu_state *rsp) |
| 976 | { |
| 977 | int cpu; |
| 978 | unsigned long flags; |
| 979 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 980 | long totqlen = 0; |
| 981 | |
| 982 | /* |
| 983 | * OK, time to rat on ourselves... |
| 984 | * See Documentation/RCU/stallwarn.txt for info on how to debug |
| 985 | * RCU CPU stall warnings. |
| 986 | */ |
| 987 | pr_err("INFO: %s self-detected stall on CPU", rsp->name); |
| 988 | print_cpu_stall_info_begin(); |
| 989 | print_cpu_stall_info(rsp, smp_processor_id()); |
| 990 | print_cpu_stall_info_end(); |
| 991 | for_each_possible_cpu(cpu) |
| 992 | totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; |
| 993 | pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", |
| 994 | jiffies - rsp->gp_start, |
| 995 | (long)rsp->gpnum, (long)rsp->completed, totqlen); |
| 996 | if (!trigger_all_cpu_backtrace()) |
| 997 | dump_stack(); |
| 998 | |
| 999 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 1000 | if (ULONG_CMP_GE(jiffies, ACCESS_ONCE(rsp->jiffies_stall))) |
| 1001 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + |
| 1002 | 3 * rcu_jiffies_till_stall_check() + 3; |
| 1003 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1004 | |
| 1005 | /* |
| 1006 | * Attempt to revive the RCU machinery by forcing a context switch. |
| 1007 | * |
| 1008 | * A context switch would normally allow the RCU state machine to make |
| 1009 | * progress and it could be we're stuck in kernel space without context |
| 1010 | * switches for an entirely unreasonable amount of time. |
| 1011 | */ |
| 1012 | resched_cpu(smp_processor_id()); |
| 1013 | } |
| 1014 | |
| 1015 | static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) |
| 1016 | { |
| 1017 | unsigned long completed; |
| 1018 | unsigned long gpnum; |
| 1019 | unsigned long gps; |
| 1020 | unsigned long j; |
| 1021 | unsigned long js; |
| 1022 | struct rcu_node *rnp; |
| 1023 | |
| 1024 | if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp)) |
| 1025 | return; |
| 1026 | j = jiffies; |
| 1027 | |
| 1028 | /* |
| 1029 | * Lots of memory barriers to reject false positives. |
| 1030 | * |
| 1031 | * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, |
| 1032 | * then rsp->gp_start, and finally rsp->completed. These values |
| 1033 | * are updated in the opposite order with memory barriers (or |
| 1034 | * equivalent) during grace-period initialization and cleanup. |
| 1035 | * Now, a false positive can occur if we get an new value of |
| 1036 | * rsp->gp_start and a old value of rsp->jiffies_stall. But given |
| 1037 | * the memory barriers, the only way that this can happen is if one |
| 1038 | * grace period ends and another starts between these two fetches. |
| 1039 | * Detect this by comparing rsp->completed with the previous fetch |
| 1040 | * from rsp->gpnum. |
| 1041 | * |
| 1042 | * Given this check, comparisons of jiffies, rsp->jiffies_stall, |
| 1043 | * and rsp->gp_start suffice to forestall false positives. |
| 1044 | */ |
| 1045 | gpnum = ACCESS_ONCE(rsp->gpnum); |
| 1046 | smp_rmb(); /* Pick up ->gpnum first... */ |
| 1047 | js = ACCESS_ONCE(rsp->jiffies_stall); |
| 1048 | smp_rmb(); /* ...then ->jiffies_stall before the rest... */ |
| 1049 | gps = ACCESS_ONCE(rsp->gp_start); |
| 1050 | smp_rmb(); /* ...and finally ->gp_start before ->completed. */ |
| 1051 | completed = ACCESS_ONCE(rsp->completed); |
| 1052 | if (ULONG_CMP_GE(completed, gpnum) || |
| 1053 | ULONG_CMP_LT(j, js) || |
| 1054 | ULONG_CMP_GE(gps, js)) |
| 1055 | return; /* No stall or GP completed since entering function. */ |
| 1056 | rnp = rdp->mynode; |
| 1057 | if (rcu_gp_in_progress(rsp) && |
| 1058 | (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) { |
| 1059 | |
| 1060 | /* We haven't checked in, so go dump stack. */ |
| 1061 | print_cpu_stall(rsp); |
| 1062 | |
| 1063 | } else if (rcu_gp_in_progress(rsp) && |
| 1064 | ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { |
| 1065 | |
| 1066 | /* They had a few time units to dump stack, so complain. */ |
| 1067 | print_other_cpu_stall(rsp); |
| 1068 | } |
| 1069 | } |
| 1070 | |
| 1071 | /** |
| 1072 | * rcu_cpu_stall_reset - prevent further stall warnings in current grace period |
| 1073 | * |
| 1074 | * Set the stall-warning timeout way off into the future, thus preventing |
| 1075 | * any RCU CPU stall-warning messages from appearing in the current set of |
| 1076 | * RCU grace periods. |
| 1077 | * |
| 1078 | * The caller must disable hard irqs. |
| 1079 | */ |
| 1080 | void rcu_cpu_stall_reset(void) |
| 1081 | { |
| 1082 | struct rcu_state *rsp; |
| 1083 | |
| 1084 | for_each_rcu_flavor(rsp) |
| 1085 | ACCESS_ONCE(rsp->jiffies_stall) = jiffies + ULONG_MAX / 2; |
| 1086 | } |
| 1087 | |
| 1088 | /* |
| 1089 | * Initialize the specified rcu_data structure's callback list to empty. |
| 1090 | */ |
| 1091 | static void init_callback_list(struct rcu_data *rdp) |
| 1092 | { |
| 1093 | int i; |
| 1094 | |
| 1095 | if (init_nocb_callback_list(rdp)) |
| 1096 | return; |
| 1097 | rdp->nxtlist = NULL; |
| 1098 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
| 1099 | rdp->nxttail[i] = &rdp->nxtlist; |
| 1100 | } |
| 1101 | |
| 1102 | /* |
| 1103 | * Determine the value that ->completed will have at the end of the |
| 1104 | * next subsequent grace period. This is used to tag callbacks so that |
| 1105 | * a CPU can invoke callbacks in a timely fashion even if that CPU has |
| 1106 | * been dyntick-idle for an extended period with callbacks under the |
| 1107 | * influence of RCU_FAST_NO_HZ. |
| 1108 | * |
| 1109 | * The caller must hold rnp->lock with interrupts disabled. |
| 1110 | */ |
| 1111 | static unsigned long rcu_cbs_completed(struct rcu_state *rsp, |
| 1112 | struct rcu_node *rnp) |
| 1113 | { |
| 1114 | /* |
| 1115 | * If RCU is idle, we just wait for the next grace period. |
| 1116 | * But we can only be sure that RCU is idle if we are looking |
| 1117 | * at the root rcu_node structure -- otherwise, a new grace |
| 1118 | * period might have started, but just not yet gotten around |
| 1119 | * to initializing the current non-root rcu_node structure. |
| 1120 | */ |
| 1121 | if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) |
| 1122 | return rnp->completed + 1; |
| 1123 | |
| 1124 | /* |
| 1125 | * Otherwise, wait for a possible partial grace period and |
| 1126 | * then the subsequent full grace period. |
| 1127 | */ |
| 1128 | return rnp->completed + 2; |
| 1129 | } |
| 1130 | |
| 1131 | /* |
| 1132 | * Trace-event helper function for rcu_start_future_gp() and |
| 1133 | * rcu_nocb_wait_gp(). |
| 1134 | */ |
| 1135 | static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, |
| 1136 | unsigned long c, const char *s) |
| 1137 | { |
| 1138 | trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, |
| 1139 | rnp->completed, c, rnp->level, |
| 1140 | rnp->grplo, rnp->grphi, s); |
| 1141 | } |
| 1142 | |
| 1143 | /* |
| 1144 | * Start some future grace period, as needed to handle newly arrived |
| 1145 | * callbacks. The required future grace periods are recorded in each |
| 1146 | * rcu_node structure's ->need_future_gp field. Returns true if there |
| 1147 | * is reason to awaken the grace-period kthread. |
| 1148 | * |
| 1149 | * The caller must hold the specified rcu_node structure's ->lock. |
| 1150 | */ |
| 1151 | static bool __maybe_unused |
| 1152 | rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, |
| 1153 | unsigned long *c_out) |
| 1154 | { |
| 1155 | unsigned long c; |
| 1156 | int i; |
| 1157 | bool ret = false; |
| 1158 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); |
| 1159 | |
| 1160 | /* |
| 1161 | * Pick up grace-period number for new callbacks. If this |
| 1162 | * grace period is already marked as needed, return to the caller. |
| 1163 | */ |
| 1164 | c = rcu_cbs_completed(rdp->rsp, rnp); |
| 1165 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); |
| 1166 | if (rnp->need_future_gp[c & 0x1]) { |
| 1167 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); |
| 1168 | goto out; |
| 1169 | } |
| 1170 | |
| 1171 | /* |
| 1172 | * If either this rcu_node structure or the root rcu_node structure |
| 1173 | * believe that a grace period is in progress, then we must wait |
| 1174 | * for the one following, which is in "c". Because our request |
| 1175 | * will be noticed at the end of the current grace period, we don't |
| 1176 | * need to explicitly start one. |
| 1177 | */ |
| 1178 | if (rnp->gpnum != rnp->completed || |
| 1179 | ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) { |
| 1180 | rnp->need_future_gp[c & 0x1]++; |
| 1181 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); |
| 1182 | goto out; |
| 1183 | } |
| 1184 | |
| 1185 | /* |
| 1186 | * There might be no grace period in progress. If we don't already |
| 1187 | * hold it, acquire the root rcu_node structure's lock in order to |
| 1188 | * start one (if needed). |
| 1189 | */ |
| 1190 | if (rnp != rnp_root) { |
| 1191 | raw_spin_lock(&rnp_root->lock); |
| 1192 | smp_mb__after_unlock_lock(); |
| 1193 | } |
| 1194 | |
| 1195 | /* |
| 1196 | * Get a new grace-period number. If there really is no grace |
| 1197 | * period in progress, it will be smaller than the one we obtained |
| 1198 | * earlier. Adjust callbacks as needed. Note that even no-CBs |
| 1199 | * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. |
| 1200 | */ |
| 1201 | c = rcu_cbs_completed(rdp->rsp, rnp_root); |
| 1202 | for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) |
| 1203 | if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) |
| 1204 | rdp->nxtcompleted[i] = c; |
| 1205 | |
| 1206 | /* |
| 1207 | * If the needed for the required grace period is already |
| 1208 | * recorded, trace and leave. |
| 1209 | */ |
| 1210 | if (rnp_root->need_future_gp[c & 0x1]) { |
| 1211 | trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); |
| 1212 | goto unlock_out; |
| 1213 | } |
| 1214 | |
| 1215 | /* Record the need for the future grace period. */ |
| 1216 | rnp_root->need_future_gp[c & 0x1]++; |
| 1217 | |
| 1218 | /* If a grace period is not already in progress, start one. */ |
| 1219 | if (rnp_root->gpnum != rnp_root->completed) { |
| 1220 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); |
| 1221 | } else { |
| 1222 | trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); |
| 1223 | ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); |
| 1224 | } |
| 1225 | unlock_out: |
| 1226 | if (rnp != rnp_root) |
| 1227 | raw_spin_unlock(&rnp_root->lock); |
| 1228 | out: |
| 1229 | if (c_out != NULL) |
| 1230 | *c_out = c; |
| 1231 | return ret; |
| 1232 | } |
| 1233 | |
| 1234 | /* |
| 1235 | * Clean up any old requests for the just-ended grace period. Also return |
| 1236 | * whether any additional grace periods have been requested. Also invoke |
| 1237 | * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads |
| 1238 | * waiting for this grace period to complete. |
| 1239 | */ |
| 1240 | static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
| 1241 | { |
| 1242 | int c = rnp->completed; |
| 1243 | int needmore; |
| 1244 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); |
| 1245 | |
| 1246 | rcu_nocb_gp_cleanup(rsp, rnp); |
| 1247 | rnp->need_future_gp[c & 0x1] = 0; |
| 1248 | needmore = rnp->need_future_gp[(c + 1) & 0x1]; |
| 1249 | trace_rcu_future_gp(rnp, rdp, c, |
| 1250 | needmore ? TPS("CleanupMore") : TPS("Cleanup")); |
| 1251 | return needmore; |
| 1252 | } |
| 1253 | |
| 1254 | /* |
| 1255 | * Awaken the grace-period kthread for the specified flavor of RCU. |
| 1256 | * Don't do a self-awaken, and don't bother awakening when there is |
| 1257 | * nothing for the grace-period kthread to do (as in several CPUs |
| 1258 | * raced to awaken, and we lost), and finally don't try to awaken |
| 1259 | * a kthread that has not yet been created. |
| 1260 | */ |
| 1261 | static void rcu_gp_kthread_wake(struct rcu_state *rsp) |
| 1262 | { |
| 1263 | if (current == rsp->gp_kthread || |
| 1264 | !ACCESS_ONCE(rsp->gp_flags) || |
| 1265 | !rsp->gp_kthread) |
| 1266 | return; |
| 1267 | wake_up(&rsp->gp_wq); |
| 1268 | } |
| 1269 | |
| 1270 | /* |
| 1271 | * If there is room, assign a ->completed number to any callbacks on |
| 1272 | * this CPU that have not already been assigned. Also accelerate any |
| 1273 | * callbacks that were previously assigned a ->completed number that has |
| 1274 | * since proven to be too conservative, which can happen if callbacks get |
| 1275 | * assigned a ->completed number while RCU is idle, but with reference to |
| 1276 | * a non-root rcu_node structure. This function is idempotent, so it does |
| 1277 | * not hurt to call it repeatedly. Returns an flag saying that we should |
| 1278 | * awaken the RCU grace-period kthread. |
| 1279 | * |
| 1280 | * The caller must hold rnp->lock with interrupts disabled. |
| 1281 | */ |
| 1282 | static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
| 1283 | struct rcu_data *rdp) |
| 1284 | { |
| 1285 | unsigned long c; |
| 1286 | int i; |
| 1287 | bool ret; |
| 1288 | |
| 1289 | /* If the CPU has no callbacks, nothing to do. */ |
| 1290 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) |
| 1291 | return false; |
| 1292 | |
| 1293 | /* |
| 1294 | * Starting from the sublist containing the callbacks most |
| 1295 | * recently assigned a ->completed number and working down, find the |
| 1296 | * first sublist that is not assignable to an upcoming grace period. |
| 1297 | * Such a sublist has something in it (first two tests) and has |
| 1298 | * a ->completed number assigned that will complete sooner than |
| 1299 | * the ->completed number for newly arrived callbacks (last test). |
| 1300 | * |
| 1301 | * The key point is that any later sublist can be assigned the |
| 1302 | * same ->completed number as the newly arrived callbacks, which |
| 1303 | * means that the callbacks in any of these later sublist can be |
| 1304 | * grouped into a single sublist, whether or not they have already |
| 1305 | * been assigned a ->completed number. |
| 1306 | */ |
| 1307 | c = rcu_cbs_completed(rsp, rnp); |
| 1308 | for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) |
| 1309 | if (rdp->nxttail[i] != rdp->nxttail[i - 1] && |
| 1310 | !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) |
| 1311 | break; |
| 1312 | |
| 1313 | /* |
| 1314 | * If there are no sublist for unassigned callbacks, leave. |
| 1315 | * At the same time, advance "i" one sublist, so that "i" will |
| 1316 | * index into the sublist where all the remaining callbacks should |
| 1317 | * be grouped into. |
| 1318 | */ |
| 1319 | if (++i >= RCU_NEXT_TAIL) |
| 1320 | return false; |
| 1321 | |
| 1322 | /* |
| 1323 | * Assign all subsequent callbacks' ->completed number to the next |
| 1324 | * full grace period and group them all in the sublist initially |
| 1325 | * indexed by "i". |
| 1326 | */ |
| 1327 | for (; i <= RCU_NEXT_TAIL; i++) { |
| 1328 | rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; |
| 1329 | rdp->nxtcompleted[i] = c; |
| 1330 | } |
| 1331 | /* Record any needed additional grace periods. */ |
| 1332 | ret = rcu_start_future_gp(rnp, rdp, NULL); |
| 1333 | |
| 1334 | /* Trace depending on how much we were able to accelerate. */ |
| 1335 | if (!*rdp->nxttail[RCU_WAIT_TAIL]) |
| 1336 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); |
| 1337 | else |
| 1338 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); |
| 1339 | return ret; |
| 1340 | } |
| 1341 | |
| 1342 | /* |
| 1343 | * Move any callbacks whose grace period has completed to the |
| 1344 | * RCU_DONE_TAIL sublist, then compact the remaining sublists and |
| 1345 | * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL |
| 1346 | * sublist. This function is idempotent, so it does not hurt to |
| 1347 | * invoke it repeatedly. As long as it is not invoked -too- often... |
| 1348 | * Returns true if the RCU grace-period kthread needs to be awakened. |
| 1349 | * |
| 1350 | * The caller must hold rnp->lock with interrupts disabled. |
| 1351 | */ |
| 1352 | static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, |
| 1353 | struct rcu_data *rdp) |
| 1354 | { |
| 1355 | int i, j; |
| 1356 | |
| 1357 | /* If the CPU has no callbacks, nothing to do. */ |
| 1358 | if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) |
| 1359 | return false; |
| 1360 | |
| 1361 | /* |
| 1362 | * Find all callbacks whose ->completed numbers indicate that they |
| 1363 | * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. |
| 1364 | */ |
| 1365 | for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { |
| 1366 | if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) |
| 1367 | break; |
| 1368 | rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; |
| 1369 | } |
| 1370 | /* Clean up any sublist tail pointers that were misordered above. */ |
| 1371 | for (j = RCU_WAIT_TAIL; j < i; j++) |
| 1372 | rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; |
| 1373 | |
| 1374 | /* Copy down callbacks to fill in empty sublists. */ |
| 1375 | for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { |
| 1376 | if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) |
| 1377 | break; |
| 1378 | rdp->nxttail[j] = rdp->nxttail[i]; |
| 1379 | rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; |
| 1380 | } |
| 1381 | |
| 1382 | /* Classify any remaining callbacks. */ |
| 1383 | return rcu_accelerate_cbs(rsp, rnp, rdp); |
| 1384 | } |
| 1385 | |
| 1386 | /* |
| 1387 | * Update CPU-local rcu_data state to record the beginnings and ends of |
| 1388 | * grace periods. The caller must hold the ->lock of the leaf rcu_node |
| 1389 | * structure corresponding to the current CPU, and must have irqs disabled. |
| 1390 | * Returns true if the grace-period kthread needs to be awakened. |
| 1391 | */ |
| 1392 | static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, |
| 1393 | struct rcu_data *rdp) |
| 1394 | { |
| 1395 | bool ret; |
| 1396 | |
| 1397 | /* Handle the ends of any preceding grace periods first. */ |
| 1398 | if (rdp->completed == rnp->completed) { |
| 1399 | |
| 1400 | /* No grace period end, so just accelerate recent callbacks. */ |
| 1401 | ret = rcu_accelerate_cbs(rsp, rnp, rdp); |
| 1402 | |
| 1403 | } else { |
| 1404 | |
| 1405 | /* Advance callbacks. */ |
| 1406 | ret = rcu_advance_cbs(rsp, rnp, rdp); |
| 1407 | |
| 1408 | /* Remember that we saw this grace-period completion. */ |
| 1409 | rdp->completed = rnp->completed; |
| 1410 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); |
| 1411 | } |
| 1412 | |
| 1413 | if (rdp->gpnum != rnp->gpnum) { |
| 1414 | /* |
| 1415 | * If the current grace period is waiting for this CPU, |
| 1416 | * set up to detect a quiescent state, otherwise don't |
| 1417 | * go looking for one. |
| 1418 | */ |
| 1419 | rdp->gpnum = rnp->gpnum; |
| 1420 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); |
| 1421 | rdp->passed_quiesce = 0; |
| 1422 | rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask); |
| 1423 | zero_cpu_stall_ticks(rdp); |
| 1424 | } |
| 1425 | return ret; |
| 1426 | } |
| 1427 | |
| 1428 | static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) |
| 1429 | { |
| 1430 | unsigned long flags; |
| 1431 | bool needwake; |
| 1432 | struct rcu_node *rnp; |
| 1433 | |
| 1434 | local_irq_save(flags); |
| 1435 | rnp = rdp->mynode; |
| 1436 | if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) && |
| 1437 | rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */ |
| 1438 | !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */ |
| 1439 | local_irq_restore(flags); |
| 1440 | return; |
| 1441 | } |
| 1442 | smp_mb__after_unlock_lock(); |
| 1443 | needwake = __note_gp_changes(rsp, rnp, rdp); |
| 1444 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1445 | if (needwake) |
| 1446 | rcu_gp_kthread_wake(rsp); |
| 1447 | } |
| 1448 | |
| 1449 | /* |
| 1450 | * Initialize a new grace period. Return 0 if no grace period required. |
| 1451 | */ |
| 1452 | static int rcu_gp_init(struct rcu_state *rsp) |
| 1453 | { |
| 1454 | struct rcu_data *rdp; |
| 1455 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 1456 | |
| 1457 | rcu_bind_gp_kthread(); |
| 1458 | raw_spin_lock_irq(&rnp->lock); |
| 1459 | smp_mb__after_unlock_lock(); |
| 1460 | if (!ACCESS_ONCE(rsp->gp_flags)) { |
| 1461 | /* Spurious wakeup, tell caller to go back to sleep. */ |
| 1462 | raw_spin_unlock_irq(&rnp->lock); |
| 1463 | return 0; |
| 1464 | } |
| 1465 | ACCESS_ONCE(rsp->gp_flags) = 0; /* Clear all flags: New grace period. */ |
| 1466 | |
| 1467 | if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { |
| 1468 | /* |
| 1469 | * Grace period already in progress, don't start another. |
| 1470 | * Not supposed to be able to happen. |
| 1471 | */ |
| 1472 | raw_spin_unlock_irq(&rnp->lock); |
| 1473 | return 0; |
| 1474 | } |
| 1475 | |
| 1476 | /* Advance to a new grace period and initialize state. */ |
| 1477 | record_gp_stall_check_time(rsp); |
| 1478 | /* Record GP times before starting GP, hence smp_store_release(). */ |
| 1479 | smp_store_release(&rsp->gpnum, rsp->gpnum + 1); |
| 1480 | trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); |
| 1481 | raw_spin_unlock_irq(&rnp->lock); |
| 1482 | |
| 1483 | /* Exclude any concurrent CPU-hotplug operations. */ |
| 1484 | mutex_lock(&rsp->onoff_mutex); |
| 1485 | smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */ |
| 1486 | |
| 1487 | /* |
| 1488 | * Set the quiescent-state-needed bits in all the rcu_node |
| 1489 | * structures for all currently online CPUs in breadth-first order, |
| 1490 | * starting from the root rcu_node structure, relying on the layout |
| 1491 | * of the tree within the rsp->node[] array. Note that other CPUs |
| 1492 | * will access only the leaves of the hierarchy, thus seeing that no |
| 1493 | * grace period is in progress, at least until the corresponding |
| 1494 | * leaf node has been initialized. In addition, we have excluded |
| 1495 | * CPU-hotplug operations. |
| 1496 | * |
| 1497 | * The grace period cannot complete until the initialization |
| 1498 | * process finishes, because this kthread handles both. |
| 1499 | */ |
| 1500 | rcu_for_each_node_breadth_first(rsp, rnp) { |
| 1501 | raw_spin_lock_irq(&rnp->lock); |
| 1502 | smp_mb__after_unlock_lock(); |
| 1503 | rdp = this_cpu_ptr(rsp->rda); |
| 1504 | rcu_preempt_check_blocked_tasks(rnp); |
| 1505 | rnp->qsmask = rnp->qsmaskinit; |
| 1506 | ACCESS_ONCE(rnp->gpnum) = rsp->gpnum; |
| 1507 | WARN_ON_ONCE(rnp->completed != rsp->completed); |
| 1508 | ACCESS_ONCE(rnp->completed) = rsp->completed; |
| 1509 | if (rnp == rdp->mynode) |
| 1510 | (void)__note_gp_changes(rsp, rnp, rdp); |
| 1511 | rcu_preempt_boost_start_gp(rnp); |
| 1512 | trace_rcu_grace_period_init(rsp->name, rnp->gpnum, |
| 1513 | rnp->level, rnp->grplo, |
| 1514 | rnp->grphi, rnp->qsmask); |
| 1515 | raw_spin_unlock_irq(&rnp->lock); |
| 1516 | #ifdef CONFIG_PROVE_RCU_DELAY |
| 1517 | if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 && |
| 1518 | system_state == SYSTEM_RUNNING) |
| 1519 | udelay(200); |
| 1520 | #endif /* #ifdef CONFIG_PROVE_RCU_DELAY */ |
| 1521 | cond_resched(); |
| 1522 | } |
| 1523 | |
| 1524 | mutex_unlock(&rsp->onoff_mutex); |
| 1525 | return 1; |
| 1526 | } |
| 1527 | |
| 1528 | /* |
| 1529 | * Do one round of quiescent-state forcing. |
| 1530 | */ |
| 1531 | static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in) |
| 1532 | { |
| 1533 | int fqs_state = fqs_state_in; |
| 1534 | bool isidle = false; |
| 1535 | unsigned long maxj; |
| 1536 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 1537 | |
| 1538 | rsp->n_force_qs++; |
| 1539 | if (fqs_state == RCU_SAVE_DYNTICK) { |
| 1540 | /* Collect dyntick-idle snapshots. */ |
| 1541 | if (is_sysidle_rcu_state(rsp)) { |
| 1542 | isidle = 1; |
| 1543 | maxj = jiffies - ULONG_MAX / 4; |
| 1544 | } |
| 1545 | force_qs_rnp(rsp, dyntick_save_progress_counter, |
| 1546 | &isidle, &maxj); |
| 1547 | rcu_sysidle_report_gp(rsp, isidle, maxj); |
| 1548 | fqs_state = RCU_FORCE_QS; |
| 1549 | } else { |
| 1550 | /* Handle dyntick-idle and offline CPUs. */ |
| 1551 | isidle = 0; |
| 1552 | force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); |
| 1553 | } |
| 1554 | /* Clear flag to prevent immediate re-entry. */ |
| 1555 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { |
| 1556 | raw_spin_lock_irq(&rnp->lock); |
| 1557 | smp_mb__after_unlock_lock(); |
| 1558 | ACCESS_ONCE(rsp->gp_flags) &= ~RCU_GP_FLAG_FQS; |
| 1559 | raw_spin_unlock_irq(&rnp->lock); |
| 1560 | } |
| 1561 | return fqs_state; |
| 1562 | } |
| 1563 | |
| 1564 | /* |
| 1565 | * Clean up after the old grace period. |
| 1566 | */ |
| 1567 | static void rcu_gp_cleanup(struct rcu_state *rsp) |
| 1568 | { |
| 1569 | unsigned long gp_duration; |
| 1570 | bool needgp = false; |
| 1571 | int nocb = 0; |
| 1572 | struct rcu_data *rdp; |
| 1573 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 1574 | |
| 1575 | raw_spin_lock_irq(&rnp->lock); |
| 1576 | smp_mb__after_unlock_lock(); |
| 1577 | gp_duration = jiffies - rsp->gp_start; |
| 1578 | if (gp_duration > rsp->gp_max) |
| 1579 | rsp->gp_max = gp_duration; |
| 1580 | |
| 1581 | /* |
| 1582 | * We know the grace period is complete, but to everyone else |
| 1583 | * it appears to still be ongoing. But it is also the case |
| 1584 | * that to everyone else it looks like there is nothing that |
| 1585 | * they can do to advance the grace period. It is therefore |
| 1586 | * safe for us to drop the lock in order to mark the grace |
| 1587 | * period as completed in all of the rcu_node structures. |
| 1588 | */ |
| 1589 | raw_spin_unlock_irq(&rnp->lock); |
| 1590 | |
| 1591 | /* |
| 1592 | * Propagate new ->completed value to rcu_node structures so |
| 1593 | * that other CPUs don't have to wait until the start of the next |
| 1594 | * grace period to process their callbacks. This also avoids |
| 1595 | * some nasty RCU grace-period initialization races by forcing |
| 1596 | * the end of the current grace period to be completely recorded in |
| 1597 | * all of the rcu_node structures before the beginning of the next |
| 1598 | * grace period is recorded in any of the rcu_node structures. |
| 1599 | */ |
| 1600 | rcu_for_each_node_breadth_first(rsp, rnp) { |
| 1601 | raw_spin_lock_irq(&rnp->lock); |
| 1602 | smp_mb__after_unlock_lock(); |
| 1603 | ACCESS_ONCE(rnp->completed) = rsp->gpnum; |
| 1604 | rdp = this_cpu_ptr(rsp->rda); |
| 1605 | if (rnp == rdp->mynode) |
| 1606 | needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; |
| 1607 | /* smp_mb() provided by prior unlock-lock pair. */ |
| 1608 | nocb += rcu_future_gp_cleanup(rsp, rnp); |
| 1609 | raw_spin_unlock_irq(&rnp->lock); |
| 1610 | cond_resched(); |
| 1611 | } |
| 1612 | rnp = rcu_get_root(rsp); |
| 1613 | raw_spin_lock_irq(&rnp->lock); |
| 1614 | smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */ |
| 1615 | rcu_nocb_gp_set(rnp, nocb); |
| 1616 | |
| 1617 | /* Declare grace period done. */ |
| 1618 | ACCESS_ONCE(rsp->completed) = rsp->gpnum; |
| 1619 | trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); |
| 1620 | rsp->fqs_state = RCU_GP_IDLE; |
| 1621 | rdp = this_cpu_ptr(rsp->rda); |
| 1622 | /* Advance CBs to reduce false positives below. */ |
| 1623 | needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; |
| 1624 | if (needgp || cpu_needs_another_gp(rsp, rdp)) { |
| 1625 | ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT; |
| 1626 | trace_rcu_grace_period(rsp->name, |
| 1627 | ACCESS_ONCE(rsp->gpnum), |
| 1628 | TPS("newreq")); |
| 1629 | } |
| 1630 | raw_spin_unlock_irq(&rnp->lock); |
| 1631 | } |
| 1632 | |
| 1633 | /* |
| 1634 | * Body of kthread that handles grace periods. |
| 1635 | */ |
| 1636 | static int __noreturn rcu_gp_kthread(void *arg) |
| 1637 | { |
| 1638 | int fqs_state; |
| 1639 | int gf; |
| 1640 | unsigned long j; |
| 1641 | int ret; |
| 1642 | struct rcu_state *rsp = arg; |
| 1643 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 1644 | |
| 1645 | for (;;) { |
| 1646 | |
| 1647 | /* Handle grace-period start. */ |
| 1648 | for (;;) { |
| 1649 | trace_rcu_grace_period(rsp->name, |
| 1650 | ACCESS_ONCE(rsp->gpnum), |
| 1651 | TPS("reqwait")); |
| 1652 | wait_event_interruptible(rsp->gp_wq, |
| 1653 | ACCESS_ONCE(rsp->gp_flags) & |
| 1654 | RCU_GP_FLAG_INIT); |
| 1655 | /* Locking provides needed memory barrier. */ |
| 1656 | if (rcu_gp_init(rsp)) |
| 1657 | break; |
| 1658 | cond_resched(); |
| 1659 | flush_signals(current); |
| 1660 | trace_rcu_grace_period(rsp->name, |
| 1661 | ACCESS_ONCE(rsp->gpnum), |
| 1662 | TPS("reqwaitsig")); |
| 1663 | } |
| 1664 | |
| 1665 | /* Handle quiescent-state forcing. */ |
| 1666 | fqs_state = RCU_SAVE_DYNTICK; |
| 1667 | j = jiffies_till_first_fqs; |
| 1668 | if (j > HZ) { |
| 1669 | j = HZ; |
| 1670 | jiffies_till_first_fqs = HZ; |
| 1671 | } |
| 1672 | ret = 0; |
| 1673 | for (;;) { |
| 1674 | if (!ret) |
| 1675 | rsp->jiffies_force_qs = jiffies + j; |
| 1676 | trace_rcu_grace_period(rsp->name, |
| 1677 | ACCESS_ONCE(rsp->gpnum), |
| 1678 | TPS("fqswait")); |
| 1679 | ret = wait_event_interruptible_timeout(rsp->gp_wq, |
| 1680 | ((gf = ACCESS_ONCE(rsp->gp_flags)) & |
| 1681 | RCU_GP_FLAG_FQS) || |
| 1682 | (!ACCESS_ONCE(rnp->qsmask) && |
| 1683 | !rcu_preempt_blocked_readers_cgp(rnp)), |
| 1684 | j); |
| 1685 | /* Locking provides needed memory barriers. */ |
| 1686 | /* If grace period done, leave loop. */ |
| 1687 | if (!ACCESS_ONCE(rnp->qsmask) && |
| 1688 | !rcu_preempt_blocked_readers_cgp(rnp)) |
| 1689 | break; |
| 1690 | /* If time for quiescent-state forcing, do it. */ |
| 1691 | if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || |
| 1692 | (gf & RCU_GP_FLAG_FQS)) { |
| 1693 | trace_rcu_grace_period(rsp->name, |
| 1694 | ACCESS_ONCE(rsp->gpnum), |
| 1695 | TPS("fqsstart")); |
| 1696 | fqs_state = rcu_gp_fqs(rsp, fqs_state); |
| 1697 | trace_rcu_grace_period(rsp->name, |
| 1698 | ACCESS_ONCE(rsp->gpnum), |
| 1699 | TPS("fqsend")); |
| 1700 | cond_resched(); |
| 1701 | } else { |
| 1702 | /* Deal with stray signal. */ |
| 1703 | cond_resched(); |
| 1704 | flush_signals(current); |
| 1705 | trace_rcu_grace_period(rsp->name, |
| 1706 | ACCESS_ONCE(rsp->gpnum), |
| 1707 | TPS("fqswaitsig")); |
| 1708 | } |
| 1709 | j = jiffies_till_next_fqs; |
| 1710 | if (j > HZ) { |
| 1711 | j = HZ; |
| 1712 | jiffies_till_next_fqs = HZ; |
| 1713 | } else if (j < 1) { |
| 1714 | j = 1; |
| 1715 | jiffies_till_next_fqs = 1; |
| 1716 | } |
| 1717 | } |
| 1718 | |
| 1719 | /* Handle grace-period end. */ |
| 1720 | rcu_gp_cleanup(rsp); |
| 1721 | } |
| 1722 | } |
| 1723 | |
| 1724 | /* |
| 1725 | * Start a new RCU grace period if warranted, re-initializing the hierarchy |
| 1726 | * in preparation for detecting the next grace period. The caller must hold |
| 1727 | * the root node's ->lock and hard irqs must be disabled. |
| 1728 | * |
| 1729 | * Note that it is legal for a dying CPU (which is marked as offline) to |
| 1730 | * invoke this function. This can happen when the dying CPU reports its |
| 1731 | * quiescent state. |
| 1732 | * |
| 1733 | * Returns true if the grace-period kthread must be awakened. |
| 1734 | */ |
| 1735 | static bool |
| 1736 | rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, |
| 1737 | struct rcu_data *rdp) |
| 1738 | { |
| 1739 | if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { |
| 1740 | /* |
| 1741 | * Either we have not yet spawned the grace-period |
| 1742 | * task, this CPU does not need another grace period, |
| 1743 | * or a grace period is already in progress. |
| 1744 | * Either way, don't start a new grace period. |
| 1745 | */ |
| 1746 | return false; |
| 1747 | } |
| 1748 | ACCESS_ONCE(rsp->gp_flags) = RCU_GP_FLAG_INIT; |
| 1749 | trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum), |
| 1750 | TPS("newreq")); |
| 1751 | |
| 1752 | /* |
| 1753 | * We can't do wakeups while holding the rnp->lock, as that |
| 1754 | * could cause possible deadlocks with the rq->lock. Defer |
| 1755 | * the wakeup to our caller. |
| 1756 | */ |
| 1757 | return true; |
| 1758 | } |
| 1759 | |
| 1760 | /* |
| 1761 | * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's |
| 1762 | * callbacks. Note that rcu_start_gp_advanced() cannot do this because it |
| 1763 | * is invoked indirectly from rcu_advance_cbs(), which would result in |
| 1764 | * endless recursion -- or would do so if it wasn't for the self-deadlock |
| 1765 | * that is encountered beforehand. |
| 1766 | * |
| 1767 | * Returns true if the grace-period kthread needs to be awakened. |
| 1768 | */ |
| 1769 | static bool rcu_start_gp(struct rcu_state *rsp) |
| 1770 | { |
| 1771 | struct rcu_data *rdp = this_cpu_ptr(rsp->rda); |
| 1772 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 1773 | bool ret = false; |
| 1774 | |
| 1775 | /* |
| 1776 | * If there is no grace period in progress right now, any |
| 1777 | * callbacks we have up to this point will be satisfied by the |
| 1778 | * next grace period. Also, advancing the callbacks reduces the |
| 1779 | * probability of false positives from cpu_needs_another_gp() |
| 1780 | * resulting in pointless grace periods. So, advance callbacks |
| 1781 | * then start the grace period! |
| 1782 | */ |
| 1783 | ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; |
| 1784 | ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; |
| 1785 | return ret; |
| 1786 | } |
| 1787 | |
| 1788 | /* |
| 1789 | * Report a full set of quiescent states to the specified rcu_state |
| 1790 | * data structure. This involves cleaning up after the prior grace |
| 1791 | * period and letting rcu_start_gp() start up the next grace period |
| 1792 | * if one is needed. Note that the caller must hold rnp->lock, which |
| 1793 | * is released before return. |
| 1794 | */ |
| 1795 | static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) |
| 1796 | __releases(rcu_get_root(rsp)->lock) |
| 1797 | { |
| 1798 | WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); |
| 1799 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
| 1800 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ |
| 1801 | } |
| 1802 | |
| 1803 | /* |
| 1804 | * Similar to rcu_report_qs_rdp(), for which it is a helper function. |
| 1805 | * Allows quiescent states for a group of CPUs to be reported at one go |
| 1806 | * to the specified rcu_node structure, though all the CPUs in the group |
| 1807 | * must be represented by the same rcu_node structure (which need not be |
| 1808 | * a leaf rcu_node structure, though it often will be). That structure's |
| 1809 | * lock must be held upon entry, and it is released before return. |
| 1810 | */ |
| 1811 | static void |
| 1812 | rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, |
| 1813 | struct rcu_node *rnp, unsigned long flags) |
| 1814 | __releases(rnp->lock) |
| 1815 | { |
| 1816 | struct rcu_node *rnp_c; |
| 1817 | |
| 1818 | /* Walk up the rcu_node hierarchy. */ |
| 1819 | for (;;) { |
| 1820 | if (!(rnp->qsmask & mask)) { |
| 1821 | |
| 1822 | /* Our bit has already been cleared, so done. */ |
| 1823 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1824 | return; |
| 1825 | } |
| 1826 | rnp->qsmask &= ~mask; |
| 1827 | trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, |
| 1828 | mask, rnp->qsmask, rnp->level, |
| 1829 | rnp->grplo, rnp->grphi, |
| 1830 | !!rnp->gp_tasks); |
| 1831 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
| 1832 | |
| 1833 | /* Other bits still set at this level, so done. */ |
| 1834 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1835 | return; |
| 1836 | } |
| 1837 | mask = rnp->grpmask; |
| 1838 | if (rnp->parent == NULL) { |
| 1839 | |
| 1840 | /* No more levels. Exit loop holding root lock. */ |
| 1841 | |
| 1842 | break; |
| 1843 | } |
| 1844 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1845 | rnp_c = rnp; |
| 1846 | rnp = rnp->parent; |
| 1847 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 1848 | smp_mb__after_unlock_lock(); |
| 1849 | WARN_ON_ONCE(rnp_c->qsmask); |
| 1850 | } |
| 1851 | |
| 1852 | /* |
| 1853 | * Get here if we are the last CPU to pass through a quiescent |
| 1854 | * state for this grace period. Invoke rcu_report_qs_rsp() |
| 1855 | * to clean up and start the next grace period if one is needed. |
| 1856 | */ |
| 1857 | rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ |
| 1858 | } |
| 1859 | |
| 1860 | /* |
| 1861 | * Record a quiescent state for the specified CPU to that CPU's rcu_data |
| 1862 | * structure. This must be either called from the specified CPU, or |
| 1863 | * called when the specified CPU is known to be offline (and when it is |
| 1864 | * also known that no other CPU is concurrently trying to help the offline |
| 1865 | * CPU). The lastcomp argument is used to make sure we are still in the |
| 1866 | * grace period of interest. We don't want to end the current grace period |
| 1867 | * based on quiescent states detected in an earlier grace period! |
| 1868 | */ |
| 1869 | static void |
| 1870 | rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) |
| 1871 | { |
| 1872 | unsigned long flags; |
| 1873 | unsigned long mask; |
| 1874 | bool needwake; |
| 1875 | struct rcu_node *rnp; |
| 1876 | |
| 1877 | rnp = rdp->mynode; |
| 1878 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 1879 | smp_mb__after_unlock_lock(); |
| 1880 | if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum || |
| 1881 | rnp->completed == rnp->gpnum) { |
| 1882 | |
| 1883 | /* |
| 1884 | * The grace period in which this quiescent state was |
| 1885 | * recorded has ended, so don't report it upwards. |
| 1886 | * We will instead need a new quiescent state that lies |
| 1887 | * within the current grace period. |
| 1888 | */ |
| 1889 | rdp->passed_quiesce = 0; /* need qs for new gp. */ |
| 1890 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1891 | return; |
| 1892 | } |
| 1893 | mask = rdp->grpmask; |
| 1894 | if ((rnp->qsmask & mask) == 0) { |
| 1895 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1896 | } else { |
| 1897 | rdp->qs_pending = 0; |
| 1898 | |
| 1899 | /* |
| 1900 | * This GP can't end until cpu checks in, so all of our |
| 1901 | * callbacks can be processed during the next GP. |
| 1902 | */ |
| 1903 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
| 1904 | |
| 1905 | rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */ |
| 1906 | if (needwake) |
| 1907 | rcu_gp_kthread_wake(rsp); |
| 1908 | } |
| 1909 | } |
| 1910 | |
| 1911 | /* |
| 1912 | * Check to see if there is a new grace period of which this CPU |
| 1913 | * is not yet aware, and if so, set up local rcu_data state for it. |
| 1914 | * Otherwise, see if this CPU has just passed through its first |
| 1915 | * quiescent state for this grace period, and record that fact if so. |
| 1916 | */ |
| 1917 | static void |
| 1918 | rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) |
| 1919 | { |
| 1920 | /* Check for grace-period ends and beginnings. */ |
| 1921 | note_gp_changes(rsp, rdp); |
| 1922 | |
| 1923 | /* |
| 1924 | * Does this CPU still need to do its part for current grace period? |
| 1925 | * If no, return and let the other CPUs do their part as well. |
| 1926 | */ |
| 1927 | if (!rdp->qs_pending) |
| 1928 | return; |
| 1929 | |
| 1930 | /* |
| 1931 | * Was there a quiescent state since the beginning of the grace |
| 1932 | * period? If no, then exit and wait for the next call. |
| 1933 | */ |
| 1934 | if (!rdp->passed_quiesce) |
| 1935 | return; |
| 1936 | |
| 1937 | /* |
| 1938 | * Tell RCU we are done (but rcu_report_qs_rdp() will be the |
| 1939 | * judge of that). |
| 1940 | */ |
| 1941 | rcu_report_qs_rdp(rdp->cpu, rsp, rdp); |
| 1942 | } |
| 1943 | |
| 1944 | #ifdef CONFIG_HOTPLUG_CPU |
| 1945 | |
| 1946 | /* |
| 1947 | * Send the specified CPU's RCU callbacks to the orphanage. The |
| 1948 | * specified CPU must be offline, and the caller must hold the |
| 1949 | * ->orphan_lock. |
| 1950 | */ |
| 1951 | static void |
| 1952 | rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, |
| 1953 | struct rcu_node *rnp, struct rcu_data *rdp) |
| 1954 | { |
| 1955 | /* No-CBs CPUs do not have orphanable callbacks. */ |
| 1956 | if (rcu_is_nocb_cpu(rdp->cpu)) |
| 1957 | return; |
| 1958 | |
| 1959 | /* |
| 1960 | * Orphan the callbacks. First adjust the counts. This is safe |
| 1961 | * because _rcu_barrier() excludes CPU-hotplug operations, so it |
| 1962 | * cannot be running now. Thus no memory barrier is required. |
| 1963 | */ |
| 1964 | if (rdp->nxtlist != NULL) { |
| 1965 | rsp->qlen_lazy += rdp->qlen_lazy; |
| 1966 | rsp->qlen += rdp->qlen; |
| 1967 | rdp->n_cbs_orphaned += rdp->qlen; |
| 1968 | rdp->qlen_lazy = 0; |
| 1969 | ACCESS_ONCE(rdp->qlen) = 0; |
| 1970 | } |
| 1971 | |
| 1972 | /* |
| 1973 | * Next, move those callbacks still needing a grace period to |
| 1974 | * the orphanage, where some other CPU will pick them up. |
| 1975 | * Some of the callbacks might have gone partway through a grace |
| 1976 | * period, but that is too bad. They get to start over because we |
| 1977 | * cannot assume that grace periods are synchronized across CPUs. |
| 1978 | * We don't bother updating the ->nxttail[] array yet, instead |
| 1979 | * we just reset the whole thing later on. |
| 1980 | */ |
| 1981 | if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { |
| 1982 | *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; |
| 1983 | rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; |
| 1984 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; |
| 1985 | } |
| 1986 | |
| 1987 | /* |
| 1988 | * Then move the ready-to-invoke callbacks to the orphanage, |
| 1989 | * where some other CPU will pick them up. These will not be |
| 1990 | * required to pass though another grace period: They are done. |
| 1991 | */ |
| 1992 | if (rdp->nxtlist != NULL) { |
| 1993 | *rsp->orphan_donetail = rdp->nxtlist; |
| 1994 | rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; |
| 1995 | } |
| 1996 | |
| 1997 | /* Finally, initialize the rcu_data structure's list to empty. */ |
| 1998 | init_callback_list(rdp); |
| 1999 | } |
| 2000 | |
| 2001 | /* |
| 2002 | * Adopt the RCU callbacks from the specified rcu_state structure's |
| 2003 | * orphanage. The caller must hold the ->orphan_lock. |
| 2004 | */ |
| 2005 | static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) |
| 2006 | { |
| 2007 | int i; |
| 2008 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
| 2009 | |
| 2010 | /* No-CBs CPUs are handled specially. */ |
| 2011 | if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) |
| 2012 | return; |
| 2013 | |
| 2014 | /* Do the accounting first. */ |
| 2015 | rdp->qlen_lazy += rsp->qlen_lazy; |
| 2016 | rdp->qlen += rsp->qlen; |
| 2017 | rdp->n_cbs_adopted += rsp->qlen; |
| 2018 | if (rsp->qlen_lazy != rsp->qlen) |
| 2019 | rcu_idle_count_callbacks_posted(); |
| 2020 | rsp->qlen_lazy = 0; |
| 2021 | rsp->qlen = 0; |
| 2022 | |
| 2023 | /* |
| 2024 | * We do not need a memory barrier here because the only way we |
| 2025 | * can get here if there is an rcu_barrier() in flight is if |
| 2026 | * we are the task doing the rcu_barrier(). |
| 2027 | */ |
| 2028 | |
| 2029 | /* First adopt the ready-to-invoke callbacks. */ |
| 2030 | if (rsp->orphan_donelist != NULL) { |
| 2031 | *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; |
| 2032 | *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; |
| 2033 | for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) |
| 2034 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) |
| 2035 | rdp->nxttail[i] = rsp->orphan_donetail; |
| 2036 | rsp->orphan_donelist = NULL; |
| 2037 | rsp->orphan_donetail = &rsp->orphan_donelist; |
| 2038 | } |
| 2039 | |
| 2040 | /* And then adopt the callbacks that still need a grace period. */ |
| 2041 | if (rsp->orphan_nxtlist != NULL) { |
| 2042 | *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; |
| 2043 | rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; |
| 2044 | rsp->orphan_nxtlist = NULL; |
| 2045 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; |
| 2046 | } |
| 2047 | } |
| 2048 | |
| 2049 | /* |
| 2050 | * Trace the fact that this CPU is going offline. |
| 2051 | */ |
| 2052 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
| 2053 | { |
| 2054 | RCU_TRACE(unsigned long mask); |
| 2055 | RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); |
| 2056 | RCU_TRACE(struct rcu_node *rnp = rdp->mynode); |
| 2057 | |
| 2058 | RCU_TRACE(mask = rdp->grpmask); |
| 2059 | trace_rcu_grace_period(rsp->name, |
| 2060 | rnp->gpnum + 1 - !!(rnp->qsmask & mask), |
| 2061 | TPS("cpuofl")); |
| 2062 | } |
| 2063 | |
| 2064 | /* |
| 2065 | * The CPU has been completely removed, and some other CPU is reporting |
| 2066 | * this fact from process context. Do the remainder of the cleanup, |
| 2067 | * including orphaning the outgoing CPU's RCU callbacks, and also |
| 2068 | * adopting them. There can only be one CPU hotplug operation at a time, |
| 2069 | * so no other CPU can be attempting to update rcu_cpu_kthread_task. |
| 2070 | */ |
| 2071 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
| 2072 | { |
| 2073 | unsigned long flags; |
| 2074 | unsigned long mask; |
| 2075 | int need_report = 0; |
| 2076 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
| 2077 | struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ |
| 2078 | |
| 2079 | /* Adjust any no-longer-needed kthreads. */ |
| 2080 | rcu_boost_kthread_setaffinity(rnp, -1); |
| 2081 | |
| 2082 | /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */ |
| 2083 | |
| 2084 | /* Exclude any attempts to start a new grace period. */ |
| 2085 | mutex_lock(&rsp->onoff_mutex); |
| 2086 | raw_spin_lock_irqsave(&rsp->orphan_lock, flags); |
| 2087 | |
| 2088 | /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ |
| 2089 | rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); |
| 2090 | rcu_adopt_orphan_cbs(rsp, flags); |
| 2091 | |
| 2092 | /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */ |
| 2093 | mask = rdp->grpmask; /* rnp->grplo is constant. */ |
| 2094 | do { |
| 2095 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
| 2096 | smp_mb__after_unlock_lock(); |
| 2097 | rnp->qsmaskinit &= ~mask; |
| 2098 | if (rnp->qsmaskinit != 0) { |
| 2099 | if (rnp != rdp->mynode) |
| 2100 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 2101 | break; |
| 2102 | } |
| 2103 | if (rnp == rdp->mynode) |
| 2104 | need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp); |
| 2105 | else |
| 2106 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 2107 | mask = rnp->grpmask; |
| 2108 | rnp = rnp->parent; |
| 2109 | } while (rnp != NULL); |
| 2110 | |
| 2111 | /* |
| 2112 | * We still hold the leaf rcu_node structure lock here, and |
| 2113 | * irqs are still disabled. The reason for this subterfuge is |
| 2114 | * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock |
| 2115 | * held leads to deadlock. |
| 2116 | */ |
| 2117 | raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */ |
| 2118 | rnp = rdp->mynode; |
| 2119 | if (need_report & RCU_OFL_TASKS_NORM_GP) |
| 2120 | rcu_report_unblock_qs_rnp(rnp, flags); |
| 2121 | else |
| 2122 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 2123 | if (need_report & RCU_OFL_TASKS_EXP_GP) |
| 2124 | rcu_report_exp_rnp(rsp, rnp, true); |
| 2125 | WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, |
| 2126 | "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", |
| 2127 | cpu, rdp->qlen, rdp->nxtlist); |
| 2128 | init_callback_list(rdp); |
| 2129 | /* Disallow further callbacks on this CPU. */ |
| 2130 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
| 2131 | mutex_unlock(&rsp->onoff_mutex); |
| 2132 | } |
| 2133 | |
| 2134 | #else /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 2135 | |
| 2136 | static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) |
| 2137 | { |
| 2138 | } |
| 2139 | |
| 2140 | static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) |
| 2141 | { |
| 2142 | } |
| 2143 | |
| 2144 | #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */ |
| 2145 | |
| 2146 | /* |
| 2147 | * Invoke any RCU callbacks that have made it to the end of their grace |
| 2148 | * period. Thottle as specified by rdp->blimit. |
| 2149 | */ |
| 2150 | static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) |
| 2151 | { |
| 2152 | unsigned long flags; |
| 2153 | struct rcu_head *next, *list, **tail; |
| 2154 | long bl, count, count_lazy; |
| 2155 | int i; |
| 2156 | |
| 2157 | /* If no callbacks are ready, just return. */ |
| 2158 | if (!cpu_has_callbacks_ready_to_invoke(rdp)) { |
| 2159 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); |
| 2160 | trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist), |
| 2161 | need_resched(), is_idle_task(current), |
| 2162 | rcu_is_callbacks_kthread()); |
| 2163 | return; |
| 2164 | } |
| 2165 | |
| 2166 | /* |
| 2167 | * Extract the list of ready callbacks, disabling to prevent |
| 2168 | * races with call_rcu() from interrupt handlers. |
| 2169 | */ |
| 2170 | local_irq_save(flags); |
| 2171 | WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); |
| 2172 | bl = rdp->blimit; |
| 2173 | trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); |
| 2174 | list = rdp->nxtlist; |
| 2175 | rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; |
| 2176 | *rdp->nxttail[RCU_DONE_TAIL] = NULL; |
| 2177 | tail = rdp->nxttail[RCU_DONE_TAIL]; |
| 2178 | for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) |
| 2179 | if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) |
| 2180 | rdp->nxttail[i] = &rdp->nxtlist; |
| 2181 | local_irq_restore(flags); |
| 2182 | |
| 2183 | /* Invoke callbacks. */ |
| 2184 | count = count_lazy = 0; |
| 2185 | while (list) { |
| 2186 | next = list->next; |
| 2187 | prefetch(next); |
| 2188 | debug_rcu_head_unqueue(list); |
| 2189 | if (__rcu_reclaim(rsp->name, list)) |
| 2190 | count_lazy++; |
| 2191 | list = next; |
| 2192 | /* Stop only if limit reached and CPU has something to do. */ |
| 2193 | if (++count >= bl && |
| 2194 | (need_resched() || |
| 2195 | (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) |
| 2196 | break; |
| 2197 | } |
| 2198 | |
| 2199 | local_irq_save(flags); |
| 2200 | trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), |
| 2201 | is_idle_task(current), |
| 2202 | rcu_is_callbacks_kthread()); |
| 2203 | |
| 2204 | /* Update count, and requeue any remaining callbacks. */ |
| 2205 | if (list != NULL) { |
| 2206 | *tail = rdp->nxtlist; |
| 2207 | rdp->nxtlist = list; |
| 2208 | for (i = 0; i < RCU_NEXT_SIZE; i++) |
| 2209 | if (&rdp->nxtlist == rdp->nxttail[i]) |
| 2210 | rdp->nxttail[i] = tail; |
| 2211 | else |
| 2212 | break; |
| 2213 | } |
| 2214 | smp_mb(); /* List handling before counting for rcu_barrier(). */ |
| 2215 | rdp->qlen_lazy -= count_lazy; |
| 2216 | ACCESS_ONCE(rdp->qlen) -= count; |
| 2217 | rdp->n_cbs_invoked += count; |
| 2218 | |
| 2219 | /* Reinstate batch limit if we have worked down the excess. */ |
| 2220 | if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) |
| 2221 | rdp->blimit = blimit; |
| 2222 | |
| 2223 | /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ |
| 2224 | if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { |
| 2225 | rdp->qlen_last_fqs_check = 0; |
| 2226 | rdp->n_force_qs_snap = rsp->n_force_qs; |
| 2227 | } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) |
| 2228 | rdp->qlen_last_fqs_check = rdp->qlen; |
| 2229 | WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); |
| 2230 | |
| 2231 | local_irq_restore(flags); |
| 2232 | |
| 2233 | /* Re-invoke RCU core processing if there are callbacks remaining. */ |
| 2234 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
| 2235 | invoke_rcu_core(); |
| 2236 | } |
| 2237 | |
| 2238 | /* |
| 2239 | * Check to see if this CPU is in a non-context-switch quiescent state |
| 2240 | * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). |
| 2241 | * Also schedule RCU core processing. |
| 2242 | * |
| 2243 | * This function must be called from hardirq context. It is normally |
| 2244 | * invoked from the scheduling-clock interrupt. If rcu_pending returns |
| 2245 | * false, there is no point in invoking rcu_check_callbacks(). |
| 2246 | */ |
| 2247 | void rcu_check_callbacks(int cpu, int user) |
| 2248 | { |
| 2249 | trace_rcu_utilization(TPS("Start scheduler-tick")); |
| 2250 | increment_cpu_stall_ticks(); |
| 2251 | if (user || rcu_is_cpu_rrupt_from_idle()) { |
| 2252 | |
| 2253 | /* |
| 2254 | * Get here if this CPU took its interrupt from user |
| 2255 | * mode or from the idle loop, and if this is not a |
| 2256 | * nested interrupt. In this case, the CPU is in |
| 2257 | * a quiescent state, so note it. |
| 2258 | * |
| 2259 | * No memory barrier is required here because both |
| 2260 | * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local |
| 2261 | * variables that other CPUs neither access nor modify, |
| 2262 | * at least not while the corresponding CPU is online. |
| 2263 | */ |
| 2264 | |
| 2265 | rcu_sched_qs(cpu); |
| 2266 | rcu_bh_qs(cpu); |
| 2267 | |
| 2268 | } else if (!in_softirq()) { |
| 2269 | |
| 2270 | /* |
| 2271 | * Get here if this CPU did not take its interrupt from |
| 2272 | * softirq, in other words, if it is not interrupting |
| 2273 | * a rcu_bh read-side critical section. This is an _bh |
| 2274 | * critical section, so note it. |
| 2275 | */ |
| 2276 | |
| 2277 | rcu_bh_qs(cpu); |
| 2278 | } |
| 2279 | rcu_preempt_check_callbacks(cpu); |
| 2280 | if (rcu_pending(cpu)) |
| 2281 | invoke_rcu_core(); |
| 2282 | trace_rcu_utilization(TPS("End scheduler-tick")); |
| 2283 | } |
| 2284 | |
| 2285 | /* |
| 2286 | * Scan the leaf rcu_node structures, processing dyntick state for any that |
| 2287 | * have not yet encountered a quiescent state, using the function specified. |
| 2288 | * Also initiate boosting for any threads blocked on the root rcu_node. |
| 2289 | * |
| 2290 | * The caller must have suppressed start of new grace periods. |
| 2291 | */ |
| 2292 | static void force_qs_rnp(struct rcu_state *rsp, |
| 2293 | int (*f)(struct rcu_data *rsp, bool *isidle, |
| 2294 | unsigned long *maxj), |
| 2295 | bool *isidle, unsigned long *maxj) |
| 2296 | { |
| 2297 | unsigned long bit; |
| 2298 | int cpu; |
| 2299 | unsigned long flags; |
| 2300 | unsigned long mask; |
| 2301 | struct rcu_node *rnp; |
| 2302 | |
| 2303 | rcu_for_each_leaf_node(rsp, rnp) { |
| 2304 | cond_resched(); |
| 2305 | mask = 0; |
| 2306 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 2307 | smp_mb__after_unlock_lock(); |
| 2308 | if (!rcu_gp_in_progress(rsp)) { |
| 2309 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 2310 | return; |
| 2311 | } |
| 2312 | if (rnp->qsmask == 0) { |
| 2313 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
| 2314 | continue; |
| 2315 | } |
| 2316 | cpu = rnp->grplo; |
| 2317 | bit = 1; |
| 2318 | for (; cpu <= rnp->grphi; cpu++, bit <<= 1) { |
| 2319 | if ((rnp->qsmask & bit) != 0) { |
| 2320 | if ((rnp->qsmaskinit & bit) != 0) |
| 2321 | *isidle = 0; |
| 2322 | if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) |
| 2323 | mask |= bit; |
| 2324 | } |
| 2325 | } |
| 2326 | if (mask != 0) { |
| 2327 | |
| 2328 | /* rcu_report_qs_rnp() releases rnp->lock. */ |
| 2329 | rcu_report_qs_rnp(mask, rsp, rnp, flags); |
| 2330 | continue; |
| 2331 | } |
| 2332 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 2333 | } |
| 2334 | rnp = rcu_get_root(rsp); |
| 2335 | if (rnp->qsmask == 0) { |
| 2336 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 2337 | smp_mb__after_unlock_lock(); |
| 2338 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ |
| 2339 | } |
| 2340 | } |
| 2341 | |
| 2342 | /* |
| 2343 | * Force quiescent states on reluctant CPUs, and also detect which |
| 2344 | * CPUs are in dyntick-idle mode. |
| 2345 | */ |
| 2346 | static void force_quiescent_state(struct rcu_state *rsp) |
| 2347 | { |
| 2348 | unsigned long flags; |
| 2349 | bool ret; |
| 2350 | struct rcu_node *rnp; |
| 2351 | struct rcu_node *rnp_old = NULL; |
| 2352 | |
| 2353 | /* Funnel through hierarchy to reduce memory contention. */ |
| 2354 | rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode; |
| 2355 | for (; rnp != NULL; rnp = rnp->parent) { |
| 2356 | ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || |
| 2357 | !raw_spin_trylock(&rnp->fqslock); |
| 2358 | if (rnp_old != NULL) |
| 2359 | raw_spin_unlock(&rnp_old->fqslock); |
| 2360 | if (ret) { |
| 2361 | ACCESS_ONCE(rsp->n_force_qs_lh)++; |
| 2362 | return; |
| 2363 | } |
| 2364 | rnp_old = rnp; |
| 2365 | } |
| 2366 | /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ |
| 2367 | |
| 2368 | /* Reached the root of the rcu_node tree, acquire lock. */ |
| 2369 | raw_spin_lock_irqsave(&rnp_old->lock, flags); |
| 2370 | smp_mb__after_unlock_lock(); |
| 2371 | raw_spin_unlock(&rnp_old->fqslock); |
| 2372 | if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { |
| 2373 | ACCESS_ONCE(rsp->n_force_qs_lh)++; |
| 2374 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
| 2375 | return; /* Someone beat us to it. */ |
| 2376 | } |
| 2377 | ACCESS_ONCE(rsp->gp_flags) |= RCU_GP_FLAG_FQS; |
| 2378 | raw_spin_unlock_irqrestore(&rnp_old->lock, flags); |
| 2379 | wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */ |
| 2380 | } |
| 2381 | |
| 2382 | /* |
| 2383 | * This does the RCU core processing work for the specified rcu_state |
| 2384 | * and rcu_data structures. This may be called only from the CPU to |
| 2385 | * whom the rdp belongs. |
| 2386 | */ |
| 2387 | static void |
| 2388 | __rcu_process_callbacks(struct rcu_state *rsp) |
| 2389 | { |
| 2390 | unsigned long flags; |
| 2391 | bool needwake; |
| 2392 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
| 2393 | |
| 2394 | WARN_ON_ONCE(rdp->beenonline == 0); |
| 2395 | |
| 2396 | /* Update RCU state based on any recent quiescent states. */ |
| 2397 | rcu_check_quiescent_state(rsp, rdp); |
| 2398 | |
| 2399 | /* Does this CPU require a not-yet-started grace period? */ |
| 2400 | local_irq_save(flags); |
| 2401 | if (cpu_needs_another_gp(rsp, rdp)) { |
| 2402 | raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */ |
| 2403 | needwake = rcu_start_gp(rsp); |
| 2404 | raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags); |
| 2405 | if (needwake) |
| 2406 | rcu_gp_kthread_wake(rsp); |
| 2407 | } else { |
| 2408 | local_irq_restore(flags); |
| 2409 | } |
| 2410 | |
| 2411 | /* If there are callbacks ready, invoke them. */ |
| 2412 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
| 2413 | invoke_rcu_callbacks(rsp, rdp); |
| 2414 | |
| 2415 | /* Do any needed deferred wakeups of rcuo kthreads. */ |
| 2416 | do_nocb_deferred_wakeup(rdp); |
| 2417 | } |
| 2418 | |
| 2419 | /* |
| 2420 | * Do RCU core processing for the current CPU. |
| 2421 | */ |
| 2422 | static void rcu_process_callbacks(struct softirq_action *unused) |
| 2423 | { |
| 2424 | struct rcu_state *rsp; |
| 2425 | |
| 2426 | if (cpu_is_offline(smp_processor_id())) |
| 2427 | return; |
| 2428 | trace_rcu_utilization(TPS("Start RCU core")); |
| 2429 | for_each_rcu_flavor(rsp) |
| 2430 | __rcu_process_callbacks(rsp); |
| 2431 | trace_rcu_utilization(TPS("End RCU core")); |
| 2432 | } |
| 2433 | |
| 2434 | /* |
| 2435 | * Schedule RCU callback invocation. If the specified type of RCU |
| 2436 | * does not support RCU priority boosting, just do a direct call, |
| 2437 | * otherwise wake up the per-CPU kernel kthread. Note that because we |
| 2438 | * are running on the current CPU with interrupts disabled, the |
| 2439 | * rcu_cpu_kthread_task cannot disappear out from under us. |
| 2440 | */ |
| 2441 | static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) |
| 2442 | { |
| 2443 | if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active))) |
| 2444 | return; |
| 2445 | if (likely(!rsp->boost)) { |
| 2446 | rcu_do_batch(rsp, rdp); |
| 2447 | return; |
| 2448 | } |
| 2449 | invoke_rcu_callbacks_kthread(); |
| 2450 | } |
| 2451 | |
| 2452 | static void invoke_rcu_core(void) |
| 2453 | { |
| 2454 | if (cpu_online(smp_processor_id())) |
| 2455 | raise_softirq(RCU_SOFTIRQ); |
| 2456 | } |
| 2457 | |
| 2458 | /* |
| 2459 | * Handle any core-RCU processing required by a call_rcu() invocation. |
| 2460 | */ |
| 2461 | static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, |
| 2462 | struct rcu_head *head, unsigned long flags) |
| 2463 | { |
| 2464 | bool needwake; |
| 2465 | |
| 2466 | /* |
| 2467 | * If called from an extended quiescent state, invoke the RCU |
| 2468 | * core in order to force a re-evaluation of RCU's idleness. |
| 2469 | */ |
| 2470 | if (!rcu_is_watching() && cpu_online(smp_processor_id())) |
| 2471 | invoke_rcu_core(); |
| 2472 | |
| 2473 | /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ |
| 2474 | if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) |
| 2475 | return; |
| 2476 | |
| 2477 | /* |
| 2478 | * Force the grace period if too many callbacks or too long waiting. |
| 2479 | * Enforce hysteresis, and don't invoke force_quiescent_state() |
| 2480 | * if some other CPU has recently done so. Also, don't bother |
| 2481 | * invoking force_quiescent_state() if the newly enqueued callback |
| 2482 | * is the only one waiting for a grace period to complete. |
| 2483 | */ |
| 2484 | if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { |
| 2485 | |
| 2486 | /* Are we ignoring a completed grace period? */ |
| 2487 | note_gp_changes(rsp, rdp); |
| 2488 | |
| 2489 | /* Start a new grace period if one not already started. */ |
| 2490 | if (!rcu_gp_in_progress(rsp)) { |
| 2491 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
| 2492 | |
| 2493 | raw_spin_lock(&rnp_root->lock); |
| 2494 | smp_mb__after_unlock_lock(); |
| 2495 | needwake = rcu_start_gp(rsp); |
| 2496 | raw_spin_unlock(&rnp_root->lock); |
| 2497 | if (needwake) |
| 2498 | rcu_gp_kthread_wake(rsp); |
| 2499 | } else { |
| 2500 | /* Give the grace period a kick. */ |
| 2501 | rdp->blimit = LONG_MAX; |
| 2502 | if (rsp->n_force_qs == rdp->n_force_qs_snap && |
| 2503 | *rdp->nxttail[RCU_DONE_TAIL] != head) |
| 2504 | force_quiescent_state(rsp); |
| 2505 | rdp->n_force_qs_snap = rsp->n_force_qs; |
| 2506 | rdp->qlen_last_fqs_check = rdp->qlen; |
| 2507 | } |
| 2508 | } |
| 2509 | } |
| 2510 | |
| 2511 | /* |
| 2512 | * RCU callback function to leak a callback. |
| 2513 | */ |
| 2514 | static void rcu_leak_callback(struct rcu_head *rhp) |
| 2515 | { |
| 2516 | } |
| 2517 | |
| 2518 | /* |
| 2519 | * Helper function for call_rcu() and friends. The cpu argument will |
| 2520 | * normally be -1, indicating "currently running CPU". It may specify |
| 2521 | * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() |
| 2522 | * is expected to specify a CPU. |
| 2523 | */ |
| 2524 | static void |
| 2525 | __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu), |
| 2526 | struct rcu_state *rsp, int cpu, bool lazy) |
| 2527 | { |
| 2528 | unsigned long flags; |
| 2529 | struct rcu_data *rdp; |
| 2530 | |
| 2531 | WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */ |
| 2532 | if (debug_rcu_head_queue(head)) { |
| 2533 | /* Probable double call_rcu(), so leak the callback. */ |
| 2534 | ACCESS_ONCE(head->func) = rcu_leak_callback; |
| 2535 | WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); |
| 2536 | return; |
| 2537 | } |
| 2538 | head->func = func; |
| 2539 | head->next = NULL; |
| 2540 | |
| 2541 | /* |
| 2542 | * Opportunistically note grace-period endings and beginnings. |
| 2543 | * Note that we might see a beginning right after we see an |
| 2544 | * end, but never vice versa, since this CPU has to pass through |
| 2545 | * a quiescent state betweentimes. |
| 2546 | */ |
| 2547 | local_irq_save(flags); |
| 2548 | rdp = this_cpu_ptr(rsp->rda); |
| 2549 | |
| 2550 | /* Add the callback to our list. */ |
| 2551 | if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { |
| 2552 | int offline; |
| 2553 | |
| 2554 | if (cpu != -1) |
| 2555 | rdp = per_cpu_ptr(rsp->rda, cpu); |
| 2556 | offline = !__call_rcu_nocb(rdp, head, lazy, flags); |
| 2557 | WARN_ON_ONCE(offline); |
| 2558 | /* _call_rcu() is illegal on offline CPU; leak the callback. */ |
| 2559 | local_irq_restore(flags); |
| 2560 | return; |
| 2561 | } |
| 2562 | ACCESS_ONCE(rdp->qlen)++; |
| 2563 | if (lazy) |
| 2564 | rdp->qlen_lazy++; |
| 2565 | else |
| 2566 | rcu_idle_count_callbacks_posted(); |
| 2567 | smp_mb(); /* Count before adding callback for rcu_barrier(). */ |
| 2568 | *rdp->nxttail[RCU_NEXT_TAIL] = head; |
| 2569 | rdp->nxttail[RCU_NEXT_TAIL] = &head->next; |
| 2570 | |
| 2571 | if (__is_kfree_rcu_offset((unsigned long)func)) |
| 2572 | trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, |
| 2573 | rdp->qlen_lazy, rdp->qlen); |
| 2574 | else |
| 2575 | trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); |
| 2576 | |
| 2577 | /* Go handle any RCU core processing required. */ |
| 2578 | __call_rcu_core(rsp, rdp, head, flags); |
| 2579 | local_irq_restore(flags); |
| 2580 | } |
| 2581 | |
| 2582 | /* |
| 2583 | * Queue an RCU-sched callback for invocation after a grace period. |
| 2584 | */ |
| 2585 | void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
| 2586 | { |
| 2587 | __call_rcu(head, func, &rcu_sched_state, -1, 0); |
| 2588 | } |
| 2589 | EXPORT_SYMBOL_GPL(call_rcu_sched); |
| 2590 | |
| 2591 | /* |
| 2592 | * Queue an RCU callback for invocation after a quicker grace period. |
| 2593 | */ |
| 2594 | void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
| 2595 | { |
| 2596 | __call_rcu(head, func, &rcu_bh_state, -1, 0); |
| 2597 | } |
| 2598 | EXPORT_SYMBOL_GPL(call_rcu_bh); |
| 2599 | |
| 2600 | /* |
| 2601 | * Queue an RCU callback for lazy invocation after a grace period. |
| 2602 | * This will likely be later named something like "call_rcu_lazy()", |
| 2603 | * but this change will require some way of tagging the lazy RCU |
| 2604 | * callbacks in the list of pending callbacks. Until then, this |
| 2605 | * function may only be called from __kfree_rcu(). |
| 2606 | */ |
| 2607 | void kfree_call_rcu(struct rcu_head *head, |
| 2608 | void (*func)(struct rcu_head *rcu)) |
| 2609 | { |
| 2610 | __call_rcu(head, func, rcu_state, -1, 1); |
| 2611 | } |
| 2612 | EXPORT_SYMBOL_GPL(kfree_call_rcu); |
| 2613 | |
| 2614 | /* |
| 2615 | * Because a context switch is a grace period for RCU-sched and RCU-bh, |
| 2616 | * any blocking grace-period wait automatically implies a grace period |
| 2617 | * if there is only one CPU online at any point time during execution |
| 2618 | * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to |
| 2619 | * occasionally incorrectly indicate that there are multiple CPUs online |
| 2620 | * when there was in fact only one the whole time, as this just adds |
| 2621 | * some overhead: RCU still operates correctly. |
| 2622 | */ |
| 2623 | static inline int rcu_blocking_is_gp(void) |
| 2624 | { |
| 2625 | int ret; |
| 2626 | |
| 2627 | might_sleep(); /* Check for RCU read-side critical section. */ |
| 2628 | preempt_disable(); |
| 2629 | ret = num_online_cpus() <= 1; |
| 2630 | preempt_enable(); |
| 2631 | return ret; |
| 2632 | } |
| 2633 | |
| 2634 | /** |
| 2635 | * synchronize_sched - wait until an rcu-sched grace period has elapsed. |
| 2636 | * |
| 2637 | * Control will return to the caller some time after a full rcu-sched |
| 2638 | * grace period has elapsed, in other words after all currently executing |
| 2639 | * rcu-sched read-side critical sections have completed. These read-side |
| 2640 | * critical sections are delimited by rcu_read_lock_sched() and |
| 2641 | * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), |
| 2642 | * local_irq_disable(), and so on may be used in place of |
| 2643 | * rcu_read_lock_sched(). |
| 2644 | * |
| 2645 | * This means that all preempt_disable code sequences, including NMI and |
| 2646 | * non-threaded hardware-interrupt handlers, in progress on entry will |
| 2647 | * have completed before this primitive returns. However, this does not |
| 2648 | * guarantee that softirq handlers will have completed, since in some |
| 2649 | * kernels, these handlers can run in process context, and can block. |
| 2650 | * |
| 2651 | * Note that this guarantee implies further memory-ordering guarantees. |
| 2652 | * On systems with more than one CPU, when synchronize_sched() returns, |
| 2653 | * each CPU is guaranteed to have executed a full memory barrier since the |
| 2654 | * end of its last RCU-sched read-side critical section whose beginning |
| 2655 | * preceded the call to synchronize_sched(). In addition, each CPU having |
| 2656 | * an RCU read-side critical section that extends beyond the return from |
| 2657 | * synchronize_sched() is guaranteed to have executed a full memory barrier |
| 2658 | * after the beginning of synchronize_sched() and before the beginning of |
| 2659 | * that RCU read-side critical section. Note that these guarantees include |
| 2660 | * CPUs that are offline, idle, or executing in user mode, as well as CPUs |
| 2661 | * that are executing in the kernel. |
| 2662 | * |
| 2663 | * Furthermore, if CPU A invoked synchronize_sched(), which returned |
| 2664 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed |
| 2665 | * to have executed a full memory barrier during the execution of |
| 2666 | * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but |
| 2667 | * again only if the system has more than one CPU). |
| 2668 | * |
| 2669 | * This primitive provides the guarantees made by the (now removed) |
| 2670 | * synchronize_kernel() API. In contrast, synchronize_rcu() only |
| 2671 | * guarantees that rcu_read_lock() sections will have completed. |
| 2672 | * In "classic RCU", these two guarantees happen to be one and |
| 2673 | * the same, but can differ in realtime RCU implementations. |
| 2674 | */ |
| 2675 | void synchronize_sched(void) |
| 2676 | { |
| 2677 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
| 2678 | !lock_is_held(&rcu_lock_map) && |
| 2679 | !lock_is_held(&rcu_sched_lock_map), |
| 2680 | "Illegal synchronize_sched() in RCU-sched read-side critical section"); |
| 2681 | if (rcu_blocking_is_gp()) |
| 2682 | return; |
| 2683 | if (rcu_expedited) |
| 2684 | synchronize_sched_expedited(); |
| 2685 | else |
| 2686 | wait_rcu_gp(call_rcu_sched); |
| 2687 | } |
| 2688 | EXPORT_SYMBOL_GPL(synchronize_sched); |
| 2689 | |
| 2690 | /** |
| 2691 | * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. |
| 2692 | * |
| 2693 | * Control will return to the caller some time after a full rcu_bh grace |
| 2694 | * period has elapsed, in other words after all currently executing rcu_bh |
| 2695 | * read-side critical sections have completed. RCU read-side critical |
| 2696 | * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), |
| 2697 | * and may be nested. |
| 2698 | * |
| 2699 | * See the description of synchronize_sched() for more detailed information |
| 2700 | * on memory ordering guarantees. |
| 2701 | */ |
| 2702 | void synchronize_rcu_bh(void) |
| 2703 | { |
| 2704 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
| 2705 | !lock_is_held(&rcu_lock_map) && |
| 2706 | !lock_is_held(&rcu_sched_lock_map), |
| 2707 | "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); |
| 2708 | if (rcu_blocking_is_gp()) |
| 2709 | return; |
| 2710 | if (rcu_expedited) |
| 2711 | synchronize_rcu_bh_expedited(); |
| 2712 | else |
| 2713 | wait_rcu_gp(call_rcu_bh); |
| 2714 | } |
| 2715 | EXPORT_SYMBOL_GPL(synchronize_rcu_bh); |
| 2716 | |
| 2717 | /** |
| 2718 | * get_state_synchronize_rcu - Snapshot current RCU state |
| 2719 | * |
| 2720 | * Returns a cookie that is used by a later call to cond_synchronize_rcu() |
| 2721 | * to determine whether or not a full grace period has elapsed in the |
| 2722 | * meantime. |
| 2723 | */ |
| 2724 | unsigned long get_state_synchronize_rcu(void) |
| 2725 | { |
| 2726 | /* |
| 2727 | * Any prior manipulation of RCU-protected data must happen |
| 2728 | * before the load from ->gpnum. |
| 2729 | */ |
| 2730 | smp_mb(); /* ^^^ */ |
| 2731 | |
| 2732 | /* |
| 2733 | * Make sure this load happens before the purportedly |
| 2734 | * time-consuming work between get_state_synchronize_rcu() |
| 2735 | * and cond_synchronize_rcu(). |
| 2736 | */ |
| 2737 | return smp_load_acquire(&rcu_state->gpnum); |
| 2738 | } |
| 2739 | EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); |
| 2740 | |
| 2741 | /** |
| 2742 | * cond_synchronize_rcu - Conditionally wait for an RCU grace period |
| 2743 | * |
| 2744 | * @oldstate: return value from earlier call to get_state_synchronize_rcu() |
| 2745 | * |
| 2746 | * If a full RCU grace period has elapsed since the earlier call to |
| 2747 | * get_state_synchronize_rcu(), just return. Otherwise, invoke |
| 2748 | * synchronize_rcu() to wait for a full grace period. |
| 2749 | * |
| 2750 | * Yes, this function does not take counter wrap into account. But |
| 2751 | * counter wrap is harmless. If the counter wraps, we have waited for |
| 2752 | * more than 2 billion grace periods (and way more on a 64-bit system!), |
| 2753 | * so waiting for one additional grace period should be just fine. |
| 2754 | */ |
| 2755 | void cond_synchronize_rcu(unsigned long oldstate) |
| 2756 | { |
| 2757 | unsigned long newstate; |
| 2758 | |
| 2759 | /* |
| 2760 | * Ensure that this load happens before any RCU-destructive |
| 2761 | * actions the caller might carry out after we return. |
| 2762 | */ |
| 2763 | newstate = smp_load_acquire(&rcu_state->completed); |
| 2764 | if (ULONG_CMP_GE(oldstate, newstate)) |
| 2765 | synchronize_rcu(); |
| 2766 | } |
| 2767 | EXPORT_SYMBOL_GPL(cond_synchronize_rcu); |
| 2768 | |
| 2769 | static int synchronize_sched_expedited_cpu_stop(void *data) |
| 2770 | { |
| 2771 | /* |
| 2772 | * There must be a full memory barrier on each affected CPU |
| 2773 | * between the time that try_stop_cpus() is called and the |
| 2774 | * time that it returns. |
| 2775 | * |
| 2776 | * In the current initial implementation of cpu_stop, the |
| 2777 | * above condition is already met when the control reaches |
| 2778 | * this point and the following smp_mb() is not strictly |
| 2779 | * necessary. Do smp_mb() anyway for documentation and |
| 2780 | * robustness against future implementation changes. |
| 2781 | */ |
| 2782 | smp_mb(); /* See above comment block. */ |
| 2783 | return 0; |
| 2784 | } |
| 2785 | |
| 2786 | /** |
| 2787 | * synchronize_sched_expedited - Brute-force RCU-sched grace period |
| 2788 | * |
| 2789 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" |
| 2790 | * approach to force the grace period to end quickly. This consumes |
| 2791 | * significant time on all CPUs and is unfriendly to real-time workloads, |
| 2792 | * so is thus not recommended for any sort of common-case code. In fact, |
| 2793 | * if you are using synchronize_sched_expedited() in a loop, please |
| 2794 | * restructure your code to batch your updates, and then use a single |
| 2795 | * synchronize_sched() instead. |
| 2796 | * |
| 2797 | * Note that it is illegal to call this function while holding any lock |
| 2798 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal |
| 2799 | * to call this function from a CPU-hotplug notifier. Failing to observe |
| 2800 | * these restriction will result in deadlock. |
| 2801 | * |
| 2802 | * This implementation can be thought of as an application of ticket |
| 2803 | * locking to RCU, with sync_sched_expedited_started and |
| 2804 | * sync_sched_expedited_done taking on the roles of the halves |
| 2805 | * of the ticket-lock word. Each task atomically increments |
| 2806 | * sync_sched_expedited_started upon entry, snapshotting the old value, |
| 2807 | * then attempts to stop all the CPUs. If this succeeds, then each |
| 2808 | * CPU will have executed a context switch, resulting in an RCU-sched |
| 2809 | * grace period. We are then done, so we use atomic_cmpxchg() to |
| 2810 | * update sync_sched_expedited_done to match our snapshot -- but |
| 2811 | * only if someone else has not already advanced past our snapshot. |
| 2812 | * |
| 2813 | * On the other hand, if try_stop_cpus() fails, we check the value |
| 2814 | * of sync_sched_expedited_done. If it has advanced past our |
| 2815 | * initial snapshot, then someone else must have forced a grace period |
| 2816 | * some time after we took our snapshot. In this case, our work is |
| 2817 | * done for us, and we can simply return. Otherwise, we try again, |
| 2818 | * but keep our initial snapshot for purposes of checking for someone |
| 2819 | * doing our work for us. |
| 2820 | * |
| 2821 | * If we fail too many times in a row, we fall back to synchronize_sched(). |
| 2822 | */ |
| 2823 | void synchronize_sched_expedited(void) |
| 2824 | { |
| 2825 | long firstsnap, s, snap; |
| 2826 | int trycount = 0; |
| 2827 | struct rcu_state *rsp = &rcu_sched_state; |
| 2828 | |
| 2829 | /* |
| 2830 | * If we are in danger of counter wrap, just do synchronize_sched(). |
| 2831 | * By allowing sync_sched_expedited_started to advance no more than |
| 2832 | * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring |
| 2833 | * that more than 3.5 billion CPUs would be required to force a |
| 2834 | * counter wrap on a 32-bit system. Quite a few more CPUs would of |
| 2835 | * course be required on a 64-bit system. |
| 2836 | */ |
| 2837 | if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start), |
| 2838 | (ulong)atomic_long_read(&rsp->expedited_done) + |
| 2839 | ULONG_MAX / 8)) { |
| 2840 | synchronize_sched(); |
| 2841 | atomic_long_inc(&rsp->expedited_wrap); |
| 2842 | return; |
| 2843 | } |
| 2844 | |
| 2845 | /* |
| 2846 | * Take a ticket. Note that atomic_inc_return() implies a |
| 2847 | * full memory barrier. |
| 2848 | */ |
| 2849 | snap = atomic_long_inc_return(&rsp->expedited_start); |
| 2850 | firstsnap = snap; |
| 2851 | get_online_cpus(); |
| 2852 | WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id())); |
| 2853 | |
| 2854 | /* |
| 2855 | * Each pass through the following loop attempts to force a |
| 2856 | * context switch on each CPU. |
| 2857 | */ |
| 2858 | while (try_stop_cpus(cpu_online_mask, |
| 2859 | synchronize_sched_expedited_cpu_stop, |
| 2860 | NULL) == -EAGAIN) { |
| 2861 | put_online_cpus(); |
| 2862 | atomic_long_inc(&rsp->expedited_tryfail); |
| 2863 | |
| 2864 | /* Check to see if someone else did our work for us. */ |
| 2865 | s = atomic_long_read(&rsp->expedited_done); |
| 2866 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
| 2867 | /* ensure test happens before caller kfree */ |
| 2868 | smp_mb__before_atomic_inc(); /* ^^^ */ |
| 2869 | atomic_long_inc(&rsp->expedited_workdone1); |
| 2870 | return; |
| 2871 | } |
| 2872 | |
| 2873 | /* No joy, try again later. Or just synchronize_sched(). */ |
| 2874 | if (trycount++ < 10) { |
| 2875 | udelay(trycount * num_online_cpus()); |
| 2876 | } else { |
| 2877 | wait_rcu_gp(call_rcu_sched); |
| 2878 | atomic_long_inc(&rsp->expedited_normal); |
| 2879 | return; |
| 2880 | } |
| 2881 | |
| 2882 | /* Recheck to see if someone else did our work for us. */ |
| 2883 | s = atomic_long_read(&rsp->expedited_done); |
| 2884 | if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) { |
| 2885 | /* ensure test happens before caller kfree */ |
| 2886 | smp_mb__before_atomic_inc(); /* ^^^ */ |
| 2887 | atomic_long_inc(&rsp->expedited_workdone2); |
| 2888 | return; |
| 2889 | } |
| 2890 | |
| 2891 | /* |
| 2892 | * Refetching sync_sched_expedited_started allows later |
| 2893 | * callers to piggyback on our grace period. We retry |
| 2894 | * after they started, so our grace period works for them, |
| 2895 | * and they started after our first try, so their grace |
| 2896 | * period works for us. |
| 2897 | */ |
| 2898 | get_online_cpus(); |
| 2899 | snap = atomic_long_read(&rsp->expedited_start); |
| 2900 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
| 2901 | } |
| 2902 | atomic_long_inc(&rsp->expedited_stoppedcpus); |
| 2903 | |
| 2904 | /* |
| 2905 | * Everyone up to our most recent fetch is covered by our grace |
| 2906 | * period. Update the counter, but only if our work is still |
| 2907 | * relevant -- which it won't be if someone who started later |
| 2908 | * than we did already did their update. |
| 2909 | */ |
| 2910 | do { |
| 2911 | atomic_long_inc(&rsp->expedited_done_tries); |
| 2912 | s = atomic_long_read(&rsp->expedited_done); |
| 2913 | if (ULONG_CMP_GE((ulong)s, (ulong)snap)) { |
| 2914 | /* ensure test happens before caller kfree */ |
| 2915 | smp_mb__before_atomic_inc(); /* ^^^ */ |
| 2916 | atomic_long_inc(&rsp->expedited_done_lost); |
| 2917 | break; |
| 2918 | } |
| 2919 | } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s); |
| 2920 | atomic_long_inc(&rsp->expedited_done_exit); |
| 2921 | |
| 2922 | put_online_cpus(); |
| 2923 | } |
| 2924 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); |
| 2925 | |
| 2926 | /* |
| 2927 | * Check to see if there is any immediate RCU-related work to be done |
| 2928 | * by the current CPU, for the specified type of RCU, returning 1 if so. |
| 2929 | * The checks are in order of increasing expense: checks that can be |
| 2930 | * carried out against CPU-local state are performed first. However, |
| 2931 | * we must check for CPU stalls first, else we might not get a chance. |
| 2932 | */ |
| 2933 | static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) |
| 2934 | { |
| 2935 | struct rcu_node *rnp = rdp->mynode; |
| 2936 | |
| 2937 | rdp->n_rcu_pending++; |
| 2938 | |
| 2939 | /* Check for CPU stalls, if enabled. */ |
| 2940 | check_cpu_stall(rsp, rdp); |
| 2941 | |
| 2942 | /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ |
| 2943 | if (rcu_nohz_full_cpu(rsp)) |
| 2944 | return 0; |
| 2945 | |
| 2946 | /* Is the RCU core waiting for a quiescent state from this CPU? */ |
| 2947 | if (rcu_scheduler_fully_active && |
| 2948 | rdp->qs_pending && !rdp->passed_quiesce) { |
| 2949 | rdp->n_rp_qs_pending++; |
| 2950 | } else if (rdp->qs_pending && rdp->passed_quiesce) { |
| 2951 | rdp->n_rp_report_qs++; |
| 2952 | return 1; |
| 2953 | } |
| 2954 | |
| 2955 | /* Does this CPU have callbacks ready to invoke? */ |
| 2956 | if (cpu_has_callbacks_ready_to_invoke(rdp)) { |
| 2957 | rdp->n_rp_cb_ready++; |
| 2958 | return 1; |
| 2959 | } |
| 2960 | |
| 2961 | /* Has RCU gone idle with this CPU needing another grace period? */ |
| 2962 | if (cpu_needs_another_gp(rsp, rdp)) { |
| 2963 | rdp->n_rp_cpu_needs_gp++; |
| 2964 | return 1; |
| 2965 | } |
| 2966 | |
| 2967 | /* Has another RCU grace period completed? */ |
| 2968 | if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ |
| 2969 | rdp->n_rp_gp_completed++; |
| 2970 | return 1; |
| 2971 | } |
| 2972 | |
| 2973 | /* Has a new RCU grace period started? */ |
| 2974 | if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */ |
| 2975 | rdp->n_rp_gp_started++; |
| 2976 | return 1; |
| 2977 | } |
| 2978 | |
| 2979 | /* Does this CPU need a deferred NOCB wakeup? */ |
| 2980 | if (rcu_nocb_need_deferred_wakeup(rdp)) { |
| 2981 | rdp->n_rp_nocb_defer_wakeup++; |
| 2982 | return 1; |
| 2983 | } |
| 2984 | |
| 2985 | /* nothing to do */ |
| 2986 | rdp->n_rp_need_nothing++; |
| 2987 | return 0; |
| 2988 | } |
| 2989 | |
| 2990 | /* |
| 2991 | * Check to see if there is any immediate RCU-related work to be done |
| 2992 | * by the current CPU, returning 1 if so. This function is part of the |
| 2993 | * RCU implementation; it is -not- an exported member of the RCU API. |
| 2994 | */ |
| 2995 | static int rcu_pending(int cpu) |
| 2996 | { |
| 2997 | struct rcu_state *rsp; |
| 2998 | |
| 2999 | for_each_rcu_flavor(rsp) |
| 3000 | if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu))) |
| 3001 | return 1; |
| 3002 | return 0; |
| 3003 | } |
| 3004 | |
| 3005 | /* |
| 3006 | * Return true if the specified CPU has any callback. If all_lazy is |
| 3007 | * non-NULL, store an indication of whether all callbacks are lazy. |
| 3008 | * (If there are no callbacks, all of them are deemed to be lazy.) |
| 3009 | */ |
| 3010 | static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy) |
| 3011 | { |
| 3012 | bool al = true; |
| 3013 | bool hc = false; |
| 3014 | struct rcu_data *rdp; |
| 3015 | struct rcu_state *rsp; |
| 3016 | |
| 3017 | for_each_rcu_flavor(rsp) { |
| 3018 | rdp = per_cpu_ptr(rsp->rda, cpu); |
| 3019 | if (!rdp->nxtlist) |
| 3020 | continue; |
| 3021 | hc = true; |
| 3022 | if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { |
| 3023 | al = false; |
| 3024 | break; |
| 3025 | } |
| 3026 | } |
| 3027 | if (all_lazy) |
| 3028 | *all_lazy = al; |
| 3029 | return hc; |
| 3030 | } |
| 3031 | |
| 3032 | /* |
| 3033 | * Helper function for _rcu_barrier() tracing. If tracing is disabled, |
| 3034 | * the compiler is expected to optimize this away. |
| 3035 | */ |
| 3036 | static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, |
| 3037 | int cpu, unsigned long done) |
| 3038 | { |
| 3039 | trace_rcu_barrier(rsp->name, s, cpu, |
| 3040 | atomic_read(&rsp->barrier_cpu_count), done); |
| 3041 | } |
| 3042 | |
| 3043 | /* |
| 3044 | * RCU callback function for _rcu_barrier(). If we are last, wake |
| 3045 | * up the task executing _rcu_barrier(). |
| 3046 | */ |
| 3047 | static void rcu_barrier_callback(struct rcu_head *rhp) |
| 3048 | { |
| 3049 | struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); |
| 3050 | struct rcu_state *rsp = rdp->rsp; |
| 3051 | |
| 3052 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { |
| 3053 | _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done); |
| 3054 | complete(&rsp->barrier_completion); |
| 3055 | } else { |
| 3056 | _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done); |
| 3057 | } |
| 3058 | } |
| 3059 | |
| 3060 | /* |
| 3061 | * Called with preemption disabled, and from cross-cpu IRQ context. |
| 3062 | */ |
| 3063 | static void rcu_barrier_func(void *type) |
| 3064 | { |
| 3065 | struct rcu_state *rsp = type; |
| 3066 | struct rcu_data *rdp = __this_cpu_ptr(rsp->rda); |
| 3067 | |
| 3068 | _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done); |
| 3069 | atomic_inc(&rsp->barrier_cpu_count); |
| 3070 | rsp->call(&rdp->barrier_head, rcu_barrier_callback); |
| 3071 | } |
| 3072 | |
| 3073 | /* |
| 3074 | * Orchestrate the specified type of RCU barrier, waiting for all |
| 3075 | * RCU callbacks of the specified type to complete. |
| 3076 | */ |
| 3077 | static void _rcu_barrier(struct rcu_state *rsp) |
| 3078 | { |
| 3079 | int cpu; |
| 3080 | struct rcu_data *rdp; |
| 3081 | unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done); |
| 3082 | unsigned long snap_done; |
| 3083 | |
| 3084 | _rcu_barrier_trace(rsp, "Begin", -1, snap); |
| 3085 | |
| 3086 | /* Take mutex to serialize concurrent rcu_barrier() requests. */ |
| 3087 | mutex_lock(&rsp->barrier_mutex); |
| 3088 | |
| 3089 | /* |
| 3090 | * Ensure that all prior references, including to ->n_barrier_done, |
| 3091 | * are ordered before the _rcu_barrier() machinery. |
| 3092 | */ |
| 3093 | smp_mb(); /* See above block comment. */ |
| 3094 | |
| 3095 | /* |
| 3096 | * Recheck ->n_barrier_done to see if others did our work for us. |
| 3097 | * This means checking ->n_barrier_done for an even-to-odd-to-even |
| 3098 | * transition. The "if" expression below therefore rounds the old |
| 3099 | * value up to the next even number and adds two before comparing. |
| 3100 | */ |
| 3101 | snap_done = rsp->n_barrier_done; |
| 3102 | _rcu_barrier_trace(rsp, "Check", -1, snap_done); |
| 3103 | |
| 3104 | /* |
| 3105 | * If the value in snap is odd, we needed to wait for the current |
| 3106 | * rcu_barrier() to complete, then wait for the next one, in other |
| 3107 | * words, we need the value of snap_done to be three larger than |
| 3108 | * the value of snap. On the other hand, if the value in snap is |
| 3109 | * even, we only had to wait for the next rcu_barrier() to complete, |
| 3110 | * in other words, we need the value of snap_done to be only two |
| 3111 | * greater than the value of snap. The "(snap + 3) & ~0x1" computes |
| 3112 | * this for us (thank you, Linus!). |
| 3113 | */ |
| 3114 | if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) { |
| 3115 | _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done); |
| 3116 | smp_mb(); /* caller's subsequent code after above check. */ |
| 3117 | mutex_unlock(&rsp->barrier_mutex); |
| 3118 | return; |
| 3119 | } |
| 3120 | |
| 3121 | /* |
| 3122 | * Increment ->n_barrier_done to avoid duplicate work. Use |
| 3123 | * ACCESS_ONCE() to prevent the compiler from speculating |
| 3124 | * the increment to precede the early-exit check. |
| 3125 | */ |
| 3126 | ACCESS_ONCE(rsp->n_barrier_done)++; |
| 3127 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1); |
| 3128 | _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done); |
| 3129 | smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */ |
| 3130 | |
| 3131 | /* |
| 3132 | * Initialize the count to one rather than to zero in order to |
| 3133 | * avoid a too-soon return to zero in case of a short grace period |
| 3134 | * (or preemption of this task). Exclude CPU-hotplug operations |
| 3135 | * to ensure that no offline CPU has callbacks queued. |
| 3136 | */ |
| 3137 | init_completion(&rsp->barrier_completion); |
| 3138 | atomic_set(&rsp->barrier_cpu_count, 1); |
| 3139 | get_online_cpus(); |
| 3140 | |
| 3141 | /* |
| 3142 | * Force each CPU with callbacks to register a new callback. |
| 3143 | * When that callback is invoked, we will know that all of the |
| 3144 | * corresponding CPU's preceding callbacks have been invoked. |
| 3145 | */ |
| 3146 | for_each_possible_cpu(cpu) { |
| 3147 | if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) |
| 3148 | continue; |
| 3149 | rdp = per_cpu_ptr(rsp->rda, cpu); |
| 3150 | if (rcu_is_nocb_cpu(cpu)) { |
| 3151 | _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, |
| 3152 | rsp->n_barrier_done); |
| 3153 | atomic_inc(&rsp->barrier_cpu_count); |
| 3154 | __call_rcu(&rdp->barrier_head, rcu_barrier_callback, |
| 3155 | rsp, cpu, 0); |
| 3156 | } else if (ACCESS_ONCE(rdp->qlen)) { |
| 3157 | _rcu_barrier_trace(rsp, "OnlineQ", cpu, |
| 3158 | rsp->n_barrier_done); |
| 3159 | smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); |
| 3160 | } else { |
| 3161 | _rcu_barrier_trace(rsp, "OnlineNQ", cpu, |
| 3162 | rsp->n_barrier_done); |
| 3163 | } |
| 3164 | } |
| 3165 | put_online_cpus(); |
| 3166 | |
| 3167 | /* |
| 3168 | * Now that we have an rcu_barrier_callback() callback on each |
| 3169 | * CPU, and thus each counted, remove the initial count. |
| 3170 | */ |
| 3171 | if (atomic_dec_and_test(&rsp->barrier_cpu_count)) |
| 3172 | complete(&rsp->barrier_completion); |
| 3173 | |
| 3174 | /* Increment ->n_barrier_done to prevent duplicate work. */ |
| 3175 | smp_mb(); /* Keep increment after above mechanism. */ |
| 3176 | ACCESS_ONCE(rsp->n_barrier_done)++; |
| 3177 | WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0); |
| 3178 | _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done); |
| 3179 | smp_mb(); /* Keep increment before caller's subsequent code. */ |
| 3180 | |
| 3181 | /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ |
| 3182 | wait_for_completion(&rsp->barrier_completion); |
| 3183 | |
| 3184 | /* Other rcu_barrier() invocations can now safely proceed. */ |
| 3185 | mutex_unlock(&rsp->barrier_mutex); |
| 3186 | } |
| 3187 | |
| 3188 | /** |
| 3189 | * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. |
| 3190 | */ |
| 3191 | void rcu_barrier_bh(void) |
| 3192 | { |
| 3193 | _rcu_barrier(&rcu_bh_state); |
| 3194 | } |
| 3195 | EXPORT_SYMBOL_GPL(rcu_barrier_bh); |
| 3196 | |
| 3197 | /** |
| 3198 | * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. |
| 3199 | */ |
| 3200 | void rcu_barrier_sched(void) |
| 3201 | { |
| 3202 | _rcu_barrier(&rcu_sched_state); |
| 3203 | } |
| 3204 | EXPORT_SYMBOL_GPL(rcu_barrier_sched); |
| 3205 | |
| 3206 | /* |
| 3207 | * Do boot-time initialization of a CPU's per-CPU RCU data. |
| 3208 | */ |
| 3209 | static void __init |
| 3210 | rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) |
| 3211 | { |
| 3212 | unsigned long flags; |
| 3213 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
| 3214 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 3215 | |
| 3216 | /* Set up local state, ensuring consistent view of global state. */ |
| 3217 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 3218 | rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo); |
| 3219 | init_callback_list(rdp); |
| 3220 | rdp->qlen_lazy = 0; |
| 3221 | ACCESS_ONCE(rdp->qlen) = 0; |
| 3222 | rdp->dynticks = &per_cpu(rcu_dynticks, cpu); |
| 3223 | WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); |
| 3224 | WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1); |
| 3225 | rdp->cpu = cpu; |
| 3226 | rdp->rsp = rsp; |
| 3227 | rcu_boot_init_nocb_percpu_data(rdp); |
| 3228 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 3229 | } |
| 3230 | |
| 3231 | /* |
| 3232 | * Initialize a CPU's per-CPU RCU data. Note that only one online or |
| 3233 | * offline event can be happening at a given time. Note also that we |
| 3234 | * can accept some slop in the rsp->completed access due to the fact |
| 3235 | * that this CPU cannot possibly have any RCU callbacks in flight yet. |
| 3236 | */ |
| 3237 | static void |
| 3238 | rcu_init_percpu_data(int cpu, struct rcu_state *rsp) |
| 3239 | { |
| 3240 | unsigned long flags; |
| 3241 | unsigned long mask; |
| 3242 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
| 3243 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 3244 | |
| 3245 | /* Exclude new grace periods. */ |
| 3246 | mutex_lock(&rsp->onoff_mutex); |
| 3247 | |
| 3248 | /* Set up local state, ensuring consistent view of global state. */ |
| 3249 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 3250 | rdp->beenonline = 1; /* We have now been online. */ |
| 3251 | rdp->qlen_last_fqs_check = 0; |
| 3252 | rdp->n_force_qs_snap = rsp->n_force_qs; |
| 3253 | rdp->blimit = blimit; |
| 3254 | init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ |
| 3255 | rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; |
| 3256 | rcu_sysidle_init_percpu_data(rdp->dynticks); |
| 3257 | atomic_set(&rdp->dynticks->dynticks, |
| 3258 | (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1); |
| 3259 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 3260 | |
| 3261 | /* Add CPU to rcu_node bitmasks. */ |
| 3262 | rnp = rdp->mynode; |
| 3263 | mask = rdp->grpmask; |
| 3264 | do { |
| 3265 | /* Exclude any attempts to start a new GP on small systems. */ |
| 3266 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
| 3267 | rnp->qsmaskinit |= mask; |
| 3268 | mask = rnp->grpmask; |
| 3269 | if (rnp == rdp->mynode) { |
| 3270 | /* |
| 3271 | * If there is a grace period in progress, we will |
| 3272 | * set up to wait for it next time we run the |
| 3273 | * RCU core code. |
| 3274 | */ |
| 3275 | rdp->gpnum = rnp->completed; |
| 3276 | rdp->completed = rnp->completed; |
| 3277 | rdp->passed_quiesce = 0; |
| 3278 | rdp->qs_pending = 0; |
| 3279 | trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); |
| 3280 | } |
| 3281 | raw_spin_unlock(&rnp->lock); /* irqs already disabled. */ |
| 3282 | rnp = rnp->parent; |
| 3283 | } while (rnp != NULL && !(rnp->qsmaskinit & mask)); |
| 3284 | local_irq_restore(flags); |
| 3285 | |
| 3286 | mutex_unlock(&rsp->onoff_mutex); |
| 3287 | } |
| 3288 | |
| 3289 | static void rcu_prepare_cpu(int cpu) |
| 3290 | { |
| 3291 | struct rcu_state *rsp; |
| 3292 | |
| 3293 | for_each_rcu_flavor(rsp) |
| 3294 | rcu_init_percpu_data(cpu, rsp); |
| 3295 | } |
| 3296 | |
| 3297 | /* |
| 3298 | * Handle CPU online/offline notification events. |
| 3299 | */ |
| 3300 | static int rcu_cpu_notify(struct notifier_block *self, |
| 3301 | unsigned long action, void *hcpu) |
| 3302 | { |
| 3303 | long cpu = (long)hcpu; |
| 3304 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); |
| 3305 | struct rcu_node *rnp = rdp->mynode; |
| 3306 | struct rcu_state *rsp; |
| 3307 | |
| 3308 | trace_rcu_utilization(TPS("Start CPU hotplug")); |
| 3309 | switch (action) { |
| 3310 | case CPU_UP_PREPARE: |
| 3311 | case CPU_UP_PREPARE_FROZEN: |
| 3312 | rcu_prepare_cpu(cpu); |
| 3313 | rcu_prepare_kthreads(cpu); |
| 3314 | break; |
| 3315 | case CPU_ONLINE: |
| 3316 | case CPU_DOWN_FAILED: |
| 3317 | rcu_boost_kthread_setaffinity(rnp, -1); |
| 3318 | break; |
| 3319 | case CPU_DOWN_PREPARE: |
| 3320 | rcu_boost_kthread_setaffinity(rnp, cpu); |
| 3321 | break; |
| 3322 | case CPU_DYING: |
| 3323 | case CPU_DYING_FROZEN: |
| 3324 | for_each_rcu_flavor(rsp) |
| 3325 | rcu_cleanup_dying_cpu(rsp); |
| 3326 | break; |
| 3327 | case CPU_DEAD: |
| 3328 | case CPU_DEAD_FROZEN: |
| 3329 | case CPU_UP_CANCELED: |
| 3330 | case CPU_UP_CANCELED_FROZEN: |
| 3331 | for_each_rcu_flavor(rsp) |
| 3332 | rcu_cleanup_dead_cpu(cpu, rsp); |
| 3333 | break; |
| 3334 | default: |
| 3335 | break; |
| 3336 | } |
| 3337 | trace_rcu_utilization(TPS("End CPU hotplug")); |
| 3338 | return NOTIFY_OK; |
| 3339 | } |
| 3340 | |
| 3341 | static int rcu_pm_notify(struct notifier_block *self, |
| 3342 | unsigned long action, void *hcpu) |
| 3343 | { |
| 3344 | switch (action) { |
| 3345 | case PM_HIBERNATION_PREPARE: |
| 3346 | case PM_SUSPEND_PREPARE: |
| 3347 | if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ |
| 3348 | rcu_expedited = 1; |
| 3349 | break; |
| 3350 | case PM_POST_HIBERNATION: |
| 3351 | case PM_POST_SUSPEND: |
| 3352 | rcu_expedited = 0; |
| 3353 | break; |
| 3354 | default: |
| 3355 | break; |
| 3356 | } |
| 3357 | return NOTIFY_OK; |
| 3358 | } |
| 3359 | |
| 3360 | /* |
| 3361 | * Spawn the kthread that handles this RCU flavor's grace periods. |
| 3362 | */ |
| 3363 | static int __init rcu_spawn_gp_kthread(void) |
| 3364 | { |
| 3365 | unsigned long flags; |
| 3366 | struct rcu_node *rnp; |
| 3367 | struct rcu_state *rsp; |
| 3368 | struct task_struct *t; |
| 3369 | |
| 3370 | for_each_rcu_flavor(rsp) { |
| 3371 | t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name); |
| 3372 | BUG_ON(IS_ERR(t)); |
| 3373 | rnp = rcu_get_root(rsp); |
| 3374 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 3375 | rsp->gp_kthread = t; |
| 3376 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 3377 | rcu_spawn_nocb_kthreads(rsp); |
| 3378 | } |
| 3379 | return 0; |
| 3380 | } |
| 3381 | early_initcall(rcu_spawn_gp_kthread); |
| 3382 | |
| 3383 | /* |
| 3384 | * This function is invoked towards the end of the scheduler's initialization |
| 3385 | * process. Before this is called, the idle task might contain |
| 3386 | * RCU read-side critical sections (during which time, this idle |
| 3387 | * task is booting the system). After this function is called, the |
| 3388 | * idle tasks are prohibited from containing RCU read-side critical |
| 3389 | * sections. This function also enables RCU lockdep checking. |
| 3390 | */ |
| 3391 | void rcu_scheduler_starting(void) |
| 3392 | { |
| 3393 | WARN_ON(num_online_cpus() != 1); |
| 3394 | WARN_ON(nr_context_switches() > 0); |
| 3395 | rcu_scheduler_active = 1; |
| 3396 | } |
| 3397 | |
| 3398 | /* |
| 3399 | * Compute the per-level fanout, either using the exact fanout specified |
| 3400 | * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT. |
| 3401 | */ |
| 3402 | #ifdef CONFIG_RCU_FANOUT_EXACT |
| 3403 | static void __init rcu_init_levelspread(struct rcu_state *rsp) |
| 3404 | { |
| 3405 | int i; |
| 3406 | |
| 3407 | rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; |
| 3408 | for (i = rcu_num_lvls - 2; i >= 0; i--) |
| 3409 | rsp->levelspread[i] = CONFIG_RCU_FANOUT; |
| 3410 | } |
| 3411 | #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */ |
| 3412 | static void __init rcu_init_levelspread(struct rcu_state *rsp) |
| 3413 | { |
| 3414 | int ccur; |
| 3415 | int cprv; |
| 3416 | int i; |
| 3417 | |
| 3418 | cprv = nr_cpu_ids; |
| 3419 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
| 3420 | ccur = rsp->levelcnt[i]; |
| 3421 | rsp->levelspread[i] = (cprv + ccur - 1) / ccur; |
| 3422 | cprv = ccur; |
| 3423 | } |
| 3424 | } |
| 3425 | #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */ |
| 3426 | |
| 3427 | /* |
| 3428 | * Helper function for rcu_init() that initializes one rcu_state structure. |
| 3429 | */ |
| 3430 | static void __init rcu_init_one(struct rcu_state *rsp, |
| 3431 | struct rcu_data __percpu *rda) |
| 3432 | { |
| 3433 | static char *buf[] = { "rcu_node_0", |
| 3434 | "rcu_node_1", |
| 3435 | "rcu_node_2", |
| 3436 | "rcu_node_3" }; /* Match MAX_RCU_LVLS */ |
| 3437 | static char *fqs[] = { "rcu_node_fqs_0", |
| 3438 | "rcu_node_fqs_1", |
| 3439 | "rcu_node_fqs_2", |
| 3440 | "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */ |
| 3441 | int cpustride = 1; |
| 3442 | int i; |
| 3443 | int j; |
| 3444 | struct rcu_node *rnp; |
| 3445 | |
| 3446 | BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ |
| 3447 | |
| 3448 | /* Silence gcc 4.8 warning about array index out of range. */ |
| 3449 | if (rcu_num_lvls > RCU_NUM_LVLS) |
| 3450 | panic("rcu_init_one: rcu_num_lvls overflow"); |
| 3451 | |
| 3452 | /* Initialize the level-tracking arrays. */ |
| 3453 | |
| 3454 | for (i = 0; i < rcu_num_lvls; i++) |
| 3455 | rsp->levelcnt[i] = num_rcu_lvl[i]; |
| 3456 | for (i = 1; i < rcu_num_lvls; i++) |
| 3457 | rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1]; |
| 3458 | rcu_init_levelspread(rsp); |
| 3459 | |
| 3460 | /* Initialize the elements themselves, starting from the leaves. */ |
| 3461 | |
| 3462 | for (i = rcu_num_lvls - 1; i >= 0; i--) { |
| 3463 | cpustride *= rsp->levelspread[i]; |
| 3464 | rnp = rsp->level[i]; |
| 3465 | for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) { |
| 3466 | raw_spin_lock_init(&rnp->lock); |
| 3467 | lockdep_set_class_and_name(&rnp->lock, |
| 3468 | &rcu_node_class[i], buf[i]); |
| 3469 | raw_spin_lock_init(&rnp->fqslock); |
| 3470 | lockdep_set_class_and_name(&rnp->fqslock, |
| 3471 | &rcu_fqs_class[i], fqs[i]); |
| 3472 | rnp->gpnum = rsp->gpnum; |
| 3473 | rnp->completed = rsp->completed; |
| 3474 | rnp->qsmask = 0; |
| 3475 | rnp->qsmaskinit = 0; |
| 3476 | rnp->grplo = j * cpustride; |
| 3477 | rnp->grphi = (j + 1) * cpustride - 1; |
| 3478 | if (rnp->grphi >= nr_cpu_ids) |
| 3479 | rnp->grphi = nr_cpu_ids - 1; |
| 3480 | if (i == 0) { |
| 3481 | rnp->grpnum = 0; |
| 3482 | rnp->grpmask = 0; |
| 3483 | rnp->parent = NULL; |
| 3484 | } else { |
| 3485 | rnp->grpnum = j % rsp->levelspread[i - 1]; |
| 3486 | rnp->grpmask = 1UL << rnp->grpnum; |
| 3487 | rnp->parent = rsp->level[i - 1] + |
| 3488 | j / rsp->levelspread[i - 1]; |
| 3489 | } |
| 3490 | rnp->level = i; |
| 3491 | INIT_LIST_HEAD(&rnp->blkd_tasks); |
| 3492 | rcu_init_one_nocb(rnp); |
| 3493 | } |
| 3494 | } |
| 3495 | |
| 3496 | rsp->rda = rda; |
| 3497 | init_waitqueue_head(&rsp->gp_wq); |
| 3498 | rnp = rsp->level[rcu_num_lvls - 1]; |
| 3499 | for_each_possible_cpu(i) { |
| 3500 | while (i > rnp->grphi) |
| 3501 | rnp++; |
| 3502 | per_cpu_ptr(rsp->rda, i)->mynode = rnp; |
| 3503 | rcu_boot_init_percpu_data(i, rsp); |
| 3504 | } |
| 3505 | list_add(&rsp->flavors, &rcu_struct_flavors); |
| 3506 | } |
| 3507 | |
| 3508 | /* |
| 3509 | * Compute the rcu_node tree geometry from kernel parameters. This cannot |
| 3510 | * replace the definitions in tree.h because those are needed to size |
| 3511 | * the ->node array in the rcu_state structure. |
| 3512 | */ |
| 3513 | static void __init rcu_init_geometry(void) |
| 3514 | { |
| 3515 | ulong d; |
| 3516 | int i; |
| 3517 | int j; |
| 3518 | int n = nr_cpu_ids; |
| 3519 | int rcu_capacity[MAX_RCU_LVLS + 1]; |
| 3520 | |
| 3521 | /* |
| 3522 | * Initialize any unspecified boot parameters. |
| 3523 | * The default values of jiffies_till_first_fqs and |
| 3524 | * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS |
| 3525 | * value, which is a function of HZ, then adding one for each |
| 3526 | * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. |
| 3527 | */ |
| 3528 | d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; |
| 3529 | if (jiffies_till_first_fqs == ULONG_MAX) |
| 3530 | jiffies_till_first_fqs = d; |
| 3531 | if (jiffies_till_next_fqs == ULONG_MAX) |
| 3532 | jiffies_till_next_fqs = d; |
| 3533 | |
| 3534 | /* If the compile-time values are accurate, just leave. */ |
| 3535 | if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF && |
| 3536 | nr_cpu_ids == NR_CPUS) |
| 3537 | return; |
| 3538 | pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", |
| 3539 | rcu_fanout_leaf, nr_cpu_ids); |
| 3540 | |
| 3541 | /* |
| 3542 | * Compute number of nodes that can be handled an rcu_node tree |
| 3543 | * with the given number of levels. Setting rcu_capacity[0] makes |
| 3544 | * some of the arithmetic easier. |
| 3545 | */ |
| 3546 | rcu_capacity[0] = 1; |
| 3547 | rcu_capacity[1] = rcu_fanout_leaf; |
| 3548 | for (i = 2; i <= MAX_RCU_LVLS; i++) |
| 3549 | rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT; |
| 3550 | |
| 3551 | /* |
| 3552 | * The boot-time rcu_fanout_leaf parameter is only permitted |
| 3553 | * to increase the leaf-level fanout, not decrease it. Of course, |
| 3554 | * the leaf-level fanout cannot exceed the number of bits in |
| 3555 | * the rcu_node masks. Finally, the tree must be able to accommodate |
| 3556 | * the configured number of CPUs. Complain and fall back to the |
| 3557 | * compile-time values if these limits are exceeded. |
| 3558 | */ |
| 3559 | if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF || |
| 3560 | rcu_fanout_leaf > sizeof(unsigned long) * 8 || |
| 3561 | n > rcu_capacity[MAX_RCU_LVLS]) { |
| 3562 | WARN_ON(1); |
| 3563 | return; |
| 3564 | } |
| 3565 | |
| 3566 | /* Calculate the number of rcu_nodes at each level of the tree. */ |
| 3567 | for (i = 1; i <= MAX_RCU_LVLS; i++) |
| 3568 | if (n <= rcu_capacity[i]) { |
| 3569 | for (j = 0; j <= i; j++) |
| 3570 | num_rcu_lvl[j] = |
| 3571 | DIV_ROUND_UP(n, rcu_capacity[i - j]); |
| 3572 | rcu_num_lvls = i; |
| 3573 | for (j = i + 1; j <= MAX_RCU_LVLS; j++) |
| 3574 | num_rcu_lvl[j] = 0; |
| 3575 | break; |
| 3576 | } |
| 3577 | |
| 3578 | /* Calculate the total number of rcu_node structures. */ |
| 3579 | rcu_num_nodes = 0; |
| 3580 | for (i = 0; i <= MAX_RCU_LVLS; i++) |
| 3581 | rcu_num_nodes += num_rcu_lvl[i]; |
| 3582 | rcu_num_nodes -= n; |
| 3583 | } |
| 3584 | |
| 3585 | void __init rcu_init(void) |
| 3586 | { |
| 3587 | int cpu; |
| 3588 | |
| 3589 | rcu_bootup_announce(); |
| 3590 | rcu_init_geometry(); |
| 3591 | rcu_init_one(&rcu_bh_state, &rcu_bh_data); |
| 3592 | rcu_init_one(&rcu_sched_state, &rcu_sched_data); |
| 3593 | __rcu_init_preempt(); |
| 3594 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
| 3595 | |
| 3596 | /* |
| 3597 | * We don't need protection against CPU-hotplug here because |
| 3598 | * this is called early in boot, before either interrupts |
| 3599 | * or the scheduler are operational. |
| 3600 | */ |
| 3601 | cpu_notifier(rcu_cpu_notify, 0); |
| 3602 | pm_notifier(rcu_pm_notify, 0); |
| 3603 | for_each_online_cpu(cpu) |
| 3604 | rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu); |
| 3605 | } |
| 3606 | |
| 3607 | #include "tree_plugin.h" |