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