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_wait.h> 36 #include <linux/interrupt.h> 37 #include <linux/sched.h> 38 #include <linux/sched/debug.h> 39 #include <linux/nmi.h> 40 #include <linux/atomic.h> 41 #include <linux/bitops.h> 42 #include <linux/export.h> 43 #include <linux/completion.h> 44 #include <linux/moduleparam.h> 45 #include <linux/percpu.h> 46 #include <linux/notifier.h> 47 #include <linux/cpu.h> 48 #include <linux/mutex.h> 49 #include <linux/time.h> 50 #include <linux/kernel_stat.h> 51 #include <linux/wait.h> 52 #include <linux/kthread.h> 53 #include <uapi/linux/sched/types.h> 54 #include <linux/prefetch.h> 55 #include <linux/delay.h> 56 #include <linux/stop_machine.h> 57 #include <linux/random.h> 58 #include <linux/trace_events.h> 59 #include <linux/suspend.h> 60 61 #include "tree.h" 62 #include "rcu.h" 63 64 #ifdef MODULE_PARAM_PREFIX 65 #undef MODULE_PARAM_PREFIX 66 #endif 67 #define MODULE_PARAM_PREFIX "rcutree." 68 69 /* Data structures. */ 70 71 /* 72 * In order to export the rcu_state name to the tracing tools, it 73 * needs to be added in the __tracepoint_string section. 74 * This requires defining a separate variable tp_<sname>_varname 75 * that points to the string being used, and this will allow 76 * the tracing userspace tools to be able to decipher the string 77 * address to the matching string. 78 */ 79 #ifdef CONFIG_TRACING 80 # define DEFINE_RCU_TPS(sname) \ 81 static char sname##_varname[] = #sname; \ 82 static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; 83 # define RCU_STATE_NAME(sname) sname##_varname 84 #else 85 # define DEFINE_RCU_TPS(sname) 86 # define RCU_STATE_NAME(sname) __stringify(sname) 87 #endif 88 89 #define RCU_STATE_INITIALIZER(sname, sabbr, cr) \ 90 DEFINE_RCU_TPS(sname) \ 91 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \ 92 struct rcu_state sname##_state = { \ 93 .level = { &sname##_state.node[0] }, \ 94 .rda = &sname##_data, \ 95 .call = cr, \ 96 .gp_state = RCU_GP_IDLE, \ 97 .gpnum = 0UL - 300UL, \ 98 .completed = 0UL - 300UL, \ 99 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \ 100 .orphan_nxttail = &sname##_state.orphan_nxtlist, \ 101 .orphan_donetail = &sname##_state.orphan_donelist, \ 102 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \ 103 .name = RCU_STATE_NAME(sname), \ 104 .abbr = sabbr, \ 105 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \ 106 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \ 107 } 108 109 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched); 110 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh); 111 112 static struct rcu_state *const rcu_state_p; 113 LIST_HEAD(rcu_struct_flavors); 114 115 /* Dump rcu_node combining tree at boot to verify correct setup. */ 116 static bool dump_tree; 117 module_param(dump_tree, bool, 0444); 118 /* Control rcu_node-tree auto-balancing at boot time. */ 119 static bool rcu_fanout_exact; 120 module_param(rcu_fanout_exact, bool, 0444); 121 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */ 122 static int rcu_fanout_leaf = RCU_FANOUT_LEAF; 123 module_param(rcu_fanout_leaf, int, 0444); 124 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS; 125 /* Number of rcu_nodes at specified level. */ 126 static int num_rcu_lvl[] = NUM_RCU_LVL_INIT; 127 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */ 128 /* panic() on RCU Stall sysctl. */ 129 int sysctl_panic_on_rcu_stall __read_mostly; 130 131 /* 132 * The rcu_scheduler_active variable is initialized to the value 133 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the 134 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE, 135 * RCU can assume that there is but one task, allowing RCU to (for example) 136 * optimize synchronize_rcu() to a simple barrier(). When this variable 137 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required 138 * to detect real grace periods. This variable is also used to suppress 139 * boot-time false positives from lockdep-RCU error checking. Finally, it 140 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU 141 * is fully initialized, including all of its kthreads having been spawned. 142 */ 143 int rcu_scheduler_active __read_mostly; 144 EXPORT_SYMBOL_GPL(rcu_scheduler_active); 145 146 /* 147 * The rcu_scheduler_fully_active variable transitions from zero to one 148 * during the early_initcall() processing, which is after the scheduler 149 * is capable of creating new tasks. So RCU processing (for example, 150 * creating tasks for RCU priority boosting) must be delayed until after 151 * rcu_scheduler_fully_active transitions from zero to one. We also 152 * currently delay invocation of any RCU callbacks until after this point. 153 * 154 * It might later prove better for people registering RCU callbacks during 155 * early boot to take responsibility for these callbacks, but one step at 156 * a time. 157 */ 158 static int rcu_scheduler_fully_active __read_mostly; 159 160 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf); 161 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf); 162 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu); 163 static void invoke_rcu_core(void); 164 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp); 165 static void rcu_report_exp_rdp(struct rcu_state *rsp, 166 struct rcu_data *rdp, bool wake); 167 static void sync_sched_exp_online_cleanup(int cpu); 168 169 /* rcuc/rcub kthread realtime priority */ 170 #ifdef CONFIG_RCU_KTHREAD_PRIO 171 static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO; 172 #else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */ 173 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0; 174 #endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */ 175 module_param(kthread_prio, int, 0644); 176 177 /* Delay in jiffies for grace-period initialization delays, debug only. */ 178 179 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT 180 static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY; 181 module_param(gp_preinit_delay, int, 0644); 182 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ 183 static const int gp_preinit_delay; 184 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */ 185 186 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT 187 static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY; 188 module_param(gp_init_delay, int, 0644); 189 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ 190 static const int gp_init_delay; 191 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */ 192 193 #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP 194 static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY; 195 module_param(gp_cleanup_delay, int, 0644); 196 #else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ 197 static const int gp_cleanup_delay; 198 #endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */ 199 200 /* 201 * Number of grace periods between delays, normalized by the duration of 202 * the delay. The longer the the delay, the more the grace periods between 203 * each delay. The reason for this normalization is that it means that, 204 * for non-zero delays, the overall slowdown of grace periods is constant 205 * regardless of the duration of the delay. This arrangement balances 206 * the need for long delays to increase some race probabilities with the 207 * need for fast grace periods to increase other race probabilities. 208 */ 209 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */ 210 211 /* 212 * Track the rcutorture test sequence number and the update version 213 * number within a given test. The rcutorture_testseq is incremented 214 * on every rcutorture module load and unload, so has an odd value 215 * when a test is running. The rcutorture_vernum is set to zero 216 * when rcutorture starts and is incremented on each rcutorture update. 217 * These variables enable correlating rcutorture output with the 218 * RCU tracing information. 219 */ 220 unsigned long rcutorture_testseq; 221 unsigned long rcutorture_vernum; 222 223 /* 224 * Compute the mask of online CPUs for the specified rcu_node structure. 225 * This will not be stable unless the rcu_node structure's ->lock is 226 * held, but the bit corresponding to the current CPU will be stable 227 * in most contexts. 228 */ 229 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp) 230 { 231 return READ_ONCE(rnp->qsmaskinitnext); 232 } 233 234 /* 235 * Return true if an RCU grace period is in progress. The READ_ONCE()s 236 * permit this function to be invoked without holding the root rcu_node 237 * structure's ->lock, but of course results can be subject to change. 238 */ 239 static int rcu_gp_in_progress(struct rcu_state *rsp) 240 { 241 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum); 242 } 243 244 /* 245 * Note a quiescent state. Because we do not need to know 246 * how many quiescent states passed, just if there was at least 247 * one since the start of the grace period, this just sets a flag. 248 * The caller must have disabled preemption. 249 */ 250 void rcu_sched_qs(void) 251 { 252 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s)) 253 return; 254 trace_rcu_grace_period(TPS("rcu_sched"), 255 __this_cpu_read(rcu_sched_data.gpnum), 256 TPS("cpuqs")); 257 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false); 258 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) 259 return; 260 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false); 261 rcu_report_exp_rdp(&rcu_sched_state, 262 this_cpu_ptr(&rcu_sched_data), true); 263 } 264 265 void rcu_bh_qs(void) 266 { 267 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) { 268 trace_rcu_grace_period(TPS("rcu_bh"), 269 __this_cpu_read(rcu_bh_data.gpnum), 270 TPS("cpuqs")); 271 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false); 272 } 273 } 274 275 static DEFINE_PER_CPU(int, rcu_sched_qs_mask); 276 277 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = { 278 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE, 279 .dynticks = ATOMIC_INIT(1), 280 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE 281 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE, 282 .dynticks_idle = ATOMIC_INIT(1), 283 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ 284 }; 285 286 /* 287 * Record entry into an extended quiescent state. This is only to be 288 * called when not already in an extended quiescent state. 289 */ 290 static void rcu_dynticks_eqs_enter(void) 291 { 292 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); 293 int special; 294 295 /* 296 * CPUs seeing atomic_inc_return() must see prior RCU read-side 297 * critical sections, and we also must force ordering with the 298 * next idle sojourn. 299 */ 300 special = atomic_inc_return(&rdtp->dynticks); 301 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && special & 0x1); 302 } 303 304 /* 305 * Record exit from an extended quiescent state. This is only to be 306 * called from an extended quiescent state. 307 */ 308 static void rcu_dynticks_eqs_exit(void) 309 { 310 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); 311 int special; 312 313 /* 314 * CPUs seeing atomic_inc_return() must see prior idle sojourns, 315 * and we also must force ordering with the next RCU read-side 316 * critical section. 317 */ 318 special = atomic_inc_return(&rdtp->dynticks); 319 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !(special & 0x1)); 320 } 321 322 /* 323 * Reset the current CPU's ->dynticks counter to indicate that the 324 * newly onlined CPU is no longer in an extended quiescent state. 325 * This will either leave the counter unchanged, or increment it 326 * to the next non-quiescent value. 327 * 328 * The non-atomic test/increment sequence works because the upper bits 329 * of the ->dynticks counter are manipulated only by the corresponding CPU, 330 * or when the corresponding CPU is offline. 331 */ 332 static void rcu_dynticks_eqs_online(void) 333 { 334 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); 335 336 if (atomic_read(&rdtp->dynticks) & 0x1) 337 return; 338 atomic_add(0x1, &rdtp->dynticks); 339 } 340 341 /* 342 * Is the current CPU in an extended quiescent state? 343 * 344 * No ordering, as we are sampling CPU-local information. 345 */ 346 bool rcu_dynticks_curr_cpu_in_eqs(void) 347 { 348 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); 349 350 return !(atomic_read(&rdtp->dynticks) & 0x1); 351 } 352 353 /* 354 * Snapshot the ->dynticks counter with full ordering so as to allow 355 * stable comparison of this counter with past and future snapshots. 356 */ 357 int rcu_dynticks_snap(struct rcu_dynticks *rdtp) 358 { 359 int snap = atomic_add_return(0, &rdtp->dynticks); 360 361 return snap; 362 } 363 364 /* 365 * Return true if the snapshot returned from rcu_dynticks_snap() 366 * indicates that RCU is in an extended quiescent state. 367 */ 368 static bool rcu_dynticks_in_eqs(int snap) 369 { 370 return !(snap & 0x1); 371 } 372 373 /* 374 * Return true if the CPU corresponding to the specified rcu_dynticks 375 * structure has spent some time in an extended quiescent state since 376 * rcu_dynticks_snap() returned the specified snapshot. 377 */ 378 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap) 379 { 380 return snap != rcu_dynticks_snap(rdtp); 381 } 382 383 /* 384 * Do a double-increment of the ->dynticks counter to emulate a 385 * momentary idle-CPU quiescent state. 386 */ 387 static void rcu_dynticks_momentary_idle(void) 388 { 389 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); 390 int special = atomic_add_return(2, &rdtp->dynticks); 391 392 /* It is illegal to call this from idle state. */ 393 WARN_ON_ONCE(!(special & 0x1)); 394 } 395 396 DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr); 397 EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr); 398 399 /* 400 * Let the RCU core know that this CPU has gone through the scheduler, 401 * which is a quiescent state. This is called when the need for a 402 * quiescent state is urgent, so we burn an atomic operation and full 403 * memory barriers to let the RCU core know about it, regardless of what 404 * this CPU might (or might not) do in the near future. 405 * 406 * We inform the RCU core by emulating a zero-duration dyntick-idle 407 * period, which we in turn do by incrementing the ->dynticks counter 408 * by two. 409 * 410 * The caller must have disabled interrupts. 411 */ 412 static void rcu_momentary_dyntick_idle(void) 413 { 414 struct rcu_data *rdp; 415 int resched_mask; 416 struct rcu_state *rsp; 417 418 /* 419 * Yes, we can lose flag-setting operations. This is OK, because 420 * the flag will be set again after some delay. 421 */ 422 resched_mask = raw_cpu_read(rcu_sched_qs_mask); 423 raw_cpu_write(rcu_sched_qs_mask, 0); 424 425 /* Find the flavor that needs a quiescent state. */ 426 for_each_rcu_flavor(rsp) { 427 rdp = raw_cpu_ptr(rsp->rda); 428 if (!(resched_mask & rsp->flavor_mask)) 429 continue; 430 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */ 431 if (READ_ONCE(rdp->mynode->completed) != 432 READ_ONCE(rdp->cond_resched_completed)) 433 continue; 434 435 /* 436 * Pretend to be momentarily idle for the quiescent state. 437 * This allows the grace-period kthread to record the 438 * quiescent state, with no need for this CPU to do anything 439 * further. 440 */ 441 rcu_dynticks_momentary_idle(); 442 break; 443 } 444 } 445 446 /* 447 * Note a context switch. This is a quiescent state for RCU-sched, 448 * and requires special handling for preemptible RCU. 449 * The caller must have disabled interrupts. 450 */ 451 void rcu_note_context_switch(void) 452 { 453 barrier(); /* Avoid RCU read-side critical sections leaking down. */ 454 trace_rcu_utilization(TPS("Start context switch")); 455 rcu_sched_qs(); 456 rcu_preempt_note_context_switch(); 457 if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) 458 rcu_momentary_dyntick_idle(); 459 trace_rcu_utilization(TPS("End context switch")); 460 barrier(); /* Avoid RCU read-side critical sections leaking up. */ 461 } 462 EXPORT_SYMBOL_GPL(rcu_note_context_switch); 463 464 /* 465 * Register a quiescent state for all RCU flavors. If there is an 466 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight 467 * dyntick-idle quiescent state visible to other CPUs (but only for those 468 * RCU flavors in desperate need of a quiescent state, which will normally 469 * be none of them). Either way, do a lightweight quiescent state for 470 * all RCU flavors. 471 * 472 * The barrier() calls are redundant in the common case when this is 473 * called externally, but just in case this is called from within this 474 * file. 475 * 476 */ 477 void rcu_all_qs(void) 478 { 479 unsigned long flags; 480 481 barrier(); /* Avoid RCU read-side critical sections leaking down. */ 482 if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) { 483 local_irq_save(flags); 484 rcu_momentary_dyntick_idle(); 485 local_irq_restore(flags); 486 } 487 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))) { 488 /* 489 * Yes, we just checked a per-CPU variable with preemption 490 * enabled, so we might be migrated to some other CPU at 491 * this point. That is OK because in that case, the 492 * migration will supply the needed quiescent state. 493 * We might end up needlessly disabling preemption and 494 * invoking rcu_sched_qs() on the destination CPU, but 495 * the probability and cost are both quite low, so this 496 * should not be a problem in practice. 497 */ 498 preempt_disable(); 499 rcu_sched_qs(); 500 preempt_enable(); 501 } 502 this_cpu_inc(rcu_qs_ctr); 503 barrier(); /* Avoid RCU read-side critical sections leaking up. */ 504 } 505 EXPORT_SYMBOL_GPL(rcu_all_qs); 506 507 static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */ 508 static long qhimark = 10000; /* If this many pending, ignore blimit. */ 509 static long qlowmark = 100; /* Once only this many pending, use blimit. */ 510 511 module_param(blimit, long, 0444); 512 module_param(qhimark, long, 0444); 513 module_param(qlowmark, long, 0444); 514 515 static ulong jiffies_till_first_fqs = ULONG_MAX; 516 static ulong jiffies_till_next_fqs = ULONG_MAX; 517 static bool rcu_kick_kthreads; 518 519 module_param(jiffies_till_first_fqs, ulong, 0644); 520 module_param(jiffies_till_next_fqs, ulong, 0644); 521 module_param(rcu_kick_kthreads, bool, 0644); 522 523 /* 524 * How long the grace period must be before we start recruiting 525 * quiescent-state help from rcu_note_context_switch(). 526 */ 527 static ulong jiffies_till_sched_qs = HZ / 20; 528 module_param(jiffies_till_sched_qs, ulong, 0644); 529 530 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, 531 struct rcu_data *rdp); 532 static void force_qs_rnp(struct rcu_state *rsp, 533 int (*f)(struct rcu_data *rsp, bool *isidle, 534 unsigned long *maxj), 535 bool *isidle, unsigned long *maxj); 536 static void force_quiescent_state(struct rcu_state *rsp); 537 static int rcu_pending(void); 538 539 /* 540 * Return the number of RCU batches started thus far for debug & stats. 541 */ 542 unsigned long rcu_batches_started(void) 543 { 544 return rcu_state_p->gpnum; 545 } 546 EXPORT_SYMBOL_GPL(rcu_batches_started); 547 548 /* 549 * Return the number of RCU-sched batches started thus far for debug & stats. 550 */ 551 unsigned long rcu_batches_started_sched(void) 552 { 553 return rcu_sched_state.gpnum; 554 } 555 EXPORT_SYMBOL_GPL(rcu_batches_started_sched); 556 557 /* 558 * Return the number of RCU BH batches started thus far for debug & stats. 559 */ 560 unsigned long rcu_batches_started_bh(void) 561 { 562 return rcu_bh_state.gpnum; 563 } 564 EXPORT_SYMBOL_GPL(rcu_batches_started_bh); 565 566 /* 567 * Return the number of RCU batches completed thus far for debug & stats. 568 */ 569 unsigned long rcu_batches_completed(void) 570 { 571 return rcu_state_p->completed; 572 } 573 EXPORT_SYMBOL_GPL(rcu_batches_completed); 574 575 /* 576 * Return the number of RCU-sched batches completed thus far for debug & stats. 577 */ 578 unsigned long rcu_batches_completed_sched(void) 579 { 580 return rcu_sched_state.completed; 581 } 582 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched); 583 584 /* 585 * Return the number of RCU BH batches completed thus far for debug & stats. 586 */ 587 unsigned long rcu_batches_completed_bh(void) 588 { 589 return rcu_bh_state.completed; 590 } 591 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); 592 593 /* 594 * Return the number of RCU expedited batches completed thus far for 595 * debug & stats. Odd numbers mean that a batch is in progress, even 596 * numbers mean idle. The value returned will thus be roughly double 597 * the cumulative batches since boot. 598 */ 599 unsigned long rcu_exp_batches_completed(void) 600 { 601 return rcu_state_p->expedited_sequence; 602 } 603 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed); 604 605 /* 606 * Return the number of RCU-sched expedited batches completed thus far 607 * for debug & stats. Similar to rcu_exp_batches_completed(). 608 */ 609 unsigned long rcu_exp_batches_completed_sched(void) 610 { 611 return rcu_sched_state.expedited_sequence; 612 } 613 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched); 614 615 /* 616 * Force a quiescent state. 617 */ 618 void rcu_force_quiescent_state(void) 619 { 620 force_quiescent_state(rcu_state_p); 621 } 622 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); 623 624 /* 625 * Force a quiescent state for RCU BH. 626 */ 627 void rcu_bh_force_quiescent_state(void) 628 { 629 force_quiescent_state(&rcu_bh_state); 630 } 631 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state); 632 633 /* 634 * Force a quiescent state for RCU-sched. 635 */ 636 void rcu_sched_force_quiescent_state(void) 637 { 638 force_quiescent_state(&rcu_sched_state); 639 } 640 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state); 641 642 /* 643 * Show the state of the grace-period kthreads. 644 */ 645 void show_rcu_gp_kthreads(void) 646 { 647 struct rcu_state *rsp; 648 649 for_each_rcu_flavor(rsp) { 650 pr_info("%s: wait state: %d ->state: %#lx\n", 651 rsp->name, rsp->gp_state, rsp->gp_kthread->state); 652 /* sched_show_task(rsp->gp_kthread); */ 653 } 654 } 655 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads); 656 657 /* 658 * Record the number of times rcutorture tests have been initiated and 659 * terminated. This information allows the debugfs tracing stats to be 660 * correlated to the rcutorture messages, even when the rcutorture module 661 * is being repeatedly loaded and unloaded. In other words, we cannot 662 * store this state in rcutorture itself. 663 */ 664 void rcutorture_record_test_transition(void) 665 { 666 rcutorture_testseq++; 667 rcutorture_vernum = 0; 668 } 669 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition); 670 671 /* 672 * Send along grace-period-related data for rcutorture diagnostics. 673 */ 674 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags, 675 unsigned long *gpnum, unsigned long *completed) 676 { 677 struct rcu_state *rsp = NULL; 678 679 switch (test_type) { 680 case RCU_FLAVOR: 681 rsp = rcu_state_p; 682 break; 683 case RCU_BH_FLAVOR: 684 rsp = &rcu_bh_state; 685 break; 686 case RCU_SCHED_FLAVOR: 687 rsp = &rcu_sched_state; 688 break; 689 default: 690 break; 691 } 692 if (rsp != NULL) { 693 *flags = READ_ONCE(rsp->gp_flags); 694 *gpnum = READ_ONCE(rsp->gpnum); 695 *completed = READ_ONCE(rsp->completed); 696 return; 697 } 698 *flags = 0; 699 *gpnum = 0; 700 *completed = 0; 701 } 702 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data); 703 704 /* 705 * Record the number of writer passes through the current rcutorture test. 706 * This is also used to correlate debugfs tracing stats with the rcutorture 707 * messages. 708 */ 709 void rcutorture_record_progress(unsigned long vernum) 710 { 711 rcutorture_vernum++; 712 } 713 EXPORT_SYMBOL_GPL(rcutorture_record_progress); 714 715 /* 716 * Does the CPU have callbacks ready to be invoked? 717 */ 718 static int 719 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp) 720 { 721 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] && 722 rdp->nxttail[RCU_NEXT_TAIL] != NULL; 723 } 724 725 /* 726 * Return the root node of the specified rcu_state structure. 727 */ 728 static struct rcu_node *rcu_get_root(struct rcu_state *rsp) 729 { 730 return &rsp->node[0]; 731 } 732 733 /* 734 * Is there any need for future grace periods? 735 * Interrupts must be disabled. If the caller does not hold the root 736 * rnp_node structure's ->lock, the results are advisory only. 737 */ 738 static int rcu_future_needs_gp(struct rcu_state *rsp) 739 { 740 struct rcu_node *rnp = rcu_get_root(rsp); 741 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1; 742 int *fp = &rnp->need_future_gp[idx]; 743 744 return READ_ONCE(*fp); 745 } 746 747 /* 748 * Does the current CPU require a not-yet-started grace period? 749 * The caller must have disabled interrupts to prevent races with 750 * normal callback registry. 751 */ 752 static bool 753 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp) 754 { 755 int i; 756 757 if (rcu_gp_in_progress(rsp)) 758 return false; /* No, a grace period is already in progress. */ 759 if (rcu_future_needs_gp(rsp)) 760 return true; /* Yes, a no-CBs CPU needs one. */ 761 if (!rdp->nxttail[RCU_NEXT_TAIL]) 762 return false; /* No, this is a no-CBs (or offline) CPU. */ 763 if (*rdp->nxttail[RCU_NEXT_READY_TAIL]) 764 return true; /* Yes, CPU has newly registered callbacks. */ 765 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) 766 if (rdp->nxttail[i - 1] != rdp->nxttail[i] && 767 ULONG_CMP_LT(READ_ONCE(rsp->completed), 768 rdp->nxtcompleted[i])) 769 return true; /* Yes, CBs for future grace period. */ 770 return false; /* No grace period needed. */ 771 } 772 773 /* 774 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state 775 * 776 * If the new value of the ->dynticks_nesting counter now is zero, 777 * we really have entered idle, and must do the appropriate accounting. 778 * The caller must have disabled interrupts. 779 */ 780 static void rcu_eqs_enter_common(long long oldval, bool user) 781 { 782 struct rcu_state *rsp; 783 struct rcu_data *rdp; 784 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);) 785 786 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting); 787 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && 788 !user && !is_idle_task(current)) { 789 struct task_struct *idle __maybe_unused = 790 idle_task(smp_processor_id()); 791 792 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0); 793 rcu_ftrace_dump(DUMP_ORIG); 794 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", 795 current->pid, current->comm, 796 idle->pid, idle->comm); /* must be idle task! */ 797 } 798 for_each_rcu_flavor(rsp) { 799 rdp = this_cpu_ptr(rsp->rda); 800 do_nocb_deferred_wakeup(rdp); 801 } 802 rcu_prepare_for_idle(); 803 rcu_dynticks_eqs_enter(); 804 rcu_dynticks_task_enter(); 805 806 /* 807 * It is illegal to enter an extended quiescent state while 808 * in an RCU read-side critical section. 809 */ 810 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map), 811 "Illegal idle entry in RCU read-side critical section."); 812 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map), 813 "Illegal idle entry in RCU-bh read-side critical section."); 814 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map), 815 "Illegal idle entry in RCU-sched read-side critical section."); 816 } 817 818 /* 819 * Enter an RCU extended quiescent state, which can be either the 820 * idle loop or adaptive-tickless usermode execution. 821 */ 822 static void rcu_eqs_enter(bool user) 823 { 824 long long oldval; 825 struct rcu_dynticks *rdtp; 826 827 rdtp = this_cpu_ptr(&rcu_dynticks); 828 oldval = rdtp->dynticks_nesting; 829 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && 830 (oldval & DYNTICK_TASK_NEST_MASK) == 0); 831 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) { 832 rdtp->dynticks_nesting = 0; 833 rcu_eqs_enter_common(oldval, user); 834 } else { 835 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE; 836 } 837 } 838 839 /** 840 * rcu_idle_enter - inform RCU that current CPU is entering idle 841 * 842 * Enter idle mode, in other words, -leave- the mode in which RCU 843 * read-side critical sections can occur. (Though RCU read-side 844 * critical sections can occur in irq handlers in idle, a possibility 845 * handled by irq_enter() and irq_exit().) 846 * 847 * We crowbar the ->dynticks_nesting field to zero to allow for 848 * the possibility of usermode upcalls having messed up our count 849 * of interrupt nesting level during the prior busy period. 850 */ 851 void rcu_idle_enter(void) 852 { 853 unsigned long flags; 854 855 local_irq_save(flags); 856 rcu_eqs_enter(false); 857 rcu_sysidle_enter(0); 858 local_irq_restore(flags); 859 } 860 EXPORT_SYMBOL_GPL(rcu_idle_enter); 861 862 #ifdef CONFIG_NO_HZ_FULL 863 /** 864 * rcu_user_enter - inform RCU that we are resuming userspace. 865 * 866 * Enter RCU idle mode right before resuming userspace. No use of RCU 867 * is permitted between this call and rcu_user_exit(). This way the 868 * CPU doesn't need to maintain the tick for RCU maintenance purposes 869 * when the CPU runs in userspace. 870 */ 871 void rcu_user_enter(void) 872 { 873 rcu_eqs_enter(1); 874 } 875 #endif /* CONFIG_NO_HZ_FULL */ 876 877 /** 878 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle 879 * 880 * Exit from an interrupt handler, which might possibly result in entering 881 * idle mode, in other words, leaving the mode in which read-side critical 882 * sections can occur. The caller must have disabled interrupts. 883 * 884 * This code assumes that the idle loop never does anything that might 885 * result in unbalanced calls to irq_enter() and irq_exit(). If your 886 * architecture violates this assumption, RCU will give you what you 887 * deserve, good and hard. But very infrequently and irreproducibly. 888 * 889 * Use things like work queues to work around this limitation. 890 * 891 * You have been warned. 892 */ 893 void rcu_irq_exit(void) 894 { 895 long long oldval; 896 struct rcu_dynticks *rdtp; 897 898 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!"); 899 rdtp = this_cpu_ptr(&rcu_dynticks); 900 oldval = rdtp->dynticks_nesting; 901 rdtp->dynticks_nesting--; 902 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && 903 rdtp->dynticks_nesting < 0); 904 if (rdtp->dynticks_nesting) 905 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting); 906 else 907 rcu_eqs_enter_common(oldval, true); 908 rcu_sysidle_enter(1); 909 } 910 911 /* 912 * Wrapper for rcu_irq_exit() where interrupts are enabled. 913 */ 914 void rcu_irq_exit_irqson(void) 915 { 916 unsigned long flags; 917 918 local_irq_save(flags); 919 rcu_irq_exit(); 920 local_irq_restore(flags); 921 } 922 923 /* 924 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state 925 * 926 * If the new value of the ->dynticks_nesting counter was previously zero, 927 * we really have exited idle, and must do the appropriate accounting. 928 * The caller must have disabled interrupts. 929 */ 930 static void rcu_eqs_exit_common(long long oldval, int user) 931 { 932 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);) 933 934 rcu_dynticks_task_exit(); 935 rcu_dynticks_eqs_exit(); 936 rcu_cleanup_after_idle(); 937 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting); 938 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && 939 !user && !is_idle_task(current)) { 940 struct task_struct *idle __maybe_unused = 941 idle_task(smp_processor_id()); 942 943 trace_rcu_dyntick(TPS("Error on exit: not idle task"), 944 oldval, rdtp->dynticks_nesting); 945 rcu_ftrace_dump(DUMP_ORIG); 946 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s", 947 current->pid, current->comm, 948 idle->pid, idle->comm); /* must be idle task! */ 949 } 950 } 951 952 /* 953 * Exit an RCU extended quiescent state, which can be either the 954 * idle loop or adaptive-tickless usermode execution. 955 */ 956 static void rcu_eqs_exit(bool user) 957 { 958 struct rcu_dynticks *rdtp; 959 long long oldval; 960 961 rdtp = this_cpu_ptr(&rcu_dynticks); 962 oldval = rdtp->dynticks_nesting; 963 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0); 964 if (oldval & DYNTICK_TASK_NEST_MASK) { 965 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE; 966 } else { 967 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; 968 rcu_eqs_exit_common(oldval, user); 969 } 970 } 971 972 /** 973 * rcu_idle_exit - inform RCU that current CPU is leaving idle 974 * 975 * Exit idle mode, in other words, -enter- the mode in which RCU 976 * read-side critical sections can occur. 977 * 978 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to 979 * allow for the possibility of usermode upcalls messing up our count 980 * of interrupt nesting level during the busy period that is just 981 * now starting. 982 */ 983 void rcu_idle_exit(void) 984 { 985 unsigned long flags; 986 987 local_irq_save(flags); 988 rcu_eqs_exit(false); 989 rcu_sysidle_exit(0); 990 local_irq_restore(flags); 991 } 992 EXPORT_SYMBOL_GPL(rcu_idle_exit); 993 994 #ifdef CONFIG_NO_HZ_FULL 995 /** 996 * rcu_user_exit - inform RCU that we are exiting userspace. 997 * 998 * Exit RCU idle mode while entering the kernel because it can 999 * run a RCU read side critical section anytime. 1000 */ 1001 void rcu_user_exit(void) 1002 { 1003 rcu_eqs_exit(1); 1004 } 1005 #endif /* CONFIG_NO_HZ_FULL */ 1006 1007 /** 1008 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle 1009 * 1010 * Enter an interrupt handler, which might possibly result in exiting 1011 * idle mode, in other words, entering the mode in which read-side critical 1012 * sections can occur. The caller must have disabled interrupts. 1013 * 1014 * Note that the Linux kernel is fully capable of entering an interrupt 1015 * handler that it never exits, for example when doing upcalls to 1016 * user mode! This code assumes that the idle loop never does upcalls to 1017 * user mode. If your architecture does do upcalls from the idle loop (or 1018 * does anything else that results in unbalanced calls to the irq_enter() 1019 * and irq_exit() functions), RCU will give you what you deserve, good 1020 * and hard. But very infrequently and irreproducibly. 1021 * 1022 * Use things like work queues to work around this limitation. 1023 * 1024 * You have been warned. 1025 */ 1026 void rcu_irq_enter(void) 1027 { 1028 struct rcu_dynticks *rdtp; 1029 long long oldval; 1030 1031 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!"); 1032 rdtp = this_cpu_ptr(&rcu_dynticks); 1033 oldval = rdtp->dynticks_nesting; 1034 rdtp->dynticks_nesting++; 1035 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && 1036 rdtp->dynticks_nesting == 0); 1037 if (oldval) 1038 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting); 1039 else 1040 rcu_eqs_exit_common(oldval, true); 1041 rcu_sysidle_exit(1); 1042 } 1043 1044 /* 1045 * Wrapper for rcu_irq_enter() where interrupts are enabled. 1046 */ 1047 void rcu_irq_enter_irqson(void) 1048 { 1049 unsigned long flags; 1050 1051 local_irq_save(flags); 1052 rcu_irq_enter(); 1053 local_irq_restore(flags); 1054 } 1055 1056 /** 1057 * rcu_nmi_enter - inform RCU of entry to NMI context 1058 * 1059 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and 1060 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know 1061 * that the CPU is active. This implementation permits nested NMIs, as 1062 * long as the nesting level does not overflow an int. (You will probably 1063 * run out of stack space first.) 1064 */ 1065 void rcu_nmi_enter(void) 1066 { 1067 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); 1068 int incby = 2; 1069 1070 /* Complain about underflow. */ 1071 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0); 1072 1073 /* 1074 * If idle from RCU viewpoint, atomically increment ->dynticks 1075 * to mark non-idle and increment ->dynticks_nmi_nesting by one. 1076 * Otherwise, increment ->dynticks_nmi_nesting by two. This means 1077 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed 1078 * to be in the outermost NMI handler that interrupted an RCU-idle 1079 * period (observation due to Andy Lutomirski). 1080 */ 1081 if (rcu_dynticks_curr_cpu_in_eqs()) { 1082 rcu_dynticks_eqs_exit(); 1083 incby = 1; 1084 } 1085 rdtp->dynticks_nmi_nesting += incby; 1086 barrier(); 1087 } 1088 1089 /** 1090 * rcu_nmi_exit - inform RCU of exit from NMI context 1091 * 1092 * If we are returning from the outermost NMI handler that interrupted an 1093 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting 1094 * to let the RCU grace-period handling know that the CPU is back to 1095 * being RCU-idle. 1096 */ 1097 void rcu_nmi_exit(void) 1098 { 1099 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); 1100 1101 /* 1102 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks. 1103 * (We are exiting an NMI handler, so RCU better be paying attention 1104 * to us!) 1105 */ 1106 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0); 1107 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs()); 1108 1109 /* 1110 * If the nesting level is not 1, the CPU wasn't RCU-idle, so 1111 * leave it in non-RCU-idle state. 1112 */ 1113 if (rdtp->dynticks_nmi_nesting != 1) { 1114 rdtp->dynticks_nmi_nesting -= 2; 1115 return; 1116 } 1117 1118 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */ 1119 rdtp->dynticks_nmi_nesting = 0; 1120 rcu_dynticks_eqs_enter(); 1121 } 1122 1123 /** 1124 * __rcu_is_watching - are RCU read-side critical sections safe? 1125 * 1126 * Return true if RCU is watching the running CPU, which means that 1127 * this CPU can safely enter RCU read-side critical sections. Unlike 1128 * rcu_is_watching(), the caller of __rcu_is_watching() must have at 1129 * least disabled preemption. 1130 */ 1131 bool notrace __rcu_is_watching(void) 1132 { 1133 return !rcu_dynticks_curr_cpu_in_eqs(); 1134 } 1135 1136 /** 1137 * rcu_is_watching - see if RCU thinks that the current CPU is idle 1138 * 1139 * If the current CPU is in its idle loop and is neither in an interrupt 1140 * or NMI handler, return true. 1141 */ 1142 bool notrace rcu_is_watching(void) 1143 { 1144 bool ret; 1145 1146 preempt_disable_notrace(); 1147 ret = __rcu_is_watching(); 1148 preempt_enable_notrace(); 1149 return ret; 1150 } 1151 EXPORT_SYMBOL_GPL(rcu_is_watching); 1152 1153 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) 1154 1155 /* 1156 * Is the current CPU online? Disable preemption to avoid false positives 1157 * that could otherwise happen due to the current CPU number being sampled, 1158 * this task being preempted, its old CPU being taken offline, resuming 1159 * on some other CPU, then determining that its old CPU is now offline. 1160 * It is OK to use RCU on an offline processor during initial boot, hence 1161 * the check for rcu_scheduler_fully_active. Note also that it is OK 1162 * for a CPU coming online to use RCU for one jiffy prior to marking itself 1163 * online in the cpu_online_mask. Similarly, it is OK for a CPU going 1164 * offline to continue to use RCU for one jiffy after marking itself 1165 * offline in the cpu_online_mask. This leniency is necessary given the 1166 * non-atomic nature of the online and offline processing, for example, 1167 * the fact that a CPU enters the scheduler after completing the teardown 1168 * of the CPU. 1169 * 1170 * This is also why RCU internally marks CPUs online during in the 1171 * preparation phase and offline after the CPU has been taken down. 1172 * 1173 * Disable checking if in an NMI handler because we cannot safely report 1174 * errors from NMI handlers anyway. 1175 */ 1176 bool rcu_lockdep_current_cpu_online(void) 1177 { 1178 struct rcu_data *rdp; 1179 struct rcu_node *rnp; 1180 bool ret; 1181 1182 if (in_nmi()) 1183 return true; 1184 preempt_disable(); 1185 rdp = this_cpu_ptr(&rcu_sched_data); 1186 rnp = rdp->mynode; 1187 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) || 1188 !rcu_scheduler_fully_active; 1189 preempt_enable(); 1190 return ret; 1191 } 1192 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online); 1193 1194 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */ 1195 1196 /** 1197 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle 1198 * 1199 * If the current CPU is idle or running at a first-level (not nested) 1200 * interrupt from idle, return true. The caller must have at least 1201 * disabled preemption. 1202 */ 1203 static int rcu_is_cpu_rrupt_from_idle(void) 1204 { 1205 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1; 1206 } 1207 1208 /* 1209 * Snapshot the specified CPU's dynticks counter so that we can later 1210 * credit them with an implicit quiescent state. Return 1 if this CPU 1211 * is in dynticks idle mode, which is an extended quiescent state. 1212 */ 1213 static int dyntick_save_progress_counter(struct rcu_data *rdp, 1214 bool *isidle, unsigned long *maxj) 1215 { 1216 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks); 1217 rcu_sysidle_check_cpu(rdp, isidle, maxj); 1218 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) { 1219 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); 1220 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, 1221 rdp->mynode->gpnum)) 1222 WRITE_ONCE(rdp->gpwrap, true); 1223 return 1; 1224 } 1225 return 0; 1226 } 1227 1228 /* 1229 * Return true if the specified CPU has passed through a quiescent 1230 * state by virtue of being in or having passed through an dynticks 1231 * idle state since the last call to dyntick_save_progress_counter() 1232 * for this same CPU, or by virtue of having been offline. 1233 */ 1234 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp, 1235 bool *isidle, unsigned long *maxj) 1236 { 1237 unsigned long jtsq; 1238 int *rcrmp; 1239 unsigned long rjtsc; 1240 struct rcu_node *rnp; 1241 1242 /* 1243 * If the CPU passed through or entered a dynticks idle phase with 1244 * no active irq/NMI handlers, then we can safely pretend that the CPU 1245 * already acknowledged the request to pass through a quiescent 1246 * state. Either way, that CPU cannot possibly be in an RCU 1247 * read-side critical section that started before the beginning 1248 * of the current RCU grace period. 1249 */ 1250 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) { 1251 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti")); 1252 rdp->dynticks_fqs++; 1253 return 1; 1254 } 1255 1256 /* Compute and saturate jiffies_till_sched_qs. */ 1257 jtsq = jiffies_till_sched_qs; 1258 rjtsc = rcu_jiffies_till_stall_check(); 1259 if (jtsq > rjtsc / 2) { 1260 WRITE_ONCE(jiffies_till_sched_qs, rjtsc); 1261 jtsq = rjtsc / 2; 1262 } else if (jtsq < 1) { 1263 WRITE_ONCE(jiffies_till_sched_qs, 1); 1264 jtsq = 1; 1265 } 1266 1267 /* 1268 * Has this CPU encountered a cond_resched_rcu_qs() since the 1269 * beginning of the grace period? For this to be the case, 1270 * the CPU has to have noticed the current grace period. This 1271 * might not be the case for nohz_full CPUs looping in the kernel. 1272 */ 1273 rnp = rdp->mynode; 1274 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) && 1275 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_qs_ctr, rdp->cpu) && 1276 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) { 1277 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc")); 1278 return 1; 1279 } 1280 1281 /* Check for the CPU being offline. */ 1282 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) { 1283 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl")); 1284 rdp->offline_fqs++; 1285 return 1; 1286 } 1287 1288 /* 1289 * A CPU running for an extended time within the kernel can 1290 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode, 1291 * even context-switching back and forth between a pair of 1292 * in-kernel CPU-bound tasks cannot advance grace periods. 1293 * So if the grace period is old enough, make the CPU pay attention. 1294 * Note that the unsynchronized assignments to the per-CPU 1295 * rcu_sched_qs_mask variable are safe. Yes, setting of 1296 * bits can be lost, but they will be set again on the next 1297 * force-quiescent-state pass. So lost bit sets do not result 1298 * in incorrect behavior, merely in a grace period lasting 1299 * a few jiffies longer than it might otherwise. Because 1300 * there are at most four threads involved, and because the 1301 * updates are only once every few jiffies, the probability of 1302 * lossage (and thus of slight grace-period extension) is 1303 * quite low. 1304 * 1305 * Note that if the jiffies_till_sched_qs boot/sysfs parameter 1306 * is set too high, we override with half of the RCU CPU stall 1307 * warning delay. 1308 */ 1309 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu); 1310 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) || 1311 time_after(jiffies, rdp->rsp->jiffies_resched)) { 1312 if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) { 1313 WRITE_ONCE(rdp->cond_resched_completed, 1314 READ_ONCE(rdp->mynode->completed)); 1315 smp_mb(); /* ->cond_resched_completed before *rcrmp. */ 1316 WRITE_ONCE(*rcrmp, 1317 READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask); 1318 } 1319 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */ 1320 } 1321 1322 /* 1323 * If more than halfway to RCU CPU stall-warning time, do 1324 * a resched_cpu() to try to loosen things up a bit. 1325 */ 1326 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) 1327 resched_cpu(rdp->cpu); 1328 1329 return 0; 1330 } 1331 1332 static void record_gp_stall_check_time(struct rcu_state *rsp) 1333 { 1334 unsigned long j = jiffies; 1335 unsigned long j1; 1336 1337 rsp->gp_start = j; 1338 smp_wmb(); /* Record start time before stall time. */ 1339 j1 = rcu_jiffies_till_stall_check(); 1340 WRITE_ONCE(rsp->jiffies_stall, j + j1); 1341 rsp->jiffies_resched = j + j1 / 2; 1342 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs); 1343 } 1344 1345 /* 1346 * Convert a ->gp_state value to a character string. 1347 */ 1348 static const char *gp_state_getname(short gs) 1349 { 1350 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names)) 1351 return "???"; 1352 return gp_state_names[gs]; 1353 } 1354 1355 /* 1356 * Complain about starvation of grace-period kthread. 1357 */ 1358 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp) 1359 { 1360 unsigned long gpa; 1361 unsigned long j; 1362 1363 j = jiffies; 1364 gpa = READ_ONCE(rsp->gp_activity); 1365 if (j - gpa > 2 * HZ) { 1366 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n", 1367 rsp->name, j - gpa, 1368 rsp->gpnum, rsp->completed, 1369 rsp->gp_flags, 1370 gp_state_getname(rsp->gp_state), rsp->gp_state, 1371 rsp->gp_kthread ? rsp->gp_kthread->state : ~0); 1372 if (rsp->gp_kthread) { 1373 sched_show_task(rsp->gp_kthread); 1374 wake_up_process(rsp->gp_kthread); 1375 } 1376 } 1377 } 1378 1379 /* 1380 * Dump stacks of all tasks running on stalled CPUs. First try using 1381 * NMIs, but fall back to manual remote stack tracing on architectures 1382 * that don't support NMI-based stack dumps. The NMI-triggered stack 1383 * traces are more accurate because they are printed by the target CPU. 1384 */ 1385 static void rcu_dump_cpu_stacks(struct rcu_state *rsp) 1386 { 1387 int cpu; 1388 unsigned long flags; 1389 struct rcu_node *rnp; 1390 1391 rcu_for_each_leaf_node(rsp, rnp) { 1392 raw_spin_lock_irqsave_rcu_node(rnp, flags); 1393 for_each_leaf_node_possible_cpu(rnp, cpu) 1394 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) 1395 if (!trigger_single_cpu_backtrace(cpu)) 1396 dump_cpu_task(cpu); 1397 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 1398 } 1399 } 1400 1401 /* 1402 * If too much time has passed in the current grace period, and if 1403 * so configured, go kick the relevant kthreads. 1404 */ 1405 static void rcu_stall_kick_kthreads(struct rcu_state *rsp) 1406 { 1407 unsigned long j; 1408 1409 if (!rcu_kick_kthreads) 1410 return; 1411 j = READ_ONCE(rsp->jiffies_kick_kthreads); 1412 if (time_after(jiffies, j) && rsp->gp_kthread && 1413 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) { 1414 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name); 1415 rcu_ftrace_dump(DUMP_ALL); 1416 wake_up_process(rsp->gp_kthread); 1417 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ); 1418 } 1419 } 1420 1421 static inline void panic_on_rcu_stall(void) 1422 { 1423 if (sysctl_panic_on_rcu_stall) 1424 panic("RCU Stall\n"); 1425 } 1426 1427 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum) 1428 { 1429 int cpu; 1430 long delta; 1431 unsigned long flags; 1432 unsigned long gpa; 1433 unsigned long j; 1434 int ndetected = 0; 1435 struct rcu_node *rnp = rcu_get_root(rsp); 1436 long totqlen = 0; 1437 1438 /* Kick and suppress, if so configured. */ 1439 rcu_stall_kick_kthreads(rsp); 1440 if (rcu_cpu_stall_suppress) 1441 return; 1442 1443 /* Only let one CPU complain about others per time interval. */ 1444 1445 raw_spin_lock_irqsave_rcu_node(rnp, flags); 1446 delta = jiffies - READ_ONCE(rsp->jiffies_stall); 1447 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) { 1448 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 1449 return; 1450 } 1451 WRITE_ONCE(rsp->jiffies_stall, 1452 jiffies + 3 * rcu_jiffies_till_stall_check() + 3); 1453 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 1454 1455 /* 1456 * OK, time to rat on our buddy... 1457 * See Documentation/RCU/stallwarn.txt for info on how to debug 1458 * RCU CPU stall warnings. 1459 */ 1460 pr_err("INFO: %s detected stalls on CPUs/tasks:", 1461 rsp->name); 1462 print_cpu_stall_info_begin(); 1463 rcu_for_each_leaf_node(rsp, rnp) { 1464 raw_spin_lock_irqsave_rcu_node(rnp, flags); 1465 ndetected += rcu_print_task_stall(rnp); 1466 if (rnp->qsmask != 0) { 1467 for_each_leaf_node_possible_cpu(rnp, cpu) 1468 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) { 1469 print_cpu_stall_info(rsp, cpu); 1470 ndetected++; 1471 } 1472 } 1473 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 1474 } 1475 1476 print_cpu_stall_info_end(); 1477 for_each_possible_cpu(cpu) 1478 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; 1479 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n", 1480 smp_processor_id(), (long)(jiffies - rsp->gp_start), 1481 (long)rsp->gpnum, (long)rsp->completed, totqlen); 1482 if (ndetected) { 1483 rcu_dump_cpu_stacks(rsp); 1484 1485 /* Complain about tasks blocking the grace period. */ 1486 rcu_print_detail_task_stall(rsp); 1487 } else { 1488 if (READ_ONCE(rsp->gpnum) != gpnum || 1489 READ_ONCE(rsp->completed) == gpnum) { 1490 pr_err("INFO: Stall ended before state dump start\n"); 1491 } else { 1492 j = jiffies; 1493 gpa = READ_ONCE(rsp->gp_activity); 1494 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n", 1495 rsp->name, j - gpa, j, gpa, 1496 jiffies_till_next_fqs, 1497 rcu_get_root(rsp)->qsmask); 1498 /* In this case, the current CPU might be at fault. */ 1499 sched_show_task(current); 1500 } 1501 } 1502 1503 rcu_check_gp_kthread_starvation(rsp); 1504 1505 panic_on_rcu_stall(); 1506 1507 force_quiescent_state(rsp); /* Kick them all. */ 1508 } 1509 1510 static void print_cpu_stall(struct rcu_state *rsp) 1511 { 1512 int cpu; 1513 unsigned long flags; 1514 struct rcu_node *rnp = rcu_get_root(rsp); 1515 long totqlen = 0; 1516 1517 /* Kick and suppress, if so configured. */ 1518 rcu_stall_kick_kthreads(rsp); 1519 if (rcu_cpu_stall_suppress) 1520 return; 1521 1522 /* 1523 * OK, time to rat on ourselves... 1524 * See Documentation/RCU/stallwarn.txt for info on how to debug 1525 * RCU CPU stall warnings. 1526 */ 1527 pr_err("INFO: %s self-detected stall on CPU", rsp->name); 1528 print_cpu_stall_info_begin(); 1529 print_cpu_stall_info(rsp, smp_processor_id()); 1530 print_cpu_stall_info_end(); 1531 for_each_possible_cpu(cpu) 1532 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen; 1533 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n", 1534 jiffies - rsp->gp_start, 1535 (long)rsp->gpnum, (long)rsp->completed, totqlen); 1536 1537 rcu_check_gp_kthread_starvation(rsp); 1538 1539 rcu_dump_cpu_stacks(rsp); 1540 1541 raw_spin_lock_irqsave_rcu_node(rnp, flags); 1542 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall))) 1543 WRITE_ONCE(rsp->jiffies_stall, 1544 jiffies + 3 * rcu_jiffies_till_stall_check() + 3); 1545 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 1546 1547 panic_on_rcu_stall(); 1548 1549 /* 1550 * Attempt to revive the RCU machinery by forcing a context switch. 1551 * 1552 * A context switch would normally allow the RCU state machine to make 1553 * progress and it could be we're stuck in kernel space without context 1554 * switches for an entirely unreasonable amount of time. 1555 */ 1556 resched_cpu(smp_processor_id()); 1557 } 1558 1559 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp) 1560 { 1561 unsigned long completed; 1562 unsigned long gpnum; 1563 unsigned long gps; 1564 unsigned long j; 1565 unsigned long js; 1566 struct rcu_node *rnp; 1567 1568 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) || 1569 !rcu_gp_in_progress(rsp)) 1570 return; 1571 rcu_stall_kick_kthreads(rsp); 1572 j = jiffies; 1573 1574 /* 1575 * Lots of memory barriers to reject false positives. 1576 * 1577 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall, 1578 * then rsp->gp_start, and finally rsp->completed. These values 1579 * are updated in the opposite order with memory barriers (or 1580 * equivalent) during grace-period initialization and cleanup. 1581 * Now, a false positive can occur if we get an new value of 1582 * rsp->gp_start and a old value of rsp->jiffies_stall. But given 1583 * the memory barriers, the only way that this can happen is if one 1584 * grace period ends and another starts between these two fetches. 1585 * Detect this by comparing rsp->completed with the previous fetch 1586 * from rsp->gpnum. 1587 * 1588 * Given this check, comparisons of jiffies, rsp->jiffies_stall, 1589 * and rsp->gp_start suffice to forestall false positives. 1590 */ 1591 gpnum = READ_ONCE(rsp->gpnum); 1592 smp_rmb(); /* Pick up ->gpnum first... */ 1593 js = READ_ONCE(rsp->jiffies_stall); 1594 smp_rmb(); /* ...then ->jiffies_stall before the rest... */ 1595 gps = READ_ONCE(rsp->gp_start); 1596 smp_rmb(); /* ...and finally ->gp_start before ->completed. */ 1597 completed = READ_ONCE(rsp->completed); 1598 if (ULONG_CMP_GE(completed, gpnum) || 1599 ULONG_CMP_LT(j, js) || 1600 ULONG_CMP_GE(gps, js)) 1601 return; /* No stall or GP completed since entering function. */ 1602 rnp = rdp->mynode; 1603 if (rcu_gp_in_progress(rsp) && 1604 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) { 1605 1606 /* We haven't checked in, so go dump stack. */ 1607 print_cpu_stall(rsp); 1608 1609 } else if (rcu_gp_in_progress(rsp) && 1610 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) { 1611 1612 /* They had a few time units to dump stack, so complain. */ 1613 print_other_cpu_stall(rsp, gpnum); 1614 } 1615 } 1616 1617 /** 1618 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period 1619 * 1620 * Set the stall-warning timeout way off into the future, thus preventing 1621 * any RCU CPU stall-warning messages from appearing in the current set of 1622 * RCU grace periods. 1623 * 1624 * The caller must disable hard irqs. 1625 */ 1626 void rcu_cpu_stall_reset(void) 1627 { 1628 struct rcu_state *rsp; 1629 1630 for_each_rcu_flavor(rsp) 1631 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2); 1632 } 1633 1634 /* 1635 * Initialize the specified rcu_data structure's default callback list 1636 * to empty. The default callback list is the one that is not used by 1637 * no-callbacks CPUs. 1638 */ 1639 static void init_default_callback_list(struct rcu_data *rdp) 1640 { 1641 int i; 1642 1643 rdp->nxtlist = NULL; 1644 for (i = 0; i < RCU_NEXT_SIZE; i++) 1645 rdp->nxttail[i] = &rdp->nxtlist; 1646 } 1647 1648 /* 1649 * Initialize the specified rcu_data structure's callback list to empty. 1650 */ 1651 static void init_callback_list(struct rcu_data *rdp) 1652 { 1653 if (init_nocb_callback_list(rdp)) 1654 return; 1655 init_default_callback_list(rdp); 1656 } 1657 1658 /* 1659 * Determine the value that ->completed will have at the end of the 1660 * next subsequent grace period. This is used to tag callbacks so that 1661 * a CPU can invoke callbacks in a timely fashion even if that CPU has 1662 * been dyntick-idle for an extended period with callbacks under the 1663 * influence of RCU_FAST_NO_HZ. 1664 * 1665 * The caller must hold rnp->lock with interrupts disabled. 1666 */ 1667 static unsigned long rcu_cbs_completed(struct rcu_state *rsp, 1668 struct rcu_node *rnp) 1669 { 1670 /* 1671 * If RCU is idle, we just wait for the next grace period. 1672 * But we can only be sure that RCU is idle if we are looking 1673 * at the root rcu_node structure -- otherwise, a new grace 1674 * period might have started, but just not yet gotten around 1675 * to initializing the current non-root rcu_node structure. 1676 */ 1677 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed) 1678 return rnp->completed + 1; 1679 1680 /* 1681 * Otherwise, wait for a possible partial grace period and 1682 * then the subsequent full grace period. 1683 */ 1684 return rnp->completed + 2; 1685 } 1686 1687 /* 1688 * Trace-event helper function for rcu_start_future_gp() and 1689 * rcu_nocb_wait_gp(). 1690 */ 1691 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, 1692 unsigned long c, const char *s) 1693 { 1694 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum, 1695 rnp->completed, c, rnp->level, 1696 rnp->grplo, rnp->grphi, s); 1697 } 1698 1699 /* 1700 * Start some future grace period, as needed to handle newly arrived 1701 * callbacks. The required future grace periods are recorded in each 1702 * rcu_node structure's ->need_future_gp field. Returns true if there 1703 * is reason to awaken the grace-period kthread. 1704 * 1705 * The caller must hold the specified rcu_node structure's ->lock. 1706 */ 1707 static bool __maybe_unused 1708 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp, 1709 unsigned long *c_out) 1710 { 1711 unsigned long c; 1712 int i; 1713 bool ret = false; 1714 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); 1715 1716 /* 1717 * Pick up grace-period number for new callbacks. If this 1718 * grace period is already marked as needed, return to the caller. 1719 */ 1720 c = rcu_cbs_completed(rdp->rsp, rnp); 1721 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf")); 1722 if (rnp->need_future_gp[c & 0x1]) { 1723 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf")); 1724 goto out; 1725 } 1726 1727 /* 1728 * If either this rcu_node structure or the root rcu_node structure 1729 * believe that a grace period is in progress, then we must wait 1730 * for the one following, which is in "c". Because our request 1731 * will be noticed at the end of the current grace period, we don't 1732 * need to explicitly start one. We only do the lockless check 1733 * of rnp_root's fields if the current rcu_node structure thinks 1734 * there is no grace period in flight, and because we hold rnp->lock, 1735 * the only possible change is when rnp_root's two fields are 1736 * equal, in which case rnp_root->gpnum might be concurrently 1737 * incremented. But that is OK, as it will just result in our 1738 * doing some extra useless work. 1739 */ 1740 if (rnp->gpnum != rnp->completed || 1741 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) { 1742 rnp->need_future_gp[c & 0x1]++; 1743 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf")); 1744 goto out; 1745 } 1746 1747 /* 1748 * There might be no grace period in progress. If we don't already 1749 * hold it, acquire the root rcu_node structure's lock in order to 1750 * start one (if needed). 1751 */ 1752 if (rnp != rnp_root) 1753 raw_spin_lock_rcu_node(rnp_root); 1754 1755 /* 1756 * Get a new grace-period number. If there really is no grace 1757 * period in progress, it will be smaller than the one we obtained 1758 * earlier. Adjust callbacks as needed. Note that even no-CBs 1759 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed. 1760 */ 1761 c = rcu_cbs_completed(rdp->rsp, rnp_root); 1762 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++) 1763 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i])) 1764 rdp->nxtcompleted[i] = c; 1765 1766 /* 1767 * If the needed for the required grace period is already 1768 * recorded, trace and leave. 1769 */ 1770 if (rnp_root->need_future_gp[c & 0x1]) { 1771 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot")); 1772 goto unlock_out; 1773 } 1774 1775 /* Record the need for the future grace period. */ 1776 rnp_root->need_future_gp[c & 0x1]++; 1777 1778 /* If a grace period is not already in progress, start one. */ 1779 if (rnp_root->gpnum != rnp_root->completed) { 1780 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot")); 1781 } else { 1782 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot")); 1783 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp); 1784 } 1785 unlock_out: 1786 if (rnp != rnp_root) 1787 raw_spin_unlock_rcu_node(rnp_root); 1788 out: 1789 if (c_out != NULL) 1790 *c_out = c; 1791 return ret; 1792 } 1793 1794 /* 1795 * Clean up any old requests for the just-ended grace period. Also return 1796 * whether any additional grace periods have been requested. Also invoke 1797 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads 1798 * waiting for this grace period to complete. 1799 */ 1800 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) 1801 { 1802 int c = rnp->completed; 1803 int needmore; 1804 struct rcu_data *rdp = this_cpu_ptr(rsp->rda); 1805 1806 rnp->need_future_gp[c & 0x1] = 0; 1807 needmore = rnp->need_future_gp[(c + 1) & 0x1]; 1808 trace_rcu_future_gp(rnp, rdp, c, 1809 needmore ? TPS("CleanupMore") : TPS("Cleanup")); 1810 return needmore; 1811 } 1812 1813 /* 1814 * Awaken the grace-period kthread for the specified flavor of RCU. 1815 * Don't do a self-awaken, and don't bother awakening when there is 1816 * nothing for the grace-period kthread to do (as in several CPUs 1817 * raced to awaken, and we lost), and finally don't try to awaken 1818 * a kthread that has not yet been created. 1819 */ 1820 static void rcu_gp_kthread_wake(struct rcu_state *rsp) 1821 { 1822 if (current == rsp->gp_kthread || 1823 !READ_ONCE(rsp->gp_flags) || 1824 !rsp->gp_kthread) 1825 return; 1826 swake_up(&rsp->gp_wq); 1827 } 1828 1829 /* 1830 * If there is room, assign a ->completed number to any callbacks on 1831 * this CPU that have not already been assigned. Also accelerate any 1832 * callbacks that were previously assigned a ->completed number that has 1833 * since proven to be too conservative, which can happen if callbacks get 1834 * assigned a ->completed number while RCU is idle, but with reference to 1835 * a non-root rcu_node structure. This function is idempotent, so it does 1836 * not hurt to call it repeatedly. Returns an flag saying that we should 1837 * awaken the RCU grace-period kthread. 1838 * 1839 * The caller must hold rnp->lock with interrupts disabled. 1840 */ 1841 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp, 1842 struct rcu_data *rdp) 1843 { 1844 unsigned long c; 1845 int i; 1846 bool ret; 1847 1848 /* If the CPU has no callbacks, nothing to do. */ 1849 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) 1850 return false; 1851 1852 /* 1853 * Starting from the sublist containing the callbacks most 1854 * recently assigned a ->completed number and working down, find the 1855 * first sublist that is not assignable to an upcoming grace period. 1856 * Such a sublist has something in it (first two tests) and has 1857 * a ->completed number assigned that will complete sooner than 1858 * the ->completed number for newly arrived callbacks (last test). 1859 * 1860 * The key point is that any later sublist can be assigned the 1861 * same ->completed number as the newly arrived callbacks, which 1862 * means that the callbacks in any of these later sublist can be 1863 * grouped into a single sublist, whether or not they have already 1864 * been assigned a ->completed number. 1865 */ 1866 c = rcu_cbs_completed(rsp, rnp); 1867 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--) 1868 if (rdp->nxttail[i] != rdp->nxttail[i - 1] && 1869 !ULONG_CMP_GE(rdp->nxtcompleted[i], c)) 1870 break; 1871 1872 /* 1873 * If there are no sublist for unassigned callbacks, leave. 1874 * At the same time, advance "i" one sublist, so that "i" will 1875 * index into the sublist where all the remaining callbacks should 1876 * be grouped into. 1877 */ 1878 if (++i >= RCU_NEXT_TAIL) 1879 return false; 1880 1881 /* 1882 * Assign all subsequent callbacks' ->completed number to the next 1883 * full grace period and group them all in the sublist initially 1884 * indexed by "i". 1885 */ 1886 for (; i <= RCU_NEXT_TAIL; i++) { 1887 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL]; 1888 rdp->nxtcompleted[i] = c; 1889 } 1890 /* Record any needed additional grace periods. */ 1891 ret = rcu_start_future_gp(rnp, rdp, NULL); 1892 1893 /* Trace depending on how much we were able to accelerate. */ 1894 if (!*rdp->nxttail[RCU_WAIT_TAIL]) 1895 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB")); 1896 else 1897 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB")); 1898 return ret; 1899 } 1900 1901 /* 1902 * Move any callbacks whose grace period has completed to the 1903 * RCU_DONE_TAIL sublist, then compact the remaining sublists and 1904 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL 1905 * sublist. This function is idempotent, so it does not hurt to 1906 * invoke it repeatedly. As long as it is not invoked -too- often... 1907 * Returns true if the RCU grace-period kthread needs to be awakened. 1908 * 1909 * The caller must hold rnp->lock with interrupts disabled. 1910 */ 1911 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp, 1912 struct rcu_data *rdp) 1913 { 1914 int i, j; 1915 1916 /* If the CPU has no callbacks, nothing to do. */ 1917 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL]) 1918 return false; 1919 1920 /* 1921 * Find all callbacks whose ->completed numbers indicate that they 1922 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist. 1923 */ 1924 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) { 1925 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i])) 1926 break; 1927 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i]; 1928 } 1929 /* Clean up any sublist tail pointers that were misordered above. */ 1930 for (j = RCU_WAIT_TAIL; j < i; j++) 1931 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL]; 1932 1933 /* Copy down callbacks to fill in empty sublists. */ 1934 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) { 1935 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL]) 1936 break; 1937 rdp->nxttail[j] = rdp->nxttail[i]; 1938 rdp->nxtcompleted[j] = rdp->nxtcompleted[i]; 1939 } 1940 1941 /* Classify any remaining callbacks. */ 1942 return rcu_accelerate_cbs(rsp, rnp, rdp); 1943 } 1944 1945 /* 1946 * Update CPU-local rcu_data state to record the beginnings and ends of 1947 * grace periods. The caller must hold the ->lock of the leaf rcu_node 1948 * structure corresponding to the current CPU, and must have irqs disabled. 1949 * Returns true if the grace-period kthread needs to be awakened. 1950 */ 1951 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, 1952 struct rcu_data *rdp) 1953 { 1954 bool ret; 1955 bool need_gp; 1956 1957 /* Handle the ends of any preceding grace periods first. */ 1958 if (rdp->completed == rnp->completed && 1959 !unlikely(READ_ONCE(rdp->gpwrap))) { 1960 1961 /* No grace period end, so just accelerate recent callbacks. */ 1962 ret = rcu_accelerate_cbs(rsp, rnp, rdp); 1963 1964 } else { 1965 1966 /* Advance callbacks. */ 1967 ret = rcu_advance_cbs(rsp, rnp, rdp); 1968 1969 /* Remember that we saw this grace-period completion. */ 1970 rdp->completed = rnp->completed; 1971 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend")); 1972 } 1973 1974 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) { 1975 /* 1976 * If the current grace period is waiting for this CPU, 1977 * set up to detect a quiescent state, otherwise don't 1978 * go looking for one. 1979 */ 1980 rdp->gpnum = rnp->gpnum; 1981 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart")); 1982 need_gp = !!(rnp->qsmask & rdp->grpmask); 1983 rdp->cpu_no_qs.b.norm = need_gp; 1984 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); 1985 rdp->core_needs_qs = need_gp; 1986 zero_cpu_stall_ticks(rdp); 1987 WRITE_ONCE(rdp->gpwrap, false); 1988 } 1989 return ret; 1990 } 1991 1992 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp) 1993 { 1994 unsigned long flags; 1995 bool needwake; 1996 struct rcu_node *rnp; 1997 1998 local_irq_save(flags); 1999 rnp = rdp->mynode; 2000 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) && 2001 rdp->completed == READ_ONCE(rnp->completed) && 2002 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */ 2003 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */ 2004 local_irq_restore(flags); 2005 return; 2006 } 2007 needwake = __note_gp_changes(rsp, rnp, rdp); 2008 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 2009 if (needwake) 2010 rcu_gp_kthread_wake(rsp); 2011 } 2012 2013 static void rcu_gp_slow(struct rcu_state *rsp, int delay) 2014 { 2015 if (delay > 0 && 2016 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay))) 2017 schedule_timeout_uninterruptible(delay); 2018 } 2019 2020 /* 2021 * Initialize a new grace period. Return false if no grace period required. 2022 */ 2023 static bool rcu_gp_init(struct rcu_state *rsp) 2024 { 2025 unsigned long oldmask; 2026 struct rcu_data *rdp; 2027 struct rcu_node *rnp = rcu_get_root(rsp); 2028 2029 WRITE_ONCE(rsp->gp_activity, jiffies); 2030 raw_spin_lock_irq_rcu_node(rnp); 2031 if (!READ_ONCE(rsp->gp_flags)) { 2032 /* Spurious wakeup, tell caller to go back to sleep. */ 2033 raw_spin_unlock_irq_rcu_node(rnp); 2034 return false; 2035 } 2036 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */ 2037 2038 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) { 2039 /* 2040 * Grace period already in progress, don't start another. 2041 * Not supposed to be able to happen. 2042 */ 2043 raw_spin_unlock_irq_rcu_node(rnp); 2044 return false; 2045 } 2046 2047 /* Advance to a new grace period and initialize state. */ 2048 record_gp_stall_check_time(rsp); 2049 /* Record GP times before starting GP, hence smp_store_release(). */ 2050 smp_store_release(&rsp->gpnum, rsp->gpnum + 1); 2051 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start")); 2052 raw_spin_unlock_irq_rcu_node(rnp); 2053 2054 /* 2055 * Apply per-leaf buffered online and offline operations to the 2056 * rcu_node tree. Note that this new grace period need not wait 2057 * for subsequent online CPUs, and that quiescent-state forcing 2058 * will handle subsequent offline CPUs. 2059 */ 2060 rcu_for_each_leaf_node(rsp, rnp) { 2061 rcu_gp_slow(rsp, gp_preinit_delay); 2062 raw_spin_lock_irq_rcu_node(rnp); 2063 if (rnp->qsmaskinit == rnp->qsmaskinitnext && 2064 !rnp->wait_blkd_tasks) { 2065 /* Nothing to do on this leaf rcu_node structure. */ 2066 raw_spin_unlock_irq_rcu_node(rnp); 2067 continue; 2068 } 2069 2070 /* Record old state, apply changes to ->qsmaskinit field. */ 2071 oldmask = rnp->qsmaskinit; 2072 rnp->qsmaskinit = rnp->qsmaskinitnext; 2073 2074 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */ 2075 if (!oldmask != !rnp->qsmaskinit) { 2076 if (!oldmask) /* First online CPU for this rcu_node. */ 2077 rcu_init_new_rnp(rnp); 2078 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */ 2079 rnp->wait_blkd_tasks = true; 2080 else /* Last offline CPU and can propagate. */ 2081 rcu_cleanup_dead_rnp(rnp); 2082 } 2083 2084 /* 2085 * If all waited-on tasks from prior grace period are 2086 * done, and if all this rcu_node structure's CPUs are 2087 * still offline, propagate up the rcu_node tree and 2088 * clear ->wait_blkd_tasks. Otherwise, if one of this 2089 * rcu_node structure's CPUs has since come back online, 2090 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp() 2091 * checks for this, so just call it unconditionally). 2092 */ 2093 if (rnp->wait_blkd_tasks && 2094 (!rcu_preempt_has_tasks(rnp) || 2095 rnp->qsmaskinit)) { 2096 rnp->wait_blkd_tasks = false; 2097 rcu_cleanup_dead_rnp(rnp); 2098 } 2099 2100 raw_spin_unlock_irq_rcu_node(rnp); 2101 } 2102 2103 /* 2104 * Set the quiescent-state-needed bits in all the rcu_node 2105 * structures for all currently online CPUs in breadth-first order, 2106 * starting from the root rcu_node structure, relying on the layout 2107 * of the tree within the rsp->node[] array. Note that other CPUs 2108 * will access only the leaves of the hierarchy, thus seeing that no 2109 * grace period is in progress, at least until the corresponding 2110 * leaf node has been initialized. 2111 * 2112 * The grace period cannot complete until the initialization 2113 * process finishes, because this kthread handles both. 2114 */ 2115 rcu_for_each_node_breadth_first(rsp, rnp) { 2116 rcu_gp_slow(rsp, gp_init_delay); 2117 raw_spin_lock_irq_rcu_node(rnp); 2118 rdp = this_cpu_ptr(rsp->rda); 2119 rcu_preempt_check_blocked_tasks(rnp); 2120 rnp->qsmask = rnp->qsmaskinit; 2121 WRITE_ONCE(rnp->gpnum, rsp->gpnum); 2122 if (WARN_ON_ONCE(rnp->completed != rsp->completed)) 2123 WRITE_ONCE(rnp->completed, rsp->completed); 2124 if (rnp == rdp->mynode) 2125 (void)__note_gp_changes(rsp, rnp, rdp); 2126 rcu_preempt_boost_start_gp(rnp); 2127 trace_rcu_grace_period_init(rsp->name, rnp->gpnum, 2128 rnp->level, rnp->grplo, 2129 rnp->grphi, rnp->qsmask); 2130 raw_spin_unlock_irq_rcu_node(rnp); 2131 cond_resched_rcu_qs(); 2132 WRITE_ONCE(rsp->gp_activity, jiffies); 2133 } 2134 2135 return true; 2136 } 2137 2138 /* 2139 * Helper function for wait_event_interruptible_timeout() wakeup 2140 * at force-quiescent-state time. 2141 */ 2142 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp) 2143 { 2144 struct rcu_node *rnp = rcu_get_root(rsp); 2145 2146 /* Someone like call_rcu() requested a force-quiescent-state scan. */ 2147 *gfp = READ_ONCE(rsp->gp_flags); 2148 if (*gfp & RCU_GP_FLAG_FQS) 2149 return true; 2150 2151 /* The current grace period has completed. */ 2152 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp)) 2153 return true; 2154 2155 return false; 2156 } 2157 2158 /* 2159 * Do one round of quiescent-state forcing. 2160 */ 2161 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time) 2162 { 2163 bool isidle = false; 2164 unsigned long maxj; 2165 struct rcu_node *rnp = rcu_get_root(rsp); 2166 2167 WRITE_ONCE(rsp->gp_activity, jiffies); 2168 rsp->n_force_qs++; 2169 if (first_time) { 2170 /* Collect dyntick-idle snapshots. */ 2171 if (is_sysidle_rcu_state(rsp)) { 2172 isidle = true; 2173 maxj = jiffies - ULONG_MAX / 4; 2174 } 2175 force_qs_rnp(rsp, dyntick_save_progress_counter, 2176 &isidle, &maxj); 2177 rcu_sysidle_report_gp(rsp, isidle, maxj); 2178 } else { 2179 /* Handle dyntick-idle and offline CPUs. */ 2180 isidle = true; 2181 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj); 2182 } 2183 /* Clear flag to prevent immediate re-entry. */ 2184 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { 2185 raw_spin_lock_irq_rcu_node(rnp); 2186 WRITE_ONCE(rsp->gp_flags, 2187 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS); 2188 raw_spin_unlock_irq_rcu_node(rnp); 2189 } 2190 } 2191 2192 /* 2193 * Clean up after the old grace period. 2194 */ 2195 static void rcu_gp_cleanup(struct rcu_state *rsp) 2196 { 2197 unsigned long gp_duration; 2198 bool needgp = false; 2199 int nocb = 0; 2200 struct rcu_data *rdp; 2201 struct rcu_node *rnp = rcu_get_root(rsp); 2202 struct swait_queue_head *sq; 2203 2204 WRITE_ONCE(rsp->gp_activity, jiffies); 2205 raw_spin_lock_irq_rcu_node(rnp); 2206 gp_duration = jiffies - rsp->gp_start; 2207 if (gp_duration > rsp->gp_max) 2208 rsp->gp_max = gp_duration; 2209 2210 /* 2211 * We know the grace period is complete, but to everyone else 2212 * it appears to still be ongoing. But it is also the case 2213 * that to everyone else it looks like there is nothing that 2214 * they can do to advance the grace period. It is therefore 2215 * safe for us to drop the lock in order to mark the grace 2216 * period as completed in all of the rcu_node structures. 2217 */ 2218 raw_spin_unlock_irq_rcu_node(rnp); 2219 2220 /* 2221 * Propagate new ->completed value to rcu_node structures so 2222 * that other CPUs don't have to wait until the start of the next 2223 * grace period to process their callbacks. This also avoids 2224 * some nasty RCU grace-period initialization races by forcing 2225 * the end of the current grace period to be completely recorded in 2226 * all of the rcu_node structures before the beginning of the next 2227 * grace period is recorded in any of the rcu_node structures. 2228 */ 2229 rcu_for_each_node_breadth_first(rsp, rnp) { 2230 raw_spin_lock_irq_rcu_node(rnp); 2231 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); 2232 WARN_ON_ONCE(rnp->qsmask); 2233 WRITE_ONCE(rnp->completed, rsp->gpnum); 2234 rdp = this_cpu_ptr(rsp->rda); 2235 if (rnp == rdp->mynode) 2236 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp; 2237 /* smp_mb() provided by prior unlock-lock pair. */ 2238 nocb += rcu_future_gp_cleanup(rsp, rnp); 2239 sq = rcu_nocb_gp_get(rnp); 2240 raw_spin_unlock_irq_rcu_node(rnp); 2241 rcu_nocb_gp_cleanup(sq); 2242 cond_resched_rcu_qs(); 2243 WRITE_ONCE(rsp->gp_activity, jiffies); 2244 rcu_gp_slow(rsp, gp_cleanup_delay); 2245 } 2246 rnp = rcu_get_root(rsp); 2247 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */ 2248 rcu_nocb_gp_set(rnp, nocb); 2249 2250 /* Declare grace period done. */ 2251 WRITE_ONCE(rsp->completed, rsp->gpnum); 2252 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end")); 2253 rsp->gp_state = RCU_GP_IDLE; 2254 rdp = this_cpu_ptr(rsp->rda); 2255 /* Advance CBs to reduce false positives below. */ 2256 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp; 2257 if (needgp || cpu_needs_another_gp(rsp, rdp)) { 2258 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); 2259 trace_rcu_grace_period(rsp->name, 2260 READ_ONCE(rsp->gpnum), 2261 TPS("newreq")); 2262 } 2263 raw_spin_unlock_irq_rcu_node(rnp); 2264 } 2265 2266 /* 2267 * Body of kthread that handles grace periods. 2268 */ 2269 static int __noreturn rcu_gp_kthread(void *arg) 2270 { 2271 bool first_gp_fqs; 2272 int gf; 2273 unsigned long j; 2274 int ret; 2275 struct rcu_state *rsp = arg; 2276 struct rcu_node *rnp = rcu_get_root(rsp); 2277 2278 rcu_bind_gp_kthread(); 2279 for (;;) { 2280 2281 /* Handle grace-period start. */ 2282 for (;;) { 2283 trace_rcu_grace_period(rsp->name, 2284 READ_ONCE(rsp->gpnum), 2285 TPS("reqwait")); 2286 rsp->gp_state = RCU_GP_WAIT_GPS; 2287 swait_event_interruptible(rsp->gp_wq, 2288 READ_ONCE(rsp->gp_flags) & 2289 RCU_GP_FLAG_INIT); 2290 rsp->gp_state = RCU_GP_DONE_GPS; 2291 /* Locking provides needed memory barrier. */ 2292 if (rcu_gp_init(rsp)) 2293 break; 2294 cond_resched_rcu_qs(); 2295 WRITE_ONCE(rsp->gp_activity, jiffies); 2296 WARN_ON(signal_pending(current)); 2297 trace_rcu_grace_period(rsp->name, 2298 READ_ONCE(rsp->gpnum), 2299 TPS("reqwaitsig")); 2300 } 2301 2302 /* Handle quiescent-state forcing. */ 2303 first_gp_fqs = true; 2304 j = jiffies_till_first_fqs; 2305 if (j > HZ) { 2306 j = HZ; 2307 jiffies_till_first_fqs = HZ; 2308 } 2309 ret = 0; 2310 for (;;) { 2311 if (!ret) { 2312 rsp->jiffies_force_qs = jiffies + j; 2313 WRITE_ONCE(rsp->jiffies_kick_kthreads, 2314 jiffies + 3 * j); 2315 } 2316 trace_rcu_grace_period(rsp->name, 2317 READ_ONCE(rsp->gpnum), 2318 TPS("fqswait")); 2319 rsp->gp_state = RCU_GP_WAIT_FQS; 2320 ret = swait_event_interruptible_timeout(rsp->gp_wq, 2321 rcu_gp_fqs_check_wake(rsp, &gf), j); 2322 rsp->gp_state = RCU_GP_DOING_FQS; 2323 /* Locking provides needed memory barriers. */ 2324 /* If grace period done, leave loop. */ 2325 if (!READ_ONCE(rnp->qsmask) && 2326 !rcu_preempt_blocked_readers_cgp(rnp)) 2327 break; 2328 /* If time for quiescent-state forcing, do it. */ 2329 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) || 2330 (gf & RCU_GP_FLAG_FQS)) { 2331 trace_rcu_grace_period(rsp->name, 2332 READ_ONCE(rsp->gpnum), 2333 TPS("fqsstart")); 2334 rcu_gp_fqs(rsp, first_gp_fqs); 2335 first_gp_fqs = false; 2336 trace_rcu_grace_period(rsp->name, 2337 READ_ONCE(rsp->gpnum), 2338 TPS("fqsend")); 2339 cond_resched_rcu_qs(); 2340 WRITE_ONCE(rsp->gp_activity, jiffies); 2341 ret = 0; /* Force full wait till next FQS. */ 2342 j = jiffies_till_next_fqs; 2343 if (j > HZ) { 2344 j = HZ; 2345 jiffies_till_next_fqs = HZ; 2346 } else if (j < 1) { 2347 j = 1; 2348 jiffies_till_next_fqs = 1; 2349 } 2350 } else { 2351 /* Deal with stray signal. */ 2352 cond_resched_rcu_qs(); 2353 WRITE_ONCE(rsp->gp_activity, jiffies); 2354 WARN_ON(signal_pending(current)); 2355 trace_rcu_grace_period(rsp->name, 2356 READ_ONCE(rsp->gpnum), 2357 TPS("fqswaitsig")); 2358 ret = 1; /* Keep old FQS timing. */ 2359 j = jiffies; 2360 if (time_after(jiffies, rsp->jiffies_force_qs)) 2361 j = 1; 2362 else 2363 j = rsp->jiffies_force_qs - j; 2364 } 2365 } 2366 2367 /* Handle grace-period end. */ 2368 rsp->gp_state = RCU_GP_CLEANUP; 2369 rcu_gp_cleanup(rsp); 2370 rsp->gp_state = RCU_GP_CLEANED; 2371 } 2372 } 2373 2374 /* 2375 * Start a new RCU grace period if warranted, re-initializing the hierarchy 2376 * in preparation for detecting the next grace period. The caller must hold 2377 * the root node's ->lock and hard irqs must be disabled. 2378 * 2379 * Note that it is legal for a dying CPU (which is marked as offline) to 2380 * invoke this function. This can happen when the dying CPU reports its 2381 * quiescent state. 2382 * 2383 * Returns true if the grace-period kthread must be awakened. 2384 */ 2385 static bool 2386 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp, 2387 struct rcu_data *rdp) 2388 { 2389 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) { 2390 /* 2391 * Either we have not yet spawned the grace-period 2392 * task, this CPU does not need another grace period, 2393 * or a grace period is already in progress. 2394 * Either way, don't start a new grace period. 2395 */ 2396 return false; 2397 } 2398 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT); 2399 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum), 2400 TPS("newreq")); 2401 2402 /* 2403 * We can't do wakeups while holding the rnp->lock, as that 2404 * could cause possible deadlocks with the rq->lock. Defer 2405 * the wakeup to our caller. 2406 */ 2407 return true; 2408 } 2409 2410 /* 2411 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's 2412 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it 2413 * is invoked indirectly from rcu_advance_cbs(), which would result in 2414 * endless recursion -- or would do so if it wasn't for the self-deadlock 2415 * that is encountered beforehand. 2416 * 2417 * Returns true if the grace-period kthread needs to be awakened. 2418 */ 2419 static bool rcu_start_gp(struct rcu_state *rsp) 2420 { 2421 struct rcu_data *rdp = this_cpu_ptr(rsp->rda); 2422 struct rcu_node *rnp = rcu_get_root(rsp); 2423 bool ret = false; 2424 2425 /* 2426 * If there is no grace period in progress right now, any 2427 * callbacks we have up to this point will be satisfied by the 2428 * next grace period. Also, advancing the callbacks reduces the 2429 * probability of false positives from cpu_needs_another_gp() 2430 * resulting in pointless grace periods. So, advance callbacks 2431 * then start the grace period! 2432 */ 2433 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret; 2434 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret; 2435 return ret; 2436 } 2437 2438 /* 2439 * Report a full set of quiescent states to the specified rcu_state data 2440 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period 2441 * kthread if another grace period is required. Whether we wake 2442 * the grace-period kthread or it awakens itself for the next round 2443 * of quiescent-state forcing, that kthread will clean up after the 2444 * just-completed grace period. Note that the caller must hold rnp->lock, 2445 * which is released before return. 2446 */ 2447 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags) 2448 __releases(rcu_get_root(rsp)->lock) 2449 { 2450 WARN_ON_ONCE(!rcu_gp_in_progress(rsp)); 2451 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); 2452 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags); 2453 rcu_gp_kthread_wake(rsp); 2454 } 2455 2456 /* 2457 * Similar to rcu_report_qs_rdp(), for which it is a helper function. 2458 * Allows quiescent states for a group of CPUs to be reported at one go 2459 * to the specified rcu_node structure, though all the CPUs in the group 2460 * must be represented by the same rcu_node structure (which need not be a 2461 * leaf rcu_node structure, though it often will be). The gps parameter 2462 * is the grace-period snapshot, which means that the quiescent states 2463 * are valid only if rnp->gpnum is equal to gps. That structure's lock 2464 * must be held upon entry, and it is released before return. 2465 */ 2466 static void 2467 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp, 2468 struct rcu_node *rnp, unsigned long gps, unsigned long flags) 2469 __releases(rnp->lock) 2470 { 2471 unsigned long oldmask = 0; 2472 struct rcu_node *rnp_c; 2473 2474 /* Walk up the rcu_node hierarchy. */ 2475 for (;;) { 2476 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) { 2477 2478 /* 2479 * Our bit has already been cleared, or the 2480 * relevant grace period is already over, so done. 2481 */ 2482 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 2483 return; 2484 } 2485 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */ 2486 rnp->qsmask &= ~mask; 2487 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum, 2488 mask, rnp->qsmask, rnp->level, 2489 rnp->grplo, rnp->grphi, 2490 !!rnp->gp_tasks); 2491 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { 2492 2493 /* Other bits still set at this level, so done. */ 2494 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 2495 return; 2496 } 2497 mask = rnp->grpmask; 2498 if (rnp->parent == NULL) { 2499 2500 /* No more levels. Exit loop holding root lock. */ 2501 2502 break; 2503 } 2504 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 2505 rnp_c = rnp; 2506 rnp = rnp->parent; 2507 raw_spin_lock_irqsave_rcu_node(rnp, flags); 2508 oldmask = rnp_c->qsmask; 2509 } 2510 2511 /* 2512 * Get here if we are the last CPU to pass through a quiescent 2513 * state for this grace period. Invoke rcu_report_qs_rsp() 2514 * to clean up and start the next grace period if one is needed. 2515 */ 2516 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */ 2517 } 2518 2519 /* 2520 * Record a quiescent state for all tasks that were previously queued 2521 * on the specified rcu_node structure and that were blocking the current 2522 * RCU grace period. The caller must hold the specified rnp->lock with 2523 * irqs disabled, and this lock is released upon return, but irqs remain 2524 * disabled. 2525 */ 2526 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp, 2527 struct rcu_node *rnp, unsigned long flags) 2528 __releases(rnp->lock) 2529 { 2530 unsigned long gps; 2531 unsigned long mask; 2532 struct rcu_node *rnp_p; 2533 2534 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p || 2535 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { 2536 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 2537 return; /* Still need more quiescent states! */ 2538 } 2539 2540 rnp_p = rnp->parent; 2541 if (rnp_p == NULL) { 2542 /* 2543 * Only one rcu_node structure in the tree, so don't 2544 * try to report up to its nonexistent parent! 2545 */ 2546 rcu_report_qs_rsp(rsp, flags); 2547 return; 2548 } 2549 2550 /* Report up the rest of the hierarchy, tracking current ->gpnum. */ 2551 gps = rnp->gpnum; 2552 mask = rnp->grpmask; 2553 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ 2554 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */ 2555 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags); 2556 } 2557 2558 /* 2559 * Record a quiescent state for the specified CPU to that CPU's rcu_data 2560 * structure. This must be called from the specified CPU. 2561 */ 2562 static void 2563 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp) 2564 { 2565 unsigned long flags; 2566 unsigned long mask; 2567 bool needwake; 2568 struct rcu_node *rnp; 2569 2570 rnp = rdp->mynode; 2571 raw_spin_lock_irqsave_rcu_node(rnp, flags); 2572 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum || 2573 rnp->completed == rnp->gpnum || rdp->gpwrap) { 2574 2575 /* 2576 * The grace period in which this quiescent state was 2577 * recorded has ended, so don't report it upwards. 2578 * We will instead need a new quiescent state that lies 2579 * within the current grace period. 2580 */ 2581 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */ 2582 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr); 2583 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 2584 return; 2585 } 2586 mask = rdp->grpmask; 2587 if ((rnp->qsmask & mask) == 0) { 2588 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 2589 } else { 2590 rdp->core_needs_qs = false; 2591 2592 /* 2593 * This GP can't end until cpu checks in, so all of our 2594 * callbacks can be processed during the next GP. 2595 */ 2596 needwake = rcu_accelerate_cbs(rsp, rnp, rdp); 2597 2598 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); 2599 /* ^^^ Released rnp->lock */ 2600 if (needwake) 2601 rcu_gp_kthread_wake(rsp); 2602 } 2603 } 2604 2605 /* 2606 * Check to see if there is a new grace period of which this CPU 2607 * is not yet aware, and if so, set up local rcu_data state for it. 2608 * Otherwise, see if this CPU has just passed through its first 2609 * quiescent state for this grace period, and record that fact if so. 2610 */ 2611 static void 2612 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp) 2613 { 2614 /* Check for grace-period ends and beginnings. */ 2615 note_gp_changes(rsp, rdp); 2616 2617 /* 2618 * Does this CPU still need to do its part for current grace period? 2619 * If no, return and let the other CPUs do their part as well. 2620 */ 2621 if (!rdp->core_needs_qs) 2622 return; 2623 2624 /* 2625 * Was there a quiescent state since the beginning of the grace 2626 * period? If no, then exit and wait for the next call. 2627 */ 2628 if (rdp->cpu_no_qs.b.norm) 2629 return; 2630 2631 /* 2632 * Tell RCU we are done (but rcu_report_qs_rdp() will be the 2633 * judge of that). 2634 */ 2635 rcu_report_qs_rdp(rdp->cpu, rsp, rdp); 2636 } 2637 2638 /* 2639 * Send the specified CPU's RCU callbacks to the orphanage. The 2640 * specified CPU must be offline, and the caller must hold the 2641 * ->orphan_lock. 2642 */ 2643 static void 2644 rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp, 2645 struct rcu_node *rnp, struct rcu_data *rdp) 2646 { 2647 /* No-CBs CPUs do not have orphanable callbacks. */ 2648 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu)) 2649 return; 2650 2651 /* 2652 * Orphan the callbacks. First adjust the counts. This is safe 2653 * because _rcu_barrier() excludes CPU-hotplug operations, so it 2654 * cannot be running now. Thus no memory barrier is required. 2655 */ 2656 if (rdp->nxtlist != NULL) { 2657 rsp->qlen_lazy += rdp->qlen_lazy; 2658 rsp->qlen += rdp->qlen; 2659 rdp->n_cbs_orphaned += rdp->qlen; 2660 rdp->qlen_lazy = 0; 2661 WRITE_ONCE(rdp->qlen, 0); 2662 } 2663 2664 /* 2665 * Next, move those callbacks still needing a grace period to 2666 * the orphanage, where some other CPU will pick them up. 2667 * Some of the callbacks might have gone partway through a grace 2668 * period, but that is too bad. They get to start over because we 2669 * cannot assume that grace periods are synchronized across CPUs. 2670 * We don't bother updating the ->nxttail[] array yet, instead 2671 * we just reset the whole thing later on. 2672 */ 2673 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) { 2674 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL]; 2675 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL]; 2676 *rdp->nxttail[RCU_DONE_TAIL] = NULL; 2677 } 2678 2679 /* 2680 * Then move the ready-to-invoke callbacks to the orphanage, 2681 * where some other CPU will pick them up. These will not be 2682 * required to pass though another grace period: They are done. 2683 */ 2684 if (rdp->nxtlist != NULL) { 2685 *rsp->orphan_donetail = rdp->nxtlist; 2686 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL]; 2687 } 2688 2689 /* 2690 * Finally, initialize the rcu_data structure's list to empty and 2691 * disallow further callbacks on this CPU. 2692 */ 2693 init_callback_list(rdp); 2694 rdp->nxttail[RCU_NEXT_TAIL] = NULL; 2695 } 2696 2697 /* 2698 * Adopt the RCU callbacks from the specified rcu_state structure's 2699 * orphanage. The caller must hold the ->orphan_lock. 2700 */ 2701 static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags) 2702 { 2703 int i; 2704 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); 2705 2706 /* No-CBs CPUs are handled specially. */ 2707 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || 2708 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags)) 2709 return; 2710 2711 /* Do the accounting first. */ 2712 rdp->qlen_lazy += rsp->qlen_lazy; 2713 rdp->qlen += rsp->qlen; 2714 rdp->n_cbs_adopted += rsp->qlen; 2715 if (rsp->qlen_lazy != rsp->qlen) 2716 rcu_idle_count_callbacks_posted(); 2717 rsp->qlen_lazy = 0; 2718 rsp->qlen = 0; 2719 2720 /* 2721 * We do not need a memory barrier here because the only way we 2722 * can get here if there is an rcu_barrier() in flight is if 2723 * we are the task doing the rcu_barrier(). 2724 */ 2725 2726 /* First adopt the ready-to-invoke callbacks. */ 2727 if (rsp->orphan_donelist != NULL) { 2728 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL]; 2729 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist; 2730 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--) 2731 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) 2732 rdp->nxttail[i] = rsp->orphan_donetail; 2733 rsp->orphan_donelist = NULL; 2734 rsp->orphan_donetail = &rsp->orphan_donelist; 2735 } 2736 2737 /* And then adopt the callbacks that still need a grace period. */ 2738 if (rsp->orphan_nxtlist != NULL) { 2739 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist; 2740 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail; 2741 rsp->orphan_nxtlist = NULL; 2742 rsp->orphan_nxttail = &rsp->orphan_nxtlist; 2743 } 2744 } 2745 2746 /* 2747 * Trace the fact that this CPU is going offline. 2748 */ 2749 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp) 2750 { 2751 RCU_TRACE(unsigned long mask); 2752 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda)); 2753 RCU_TRACE(struct rcu_node *rnp = rdp->mynode); 2754 2755 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) 2756 return; 2757 2758 RCU_TRACE(mask = rdp->grpmask); 2759 trace_rcu_grace_period(rsp->name, 2760 rnp->gpnum + 1 - !!(rnp->qsmask & mask), 2761 TPS("cpuofl")); 2762 } 2763 2764 /* 2765 * All CPUs for the specified rcu_node structure have gone offline, 2766 * and all tasks that were preempted within an RCU read-side critical 2767 * section while running on one of those CPUs have since exited their RCU 2768 * read-side critical section. Some other CPU is reporting this fact with 2769 * the specified rcu_node structure's ->lock held and interrupts disabled. 2770 * This function therefore goes up the tree of rcu_node structures, 2771 * clearing the corresponding bits in the ->qsmaskinit fields. Note that 2772 * the leaf rcu_node structure's ->qsmaskinit field has already been 2773 * updated 2774 * 2775 * This function does check that the specified rcu_node structure has 2776 * all CPUs offline and no blocked tasks, so it is OK to invoke it 2777 * prematurely. That said, invoking it after the fact will cost you 2778 * a needless lock acquisition. So once it has done its work, don't 2779 * invoke it again. 2780 */ 2781 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf) 2782 { 2783 long mask; 2784 struct rcu_node *rnp = rnp_leaf; 2785 2786 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || 2787 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp)) 2788 return; 2789 for (;;) { 2790 mask = rnp->grpmask; 2791 rnp = rnp->parent; 2792 if (!rnp) 2793 break; 2794 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ 2795 rnp->qsmaskinit &= ~mask; 2796 rnp->qsmask &= ~mask; 2797 if (rnp->qsmaskinit) { 2798 raw_spin_unlock_rcu_node(rnp); 2799 /* irqs remain disabled. */ 2800 return; 2801 } 2802 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ 2803 } 2804 } 2805 2806 /* 2807 * The CPU has been completely removed, and some other CPU is reporting 2808 * this fact from process context. Do the remainder of the cleanup, 2809 * including orphaning the outgoing CPU's RCU callbacks, and also 2810 * adopting them. There can only be one CPU hotplug operation at a time, 2811 * so no other CPU can be attempting to update rcu_cpu_kthread_task. 2812 */ 2813 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp) 2814 { 2815 unsigned long flags; 2816 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); 2817 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ 2818 2819 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) 2820 return; 2821 2822 /* Adjust any no-longer-needed kthreads. */ 2823 rcu_boost_kthread_setaffinity(rnp, -1); 2824 2825 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */ 2826 raw_spin_lock_irqsave(&rsp->orphan_lock, flags); 2827 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp); 2828 rcu_adopt_orphan_cbs(rsp, flags); 2829 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags); 2830 2831 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL, 2832 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n", 2833 cpu, rdp->qlen, rdp->nxtlist); 2834 } 2835 2836 /* 2837 * Invoke any RCU callbacks that have made it to the end of their grace 2838 * period. Thottle as specified by rdp->blimit. 2839 */ 2840 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp) 2841 { 2842 unsigned long flags; 2843 struct rcu_head *next, *list, **tail; 2844 long bl, count, count_lazy; 2845 int i; 2846 2847 /* If no callbacks are ready, just return. */ 2848 if (!cpu_has_callbacks_ready_to_invoke(rdp)) { 2849 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0); 2850 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist), 2851 need_resched(), is_idle_task(current), 2852 rcu_is_callbacks_kthread()); 2853 return; 2854 } 2855 2856 /* 2857 * Extract the list of ready callbacks, disabling to prevent 2858 * races with call_rcu() from interrupt handlers. 2859 */ 2860 local_irq_save(flags); 2861 WARN_ON_ONCE(cpu_is_offline(smp_processor_id())); 2862 bl = rdp->blimit; 2863 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl); 2864 list = rdp->nxtlist; 2865 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL]; 2866 *rdp->nxttail[RCU_DONE_TAIL] = NULL; 2867 tail = rdp->nxttail[RCU_DONE_TAIL]; 2868 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--) 2869 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL]) 2870 rdp->nxttail[i] = &rdp->nxtlist; 2871 local_irq_restore(flags); 2872 2873 /* Invoke callbacks. */ 2874 count = count_lazy = 0; 2875 while (list) { 2876 next = list->next; 2877 prefetch(next); 2878 debug_rcu_head_unqueue(list); 2879 if (__rcu_reclaim(rsp->name, list)) 2880 count_lazy++; 2881 list = next; 2882 /* Stop only if limit reached and CPU has something to do. */ 2883 if (++count >= bl && 2884 (need_resched() || 2885 (!is_idle_task(current) && !rcu_is_callbacks_kthread()))) 2886 break; 2887 } 2888 2889 local_irq_save(flags); 2890 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(), 2891 is_idle_task(current), 2892 rcu_is_callbacks_kthread()); 2893 2894 /* Update count, and requeue any remaining callbacks. */ 2895 if (list != NULL) { 2896 *tail = rdp->nxtlist; 2897 rdp->nxtlist = list; 2898 for (i = 0; i < RCU_NEXT_SIZE; i++) 2899 if (&rdp->nxtlist == rdp->nxttail[i]) 2900 rdp->nxttail[i] = tail; 2901 else 2902 break; 2903 } 2904 smp_mb(); /* List handling before counting for rcu_barrier(). */ 2905 rdp->qlen_lazy -= count_lazy; 2906 WRITE_ONCE(rdp->qlen, rdp->qlen - count); 2907 rdp->n_cbs_invoked += count; 2908 2909 /* Reinstate batch limit if we have worked down the excess. */ 2910 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark) 2911 rdp->blimit = blimit; 2912 2913 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */ 2914 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) { 2915 rdp->qlen_last_fqs_check = 0; 2916 rdp->n_force_qs_snap = rsp->n_force_qs; 2917 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark) 2918 rdp->qlen_last_fqs_check = rdp->qlen; 2919 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0)); 2920 2921 local_irq_restore(flags); 2922 2923 /* Re-invoke RCU core processing if there are callbacks remaining. */ 2924 if (cpu_has_callbacks_ready_to_invoke(rdp)) 2925 invoke_rcu_core(); 2926 } 2927 2928 /* 2929 * Check to see if this CPU is in a non-context-switch quiescent state 2930 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh). 2931 * Also schedule RCU core processing. 2932 * 2933 * This function must be called from hardirq context. It is normally 2934 * invoked from the scheduling-clock interrupt. 2935 */ 2936 void rcu_check_callbacks(int user) 2937 { 2938 trace_rcu_utilization(TPS("Start scheduler-tick")); 2939 increment_cpu_stall_ticks(); 2940 if (user || rcu_is_cpu_rrupt_from_idle()) { 2941 2942 /* 2943 * Get here if this CPU took its interrupt from user 2944 * mode or from the idle loop, and if this is not a 2945 * nested interrupt. In this case, the CPU is in 2946 * a quiescent state, so note it. 2947 * 2948 * No memory barrier is required here because both 2949 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local 2950 * variables that other CPUs neither access nor modify, 2951 * at least not while the corresponding CPU is online. 2952 */ 2953 2954 rcu_sched_qs(); 2955 rcu_bh_qs(); 2956 2957 } else if (!in_softirq()) { 2958 2959 /* 2960 * Get here if this CPU did not take its interrupt from 2961 * softirq, in other words, if it is not interrupting 2962 * a rcu_bh read-side critical section. This is an _bh 2963 * critical section, so note it. 2964 */ 2965 2966 rcu_bh_qs(); 2967 } 2968 rcu_preempt_check_callbacks(); 2969 if (rcu_pending()) 2970 invoke_rcu_core(); 2971 if (user) 2972 rcu_note_voluntary_context_switch(current); 2973 trace_rcu_utilization(TPS("End scheduler-tick")); 2974 } 2975 2976 /* 2977 * Scan the leaf rcu_node structures, processing dyntick state for any that 2978 * have not yet encountered a quiescent state, using the function specified. 2979 * Also initiate boosting for any threads blocked on the root rcu_node. 2980 * 2981 * The caller must have suppressed start of new grace periods. 2982 */ 2983 static void force_qs_rnp(struct rcu_state *rsp, 2984 int (*f)(struct rcu_data *rsp, bool *isidle, 2985 unsigned long *maxj), 2986 bool *isidle, unsigned long *maxj) 2987 { 2988 int cpu; 2989 unsigned long flags; 2990 unsigned long mask; 2991 struct rcu_node *rnp; 2992 2993 rcu_for_each_leaf_node(rsp, rnp) { 2994 cond_resched_rcu_qs(); 2995 mask = 0; 2996 raw_spin_lock_irqsave_rcu_node(rnp, flags); 2997 if (rnp->qsmask == 0) { 2998 if (rcu_state_p == &rcu_sched_state || 2999 rsp != rcu_state_p || 3000 rcu_preempt_blocked_readers_cgp(rnp)) { 3001 /* 3002 * No point in scanning bits because they 3003 * are all zero. But we might need to 3004 * priority-boost blocked readers. 3005 */ 3006 rcu_initiate_boost(rnp, flags); 3007 /* rcu_initiate_boost() releases rnp->lock */ 3008 continue; 3009 } 3010 if (rnp->parent && 3011 (rnp->parent->qsmask & rnp->grpmask)) { 3012 /* 3013 * Race between grace-period 3014 * initialization and task exiting RCU 3015 * read-side critical section: Report. 3016 */ 3017 rcu_report_unblock_qs_rnp(rsp, rnp, flags); 3018 /* rcu_report_unblock_qs_rnp() rlses ->lock */ 3019 continue; 3020 } 3021 } 3022 for_each_leaf_node_possible_cpu(rnp, cpu) { 3023 unsigned long bit = leaf_node_cpu_bit(rnp, cpu); 3024 if ((rnp->qsmask & bit) != 0) { 3025 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj)) 3026 mask |= bit; 3027 } 3028 } 3029 if (mask != 0) { 3030 /* Idle/offline CPUs, report (releases rnp->lock. */ 3031 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags); 3032 } else { 3033 /* Nothing to do here, so just drop the lock. */ 3034 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 3035 } 3036 } 3037 } 3038 3039 /* 3040 * Force quiescent states on reluctant CPUs, and also detect which 3041 * CPUs are in dyntick-idle mode. 3042 */ 3043 static void force_quiescent_state(struct rcu_state *rsp) 3044 { 3045 unsigned long flags; 3046 bool ret; 3047 struct rcu_node *rnp; 3048 struct rcu_node *rnp_old = NULL; 3049 3050 /* Funnel through hierarchy to reduce memory contention. */ 3051 rnp = __this_cpu_read(rsp->rda->mynode); 3052 for (; rnp != NULL; rnp = rnp->parent) { 3053 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) || 3054 !raw_spin_trylock(&rnp->fqslock); 3055 if (rnp_old != NULL) 3056 raw_spin_unlock(&rnp_old->fqslock); 3057 if (ret) { 3058 rsp->n_force_qs_lh++; 3059 return; 3060 } 3061 rnp_old = rnp; 3062 } 3063 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */ 3064 3065 /* Reached the root of the rcu_node tree, acquire lock. */ 3066 raw_spin_lock_irqsave_rcu_node(rnp_old, flags); 3067 raw_spin_unlock(&rnp_old->fqslock); 3068 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) { 3069 rsp->n_force_qs_lh++; 3070 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags); 3071 return; /* Someone beat us to it. */ 3072 } 3073 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS); 3074 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags); 3075 rcu_gp_kthread_wake(rsp); 3076 } 3077 3078 /* 3079 * This does the RCU core processing work for the specified rcu_state 3080 * and rcu_data structures. This may be called only from the CPU to 3081 * whom the rdp belongs. 3082 */ 3083 static void 3084 __rcu_process_callbacks(struct rcu_state *rsp) 3085 { 3086 unsigned long flags; 3087 bool needwake; 3088 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); 3089 3090 WARN_ON_ONCE(rdp->beenonline == 0); 3091 3092 /* Update RCU state based on any recent quiescent states. */ 3093 rcu_check_quiescent_state(rsp, rdp); 3094 3095 /* Does this CPU require a not-yet-started grace period? */ 3096 local_irq_save(flags); 3097 if (cpu_needs_another_gp(rsp, rdp)) { 3098 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */ 3099 needwake = rcu_start_gp(rsp); 3100 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags); 3101 if (needwake) 3102 rcu_gp_kthread_wake(rsp); 3103 } else { 3104 local_irq_restore(flags); 3105 } 3106 3107 /* If there are callbacks ready, invoke them. */ 3108 if (cpu_has_callbacks_ready_to_invoke(rdp)) 3109 invoke_rcu_callbacks(rsp, rdp); 3110 3111 /* Do any needed deferred wakeups of rcuo kthreads. */ 3112 do_nocb_deferred_wakeup(rdp); 3113 } 3114 3115 /* 3116 * Do RCU core processing for the current CPU. 3117 */ 3118 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused) 3119 { 3120 struct rcu_state *rsp; 3121 3122 if (cpu_is_offline(smp_processor_id())) 3123 return; 3124 trace_rcu_utilization(TPS("Start RCU core")); 3125 for_each_rcu_flavor(rsp) 3126 __rcu_process_callbacks(rsp); 3127 trace_rcu_utilization(TPS("End RCU core")); 3128 } 3129 3130 /* 3131 * Schedule RCU callback invocation. If the specified type of RCU 3132 * does not support RCU priority boosting, just do a direct call, 3133 * otherwise wake up the per-CPU kernel kthread. Note that because we 3134 * are running on the current CPU with softirqs disabled, the 3135 * rcu_cpu_kthread_task cannot disappear out from under us. 3136 */ 3137 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp) 3138 { 3139 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active))) 3140 return; 3141 if (likely(!rsp->boost)) { 3142 rcu_do_batch(rsp, rdp); 3143 return; 3144 } 3145 invoke_rcu_callbacks_kthread(); 3146 } 3147 3148 static void invoke_rcu_core(void) 3149 { 3150 if (cpu_online(smp_processor_id())) 3151 raise_softirq(RCU_SOFTIRQ); 3152 } 3153 3154 /* 3155 * Handle any core-RCU processing required by a call_rcu() invocation. 3156 */ 3157 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, 3158 struct rcu_head *head, unsigned long flags) 3159 { 3160 bool needwake; 3161 3162 /* 3163 * If called from an extended quiescent state, invoke the RCU 3164 * core in order to force a re-evaluation of RCU's idleness. 3165 */ 3166 if (!rcu_is_watching()) 3167 invoke_rcu_core(); 3168 3169 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */ 3170 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id())) 3171 return; 3172 3173 /* 3174 * Force the grace period if too many callbacks or too long waiting. 3175 * Enforce hysteresis, and don't invoke force_quiescent_state() 3176 * if some other CPU has recently done so. Also, don't bother 3177 * invoking force_quiescent_state() if the newly enqueued callback 3178 * is the only one waiting for a grace period to complete. 3179 */ 3180 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) { 3181 3182 /* Are we ignoring a completed grace period? */ 3183 note_gp_changes(rsp, rdp); 3184 3185 /* Start a new grace period if one not already started. */ 3186 if (!rcu_gp_in_progress(rsp)) { 3187 struct rcu_node *rnp_root = rcu_get_root(rsp); 3188 3189 raw_spin_lock_rcu_node(rnp_root); 3190 needwake = rcu_start_gp(rsp); 3191 raw_spin_unlock_rcu_node(rnp_root); 3192 if (needwake) 3193 rcu_gp_kthread_wake(rsp); 3194 } else { 3195 /* Give the grace period a kick. */ 3196 rdp->blimit = LONG_MAX; 3197 if (rsp->n_force_qs == rdp->n_force_qs_snap && 3198 *rdp->nxttail[RCU_DONE_TAIL] != head) 3199 force_quiescent_state(rsp); 3200 rdp->n_force_qs_snap = rsp->n_force_qs; 3201 rdp->qlen_last_fqs_check = rdp->qlen; 3202 } 3203 } 3204 } 3205 3206 /* 3207 * RCU callback function to leak a callback. 3208 */ 3209 static void rcu_leak_callback(struct rcu_head *rhp) 3210 { 3211 } 3212 3213 /* 3214 * Helper function for call_rcu() and friends. The cpu argument will 3215 * normally be -1, indicating "currently running CPU". It may specify 3216 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier() 3217 * is expected to specify a CPU. 3218 */ 3219 static void 3220 __call_rcu(struct rcu_head *head, rcu_callback_t func, 3221 struct rcu_state *rsp, int cpu, bool lazy) 3222 { 3223 unsigned long flags; 3224 struct rcu_data *rdp; 3225 3226 /* Misaligned rcu_head! */ 3227 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1)); 3228 3229 if (debug_rcu_head_queue(head)) { 3230 /* Probable double call_rcu(), so leak the callback. */ 3231 WRITE_ONCE(head->func, rcu_leak_callback); 3232 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n"); 3233 return; 3234 } 3235 head->func = func; 3236 head->next = NULL; 3237 local_irq_save(flags); 3238 rdp = this_cpu_ptr(rsp->rda); 3239 3240 /* Add the callback to our list. */ 3241 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) { 3242 int offline; 3243 3244 if (cpu != -1) 3245 rdp = per_cpu_ptr(rsp->rda, cpu); 3246 if (likely(rdp->mynode)) { 3247 /* Post-boot, so this should be for a no-CBs CPU. */ 3248 offline = !__call_rcu_nocb(rdp, head, lazy, flags); 3249 WARN_ON_ONCE(offline); 3250 /* Offline CPU, _call_rcu() illegal, leak callback. */ 3251 local_irq_restore(flags); 3252 return; 3253 } 3254 /* 3255 * Very early boot, before rcu_init(). Initialize if needed 3256 * and then drop through to queue the callback. 3257 */ 3258 BUG_ON(cpu != -1); 3259 WARN_ON_ONCE(!rcu_is_watching()); 3260 if (!likely(rdp->nxtlist)) 3261 init_default_callback_list(rdp); 3262 } 3263 WRITE_ONCE(rdp->qlen, rdp->qlen + 1); 3264 if (lazy) 3265 rdp->qlen_lazy++; 3266 else 3267 rcu_idle_count_callbacks_posted(); 3268 smp_mb(); /* Count before adding callback for rcu_barrier(). */ 3269 *rdp->nxttail[RCU_NEXT_TAIL] = head; 3270 rdp->nxttail[RCU_NEXT_TAIL] = &head->next; 3271 3272 if (__is_kfree_rcu_offset((unsigned long)func)) 3273 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func, 3274 rdp->qlen_lazy, rdp->qlen); 3275 else 3276 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen); 3277 3278 /* Go handle any RCU core processing required. */ 3279 __call_rcu_core(rsp, rdp, head, flags); 3280 local_irq_restore(flags); 3281 } 3282 3283 /* 3284 * Queue an RCU-sched callback for invocation after a grace period. 3285 */ 3286 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func) 3287 { 3288 __call_rcu(head, func, &rcu_sched_state, -1, 0); 3289 } 3290 EXPORT_SYMBOL_GPL(call_rcu_sched); 3291 3292 /* 3293 * Queue an RCU callback for invocation after a quicker grace period. 3294 */ 3295 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func) 3296 { 3297 __call_rcu(head, func, &rcu_bh_state, -1, 0); 3298 } 3299 EXPORT_SYMBOL_GPL(call_rcu_bh); 3300 3301 /* 3302 * Queue an RCU callback for lazy invocation after a grace period. 3303 * This will likely be later named something like "call_rcu_lazy()", 3304 * but this change will require some way of tagging the lazy RCU 3305 * callbacks in the list of pending callbacks. Until then, this 3306 * function may only be called from __kfree_rcu(). 3307 */ 3308 void kfree_call_rcu(struct rcu_head *head, 3309 rcu_callback_t func) 3310 { 3311 __call_rcu(head, func, rcu_state_p, -1, 1); 3312 } 3313 EXPORT_SYMBOL_GPL(kfree_call_rcu); 3314 3315 /* 3316 * Because a context switch is a grace period for RCU-sched and RCU-bh, 3317 * any blocking grace-period wait automatically implies a grace period 3318 * if there is only one CPU online at any point time during execution 3319 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to 3320 * occasionally incorrectly indicate that there are multiple CPUs online 3321 * when there was in fact only one the whole time, as this just adds 3322 * some overhead: RCU still operates correctly. 3323 */ 3324 static inline int rcu_blocking_is_gp(void) 3325 { 3326 int ret; 3327 3328 might_sleep(); /* Check for RCU read-side critical section. */ 3329 preempt_disable(); 3330 ret = num_online_cpus() <= 1; 3331 preempt_enable(); 3332 return ret; 3333 } 3334 3335 /** 3336 * synchronize_sched - wait until an rcu-sched grace period has elapsed. 3337 * 3338 * Control will return to the caller some time after a full rcu-sched 3339 * grace period has elapsed, in other words after all currently executing 3340 * rcu-sched read-side critical sections have completed. These read-side 3341 * critical sections are delimited by rcu_read_lock_sched() and 3342 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(), 3343 * local_irq_disable(), and so on may be used in place of 3344 * rcu_read_lock_sched(). 3345 * 3346 * This means that all preempt_disable code sequences, including NMI and 3347 * non-threaded hardware-interrupt handlers, in progress on entry will 3348 * have completed before this primitive returns. However, this does not 3349 * guarantee that softirq handlers will have completed, since in some 3350 * kernels, these handlers can run in process context, and can block. 3351 * 3352 * Note that this guarantee implies further memory-ordering guarantees. 3353 * On systems with more than one CPU, when synchronize_sched() returns, 3354 * each CPU is guaranteed to have executed a full memory barrier since the 3355 * end of its last RCU-sched read-side critical section whose beginning 3356 * preceded the call to synchronize_sched(). In addition, each CPU having 3357 * an RCU read-side critical section that extends beyond the return from 3358 * synchronize_sched() is guaranteed to have executed a full memory barrier 3359 * after the beginning of synchronize_sched() and before the beginning of 3360 * that RCU read-side critical section. Note that these guarantees include 3361 * CPUs that are offline, idle, or executing in user mode, as well as CPUs 3362 * that are executing in the kernel. 3363 * 3364 * Furthermore, if CPU A invoked synchronize_sched(), which returned 3365 * to its caller on CPU B, then both CPU A and CPU B are guaranteed 3366 * to have executed a full memory barrier during the execution of 3367 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but 3368 * again only if the system has more than one CPU). 3369 * 3370 * This primitive provides the guarantees made by the (now removed) 3371 * synchronize_kernel() API. In contrast, synchronize_rcu() only 3372 * guarantees that rcu_read_lock() sections will have completed. 3373 * In "classic RCU", these two guarantees happen to be one and 3374 * the same, but can differ in realtime RCU implementations. 3375 */ 3376 void synchronize_sched(void) 3377 { 3378 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || 3379 lock_is_held(&rcu_lock_map) || 3380 lock_is_held(&rcu_sched_lock_map), 3381 "Illegal synchronize_sched() in RCU-sched read-side critical section"); 3382 if (rcu_blocking_is_gp()) 3383 return; 3384 if (rcu_gp_is_expedited()) 3385 synchronize_sched_expedited(); 3386 else 3387 wait_rcu_gp(call_rcu_sched); 3388 } 3389 EXPORT_SYMBOL_GPL(synchronize_sched); 3390 3391 /** 3392 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed. 3393 * 3394 * Control will return to the caller some time after a full rcu_bh grace 3395 * period has elapsed, in other words after all currently executing rcu_bh 3396 * read-side critical sections have completed. RCU read-side critical 3397 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(), 3398 * and may be nested. 3399 * 3400 * See the description of synchronize_sched() for more detailed information 3401 * on memory ordering guarantees. 3402 */ 3403 void synchronize_rcu_bh(void) 3404 { 3405 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || 3406 lock_is_held(&rcu_lock_map) || 3407 lock_is_held(&rcu_sched_lock_map), 3408 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section"); 3409 if (rcu_blocking_is_gp()) 3410 return; 3411 if (rcu_gp_is_expedited()) 3412 synchronize_rcu_bh_expedited(); 3413 else 3414 wait_rcu_gp(call_rcu_bh); 3415 } 3416 EXPORT_SYMBOL_GPL(synchronize_rcu_bh); 3417 3418 /** 3419 * get_state_synchronize_rcu - Snapshot current RCU state 3420 * 3421 * Returns a cookie that is used by a later call to cond_synchronize_rcu() 3422 * to determine whether or not a full grace period has elapsed in the 3423 * meantime. 3424 */ 3425 unsigned long get_state_synchronize_rcu(void) 3426 { 3427 /* 3428 * Any prior manipulation of RCU-protected data must happen 3429 * before the load from ->gpnum. 3430 */ 3431 smp_mb(); /* ^^^ */ 3432 3433 /* 3434 * Make sure this load happens before the purportedly 3435 * time-consuming work between get_state_synchronize_rcu() 3436 * and cond_synchronize_rcu(). 3437 */ 3438 return smp_load_acquire(&rcu_state_p->gpnum); 3439 } 3440 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu); 3441 3442 /** 3443 * cond_synchronize_rcu - Conditionally wait for an RCU grace period 3444 * 3445 * @oldstate: return value from earlier call to get_state_synchronize_rcu() 3446 * 3447 * If a full RCU grace period has elapsed since the earlier call to 3448 * get_state_synchronize_rcu(), just return. Otherwise, invoke 3449 * synchronize_rcu() to wait for a full grace period. 3450 * 3451 * Yes, this function does not take counter wrap into account. But 3452 * counter wrap is harmless. If the counter wraps, we have waited for 3453 * more than 2 billion grace periods (and way more on a 64-bit system!), 3454 * so waiting for one additional grace period should be just fine. 3455 */ 3456 void cond_synchronize_rcu(unsigned long oldstate) 3457 { 3458 unsigned long newstate; 3459 3460 /* 3461 * Ensure that this load happens before any RCU-destructive 3462 * actions the caller might carry out after we return. 3463 */ 3464 newstate = smp_load_acquire(&rcu_state_p->completed); 3465 if (ULONG_CMP_GE(oldstate, newstate)) 3466 synchronize_rcu(); 3467 } 3468 EXPORT_SYMBOL_GPL(cond_synchronize_rcu); 3469 3470 /** 3471 * get_state_synchronize_sched - Snapshot current RCU-sched state 3472 * 3473 * Returns a cookie that is used by a later call to cond_synchronize_sched() 3474 * to determine whether or not a full grace period has elapsed in the 3475 * meantime. 3476 */ 3477 unsigned long get_state_synchronize_sched(void) 3478 { 3479 /* 3480 * Any prior manipulation of RCU-protected data must happen 3481 * before the load from ->gpnum. 3482 */ 3483 smp_mb(); /* ^^^ */ 3484 3485 /* 3486 * Make sure this load happens before the purportedly 3487 * time-consuming work between get_state_synchronize_sched() 3488 * and cond_synchronize_sched(). 3489 */ 3490 return smp_load_acquire(&rcu_sched_state.gpnum); 3491 } 3492 EXPORT_SYMBOL_GPL(get_state_synchronize_sched); 3493 3494 /** 3495 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period 3496 * 3497 * @oldstate: return value from earlier call to get_state_synchronize_sched() 3498 * 3499 * If a full RCU-sched grace period has elapsed since the earlier call to 3500 * get_state_synchronize_sched(), just return. Otherwise, invoke 3501 * synchronize_sched() to wait for a full grace period. 3502 * 3503 * Yes, this function does not take counter wrap into account. But 3504 * counter wrap is harmless. If the counter wraps, we have waited for 3505 * more than 2 billion grace periods (and way more on a 64-bit system!), 3506 * so waiting for one additional grace period should be just fine. 3507 */ 3508 void cond_synchronize_sched(unsigned long oldstate) 3509 { 3510 unsigned long newstate; 3511 3512 /* 3513 * Ensure that this load happens before any RCU-destructive 3514 * actions the caller might carry out after we return. 3515 */ 3516 newstate = smp_load_acquire(&rcu_sched_state.completed); 3517 if (ULONG_CMP_GE(oldstate, newstate)) 3518 synchronize_sched(); 3519 } 3520 EXPORT_SYMBOL_GPL(cond_synchronize_sched); 3521 3522 /* Adjust sequence number for start of update-side operation. */ 3523 static void rcu_seq_start(unsigned long *sp) 3524 { 3525 WRITE_ONCE(*sp, *sp + 1); 3526 smp_mb(); /* Ensure update-side operation after counter increment. */ 3527 WARN_ON_ONCE(!(*sp & 0x1)); 3528 } 3529 3530 /* Adjust sequence number for end of update-side operation. */ 3531 static void rcu_seq_end(unsigned long *sp) 3532 { 3533 smp_mb(); /* Ensure update-side operation before counter increment. */ 3534 WRITE_ONCE(*sp, *sp + 1); 3535 WARN_ON_ONCE(*sp & 0x1); 3536 } 3537 3538 /* Take a snapshot of the update side's sequence number. */ 3539 static unsigned long rcu_seq_snap(unsigned long *sp) 3540 { 3541 unsigned long s; 3542 3543 s = (READ_ONCE(*sp) + 3) & ~0x1; 3544 smp_mb(); /* Above access must not bleed into critical section. */ 3545 return s; 3546 } 3547 3548 /* 3549 * Given a snapshot from rcu_seq_snap(), determine whether or not a 3550 * full update-side operation has occurred. 3551 */ 3552 static bool rcu_seq_done(unsigned long *sp, unsigned long s) 3553 { 3554 return ULONG_CMP_GE(READ_ONCE(*sp), s); 3555 } 3556 3557 /* 3558 * Check to see if there is any immediate RCU-related work to be done 3559 * by the current CPU, for the specified type of RCU, returning 1 if so. 3560 * The checks are in order of increasing expense: checks that can be 3561 * carried out against CPU-local state are performed first. However, 3562 * we must check for CPU stalls first, else we might not get a chance. 3563 */ 3564 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp) 3565 { 3566 struct rcu_node *rnp = rdp->mynode; 3567 3568 rdp->n_rcu_pending++; 3569 3570 /* Check for CPU stalls, if enabled. */ 3571 check_cpu_stall(rsp, rdp); 3572 3573 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */ 3574 if (rcu_nohz_full_cpu(rsp)) 3575 return 0; 3576 3577 /* Is the RCU core waiting for a quiescent state from this CPU? */ 3578 if (rcu_scheduler_fully_active && 3579 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm && 3580 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) { 3581 rdp->n_rp_core_needs_qs++; 3582 } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) { 3583 rdp->n_rp_report_qs++; 3584 return 1; 3585 } 3586 3587 /* Does this CPU have callbacks ready to invoke? */ 3588 if (cpu_has_callbacks_ready_to_invoke(rdp)) { 3589 rdp->n_rp_cb_ready++; 3590 return 1; 3591 } 3592 3593 /* Has RCU gone idle with this CPU needing another grace period? */ 3594 if (cpu_needs_another_gp(rsp, rdp)) { 3595 rdp->n_rp_cpu_needs_gp++; 3596 return 1; 3597 } 3598 3599 /* Has another RCU grace period completed? */ 3600 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */ 3601 rdp->n_rp_gp_completed++; 3602 return 1; 3603 } 3604 3605 /* Has a new RCU grace period started? */ 3606 if (READ_ONCE(rnp->gpnum) != rdp->gpnum || 3607 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */ 3608 rdp->n_rp_gp_started++; 3609 return 1; 3610 } 3611 3612 /* Does this CPU need a deferred NOCB wakeup? */ 3613 if (rcu_nocb_need_deferred_wakeup(rdp)) { 3614 rdp->n_rp_nocb_defer_wakeup++; 3615 return 1; 3616 } 3617 3618 /* nothing to do */ 3619 rdp->n_rp_need_nothing++; 3620 return 0; 3621 } 3622 3623 /* 3624 * Check to see if there is any immediate RCU-related work to be done 3625 * by the current CPU, returning 1 if so. This function is part of the 3626 * RCU implementation; it is -not- an exported member of the RCU API. 3627 */ 3628 static int rcu_pending(void) 3629 { 3630 struct rcu_state *rsp; 3631 3632 for_each_rcu_flavor(rsp) 3633 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda))) 3634 return 1; 3635 return 0; 3636 } 3637 3638 /* 3639 * Return true if the specified CPU has any callback. If all_lazy is 3640 * non-NULL, store an indication of whether all callbacks are lazy. 3641 * (If there are no callbacks, all of them are deemed to be lazy.) 3642 */ 3643 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy) 3644 { 3645 bool al = true; 3646 bool hc = false; 3647 struct rcu_data *rdp; 3648 struct rcu_state *rsp; 3649 3650 for_each_rcu_flavor(rsp) { 3651 rdp = this_cpu_ptr(rsp->rda); 3652 if (!rdp->nxtlist) 3653 continue; 3654 hc = true; 3655 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) { 3656 al = false; 3657 break; 3658 } 3659 } 3660 if (all_lazy) 3661 *all_lazy = al; 3662 return hc; 3663 } 3664 3665 /* 3666 * Helper function for _rcu_barrier() tracing. If tracing is disabled, 3667 * the compiler is expected to optimize this away. 3668 */ 3669 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s, 3670 int cpu, unsigned long done) 3671 { 3672 trace_rcu_barrier(rsp->name, s, cpu, 3673 atomic_read(&rsp->barrier_cpu_count), done); 3674 } 3675 3676 /* 3677 * RCU callback function for _rcu_barrier(). If we are last, wake 3678 * up the task executing _rcu_barrier(). 3679 */ 3680 static void rcu_barrier_callback(struct rcu_head *rhp) 3681 { 3682 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head); 3683 struct rcu_state *rsp = rdp->rsp; 3684 3685 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) { 3686 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence); 3687 complete(&rsp->barrier_completion); 3688 } else { 3689 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence); 3690 } 3691 } 3692 3693 /* 3694 * Called with preemption disabled, and from cross-cpu IRQ context. 3695 */ 3696 static void rcu_barrier_func(void *type) 3697 { 3698 struct rcu_state *rsp = type; 3699 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda); 3700 3701 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence); 3702 atomic_inc(&rsp->barrier_cpu_count); 3703 rsp->call(&rdp->barrier_head, rcu_barrier_callback); 3704 } 3705 3706 /* 3707 * Orchestrate the specified type of RCU barrier, waiting for all 3708 * RCU callbacks of the specified type to complete. 3709 */ 3710 static void _rcu_barrier(struct rcu_state *rsp) 3711 { 3712 int cpu; 3713 struct rcu_data *rdp; 3714 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence); 3715 3716 _rcu_barrier_trace(rsp, "Begin", -1, s); 3717 3718 /* Take mutex to serialize concurrent rcu_barrier() requests. */ 3719 mutex_lock(&rsp->barrier_mutex); 3720 3721 /* Did someone else do our work for us? */ 3722 if (rcu_seq_done(&rsp->barrier_sequence, s)) { 3723 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence); 3724 smp_mb(); /* caller's subsequent code after above check. */ 3725 mutex_unlock(&rsp->barrier_mutex); 3726 return; 3727 } 3728 3729 /* Mark the start of the barrier operation. */ 3730 rcu_seq_start(&rsp->barrier_sequence); 3731 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence); 3732 3733 /* 3734 * Initialize the count to one rather than to zero in order to 3735 * avoid a too-soon return to zero in case of a short grace period 3736 * (or preemption of this task). Exclude CPU-hotplug operations 3737 * to ensure that no offline CPU has callbacks queued. 3738 */ 3739 init_completion(&rsp->barrier_completion); 3740 atomic_set(&rsp->barrier_cpu_count, 1); 3741 get_online_cpus(); 3742 3743 /* 3744 * Force each CPU with callbacks to register a new callback. 3745 * When that callback is invoked, we will know that all of the 3746 * corresponding CPU's preceding callbacks have been invoked. 3747 */ 3748 for_each_possible_cpu(cpu) { 3749 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu)) 3750 continue; 3751 rdp = per_cpu_ptr(rsp->rda, cpu); 3752 if (rcu_is_nocb_cpu(cpu)) { 3753 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) { 3754 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu, 3755 rsp->barrier_sequence); 3756 } else { 3757 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu, 3758 rsp->barrier_sequence); 3759 smp_mb__before_atomic(); 3760 atomic_inc(&rsp->barrier_cpu_count); 3761 __call_rcu(&rdp->barrier_head, 3762 rcu_barrier_callback, rsp, cpu, 0); 3763 } 3764 } else if (READ_ONCE(rdp->qlen)) { 3765 _rcu_barrier_trace(rsp, "OnlineQ", cpu, 3766 rsp->barrier_sequence); 3767 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1); 3768 } else { 3769 _rcu_barrier_trace(rsp, "OnlineNQ", cpu, 3770 rsp->barrier_sequence); 3771 } 3772 } 3773 put_online_cpus(); 3774 3775 /* 3776 * Now that we have an rcu_barrier_callback() callback on each 3777 * CPU, and thus each counted, remove the initial count. 3778 */ 3779 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) 3780 complete(&rsp->barrier_completion); 3781 3782 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */ 3783 wait_for_completion(&rsp->barrier_completion); 3784 3785 /* Mark the end of the barrier operation. */ 3786 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence); 3787 rcu_seq_end(&rsp->barrier_sequence); 3788 3789 /* Other rcu_barrier() invocations can now safely proceed. */ 3790 mutex_unlock(&rsp->barrier_mutex); 3791 } 3792 3793 /** 3794 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete. 3795 */ 3796 void rcu_barrier_bh(void) 3797 { 3798 _rcu_barrier(&rcu_bh_state); 3799 } 3800 EXPORT_SYMBOL_GPL(rcu_barrier_bh); 3801 3802 /** 3803 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks. 3804 */ 3805 void rcu_barrier_sched(void) 3806 { 3807 _rcu_barrier(&rcu_sched_state); 3808 } 3809 EXPORT_SYMBOL_GPL(rcu_barrier_sched); 3810 3811 /* 3812 * Propagate ->qsinitmask bits up the rcu_node tree to account for the 3813 * first CPU in a given leaf rcu_node structure coming online. The caller 3814 * must hold the corresponding leaf rcu_node ->lock with interrrupts 3815 * disabled. 3816 */ 3817 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf) 3818 { 3819 long mask; 3820 struct rcu_node *rnp = rnp_leaf; 3821 3822 for (;;) { 3823 mask = rnp->grpmask; 3824 rnp = rnp->parent; 3825 if (rnp == NULL) 3826 return; 3827 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */ 3828 rnp->qsmaskinit |= mask; 3829 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */ 3830 } 3831 } 3832 3833 /* 3834 * Do boot-time initialization of a CPU's per-CPU RCU data. 3835 */ 3836 static void __init 3837 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp) 3838 { 3839 unsigned long flags; 3840 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); 3841 struct rcu_node *rnp = rcu_get_root(rsp); 3842 3843 /* Set up local state, ensuring consistent view of global state. */ 3844 raw_spin_lock_irqsave_rcu_node(rnp, flags); 3845 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu); 3846 rdp->dynticks = &per_cpu(rcu_dynticks, cpu); 3847 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE); 3848 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks))); 3849 rdp->cpu = cpu; 3850 rdp->rsp = rsp; 3851 rcu_boot_init_nocb_percpu_data(rdp); 3852 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 3853 } 3854 3855 /* 3856 * Initialize a CPU's per-CPU RCU data. Note that only one online or 3857 * offline event can be happening at a given time. Note also that we 3858 * can accept some slop in the rsp->completed access due to the fact 3859 * that this CPU cannot possibly have any RCU callbacks in flight yet. 3860 */ 3861 static void 3862 rcu_init_percpu_data(int cpu, struct rcu_state *rsp) 3863 { 3864 unsigned long flags; 3865 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); 3866 struct rcu_node *rnp = rcu_get_root(rsp); 3867 3868 /* Set up local state, ensuring consistent view of global state. */ 3869 raw_spin_lock_irqsave_rcu_node(rnp, flags); 3870 rdp->qlen_last_fqs_check = 0; 3871 rdp->n_force_qs_snap = rsp->n_force_qs; 3872 rdp->blimit = blimit; 3873 if (!rdp->nxtlist) 3874 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */ 3875 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE; 3876 rcu_sysidle_init_percpu_data(rdp->dynticks); 3877 rcu_dynticks_eqs_online(); 3878 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ 3879 3880 /* 3881 * Add CPU to leaf rcu_node pending-online bitmask. Any needed 3882 * propagation up the rcu_node tree will happen at the beginning 3883 * of the next grace period. 3884 */ 3885 rnp = rdp->mynode; 3886 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ 3887 if (!rdp->beenonline) 3888 WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1); 3889 rdp->beenonline = true; /* We have now been online. */ 3890 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */ 3891 rdp->completed = rnp->completed; 3892 rdp->cpu_no_qs.b.norm = true; 3893 rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu); 3894 rdp->core_needs_qs = false; 3895 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl")); 3896 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 3897 } 3898 3899 int rcutree_prepare_cpu(unsigned int cpu) 3900 { 3901 struct rcu_state *rsp; 3902 3903 for_each_rcu_flavor(rsp) 3904 rcu_init_percpu_data(cpu, rsp); 3905 3906 rcu_prepare_kthreads(cpu); 3907 rcu_spawn_all_nocb_kthreads(cpu); 3908 3909 return 0; 3910 } 3911 3912 static void rcutree_affinity_setting(unsigned int cpu, int outgoing) 3913 { 3914 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); 3915 3916 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing); 3917 } 3918 3919 int rcutree_online_cpu(unsigned int cpu) 3920 { 3921 sync_sched_exp_online_cleanup(cpu); 3922 rcutree_affinity_setting(cpu, -1); 3923 return 0; 3924 } 3925 3926 int rcutree_offline_cpu(unsigned int cpu) 3927 { 3928 rcutree_affinity_setting(cpu, cpu); 3929 return 0; 3930 } 3931 3932 3933 int rcutree_dying_cpu(unsigned int cpu) 3934 { 3935 struct rcu_state *rsp; 3936 3937 for_each_rcu_flavor(rsp) 3938 rcu_cleanup_dying_cpu(rsp); 3939 return 0; 3940 } 3941 3942 int rcutree_dead_cpu(unsigned int cpu) 3943 { 3944 struct rcu_state *rsp; 3945 3946 for_each_rcu_flavor(rsp) { 3947 rcu_cleanup_dead_cpu(cpu, rsp); 3948 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu)); 3949 } 3950 return 0; 3951 } 3952 3953 /* 3954 * Mark the specified CPU as being online so that subsequent grace periods 3955 * (both expedited and normal) will wait on it. Note that this means that 3956 * incoming CPUs are not allowed to use RCU read-side critical sections 3957 * until this function is called. Failing to observe this restriction 3958 * will result in lockdep splats. 3959 */ 3960 void rcu_cpu_starting(unsigned int cpu) 3961 { 3962 unsigned long flags; 3963 unsigned long mask; 3964 struct rcu_data *rdp; 3965 struct rcu_node *rnp; 3966 struct rcu_state *rsp; 3967 3968 for_each_rcu_flavor(rsp) { 3969 rdp = per_cpu_ptr(rsp->rda, cpu); 3970 rnp = rdp->mynode; 3971 mask = rdp->grpmask; 3972 raw_spin_lock_irqsave_rcu_node(rnp, flags); 3973 rnp->qsmaskinitnext |= mask; 3974 rnp->expmaskinitnext |= mask; 3975 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 3976 } 3977 } 3978 3979 #ifdef CONFIG_HOTPLUG_CPU 3980 /* 3981 * The CPU is exiting the idle loop into the arch_cpu_idle_dead() 3982 * function. We now remove it from the rcu_node tree's ->qsmaskinit 3983 * bit masks. 3984 * The CPU is exiting the idle loop into the arch_cpu_idle_dead() 3985 * function. We now remove it from the rcu_node tree's ->qsmaskinit 3986 * bit masks. 3987 */ 3988 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp) 3989 { 3990 unsigned long flags; 3991 unsigned long mask; 3992 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); 3993 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */ 3994 3995 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */ 3996 mask = rdp->grpmask; 3997 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */ 3998 rnp->qsmaskinitnext &= ~mask; 3999 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 4000 } 4001 4002 void rcu_report_dead(unsigned int cpu) 4003 { 4004 struct rcu_state *rsp; 4005 4006 /* QS for any half-done expedited RCU-sched GP. */ 4007 preempt_disable(); 4008 rcu_report_exp_rdp(&rcu_sched_state, 4009 this_cpu_ptr(rcu_sched_state.rda), true); 4010 preempt_enable(); 4011 for_each_rcu_flavor(rsp) 4012 rcu_cleanup_dying_idle_cpu(cpu, rsp); 4013 } 4014 #endif 4015 4016 static int rcu_pm_notify(struct notifier_block *self, 4017 unsigned long action, void *hcpu) 4018 { 4019 switch (action) { 4020 case PM_HIBERNATION_PREPARE: 4021 case PM_SUSPEND_PREPARE: 4022 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ 4023 rcu_expedite_gp(); 4024 break; 4025 case PM_POST_HIBERNATION: 4026 case PM_POST_SUSPEND: 4027 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */ 4028 rcu_unexpedite_gp(); 4029 break; 4030 default: 4031 break; 4032 } 4033 return NOTIFY_OK; 4034 } 4035 4036 /* 4037 * Spawn the kthreads that handle each RCU flavor's grace periods. 4038 */ 4039 static int __init rcu_spawn_gp_kthread(void) 4040 { 4041 unsigned long flags; 4042 int kthread_prio_in = kthread_prio; 4043 struct rcu_node *rnp; 4044 struct rcu_state *rsp; 4045 struct sched_param sp; 4046 struct task_struct *t; 4047 4048 /* Force priority into range. */ 4049 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1) 4050 kthread_prio = 1; 4051 else if (kthread_prio < 0) 4052 kthread_prio = 0; 4053 else if (kthread_prio > 99) 4054 kthread_prio = 99; 4055 if (kthread_prio != kthread_prio_in) 4056 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n", 4057 kthread_prio, kthread_prio_in); 4058 4059 rcu_scheduler_fully_active = 1; 4060 for_each_rcu_flavor(rsp) { 4061 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name); 4062 BUG_ON(IS_ERR(t)); 4063 rnp = rcu_get_root(rsp); 4064 raw_spin_lock_irqsave_rcu_node(rnp, flags); 4065 rsp->gp_kthread = t; 4066 if (kthread_prio) { 4067 sp.sched_priority = kthread_prio; 4068 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); 4069 } 4070 raw_spin_unlock_irqrestore_rcu_node(rnp, flags); 4071 wake_up_process(t); 4072 } 4073 rcu_spawn_nocb_kthreads(); 4074 rcu_spawn_boost_kthreads(); 4075 return 0; 4076 } 4077 early_initcall(rcu_spawn_gp_kthread); 4078 4079 /* 4080 * This function is invoked towards the end of the scheduler's 4081 * initialization process. Before this is called, the idle task might 4082 * contain synchronous grace-period primitives (during which time, this idle 4083 * task is booting the system, and such primitives are no-ops). After this 4084 * function is called, any synchronous grace-period primitives are run as 4085 * expedited, with the requesting task driving the grace period forward. 4086 * A later core_initcall() rcu_exp_runtime_mode() will switch to full 4087 * runtime RCU functionality. 4088 */ 4089 void rcu_scheduler_starting(void) 4090 { 4091 WARN_ON(num_online_cpus() != 1); 4092 WARN_ON(nr_context_switches() > 0); 4093 rcu_test_sync_prims(); 4094 rcu_scheduler_active = RCU_SCHEDULER_INIT; 4095 rcu_test_sync_prims(); 4096 } 4097 4098 /* 4099 * Compute the per-level fanout, either using the exact fanout specified 4100 * or balancing the tree, depending on the rcu_fanout_exact boot parameter. 4101 */ 4102 static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt) 4103 { 4104 int i; 4105 4106 if (rcu_fanout_exact) { 4107 levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf; 4108 for (i = rcu_num_lvls - 2; i >= 0; i--) 4109 levelspread[i] = RCU_FANOUT; 4110 } else { 4111 int ccur; 4112 int cprv; 4113 4114 cprv = nr_cpu_ids; 4115 for (i = rcu_num_lvls - 1; i >= 0; i--) { 4116 ccur = levelcnt[i]; 4117 levelspread[i] = (cprv + ccur - 1) / ccur; 4118 cprv = ccur; 4119 } 4120 } 4121 } 4122 4123 /* 4124 * Helper function for rcu_init() that initializes one rcu_state structure. 4125 */ 4126 static void __init rcu_init_one(struct rcu_state *rsp) 4127 { 4128 static const char * const buf[] = RCU_NODE_NAME_INIT; 4129 static const char * const fqs[] = RCU_FQS_NAME_INIT; 4130 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS]; 4131 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS]; 4132 static u8 fl_mask = 0x1; 4133 4134 int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */ 4135 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */ 4136 int cpustride = 1; 4137 int i; 4138 int j; 4139 struct rcu_node *rnp; 4140 4141 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */ 4142 4143 /* Silence gcc 4.8 false positive about array index out of range. */ 4144 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS) 4145 panic("rcu_init_one: rcu_num_lvls out of range"); 4146 4147 /* Initialize the level-tracking arrays. */ 4148 4149 for (i = 0; i < rcu_num_lvls; i++) 4150 levelcnt[i] = num_rcu_lvl[i]; 4151 for (i = 1; i < rcu_num_lvls; i++) 4152 rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1]; 4153 rcu_init_levelspread(levelspread, levelcnt); 4154 rsp->flavor_mask = fl_mask; 4155 fl_mask <<= 1; 4156 4157 /* Initialize the elements themselves, starting from the leaves. */ 4158 4159 for (i = rcu_num_lvls - 1; i >= 0; i--) { 4160 cpustride *= levelspread[i]; 4161 rnp = rsp->level[i]; 4162 for (j = 0; j < levelcnt[i]; j++, rnp++) { 4163 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock)); 4164 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock), 4165 &rcu_node_class[i], buf[i]); 4166 raw_spin_lock_init(&rnp->fqslock); 4167 lockdep_set_class_and_name(&rnp->fqslock, 4168 &rcu_fqs_class[i], fqs[i]); 4169 rnp->gpnum = rsp->gpnum; 4170 rnp->completed = rsp->completed; 4171 rnp->qsmask = 0; 4172 rnp->qsmaskinit = 0; 4173 rnp->grplo = j * cpustride; 4174 rnp->grphi = (j + 1) * cpustride - 1; 4175 if (rnp->grphi >= nr_cpu_ids) 4176 rnp->grphi = nr_cpu_ids - 1; 4177 if (i == 0) { 4178 rnp->grpnum = 0; 4179 rnp->grpmask = 0; 4180 rnp->parent = NULL; 4181 } else { 4182 rnp->grpnum = j % levelspread[i - 1]; 4183 rnp->grpmask = 1UL << rnp->grpnum; 4184 rnp->parent = rsp->level[i - 1] + 4185 j / levelspread[i - 1]; 4186 } 4187 rnp->level = i; 4188 INIT_LIST_HEAD(&rnp->blkd_tasks); 4189 rcu_init_one_nocb(rnp); 4190 init_waitqueue_head(&rnp->exp_wq[0]); 4191 init_waitqueue_head(&rnp->exp_wq[1]); 4192 init_waitqueue_head(&rnp->exp_wq[2]); 4193 init_waitqueue_head(&rnp->exp_wq[3]); 4194 spin_lock_init(&rnp->exp_lock); 4195 } 4196 } 4197 4198 init_swait_queue_head(&rsp->gp_wq); 4199 init_swait_queue_head(&rsp->expedited_wq); 4200 rnp = rsp->level[rcu_num_lvls - 1]; 4201 for_each_possible_cpu(i) { 4202 while (i > rnp->grphi) 4203 rnp++; 4204 per_cpu_ptr(rsp->rda, i)->mynode = rnp; 4205 rcu_boot_init_percpu_data(i, rsp); 4206 } 4207 list_add(&rsp->flavors, &rcu_struct_flavors); 4208 } 4209 4210 /* 4211 * Compute the rcu_node tree geometry from kernel parameters. This cannot 4212 * replace the definitions in tree.h because those are needed to size 4213 * the ->node array in the rcu_state structure. 4214 */ 4215 static void __init rcu_init_geometry(void) 4216 { 4217 ulong d; 4218 int i; 4219 int rcu_capacity[RCU_NUM_LVLS]; 4220 4221 /* 4222 * Initialize any unspecified boot parameters. 4223 * The default values of jiffies_till_first_fqs and 4224 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS 4225 * value, which is a function of HZ, then adding one for each 4226 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system. 4227 */ 4228 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV; 4229 if (jiffies_till_first_fqs == ULONG_MAX) 4230 jiffies_till_first_fqs = d; 4231 if (jiffies_till_next_fqs == ULONG_MAX) 4232 jiffies_till_next_fqs = d; 4233 4234 /* If the compile-time values are accurate, just leave. */ 4235 if (rcu_fanout_leaf == RCU_FANOUT_LEAF && 4236 nr_cpu_ids == NR_CPUS) 4237 return; 4238 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n", 4239 rcu_fanout_leaf, nr_cpu_ids); 4240 4241 /* 4242 * The boot-time rcu_fanout_leaf parameter must be at least two 4243 * and cannot exceed the number of bits in the rcu_node masks. 4244 * Complain and fall back to the compile-time values if this 4245 * limit is exceeded. 4246 */ 4247 if (rcu_fanout_leaf < 2 || 4248 rcu_fanout_leaf > sizeof(unsigned long) * 8) { 4249 rcu_fanout_leaf = RCU_FANOUT_LEAF; 4250 WARN_ON(1); 4251 return; 4252 } 4253 4254 /* 4255 * Compute number of nodes that can be handled an rcu_node tree 4256 * with the given number of levels. 4257 */ 4258 rcu_capacity[0] = rcu_fanout_leaf; 4259 for (i = 1; i < RCU_NUM_LVLS; i++) 4260 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT; 4261 4262 /* 4263 * The tree must be able to accommodate the configured number of CPUs. 4264 * If this limit is exceeded, fall back to the compile-time values. 4265 */ 4266 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) { 4267 rcu_fanout_leaf = RCU_FANOUT_LEAF; 4268 WARN_ON(1); 4269 return; 4270 } 4271 4272 /* Calculate the number of levels in the tree. */ 4273 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) { 4274 } 4275 rcu_num_lvls = i + 1; 4276 4277 /* Calculate the number of rcu_nodes at each level of the tree. */ 4278 for (i = 0; i < rcu_num_lvls; i++) { 4279 int cap = rcu_capacity[(rcu_num_lvls - 1) - i]; 4280 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap); 4281 } 4282 4283 /* Calculate the total number of rcu_node structures. */ 4284 rcu_num_nodes = 0; 4285 for (i = 0; i < rcu_num_lvls; i++) 4286 rcu_num_nodes += num_rcu_lvl[i]; 4287 } 4288 4289 /* 4290 * Dump out the structure of the rcu_node combining tree associated 4291 * with the rcu_state structure referenced by rsp. 4292 */ 4293 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp) 4294 { 4295 int level = 0; 4296 struct rcu_node *rnp; 4297 4298 pr_info("rcu_node tree layout dump\n"); 4299 pr_info(" "); 4300 rcu_for_each_node_breadth_first(rsp, rnp) { 4301 if (rnp->level != level) { 4302 pr_cont("\n"); 4303 pr_info(" "); 4304 level = rnp->level; 4305 } 4306 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum); 4307 } 4308 pr_cont("\n"); 4309 } 4310 4311 void __init rcu_init(void) 4312 { 4313 int cpu; 4314 4315 rcu_early_boot_tests(); 4316 4317 rcu_bootup_announce(); 4318 rcu_init_geometry(); 4319 rcu_init_one(&rcu_bh_state); 4320 rcu_init_one(&rcu_sched_state); 4321 if (dump_tree) 4322 rcu_dump_rcu_node_tree(&rcu_sched_state); 4323 __rcu_init_preempt(); 4324 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); 4325 4326 /* 4327 * We don't need protection against CPU-hotplug here because 4328 * this is called early in boot, before either interrupts 4329 * or the scheduler are operational. 4330 */ 4331 pm_notifier(rcu_pm_notify, 0); 4332 for_each_online_cpu(cpu) { 4333 rcutree_prepare_cpu(cpu); 4334 rcu_cpu_starting(cpu); 4335 } 4336 } 4337 4338 #include "tree_exp.h" 4339 #include "tree_plugin.h" 4340