| 1 | /* |
| 2 | * Read-Copy Update mechanism for mutual exclusion (tree-based version) |
| 3 | * Internal non-public definitions that provide either classic |
| 4 | * or preemptible semantics. |
| 5 | * |
| 6 | * This program is free software; you can redistribute it and/or modify |
| 7 | * it under the terms of the GNU General Public License as published by |
| 8 | * the Free Software Foundation; either version 2 of the License, or |
| 9 | * (at your option) any later version. |
| 10 | * |
| 11 | * This program is distributed in the hope that it will be useful, |
| 12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 14 | * GNU General Public License for more details. |
| 15 | * |
| 16 | * You should have received a copy of the GNU General Public License |
| 17 | * along with this program; if not, you can access it online at |
| 18 | * http://www.gnu.org/licenses/gpl-2.0.html. |
| 19 | * |
| 20 | * Copyright Red Hat, 2009 |
| 21 | * Copyright IBM Corporation, 2009 |
| 22 | * |
| 23 | * Author: Ingo Molnar <mingo@elte.hu> |
| 24 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> |
| 25 | */ |
| 26 | |
| 27 | #include <linux/delay.h> |
| 28 | #include <linux/gfp.h> |
| 29 | #include <linux/oom.h> |
| 30 | #include <linux/smpboot.h> |
| 31 | #include "../time/tick-internal.h" |
| 32 | |
| 33 | #define RCU_KTHREAD_PRIO 1 |
| 34 | |
| 35 | #ifdef CONFIG_RCU_BOOST |
| 36 | #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO |
| 37 | #else |
| 38 | #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO |
| 39 | #endif |
| 40 | |
| 41 | #ifdef CONFIG_RCU_NOCB_CPU |
| 42 | static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */ |
| 43 | static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */ |
| 44 | static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */ |
| 45 | static char __initdata nocb_buf[NR_CPUS * 5]; |
| 46 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ |
| 47 | |
| 48 | /* |
| 49 | * Check the RCU kernel configuration parameters and print informative |
| 50 | * messages about anything out of the ordinary. If you like #ifdef, you |
| 51 | * will love this function. |
| 52 | */ |
| 53 | static void __init rcu_bootup_announce_oddness(void) |
| 54 | { |
| 55 | #ifdef CONFIG_RCU_TRACE |
| 56 | pr_info("\tRCU debugfs-based tracing is enabled.\n"); |
| 57 | #endif |
| 58 | #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32) |
| 59 | pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n", |
| 60 | CONFIG_RCU_FANOUT); |
| 61 | #endif |
| 62 | #ifdef CONFIG_RCU_FANOUT_EXACT |
| 63 | pr_info("\tHierarchical RCU autobalancing is disabled.\n"); |
| 64 | #endif |
| 65 | #ifdef CONFIG_RCU_FAST_NO_HZ |
| 66 | pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n"); |
| 67 | #endif |
| 68 | #ifdef CONFIG_PROVE_RCU |
| 69 | pr_info("\tRCU lockdep checking is enabled.\n"); |
| 70 | #endif |
| 71 | #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE |
| 72 | pr_info("\tRCU torture testing starts during boot.\n"); |
| 73 | #endif |
| 74 | #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE) |
| 75 | pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n"); |
| 76 | #endif |
| 77 | #if defined(CONFIG_RCU_CPU_STALL_INFO) |
| 78 | pr_info("\tAdditional per-CPU info printed with stalls.\n"); |
| 79 | #endif |
| 80 | #if NUM_RCU_LVL_4 != 0 |
| 81 | pr_info("\tFour-level hierarchy is enabled.\n"); |
| 82 | #endif |
| 83 | if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF) |
| 84 | pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf); |
| 85 | if (nr_cpu_ids != NR_CPUS) |
| 86 | pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids); |
| 87 | #ifdef CONFIG_RCU_NOCB_CPU |
| 88 | #ifndef CONFIG_RCU_NOCB_CPU_NONE |
| 89 | if (!have_rcu_nocb_mask) { |
| 90 | zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL); |
| 91 | have_rcu_nocb_mask = true; |
| 92 | } |
| 93 | #ifdef CONFIG_RCU_NOCB_CPU_ZERO |
| 94 | pr_info("\tOffload RCU callbacks from CPU 0\n"); |
| 95 | cpumask_set_cpu(0, rcu_nocb_mask); |
| 96 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */ |
| 97 | #ifdef CONFIG_RCU_NOCB_CPU_ALL |
| 98 | pr_info("\tOffload RCU callbacks from all CPUs\n"); |
| 99 | cpumask_copy(rcu_nocb_mask, cpu_possible_mask); |
| 100 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */ |
| 101 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */ |
| 102 | if (have_rcu_nocb_mask) { |
| 103 | if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) { |
| 104 | pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n"); |
| 105 | cpumask_and(rcu_nocb_mask, cpu_possible_mask, |
| 106 | rcu_nocb_mask); |
| 107 | } |
| 108 | cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask); |
| 109 | pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf); |
| 110 | if (rcu_nocb_poll) |
| 111 | pr_info("\tPoll for callbacks from no-CBs CPUs.\n"); |
| 112 | } |
| 113 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ |
| 114 | } |
| 115 | |
| 116 | #ifdef CONFIG_TREE_PREEMPT_RCU |
| 117 | |
| 118 | RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu); |
| 119 | static struct rcu_state *rcu_state_p = &rcu_preempt_state; |
| 120 | |
| 121 | static int rcu_preempted_readers_exp(struct rcu_node *rnp); |
| 122 | |
| 123 | /* |
| 124 | * Tell them what RCU they are running. |
| 125 | */ |
| 126 | static void __init rcu_bootup_announce(void) |
| 127 | { |
| 128 | pr_info("Preemptible hierarchical RCU implementation.\n"); |
| 129 | rcu_bootup_announce_oddness(); |
| 130 | } |
| 131 | |
| 132 | /* |
| 133 | * Return the number of RCU-preempt batches processed thus far |
| 134 | * for debug and statistics. |
| 135 | */ |
| 136 | long rcu_batches_completed_preempt(void) |
| 137 | { |
| 138 | return rcu_preempt_state.completed; |
| 139 | } |
| 140 | EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt); |
| 141 | |
| 142 | /* |
| 143 | * Return the number of RCU batches processed thus far for debug & stats. |
| 144 | */ |
| 145 | long rcu_batches_completed(void) |
| 146 | { |
| 147 | return rcu_batches_completed_preempt(); |
| 148 | } |
| 149 | EXPORT_SYMBOL_GPL(rcu_batches_completed); |
| 150 | |
| 151 | /* |
| 152 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
| 153 | * that this just means that the task currently running on the CPU is |
| 154 | * not in a quiescent state. There might be any number of tasks blocked |
| 155 | * while in an RCU read-side critical section. |
| 156 | * |
| 157 | * Unlike the other rcu_*_qs() functions, callers to this function |
| 158 | * must disable irqs in order to protect the assignment to |
| 159 | * ->rcu_read_unlock_special. |
| 160 | */ |
| 161 | static void rcu_preempt_qs(int cpu) |
| 162 | { |
| 163 | struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu); |
| 164 | |
| 165 | if (rdp->passed_quiesce == 0) |
| 166 | trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs")); |
| 167 | rdp->passed_quiesce = 1; |
| 168 | current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS; |
| 169 | } |
| 170 | |
| 171 | /* |
| 172 | * We have entered the scheduler, and the current task might soon be |
| 173 | * context-switched away from. If this task is in an RCU read-side |
| 174 | * critical section, we will no longer be able to rely on the CPU to |
| 175 | * record that fact, so we enqueue the task on the blkd_tasks list. |
| 176 | * The task will dequeue itself when it exits the outermost enclosing |
| 177 | * RCU read-side critical section. Therefore, the current grace period |
| 178 | * cannot be permitted to complete until the blkd_tasks list entries |
| 179 | * predating the current grace period drain, in other words, until |
| 180 | * rnp->gp_tasks becomes NULL. |
| 181 | * |
| 182 | * Caller must disable preemption. |
| 183 | */ |
| 184 | static void rcu_preempt_note_context_switch(int cpu) |
| 185 | { |
| 186 | struct task_struct *t = current; |
| 187 | unsigned long flags; |
| 188 | struct rcu_data *rdp; |
| 189 | struct rcu_node *rnp; |
| 190 | |
| 191 | if (t->rcu_read_lock_nesting > 0 && |
| 192 | (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) { |
| 193 | |
| 194 | /* Possibly blocking in an RCU read-side critical section. */ |
| 195 | rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu); |
| 196 | rnp = rdp->mynode; |
| 197 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 198 | smp_mb__after_unlock_lock(); |
| 199 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED; |
| 200 | t->rcu_blocked_node = rnp; |
| 201 | |
| 202 | /* |
| 203 | * If this CPU has already checked in, then this task |
| 204 | * will hold up the next grace period rather than the |
| 205 | * current grace period. Queue the task accordingly. |
| 206 | * If the task is queued for the current grace period |
| 207 | * (i.e., this CPU has not yet passed through a quiescent |
| 208 | * state for the current grace period), then as long |
| 209 | * as that task remains queued, the current grace period |
| 210 | * cannot end. Note that there is some uncertainty as |
| 211 | * to exactly when the current grace period started. |
| 212 | * We take a conservative approach, which can result |
| 213 | * in unnecessarily waiting on tasks that started very |
| 214 | * slightly after the current grace period began. C'est |
| 215 | * la vie!!! |
| 216 | * |
| 217 | * But first, note that the current CPU must still be |
| 218 | * on line! |
| 219 | */ |
| 220 | WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0); |
| 221 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
| 222 | if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) { |
| 223 | list_add(&t->rcu_node_entry, rnp->gp_tasks->prev); |
| 224 | rnp->gp_tasks = &t->rcu_node_entry; |
| 225 | #ifdef CONFIG_RCU_BOOST |
| 226 | if (rnp->boost_tasks != NULL) |
| 227 | rnp->boost_tasks = rnp->gp_tasks; |
| 228 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 229 | } else { |
| 230 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); |
| 231 | if (rnp->qsmask & rdp->grpmask) |
| 232 | rnp->gp_tasks = &t->rcu_node_entry; |
| 233 | } |
| 234 | trace_rcu_preempt_task(rdp->rsp->name, |
| 235 | t->pid, |
| 236 | (rnp->qsmask & rdp->grpmask) |
| 237 | ? rnp->gpnum |
| 238 | : rnp->gpnum + 1); |
| 239 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 240 | } else if (t->rcu_read_lock_nesting < 0 && |
| 241 | t->rcu_read_unlock_special) { |
| 242 | |
| 243 | /* |
| 244 | * Complete exit from RCU read-side critical section on |
| 245 | * behalf of preempted instance of __rcu_read_unlock(). |
| 246 | */ |
| 247 | rcu_read_unlock_special(t); |
| 248 | } |
| 249 | |
| 250 | /* |
| 251 | * Either we were not in an RCU read-side critical section to |
| 252 | * begin with, or we have now recorded that critical section |
| 253 | * globally. Either way, we can now note a quiescent state |
| 254 | * for this CPU. Again, if we were in an RCU read-side critical |
| 255 | * section, and if that critical section was blocking the current |
| 256 | * grace period, then the fact that the task has been enqueued |
| 257 | * means that we continue to block the current grace period. |
| 258 | */ |
| 259 | local_irq_save(flags); |
| 260 | rcu_preempt_qs(cpu); |
| 261 | local_irq_restore(flags); |
| 262 | } |
| 263 | |
| 264 | /* |
| 265 | * Check for preempted RCU readers blocking the current grace period |
| 266 | * for the specified rcu_node structure. If the caller needs a reliable |
| 267 | * answer, it must hold the rcu_node's ->lock. |
| 268 | */ |
| 269 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
| 270 | { |
| 271 | return rnp->gp_tasks != NULL; |
| 272 | } |
| 273 | |
| 274 | /* |
| 275 | * Record a quiescent state for all tasks that were previously queued |
| 276 | * on the specified rcu_node structure and that were blocking the current |
| 277 | * RCU grace period. The caller must hold the specified rnp->lock with |
| 278 | * irqs disabled, and this lock is released upon return, but irqs remain |
| 279 | * disabled. |
| 280 | */ |
| 281 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
| 282 | __releases(rnp->lock) |
| 283 | { |
| 284 | unsigned long mask; |
| 285 | struct rcu_node *rnp_p; |
| 286 | |
| 287 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
| 288 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 289 | return; /* Still need more quiescent states! */ |
| 290 | } |
| 291 | |
| 292 | rnp_p = rnp->parent; |
| 293 | if (rnp_p == NULL) { |
| 294 | /* |
| 295 | * Either there is only one rcu_node in the tree, |
| 296 | * or tasks were kicked up to root rcu_node due to |
| 297 | * CPUs going offline. |
| 298 | */ |
| 299 | rcu_report_qs_rsp(&rcu_preempt_state, flags); |
| 300 | return; |
| 301 | } |
| 302 | |
| 303 | /* Report up the rest of the hierarchy. */ |
| 304 | mask = rnp->grpmask; |
| 305 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 306 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ |
| 307 | smp_mb__after_unlock_lock(); |
| 308 | rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags); |
| 309 | } |
| 310 | |
| 311 | /* |
| 312 | * Advance a ->blkd_tasks-list pointer to the next entry, instead |
| 313 | * returning NULL if at the end of the list. |
| 314 | */ |
| 315 | static struct list_head *rcu_next_node_entry(struct task_struct *t, |
| 316 | struct rcu_node *rnp) |
| 317 | { |
| 318 | struct list_head *np; |
| 319 | |
| 320 | np = t->rcu_node_entry.next; |
| 321 | if (np == &rnp->blkd_tasks) |
| 322 | np = NULL; |
| 323 | return np; |
| 324 | } |
| 325 | |
| 326 | /* |
| 327 | * Handle special cases during rcu_read_unlock(), such as needing to |
| 328 | * notify RCU core processing or task having blocked during the RCU |
| 329 | * read-side critical section. |
| 330 | */ |
| 331 | void rcu_read_unlock_special(struct task_struct *t) |
| 332 | { |
| 333 | int empty; |
| 334 | int empty_exp; |
| 335 | int empty_exp_now; |
| 336 | unsigned long flags; |
| 337 | struct list_head *np; |
| 338 | #ifdef CONFIG_RCU_BOOST |
| 339 | struct rt_mutex *rbmp = NULL; |
| 340 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 341 | struct rcu_node *rnp; |
| 342 | int special; |
| 343 | |
| 344 | /* NMI handlers cannot block and cannot safely manipulate state. */ |
| 345 | if (in_nmi()) |
| 346 | return; |
| 347 | |
| 348 | local_irq_save(flags); |
| 349 | |
| 350 | /* |
| 351 | * If RCU core is waiting for this CPU to exit critical section, |
| 352 | * let it know that we have done so. |
| 353 | */ |
| 354 | special = t->rcu_read_unlock_special; |
| 355 | if (special & RCU_READ_UNLOCK_NEED_QS) { |
| 356 | rcu_preempt_qs(smp_processor_id()); |
| 357 | if (!t->rcu_read_unlock_special) { |
| 358 | local_irq_restore(flags); |
| 359 | return; |
| 360 | } |
| 361 | } |
| 362 | |
| 363 | /* Hardware IRQ handlers cannot block, complain if they get here. */ |
| 364 | if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) { |
| 365 | local_irq_restore(flags); |
| 366 | return; |
| 367 | } |
| 368 | |
| 369 | /* Clean up if blocked during RCU read-side critical section. */ |
| 370 | if (special & RCU_READ_UNLOCK_BLOCKED) { |
| 371 | t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED; |
| 372 | |
| 373 | /* |
| 374 | * Remove this task from the list it blocked on. The |
| 375 | * task can migrate while we acquire the lock, but at |
| 376 | * most one time. So at most two passes through loop. |
| 377 | */ |
| 378 | for (;;) { |
| 379 | rnp = t->rcu_blocked_node; |
| 380 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
| 381 | smp_mb__after_unlock_lock(); |
| 382 | if (rnp == t->rcu_blocked_node) |
| 383 | break; |
| 384 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 385 | } |
| 386 | empty = !rcu_preempt_blocked_readers_cgp(rnp); |
| 387 | empty_exp = !rcu_preempted_readers_exp(rnp); |
| 388 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ |
| 389 | np = rcu_next_node_entry(t, rnp); |
| 390 | list_del_init(&t->rcu_node_entry); |
| 391 | t->rcu_blocked_node = NULL; |
| 392 | trace_rcu_unlock_preempted_task(TPS("rcu_preempt"), |
| 393 | rnp->gpnum, t->pid); |
| 394 | if (&t->rcu_node_entry == rnp->gp_tasks) |
| 395 | rnp->gp_tasks = np; |
| 396 | if (&t->rcu_node_entry == rnp->exp_tasks) |
| 397 | rnp->exp_tasks = np; |
| 398 | #ifdef CONFIG_RCU_BOOST |
| 399 | if (&t->rcu_node_entry == rnp->boost_tasks) |
| 400 | rnp->boost_tasks = np; |
| 401 | /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */ |
| 402 | if (t->rcu_boost_mutex) { |
| 403 | rbmp = t->rcu_boost_mutex; |
| 404 | t->rcu_boost_mutex = NULL; |
| 405 | } |
| 406 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 407 | |
| 408 | /* |
| 409 | * If this was the last task on the current list, and if |
| 410 | * we aren't waiting on any CPUs, report the quiescent state. |
| 411 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock, |
| 412 | * so we must take a snapshot of the expedited state. |
| 413 | */ |
| 414 | empty_exp_now = !rcu_preempted_readers_exp(rnp); |
| 415 | if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) { |
| 416 | trace_rcu_quiescent_state_report(TPS("preempt_rcu"), |
| 417 | rnp->gpnum, |
| 418 | 0, rnp->qsmask, |
| 419 | rnp->level, |
| 420 | rnp->grplo, |
| 421 | rnp->grphi, |
| 422 | !!rnp->gp_tasks); |
| 423 | rcu_report_unblock_qs_rnp(rnp, flags); |
| 424 | } else { |
| 425 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 426 | } |
| 427 | |
| 428 | #ifdef CONFIG_RCU_BOOST |
| 429 | /* Unboost if we were boosted. */ |
| 430 | if (rbmp) |
| 431 | rt_mutex_unlock(rbmp); |
| 432 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 433 | |
| 434 | /* |
| 435 | * If this was the last task on the expedited lists, |
| 436 | * then we need to report up the rcu_node hierarchy. |
| 437 | */ |
| 438 | if (!empty_exp && empty_exp_now) |
| 439 | rcu_report_exp_rnp(&rcu_preempt_state, rnp, true); |
| 440 | } else { |
| 441 | local_irq_restore(flags); |
| 442 | } |
| 443 | } |
| 444 | |
| 445 | #ifdef CONFIG_RCU_CPU_STALL_VERBOSE |
| 446 | |
| 447 | /* |
| 448 | * Dump detailed information for all tasks blocking the current RCU |
| 449 | * grace period on the specified rcu_node structure. |
| 450 | */ |
| 451 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) |
| 452 | { |
| 453 | unsigned long flags; |
| 454 | struct task_struct *t; |
| 455 | |
| 456 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 457 | if (!rcu_preempt_blocked_readers_cgp(rnp)) { |
| 458 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 459 | return; |
| 460 | } |
| 461 | t = list_entry(rnp->gp_tasks, |
| 462 | struct task_struct, rcu_node_entry); |
| 463 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) |
| 464 | sched_show_task(t); |
| 465 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 466 | } |
| 467 | |
| 468 | /* |
| 469 | * Dump detailed information for all tasks blocking the current RCU |
| 470 | * grace period. |
| 471 | */ |
| 472 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) |
| 473 | { |
| 474 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 475 | |
| 476 | rcu_print_detail_task_stall_rnp(rnp); |
| 477 | rcu_for_each_leaf_node(rsp, rnp) |
| 478 | rcu_print_detail_task_stall_rnp(rnp); |
| 479 | } |
| 480 | |
| 481 | #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ |
| 482 | |
| 483 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) |
| 484 | { |
| 485 | } |
| 486 | |
| 487 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ |
| 488 | |
| 489 | #ifdef CONFIG_RCU_CPU_STALL_INFO |
| 490 | |
| 491 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) |
| 492 | { |
| 493 | pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):", |
| 494 | rnp->level, rnp->grplo, rnp->grphi); |
| 495 | } |
| 496 | |
| 497 | static void rcu_print_task_stall_end(void) |
| 498 | { |
| 499 | pr_cont("\n"); |
| 500 | } |
| 501 | |
| 502 | #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */ |
| 503 | |
| 504 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) |
| 505 | { |
| 506 | } |
| 507 | |
| 508 | static void rcu_print_task_stall_end(void) |
| 509 | { |
| 510 | } |
| 511 | |
| 512 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */ |
| 513 | |
| 514 | /* |
| 515 | * Scan the current list of tasks blocked within RCU read-side critical |
| 516 | * sections, printing out the tid of each. |
| 517 | */ |
| 518 | static int rcu_print_task_stall(struct rcu_node *rnp) |
| 519 | { |
| 520 | struct task_struct *t; |
| 521 | int ndetected = 0; |
| 522 | |
| 523 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
| 524 | return 0; |
| 525 | rcu_print_task_stall_begin(rnp); |
| 526 | t = list_entry(rnp->gp_tasks, |
| 527 | struct task_struct, rcu_node_entry); |
| 528 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { |
| 529 | pr_cont(" P%d", t->pid); |
| 530 | ndetected++; |
| 531 | } |
| 532 | rcu_print_task_stall_end(); |
| 533 | return ndetected; |
| 534 | } |
| 535 | |
| 536 | /* |
| 537 | * Check that the list of blocked tasks for the newly completed grace |
| 538 | * period is in fact empty. It is a serious bug to complete a grace |
| 539 | * period that still has RCU readers blocked! This function must be |
| 540 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock |
| 541 | * must be held by the caller. |
| 542 | * |
| 543 | * Also, if there are blocked tasks on the list, they automatically |
| 544 | * block the newly created grace period, so set up ->gp_tasks accordingly. |
| 545 | */ |
| 546 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) |
| 547 | { |
| 548 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
| 549 | if (!list_empty(&rnp->blkd_tasks)) |
| 550 | rnp->gp_tasks = rnp->blkd_tasks.next; |
| 551 | WARN_ON_ONCE(rnp->qsmask); |
| 552 | } |
| 553 | |
| 554 | #ifdef CONFIG_HOTPLUG_CPU |
| 555 | |
| 556 | /* |
| 557 | * Handle tasklist migration for case in which all CPUs covered by the |
| 558 | * specified rcu_node have gone offline. Move them up to the root |
| 559 | * rcu_node. The reason for not just moving them to the immediate |
| 560 | * parent is to remove the need for rcu_read_unlock_special() to |
| 561 | * make more than two attempts to acquire the target rcu_node's lock. |
| 562 | * Returns true if there were tasks blocking the current RCU grace |
| 563 | * period. |
| 564 | * |
| 565 | * Returns 1 if there was previously a task blocking the current grace |
| 566 | * period on the specified rcu_node structure. |
| 567 | * |
| 568 | * The caller must hold rnp->lock with irqs disabled. |
| 569 | */ |
| 570 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
| 571 | struct rcu_node *rnp, |
| 572 | struct rcu_data *rdp) |
| 573 | { |
| 574 | struct list_head *lp; |
| 575 | struct list_head *lp_root; |
| 576 | int retval = 0; |
| 577 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
| 578 | struct task_struct *t; |
| 579 | |
| 580 | if (rnp == rnp_root) { |
| 581 | WARN_ONCE(1, "Last CPU thought to be offlined?"); |
| 582 | return 0; /* Shouldn't happen: at least one CPU online. */ |
| 583 | } |
| 584 | |
| 585 | /* If we are on an internal node, complain bitterly. */ |
| 586 | WARN_ON_ONCE(rnp != rdp->mynode); |
| 587 | |
| 588 | /* |
| 589 | * Move tasks up to root rcu_node. Don't try to get fancy for |
| 590 | * this corner-case operation -- just put this node's tasks |
| 591 | * at the head of the root node's list, and update the root node's |
| 592 | * ->gp_tasks and ->exp_tasks pointers to those of this node's, |
| 593 | * if non-NULL. This might result in waiting for more tasks than |
| 594 | * absolutely necessary, but this is a good performance/complexity |
| 595 | * tradeoff. |
| 596 | */ |
| 597 | if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0) |
| 598 | retval |= RCU_OFL_TASKS_NORM_GP; |
| 599 | if (rcu_preempted_readers_exp(rnp)) |
| 600 | retval |= RCU_OFL_TASKS_EXP_GP; |
| 601 | lp = &rnp->blkd_tasks; |
| 602 | lp_root = &rnp_root->blkd_tasks; |
| 603 | while (!list_empty(lp)) { |
| 604 | t = list_entry(lp->next, typeof(*t), rcu_node_entry); |
| 605 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ |
| 606 | smp_mb__after_unlock_lock(); |
| 607 | list_del(&t->rcu_node_entry); |
| 608 | t->rcu_blocked_node = rnp_root; |
| 609 | list_add(&t->rcu_node_entry, lp_root); |
| 610 | if (&t->rcu_node_entry == rnp->gp_tasks) |
| 611 | rnp_root->gp_tasks = rnp->gp_tasks; |
| 612 | if (&t->rcu_node_entry == rnp->exp_tasks) |
| 613 | rnp_root->exp_tasks = rnp->exp_tasks; |
| 614 | #ifdef CONFIG_RCU_BOOST |
| 615 | if (&t->rcu_node_entry == rnp->boost_tasks) |
| 616 | rnp_root->boost_tasks = rnp->boost_tasks; |
| 617 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 618 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ |
| 619 | } |
| 620 | |
| 621 | rnp->gp_tasks = NULL; |
| 622 | rnp->exp_tasks = NULL; |
| 623 | #ifdef CONFIG_RCU_BOOST |
| 624 | rnp->boost_tasks = NULL; |
| 625 | /* |
| 626 | * In case root is being boosted and leaf was not. Make sure |
| 627 | * that we boost the tasks blocking the current grace period |
| 628 | * in this case. |
| 629 | */ |
| 630 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ |
| 631 | smp_mb__after_unlock_lock(); |
| 632 | if (rnp_root->boost_tasks != NULL && |
| 633 | rnp_root->boost_tasks != rnp_root->gp_tasks && |
| 634 | rnp_root->boost_tasks != rnp_root->exp_tasks) |
| 635 | rnp_root->boost_tasks = rnp_root->gp_tasks; |
| 636 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ |
| 637 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 638 | |
| 639 | return retval; |
| 640 | } |
| 641 | |
| 642 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 643 | |
| 644 | /* |
| 645 | * Check for a quiescent state from the current CPU. When a task blocks, |
| 646 | * the task is recorded in the corresponding CPU's rcu_node structure, |
| 647 | * which is checked elsewhere. |
| 648 | * |
| 649 | * Caller must disable hard irqs. |
| 650 | */ |
| 651 | static void rcu_preempt_check_callbacks(int cpu) |
| 652 | { |
| 653 | struct task_struct *t = current; |
| 654 | |
| 655 | if (t->rcu_read_lock_nesting == 0) { |
| 656 | rcu_preempt_qs(cpu); |
| 657 | return; |
| 658 | } |
| 659 | if (t->rcu_read_lock_nesting > 0 && |
| 660 | per_cpu(rcu_preempt_data, cpu).qs_pending) |
| 661 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS; |
| 662 | } |
| 663 | |
| 664 | #ifdef CONFIG_RCU_BOOST |
| 665 | |
| 666 | static void rcu_preempt_do_callbacks(void) |
| 667 | { |
| 668 | rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data)); |
| 669 | } |
| 670 | |
| 671 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 672 | |
| 673 | /* |
| 674 | * Queue a preemptible-RCU callback for invocation after a grace period. |
| 675 | */ |
| 676 | void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
| 677 | { |
| 678 | __call_rcu(head, func, &rcu_preempt_state, -1, 0); |
| 679 | } |
| 680 | EXPORT_SYMBOL_GPL(call_rcu); |
| 681 | |
| 682 | /** |
| 683 | * synchronize_rcu - wait until a grace period has elapsed. |
| 684 | * |
| 685 | * Control will return to the caller some time after a full grace |
| 686 | * period has elapsed, in other words after all currently executing RCU |
| 687 | * read-side critical sections have completed. Note, however, that |
| 688 | * upon return from synchronize_rcu(), the caller might well be executing |
| 689 | * concurrently with new RCU read-side critical sections that began while |
| 690 | * synchronize_rcu() was waiting. RCU read-side critical sections are |
| 691 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. |
| 692 | * |
| 693 | * See the description of synchronize_sched() for more detailed information |
| 694 | * on memory ordering guarantees. |
| 695 | */ |
| 696 | void synchronize_rcu(void) |
| 697 | { |
| 698 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
| 699 | !lock_is_held(&rcu_lock_map) && |
| 700 | !lock_is_held(&rcu_sched_lock_map), |
| 701 | "Illegal synchronize_rcu() in RCU read-side critical section"); |
| 702 | if (!rcu_scheduler_active) |
| 703 | return; |
| 704 | if (rcu_expedited) |
| 705 | synchronize_rcu_expedited(); |
| 706 | else |
| 707 | wait_rcu_gp(call_rcu); |
| 708 | } |
| 709 | EXPORT_SYMBOL_GPL(synchronize_rcu); |
| 710 | |
| 711 | static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq); |
| 712 | static unsigned long sync_rcu_preempt_exp_count; |
| 713 | static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex); |
| 714 | |
| 715 | /* |
| 716 | * Return non-zero if there are any tasks in RCU read-side critical |
| 717 | * sections blocking the current preemptible-RCU expedited grace period. |
| 718 | * If there is no preemptible-RCU expedited grace period currently in |
| 719 | * progress, returns zero unconditionally. |
| 720 | */ |
| 721 | static int rcu_preempted_readers_exp(struct rcu_node *rnp) |
| 722 | { |
| 723 | return rnp->exp_tasks != NULL; |
| 724 | } |
| 725 | |
| 726 | /* |
| 727 | * return non-zero if there is no RCU expedited grace period in progress |
| 728 | * for the specified rcu_node structure, in other words, if all CPUs and |
| 729 | * tasks covered by the specified rcu_node structure have done their bit |
| 730 | * for the current expedited grace period. Works only for preemptible |
| 731 | * RCU -- other RCU implementation use other means. |
| 732 | * |
| 733 | * Caller must hold sync_rcu_preempt_exp_mutex. |
| 734 | */ |
| 735 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) |
| 736 | { |
| 737 | return !rcu_preempted_readers_exp(rnp) && |
| 738 | ACCESS_ONCE(rnp->expmask) == 0; |
| 739 | } |
| 740 | |
| 741 | /* |
| 742 | * Report the exit from RCU read-side critical section for the last task |
| 743 | * that queued itself during or before the current expedited preemptible-RCU |
| 744 | * grace period. This event is reported either to the rcu_node structure on |
| 745 | * which the task was queued or to one of that rcu_node structure's ancestors, |
| 746 | * recursively up the tree. (Calm down, calm down, we do the recursion |
| 747 | * iteratively!) |
| 748 | * |
| 749 | * Most callers will set the "wake" flag, but the task initiating the |
| 750 | * expedited grace period need not wake itself. |
| 751 | * |
| 752 | * Caller must hold sync_rcu_preempt_exp_mutex. |
| 753 | */ |
| 754 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
| 755 | bool wake) |
| 756 | { |
| 757 | unsigned long flags; |
| 758 | unsigned long mask; |
| 759 | |
| 760 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 761 | smp_mb__after_unlock_lock(); |
| 762 | for (;;) { |
| 763 | if (!sync_rcu_preempt_exp_done(rnp)) { |
| 764 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 765 | break; |
| 766 | } |
| 767 | if (rnp->parent == NULL) { |
| 768 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 769 | if (wake) { |
| 770 | smp_mb(); /* EGP done before wake_up(). */ |
| 771 | wake_up(&sync_rcu_preempt_exp_wq); |
| 772 | } |
| 773 | break; |
| 774 | } |
| 775 | mask = rnp->grpmask; |
| 776 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
| 777 | rnp = rnp->parent; |
| 778 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
| 779 | smp_mb__after_unlock_lock(); |
| 780 | rnp->expmask &= ~mask; |
| 781 | } |
| 782 | } |
| 783 | |
| 784 | /* |
| 785 | * Snapshot the tasks blocking the newly started preemptible-RCU expedited |
| 786 | * grace period for the specified rcu_node structure. If there are no such |
| 787 | * tasks, report it up the rcu_node hierarchy. |
| 788 | * |
| 789 | * Caller must hold sync_rcu_preempt_exp_mutex and must exclude |
| 790 | * CPU hotplug operations. |
| 791 | */ |
| 792 | static void |
| 793 | sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp) |
| 794 | { |
| 795 | unsigned long flags; |
| 796 | int must_wait = 0; |
| 797 | |
| 798 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 799 | smp_mb__after_unlock_lock(); |
| 800 | if (list_empty(&rnp->blkd_tasks)) { |
| 801 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 802 | } else { |
| 803 | rnp->exp_tasks = rnp->blkd_tasks.next; |
| 804 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
| 805 | must_wait = 1; |
| 806 | } |
| 807 | if (!must_wait) |
| 808 | rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */ |
| 809 | } |
| 810 | |
| 811 | /** |
| 812 | * synchronize_rcu_expedited - Brute-force RCU grace period |
| 813 | * |
| 814 | * Wait for an RCU-preempt grace period, but expedite it. The basic |
| 815 | * idea is to invoke synchronize_sched_expedited() to push all the tasks to |
| 816 | * the ->blkd_tasks lists and wait for this list to drain. This consumes |
| 817 | * significant time on all CPUs and is unfriendly to real-time workloads, |
| 818 | * so is thus not recommended for any sort of common-case code. |
| 819 | * In fact, if you are using synchronize_rcu_expedited() in a loop, |
| 820 | * please restructure your code to batch your updates, and then Use a |
| 821 | * single synchronize_rcu() instead. |
| 822 | * |
| 823 | * Note that it is illegal to call this function while holding any lock |
| 824 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal |
| 825 | * to call this function from a CPU-hotplug notifier. Failing to observe |
| 826 | * these restriction will result in deadlock. |
| 827 | */ |
| 828 | void synchronize_rcu_expedited(void) |
| 829 | { |
| 830 | unsigned long flags; |
| 831 | struct rcu_node *rnp; |
| 832 | struct rcu_state *rsp = &rcu_preempt_state; |
| 833 | unsigned long snap; |
| 834 | int trycount = 0; |
| 835 | |
| 836 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ |
| 837 | snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1; |
| 838 | smp_mb(); /* Above access cannot bleed into critical section. */ |
| 839 | |
| 840 | /* |
| 841 | * Block CPU-hotplug operations. This means that any CPU-hotplug |
| 842 | * operation that finds an rcu_node structure with tasks in the |
| 843 | * process of being boosted will know that all tasks blocking |
| 844 | * this expedited grace period will already be in the process of |
| 845 | * being boosted. This simplifies the process of moving tasks |
| 846 | * from leaf to root rcu_node structures. |
| 847 | */ |
| 848 | get_online_cpus(); |
| 849 | |
| 850 | /* |
| 851 | * Acquire lock, falling back to synchronize_rcu() if too many |
| 852 | * lock-acquisition failures. Of course, if someone does the |
| 853 | * expedited grace period for us, just leave. |
| 854 | */ |
| 855 | while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) { |
| 856 | if (ULONG_CMP_LT(snap, |
| 857 | ACCESS_ONCE(sync_rcu_preempt_exp_count))) { |
| 858 | put_online_cpus(); |
| 859 | goto mb_ret; /* Others did our work for us. */ |
| 860 | } |
| 861 | if (trycount++ < 10) { |
| 862 | udelay(trycount * num_online_cpus()); |
| 863 | } else { |
| 864 | put_online_cpus(); |
| 865 | wait_rcu_gp(call_rcu); |
| 866 | return; |
| 867 | } |
| 868 | } |
| 869 | if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) { |
| 870 | put_online_cpus(); |
| 871 | goto unlock_mb_ret; /* Others did our work for us. */ |
| 872 | } |
| 873 | |
| 874 | /* force all RCU readers onto ->blkd_tasks lists. */ |
| 875 | synchronize_sched_expedited(); |
| 876 | |
| 877 | /* Initialize ->expmask for all non-leaf rcu_node structures. */ |
| 878 | rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) { |
| 879 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 880 | smp_mb__after_unlock_lock(); |
| 881 | rnp->expmask = rnp->qsmaskinit; |
| 882 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 883 | } |
| 884 | |
| 885 | /* Snapshot current state of ->blkd_tasks lists. */ |
| 886 | rcu_for_each_leaf_node(rsp, rnp) |
| 887 | sync_rcu_preempt_exp_init(rsp, rnp); |
| 888 | if (NUM_RCU_NODES > 1) |
| 889 | sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp)); |
| 890 | |
| 891 | put_online_cpus(); |
| 892 | |
| 893 | /* Wait for snapshotted ->blkd_tasks lists to drain. */ |
| 894 | rnp = rcu_get_root(rsp); |
| 895 | wait_event(sync_rcu_preempt_exp_wq, |
| 896 | sync_rcu_preempt_exp_done(rnp)); |
| 897 | |
| 898 | /* Clean up and exit. */ |
| 899 | smp_mb(); /* ensure expedited GP seen before counter increment. */ |
| 900 | ACCESS_ONCE(sync_rcu_preempt_exp_count)++; |
| 901 | unlock_mb_ret: |
| 902 | mutex_unlock(&sync_rcu_preempt_exp_mutex); |
| 903 | mb_ret: |
| 904 | smp_mb(); /* ensure subsequent action seen after grace period. */ |
| 905 | } |
| 906 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); |
| 907 | |
| 908 | /** |
| 909 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. |
| 910 | * |
| 911 | * Note that this primitive does not necessarily wait for an RCU grace period |
| 912 | * to complete. For example, if there are no RCU callbacks queued anywhere |
| 913 | * in the system, then rcu_barrier() is within its rights to return |
| 914 | * immediately, without waiting for anything, much less an RCU grace period. |
| 915 | */ |
| 916 | void rcu_barrier(void) |
| 917 | { |
| 918 | _rcu_barrier(&rcu_preempt_state); |
| 919 | } |
| 920 | EXPORT_SYMBOL_GPL(rcu_barrier); |
| 921 | |
| 922 | /* |
| 923 | * Initialize preemptible RCU's state structures. |
| 924 | */ |
| 925 | static void __init __rcu_init_preempt(void) |
| 926 | { |
| 927 | rcu_init_one(&rcu_preempt_state, &rcu_preempt_data); |
| 928 | } |
| 929 | |
| 930 | /* |
| 931 | * Check for a task exiting while in a preemptible-RCU read-side |
| 932 | * critical section, clean up if so. No need to issue warnings, |
| 933 | * as debug_check_no_locks_held() already does this if lockdep |
| 934 | * is enabled. |
| 935 | */ |
| 936 | void exit_rcu(void) |
| 937 | { |
| 938 | struct task_struct *t = current; |
| 939 | |
| 940 | if (likely(list_empty(¤t->rcu_node_entry))) |
| 941 | return; |
| 942 | t->rcu_read_lock_nesting = 1; |
| 943 | barrier(); |
| 944 | t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED; |
| 945 | __rcu_read_unlock(); |
| 946 | } |
| 947 | |
| 948 | #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */ |
| 949 | |
| 950 | static struct rcu_state *rcu_state_p = &rcu_sched_state; |
| 951 | |
| 952 | /* |
| 953 | * Tell them what RCU they are running. |
| 954 | */ |
| 955 | static void __init rcu_bootup_announce(void) |
| 956 | { |
| 957 | pr_info("Hierarchical RCU implementation.\n"); |
| 958 | rcu_bootup_announce_oddness(); |
| 959 | } |
| 960 | |
| 961 | /* |
| 962 | * Return the number of RCU batches processed thus far for debug & stats. |
| 963 | */ |
| 964 | long rcu_batches_completed(void) |
| 965 | { |
| 966 | return rcu_batches_completed_sched(); |
| 967 | } |
| 968 | EXPORT_SYMBOL_GPL(rcu_batches_completed); |
| 969 | |
| 970 | /* |
| 971 | * Because preemptible RCU does not exist, we never have to check for |
| 972 | * CPUs being in quiescent states. |
| 973 | */ |
| 974 | static void rcu_preempt_note_context_switch(int cpu) |
| 975 | { |
| 976 | } |
| 977 | |
| 978 | /* |
| 979 | * Because preemptible RCU does not exist, there are never any preempted |
| 980 | * RCU readers. |
| 981 | */ |
| 982 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
| 983 | { |
| 984 | return 0; |
| 985 | } |
| 986 | |
| 987 | #ifdef CONFIG_HOTPLUG_CPU |
| 988 | |
| 989 | /* Because preemptible RCU does not exist, no quieting of tasks. */ |
| 990 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
| 991 | { |
| 992 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 993 | } |
| 994 | |
| 995 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 996 | |
| 997 | /* |
| 998 | * Because preemptible RCU does not exist, we never have to check for |
| 999 | * tasks blocked within RCU read-side critical sections. |
| 1000 | */ |
| 1001 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) |
| 1002 | { |
| 1003 | } |
| 1004 | |
| 1005 | /* |
| 1006 | * Because preemptible RCU does not exist, we never have to check for |
| 1007 | * tasks blocked within RCU read-side critical sections. |
| 1008 | */ |
| 1009 | static int rcu_print_task_stall(struct rcu_node *rnp) |
| 1010 | { |
| 1011 | return 0; |
| 1012 | } |
| 1013 | |
| 1014 | /* |
| 1015 | * Because there is no preemptible RCU, there can be no readers blocked, |
| 1016 | * so there is no need to check for blocked tasks. So check only for |
| 1017 | * bogus qsmask values. |
| 1018 | */ |
| 1019 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) |
| 1020 | { |
| 1021 | WARN_ON_ONCE(rnp->qsmask); |
| 1022 | } |
| 1023 | |
| 1024 | #ifdef CONFIG_HOTPLUG_CPU |
| 1025 | |
| 1026 | /* |
| 1027 | * Because preemptible RCU does not exist, it never needs to migrate |
| 1028 | * tasks that were blocked within RCU read-side critical sections, and |
| 1029 | * such non-existent tasks cannot possibly have been blocking the current |
| 1030 | * grace period. |
| 1031 | */ |
| 1032 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
| 1033 | struct rcu_node *rnp, |
| 1034 | struct rcu_data *rdp) |
| 1035 | { |
| 1036 | return 0; |
| 1037 | } |
| 1038 | |
| 1039 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 1040 | |
| 1041 | /* |
| 1042 | * Because preemptible RCU does not exist, it never has any callbacks |
| 1043 | * to check. |
| 1044 | */ |
| 1045 | static void rcu_preempt_check_callbacks(int cpu) |
| 1046 | { |
| 1047 | } |
| 1048 | |
| 1049 | /* |
| 1050 | * Wait for an rcu-preempt grace period, but make it happen quickly. |
| 1051 | * But because preemptible RCU does not exist, map to rcu-sched. |
| 1052 | */ |
| 1053 | void synchronize_rcu_expedited(void) |
| 1054 | { |
| 1055 | synchronize_sched_expedited(); |
| 1056 | } |
| 1057 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); |
| 1058 | |
| 1059 | #ifdef CONFIG_HOTPLUG_CPU |
| 1060 | |
| 1061 | /* |
| 1062 | * Because preemptible RCU does not exist, there is never any need to |
| 1063 | * report on tasks preempted in RCU read-side critical sections during |
| 1064 | * expedited RCU grace periods. |
| 1065 | */ |
| 1066 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
| 1067 | bool wake) |
| 1068 | { |
| 1069 | } |
| 1070 | |
| 1071 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 1072 | |
| 1073 | /* |
| 1074 | * Because preemptible RCU does not exist, rcu_barrier() is just |
| 1075 | * another name for rcu_barrier_sched(). |
| 1076 | */ |
| 1077 | void rcu_barrier(void) |
| 1078 | { |
| 1079 | rcu_barrier_sched(); |
| 1080 | } |
| 1081 | EXPORT_SYMBOL_GPL(rcu_barrier); |
| 1082 | |
| 1083 | /* |
| 1084 | * Because preemptible RCU does not exist, it need not be initialized. |
| 1085 | */ |
| 1086 | static void __init __rcu_init_preempt(void) |
| 1087 | { |
| 1088 | } |
| 1089 | |
| 1090 | /* |
| 1091 | * Because preemptible RCU does not exist, tasks cannot possibly exit |
| 1092 | * while in preemptible RCU read-side critical sections. |
| 1093 | */ |
| 1094 | void exit_rcu(void) |
| 1095 | { |
| 1096 | } |
| 1097 | |
| 1098 | #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */ |
| 1099 | |
| 1100 | #ifdef CONFIG_RCU_BOOST |
| 1101 | |
| 1102 | #include "../locking/rtmutex_common.h" |
| 1103 | |
| 1104 | #ifdef CONFIG_RCU_TRACE |
| 1105 | |
| 1106 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) |
| 1107 | { |
| 1108 | if (list_empty(&rnp->blkd_tasks)) |
| 1109 | rnp->n_balk_blkd_tasks++; |
| 1110 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) |
| 1111 | rnp->n_balk_exp_gp_tasks++; |
| 1112 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) |
| 1113 | rnp->n_balk_boost_tasks++; |
| 1114 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) |
| 1115 | rnp->n_balk_notblocked++; |
| 1116 | else if (rnp->gp_tasks != NULL && |
| 1117 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
| 1118 | rnp->n_balk_notyet++; |
| 1119 | else |
| 1120 | rnp->n_balk_nos++; |
| 1121 | } |
| 1122 | |
| 1123 | #else /* #ifdef CONFIG_RCU_TRACE */ |
| 1124 | |
| 1125 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) |
| 1126 | { |
| 1127 | } |
| 1128 | |
| 1129 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ |
| 1130 | |
| 1131 | static void rcu_wake_cond(struct task_struct *t, int status) |
| 1132 | { |
| 1133 | /* |
| 1134 | * If the thread is yielding, only wake it when this |
| 1135 | * is invoked from idle |
| 1136 | */ |
| 1137 | if (status != RCU_KTHREAD_YIELDING || is_idle_task(current)) |
| 1138 | wake_up_process(t); |
| 1139 | } |
| 1140 | |
| 1141 | /* |
| 1142 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks |
| 1143 | * or ->boost_tasks, advancing the pointer to the next task in the |
| 1144 | * ->blkd_tasks list. |
| 1145 | * |
| 1146 | * Note that irqs must be enabled: boosting the task can block. |
| 1147 | * Returns 1 if there are more tasks needing to be boosted. |
| 1148 | */ |
| 1149 | static int rcu_boost(struct rcu_node *rnp) |
| 1150 | { |
| 1151 | unsigned long flags; |
| 1152 | struct rt_mutex mtx; |
| 1153 | struct task_struct *t; |
| 1154 | struct list_head *tb; |
| 1155 | |
| 1156 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) |
| 1157 | return 0; /* Nothing left to boost. */ |
| 1158 | |
| 1159 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 1160 | smp_mb__after_unlock_lock(); |
| 1161 | |
| 1162 | /* |
| 1163 | * Recheck under the lock: all tasks in need of boosting |
| 1164 | * might exit their RCU read-side critical sections on their own. |
| 1165 | */ |
| 1166 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { |
| 1167 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1168 | return 0; |
| 1169 | } |
| 1170 | |
| 1171 | /* |
| 1172 | * Preferentially boost tasks blocking expedited grace periods. |
| 1173 | * This cannot starve the normal grace periods because a second |
| 1174 | * expedited grace period must boost all blocked tasks, including |
| 1175 | * those blocking the pre-existing normal grace period. |
| 1176 | */ |
| 1177 | if (rnp->exp_tasks != NULL) { |
| 1178 | tb = rnp->exp_tasks; |
| 1179 | rnp->n_exp_boosts++; |
| 1180 | } else { |
| 1181 | tb = rnp->boost_tasks; |
| 1182 | rnp->n_normal_boosts++; |
| 1183 | } |
| 1184 | rnp->n_tasks_boosted++; |
| 1185 | |
| 1186 | /* |
| 1187 | * We boost task t by manufacturing an rt_mutex that appears to |
| 1188 | * be held by task t. We leave a pointer to that rt_mutex where |
| 1189 | * task t can find it, and task t will release the mutex when it |
| 1190 | * exits its outermost RCU read-side critical section. Then |
| 1191 | * simply acquiring this artificial rt_mutex will boost task |
| 1192 | * t's priority. (Thanks to tglx for suggesting this approach!) |
| 1193 | * |
| 1194 | * Note that task t must acquire rnp->lock to remove itself from |
| 1195 | * the ->blkd_tasks list, which it will do from exit() if from |
| 1196 | * nowhere else. We therefore are guaranteed that task t will |
| 1197 | * stay around at least until we drop rnp->lock. Note that |
| 1198 | * rnp->lock also resolves races between our priority boosting |
| 1199 | * and task t's exiting its outermost RCU read-side critical |
| 1200 | * section. |
| 1201 | */ |
| 1202 | t = container_of(tb, struct task_struct, rcu_node_entry); |
| 1203 | rt_mutex_init_proxy_locked(&mtx, t); |
| 1204 | t->rcu_boost_mutex = &mtx; |
| 1205 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1206 | rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */ |
| 1207 | rt_mutex_unlock(&mtx); /* Keep lockdep happy. */ |
| 1208 | |
| 1209 | return ACCESS_ONCE(rnp->exp_tasks) != NULL || |
| 1210 | ACCESS_ONCE(rnp->boost_tasks) != NULL; |
| 1211 | } |
| 1212 | |
| 1213 | /* |
| 1214 | * Priority-boosting kthread. One per leaf rcu_node and one for the |
| 1215 | * root rcu_node. |
| 1216 | */ |
| 1217 | static int rcu_boost_kthread(void *arg) |
| 1218 | { |
| 1219 | struct rcu_node *rnp = (struct rcu_node *)arg; |
| 1220 | int spincnt = 0; |
| 1221 | int more2boost; |
| 1222 | |
| 1223 | trace_rcu_utilization(TPS("Start boost kthread@init")); |
| 1224 | for (;;) { |
| 1225 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
| 1226 | trace_rcu_utilization(TPS("End boost kthread@rcu_wait")); |
| 1227 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
| 1228 | trace_rcu_utilization(TPS("Start boost kthread@rcu_wait")); |
| 1229 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
| 1230 | more2boost = rcu_boost(rnp); |
| 1231 | if (more2boost) |
| 1232 | spincnt++; |
| 1233 | else |
| 1234 | spincnt = 0; |
| 1235 | if (spincnt > 10) { |
| 1236 | rnp->boost_kthread_status = RCU_KTHREAD_YIELDING; |
| 1237 | trace_rcu_utilization(TPS("End boost kthread@rcu_yield")); |
| 1238 | schedule_timeout_interruptible(2); |
| 1239 | trace_rcu_utilization(TPS("Start boost kthread@rcu_yield")); |
| 1240 | spincnt = 0; |
| 1241 | } |
| 1242 | } |
| 1243 | /* NOTREACHED */ |
| 1244 | trace_rcu_utilization(TPS("End boost kthread@notreached")); |
| 1245 | return 0; |
| 1246 | } |
| 1247 | |
| 1248 | /* |
| 1249 | * Check to see if it is time to start boosting RCU readers that are |
| 1250 | * blocking the current grace period, and, if so, tell the per-rcu_node |
| 1251 | * kthread to start boosting them. If there is an expedited grace |
| 1252 | * period in progress, it is always time to boost. |
| 1253 | * |
| 1254 | * The caller must hold rnp->lock, which this function releases. |
| 1255 | * The ->boost_kthread_task is immortal, so we don't need to worry |
| 1256 | * about it going away. |
| 1257 | */ |
| 1258 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
| 1259 | { |
| 1260 | struct task_struct *t; |
| 1261 | |
| 1262 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
| 1263 | rnp->n_balk_exp_gp_tasks++; |
| 1264 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1265 | return; |
| 1266 | } |
| 1267 | if (rnp->exp_tasks != NULL || |
| 1268 | (rnp->gp_tasks != NULL && |
| 1269 | rnp->boost_tasks == NULL && |
| 1270 | rnp->qsmask == 0 && |
| 1271 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { |
| 1272 | if (rnp->exp_tasks == NULL) |
| 1273 | rnp->boost_tasks = rnp->gp_tasks; |
| 1274 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1275 | t = rnp->boost_kthread_task; |
| 1276 | if (t) |
| 1277 | rcu_wake_cond(t, rnp->boost_kthread_status); |
| 1278 | } else { |
| 1279 | rcu_initiate_boost_trace(rnp); |
| 1280 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1281 | } |
| 1282 | } |
| 1283 | |
| 1284 | /* |
| 1285 | * Wake up the per-CPU kthread to invoke RCU callbacks. |
| 1286 | */ |
| 1287 | static void invoke_rcu_callbacks_kthread(void) |
| 1288 | { |
| 1289 | unsigned long flags; |
| 1290 | |
| 1291 | local_irq_save(flags); |
| 1292 | __this_cpu_write(rcu_cpu_has_work, 1); |
| 1293 | if (__this_cpu_read(rcu_cpu_kthread_task) != NULL && |
| 1294 | current != __this_cpu_read(rcu_cpu_kthread_task)) { |
| 1295 | rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task), |
| 1296 | __this_cpu_read(rcu_cpu_kthread_status)); |
| 1297 | } |
| 1298 | local_irq_restore(flags); |
| 1299 | } |
| 1300 | |
| 1301 | /* |
| 1302 | * Is the current CPU running the RCU-callbacks kthread? |
| 1303 | * Caller must have preemption disabled. |
| 1304 | */ |
| 1305 | static bool rcu_is_callbacks_kthread(void) |
| 1306 | { |
| 1307 | return __this_cpu_read(rcu_cpu_kthread_task) == current; |
| 1308 | } |
| 1309 | |
| 1310 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) |
| 1311 | |
| 1312 | /* |
| 1313 | * Do priority-boost accounting for the start of a new grace period. |
| 1314 | */ |
| 1315 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
| 1316 | { |
| 1317 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; |
| 1318 | } |
| 1319 | |
| 1320 | /* |
| 1321 | * Create an RCU-boost kthread for the specified node if one does not |
| 1322 | * already exist. We only create this kthread for preemptible RCU. |
| 1323 | * Returns zero if all is well, a negated errno otherwise. |
| 1324 | */ |
| 1325 | static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp, |
| 1326 | struct rcu_node *rnp) |
| 1327 | { |
| 1328 | int rnp_index = rnp - &rsp->node[0]; |
| 1329 | unsigned long flags; |
| 1330 | struct sched_param sp; |
| 1331 | struct task_struct *t; |
| 1332 | |
| 1333 | if (&rcu_preempt_state != rsp) |
| 1334 | return 0; |
| 1335 | |
| 1336 | if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0) |
| 1337 | return 0; |
| 1338 | |
| 1339 | rsp->boost = 1; |
| 1340 | if (rnp->boost_kthread_task != NULL) |
| 1341 | return 0; |
| 1342 | t = kthread_create(rcu_boost_kthread, (void *)rnp, |
| 1343 | "rcub/%d", rnp_index); |
| 1344 | if (IS_ERR(t)) |
| 1345 | return PTR_ERR(t); |
| 1346 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 1347 | smp_mb__after_unlock_lock(); |
| 1348 | rnp->boost_kthread_task = t; |
| 1349 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1350 | sp.sched_priority = RCU_BOOST_PRIO; |
| 1351 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); |
| 1352 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
| 1353 | return 0; |
| 1354 | } |
| 1355 | |
| 1356 | static void rcu_kthread_do_work(void) |
| 1357 | { |
| 1358 | rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data)); |
| 1359 | rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data)); |
| 1360 | rcu_preempt_do_callbacks(); |
| 1361 | } |
| 1362 | |
| 1363 | static void rcu_cpu_kthread_setup(unsigned int cpu) |
| 1364 | { |
| 1365 | struct sched_param sp; |
| 1366 | |
| 1367 | sp.sched_priority = RCU_KTHREAD_PRIO; |
| 1368 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); |
| 1369 | } |
| 1370 | |
| 1371 | static void rcu_cpu_kthread_park(unsigned int cpu) |
| 1372 | { |
| 1373 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; |
| 1374 | } |
| 1375 | |
| 1376 | static int rcu_cpu_kthread_should_run(unsigned int cpu) |
| 1377 | { |
| 1378 | return __this_cpu_read(rcu_cpu_has_work); |
| 1379 | } |
| 1380 | |
| 1381 | /* |
| 1382 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the |
| 1383 | * RCU softirq used in flavors and configurations of RCU that do not |
| 1384 | * support RCU priority boosting. |
| 1385 | */ |
| 1386 | static void rcu_cpu_kthread(unsigned int cpu) |
| 1387 | { |
| 1388 | unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status); |
| 1389 | char work, *workp = this_cpu_ptr(&rcu_cpu_has_work); |
| 1390 | int spincnt; |
| 1391 | |
| 1392 | for (spincnt = 0; spincnt < 10; spincnt++) { |
| 1393 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait")); |
| 1394 | local_bh_disable(); |
| 1395 | *statusp = RCU_KTHREAD_RUNNING; |
| 1396 | this_cpu_inc(rcu_cpu_kthread_loops); |
| 1397 | local_irq_disable(); |
| 1398 | work = *workp; |
| 1399 | *workp = 0; |
| 1400 | local_irq_enable(); |
| 1401 | if (work) |
| 1402 | rcu_kthread_do_work(); |
| 1403 | local_bh_enable(); |
| 1404 | if (*workp == 0) { |
| 1405 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); |
| 1406 | *statusp = RCU_KTHREAD_WAITING; |
| 1407 | return; |
| 1408 | } |
| 1409 | } |
| 1410 | *statusp = RCU_KTHREAD_YIELDING; |
| 1411 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); |
| 1412 | schedule_timeout_interruptible(2); |
| 1413 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); |
| 1414 | *statusp = RCU_KTHREAD_WAITING; |
| 1415 | } |
| 1416 | |
| 1417 | /* |
| 1418 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are |
| 1419 | * served by the rcu_node in question. The CPU hotplug lock is still |
| 1420 | * held, so the value of rnp->qsmaskinit will be stable. |
| 1421 | * |
| 1422 | * We don't include outgoingcpu in the affinity set, use -1 if there is |
| 1423 | * no outgoing CPU. If there are no CPUs left in the affinity set, |
| 1424 | * this function allows the kthread to execute on any CPU. |
| 1425 | */ |
| 1426 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
| 1427 | { |
| 1428 | struct task_struct *t = rnp->boost_kthread_task; |
| 1429 | unsigned long mask = rnp->qsmaskinit; |
| 1430 | cpumask_var_t cm; |
| 1431 | int cpu; |
| 1432 | |
| 1433 | if (!t) |
| 1434 | return; |
| 1435 | if (!zalloc_cpumask_var(&cm, GFP_KERNEL)) |
| 1436 | return; |
| 1437 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) |
| 1438 | if ((mask & 0x1) && cpu != outgoingcpu) |
| 1439 | cpumask_set_cpu(cpu, cm); |
| 1440 | if (cpumask_weight(cm) == 0) { |
| 1441 | cpumask_setall(cm); |
| 1442 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) |
| 1443 | cpumask_clear_cpu(cpu, cm); |
| 1444 | WARN_ON_ONCE(cpumask_weight(cm) == 0); |
| 1445 | } |
| 1446 | set_cpus_allowed_ptr(t, cm); |
| 1447 | free_cpumask_var(cm); |
| 1448 | } |
| 1449 | |
| 1450 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { |
| 1451 | .store = &rcu_cpu_kthread_task, |
| 1452 | .thread_should_run = rcu_cpu_kthread_should_run, |
| 1453 | .thread_fn = rcu_cpu_kthread, |
| 1454 | .thread_comm = "rcuc/%u", |
| 1455 | .setup = rcu_cpu_kthread_setup, |
| 1456 | .park = rcu_cpu_kthread_park, |
| 1457 | }; |
| 1458 | |
| 1459 | /* |
| 1460 | * Spawn all kthreads -- called as soon as the scheduler is running. |
| 1461 | */ |
| 1462 | static int __init rcu_spawn_kthreads(void) |
| 1463 | { |
| 1464 | struct rcu_node *rnp; |
| 1465 | int cpu; |
| 1466 | |
| 1467 | rcu_scheduler_fully_active = 1; |
| 1468 | for_each_possible_cpu(cpu) |
| 1469 | per_cpu(rcu_cpu_has_work, cpu) = 0; |
| 1470 | BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec)); |
| 1471 | rnp = rcu_get_root(rcu_state_p); |
| 1472 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); |
| 1473 | if (NUM_RCU_NODES > 1) { |
| 1474 | rcu_for_each_leaf_node(rcu_state_p, rnp) |
| 1475 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); |
| 1476 | } |
| 1477 | return 0; |
| 1478 | } |
| 1479 | early_initcall(rcu_spawn_kthreads); |
| 1480 | |
| 1481 | static void rcu_prepare_kthreads(int cpu) |
| 1482 | { |
| 1483 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
| 1484 | struct rcu_node *rnp = rdp->mynode; |
| 1485 | |
| 1486 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ |
| 1487 | if (rcu_scheduler_fully_active) |
| 1488 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); |
| 1489 | } |
| 1490 | |
| 1491 | #else /* #ifdef CONFIG_RCU_BOOST */ |
| 1492 | |
| 1493 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
| 1494 | { |
| 1495 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1496 | } |
| 1497 | |
| 1498 | static void invoke_rcu_callbacks_kthread(void) |
| 1499 | { |
| 1500 | WARN_ON_ONCE(1); |
| 1501 | } |
| 1502 | |
| 1503 | static bool rcu_is_callbacks_kthread(void) |
| 1504 | { |
| 1505 | return false; |
| 1506 | } |
| 1507 | |
| 1508 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
| 1509 | { |
| 1510 | } |
| 1511 | |
| 1512 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
| 1513 | { |
| 1514 | } |
| 1515 | |
| 1516 | static int __init rcu_scheduler_really_started(void) |
| 1517 | { |
| 1518 | rcu_scheduler_fully_active = 1; |
| 1519 | return 0; |
| 1520 | } |
| 1521 | early_initcall(rcu_scheduler_really_started); |
| 1522 | |
| 1523 | static void rcu_prepare_kthreads(int cpu) |
| 1524 | { |
| 1525 | } |
| 1526 | |
| 1527 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
| 1528 | |
| 1529 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
| 1530 | |
| 1531 | /* |
| 1532 | * Check to see if any future RCU-related work will need to be done |
| 1533 | * by the current CPU, even if none need be done immediately, returning |
| 1534 | * 1 if so. This function is part of the RCU implementation; it is -not- |
| 1535 | * an exported member of the RCU API. |
| 1536 | * |
| 1537 | * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs |
| 1538 | * any flavor of RCU. |
| 1539 | */ |
| 1540 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
| 1541 | int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies) |
| 1542 | { |
| 1543 | *delta_jiffies = ULONG_MAX; |
| 1544 | return rcu_cpu_has_callbacks(cpu, NULL); |
| 1545 | } |
| 1546 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
| 1547 | |
| 1548 | /* |
| 1549 | * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up |
| 1550 | * after it. |
| 1551 | */ |
| 1552 | static void rcu_cleanup_after_idle(int cpu) |
| 1553 | { |
| 1554 | } |
| 1555 | |
| 1556 | /* |
| 1557 | * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n, |
| 1558 | * is nothing. |
| 1559 | */ |
| 1560 | static void rcu_prepare_for_idle(int cpu) |
| 1561 | { |
| 1562 | } |
| 1563 | |
| 1564 | /* |
| 1565 | * Don't bother keeping a running count of the number of RCU callbacks |
| 1566 | * posted because CONFIG_RCU_FAST_NO_HZ=n. |
| 1567 | */ |
| 1568 | static void rcu_idle_count_callbacks_posted(void) |
| 1569 | { |
| 1570 | } |
| 1571 | |
| 1572 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
| 1573 | |
| 1574 | /* |
| 1575 | * This code is invoked when a CPU goes idle, at which point we want |
| 1576 | * to have the CPU do everything required for RCU so that it can enter |
| 1577 | * the energy-efficient dyntick-idle mode. This is handled by a |
| 1578 | * state machine implemented by rcu_prepare_for_idle() below. |
| 1579 | * |
| 1580 | * The following three proprocessor symbols control this state machine: |
| 1581 | * |
| 1582 | * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted |
| 1583 | * to sleep in dyntick-idle mode with RCU callbacks pending. This |
| 1584 | * is sized to be roughly one RCU grace period. Those energy-efficiency |
| 1585 | * benchmarkers who might otherwise be tempted to set this to a large |
| 1586 | * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your |
| 1587 | * system. And if you are -that- concerned about energy efficiency, |
| 1588 | * just power the system down and be done with it! |
| 1589 | * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is |
| 1590 | * permitted to sleep in dyntick-idle mode with only lazy RCU |
| 1591 | * callbacks pending. Setting this too high can OOM your system. |
| 1592 | * |
| 1593 | * The values below work well in practice. If future workloads require |
| 1594 | * adjustment, they can be converted into kernel config parameters, though |
| 1595 | * making the state machine smarter might be a better option. |
| 1596 | */ |
| 1597 | #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */ |
| 1598 | #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */ |
| 1599 | |
| 1600 | static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY; |
| 1601 | module_param(rcu_idle_gp_delay, int, 0644); |
| 1602 | static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY; |
| 1603 | module_param(rcu_idle_lazy_gp_delay, int, 0644); |
| 1604 | |
| 1605 | extern int tick_nohz_active; |
| 1606 | |
| 1607 | /* |
| 1608 | * Try to advance callbacks for all flavors of RCU on the current CPU, but |
| 1609 | * only if it has been awhile since the last time we did so. Afterwards, |
| 1610 | * if there are any callbacks ready for immediate invocation, return true. |
| 1611 | */ |
| 1612 | static bool __maybe_unused rcu_try_advance_all_cbs(void) |
| 1613 | { |
| 1614 | bool cbs_ready = false; |
| 1615 | struct rcu_data *rdp; |
| 1616 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
| 1617 | struct rcu_node *rnp; |
| 1618 | struct rcu_state *rsp; |
| 1619 | |
| 1620 | /* Exit early if we advanced recently. */ |
| 1621 | if (jiffies == rdtp->last_advance_all) |
| 1622 | return 0; |
| 1623 | rdtp->last_advance_all = jiffies; |
| 1624 | |
| 1625 | for_each_rcu_flavor(rsp) { |
| 1626 | rdp = this_cpu_ptr(rsp->rda); |
| 1627 | rnp = rdp->mynode; |
| 1628 | |
| 1629 | /* |
| 1630 | * Don't bother checking unless a grace period has |
| 1631 | * completed since we last checked and there are |
| 1632 | * callbacks not yet ready to invoke. |
| 1633 | */ |
| 1634 | if (rdp->completed != rnp->completed && |
| 1635 | rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL]) |
| 1636 | note_gp_changes(rsp, rdp); |
| 1637 | |
| 1638 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
| 1639 | cbs_ready = true; |
| 1640 | } |
| 1641 | return cbs_ready; |
| 1642 | } |
| 1643 | |
| 1644 | /* |
| 1645 | * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready |
| 1646 | * to invoke. If the CPU has callbacks, try to advance them. Tell the |
| 1647 | * caller to set the timeout based on whether or not there are non-lazy |
| 1648 | * callbacks. |
| 1649 | * |
| 1650 | * The caller must have disabled interrupts. |
| 1651 | */ |
| 1652 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
| 1653 | int rcu_needs_cpu(int cpu, unsigned long *dj) |
| 1654 | { |
| 1655 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
| 1656 | |
| 1657 | /* Snapshot to detect later posting of non-lazy callback. */ |
| 1658 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; |
| 1659 | |
| 1660 | /* If no callbacks, RCU doesn't need the CPU. */ |
| 1661 | if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) { |
| 1662 | *dj = ULONG_MAX; |
| 1663 | return 0; |
| 1664 | } |
| 1665 | |
| 1666 | /* Attempt to advance callbacks. */ |
| 1667 | if (rcu_try_advance_all_cbs()) { |
| 1668 | /* Some ready to invoke, so initiate later invocation. */ |
| 1669 | invoke_rcu_core(); |
| 1670 | return 1; |
| 1671 | } |
| 1672 | rdtp->last_accelerate = jiffies; |
| 1673 | |
| 1674 | /* Request timer delay depending on laziness, and round. */ |
| 1675 | if (!rdtp->all_lazy) { |
| 1676 | *dj = round_up(rcu_idle_gp_delay + jiffies, |
| 1677 | rcu_idle_gp_delay) - jiffies; |
| 1678 | } else { |
| 1679 | *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies; |
| 1680 | } |
| 1681 | return 0; |
| 1682 | } |
| 1683 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
| 1684 | |
| 1685 | /* |
| 1686 | * Prepare a CPU for idle from an RCU perspective. The first major task |
| 1687 | * is to sense whether nohz mode has been enabled or disabled via sysfs. |
| 1688 | * The second major task is to check to see if a non-lazy callback has |
| 1689 | * arrived at a CPU that previously had only lazy callbacks. The third |
| 1690 | * major task is to accelerate (that is, assign grace-period numbers to) |
| 1691 | * any recently arrived callbacks. |
| 1692 | * |
| 1693 | * The caller must have disabled interrupts. |
| 1694 | */ |
| 1695 | static void rcu_prepare_for_idle(int cpu) |
| 1696 | { |
| 1697 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
| 1698 | bool needwake; |
| 1699 | struct rcu_data *rdp; |
| 1700 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
| 1701 | struct rcu_node *rnp; |
| 1702 | struct rcu_state *rsp; |
| 1703 | int tne; |
| 1704 | |
| 1705 | /* Handle nohz enablement switches conservatively. */ |
| 1706 | tne = ACCESS_ONCE(tick_nohz_active); |
| 1707 | if (tne != rdtp->tick_nohz_enabled_snap) { |
| 1708 | if (rcu_cpu_has_callbacks(cpu, NULL)) |
| 1709 | invoke_rcu_core(); /* force nohz to see update. */ |
| 1710 | rdtp->tick_nohz_enabled_snap = tne; |
| 1711 | return; |
| 1712 | } |
| 1713 | if (!tne) |
| 1714 | return; |
| 1715 | |
| 1716 | /* If this is a no-CBs CPU, no callbacks, just return. */ |
| 1717 | if (rcu_is_nocb_cpu(cpu)) |
| 1718 | return; |
| 1719 | |
| 1720 | /* |
| 1721 | * If a non-lazy callback arrived at a CPU having only lazy |
| 1722 | * callbacks, invoke RCU core for the side-effect of recalculating |
| 1723 | * idle duration on re-entry to idle. |
| 1724 | */ |
| 1725 | if (rdtp->all_lazy && |
| 1726 | rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) { |
| 1727 | rdtp->all_lazy = false; |
| 1728 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; |
| 1729 | invoke_rcu_core(); |
| 1730 | return; |
| 1731 | } |
| 1732 | |
| 1733 | /* |
| 1734 | * If we have not yet accelerated this jiffy, accelerate all |
| 1735 | * callbacks on this CPU. |
| 1736 | */ |
| 1737 | if (rdtp->last_accelerate == jiffies) |
| 1738 | return; |
| 1739 | rdtp->last_accelerate = jiffies; |
| 1740 | for_each_rcu_flavor(rsp) { |
| 1741 | rdp = per_cpu_ptr(rsp->rda, cpu); |
| 1742 | if (!*rdp->nxttail[RCU_DONE_TAIL]) |
| 1743 | continue; |
| 1744 | rnp = rdp->mynode; |
| 1745 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
| 1746 | smp_mb__after_unlock_lock(); |
| 1747 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
| 1748 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 1749 | if (needwake) |
| 1750 | rcu_gp_kthread_wake(rsp); |
| 1751 | } |
| 1752 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
| 1753 | } |
| 1754 | |
| 1755 | /* |
| 1756 | * Clean up for exit from idle. Attempt to advance callbacks based on |
| 1757 | * any grace periods that elapsed while the CPU was idle, and if any |
| 1758 | * callbacks are now ready to invoke, initiate invocation. |
| 1759 | */ |
| 1760 | static void rcu_cleanup_after_idle(int cpu) |
| 1761 | { |
| 1762 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
| 1763 | if (rcu_is_nocb_cpu(cpu)) |
| 1764 | return; |
| 1765 | if (rcu_try_advance_all_cbs()) |
| 1766 | invoke_rcu_core(); |
| 1767 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
| 1768 | } |
| 1769 | |
| 1770 | /* |
| 1771 | * Keep a running count of the number of non-lazy callbacks posted |
| 1772 | * on this CPU. This running counter (which is never decremented) allows |
| 1773 | * rcu_prepare_for_idle() to detect when something out of the idle loop |
| 1774 | * posts a callback, even if an equal number of callbacks are invoked. |
| 1775 | * Of course, callbacks should only be posted from within a trace event |
| 1776 | * designed to be called from idle or from within RCU_NONIDLE(). |
| 1777 | */ |
| 1778 | static void rcu_idle_count_callbacks_posted(void) |
| 1779 | { |
| 1780 | __this_cpu_add(rcu_dynticks.nonlazy_posted, 1); |
| 1781 | } |
| 1782 | |
| 1783 | /* |
| 1784 | * Data for flushing lazy RCU callbacks at OOM time. |
| 1785 | */ |
| 1786 | static atomic_t oom_callback_count; |
| 1787 | static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq); |
| 1788 | |
| 1789 | /* |
| 1790 | * RCU OOM callback -- decrement the outstanding count and deliver the |
| 1791 | * wake-up if we are the last one. |
| 1792 | */ |
| 1793 | static void rcu_oom_callback(struct rcu_head *rhp) |
| 1794 | { |
| 1795 | if (atomic_dec_and_test(&oom_callback_count)) |
| 1796 | wake_up(&oom_callback_wq); |
| 1797 | } |
| 1798 | |
| 1799 | /* |
| 1800 | * Post an rcu_oom_notify callback on the current CPU if it has at |
| 1801 | * least one lazy callback. This will unnecessarily post callbacks |
| 1802 | * to CPUs that already have a non-lazy callback at the end of their |
| 1803 | * callback list, but this is an infrequent operation, so accept some |
| 1804 | * extra overhead to keep things simple. |
| 1805 | */ |
| 1806 | static void rcu_oom_notify_cpu(void *unused) |
| 1807 | { |
| 1808 | struct rcu_state *rsp; |
| 1809 | struct rcu_data *rdp; |
| 1810 | |
| 1811 | for_each_rcu_flavor(rsp) { |
| 1812 | rdp = raw_cpu_ptr(rsp->rda); |
| 1813 | if (rdp->qlen_lazy != 0) { |
| 1814 | atomic_inc(&oom_callback_count); |
| 1815 | rsp->call(&rdp->oom_head, rcu_oom_callback); |
| 1816 | } |
| 1817 | } |
| 1818 | } |
| 1819 | |
| 1820 | /* |
| 1821 | * If low on memory, ensure that each CPU has a non-lazy callback. |
| 1822 | * This will wake up CPUs that have only lazy callbacks, in turn |
| 1823 | * ensuring that they free up the corresponding memory in a timely manner. |
| 1824 | * Because an uncertain amount of memory will be freed in some uncertain |
| 1825 | * timeframe, we do not claim to have freed anything. |
| 1826 | */ |
| 1827 | static int rcu_oom_notify(struct notifier_block *self, |
| 1828 | unsigned long notused, void *nfreed) |
| 1829 | { |
| 1830 | int cpu; |
| 1831 | |
| 1832 | /* Wait for callbacks from earlier instance to complete. */ |
| 1833 | wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0); |
| 1834 | smp_mb(); /* Ensure callback reuse happens after callback invocation. */ |
| 1835 | |
| 1836 | /* |
| 1837 | * Prevent premature wakeup: ensure that all increments happen |
| 1838 | * before there is a chance of the counter reaching zero. |
| 1839 | */ |
| 1840 | atomic_set(&oom_callback_count, 1); |
| 1841 | |
| 1842 | get_online_cpus(); |
| 1843 | for_each_online_cpu(cpu) { |
| 1844 | smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1); |
| 1845 | cond_resched(); |
| 1846 | } |
| 1847 | put_online_cpus(); |
| 1848 | |
| 1849 | /* Unconditionally decrement: no need to wake ourselves up. */ |
| 1850 | atomic_dec(&oom_callback_count); |
| 1851 | |
| 1852 | return NOTIFY_OK; |
| 1853 | } |
| 1854 | |
| 1855 | static struct notifier_block rcu_oom_nb = { |
| 1856 | .notifier_call = rcu_oom_notify |
| 1857 | }; |
| 1858 | |
| 1859 | static int __init rcu_register_oom_notifier(void) |
| 1860 | { |
| 1861 | register_oom_notifier(&rcu_oom_nb); |
| 1862 | return 0; |
| 1863 | } |
| 1864 | early_initcall(rcu_register_oom_notifier); |
| 1865 | |
| 1866 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
| 1867 | |
| 1868 | #ifdef CONFIG_RCU_CPU_STALL_INFO |
| 1869 | |
| 1870 | #ifdef CONFIG_RCU_FAST_NO_HZ |
| 1871 | |
| 1872 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) |
| 1873 | { |
| 1874 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
| 1875 | unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap; |
| 1876 | |
| 1877 | sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c", |
| 1878 | rdtp->last_accelerate & 0xffff, jiffies & 0xffff, |
| 1879 | ulong2long(nlpd), |
| 1880 | rdtp->all_lazy ? 'L' : '.', |
| 1881 | rdtp->tick_nohz_enabled_snap ? '.' : 'D'); |
| 1882 | } |
| 1883 | |
| 1884 | #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */ |
| 1885 | |
| 1886 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) |
| 1887 | { |
| 1888 | *cp = '\0'; |
| 1889 | } |
| 1890 | |
| 1891 | #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */ |
| 1892 | |
| 1893 | /* Initiate the stall-info list. */ |
| 1894 | static void print_cpu_stall_info_begin(void) |
| 1895 | { |
| 1896 | pr_cont("\n"); |
| 1897 | } |
| 1898 | |
| 1899 | /* |
| 1900 | * Print out diagnostic information for the specified stalled CPU. |
| 1901 | * |
| 1902 | * If the specified CPU is aware of the current RCU grace period |
| 1903 | * (flavor specified by rsp), then print the number of scheduling |
| 1904 | * clock interrupts the CPU has taken during the time that it has |
| 1905 | * been aware. Otherwise, print the number of RCU grace periods |
| 1906 | * that this CPU is ignorant of, for example, "1" if the CPU was |
| 1907 | * aware of the previous grace period. |
| 1908 | * |
| 1909 | * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info. |
| 1910 | */ |
| 1911 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) |
| 1912 | { |
| 1913 | char fast_no_hz[72]; |
| 1914 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
| 1915 | struct rcu_dynticks *rdtp = rdp->dynticks; |
| 1916 | char *ticks_title; |
| 1917 | unsigned long ticks_value; |
| 1918 | |
| 1919 | if (rsp->gpnum == rdp->gpnum) { |
| 1920 | ticks_title = "ticks this GP"; |
| 1921 | ticks_value = rdp->ticks_this_gp; |
| 1922 | } else { |
| 1923 | ticks_title = "GPs behind"; |
| 1924 | ticks_value = rsp->gpnum - rdp->gpnum; |
| 1925 | } |
| 1926 | print_cpu_stall_fast_no_hz(fast_no_hz, cpu); |
| 1927 | pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n", |
| 1928 | cpu, ticks_value, ticks_title, |
| 1929 | atomic_read(&rdtp->dynticks) & 0xfff, |
| 1930 | rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting, |
| 1931 | rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu), |
| 1932 | fast_no_hz); |
| 1933 | } |
| 1934 | |
| 1935 | /* Terminate the stall-info list. */ |
| 1936 | static void print_cpu_stall_info_end(void) |
| 1937 | { |
| 1938 | pr_err("\t"); |
| 1939 | } |
| 1940 | |
| 1941 | /* Zero ->ticks_this_gp for all flavors of RCU. */ |
| 1942 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) |
| 1943 | { |
| 1944 | rdp->ticks_this_gp = 0; |
| 1945 | rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id()); |
| 1946 | } |
| 1947 | |
| 1948 | /* Increment ->ticks_this_gp for all flavors of RCU. */ |
| 1949 | static void increment_cpu_stall_ticks(void) |
| 1950 | { |
| 1951 | struct rcu_state *rsp; |
| 1952 | |
| 1953 | for_each_rcu_flavor(rsp) |
| 1954 | raw_cpu_inc(rsp->rda->ticks_this_gp); |
| 1955 | } |
| 1956 | |
| 1957 | #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */ |
| 1958 | |
| 1959 | static void print_cpu_stall_info_begin(void) |
| 1960 | { |
| 1961 | pr_cont(" {"); |
| 1962 | } |
| 1963 | |
| 1964 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) |
| 1965 | { |
| 1966 | pr_cont(" %d", cpu); |
| 1967 | } |
| 1968 | |
| 1969 | static void print_cpu_stall_info_end(void) |
| 1970 | { |
| 1971 | pr_cont("} "); |
| 1972 | } |
| 1973 | |
| 1974 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) |
| 1975 | { |
| 1976 | } |
| 1977 | |
| 1978 | static void increment_cpu_stall_ticks(void) |
| 1979 | { |
| 1980 | } |
| 1981 | |
| 1982 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */ |
| 1983 | |
| 1984 | #ifdef CONFIG_RCU_NOCB_CPU |
| 1985 | |
| 1986 | /* |
| 1987 | * Offload callback processing from the boot-time-specified set of CPUs |
| 1988 | * specified by rcu_nocb_mask. For each CPU in the set, there is a |
| 1989 | * kthread created that pulls the callbacks from the corresponding CPU, |
| 1990 | * waits for a grace period to elapse, and invokes the callbacks. |
| 1991 | * The no-CBs CPUs do a wake_up() on their kthread when they insert |
| 1992 | * a callback into any empty list, unless the rcu_nocb_poll boot parameter |
| 1993 | * has been specified, in which case each kthread actively polls its |
| 1994 | * CPU. (Which isn't so great for energy efficiency, but which does |
| 1995 | * reduce RCU's overhead on that CPU.) |
| 1996 | * |
| 1997 | * This is intended to be used in conjunction with Frederic Weisbecker's |
| 1998 | * adaptive-idle work, which would seriously reduce OS jitter on CPUs |
| 1999 | * running CPU-bound user-mode computations. |
| 2000 | * |
| 2001 | * Offloading of callback processing could also in theory be used as |
| 2002 | * an energy-efficiency measure because CPUs with no RCU callbacks |
| 2003 | * queued are more aggressive about entering dyntick-idle mode. |
| 2004 | */ |
| 2005 | |
| 2006 | |
| 2007 | /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */ |
| 2008 | static int __init rcu_nocb_setup(char *str) |
| 2009 | { |
| 2010 | alloc_bootmem_cpumask_var(&rcu_nocb_mask); |
| 2011 | have_rcu_nocb_mask = true; |
| 2012 | cpulist_parse(str, rcu_nocb_mask); |
| 2013 | return 1; |
| 2014 | } |
| 2015 | __setup("rcu_nocbs=", rcu_nocb_setup); |
| 2016 | |
| 2017 | static int __init parse_rcu_nocb_poll(char *arg) |
| 2018 | { |
| 2019 | rcu_nocb_poll = 1; |
| 2020 | return 0; |
| 2021 | } |
| 2022 | early_param("rcu_nocb_poll", parse_rcu_nocb_poll); |
| 2023 | |
| 2024 | /* |
| 2025 | * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended |
| 2026 | * grace period. |
| 2027 | */ |
| 2028 | static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
| 2029 | { |
| 2030 | wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]); |
| 2031 | } |
| 2032 | |
| 2033 | /* |
| 2034 | * Set the root rcu_node structure's ->need_future_gp field |
| 2035 | * based on the sum of those of all rcu_node structures. This does |
| 2036 | * double-count the root rcu_node structure's requests, but this |
| 2037 | * is necessary to handle the possibility of a rcu_nocb_kthread() |
| 2038 | * having awakened during the time that the rcu_node structures |
| 2039 | * were being updated for the end of the previous grace period. |
| 2040 | */ |
| 2041 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
| 2042 | { |
| 2043 | rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq; |
| 2044 | } |
| 2045 | |
| 2046 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
| 2047 | { |
| 2048 | init_waitqueue_head(&rnp->nocb_gp_wq[0]); |
| 2049 | init_waitqueue_head(&rnp->nocb_gp_wq[1]); |
| 2050 | } |
| 2051 | |
| 2052 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
| 2053 | /* Is the specified CPU a no-CBs CPU? */ |
| 2054 | bool rcu_is_nocb_cpu(int cpu) |
| 2055 | { |
| 2056 | if (have_rcu_nocb_mask) |
| 2057 | return cpumask_test_cpu(cpu, rcu_nocb_mask); |
| 2058 | return false; |
| 2059 | } |
| 2060 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
| 2061 | |
| 2062 | /* |
| 2063 | * Kick the leader kthread for this NOCB group. |
| 2064 | */ |
| 2065 | static void wake_nocb_leader(struct rcu_data *rdp, bool force) |
| 2066 | { |
| 2067 | struct rcu_data *rdp_leader = rdp->nocb_leader; |
| 2068 | |
| 2069 | if (!ACCESS_ONCE(rdp_leader->nocb_kthread)) |
| 2070 | return; |
| 2071 | if (!ACCESS_ONCE(rdp_leader->nocb_leader_wake) || force) { |
| 2072 | /* Prior xchg orders against prior callback enqueue. */ |
| 2073 | ACCESS_ONCE(rdp_leader->nocb_leader_wake) = true; |
| 2074 | wake_up(&rdp_leader->nocb_wq); |
| 2075 | } |
| 2076 | } |
| 2077 | |
| 2078 | /* |
| 2079 | * Enqueue the specified string of rcu_head structures onto the specified |
| 2080 | * CPU's no-CBs lists. The CPU is specified by rdp, the head of the |
| 2081 | * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy |
| 2082 | * counts are supplied by rhcount and rhcount_lazy. |
| 2083 | * |
| 2084 | * If warranted, also wake up the kthread servicing this CPUs queues. |
| 2085 | */ |
| 2086 | static void __call_rcu_nocb_enqueue(struct rcu_data *rdp, |
| 2087 | struct rcu_head *rhp, |
| 2088 | struct rcu_head **rhtp, |
| 2089 | int rhcount, int rhcount_lazy, |
| 2090 | unsigned long flags) |
| 2091 | { |
| 2092 | int len; |
| 2093 | struct rcu_head **old_rhpp; |
| 2094 | struct task_struct *t; |
| 2095 | |
| 2096 | /* Enqueue the callback on the nocb list and update counts. */ |
| 2097 | old_rhpp = xchg(&rdp->nocb_tail, rhtp); |
| 2098 | ACCESS_ONCE(*old_rhpp) = rhp; |
| 2099 | atomic_long_add(rhcount, &rdp->nocb_q_count); |
| 2100 | atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy); |
| 2101 | |
| 2102 | /* If we are not being polled and there is a kthread, awaken it ... */ |
| 2103 | t = ACCESS_ONCE(rdp->nocb_kthread); |
| 2104 | if (rcu_nocb_poll || !t) { |
| 2105 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
| 2106 | TPS("WakeNotPoll")); |
| 2107 | return; |
| 2108 | } |
| 2109 | len = atomic_long_read(&rdp->nocb_q_count); |
| 2110 | if (old_rhpp == &rdp->nocb_head) { |
| 2111 | if (!irqs_disabled_flags(flags)) { |
| 2112 | /* ... if queue was empty ... */ |
| 2113 | wake_nocb_leader(rdp, false); |
| 2114 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
| 2115 | TPS("WakeEmpty")); |
| 2116 | } else { |
| 2117 | rdp->nocb_defer_wakeup = true; |
| 2118 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
| 2119 | TPS("WakeEmptyIsDeferred")); |
| 2120 | } |
| 2121 | rdp->qlen_last_fqs_check = 0; |
| 2122 | } else if (len > rdp->qlen_last_fqs_check + qhimark) { |
| 2123 | /* ... or if many callbacks queued. */ |
| 2124 | wake_nocb_leader(rdp, true); |
| 2125 | rdp->qlen_last_fqs_check = LONG_MAX / 2; |
| 2126 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf")); |
| 2127 | } else { |
| 2128 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot")); |
| 2129 | } |
| 2130 | return; |
| 2131 | } |
| 2132 | |
| 2133 | /* |
| 2134 | * This is a helper for __call_rcu(), which invokes this when the normal |
| 2135 | * callback queue is inoperable. If this is not a no-CBs CPU, this |
| 2136 | * function returns failure back to __call_rcu(), which can complain |
| 2137 | * appropriately. |
| 2138 | * |
| 2139 | * Otherwise, this function queues the callback where the corresponding |
| 2140 | * "rcuo" kthread can find it. |
| 2141 | */ |
| 2142 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
| 2143 | bool lazy, unsigned long flags) |
| 2144 | { |
| 2145 | |
| 2146 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
| 2147 | return 0; |
| 2148 | __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags); |
| 2149 | if (__is_kfree_rcu_offset((unsigned long)rhp->func)) |
| 2150 | trace_rcu_kfree_callback(rdp->rsp->name, rhp, |
| 2151 | (unsigned long)rhp->func, |
| 2152 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
| 2153 | -atomic_long_read(&rdp->nocb_q_count)); |
| 2154 | else |
| 2155 | trace_rcu_callback(rdp->rsp->name, rhp, |
| 2156 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
| 2157 | -atomic_long_read(&rdp->nocb_q_count)); |
| 2158 | return 1; |
| 2159 | } |
| 2160 | |
| 2161 | /* |
| 2162 | * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is |
| 2163 | * not a no-CBs CPU. |
| 2164 | */ |
| 2165 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, |
| 2166 | struct rcu_data *rdp, |
| 2167 | unsigned long flags) |
| 2168 | { |
| 2169 | long ql = rsp->qlen; |
| 2170 | long qll = rsp->qlen_lazy; |
| 2171 | |
| 2172 | /* If this is not a no-CBs CPU, tell the caller to do it the old way. */ |
| 2173 | if (!rcu_is_nocb_cpu(smp_processor_id())) |
| 2174 | return 0; |
| 2175 | rsp->qlen = 0; |
| 2176 | rsp->qlen_lazy = 0; |
| 2177 | |
| 2178 | /* First, enqueue the donelist, if any. This preserves CB ordering. */ |
| 2179 | if (rsp->orphan_donelist != NULL) { |
| 2180 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist, |
| 2181 | rsp->orphan_donetail, ql, qll, flags); |
| 2182 | ql = qll = 0; |
| 2183 | rsp->orphan_donelist = NULL; |
| 2184 | rsp->orphan_donetail = &rsp->orphan_donelist; |
| 2185 | } |
| 2186 | if (rsp->orphan_nxtlist != NULL) { |
| 2187 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist, |
| 2188 | rsp->orphan_nxttail, ql, qll, flags); |
| 2189 | ql = qll = 0; |
| 2190 | rsp->orphan_nxtlist = NULL; |
| 2191 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; |
| 2192 | } |
| 2193 | return 1; |
| 2194 | } |
| 2195 | |
| 2196 | /* |
| 2197 | * If necessary, kick off a new grace period, and either way wait |
| 2198 | * for a subsequent grace period to complete. |
| 2199 | */ |
| 2200 | static void rcu_nocb_wait_gp(struct rcu_data *rdp) |
| 2201 | { |
| 2202 | unsigned long c; |
| 2203 | bool d; |
| 2204 | unsigned long flags; |
| 2205 | bool needwake; |
| 2206 | struct rcu_node *rnp = rdp->mynode; |
| 2207 | |
| 2208 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 2209 | smp_mb__after_unlock_lock(); |
| 2210 | needwake = rcu_start_future_gp(rnp, rdp, &c); |
| 2211 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 2212 | if (needwake) |
| 2213 | rcu_gp_kthread_wake(rdp->rsp); |
| 2214 | |
| 2215 | /* |
| 2216 | * Wait for the grace period. Do so interruptibly to avoid messing |
| 2217 | * up the load average. |
| 2218 | */ |
| 2219 | trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait")); |
| 2220 | for (;;) { |
| 2221 | wait_event_interruptible( |
| 2222 | rnp->nocb_gp_wq[c & 0x1], |
| 2223 | (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c))); |
| 2224 | if (likely(d)) |
| 2225 | break; |
| 2226 | flush_signals(current); |
| 2227 | trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait")); |
| 2228 | } |
| 2229 | trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait")); |
| 2230 | smp_mb(); /* Ensure that CB invocation happens after GP end. */ |
| 2231 | } |
| 2232 | |
| 2233 | /* |
| 2234 | * Leaders come here to wait for additional callbacks to show up. |
| 2235 | * This function does not return until callbacks appear. |
| 2236 | */ |
| 2237 | static void nocb_leader_wait(struct rcu_data *my_rdp) |
| 2238 | { |
| 2239 | bool firsttime = true; |
| 2240 | bool gotcbs; |
| 2241 | struct rcu_data *rdp; |
| 2242 | struct rcu_head **tail; |
| 2243 | |
| 2244 | wait_again: |
| 2245 | |
| 2246 | /* Wait for callbacks to appear. */ |
| 2247 | if (!rcu_nocb_poll) { |
| 2248 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep"); |
| 2249 | wait_event_interruptible(my_rdp->nocb_wq, |
| 2250 | ACCESS_ONCE(my_rdp->nocb_leader_wake)); |
| 2251 | /* Memory barrier handled by smp_mb() calls below and repoll. */ |
| 2252 | } else if (firsttime) { |
| 2253 | firsttime = false; /* Don't drown trace log with "Poll"! */ |
| 2254 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll"); |
| 2255 | } |
| 2256 | |
| 2257 | /* |
| 2258 | * Each pass through the following loop checks a follower for CBs. |
| 2259 | * We are our own first follower. Any CBs found are moved to |
| 2260 | * nocb_gp_head, where they await a grace period. |
| 2261 | */ |
| 2262 | gotcbs = false; |
| 2263 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { |
| 2264 | rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head); |
| 2265 | if (!rdp->nocb_gp_head) |
| 2266 | continue; /* No CBs here, try next follower. */ |
| 2267 | |
| 2268 | /* Move callbacks to wait-for-GP list, which is empty. */ |
| 2269 | ACCESS_ONCE(rdp->nocb_head) = NULL; |
| 2270 | rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head); |
| 2271 | rdp->nocb_gp_count = atomic_long_xchg(&rdp->nocb_q_count, 0); |
| 2272 | rdp->nocb_gp_count_lazy = |
| 2273 | atomic_long_xchg(&rdp->nocb_q_count_lazy, 0); |
| 2274 | gotcbs = true; |
| 2275 | } |
| 2276 | |
| 2277 | /* |
| 2278 | * If there were no callbacks, sleep a bit, rescan after a |
| 2279 | * memory barrier, and go retry. |
| 2280 | */ |
| 2281 | if (unlikely(!gotcbs)) { |
| 2282 | if (!rcu_nocb_poll) |
| 2283 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, |
| 2284 | "WokeEmpty"); |
| 2285 | flush_signals(current); |
| 2286 | schedule_timeout_interruptible(1); |
| 2287 | |
| 2288 | /* Rescan in case we were a victim of memory ordering. */ |
| 2289 | my_rdp->nocb_leader_wake = false; |
| 2290 | smp_mb(); /* Ensure _wake false before scan. */ |
| 2291 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) |
| 2292 | if (ACCESS_ONCE(rdp->nocb_head)) { |
| 2293 | /* Found CB, so short-circuit next wait. */ |
| 2294 | my_rdp->nocb_leader_wake = true; |
| 2295 | break; |
| 2296 | } |
| 2297 | goto wait_again; |
| 2298 | } |
| 2299 | |
| 2300 | /* Wait for one grace period. */ |
| 2301 | rcu_nocb_wait_gp(my_rdp); |
| 2302 | |
| 2303 | /* |
| 2304 | * We left ->nocb_leader_wake set to reduce cache thrashing. |
| 2305 | * We clear it now, but recheck for new callbacks while |
| 2306 | * traversing our follower list. |
| 2307 | */ |
| 2308 | my_rdp->nocb_leader_wake = false; |
| 2309 | smp_mb(); /* Ensure _wake false before scan of ->nocb_head. */ |
| 2310 | |
| 2311 | /* Each pass through the following loop wakes a follower, if needed. */ |
| 2312 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { |
| 2313 | if (ACCESS_ONCE(rdp->nocb_head)) |
| 2314 | my_rdp->nocb_leader_wake = true; /* No need to wait. */ |
| 2315 | if (!rdp->nocb_gp_head) |
| 2316 | continue; /* No CBs, so no need to wake follower. */ |
| 2317 | |
| 2318 | /* Append callbacks to follower's "done" list. */ |
| 2319 | tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail); |
| 2320 | *tail = rdp->nocb_gp_head; |
| 2321 | atomic_long_add(rdp->nocb_gp_count, &rdp->nocb_follower_count); |
| 2322 | atomic_long_add(rdp->nocb_gp_count_lazy, |
| 2323 | &rdp->nocb_follower_count_lazy); |
| 2324 | if (rdp != my_rdp && tail == &rdp->nocb_follower_head) { |
| 2325 | /* |
| 2326 | * List was empty, wake up the follower. |
| 2327 | * Memory barriers supplied by atomic_long_add(). |
| 2328 | */ |
| 2329 | wake_up(&rdp->nocb_wq); |
| 2330 | } |
| 2331 | } |
| 2332 | |
| 2333 | /* If we (the leader) don't have CBs, go wait some more. */ |
| 2334 | if (!my_rdp->nocb_follower_head) |
| 2335 | goto wait_again; |
| 2336 | } |
| 2337 | |
| 2338 | /* |
| 2339 | * Followers come here to wait for additional callbacks to show up. |
| 2340 | * This function does not return until callbacks appear. |
| 2341 | */ |
| 2342 | static void nocb_follower_wait(struct rcu_data *rdp) |
| 2343 | { |
| 2344 | bool firsttime = true; |
| 2345 | |
| 2346 | for (;;) { |
| 2347 | if (!rcu_nocb_poll) { |
| 2348 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
| 2349 | "FollowerSleep"); |
| 2350 | wait_event_interruptible(rdp->nocb_wq, |
| 2351 | ACCESS_ONCE(rdp->nocb_follower_head)); |
| 2352 | } else if (firsttime) { |
| 2353 | /* Don't drown trace log with "Poll"! */ |
| 2354 | firsttime = false; |
| 2355 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll"); |
| 2356 | } |
| 2357 | if (smp_load_acquire(&rdp->nocb_follower_head)) { |
| 2358 | /* ^^^ Ensure CB invocation follows _head test. */ |
| 2359 | return; |
| 2360 | } |
| 2361 | if (!rcu_nocb_poll) |
| 2362 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
| 2363 | "WokeEmpty"); |
| 2364 | flush_signals(current); |
| 2365 | schedule_timeout_interruptible(1); |
| 2366 | } |
| 2367 | } |
| 2368 | |
| 2369 | /* |
| 2370 | * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes |
| 2371 | * callbacks queued by the corresponding no-CBs CPU, however, there is |
| 2372 | * an optional leader-follower relationship so that the grace-period |
| 2373 | * kthreads don't have to do quite so many wakeups. |
| 2374 | */ |
| 2375 | static int rcu_nocb_kthread(void *arg) |
| 2376 | { |
| 2377 | int c, cl; |
| 2378 | struct rcu_head *list; |
| 2379 | struct rcu_head *next; |
| 2380 | struct rcu_head **tail; |
| 2381 | struct rcu_data *rdp = arg; |
| 2382 | |
| 2383 | /* Each pass through this loop invokes one batch of callbacks */ |
| 2384 | for (;;) { |
| 2385 | /* Wait for callbacks. */ |
| 2386 | if (rdp->nocb_leader == rdp) |
| 2387 | nocb_leader_wait(rdp); |
| 2388 | else |
| 2389 | nocb_follower_wait(rdp); |
| 2390 | |
| 2391 | /* Pull the ready-to-invoke callbacks onto local list. */ |
| 2392 | list = ACCESS_ONCE(rdp->nocb_follower_head); |
| 2393 | BUG_ON(!list); |
| 2394 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty"); |
| 2395 | ACCESS_ONCE(rdp->nocb_follower_head) = NULL; |
| 2396 | tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head); |
| 2397 | c = atomic_long_xchg(&rdp->nocb_follower_count, 0); |
| 2398 | cl = atomic_long_xchg(&rdp->nocb_follower_count_lazy, 0); |
| 2399 | rdp->nocb_p_count += c; |
| 2400 | rdp->nocb_p_count_lazy += cl; |
| 2401 | |
| 2402 | /* Each pass through the following loop invokes a callback. */ |
| 2403 | trace_rcu_batch_start(rdp->rsp->name, cl, c, -1); |
| 2404 | c = cl = 0; |
| 2405 | while (list) { |
| 2406 | next = list->next; |
| 2407 | /* Wait for enqueuing to complete, if needed. */ |
| 2408 | while (next == NULL && &list->next != tail) { |
| 2409 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
| 2410 | TPS("WaitQueue")); |
| 2411 | schedule_timeout_interruptible(1); |
| 2412 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
| 2413 | TPS("WokeQueue")); |
| 2414 | next = list->next; |
| 2415 | } |
| 2416 | debug_rcu_head_unqueue(list); |
| 2417 | local_bh_disable(); |
| 2418 | if (__rcu_reclaim(rdp->rsp->name, list)) |
| 2419 | cl++; |
| 2420 | c++; |
| 2421 | local_bh_enable(); |
| 2422 | list = next; |
| 2423 | } |
| 2424 | trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); |
| 2425 | ACCESS_ONCE(rdp->nocb_p_count) -= c; |
| 2426 | ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl; |
| 2427 | rdp->n_nocbs_invoked += c; |
| 2428 | } |
| 2429 | return 0; |
| 2430 | } |
| 2431 | |
| 2432 | /* Is a deferred wakeup of rcu_nocb_kthread() required? */ |
| 2433 | static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
| 2434 | { |
| 2435 | return ACCESS_ONCE(rdp->nocb_defer_wakeup); |
| 2436 | } |
| 2437 | |
| 2438 | /* Do a deferred wakeup of rcu_nocb_kthread(). */ |
| 2439 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) |
| 2440 | { |
| 2441 | if (!rcu_nocb_need_deferred_wakeup(rdp)) |
| 2442 | return; |
| 2443 | ACCESS_ONCE(rdp->nocb_defer_wakeup) = false; |
| 2444 | wake_nocb_leader(rdp, false); |
| 2445 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWakeEmpty")); |
| 2446 | } |
| 2447 | |
| 2448 | /* Initialize per-rcu_data variables for no-CBs CPUs. */ |
| 2449 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
| 2450 | { |
| 2451 | rdp->nocb_tail = &rdp->nocb_head; |
| 2452 | init_waitqueue_head(&rdp->nocb_wq); |
| 2453 | rdp->nocb_follower_tail = &rdp->nocb_follower_head; |
| 2454 | } |
| 2455 | |
| 2456 | /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */ |
| 2457 | static int rcu_nocb_leader_stride = -1; |
| 2458 | module_param(rcu_nocb_leader_stride, int, 0444); |
| 2459 | |
| 2460 | /* |
| 2461 | * Create a kthread for each RCU flavor for each no-CBs CPU. |
| 2462 | * Also initialize leader-follower relationships. |
| 2463 | */ |
| 2464 | static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp) |
| 2465 | { |
| 2466 | int cpu; |
| 2467 | int ls = rcu_nocb_leader_stride; |
| 2468 | int nl = 0; /* Next leader. */ |
| 2469 | struct rcu_data *rdp; |
| 2470 | struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */ |
| 2471 | struct rcu_data *rdp_prev = NULL; |
| 2472 | struct task_struct *t; |
| 2473 | |
| 2474 | if (rcu_nocb_mask == NULL) |
| 2475 | return; |
| 2476 | if (ls == -1) { |
| 2477 | ls = int_sqrt(nr_cpu_ids); |
| 2478 | rcu_nocb_leader_stride = ls; |
| 2479 | } |
| 2480 | |
| 2481 | /* |
| 2482 | * Each pass through this loop sets up one rcu_data structure and |
| 2483 | * spawns one rcu_nocb_kthread(). |
| 2484 | */ |
| 2485 | for_each_cpu(cpu, rcu_nocb_mask) { |
| 2486 | rdp = per_cpu_ptr(rsp->rda, cpu); |
| 2487 | if (rdp->cpu >= nl) { |
| 2488 | /* New leader, set up for followers & next leader. */ |
| 2489 | nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls; |
| 2490 | rdp->nocb_leader = rdp; |
| 2491 | rdp_leader = rdp; |
| 2492 | } else { |
| 2493 | /* Another follower, link to previous leader. */ |
| 2494 | rdp->nocb_leader = rdp_leader; |
| 2495 | rdp_prev->nocb_next_follower = rdp; |
| 2496 | } |
| 2497 | rdp_prev = rdp; |
| 2498 | |
| 2499 | /* Spawn the kthread for this CPU. */ |
| 2500 | t = kthread_run(rcu_nocb_kthread, rdp, |
| 2501 | "rcuo%c/%d", rsp->abbr, cpu); |
| 2502 | BUG_ON(IS_ERR(t)); |
| 2503 | ACCESS_ONCE(rdp->nocb_kthread) = t; |
| 2504 | } |
| 2505 | } |
| 2506 | |
| 2507 | /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */ |
| 2508 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
| 2509 | { |
| 2510 | if (rcu_nocb_mask == NULL || |
| 2511 | !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask)) |
| 2512 | return false; |
| 2513 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
| 2514 | return true; |
| 2515 | } |
| 2516 | |
| 2517 | #else /* #ifdef CONFIG_RCU_NOCB_CPU */ |
| 2518 | |
| 2519 | static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
| 2520 | { |
| 2521 | } |
| 2522 | |
| 2523 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
| 2524 | { |
| 2525 | } |
| 2526 | |
| 2527 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
| 2528 | { |
| 2529 | } |
| 2530 | |
| 2531 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
| 2532 | bool lazy, unsigned long flags) |
| 2533 | { |
| 2534 | return 0; |
| 2535 | } |
| 2536 | |
| 2537 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, |
| 2538 | struct rcu_data *rdp, |
| 2539 | unsigned long flags) |
| 2540 | { |
| 2541 | return 0; |
| 2542 | } |
| 2543 | |
| 2544 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
| 2545 | { |
| 2546 | } |
| 2547 | |
| 2548 | static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
| 2549 | { |
| 2550 | return false; |
| 2551 | } |
| 2552 | |
| 2553 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) |
| 2554 | { |
| 2555 | } |
| 2556 | |
| 2557 | static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp) |
| 2558 | { |
| 2559 | } |
| 2560 | |
| 2561 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
| 2562 | { |
| 2563 | return false; |
| 2564 | } |
| 2565 | |
| 2566 | #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */ |
| 2567 | |
| 2568 | /* |
| 2569 | * An adaptive-ticks CPU can potentially execute in kernel mode for an |
| 2570 | * arbitrarily long period of time with the scheduling-clock tick turned |
| 2571 | * off. RCU will be paying attention to this CPU because it is in the |
| 2572 | * kernel, but the CPU cannot be guaranteed to be executing the RCU state |
| 2573 | * machine because the scheduling-clock tick has been disabled. Therefore, |
| 2574 | * if an adaptive-ticks CPU is failing to respond to the current grace |
| 2575 | * period and has not be idle from an RCU perspective, kick it. |
| 2576 | */ |
| 2577 | static void __maybe_unused rcu_kick_nohz_cpu(int cpu) |
| 2578 | { |
| 2579 | #ifdef CONFIG_NO_HZ_FULL |
| 2580 | if (tick_nohz_full_cpu(cpu)) |
| 2581 | smp_send_reschedule(cpu); |
| 2582 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ |
| 2583 | } |
| 2584 | |
| 2585 | |
| 2586 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
| 2587 | |
| 2588 | /* |
| 2589 | * Define RCU flavor that holds sysidle state. This needs to be the |
| 2590 | * most active flavor of RCU. |
| 2591 | */ |
| 2592 | #ifdef CONFIG_PREEMPT_RCU |
| 2593 | static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state; |
| 2594 | #else /* #ifdef CONFIG_PREEMPT_RCU */ |
| 2595 | static struct rcu_state *rcu_sysidle_state = &rcu_sched_state; |
| 2596 | #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ |
| 2597 | |
| 2598 | static int full_sysidle_state; /* Current system-idle state. */ |
| 2599 | #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */ |
| 2600 | #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */ |
| 2601 | #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */ |
| 2602 | #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */ |
| 2603 | #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */ |
| 2604 | |
| 2605 | /* |
| 2606 | * Invoked to note exit from irq or task transition to idle. Note that |
| 2607 | * usermode execution does -not- count as idle here! After all, we want |
| 2608 | * to detect full-system idle states, not RCU quiescent states and grace |
| 2609 | * periods. The caller must have disabled interrupts. |
| 2610 | */ |
| 2611 | static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq) |
| 2612 | { |
| 2613 | unsigned long j; |
| 2614 | |
| 2615 | /* Adjust nesting, check for fully idle. */ |
| 2616 | if (irq) { |
| 2617 | rdtp->dynticks_idle_nesting--; |
| 2618 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); |
| 2619 | if (rdtp->dynticks_idle_nesting != 0) |
| 2620 | return; /* Still not fully idle. */ |
| 2621 | } else { |
| 2622 | if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) == |
| 2623 | DYNTICK_TASK_NEST_VALUE) { |
| 2624 | rdtp->dynticks_idle_nesting = 0; |
| 2625 | } else { |
| 2626 | rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE; |
| 2627 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); |
| 2628 | return; /* Still not fully idle. */ |
| 2629 | } |
| 2630 | } |
| 2631 | |
| 2632 | /* Record start of fully idle period. */ |
| 2633 | j = jiffies; |
| 2634 | ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j; |
| 2635 | smp_mb__before_atomic(); |
| 2636 | atomic_inc(&rdtp->dynticks_idle); |
| 2637 | smp_mb__after_atomic(); |
| 2638 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1); |
| 2639 | } |
| 2640 | |
| 2641 | /* |
| 2642 | * Unconditionally force exit from full system-idle state. This is |
| 2643 | * invoked when a normal CPU exits idle, but must be called separately |
| 2644 | * for the timekeeping CPU (tick_do_timer_cpu). The reason for this |
| 2645 | * is that the timekeeping CPU is permitted to take scheduling-clock |
| 2646 | * interrupts while the system is in system-idle state, and of course |
| 2647 | * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock |
| 2648 | * interrupt from any other type of interrupt. |
| 2649 | */ |
| 2650 | void rcu_sysidle_force_exit(void) |
| 2651 | { |
| 2652 | int oldstate = ACCESS_ONCE(full_sysidle_state); |
| 2653 | int newoldstate; |
| 2654 | |
| 2655 | /* |
| 2656 | * Each pass through the following loop attempts to exit full |
| 2657 | * system-idle state. If contention proves to be a problem, |
| 2658 | * a trylock-based contention tree could be used here. |
| 2659 | */ |
| 2660 | while (oldstate > RCU_SYSIDLE_SHORT) { |
| 2661 | newoldstate = cmpxchg(&full_sysidle_state, |
| 2662 | oldstate, RCU_SYSIDLE_NOT); |
| 2663 | if (oldstate == newoldstate && |
| 2664 | oldstate == RCU_SYSIDLE_FULL_NOTED) { |
| 2665 | rcu_kick_nohz_cpu(tick_do_timer_cpu); |
| 2666 | return; /* We cleared it, done! */ |
| 2667 | } |
| 2668 | oldstate = newoldstate; |
| 2669 | } |
| 2670 | smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */ |
| 2671 | } |
| 2672 | |
| 2673 | /* |
| 2674 | * Invoked to note entry to irq or task transition from idle. Note that |
| 2675 | * usermode execution does -not- count as idle here! The caller must |
| 2676 | * have disabled interrupts. |
| 2677 | */ |
| 2678 | static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq) |
| 2679 | { |
| 2680 | /* Adjust nesting, check for already non-idle. */ |
| 2681 | if (irq) { |
| 2682 | rdtp->dynticks_idle_nesting++; |
| 2683 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); |
| 2684 | if (rdtp->dynticks_idle_nesting != 1) |
| 2685 | return; /* Already non-idle. */ |
| 2686 | } else { |
| 2687 | /* |
| 2688 | * Allow for irq misnesting. Yes, it really is possible |
| 2689 | * to enter an irq handler then never leave it, and maybe |
| 2690 | * also vice versa. Handle both possibilities. |
| 2691 | */ |
| 2692 | if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) { |
| 2693 | rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE; |
| 2694 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); |
| 2695 | return; /* Already non-idle. */ |
| 2696 | } else { |
| 2697 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE; |
| 2698 | } |
| 2699 | } |
| 2700 | |
| 2701 | /* Record end of idle period. */ |
| 2702 | smp_mb__before_atomic(); |
| 2703 | atomic_inc(&rdtp->dynticks_idle); |
| 2704 | smp_mb__after_atomic(); |
| 2705 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1)); |
| 2706 | |
| 2707 | /* |
| 2708 | * If we are the timekeeping CPU, we are permitted to be non-idle |
| 2709 | * during a system-idle state. This must be the case, because |
| 2710 | * the timekeeping CPU has to take scheduling-clock interrupts |
| 2711 | * during the time that the system is transitioning to full |
| 2712 | * system-idle state. This means that the timekeeping CPU must |
| 2713 | * invoke rcu_sysidle_force_exit() directly if it does anything |
| 2714 | * more than take a scheduling-clock interrupt. |
| 2715 | */ |
| 2716 | if (smp_processor_id() == tick_do_timer_cpu) |
| 2717 | return; |
| 2718 | |
| 2719 | /* Update system-idle state: We are clearly no longer fully idle! */ |
| 2720 | rcu_sysidle_force_exit(); |
| 2721 | } |
| 2722 | |
| 2723 | /* |
| 2724 | * Check to see if the current CPU is idle. Note that usermode execution |
| 2725 | * does not count as idle. The caller must have disabled interrupts. |
| 2726 | */ |
| 2727 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, |
| 2728 | unsigned long *maxj) |
| 2729 | { |
| 2730 | int cur; |
| 2731 | unsigned long j; |
| 2732 | struct rcu_dynticks *rdtp = rdp->dynticks; |
| 2733 | |
| 2734 | /* |
| 2735 | * If some other CPU has already reported non-idle, if this is |
| 2736 | * not the flavor of RCU that tracks sysidle state, or if this |
| 2737 | * is an offline or the timekeeping CPU, nothing to do. |
| 2738 | */ |
| 2739 | if (!*isidle || rdp->rsp != rcu_sysidle_state || |
| 2740 | cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu) |
| 2741 | return; |
| 2742 | if (rcu_gp_in_progress(rdp->rsp)) |
| 2743 | WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu); |
| 2744 | |
| 2745 | /* Pick up current idle and NMI-nesting counter and check. */ |
| 2746 | cur = atomic_read(&rdtp->dynticks_idle); |
| 2747 | if (cur & 0x1) { |
| 2748 | *isidle = false; /* We are not idle! */ |
| 2749 | return; |
| 2750 | } |
| 2751 | smp_mb(); /* Read counters before timestamps. */ |
| 2752 | |
| 2753 | /* Pick up timestamps. */ |
| 2754 | j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies); |
| 2755 | /* If this CPU entered idle more recently, update maxj timestamp. */ |
| 2756 | if (ULONG_CMP_LT(*maxj, j)) |
| 2757 | *maxj = j; |
| 2758 | } |
| 2759 | |
| 2760 | /* |
| 2761 | * Is this the flavor of RCU that is handling full-system idle? |
| 2762 | */ |
| 2763 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) |
| 2764 | { |
| 2765 | return rsp == rcu_sysidle_state; |
| 2766 | } |
| 2767 | |
| 2768 | /* |
| 2769 | * Return a delay in jiffies based on the number of CPUs, rcu_node |
| 2770 | * leaf fanout, and jiffies tick rate. The idea is to allow larger |
| 2771 | * systems more time to transition to full-idle state in order to |
| 2772 | * avoid the cache thrashing that otherwise occur on the state variable. |
| 2773 | * Really small systems (less than a couple of tens of CPUs) should |
| 2774 | * instead use a single global atomically incremented counter, and later |
| 2775 | * versions of this will automatically reconfigure themselves accordingly. |
| 2776 | */ |
| 2777 | static unsigned long rcu_sysidle_delay(void) |
| 2778 | { |
| 2779 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) |
| 2780 | return 0; |
| 2781 | return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000); |
| 2782 | } |
| 2783 | |
| 2784 | /* |
| 2785 | * Advance the full-system-idle state. This is invoked when all of |
| 2786 | * the non-timekeeping CPUs are idle. |
| 2787 | */ |
| 2788 | static void rcu_sysidle(unsigned long j) |
| 2789 | { |
| 2790 | /* Check the current state. */ |
| 2791 | switch (ACCESS_ONCE(full_sysidle_state)) { |
| 2792 | case RCU_SYSIDLE_NOT: |
| 2793 | |
| 2794 | /* First time all are idle, so note a short idle period. */ |
| 2795 | ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT; |
| 2796 | break; |
| 2797 | |
| 2798 | case RCU_SYSIDLE_SHORT: |
| 2799 | |
| 2800 | /* |
| 2801 | * Idle for a bit, time to advance to next state? |
| 2802 | * cmpxchg failure means race with non-idle, let them win. |
| 2803 | */ |
| 2804 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) |
| 2805 | (void)cmpxchg(&full_sysidle_state, |
| 2806 | RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG); |
| 2807 | break; |
| 2808 | |
| 2809 | case RCU_SYSIDLE_LONG: |
| 2810 | |
| 2811 | /* |
| 2812 | * Do an additional check pass before advancing to full. |
| 2813 | * cmpxchg failure means race with non-idle, let them win. |
| 2814 | */ |
| 2815 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) |
| 2816 | (void)cmpxchg(&full_sysidle_state, |
| 2817 | RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL); |
| 2818 | break; |
| 2819 | |
| 2820 | default: |
| 2821 | break; |
| 2822 | } |
| 2823 | } |
| 2824 | |
| 2825 | /* |
| 2826 | * Found a non-idle non-timekeeping CPU, so kick the system-idle state |
| 2827 | * back to the beginning. |
| 2828 | */ |
| 2829 | static void rcu_sysidle_cancel(void) |
| 2830 | { |
| 2831 | smp_mb(); |
| 2832 | if (full_sysidle_state > RCU_SYSIDLE_SHORT) |
| 2833 | ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT; |
| 2834 | } |
| 2835 | |
| 2836 | /* |
| 2837 | * Update the sysidle state based on the results of a force-quiescent-state |
| 2838 | * scan of the CPUs' dyntick-idle state. |
| 2839 | */ |
| 2840 | static void rcu_sysidle_report(struct rcu_state *rsp, int isidle, |
| 2841 | unsigned long maxj, bool gpkt) |
| 2842 | { |
| 2843 | if (rsp != rcu_sysidle_state) |
| 2844 | return; /* Wrong flavor, ignore. */ |
| 2845 | if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) |
| 2846 | return; /* Running state machine from timekeeping CPU. */ |
| 2847 | if (isidle) |
| 2848 | rcu_sysidle(maxj); /* More idle! */ |
| 2849 | else |
| 2850 | rcu_sysidle_cancel(); /* Idle is over. */ |
| 2851 | } |
| 2852 | |
| 2853 | /* |
| 2854 | * Wrapper for rcu_sysidle_report() when called from the grace-period |
| 2855 | * kthread's context. |
| 2856 | */ |
| 2857 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, |
| 2858 | unsigned long maxj) |
| 2859 | { |
| 2860 | rcu_sysidle_report(rsp, isidle, maxj, true); |
| 2861 | } |
| 2862 | |
| 2863 | /* Callback and function for forcing an RCU grace period. */ |
| 2864 | struct rcu_sysidle_head { |
| 2865 | struct rcu_head rh; |
| 2866 | int inuse; |
| 2867 | }; |
| 2868 | |
| 2869 | static void rcu_sysidle_cb(struct rcu_head *rhp) |
| 2870 | { |
| 2871 | struct rcu_sysidle_head *rshp; |
| 2872 | |
| 2873 | /* |
| 2874 | * The following memory barrier is needed to replace the |
| 2875 | * memory barriers that would normally be in the memory |
| 2876 | * allocator. |
| 2877 | */ |
| 2878 | smp_mb(); /* grace period precedes setting inuse. */ |
| 2879 | |
| 2880 | rshp = container_of(rhp, struct rcu_sysidle_head, rh); |
| 2881 | ACCESS_ONCE(rshp->inuse) = 0; |
| 2882 | } |
| 2883 | |
| 2884 | /* |
| 2885 | * Check to see if the system is fully idle, other than the timekeeping CPU. |
| 2886 | * The caller must have disabled interrupts. |
| 2887 | */ |
| 2888 | bool rcu_sys_is_idle(void) |
| 2889 | { |
| 2890 | static struct rcu_sysidle_head rsh; |
| 2891 | int rss = ACCESS_ONCE(full_sysidle_state); |
| 2892 | |
| 2893 | if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu)) |
| 2894 | return false; |
| 2895 | |
| 2896 | /* Handle small-system case by doing a full scan of CPUs. */ |
| 2897 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) { |
| 2898 | int oldrss = rss - 1; |
| 2899 | |
| 2900 | /* |
| 2901 | * One pass to advance to each state up to _FULL. |
| 2902 | * Give up if any pass fails to advance the state. |
| 2903 | */ |
| 2904 | while (rss < RCU_SYSIDLE_FULL && oldrss < rss) { |
| 2905 | int cpu; |
| 2906 | bool isidle = true; |
| 2907 | unsigned long maxj = jiffies - ULONG_MAX / 4; |
| 2908 | struct rcu_data *rdp; |
| 2909 | |
| 2910 | /* Scan all the CPUs looking for nonidle CPUs. */ |
| 2911 | for_each_possible_cpu(cpu) { |
| 2912 | rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu); |
| 2913 | rcu_sysidle_check_cpu(rdp, &isidle, &maxj); |
| 2914 | if (!isidle) |
| 2915 | break; |
| 2916 | } |
| 2917 | rcu_sysidle_report(rcu_sysidle_state, |
| 2918 | isidle, maxj, false); |
| 2919 | oldrss = rss; |
| 2920 | rss = ACCESS_ONCE(full_sysidle_state); |
| 2921 | } |
| 2922 | } |
| 2923 | |
| 2924 | /* If this is the first observation of an idle period, record it. */ |
| 2925 | if (rss == RCU_SYSIDLE_FULL) { |
| 2926 | rss = cmpxchg(&full_sysidle_state, |
| 2927 | RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED); |
| 2928 | return rss == RCU_SYSIDLE_FULL; |
| 2929 | } |
| 2930 | |
| 2931 | smp_mb(); /* ensure rss load happens before later caller actions. */ |
| 2932 | |
| 2933 | /* If already fully idle, tell the caller (in case of races). */ |
| 2934 | if (rss == RCU_SYSIDLE_FULL_NOTED) |
| 2935 | return true; |
| 2936 | |
| 2937 | /* |
| 2938 | * If we aren't there yet, and a grace period is not in flight, |
| 2939 | * initiate a grace period. Either way, tell the caller that |
| 2940 | * we are not there yet. We use an xchg() rather than an assignment |
| 2941 | * to make up for the memory barriers that would otherwise be |
| 2942 | * provided by the memory allocator. |
| 2943 | */ |
| 2944 | if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL && |
| 2945 | !rcu_gp_in_progress(rcu_sysidle_state) && |
| 2946 | !rsh.inuse && xchg(&rsh.inuse, 1) == 0) |
| 2947 | call_rcu(&rsh.rh, rcu_sysidle_cb); |
| 2948 | return false; |
| 2949 | } |
| 2950 | |
| 2951 | /* |
| 2952 | * Initialize dynticks sysidle state for CPUs coming online. |
| 2953 | */ |
| 2954 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) |
| 2955 | { |
| 2956 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE; |
| 2957 | } |
| 2958 | |
| 2959 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
| 2960 | |
| 2961 | static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq) |
| 2962 | { |
| 2963 | } |
| 2964 | |
| 2965 | static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq) |
| 2966 | { |
| 2967 | } |
| 2968 | |
| 2969 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, |
| 2970 | unsigned long *maxj) |
| 2971 | { |
| 2972 | } |
| 2973 | |
| 2974 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) |
| 2975 | { |
| 2976 | return false; |
| 2977 | } |
| 2978 | |
| 2979 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, |
| 2980 | unsigned long maxj) |
| 2981 | { |
| 2982 | } |
| 2983 | |
| 2984 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) |
| 2985 | { |
| 2986 | } |
| 2987 | |
| 2988 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
| 2989 | |
| 2990 | /* |
| 2991 | * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the |
| 2992 | * grace-period kthread will do force_quiescent_state() processing? |
| 2993 | * The idea is to avoid waking up RCU core processing on such a |
| 2994 | * CPU unless the grace period has extended for too long. |
| 2995 | * |
| 2996 | * This code relies on the fact that all NO_HZ_FULL CPUs are also |
| 2997 | * CONFIG_RCU_NOCB_CPU CPUs. |
| 2998 | */ |
| 2999 | static bool rcu_nohz_full_cpu(struct rcu_state *rsp) |
| 3000 | { |
| 3001 | #ifdef CONFIG_NO_HZ_FULL |
| 3002 | if (tick_nohz_full_cpu(smp_processor_id()) && |
| 3003 | (!rcu_gp_in_progress(rsp) || |
| 3004 | ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ))) |
| 3005 | return 1; |
| 3006 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ |
| 3007 | return 0; |
| 3008 | } |
| 3009 | |
| 3010 | /* |
| 3011 | * Bind the grace-period kthread for the sysidle flavor of RCU to the |
| 3012 | * timekeeping CPU. |
| 3013 | */ |
| 3014 | static void rcu_bind_gp_kthread(void) |
| 3015 | { |
| 3016 | #ifdef CONFIG_NO_HZ_FULL |
| 3017 | int cpu = ACCESS_ONCE(tick_do_timer_cpu); |
| 3018 | |
| 3019 | if (cpu < 0 || cpu >= nr_cpu_ids) |
| 3020 | return; |
| 3021 | if (raw_smp_processor_id() != cpu) |
| 3022 | set_cpus_allowed_ptr(current, cpumask_of(cpu)); |
| 3023 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ |
| 3024 | } |