Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Read-Copy Update mechanism for mutual exclusion (tree-based version) |
| 3 | * Internal non-public definitions that provide either classic |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 4 | * or preemptible semantics. |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 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, write to the Free Software |
| 18 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. |
| 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 | |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 27 | #include <linux/delay.h> |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 28 | #include <linux/gfp.h> |
Paul E. McKenney | b626c1b | 2012-06-11 17:39:43 -0700 | [diff] [blame] | 29 | #include <linux/oom.h> |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 30 | #include <linux/smpboot.h> |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 31 | |
Mike Galbraith | 5b61b0b | 2011-08-19 11:39:11 -0700 | [diff] [blame] | 32 | #define RCU_KTHREAD_PRIO 1 |
| 33 | |
| 34 | #ifdef CONFIG_RCU_BOOST |
| 35 | #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO |
| 36 | #else |
| 37 | #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO |
| 38 | #endif |
| 39 | |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 40 | #ifdef CONFIG_RCU_NOCB_CPU |
| 41 | static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */ |
| 42 | static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */ |
Paul Gortmaker | 1b0048a | 2012-12-20 13:19:22 -0800 | [diff] [blame] | 43 | static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */ |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 44 | static char __initdata nocb_buf[NR_CPUS * 5]; |
| 45 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ |
| 46 | |
Paul E. McKenney | 26845c2 | 2010-04-13 14:19:23 -0700 | [diff] [blame] | 47 | /* |
| 48 | * Check the RCU kernel configuration parameters and print informative |
| 49 | * messages about anything out of the ordinary. If you like #ifdef, you |
| 50 | * will love this function. |
| 51 | */ |
| 52 | static void __init rcu_bootup_announce_oddness(void) |
| 53 | { |
| 54 | #ifdef CONFIG_RCU_TRACE |
| 55 | printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n"); |
| 56 | #endif |
| 57 | #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32) |
| 58 | printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n", |
| 59 | CONFIG_RCU_FANOUT); |
| 60 | #endif |
| 61 | #ifdef CONFIG_RCU_FANOUT_EXACT |
| 62 | printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n"); |
| 63 | #endif |
| 64 | #ifdef CONFIG_RCU_FAST_NO_HZ |
| 65 | printk(KERN_INFO |
| 66 | "\tRCU dyntick-idle grace-period acceleration is enabled.\n"); |
| 67 | #endif |
| 68 | #ifdef CONFIG_PROVE_RCU |
| 69 | printk(KERN_INFO "\tRCU lockdep checking is enabled.\n"); |
| 70 | #endif |
| 71 | #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE |
| 72 | printk(KERN_INFO "\tRCU torture testing starts during boot.\n"); |
| 73 | #endif |
Paul E. McKenney | 81a294c | 2010-08-30 09:52:50 -0700 | [diff] [blame] | 74 | #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE) |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 75 | printk(KERN_INFO "\tDump stacks of tasks blocking RCU-preempt GP.\n"); |
| 76 | #endif |
| 77 | #if defined(CONFIG_RCU_CPU_STALL_INFO) |
| 78 | printk(KERN_INFO "\tAdditional per-CPU info printed with stalls.\n"); |
Paul E. McKenney | 26845c2 | 2010-04-13 14:19:23 -0700 | [diff] [blame] | 79 | #endif |
| 80 | #if NUM_RCU_LVL_4 != 0 |
Paul E. McKenney | cc5df65 | 2012-06-15 18:16:00 -0700 | [diff] [blame] | 81 | printk(KERN_INFO "\tFour-level hierarchy is enabled.\n"); |
Paul E. McKenney | 26845c2 | 2010-04-13 14:19:23 -0700 | [diff] [blame] | 82 | #endif |
Paul E. McKenney | f885b7f | 2012-04-23 15:52:53 -0700 | [diff] [blame] | 83 | if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF) |
| 84 | printk(KERN_INFO "\tExperimental boot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf); |
Paul E. McKenney | cca6f39 | 2012-05-08 21:00:28 -0700 | [diff] [blame] | 85 | if (nr_cpu_ids != NR_CPUS) |
| 86 | printk(KERN_INFO "\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids); |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 87 | #ifdef CONFIG_RCU_NOCB_CPU |
Paul E. McKenney | 911af50 | 2013-02-11 10:23:27 -0800 | [diff] [blame] | 88 | #ifndef CONFIG_RCU_NOCB_CPU_NONE |
| 89 | if (!have_rcu_nocb_mask) { |
| 90 | alloc_bootmem_cpumask_var(&rcu_nocb_mask); |
| 91 | have_rcu_nocb_mask = true; |
| 92 | } |
| 93 | #ifdef CONFIG_RCU_NOCB_CPU_ZERO |
| 94 | pr_info("\tExperimental no-CBs 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("\tExperimental no-CBs for all CPUs\n"); |
| 99 | cpumask_setall(rcu_nocb_mask); |
| 100 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */ |
| 101 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */ |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 102 | if (have_rcu_nocb_mask) { |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 103 | cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask); |
| 104 | pr_info("\tExperimental no-CBs CPUs: %s.\n", nocb_buf); |
| 105 | if (rcu_nocb_poll) |
| 106 | pr_info("\tExperimental polled no-CBs CPUs.\n"); |
| 107 | } |
| 108 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ |
Paul E. McKenney | 26845c2 | 2010-04-13 14:19:23 -0700 | [diff] [blame] | 109 | } |
| 110 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 111 | #ifdef CONFIG_TREE_PREEMPT_RCU |
| 112 | |
Paul E. McKenney | 037b64e | 2012-05-28 23:26:01 -0700 | [diff] [blame] | 113 | struct rcu_state rcu_preempt_state = |
Paul E. McKenney | a488985 | 2012-12-03 08:16:28 -0800 | [diff] [blame^] | 114 | RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 115 | DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 116 | static struct rcu_state *rcu_state = &rcu_preempt_state; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 117 | |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 118 | static int rcu_preempted_readers_exp(struct rcu_node *rnp); |
| 119 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 120 | /* |
| 121 | * Tell them what RCU they are running. |
| 122 | */ |
Paul E. McKenney | 0e0fc1c | 2009-11-11 11:28:06 -0800 | [diff] [blame] | 123 | static void __init rcu_bootup_announce(void) |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 124 | { |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 125 | printk(KERN_INFO "Preemptible hierarchical RCU implementation.\n"); |
Paul E. McKenney | 26845c2 | 2010-04-13 14:19:23 -0700 | [diff] [blame] | 126 | rcu_bootup_announce_oddness(); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 127 | } |
| 128 | |
| 129 | /* |
| 130 | * Return the number of RCU-preempt batches processed thus far |
| 131 | * for debug and statistics. |
| 132 | */ |
| 133 | long rcu_batches_completed_preempt(void) |
| 134 | { |
| 135 | return rcu_preempt_state.completed; |
| 136 | } |
| 137 | EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt); |
| 138 | |
| 139 | /* |
| 140 | * Return the number of RCU batches processed thus far for debug & stats. |
| 141 | */ |
| 142 | long rcu_batches_completed(void) |
| 143 | { |
| 144 | return rcu_batches_completed_preempt(); |
| 145 | } |
| 146 | EXPORT_SYMBOL_GPL(rcu_batches_completed); |
| 147 | |
| 148 | /* |
Paul E. McKenney | bf66f18 | 2010-01-04 15:09:10 -0800 | [diff] [blame] | 149 | * Force a quiescent state for preemptible RCU. |
| 150 | */ |
| 151 | void rcu_force_quiescent_state(void) |
| 152 | { |
Paul E. McKenney | 4cdfc17 | 2012-06-22 17:06:26 -0700 | [diff] [blame] | 153 | force_quiescent_state(&rcu_preempt_state); |
Paul E. McKenney | bf66f18 | 2010-01-04 15:09:10 -0800 | [diff] [blame] | 154 | } |
| 155 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); |
| 156 | |
| 157 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 158 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 159 | * that this just means that the task currently running on the CPU is |
| 160 | * not in a quiescent state. There might be any number of tasks blocked |
| 161 | * while in an RCU read-side critical section. |
Paul E. McKenney | 25502a6 | 2010-04-01 17:37:01 -0700 | [diff] [blame] | 162 | * |
| 163 | * Unlike the other rcu_*_qs() functions, callers to this function |
| 164 | * must disable irqs in order to protect the assignment to |
| 165 | * ->rcu_read_unlock_special. |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 166 | */ |
Paul E. McKenney | c3422be | 2009-09-13 09:15:10 -0700 | [diff] [blame] | 167 | static void rcu_preempt_qs(int cpu) |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 168 | { |
| 169 | struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu); |
Paul E. McKenney | 25502a6 | 2010-04-01 17:37:01 -0700 | [diff] [blame] | 170 | |
Paul E. McKenney | e4cc1f2 | 2011-06-27 00:17:43 -0700 | [diff] [blame] | 171 | if (rdp->passed_quiesce == 0) |
Paul E. McKenney | d4c08f2 | 2011-06-25 06:36:56 -0700 | [diff] [blame] | 172 | trace_rcu_grace_period("rcu_preempt", rdp->gpnum, "cpuqs"); |
Paul E. McKenney | e4cc1f2 | 2011-06-27 00:17:43 -0700 | [diff] [blame] | 173 | rdp->passed_quiesce = 1; |
Paul E. McKenney | 25502a6 | 2010-04-01 17:37:01 -0700 | [diff] [blame] | 174 | current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 175 | } |
| 176 | |
| 177 | /* |
Paul E. McKenney | c3422be | 2009-09-13 09:15:10 -0700 | [diff] [blame] | 178 | * We have entered the scheduler, and the current task might soon be |
| 179 | * context-switched away from. If this task is in an RCU read-side |
| 180 | * critical section, we will no longer be able to rely on the CPU to |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 181 | * record that fact, so we enqueue the task on the blkd_tasks list. |
| 182 | * The task will dequeue itself when it exits the outermost enclosing |
| 183 | * RCU read-side critical section. Therefore, the current grace period |
| 184 | * cannot be permitted to complete until the blkd_tasks list entries |
| 185 | * predating the current grace period drain, in other words, until |
| 186 | * rnp->gp_tasks becomes NULL. |
Paul E. McKenney | c3422be | 2009-09-13 09:15:10 -0700 | [diff] [blame] | 187 | * |
| 188 | * Caller must disable preemption. |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 189 | */ |
Paul E. McKenney | cba6d0d | 2012-07-02 07:08:42 -0700 | [diff] [blame] | 190 | static void rcu_preempt_note_context_switch(int cpu) |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 191 | { |
| 192 | struct task_struct *t = current; |
Paul E. McKenney | c3422be | 2009-09-13 09:15:10 -0700 | [diff] [blame] | 193 | unsigned long flags; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 194 | struct rcu_data *rdp; |
| 195 | struct rcu_node *rnp; |
| 196 | |
Paul E. McKenney | 10f39bb | 2011-07-17 21:14:35 -0700 | [diff] [blame] | 197 | if (t->rcu_read_lock_nesting > 0 && |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 198 | (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) { |
| 199 | |
| 200 | /* Possibly blocking in an RCU read-side critical section. */ |
Paul E. McKenney | cba6d0d | 2012-07-02 07:08:42 -0700 | [diff] [blame] | 201 | rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 202 | rnp = rdp->mynode; |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 203 | raw_spin_lock_irqsave(&rnp->lock, flags); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 204 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED; |
Paul E. McKenney | 8684896 | 2009-08-27 15:00:12 -0700 | [diff] [blame] | 205 | t->rcu_blocked_node = rnp; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 206 | |
| 207 | /* |
| 208 | * If this CPU has already checked in, then this task |
| 209 | * will hold up the next grace period rather than the |
| 210 | * current grace period. Queue the task accordingly. |
| 211 | * If the task is queued for the current grace period |
| 212 | * (i.e., this CPU has not yet passed through a quiescent |
| 213 | * state for the current grace period), then as long |
| 214 | * as that task remains queued, the current grace period |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 215 | * cannot end. Note that there is some uncertainty as |
| 216 | * to exactly when the current grace period started. |
| 217 | * We take a conservative approach, which can result |
| 218 | * in unnecessarily waiting on tasks that started very |
| 219 | * slightly after the current grace period began. C'est |
| 220 | * la vie!!! |
Paul E. McKenney | b0e165c | 2009-09-13 09:15:09 -0700 | [diff] [blame] | 221 | * |
| 222 | * But first, note that the current CPU must still be |
| 223 | * on line! |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 224 | */ |
Paul E. McKenney | b0e165c | 2009-09-13 09:15:09 -0700 | [diff] [blame] | 225 | WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0); |
Paul E. McKenney | e7d8842 | 2009-09-18 09:50:18 -0700 | [diff] [blame] | 226 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 227 | if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) { |
| 228 | list_add(&t->rcu_node_entry, rnp->gp_tasks->prev); |
| 229 | rnp->gp_tasks = &t->rcu_node_entry; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 230 | #ifdef CONFIG_RCU_BOOST |
| 231 | if (rnp->boost_tasks != NULL) |
| 232 | rnp->boost_tasks = rnp->gp_tasks; |
| 233 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 234 | } else { |
| 235 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); |
| 236 | if (rnp->qsmask & rdp->grpmask) |
| 237 | rnp->gp_tasks = &t->rcu_node_entry; |
| 238 | } |
Paul E. McKenney | d4c08f2 | 2011-06-25 06:36:56 -0700 | [diff] [blame] | 239 | trace_rcu_preempt_task(rdp->rsp->name, |
| 240 | t->pid, |
| 241 | (rnp->qsmask & rdp->grpmask) |
| 242 | ? rnp->gpnum |
| 243 | : rnp->gpnum + 1); |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 244 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | 10f39bb | 2011-07-17 21:14:35 -0700 | [diff] [blame] | 245 | } else if (t->rcu_read_lock_nesting < 0 && |
| 246 | t->rcu_read_unlock_special) { |
| 247 | |
| 248 | /* |
| 249 | * Complete exit from RCU read-side critical section on |
| 250 | * behalf of preempted instance of __rcu_read_unlock(). |
| 251 | */ |
| 252 | rcu_read_unlock_special(t); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 253 | } |
| 254 | |
| 255 | /* |
| 256 | * Either we were not in an RCU read-side critical section to |
| 257 | * begin with, or we have now recorded that critical section |
| 258 | * globally. Either way, we can now note a quiescent state |
| 259 | * for this CPU. Again, if we were in an RCU read-side critical |
| 260 | * section, and if that critical section was blocking the current |
| 261 | * grace period, then the fact that the task has been enqueued |
| 262 | * means that we continue to block the current grace period. |
| 263 | */ |
Paul E. McKenney | e7d8842 | 2009-09-18 09:50:18 -0700 | [diff] [blame] | 264 | local_irq_save(flags); |
Paul E. McKenney | cba6d0d | 2012-07-02 07:08:42 -0700 | [diff] [blame] | 265 | rcu_preempt_qs(cpu); |
Paul E. McKenney | e7d8842 | 2009-09-18 09:50:18 -0700 | [diff] [blame] | 266 | local_irq_restore(flags); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 267 | } |
| 268 | |
| 269 | /* |
Paul E. McKenney | fc2219d | 2009-09-23 09:50:41 -0700 | [diff] [blame] | 270 | * Check for preempted RCU readers blocking the current grace period |
| 271 | * for the specified rcu_node structure. If the caller needs a reliable |
| 272 | * answer, it must hold the rcu_node's ->lock. |
| 273 | */ |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 274 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
Paul E. McKenney | fc2219d | 2009-09-23 09:50:41 -0700 | [diff] [blame] | 275 | { |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 276 | return rnp->gp_tasks != NULL; |
Paul E. McKenney | fc2219d | 2009-09-23 09:50:41 -0700 | [diff] [blame] | 277 | } |
| 278 | |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 279 | /* |
| 280 | * Record a quiescent state for all tasks that were previously queued |
| 281 | * on the specified rcu_node structure and that were blocking the current |
| 282 | * RCU grace period. The caller must hold the specified rnp->lock with |
| 283 | * irqs disabled, and this lock is released upon return, but irqs remain |
| 284 | * disabled. |
| 285 | */ |
Paul E. McKenney | d3f6bad | 2009-12-02 12:10:13 -0800 | [diff] [blame] | 286 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 287 | __releases(rnp->lock) |
| 288 | { |
| 289 | unsigned long mask; |
| 290 | struct rcu_node *rnp_p; |
| 291 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 292 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 293 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 294 | return; /* Still need more quiescent states! */ |
| 295 | } |
| 296 | |
| 297 | rnp_p = rnp->parent; |
| 298 | if (rnp_p == NULL) { |
| 299 | /* |
| 300 | * Either there is only one rcu_node in the tree, |
| 301 | * or tasks were kicked up to root rcu_node due to |
| 302 | * CPUs going offline. |
| 303 | */ |
Paul E. McKenney | d3f6bad | 2009-12-02 12:10:13 -0800 | [diff] [blame] | 304 | rcu_report_qs_rsp(&rcu_preempt_state, flags); |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 305 | return; |
| 306 | } |
| 307 | |
| 308 | /* Report up the rest of the hierarchy. */ |
| 309 | mask = rnp->grpmask; |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 310 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
| 311 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ |
Paul E. McKenney | d3f6bad | 2009-12-02 12:10:13 -0800 | [diff] [blame] | 312 | rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags); |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 313 | } |
| 314 | |
| 315 | /* |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 316 | * Advance a ->blkd_tasks-list pointer to the next entry, instead |
| 317 | * returning NULL if at the end of the list. |
| 318 | */ |
| 319 | static struct list_head *rcu_next_node_entry(struct task_struct *t, |
| 320 | struct rcu_node *rnp) |
| 321 | { |
| 322 | struct list_head *np; |
| 323 | |
| 324 | np = t->rcu_node_entry.next; |
| 325 | if (np == &rnp->blkd_tasks) |
| 326 | np = NULL; |
| 327 | return np; |
| 328 | } |
| 329 | |
| 330 | /* |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 331 | * Handle special cases during rcu_read_unlock(), such as needing to |
| 332 | * notify RCU core processing or task having blocked during the RCU |
| 333 | * read-side critical section. |
| 334 | */ |
Paul E. McKenney | 2a3fa84 | 2012-05-21 11:58:36 -0700 | [diff] [blame] | 335 | void rcu_read_unlock_special(struct task_struct *t) |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 336 | { |
| 337 | int empty; |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 338 | int empty_exp; |
Paul E. McKenney | 389abd4 | 2011-09-21 14:41:37 -0700 | [diff] [blame] | 339 | int empty_exp_now; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 340 | unsigned long flags; |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 341 | struct list_head *np; |
Paul E. McKenney | 82e78d8 | 2011-08-04 07:55:34 -0700 | [diff] [blame] | 342 | #ifdef CONFIG_RCU_BOOST |
| 343 | struct rt_mutex *rbmp = NULL; |
| 344 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 345 | struct rcu_node *rnp; |
| 346 | int special; |
| 347 | |
| 348 | /* NMI handlers cannot block and cannot safely manipulate state. */ |
| 349 | if (in_nmi()) |
| 350 | return; |
| 351 | |
| 352 | local_irq_save(flags); |
| 353 | |
| 354 | /* |
| 355 | * If RCU core is waiting for this CPU to exit critical section, |
| 356 | * let it know that we have done so. |
| 357 | */ |
| 358 | special = t->rcu_read_unlock_special; |
| 359 | if (special & RCU_READ_UNLOCK_NEED_QS) { |
Paul E. McKenney | c3422be | 2009-09-13 09:15:10 -0700 | [diff] [blame] | 360 | rcu_preempt_qs(smp_processor_id()); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 361 | } |
| 362 | |
| 363 | /* Hardware IRQ handlers cannot block. */ |
Peter Zijlstra | ec433f0 | 2011-07-19 15:32:00 -0700 | [diff] [blame] | 364 | if (in_irq() || in_serving_softirq()) { |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 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 | |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 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 (;;) { |
Paul E. McKenney | 8684896 | 2009-08-27 15:00:12 -0700 | [diff] [blame] | 379 | rnp = t->rcu_blocked_node; |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 380 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
Paul E. McKenney | 8684896 | 2009-08-27 15:00:12 -0700 | [diff] [blame] | 381 | if (rnp == t->rcu_blocked_node) |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 382 | break; |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 383 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 384 | } |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 385 | empty = !rcu_preempt_blocked_readers_cgp(rnp); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 386 | empty_exp = !rcu_preempted_readers_exp(rnp); |
| 387 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 388 | np = rcu_next_node_entry(t, rnp); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 389 | list_del_init(&t->rcu_node_entry); |
Paul E. McKenney | 82e78d8 | 2011-08-04 07:55:34 -0700 | [diff] [blame] | 390 | t->rcu_blocked_node = NULL; |
Paul E. McKenney | d4c08f2 | 2011-06-25 06:36:56 -0700 | [diff] [blame] | 391 | trace_rcu_unlock_preempted_task("rcu_preempt", |
| 392 | rnp->gpnum, t->pid); |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 393 | if (&t->rcu_node_entry == rnp->gp_tasks) |
| 394 | rnp->gp_tasks = np; |
| 395 | if (&t->rcu_node_entry == rnp->exp_tasks) |
| 396 | rnp->exp_tasks = np; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 397 | #ifdef CONFIG_RCU_BOOST |
| 398 | if (&t->rcu_node_entry == rnp->boost_tasks) |
| 399 | rnp->boost_tasks = np; |
Paul E. McKenney | 82e78d8 | 2011-08-04 07:55:34 -0700 | [diff] [blame] | 400 | /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */ |
| 401 | if (t->rcu_boost_mutex) { |
| 402 | rbmp = t->rcu_boost_mutex; |
| 403 | t->rcu_boost_mutex = NULL; |
Paul E. McKenney | 7765be2 | 2011-07-14 12:24:11 -0700 | [diff] [blame] | 404 | } |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 405 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 406 | |
| 407 | /* |
| 408 | * If this was the last task on the current list, and if |
| 409 | * we aren't waiting on any CPUs, report the quiescent state. |
Paul E. McKenney | 389abd4 | 2011-09-21 14:41:37 -0700 | [diff] [blame] | 410 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock, |
| 411 | * so we must take a snapshot of the expedited state. |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 412 | */ |
Paul E. McKenney | 389abd4 | 2011-09-21 14:41:37 -0700 | [diff] [blame] | 413 | empty_exp_now = !rcu_preempted_readers_exp(rnp); |
Paul E. McKenney | d4c08f2 | 2011-06-25 06:36:56 -0700 | [diff] [blame] | 414 | if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) { |
| 415 | trace_rcu_quiescent_state_report("preempt_rcu", |
| 416 | rnp->gpnum, |
| 417 | 0, rnp->qsmask, |
| 418 | rnp->level, |
| 419 | rnp->grplo, |
| 420 | rnp->grphi, |
| 421 | !!rnp->gp_tasks); |
Paul E. McKenney | d3f6bad | 2009-12-02 12:10:13 -0800 | [diff] [blame] | 422 | rcu_report_unblock_qs_rnp(rnp, flags); |
Paul E. McKenney | c701d5d | 2012-06-28 08:08:25 -0700 | [diff] [blame] | 423 | } else { |
Paul E. McKenney | d4c08f2 | 2011-06-25 06:36:56 -0700 | [diff] [blame] | 424 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | c701d5d | 2012-06-28 08:08:25 -0700 | [diff] [blame] | 425 | } |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 426 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 427 | #ifdef CONFIG_RCU_BOOST |
| 428 | /* Unboost if we were boosted. */ |
Paul E. McKenney | 82e78d8 | 2011-08-04 07:55:34 -0700 | [diff] [blame] | 429 | if (rbmp) |
| 430 | rt_mutex_unlock(rbmp); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 431 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 432 | |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 433 | /* |
| 434 | * If this was the last task on the expedited lists, |
| 435 | * then we need to report up the rcu_node hierarchy. |
| 436 | */ |
Paul E. McKenney | 389abd4 | 2011-09-21 14:41:37 -0700 | [diff] [blame] | 437 | if (!empty_exp && empty_exp_now) |
Thomas Gleixner | b40d293 | 2011-10-22 07:12:34 -0700 | [diff] [blame] | 438 | rcu_report_exp_rnp(&rcu_preempt_state, rnp, true); |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 439 | } else { |
| 440 | local_irq_restore(flags); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 441 | } |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 442 | } |
| 443 | |
Paul E. McKenney | 1ed509a | 2010-02-22 17:05:05 -0800 | [diff] [blame] | 444 | #ifdef CONFIG_RCU_CPU_STALL_VERBOSE |
| 445 | |
| 446 | /* |
| 447 | * Dump detailed information for all tasks blocking the current RCU |
| 448 | * grace period on the specified rcu_node structure. |
| 449 | */ |
| 450 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) |
| 451 | { |
| 452 | unsigned long flags; |
Paul E. McKenney | 1ed509a | 2010-02-22 17:05:05 -0800 | [diff] [blame] | 453 | struct task_struct *t; |
| 454 | |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 455 | raw_spin_lock_irqsave(&rnp->lock, flags); |
Paul E. McKenney | 5fd4dc0 | 2012-08-10 16:00:11 -0700 | [diff] [blame] | 456 | if (!rcu_preempt_blocked_readers_cgp(rnp)) { |
| 457 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 458 | return; |
| 459 | } |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 460 | t = list_entry(rnp->gp_tasks, |
| 461 | struct task_struct, rcu_node_entry); |
| 462 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) |
| 463 | sched_show_task(t); |
| 464 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | 1ed509a | 2010-02-22 17:05:05 -0800 | [diff] [blame] | 465 | } |
| 466 | |
| 467 | /* |
| 468 | * Dump detailed information for all tasks blocking the current RCU |
| 469 | * grace period. |
| 470 | */ |
| 471 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) |
| 472 | { |
| 473 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 474 | |
| 475 | rcu_print_detail_task_stall_rnp(rnp); |
| 476 | rcu_for_each_leaf_node(rsp, rnp) |
| 477 | rcu_print_detail_task_stall_rnp(rnp); |
| 478 | } |
| 479 | |
| 480 | #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ |
| 481 | |
| 482 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) |
| 483 | { |
| 484 | } |
| 485 | |
| 486 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ |
| 487 | |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 488 | #ifdef CONFIG_RCU_CPU_STALL_INFO |
| 489 | |
| 490 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) |
| 491 | { |
| 492 | printk(KERN_ERR "\tTasks blocked on level-%d rcu_node (CPUs %d-%d):", |
| 493 | rnp->level, rnp->grplo, rnp->grphi); |
| 494 | } |
| 495 | |
| 496 | static void rcu_print_task_stall_end(void) |
| 497 | { |
| 498 | printk(KERN_CONT "\n"); |
| 499 | } |
| 500 | |
| 501 | #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */ |
| 502 | |
| 503 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) |
| 504 | { |
| 505 | } |
| 506 | |
| 507 | static void rcu_print_task_stall_end(void) |
| 508 | { |
| 509 | } |
| 510 | |
| 511 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */ |
| 512 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 513 | /* |
| 514 | * Scan the current list of tasks blocked within RCU read-side critical |
| 515 | * sections, printing out the tid of each. |
| 516 | */ |
Paul E. McKenney | 9bc8b55 | 2011-08-13 13:31:47 -0700 | [diff] [blame] | 517 | static int rcu_print_task_stall(struct rcu_node *rnp) |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 518 | { |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 519 | struct task_struct *t; |
Paul E. McKenney | 9bc8b55 | 2011-08-13 13:31:47 -0700 | [diff] [blame] | 520 | int ndetected = 0; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 521 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 522 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
Paul E. McKenney | 9bc8b55 | 2011-08-13 13:31:47 -0700 | [diff] [blame] | 523 | return 0; |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 524 | rcu_print_task_stall_begin(rnp); |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 525 | t = list_entry(rnp->gp_tasks, |
| 526 | struct task_struct, rcu_node_entry); |
Paul E. McKenney | 9bc8b55 | 2011-08-13 13:31:47 -0700 | [diff] [blame] | 527 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 528 | printk(KERN_CONT " P%d", t->pid); |
Paul E. McKenney | 9bc8b55 | 2011-08-13 13:31:47 -0700 | [diff] [blame] | 529 | ndetected++; |
| 530 | } |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 531 | rcu_print_task_stall_end(); |
Paul E. McKenney | 9bc8b55 | 2011-08-13 13:31:47 -0700 | [diff] [blame] | 532 | return ndetected; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 533 | } |
| 534 | |
Paul E. McKenney | 53d84e0 | 2010-08-10 14:28:53 -0700 | [diff] [blame] | 535 | /* |
Paul E. McKenney | b0e165c | 2009-09-13 09:15:09 -0700 | [diff] [blame] | 536 | * Check that the list of blocked tasks for the newly completed grace |
| 537 | * period is in fact empty. It is a serious bug to complete a grace |
| 538 | * period that still has RCU readers blocked! This function must be |
| 539 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock |
| 540 | * must be held by the caller. |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 541 | * |
| 542 | * Also, if there are blocked tasks on the list, they automatically |
| 543 | * block the newly created grace period, so set up ->gp_tasks accordingly. |
Paul E. McKenney | b0e165c | 2009-09-13 09:15:09 -0700 | [diff] [blame] | 544 | */ |
| 545 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) |
| 546 | { |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 547 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 548 | if (!list_empty(&rnp->blkd_tasks)) |
| 549 | rnp->gp_tasks = rnp->blkd_tasks.next; |
Paul E. McKenney | 28ecd58 | 2009-09-18 09:50:17 -0700 | [diff] [blame] | 550 | WARN_ON_ONCE(rnp->qsmask); |
Paul E. McKenney | b0e165c | 2009-09-13 09:15:09 -0700 | [diff] [blame] | 551 | } |
| 552 | |
Paul E. McKenney | 33f7614 | 2009-08-24 09:42:01 -0700 | [diff] [blame] | 553 | #ifdef CONFIG_HOTPLUG_CPU |
| 554 | |
| 555 | /* |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 556 | * Handle tasklist migration for case in which all CPUs covered by the |
| 557 | * specified rcu_node have gone offline. Move them up to the root |
| 558 | * rcu_node. The reason for not just moving them to the immediate |
| 559 | * parent is to remove the need for rcu_read_unlock_special() to |
| 560 | * make more than two attempts to acquire the target rcu_node's lock. |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 561 | * Returns true if there were tasks blocking the current RCU grace |
| 562 | * period. |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 563 | * |
Paul E. McKenney | 237c80c | 2009-10-15 09:26:14 -0700 | [diff] [blame] | 564 | * Returns 1 if there was previously a task blocking the current grace |
| 565 | * period on the specified rcu_node structure. |
| 566 | * |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 567 | * The caller must hold rnp->lock with irqs disabled. |
| 568 | */ |
Paul E. McKenney | 237c80c | 2009-10-15 09:26:14 -0700 | [diff] [blame] | 569 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
| 570 | struct rcu_node *rnp, |
| 571 | struct rcu_data *rdp) |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 572 | { |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 573 | struct list_head *lp; |
| 574 | struct list_head *lp_root; |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 575 | int retval = 0; |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 576 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 577 | struct task_struct *t; |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 578 | |
Paul E. McKenney | 8684896 | 2009-08-27 15:00:12 -0700 | [diff] [blame] | 579 | if (rnp == rnp_root) { |
| 580 | WARN_ONCE(1, "Last CPU thought to be offlined?"); |
Paul E. McKenney | 237c80c | 2009-10-15 09:26:14 -0700 | [diff] [blame] | 581 | return 0; /* Shouldn't happen: at least one CPU online. */ |
Paul E. McKenney | 8684896 | 2009-08-27 15:00:12 -0700 | [diff] [blame] | 582 | } |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 583 | |
| 584 | /* If we are on an internal node, complain bitterly. */ |
| 585 | WARN_ON_ONCE(rnp != rdp->mynode); |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 586 | |
| 587 | /* |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 588 | * Move tasks up to root rcu_node. Don't try to get fancy for |
| 589 | * this corner-case operation -- just put this node's tasks |
| 590 | * at the head of the root node's list, and update the root node's |
| 591 | * ->gp_tasks and ->exp_tasks pointers to those of this node's, |
| 592 | * if non-NULL. This might result in waiting for more tasks than |
| 593 | * absolutely necessary, but this is a good performance/complexity |
| 594 | * tradeoff. |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 595 | */ |
Paul E. McKenney | 2036d94 | 2012-01-30 17:02:47 -0800 | [diff] [blame] | 596 | if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0) |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 597 | retval |= RCU_OFL_TASKS_NORM_GP; |
| 598 | if (rcu_preempted_readers_exp(rnp)) |
| 599 | retval |= RCU_OFL_TASKS_EXP_GP; |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 600 | lp = &rnp->blkd_tasks; |
| 601 | lp_root = &rnp_root->blkd_tasks; |
| 602 | while (!list_empty(lp)) { |
| 603 | t = list_entry(lp->next, typeof(*t), rcu_node_entry); |
| 604 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ |
| 605 | list_del(&t->rcu_node_entry); |
| 606 | t->rcu_blocked_node = rnp_root; |
| 607 | list_add(&t->rcu_node_entry, lp_root); |
| 608 | if (&t->rcu_node_entry == rnp->gp_tasks) |
| 609 | rnp_root->gp_tasks = rnp->gp_tasks; |
| 610 | if (&t->rcu_node_entry == rnp->exp_tasks) |
| 611 | rnp_root->exp_tasks = rnp->exp_tasks; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 612 | #ifdef CONFIG_RCU_BOOST |
| 613 | if (&t->rcu_node_entry == rnp->boost_tasks) |
| 614 | rnp_root->boost_tasks = rnp->boost_tasks; |
| 615 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 616 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 617 | } |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 618 | |
Paul E. McKenney | 1e3fd2b | 2012-07-27 13:41:47 -0700 | [diff] [blame] | 619 | rnp->gp_tasks = NULL; |
| 620 | rnp->exp_tasks = NULL; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 621 | #ifdef CONFIG_RCU_BOOST |
Paul E. McKenney | 1e3fd2b | 2012-07-27 13:41:47 -0700 | [diff] [blame] | 622 | rnp->boost_tasks = NULL; |
Paul E. McKenney | 5cc900c | 2012-07-31 14:09:49 -0700 | [diff] [blame] | 623 | /* |
| 624 | * In case root is being boosted and leaf was not. Make sure |
| 625 | * that we boost the tasks blocking the current grace period |
| 626 | * in this case. |
| 627 | */ |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 628 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ |
| 629 | if (rnp_root->boost_tasks != NULL && |
Paul E. McKenney | 5cc900c | 2012-07-31 14:09:49 -0700 | [diff] [blame] | 630 | rnp_root->boost_tasks != rnp_root->gp_tasks && |
| 631 | rnp_root->boost_tasks != rnp_root->exp_tasks) |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 632 | rnp_root->boost_tasks = rnp_root->gp_tasks; |
| 633 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ |
| 634 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 635 | |
Paul E. McKenney | 237c80c | 2009-10-15 09:26:14 -0700 | [diff] [blame] | 636 | return retval; |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 637 | } |
| 638 | |
Paul E. McKenney | e560140 | 2012-01-07 11:03:57 -0800 | [diff] [blame] | 639 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 640 | |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 641 | /* |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 642 | * Check for a quiescent state from the current CPU. When a task blocks, |
| 643 | * the task is recorded in the corresponding CPU's rcu_node structure, |
| 644 | * which is checked elsewhere. |
| 645 | * |
| 646 | * Caller must disable hard irqs. |
| 647 | */ |
| 648 | static void rcu_preempt_check_callbacks(int cpu) |
| 649 | { |
| 650 | struct task_struct *t = current; |
| 651 | |
| 652 | if (t->rcu_read_lock_nesting == 0) { |
Paul E. McKenney | c3422be | 2009-09-13 09:15:10 -0700 | [diff] [blame] | 653 | rcu_preempt_qs(cpu); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 654 | return; |
| 655 | } |
Paul E. McKenney | 10f39bb | 2011-07-17 21:14:35 -0700 | [diff] [blame] | 656 | if (t->rcu_read_lock_nesting > 0 && |
| 657 | per_cpu(rcu_preempt_data, cpu).qs_pending) |
Paul E. McKenney | c3422be | 2009-09-13 09:15:10 -0700 | [diff] [blame] | 658 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 659 | } |
| 660 | |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 661 | #ifdef CONFIG_RCU_BOOST |
| 662 | |
Shaohua Li | 0922337 | 2011-06-14 13:26:25 +0800 | [diff] [blame] | 663 | static void rcu_preempt_do_callbacks(void) |
| 664 | { |
| 665 | rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data)); |
| 666 | } |
| 667 | |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 668 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
| 669 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 670 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 671 | * Queue a preemptible-RCU callback for invocation after a grace period. |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 672 | */ |
| 673 | void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) |
| 674 | { |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 675 | __call_rcu(head, func, &rcu_preempt_state, -1, 0); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 676 | } |
| 677 | EXPORT_SYMBOL_GPL(call_rcu); |
| 678 | |
Paul E. McKenney | 486e259 | 2012-01-06 14:11:30 -0800 | [diff] [blame] | 679 | /* |
| 680 | * Queue an RCU callback for lazy invocation after a grace period. |
| 681 | * This will likely be later named something like "call_rcu_lazy()", |
| 682 | * but this change will require some way of tagging the lazy RCU |
| 683 | * callbacks in the list of pending callbacks. Until then, this |
| 684 | * function may only be called from __kfree_rcu(). |
| 685 | */ |
| 686 | void kfree_call_rcu(struct rcu_head *head, |
| 687 | void (*func)(struct rcu_head *rcu)) |
| 688 | { |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 689 | __call_rcu(head, func, &rcu_preempt_state, -1, 1); |
Paul E. McKenney | 486e259 | 2012-01-06 14:11:30 -0800 | [diff] [blame] | 690 | } |
| 691 | EXPORT_SYMBOL_GPL(kfree_call_rcu); |
| 692 | |
Paul E. McKenney | 6ebb237 | 2009-11-22 08:53:50 -0800 | [diff] [blame] | 693 | /** |
| 694 | * synchronize_rcu - wait until a grace period has elapsed. |
| 695 | * |
| 696 | * Control will return to the caller some time after a full grace |
| 697 | * period has elapsed, in other words after all currently executing RCU |
Paul E. McKenney | 77d8485 | 2010-07-08 17:38:59 -0700 | [diff] [blame] | 698 | * read-side critical sections have completed. Note, however, that |
| 699 | * upon return from synchronize_rcu(), the caller might well be executing |
| 700 | * concurrently with new RCU read-side critical sections that began while |
| 701 | * synchronize_rcu() was waiting. RCU read-side critical sections are |
| 702 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. |
Paul E. McKenney | f0a0e6f | 2012-10-23 13:47:01 -0700 | [diff] [blame] | 703 | * |
| 704 | * See the description of synchronize_sched() for more detailed information |
| 705 | * on memory ordering guarantees. |
Paul E. McKenney | 6ebb237 | 2009-11-22 08:53:50 -0800 | [diff] [blame] | 706 | */ |
| 707 | void synchronize_rcu(void) |
| 708 | { |
Paul E. McKenney | fe15d70 | 2012-01-04 13:30:33 -0800 | [diff] [blame] | 709 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
| 710 | !lock_is_held(&rcu_lock_map) && |
| 711 | !lock_is_held(&rcu_sched_lock_map), |
| 712 | "Illegal synchronize_rcu() in RCU read-side critical section"); |
Paul E. McKenney | 6ebb237 | 2009-11-22 08:53:50 -0800 | [diff] [blame] | 713 | if (!rcu_scheduler_active) |
| 714 | return; |
Antti P Miettinen | 3705b88 | 2012-10-05 09:59:15 +0300 | [diff] [blame] | 715 | if (rcu_expedited) |
| 716 | synchronize_rcu_expedited(); |
| 717 | else |
| 718 | wait_rcu_gp(call_rcu); |
Paul E. McKenney | 6ebb237 | 2009-11-22 08:53:50 -0800 | [diff] [blame] | 719 | } |
| 720 | EXPORT_SYMBOL_GPL(synchronize_rcu); |
| 721 | |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 722 | static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq); |
Paul E. McKenney | bcfa57c | 2012-07-23 16:03:51 -0700 | [diff] [blame] | 723 | static unsigned long sync_rcu_preempt_exp_count; |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 724 | static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex); |
| 725 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 726 | /* |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 727 | * Return non-zero if there are any tasks in RCU read-side critical |
| 728 | * sections blocking the current preemptible-RCU expedited grace period. |
| 729 | * If there is no preemptible-RCU expedited grace period currently in |
| 730 | * progress, returns zero unconditionally. |
| 731 | */ |
| 732 | static int rcu_preempted_readers_exp(struct rcu_node *rnp) |
| 733 | { |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 734 | return rnp->exp_tasks != NULL; |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 735 | } |
| 736 | |
| 737 | /* |
| 738 | * return non-zero if there is no RCU expedited grace period in progress |
| 739 | * for the specified rcu_node structure, in other words, if all CPUs and |
| 740 | * tasks covered by the specified rcu_node structure have done their bit |
| 741 | * for the current expedited grace period. Works only for preemptible |
| 742 | * RCU -- other RCU implementation use other means. |
| 743 | * |
| 744 | * Caller must hold sync_rcu_preempt_exp_mutex. |
| 745 | */ |
| 746 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) |
| 747 | { |
| 748 | return !rcu_preempted_readers_exp(rnp) && |
| 749 | ACCESS_ONCE(rnp->expmask) == 0; |
| 750 | } |
| 751 | |
| 752 | /* |
| 753 | * Report the exit from RCU read-side critical section for the last task |
| 754 | * that queued itself during or before the current expedited preemptible-RCU |
| 755 | * grace period. This event is reported either to the rcu_node structure on |
| 756 | * which the task was queued or to one of that rcu_node structure's ancestors, |
| 757 | * recursively up the tree. (Calm down, calm down, we do the recursion |
| 758 | * iteratively!) |
| 759 | * |
Thomas Gleixner | b40d293 | 2011-10-22 07:12:34 -0700 | [diff] [blame] | 760 | * Most callers will set the "wake" flag, but the task initiating the |
| 761 | * expedited grace period need not wake itself. |
| 762 | * |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 763 | * Caller must hold sync_rcu_preempt_exp_mutex. |
| 764 | */ |
Thomas Gleixner | b40d293 | 2011-10-22 07:12:34 -0700 | [diff] [blame] | 765 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
| 766 | bool wake) |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 767 | { |
| 768 | unsigned long flags; |
| 769 | unsigned long mask; |
| 770 | |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 771 | raw_spin_lock_irqsave(&rnp->lock, flags); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 772 | for (;;) { |
Paul E. McKenney | 131906b | 2011-07-17 02:05:49 -0700 | [diff] [blame] | 773 | if (!sync_rcu_preempt_exp_done(rnp)) { |
| 774 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 775 | break; |
Paul E. McKenney | 131906b | 2011-07-17 02:05:49 -0700 | [diff] [blame] | 776 | } |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 777 | if (rnp->parent == NULL) { |
Paul E. McKenney | 131906b | 2011-07-17 02:05:49 -0700 | [diff] [blame] | 778 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Thomas Gleixner | b40d293 | 2011-10-22 07:12:34 -0700 | [diff] [blame] | 779 | if (wake) |
| 780 | wake_up(&sync_rcu_preempt_exp_wq); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 781 | break; |
| 782 | } |
| 783 | mask = rnp->grpmask; |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 784 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 785 | rnp = rnp->parent; |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 786 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 787 | rnp->expmask &= ~mask; |
| 788 | } |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 789 | } |
| 790 | |
| 791 | /* |
| 792 | * Snapshot the tasks blocking the newly started preemptible-RCU expedited |
| 793 | * grace period for the specified rcu_node structure. If there are no such |
| 794 | * tasks, report it up the rcu_node hierarchy. |
| 795 | * |
Paul E. McKenney | 7b2e601 | 2012-10-08 10:54:03 -0700 | [diff] [blame] | 796 | * Caller must hold sync_rcu_preempt_exp_mutex and must exclude |
| 797 | * CPU hotplug operations. |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 798 | */ |
| 799 | static void |
| 800 | sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp) |
| 801 | { |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 802 | unsigned long flags; |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 803 | int must_wait = 0; |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 804 | |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 805 | raw_spin_lock_irqsave(&rnp->lock, flags); |
Paul E. McKenney | c701d5d | 2012-06-28 08:08:25 -0700 | [diff] [blame] | 806 | if (list_empty(&rnp->blkd_tasks)) { |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 807 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | c701d5d | 2012-06-28 08:08:25 -0700 | [diff] [blame] | 808 | } else { |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 809 | rnp->exp_tasks = rnp->blkd_tasks.next; |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 810 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 811 | must_wait = 1; |
| 812 | } |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 813 | if (!must_wait) |
Thomas Gleixner | b40d293 | 2011-10-22 07:12:34 -0700 | [diff] [blame] | 814 | rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */ |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 815 | } |
| 816 | |
Paul E. McKenney | 236fefa | 2012-01-31 14:00:41 -0800 | [diff] [blame] | 817 | /** |
| 818 | * synchronize_rcu_expedited - Brute-force RCU grace period |
| 819 | * |
| 820 | * Wait for an RCU-preempt grace period, but expedite it. The basic |
| 821 | * idea is to invoke synchronize_sched_expedited() to push all the tasks to |
| 822 | * the ->blkd_tasks lists and wait for this list to drain. This consumes |
| 823 | * significant time on all CPUs and is unfriendly to real-time workloads, |
| 824 | * so is thus not recommended for any sort of common-case code. |
| 825 | * In fact, if you are using synchronize_rcu_expedited() in a loop, |
| 826 | * please restructure your code to batch your updates, and then Use a |
| 827 | * single synchronize_rcu() instead. |
| 828 | * |
| 829 | * Note that it is illegal to call this function while holding any lock |
| 830 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal |
| 831 | * to call this function from a CPU-hotplug notifier. Failing to observe |
| 832 | * these restriction will result in deadlock. |
Paul E. McKenney | 019129d5 | 2009-10-14 10:15:56 -0700 | [diff] [blame] | 833 | */ |
| 834 | void synchronize_rcu_expedited(void) |
| 835 | { |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 836 | unsigned long flags; |
| 837 | struct rcu_node *rnp; |
| 838 | struct rcu_state *rsp = &rcu_preempt_state; |
Paul E. McKenney | bcfa57c | 2012-07-23 16:03:51 -0700 | [diff] [blame] | 839 | unsigned long snap; |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 840 | int trycount = 0; |
| 841 | |
| 842 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ |
| 843 | snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1; |
| 844 | smp_mb(); /* Above access cannot bleed into critical section. */ |
| 845 | |
| 846 | /* |
Paul E. McKenney | 1943c89 | 2012-07-30 17:19:25 -0700 | [diff] [blame] | 847 | * Block CPU-hotplug operations. This means that any CPU-hotplug |
| 848 | * operation that finds an rcu_node structure with tasks in the |
| 849 | * process of being boosted will know that all tasks blocking |
| 850 | * this expedited grace period will already be in the process of |
| 851 | * being boosted. This simplifies the process of moving tasks |
| 852 | * from leaf to root rcu_node structures. |
| 853 | */ |
| 854 | get_online_cpus(); |
| 855 | |
| 856 | /* |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 857 | * Acquire lock, falling back to synchronize_rcu() if too many |
| 858 | * lock-acquisition failures. Of course, if someone does the |
| 859 | * expedited grace period for us, just leave. |
| 860 | */ |
| 861 | while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) { |
Paul E. McKenney | 1943c89 | 2012-07-30 17:19:25 -0700 | [diff] [blame] | 862 | if (ULONG_CMP_LT(snap, |
| 863 | ACCESS_ONCE(sync_rcu_preempt_exp_count))) { |
| 864 | put_online_cpus(); |
| 865 | goto mb_ret; /* Others did our work for us. */ |
| 866 | } |
Paul E. McKenney | c701d5d | 2012-06-28 08:08:25 -0700 | [diff] [blame] | 867 | if (trycount++ < 10) { |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 868 | udelay(trycount * num_online_cpus()); |
Paul E. McKenney | c701d5d | 2012-06-28 08:08:25 -0700 | [diff] [blame] | 869 | } else { |
Paul E. McKenney | 1943c89 | 2012-07-30 17:19:25 -0700 | [diff] [blame] | 870 | put_online_cpus(); |
Antti P Miettinen | 3705b88 | 2012-10-05 09:59:15 +0300 | [diff] [blame] | 871 | wait_rcu_gp(call_rcu); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 872 | return; |
| 873 | } |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 874 | } |
Paul E. McKenney | 1943c89 | 2012-07-30 17:19:25 -0700 | [diff] [blame] | 875 | if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) { |
| 876 | put_online_cpus(); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 877 | goto unlock_mb_ret; /* Others did our work for us. */ |
Paul E. McKenney | 1943c89 | 2012-07-30 17:19:25 -0700 | [diff] [blame] | 878 | } |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 879 | |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 880 | /* force all RCU readers onto ->blkd_tasks lists. */ |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 881 | synchronize_sched_expedited(); |
| 882 | |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 883 | /* Initialize ->expmask for all non-leaf rcu_node structures. */ |
| 884 | rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) { |
Paul E. McKenney | 1943c89 | 2012-07-30 17:19:25 -0700 | [diff] [blame] | 885 | raw_spin_lock_irqsave(&rnp->lock, flags); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 886 | rnp->expmask = rnp->qsmaskinit; |
Paul E. McKenney | 1943c89 | 2012-07-30 17:19:25 -0700 | [diff] [blame] | 887 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 888 | } |
| 889 | |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 890 | /* Snapshot current state of ->blkd_tasks lists. */ |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 891 | rcu_for_each_leaf_node(rsp, rnp) |
| 892 | sync_rcu_preempt_exp_init(rsp, rnp); |
| 893 | if (NUM_RCU_NODES > 1) |
| 894 | sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp)); |
| 895 | |
Paul E. McKenney | 1943c89 | 2012-07-30 17:19:25 -0700 | [diff] [blame] | 896 | put_online_cpus(); |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 897 | |
Paul E. McKenney | 12f5f52 | 2010-11-29 21:56:39 -0800 | [diff] [blame] | 898 | /* Wait for snapshotted ->blkd_tasks lists to drain. */ |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 899 | rnp = rcu_get_root(rsp); |
| 900 | wait_event(sync_rcu_preempt_exp_wq, |
| 901 | sync_rcu_preempt_exp_done(rnp)); |
| 902 | |
| 903 | /* Clean up and exit. */ |
| 904 | smp_mb(); /* ensure expedited GP seen before counter increment. */ |
| 905 | ACCESS_ONCE(sync_rcu_preempt_exp_count)++; |
| 906 | unlock_mb_ret: |
| 907 | mutex_unlock(&sync_rcu_preempt_exp_mutex); |
| 908 | mb_ret: |
| 909 | smp_mb(); /* ensure subsequent action seen after grace period. */ |
Paul E. McKenney | 019129d5 | 2009-10-14 10:15:56 -0700 | [diff] [blame] | 910 | } |
| 911 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); |
| 912 | |
Paul E. McKenney | e74f4c4 | 2009-10-06 21:48:17 -0700 | [diff] [blame] | 913 | /** |
| 914 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. |
Paul E. McKenney | f0a0e6f | 2012-10-23 13:47:01 -0700 | [diff] [blame] | 915 | * |
| 916 | * Note that this primitive does not necessarily wait for an RCU grace period |
| 917 | * to complete. For example, if there are no RCU callbacks queued anywhere |
| 918 | * in the system, then rcu_barrier() is within its rights to return |
| 919 | * immediately, without waiting for anything, much less an RCU grace period. |
Paul E. McKenney | e74f4c4 | 2009-10-06 21:48:17 -0700 | [diff] [blame] | 920 | */ |
| 921 | void rcu_barrier(void) |
| 922 | { |
Paul E. McKenney | 037b64e | 2012-05-28 23:26:01 -0700 | [diff] [blame] | 923 | _rcu_barrier(&rcu_preempt_state); |
Paul E. McKenney | e74f4c4 | 2009-10-06 21:48:17 -0700 | [diff] [blame] | 924 | } |
| 925 | EXPORT_SYMBOL_GPL(rcu_barrier); |
| 926 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 927 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 928 | * Initialize preemptible RCU's state structures. |
Paul E. McKenney | 1eba8f8 | 2009-09-23 09:50:42 -0700 | [diff] [blame] | 929 | */ |
| 930 | static void __init __rcu_init_preempt(void) |
| 931 | { |
Lai Jiangshan | 394f99a | 2010-06-28 16:25:04 +0800 | [diff] [blame] | 932 | rcu_init_one(&rcu_preempt_state, &rcu_preempt_data); |
Paul E. McKenney | 1eba8f8 | 2009-09-23 09:50:42 -0700 | [diff] [blame] | 933 | } |
| 934 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 935 | #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */ |
| 936 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 937 | static struct rcu_state *rcu_state = &rcu_sched_state; |
| 938 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 939 | /* |
| 940 | * Tell them what RCU they are running. |
| 941 | */ |
Paul E. McKenney | 0e0fc1c | 2009-11-11 11:28:06 -0800 | [diff] [blame] | 942 | static void __init rcu_bootup_announce(void) |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 943 | { |
| 944 | printk(KERN_INFO "Hierarchical RCU implementation.\n"); |
Paul E. McKenney | 26845c2 | 2010-04-13 14:19:23 -0700 | [diff] [blame] | 945 | rcu_bootup_announce_oddness(); |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 946 | } |
| 947 | |
| 948 | /* |
| 949 | * Return the number of RCU batches processed thus far for debug & stats. |
| 950 | */ |
| 951 | long rcu_batches_completed(void) |
| 952 | { |
| 953 | return rcu_batches_completed_sched(); |
| 954 | } |
| 955 | EXPORT_SYMBOL_GPL(rcu_batches_completed); |
| 956 | |
| 957 | /* |
Paul E. McKenney | bf66f18 | 2010-01-04 15:09:10 -0800 | [diff] [blame] | 958 | * Force a quiescent state for RCU, which, because there is no preemptible |
| 959 | * RCU, becomes the same as rcu-sched. |
| 960 | */ |
| 961 | void rcu_force_quiescent_state(void) |
| 962 | { |
| 963 | rcu_sched_force_quiescent_state(); |
| 964 | } |
| 965 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); |
| 966 | |
| 967 | /* |
Paul E. McKenney | cba6d0d | 2012-07-02 07:08:42 -0700 | [diff] [blame] | 968 | * Because preemptible RCU does not exist, we never have to check for |
| 969 | * CPUs being in quiescent states. |
| 970 | */ |
| 971 | static void rcu_preempt_note_context_switch(int cpu) |
| 972 | { |
| 973 | } |
| 974 | |
| 975 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 976 | * Because preemptible RCU does not exist, there are never any preempted |
Paul E. McKenney | fc2219d | 2009-09-23 09:50:41 -0700 | [diff] [blame] | 977 | * RCU readers. |
| 978 | */ |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 979 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
Paul E. McKenney | fc2219d | 2009-09-23 09:50:41 -0700 | [diff] [blame] | 980 | { |
| 981 | return 0; |
| 982 | } |
| 983 | |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 984 | #ifdef CONFIG_HOTPLUG_CPU |
| 985 | |
| 986 | /* Because preemptible RCU does not exist, no quieting of tasks. */ |
Paul E. McKenney | d3f6bad | 2009-12-02 12:10:13 -0800 | [diff] [blame] | 987 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 988 | { |
Paul E. McKenney | 1304afb | 2010-02-22 17:05:02 -0800 | [diff] [blame] | 989 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | b668c9c | 2009-11-22 08:53:48 -0800 | [diff] [blame] | 990 | } |
| 991 | |
| 992 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 993 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 994 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 995 | * Because preemptible RCU does not exist, we never have to check for |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 996 | * tasks blocked within RCU read-side critical sections. |
| 997 | */ |
Paul E. McKenney | 1ed509a | 2010-02-22 17:05:05 -0800 | [diff] [blame] | 998 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) |
| 999 | { |
| 1000 | } |
| 1001 | |
| 1002 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 1003 | * Because preemptible RCU does not exist, we never have to check for |
Paul E. McKenney | 1ed509a | 2010-02-22 17:05:05 -0800 | [diff] [blame] | 1004 | * tasks blocked within RCU read-side critical sections. |
| 1005 | */ |
Paul E. McKenney | 9bc8b55 | 2011-08-13 13:31:47 -0700 | [diff] [blame] | 1006 | static int rcu_print_task_stall(struct rcu_node *rnp) |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 1007 | { |
Paul E. McKenney | 9bc8b55 | 2011-08-13 13:31:47 -0700 | [diff] [blame] | 1008 | return 0; |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 1009 | } |
| 1010 | |
Paul E. McKenney | 53d84e0 | 2010-08-10 14:28:53 -0700 | [diff] [blame] | 1011 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 1012 | * Because there is no preemptible RCU, there can be no readers blocked, |
Paul E. McKenney | 49e2912 | 2009-09-18 09:50:19 -0700 | [diff] [blame] | 1013 | * so there is no need to check for blocked tasks. So check only for |
| 1014 | * bogus qsmask values. |
Paul E. McKenney | b0e165c | 2009-09-13 09:15:09 -0700 | [diff] [blame] | 1015 | */ |
| 1016 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) |
| 1017 | { |
Paul E. McKenney | 49e2912 | 2009-09-18 09:50:19 -0700 | [diff] [blame] | 1018 | WARN_ON_ONCE(rnp->qsmask); |
Paul E. McKenney | b0e165c | 2009-09-13 09:15:09 -0700 | [diff] [blame] | 1019 | } |
| 1020 | |
Paul E. McKenney | 33f7614 | 2009-08-24 09:42:01 -0700 | [diff] [blame] | 1021 | #ifdef CONFIG_HOTPLUG_CPU |
| 1022 | |
| 1023 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 1024 | * Because preemptible RCU does not exist, it never needs to migrate |
Paul E. McKenney | 237c80c | 2009-10-15 09:26:14 -0700 | [diff] [blame] | 1025 | * tasks that were blocked within RCU read-side critical sections, and |
| 1026 | * such non-existent tasks cannot possibly have been blocking the current |
| 1027 | * grace period. |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 1028 | */ |
Paul E. McKenney | 237c80c | 2009-10-15 09:26:14 -0700 | [diff] [blame] | 1029 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
| 1030 | struct rcu_node *rnp, |
| 1031 | struct rcu_data *rdp) |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 1032 | { |
Paul E. McKenney | 237c80c | 2009-10-15 09:26:14 -0700 | [diff] [blame] | 1033 | return 0; |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 1034 | } |
| 1035 | |
Paul E. McKenney | e560140 | 2012-01-07 11:03:57 -0800 | [diff] [blame] | 1036 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 1037 | |
Paul E. McKenney | dd5d19b | 2009-08-27 14:58:16 -0700 | [diff] [blame] | 1038 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 1039 | * Because preemptible RCU does not exist, it never has any callbacks |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 1040 | * to check. |
| 1041 | */ |
Paul E. McKenney | 1eba8f8 | 2009-09-23 09:50:42 -0700 | [diff] [blame] | 1042 | static void rcu_preempt_check_callbacks(int cpu) |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 1043 | { |
| 1044 | } |
| 1045 | |
| 1046 | /* |
Paul E. McKenney | 486e259 | 2012-01-06 14:11:30 -0800 | [diff] [blame] | 1047 | * Queue an RCU callback for lazy invocation after a grace period. |
| 1048 | * This will likely be later named something like "call_rcu_lazy()", |
| 1049 | * but this change will require some way of tagging the lazy RCU |
| 1050 | * callbacks in the list of pending callbacks. Until then, this |
| 1051 | * function may only be called from __kfree_rcu(). |
| 1052 | * |
| 1053 | * Because there is no preemptible RCU, we use RCU-sched instead. |
| 1054 | */ |
| 1055 | void kfree_call_rcu(struct rcu_head *head, |
| 1056 | void (*func)(struct rcu_head *rcu)) |
| 1057 | { |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 1058 | __call_rcu(head, func, &rcu_sched_state, -1, 1); |
Paul E. McKenney | 486e259 | 2012-01-06 14:11:30 -0800 | [diff] [blame] | 1059 | } |
| 1060 | EXPORT_SYMBOL_GPL(kfree_call_rcu); |
| 1061 | |
| 1062 | /* |
Paul E. McKenney | 019129d5 | 2009-10-14 10:15:56 -0700 | [diff] [blame] | 1063 | * Wait for an rcu-preempt grace period, but make it happen quickly. |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 1064 | * But because preemptible RCU does not exist, map to rcu-sched. |
Paul E. McKenney | 019129d5 | 2009-10-14 10:15:56 -0700 | [diff] [blame] | 1065 | */ |
| 1066 | void synchronize_rcu_expedited(void) |
| 1067 | { |
| 1068 | synchronize_sched_expedited(); |
| 1069 | } |
| 1070 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); |
| 1071 | |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 1072 | #ifdef CONFIG_HOTPLUG_CPU |
| 1073 | |
| 1074 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 1075 | * Because preemptible RCU does not exist, there is never any need to |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 1076 | * report on tasks preempted in RCU read-side critical sections during |
| 1077 | * expedited RCU grace periods. |
| 1078 | */ |
Thomas Gleixner | b40d293 | 2011-10-22 07:12:34 -0700 | [diff] [blame] | 1079 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
| 1080 | bool wake) |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 1081 | { |
Paul E. McKenney | d9a3da0 | 2009-12-02 12:10:15 -0800 | [diff] [blame] | 1082 | } |
| 1083 | |
| 1084 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
| 1085 | |
Paul E. McKenney | 019129d5 | 2009-10-14 10:15:56 -0700 | [diff] [blame] | 1086 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 1087 | * Because preemptible RCU does not exist, rcu_barrier() is just |
Paul E. McKenney | e74f4c4 | 2009-10-06 21:48:17 -0700 | [diff] [blame] | 1088 | * another name for rcu_barrier_sched(). |
| 1089 | */ |
| 1090 | void rcu_barrier(void) |
| 1091 | { |
| 1092 | rcu_barrier_sched(); |
| 1093 | } |
| 1094 | EXPORT_SYMBOL_GPL(rcu_barrier); |
| 1095 | |
| 1096 | /* |
Paul E. McKenney | 6cc6879 | 2011-03-02 13:15:15 -0800 | [diff] [blame] | 1097 | * Because preemptible RCU does not exist, it need not be initialized. |
Paul E. McKenney | 1eba8f8 | 2009-09-23 09:50:42 -0700 | [diff] [blame] | 1098 | */ |
| 1099 | static void __init __rcu_init_preempt(void) |
| 1100 | { |
| 1101 | } |
| 1102 | |
Paul E. McKenney | f41d911 | 2009-08-22 13:56:52 -0700 | [diff] [blame] | 1103 | #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */ |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1104 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1105 | #ifdef CONFIG_RCU_BOOST |
| 1106 | |
| 1107 | #include "rtmutex_common.h" |
| 1108 | |
Paul E. McKenney | 0ea1f2e | 2011-02-22 13:42:43 -0800 | [diff] [blame] | 1109 | #ifdef CONFIG_RCU_TRACE |
| 1110 | |
| 1111 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) |
| 1112 | { |
| 1113 | if (list_empty(&rnp->blkd_tasks)) |
| 1114 | rnp->n_balk_blkd_tasks++; |
| 1115 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) |
| 1116 | rnp->n_balk_exp_gp_tasks++; |
| 1117 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) |
| 1118 | rnp->n_balk_boost_tasks++; |
| 1119 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) |
| 1120 | rnp->n_balk_notblocked++; |
| 1121 | else if (rnp->gp_tasks != NULL && |
Paul E. McKenney | a9f4793 | 2011-05-02 03:46:10 -0700 | [diff] [blame] | 1122 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
Paul E. McKenney | 0ea1f2e | 2011-02-22 13:42:43 -0800 | [diff] [blame] | 1123 | rnp->n_balk_notyet++; |
| 1124 | else |
| 1125 | rnp->n_balk_nos++; |
| 1126 | } |
| 1127 | |
| 1128 | #else /* #ifdef CONFIG_RCU_TRACE */ |
| 1129 | |
| 1130 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) |
| 1131 | { |
| 1132 | } |
| 1133 | |
| 1134 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ |
| 1135 | |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1136 | static void rcu_wake_cond(struct task_struct *t, int status) |
| 1137 | { |
| 1138 | /* |
| 1139 | * If the thread is yielding, only wake it when this |
| 1140 | * is invoked from idle |
| 1141 | */ |
| 1142 | if (status != RCU_KTHREAD_YIELDING || is_idle_task(current)) |
| 1143 | wake_up_process(t); |
| 1144 | } |
| 1145 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1146 | /* |
| 1147 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks |
| 1148 | * or ->boost_tasks, advancing the pointer to the next task in the |
| 1149 | * ->blkd_tasks list. |
| 1150 | * |
| 1151 | * Note that irqs must be enabled: boosting the task can block. |
| 1152 | * Returns 1 if there are more tasks needing to be boosted. |
| 1153 | */ |
| 1154 | static int rcu_boost(struct rcu_node *rnp) |
| 1155 | { |
| 1156 | unsigned long flags; |
| 1157 | struct rt_mutex mtx; |
| 1158 | struct task_struct *t; |
| 1159 | struct list_head *tb; |
| 1160 | |
| 1161 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) |
| 1162 | return 0; /* Nothing left to boost. */ |
| 1163 | |
| 1164 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 1165 | |
| 1166 | /* |
| 1167 | * Recheck under the lock: all tasks in need of boosting |
| 1168 | * might exit their RCU read-side critical sections on their own. |
| 1169 | */ |
| 1170 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { |
| 1171 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1172 | return 0; |
| 1173 | } |
| 1174 | |
| 1175 | /* |
| 1176 | * Preferentially boost tasks blocking expedited grace periods. |
| 1177 | * This cannot starve the normal grace periods because a second |
| 1178 | * expedited grace period must boost all blocked tasks, including |
| 1179 | * those blocking the pre-existing normal grace period. |
| 1180 | */ |
Paul E. McKenney | 0ea1f2e | 2011-02-22 13:42:43 -0800 | [diff] [blame] | 1181 | if (rnp->exp_tasks != NULL) { |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1182 | tb = rnp->exp_tasks; |
Paul E. McKenney | 0ea1f2e | 2011-02-22 13:42:43 -0800 | [diff] [blame] | 1183 | rnp->n_exp_boosts++; |
| 1184 | } else { |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1185 | tb = rnp->boost_tasks; |
Paul E. McKenney | 0ea1f2e | 2011-02-22 13:42:43 -0800 | [diff] [blame] | 1186 | rnp->n_normal_boosts++; |
| 1187 | } |
| 1188 | rnp->n_tasks_boosted++; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1189 | |
| 1190 | /* |
| 1191 | * We boost task t by manufacturing an rt_mutex that appears to |
| 1192 | * be held by task t. We leave a pointer to that rt_mutex where |
| 1193 | * task t can find it, and task t will release the mutex when it |
| 1194 | * exits its outermost RCU read-side critical section. Then |
| 1195 | * simply acquiring this artificial rt_mutex will boost task |
| 1196 | * t's priority. (Thanks to tglx for suggesting this approach!) |
| 1197 | * |
| 1198 | * Note that task t must acquire rnp->lock to remove itself from |
| 1199 | * the ->blkd_tasks list, which it will do from exit() if from |
| 1200 | * nowhere else. We therefore are guaranteed that task t will |
| 1201 | * stay around at least until we drop rnp->lock. Note that |
| 1202 | * rnp->lock also resolves races between our priority boosting |
| 1203 | * and task t's exiting its outermost RCU read-side critical |
| 1204 | * section. |
| 1205 | */ |
| 1206 | t = container_of(tb, struct task_struct, rcu_node_entry); |
| 1207 | rt_mutex_init_proxy_locked(&mtx, t); |
| 1208 | t->rcu_boost_mutex = &mtx; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1209 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1210 | rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */ |
| 1211 | rt_mutex_unlock(&mtx); /* Keep lockdep happy. */ |
| 1212 | |
Paul E. McKenney | 4f89b33 | 2011-12-09 14:43:47 -0800 | [diff] [blame] | 1213 | return ACCESS_ONCE(rnp->exp_tasks) != NULL || |
| 1214 | ACCESS_ONCE(rnp->boost_tasks) != NULL; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1215 | } |
| 1216 | |
| 1217 | /* |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1218 | * Priority-boosting kthread. One per leaf rcu_node and one for the |
| 1219 | * root rcu_node. |
| 1220 | */ |
| 1221 | static int rcu_boost_kthread(void *arg) |
| 1222 | { |
| 1223 | struct rcu_node *rnp = (struct rcu_node *)arg; |
| 1224 | int spincnt = 0; |
| 1225 | int more2boost; |
| 1226 | |
Paul E. McKenney | 385680a | 2011-06-21 22:43:26 -0700 | [diff] [blame] | 1227 | trace_rcu_utilization("Start boost kthread@init"); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1228 | for (;;) { |
Paul E. McKenney | d71df90 | 2011-03-29 17:48:28 -0700 | [diff] [blame] | 1229 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
Paul E. McKenney | 385680a | 2011-06-21 22:43:26 -0700 | [diff] [blame] | 1230 | trace_rcu_utilization("End boost kthread@rcu_wait"); |
Peter Zijlstra | 08bca60 | 2011-05-20 16:06:29 -0700 | [diff] [blame] | 1231 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
Paul E. McKenney | 385680a | 2011-06-21 22:43:26 -0700 | [diff] [blame] | 1232 | trace_rcu_utilization("Start boost kthread@rcu_wait"); |
Paul E. McKenney | d71df90 | 2011-03-29 17:48:28 -0700 | [diff] [blame] | 1233 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1234 | more2boost = rcu_boost(rnp); |
| 1235 | if (more2boost) |
| 1236 | spincnt++; |
| 1237 | else |
| 1238 | spincnt = 0; |
| 1239 | if (spincnt > 10) { |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1240 | rnp->boost_kthread_status = RCU_KTHREAD_YIELDING; |
Paul E. McKenney | 385680a | 2011-06-21 22:43:26 -0700 | [diff] [blame] | 1241 | trace_rcu_utilization("End boost kthread@rcu_yield"); |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1242 | schedule_timeout_interruptible(2); |
Paul E. McKenney | 385680a | 2011-06-21 22:43:26 -0700 | [diff] [blame] | 1243 | trace_rcu_utilization("Start boost kthread@rcu_yield"); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1244 | spincnt = 0; |
| 1245 | } |
| 1246 | } |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 1247 | /* NOTREACHED */ |
Paul E. McKenney | 385680a | 2011-06-21 22:43:26 -0700 | [diff] [blame] | 1248 | trace_rcu_utilization("End boost kthread@notreached"); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1249 | return 0; |
| 1250 | } |
| 1251 | |
| 1252 | /* |
| 1253 | * Check to see if it is time to start boosting RCU readers that are |
| 1254 | * blocking the current grace period, and, if so, tell the per-rcu_node |
| 1255 | * kthread to start boosting them. If there is an expedited grace |
| 1256 | * period in progress, it is always time to boost. |
| 1257 | * |
Paul E. McKenney | b065a85 | 2012-08-01 15:57:54 -0700 | [diff] [blame] | 1258 | * The caller must hold rnp->lock, which this function releases. |
| 1259 | * The ->boost_kthread_task is immortal, so we don't need to worry |
| 1260 | * about it going away. |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1261 | */ |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 1262 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1263 | { |
| 1264 | struct task_struct *t; |
| 1265 | |
Paul E. McKenney | 0ea1f2e | 2011-02-22 13:42:43 -0800 | [diff] [blame] | 1266 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
| 1267 | rnp->n_balk_exp_gp_tasks++; |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 1268 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1269 | return; |
Paul E. McKenney | 0ea1f2e | 2011-02-22 13:42:43 -0800 | [diff] [blame] | 1270 | } |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1271 | if (rnp->exp_tasks != NULL || |
| 1272 | (rnp->gp_tasks != NULL && |
| 1273 | rnp->boost_tasks == NULL && |
| 1274 | rnp->qsmask == 0 && |
| 1275 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { |
| 1276 | if (rnp->exp_tasks == NULL) |
| 1277 | rnp->boost_tasks = rnp->gp_tasks; |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 1278 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1279 | t = rnp->boost_kthread_task; |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1280 | if (t) |
| 1281 | rcu_wake_cond(t, rnp->boost_kthread_status); |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 1282 | } else { |
Paul E. McKenney | 0ea1f2e | 2011-02-22 13:42:43 -0800 | [diff] [blame] | 1283 | rcu_initiate_boost_trace(rnp); |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 1284 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 1285 | } |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1286 | } |
| 1287 | |
Paul E. McKenney | 0f962a5 | 2011-04-14 12:13:53 -0700 | [diff] [blame] | 1288 | /* |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 1289 | * Wake up the per-CPU kthread to invoke RCU callbacks. |
| 1290 | */ |
| 1291 | static void invoke_rcu_callbacks_kthread(void) |
| 1292 | { |
| 1293 | unsigned long flags; |
| 1294 | |
| 1295 | local_irq_save(flags); |
| 1296 | __this_cpu_write(rcu_cpu_has_work, 1); |
Shaohua Li | 1eb5212 | 2011-06-16 16:02:54 -0700 | [diff] [blame] | 1297 | if (__this_cpu_read(rcu_cpu_kthread_task) != NULL && |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1298 | current != __this_cpu_read(rcu_cpu_kthread_task)) { |
| 1299 | rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task), |
| 1300 | __this_cpu_read(rcu_cpu_kthread_status)); |
| 1301 | } |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 1302 | local_irq_restore(flags); |
| 1303 | } |
| 1304 | |
| 1305 | /* |
Paul E. McKenney | dff1672 | 2011-11-29 15:57:13 -0800 | [diff] [blame] | 1306 | * Is the current CPU running the RCU-callbacks kthread? |
| 1307 | * Caller must have preemption disabled. |
| 1308 | */ |
| 1309 | static bool rcu_is_callbacks_kthread(void) |
| 1310 | { |
| 1311 | return __get_cpu_var(rcu_cpu_kthread_task) == current; |
| 1312 | } |
| 1313 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1314 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) |
| 1315 | |
| 1316 | /* |
| 1317 | * Do priority-boost accounting for the start of a new grace period. |
| 1318 | */ |
| 1319 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
| 1320 | { |
| 1321 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; |
| 1322 | } |
| 1323 | |
| 1324 | /* |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1325 | * Create an RCU-boost kthread for the specified node if one does not |
| 1326 | * already exist. We only create this kthread for preemptible RCU. |
| 1327 | * Returns zero if all is well, a negated errno otherwise. |
| 1328 | */ |
| 1329 | static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp, |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1330 | struct rcu_node *rnp) |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1331 | { |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1332 | int rnp_index = rnp - &rsp->node[0]; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1333 | unsigned long flags; |
| 1334 | struct sched_param sp; |
| 1335 | struct task_struct *t; |
| 1336 | |
| 1337 | if (&rcu_preempt_state != rsp) |
| 1338 | return 0; |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1339 | |
| 1340 | if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0) |
| 1341 | return 0; |
| 1342 | |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 1343 | rsp->boost = 1; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1344 | if (rnp->boost_kthread_task != NULL) |
| 1345 | return 0; |
| 1346 | t = kthread_create(rcu_boost_kthread, (void *)rnp, |
Mike Galbraith | 5b61b0b | 2011-08-19 11:39:11 -0700 | [diff] [blame] | 1347 | "rcub/%d", rnp_index); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1348 | if (IS_ERR(t)) |
| 1349 | return PTR_ERR(t); |
| 1350 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 1351 | rnp->boost_kthread_task = t; |
| 1352 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Mike Galbraith | 5b61b0b | 2011-08-19 11:39:11 -0700 | [diff] [blame] | 1353 | sp.sched_priority = RCU_BOOST_PRIO; |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1354 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); |
Paul E. McKenney | 9a43273 | 2011-05-30 20:38:55 -0700 | [diff] [blame] | 1355 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1356 | return 0; |
| 1357 | } |
| 1358 | |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1359 | static void rcu_kthread_do_work(void) |
| 1360 | { |
| 1361 | rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data)); |
| 1362 | rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data)); |
| 1363 | rcu_preempt_do_callbacks(); |
| 1364 | } |
| 1365 | |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1366 | static void rcu_cpu_kthread_setup(unsigned int cpu) |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1367 | { |
| 1368 | struct sched_param sp; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1369 | |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1370 | sp.sched_priority = RCU_KTHREAD_PRIO; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1371 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1372 | } |
| 1373 | |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1374 | static void rcu_cpu_kthread_park(unsigned int cpu) |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1375 | { |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1376 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; |
| 1377 | } |
| 1378 | |
| 1379 | static int rcu_cpu_kthread_should_run(unsigned int cpu) |
| 1380 | { |
| 1381 | return __get_cpu_var(rcu_cpu_has_work); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1382 | } |
| 1383 | |
| 1384 | /* |
| 1385 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the |
Paul E. McKenney | e0f2306 | 2011-06-21 01:29:39 -0700 | [diff] [blame] | 1386 | * RCU softirq used in flavors and configurations of RCU that do not |
| 1387 | * support RCU priority boosting. |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1388 | */ |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1389 | static void rcu_cpu_kthread(unsigned int cpu) |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1390 | { |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1391 | unsigned int *statusp = &__get_cpu_var(rcu_cpu_kthread_status); |
| 1392 | char work, *workp = &__get_cpu_var(rcu_cpu_has_work); |
| 1393 | int spincnt; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1394 | |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1395 | for (spincnt = 0; spincnt < 10; spincnt++) { |
Paul E. McKenney | 385680a | 2011-06-21 22:43:26 -0700 | [diff] [blame] | 1396 | trace_rcu_utilization("Start CPU kthread@rcu_wait"); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1397 | local_bh_disable(); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1398 | *statusp = RCU_KTHREAD_RUNNING; |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1399 | this_cpu_inc(rcu_cpu_kthread_loops); |
| 1400 | local_irq_disable(); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1401 | work = *workp; |
| 1402 | *workp = 0; |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1403 | local_irq_enable(); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1404 | if (work) |
| 1405 | rcu_kthread_do_work(); |
| 1406 | local_bh_enable(); |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1407 | if (*workp == 0) { |
| 1408 | trace_rcu_utilization("End CPU kthread@rcu_wait"); |
| 1409 | *statusp = RCU_KTHREAD_WAITING; |
| 1410 | return; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1411 | } |
| 1412 | } |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1413 | *statusp = RCU_KTHREAD_YIELDING; |
| 1414 | trace_rcu_utilization("Start CPU kthread@rcu_yield"); |
| 1415 | schedule_timeout_interruptible(2); |
| 1416 | trace_rcu_utilization("End CPU kthread@rcu_yield"); |
| 1417 | *statusp = RCU_KTHREAD_WAITING; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1418 | } |
| 1419 | |
| 1420 | /* |
| 1421 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are |
| 1422 | * served by the rcu_node in question. The CPU hotplug lock is still |
| 1423 | * held, so the value of rnp->qsmaskinit will be stable. |
| 1424 | * |
| 1425 | * We don't include outgoingcpu in the affinity set, use -1 if there is |
| 1426 | * no outgoing CPU. If there are no CPUs left in the affinity set, |
| 1427 | * this function allows the kthread to execute on any CPU. |
| 1428 | */ |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1429 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1430 | { |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1431 | struct task_struct *t = rnp->boost_kthread_task; |
| 1432 | unsigned long mask = rnp->qsmaskinit; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1433 | cpumask_var_t cm; |
| 1434 | int cpu; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1435 | |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1436 | if (!t) |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1437 | return; |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1438 | if (!zalloc_cpumask_var(&cm, GFP_KERNEL)) |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1439 | return; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1440 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) |
| 1441 | if ((mask & 0x1) && cpu != outgoingcpu) |
| 1442 | cpumask_set_cpu(cpu, cm); |
| 1443 | if (cpumask_weight(cm) == 0) { |
| 1444 | cpumask_setall(cm); |
| 1445 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) |
| 1446 | cpumask_clear_cpu(cpu, cm); |
| 1447 | WARN_ON_ONCE(cpumask_weight(cm) == 0); |
| 1448 | } |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1449 | set_cpus_allowed_ptr(t, cm); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1450 | free_cpumask_var(cm); |
| 1451 | } |
| 1452 | |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1453 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { |
| 1454 | .store = &rcu_cpu_kthread_task, |
| 1455 | .thread_should_run = rcu_cpu_kthread_should_run, |
| 1456 | .thread_fn = rcu_cpu_kthread, |
| 1457 | .thread_comm = "rcuc/%u", |
| 1458 | .setup = rcu_cpu_kthread_setup, |
| 1459 | .park = rcu_cpu_kthread_park, |
| 1460 | }; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1461 | |
| 1462 | /* |
| 1463 | * Spawn all kthreads -- called as soon as the scheduler is running. |
| 1464 | */ |
| 1465 | static int __init rcu_spawn_kthreads(void) |
| 1466 | { |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1467 | struct rcu_node *rnp; |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1468 | int cpu; |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1469 | |
Paul E. McKenney | b0d3041 | 2011-07-10 15:57:35 -0700 | [diff] [blame] | 1470 | rcu_scheduler_fully_active = 1; |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1471 | for_each_possible_cpu(cpu) |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1472 | per_cpu(rcu_cpu_has_work, cpu) = 0; |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1473 | BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec)); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1474 | rnp = rcu_get_root(rcu_state); |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1475 | (void)rcu_spawn_one_boost_kthread(rcu_state, rnp); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1476 | if (NUM_RCU_NODES > 1) { |
| 1477 | rcu_for_each_leaf_node(rcu_state, rnp) |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1478 | (void)rcu_spawn_one_boost_kthread(rcu_state, rnp); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1479 | } |
| 1480 | return 0; |
| 1481 | } |
| 1482 | early_initcall(rcu_spawn_kthreads); |
| 1483 | |
| 1484 | static void __cpuinit rcu_prepare_kthreads(int cpu) |
| 1485 | { |
| 1486 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); |
| 1487 | struct rcu_node *rnp = rdp->mynode; |
| 1488 | |
| 1489 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ |
Paul E. McKenney | 62ab707 | 2012-07-16 10:42:38 +0000 | [diff] [blame] | 1490 | if (rcu_scheduler_fully_active) |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1491 | (void)rcu_spawn_one_boost_kthread(rcu_state, rnp); |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1492 | } |
| 1493 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1494 | #else /* #ifdef CONFIG_RCU_BOOST */ |
| 1495 | |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 1496 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1497 | { |
Paul E. McKenney | 1217ed1 | 2011-05-04 21:43:49 -0700 | [diff] [blame] | 1498 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1499 | } |
| 1500 | |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 1501 | static void invoke_rcu_callbacks_kthread(void) |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1502 | { |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 1503 | WARN_ON_ONCE(1); |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1504 | } |
| 1505 | |
Paul E. McKenney | dff1672 | 2011-11-29 15:57:13 -0800 | [diff] [blame] | 1506 | static bool rcu_is_callbacks_kthread(void) |
| 1507 | { |
| 1508 | return false; |
| 1509 | } |
| 1510 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1511 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
| 1512 | { |
| 1513 | } |
| 1514 | |
Thomas Gleixner | 5d01bbd | 2012-07-16 10:42:35 +0000 | [diff] [blame] | 1515 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1516 | { |
| 1517 | } |
| 1518 | |
Paul E. McKenney | b0d3041 | 2011-07-10 15:57:35 -0700 | [diff] [blame] | 1519 | static int __init rcu_scheduler_really_started(void) |
| 1520 | { |
| 1521 | rcu_scheduler_fully_active = 1; |
| 1522 | return 0; |
| 1523 | } |
| 1524 | early_initcall(rcu_scheduler_really_started); |
| 1525 | |
Paul E. McKenney | f8b7fc6 | 2011-06-16 08:26:32 -0700 | [diff] [blame] | 1526 | static void __cpuinit rcu_prepare_kthreads(int cpu) |
| 1527 | { |
| 1528 | } |
| 1529 | |
Paul E. McKenney | 27f4d28 | 2011-02-07 12:47:15 -0800 | [diff] [blame] | 1530 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
| 1531 | |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1532 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
| 1533 | |
| 1534 | /* |
| 1535 | * Check to see if any future RCU-related work will need to be done |
| 1536 | * by the current CPU, even if none need be done immediately, returning |
| 1537 | * 1 if so. This function is part of the RCU implementation; it is -not- |
| 1538 | * an exported member of the RCU API. |
| 1539 | * |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1540 | * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs |
| 1541 | * any flavor of RCU. |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1542 | */ |
Paul E. McKenney | aa9b1630 | 2012-05-10 16:41:44 -0700 | [diff] [blame] | 1543 | int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies) |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1544 | { |
Paul E. McKenney | aa9b1630 | 2012-05-10 16:41:44 -0700 | [diff] [blame] | 1545 | *delta_jiffies = ULONG_MAX; |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1546 | return rcu_cpu_has_callbacks(cpu); |
| 1547 | } |
| 1548 | |
| 1549 | /* |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1550 | * Because we do not have RCU_FAST_NO_HZ, don't bother initializing for it. |
| 1551 | */ |
| 1552 | static void rcu_prepare_for_idle_init(int cpu) |
| 1553 | { |
| 1554 | } |
| 1555 | |
| 1556 | /* |
| 1557 | * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up |
| 1558 | * after it. |
| 1559 | */ |
| 1560 | static void rcu_cleanup_after_idle(int cpu) |
| 1561 | { |
| 1562 | } |
| 1563 | |
| 1564 | /* |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 1565 | * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n, |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1566 | * is nothing. |
| 1567 | */ |
| 1568 | static void rcu_prepare_for_idle(int cpu) |
| 1569 | { |
| 1570 | } |
| 1571 | |
Paul E. McKenney | c57afe8 | 2012-02-28 11:02:21 -0800 | [diff] [blame] | 1572 | /* |
| 1573 | * Don't bother keeping a running count of the number of RCU callbacks |
| 1574 | * posted because CONFIG_RCU_FAST_NO_HZ=n. |
| 1575 | */ |
| 1576 | static void rcu_idle_count_callbacks_posted(void) |
| 1577 | { |
| 1578 | } |
| 1579 | |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1580 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
| 1581 | |
Paul E. McKenney | f23f7fa | 2011-11-30 15:41:14 -0800 | [diff] [blame] | 1582 | /* |
| 1583 | * This code is invoked when a CPU goes idle, at which point we want |
| 1584 | * to have the CPU do everything required for RCU so that it can enter |
| 1585 | * the energy-efficient dyntick-idle mode. This is handled by a |
| 1586 | * state machine implemented by rcu_prepare_for_idle() below. |
| 1587 | * |
| 1588 | * The following three proprocessor symbols control this state machine: |
| 1589 | * |
| 1590 | * RCU_IDLE_FLUSHES gives the maximum number of times that we will attempt |
| 1591 | * to satisfy RCU. Beyond this point, it is better to incur a periodic |
| 1592 | * scheduling-clock interrupt than to loop through the state machine |
| 1593 | * at full power. |
| 1594 | * RCU_IDLE_OPT_FLUSHES gives the number of RCU_IDLE_FLUSHES that are |
| 1595 | * optional if RCU does not need anything immediately from this |
| 1596 | * CPU, even if this CPU still has RCU callbacks queued. The first |
| 1597 | * times through the state machine are mandatory: we need to give |
| 1598 | * the state machine a chance to communicate a quiescent state |
| 1599 | * to the RCU core. |
| 1600 | * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted |
| 1601 | * to sleep in dyntick-idle mode with RCU callbacks pending. This |
| 1602 | * is sized to be roughly one RCU grace period. Those energy-efficiency |
| 1603 | * benchmarkers who might otherwise be tempted to set this to a large |
| 1604 | * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your |
| 1605 | * system. And if you are -that- concerned about energy efficiency, |
| 1606 | * just power the system down and be done with it! |
Paul E. McKenney | 778d250 | 2012-01-10 14:13:24 -0800 | [diff] [blame] | 1607 | * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is |
| 1608 | * permitted to sleep in dyntick-idle mode with only lazy RCU |
| 1609 | * callbacks pending. Setting this too high can OOM your system. |
Paul E. McKenney | f23f7fa | 2011-11-30 15:41:14 -0800 | [diff] [blame] | 1610 | * |
| 1611 | * The values below work well in practice. If future workloads require |
| 1612 | * adjustment, they can be converted into kernel config parameters, though |
| 1613 | * making the state machine smarter might be a better option. |
| 1614 | */ |
| 1615 | #define RCU_IDLE_FLUSHES 5 /* Number of dyntick-idle tries. */ |
| 1616 | #define RCU_IDLE_OPT_FLUSHES 3 /* Optional dyntick-idle tries. */ |
Paul E. McKenney | e84c48a | 2012-06-04 20:45:10 -0700 | [diff] [blame] | 1617 | #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */ |
Paul E. McKenney | 778d250 | 2012-01-10 14:13:24 -0800 | [diff] [blame] | 1618 | #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */ |
Paul E. McKenney | f23f7fa | 2011-11-30 15:41:14 -0800 | [diff] [blame] | 1619 | |
Paul E. McKenney | 9d2ad24 | 2012-06-24 10:15:02 -0700 | [diff] [blame] | 1620 | extern int tick_nohz_enabled; |
| 1621 | |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1622 | /* |
Paul E. McKenney | 486e259 | 2012-01-06 14:11:30 -0800 | [diff] [blame] | 1623 | * Does the specified flavor of RCU have non-lazy callbacks pending on |
| 1624 | * the specified CPU? Both RCU flavor and CPU are specified by the |
| 1625 | * rcu_data structure. |
| 1626 | */ |
| 1627 | static bool __rcu_cpu_has_nonlazy_callbacks(struct rcu_data *rdp) |
| 1628 | { |
| 1629 | return rdp->qlen != rdp->qlen_lazy; |
| 1630 | } |
| 1631 | |
| 1632 | #ifdef CONFIG_TREE_PREEMPT_RCU |
| 1633 | |
| 1634 | /* |
| 1635 | * Are there non-lazy RCU-preempt callbacks? (There cannot be if there |
| 1636 | * is no RCU-preempt in the kernel.) |
| 1637 | */ |
| 1638 | static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu) |
| 1639 | { |
| 1640 | struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu); |
| 1641 | |
| 1642 | return __rcu_cpu_has_nonlazy_callbacks(rdp); |
| 1643 | } |
| 1644 | |
| 1645 | #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */ |
| 1646 | |
| 1647 | static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu) |
| 1648 | { |
| 1649 | return 0; |
| 1650 | } |
| 1651 | |
| 1652 | #endif /* else #ifdef CONFIG_TREE_PREEMPT_RCU */ |
| 1653 | |
| 1654 | /* |
| 1655 | * Does any flavor of RCU have non-lazy callbacks on the specified CPU? |
| 1656 | */ |
| 1657 | static bool rcu_cpu_has_nonlazy_callbacks(int cpu) |
| 1658 | { |
| 1659 | return __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_sched_data, cpu)) || |
| 1660 | __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_bh_data, cpu)) || |
| 1661 | rcu_preempt_cpu_has_nonlazy_callbacks(cpu); |
| 1662 | } |
| 1663 | |
| 1664 | /* |
Paul E. McKenney | aa9b1630 | 2012-05-10 16:41:44 -0700 | [diff] [blame] | 1665 | * Allow the CPU to enter dyntick-idle mode if either: (1) There are no |
| 1666 | * callbacks on this CPU, (2) this CPU has not yet attempted to enter |
| 1667 | * dyntick-idle mode, or (3) this CPU is in the process of attempting to |
| 1668 | * enter dyntick-idle mode. Otherwise, if we have recently tried and failed |
| 1669 | * to enter dyntick-idle mode, we refuse to try to enter it. After all, |
| 1670 | * it is better to incur scheduling-clock interrupts than to spin |
| 1671 | * continuously for the same time duration! |
| 1672 | * |
| 1673 | * The delta_jiffies argument is used to store the time when RCU is |
| 1674 | * going to need the CPU again if it still has callbacks. The reason |
| 1675 | * for this is that rcu_prepare_for_idle() might need to post a timer, |
| 1676 | * but if so, it will do so after tick_nohz_stop_sched_tick() has set |
| 1677 | * the wakeup time for this CPU. This means that RCU's timer can be |
| 1678 | * delayed until the wakeup time, which defeats the purpose of posting |
| 1679 | * a timer. |
| 1680 | */ |
| 1681 | int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies) |
| 1682 | { |
| 1683 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
| 1684 | |
| 1685 | /* Flag a new idle sojourn to the idle-entry state machine. */ |
| 1686 | rdtp->idle_first_pass = 1; |
| 1687 | /* If no callbacks, RCU doesn't need the CPU. */ |
| 1688 | if (!rcu_cpu_has_callbacks(cpu)) { |
| 1689 | *delta_jiffies = ULONG_MAX; |
| 1690 | return 0; |
| 1691 | } |
| 1692 | if (rdtp->dyntick_holdoff == jiffies) { |
| 1693 | /* RCU recently tried and failed, so don't try again. */ |
| 1694 | *delta_jiffies = 1; |
| 1695 | return 1; |
| 1696 | } |
| 1697 | /* Set up for the possibility that RCU will post a timer. */ |
Paul E. McKenney | e84c48a | 2012-06-04 20:45:10 -0700 | [diff] [blame] | 1698 | if (rcu_cpu_has_nonlazy_callbacks(cpu)) { |
| 1699 | *delta_jiffies = round_up(RCU_IDLE_GP_DELAY + jiffies, |
| 1700 | RCU_IDLE_GP_DELAY) - jiffies; |
| 1701 | } else { |
| 1702 | *delta_jiffies = jiffies + RCU_IDLE_LAZY_GP_DELAY; |
| 1703 | *delta_jiffies = round_jiffies(*delta_jiffies) - jiffies; |
| 1704 | } |
Paul E. McKenney | aa9b1630 | 2012-05-10 16:41:44 -0700 | [diff] [blame] | 1705 | return 0; |
| 1706 | } |
| 1707 | |
| 1708 | /* |
Paul E. McKenney | 21e52e1 | 2012-04-30 14:16:19 -0700 | [diff] [blame] | 1709 | * Handler for smp_call_function_single(). The only point of this |
| 1710 | * handler is to wake the CPU up, so the handler does only tracing. |
| 1711 | */ |
| 1712 | void rcu_idle_demigrate(void *unused) |
| 1713 | { |
| 1714 | trace_rcu_prep_idle("Demigrate"); |
| 1715 | } |
| 1716 | |
| 1717 | /* |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1718 | * Timer handler used to force CPU to start pushing its remaining RCU |
| 1719 | * callbacks in the case where it entered dyntick-idle mode with callbacks |
| 1720 | * pending. The hander doesn't really need to do anything because the |
| 1721 | * real work is done upon re-entry to idle, or by the next scheduling-clock |
| 1722 | * interrupt should idle not be re-entered. |
Paul E. McKenney | 21e52e1 | 2012-04-30 14:16:19 -0700 | [diff] [blame] | 1723 | * |
| 1724 | * One special case: the timer gets migrated without awakening the CPU |
| 1725 | * on which the timer was scheduled on. In this case, we must wake up |
| 1726 | * that CPU. We do so with smp_call_function_single(). |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1727 | */ |
Paul E. McKenney | 21e52e1 | 2012-04-30 14:16:19 -0700 | [diff] [blame] | 1728 | static void rcu_idle_gp_timer_func(unsigned long cpu_in) |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1729 | { |
Paul E. McKenney | 21e52e1 | 2012-04-30 14:16:19 -0700 | [diff] [blame] | 1730 | int cpu = (int)cpu_in; |
| 1731 | |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1732 | trace_rcu_prep_idle("Timer"); |
Paul E. McKenney | 21e52e1 | 2012-04-30 14:16:19 -0700 | [diff] [blame] | 1733 | if (cpu != smp_processor_id()) |
| 1734 | smp_call_function_single(cpu, rcu_idle_demigrate, NULL, 0); |
| 1735 | else |
| 1736 | WARN_ON_ONCE(1); /* Getting here can hang the system... */ |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1737 | } |
| 1738 | |
| 1739 | /* |
| 1740 | * Initialize the timer used to pull CPUs out of dyntick-idle mode. |
| 1741 | */ |
| 1742 | static void rcu_prepare_for_idle_init(int cpu) |
| 1743 | { |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1744 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
| 1745 | |
| 1746 | rdtp->dyntick_holdoff = jiffies - 1; |
| 1747 | setup_timer(&rdtp->idle_gp_timer, rcu_idle_gp_timer_func, cpu); |
| 1748 | rdtp->idle_gp_timer_expires = jiffies - 1; |
| 1749 | rdtp->idle_first_pass = 1; |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1750 | } |
| 1751 | |
| 1752 | /* |
| 1753 | * Clean up for exit from idle. Because we are exiting from idle, there |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1754 | * is no longer any point to ->idle_gp_timer, so cancel it. This will |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1755 | * do nothing if this timer is not active, so just cancel it unconditionally. |
| 1756 | */ |
| 1757 | static void rcu_cleanup_after_idle(int cpu) |
| 1758 | { |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1759 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
| 1760 | |
| 1761 | del_timer(&rdtp->idle_gp_timer); |
Paul E. McKenney | 2fdbb31 | 2012-02-23 15:58:29 -0800 | [diff] [blame] | 1762 | trace_rcu_prep_idle("Cleanup after idle"); |
Paul E. McKenney | 9d2ad24 | 2012-06-24 10:15:02 -0700 | [diff] [blame] | 1763 | rdtp->tick_nohz_enabled_snap = ACCESS_ONCE(tick_nohz_enabled); |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1764 | } |
| 1765 | |
| 1766 | /* |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1767 | * Check to see if any RCU-related work can be done by the current CPU, |
| 1768 | * and if so, schedule a softirq to get it done. This function is part |
| 1769 | * of the RCU implementation; it is -not- an exported member of the RCU API. |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1770 | * |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1771 | * The idea is for the current CPU to clear out all work required by the |
| 1772 | * RCU core for the current grace period, so that this CPU can be permitted |
| 1773 | * to enter dyntick-idle mode. In some cases, it will need to be awakened |
| 1774 | * at the end of the grace period by whatever CPU ends the grace period. |
| 1775 | * This allows CPUs to go dyntick-idle more quickly, and to reduce the |
| 1776 | * number of wakeups by a modest integer factor. |
Paul E. McKenney | a47cd88 | 2010-02-26 16:38:56 -0800 | [diff] [blame] | 1777 | * |
| 1778 | * Because it is not legal to invoke rcu_process_callbacks() with irqs |
| 1779 | * disabled, we do one pass of force_quiescent_state(), then do a |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 1780 | * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1781 | * later. The ->dyntick_drain field controls the sequencing. |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1782 | * |
| 1783 | * The caller must have disabled interrupts. |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1784 | */ |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1785 | static void rcu_prepare_for_idle(int cpu) |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1786 | { |
Paul E. McKenney | f511fc6 | 2012-03-15 12:16:26 -0700 | [diff] [blame] | 1787 | struct timer_list *tp; |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1788 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
Paul E. McKenney | 9d2ad24 | 2012-06-24 10:15:02 -0700 | [diff] [blame] | 1789 | int tne; |
| 1790 | |
| 1791 | /* Handle nohz enablement switches conservatively. */ |
| 1792 | tne = ACCESS_ONCE(tick_nohz_enabled); |
| 1793 | if (tne != rdtp->tick_nohz_enabled_snap) { |
| 1794 | if (rcu_cpu_has_callbacks(cpu)) |
| 1795 | invoke_rcu_core(); /* force nohz to see update. */ |
| 1796 | rdtp->tick_nohz_enabled_snap = tne; |
| 1797 | return; |
| 1798 | } |
| 1799 | if (!tne) |
| 1800 | return; |
Paul E. McKenney | f511fc6 | 2012-03-15 12:16:26 -0700 | [diff] [blame] | 1801 | |
Paul E. McKenney | 9a0c6fe | 2012-06-28 12:33:51 -0700 | [diff] [blame] | 1802 | /* Adaptive-tick mode, where usermode execution is idle to RCU. */ |
| 1803 | if (!is_idle_task(current)) { |
| 1804 | rdtp->dyntick_holdoff = jiffies - 1; |
| 1805 | if (rcu_cpu_has_nonlazy_callbacks(cpu)) { |
| 1806 | trace_rcu_prep_idle("User dyntick with callbacks"); |
| 1807 | rdtp->idle_gp_timer_expires = |
| 1808 | round_up(jiffies + RCU_IDLE_GP_DELAY, |
| 1809 | RCU_IDLE_GP_DELAY); |
| 1810 | } else if (rcu_cpu_has_callbacks(cpu)) { |
| 1811 | rdtp->idle_gp_timer_expires = |
| 1812 | round_jiffies(jiffies + RCU_IDLE_LAZY_GP_DELAY); |
| 1813 | trace_rcu_prep_idle("User dyntick with lazy callbacks"); |
| 1814 | } else { |
| 1815 | return; |
| 1816 | } |
| 1817 | tp = &rdtp->idle_gp_timer; |
| 1818 | mod_timer_pinned(tp, rdtp->idle_gp_timer_expires); |
| 1819 | return; |
| 1820 | } |
| 1821 | |
Paul E. McKenney | 3084f2f | 2011-11-22 17:07:11 -0800 | [diff] [blame] | 1822 | /* |
Paul E. McKenney | c57afe8 | 2012-02-28 11:02:21 -0800 | [diff] [blame] | 1823 | * If this is an idle re-entry, for example, due to use of |
| 1824 | * RCU_NONIDLE() or the new idle-loop tracing API within the idle |
| 1825 | * loop, then don't take any state-machine actions, unless the |
| 1826 | * momentary exit from idle queued additional non-lazy callbacks. |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1827 | * Instead, repost the ->idle_gp_timer if this CPU has callbacks |
Paul E. McKenney | c57afe8 | 2012-02-28 11:02:21 -0800 | [diff] [blame] | 1828 | * pending. |
| 1829 | */ |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1830 | if (!rdtp->idle_first_pass && |
| 1831 | (rdtp->nonlazy_posted == rdtp->nonlazy_posted_snap)) { |
Paul E. McKenney | f511fc6 | 2012-03-15 12:16:26 -0700 | [diff] [blame] | 1832 | if (rcu_cpu_has_callbacks(cpu)) { |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1833 | tp = &rdtp->idle_gp_timer; |
| 1834 | mod_timer_pinned(tp, rdtp->idle_gp_timer_expires); |
Paul E. McKenney | f511fc6 | 2012-03-15 12:16:26 -0700 | [diff] [blame] | 1835 | } |
Paul E. McKenney | c57afe8 | 2012-02-28 11:02:21 -0800 | [diff] [blame] | 1836 | return; |
| 1837 | } |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1838 | rdtp->idle_first_pass = 0; |
| 1839 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted - 1; |
Paul E. McKenney | c57afe8 | 2012-02-28 11:02:21 -0800 | [diff] [blame] | 1840 | |
| 1841 | /* |
Paul E. McKenney | f535a60 | 2011-11-22 20:43:02 -0800 | [diff] [blame] | 1842 | * If there are no callbacks on this CPU, enter dyntick-idle mode. |
| 1843 | * Also reset state to avoid prejudicing later attempts. |
Paul E. McKenney | 3084f2f | 2011-11-22 17:07:11 -0800 | [diff] [blame] | 1844 | */ |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1845 | if (!rcu_cpu_has_callbacks(cpu)) { |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1846 | rdtp->dyntick_holdoff = jiffies - 1; |
| 1847 | rdtp->dyntick_drain = 0; |
Paul E. McKenney | 433cddd | 2011-11-22 14:58:03 -0800 | [diff] [blame] | 1848 | trace_rcu_prep_idle("No callbacks"); |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1849 | return; |
Paul E. McKenney | 77e38ed | 2010-04-25 21:04:29 -0700 | [diff] [blame] | 1850 | } |
Paul E. McKenney | 3084f2f | 2011-11-22 17:07:11 -0800 | [diff] [blame] | 1851 | |
| 1852 | /* |
| 1853 | * If in holdoff mode, just return. We will presumably have |
| 1854 | * refrained from disabling the scheduling-clock tick. |
| 1855 | */ |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1856 | if (rdtp->dyntick_holdoff == jiffies) { |
Paul E. McKenney | 433cddd | 2011-11-22 14:58:03 -0800 | [diff] [blame] | 1857 | trace_rcu_prep_idle("In holdoff"); |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1858 | return; |
Paul E. McKenney | 433cddd | 2011-11-22 14:58:03 -0800 | [diff] [blame] | 1859 | } |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1860 | |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1861 | /* Check and update the ->dyntick_drain sequencing. */ |
| 1862 | if (rdtp->dyntick_drain <= 0) { |
Paul E. McKenney | a47cd88 | 2010-02-26 16:38:56 -0800 | [diff] [blame] | 1863 | /* First time through, initialize the counter. */ |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1864 | rdtp->dyntick_drain = RCU_IDLE_FLUSHES; |
| 1865 | } else if (rdtp->dyntick_drain <= RCU_IDLE_OPT_FLUSHES && |
Paul E. McKenney | c3ce910 | 2012-02-14 10:12:54 -0800 | [diff] [blame] | 1866 | !rcu_pending(cpu) && |
| 1867 | !local_softirq_pending()) { |
Paul E. McKenney | 7cb9249 | 2011-11-28 12:28:34 -0800 | [diff] [blame] | 1868 | /* Can we go dyntick-idle despite still having callbacks? */ |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1869 | rdtp->dyntick_drain = 0; |
| 1870 | rdtp->dyntick_holdoff = jiffies; |
Paul E. McKenney | fd4b352 | 2012-05-05 19:10:35 -0700 | [diff] [blame] | 1871 | if (rcu_cpu_has_nonlazy_callbacks(cpu)) { |
| 1872 | trace_rcu_prep_idle("Dyntick with callbacks"); |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1873 | rdtp->idle_gp_timer_expires = |
Paul E. McKenney | e84c48a | 2012-06-04 20:45:10 -0700 | [diff] [blame] | 1874 | round_up(jiffies + RCU_IDLE_GP_DELAY, |
| 1875 | RCU_IDLE_GP_DELAY); |
Paul E. McKenney | fd4b352 | 2012-05-05 19:10:35 -0700 | [diff] [blame] | 1876 | } else { |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1877 | rdtp->idle_gp_timer_expires = |
Paul E. McKenney | e84c48a | 2012-06-04 20:45:10 -0700 | [diff] [blame] | 1878 | round_jiffies(jiffies + RCU_IDLE_LAZY_GP_DELAY); |
Paul E. McKenney | fd4b352 | 2012-05-05 19:10:35 -0700 | [diff] [blame] | 1879 | trace_rcu_prep_idle("Dyntick with lazy callbacks"); |
| 1880 | } |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1881 | tp = &rdtp->idle_gp_timer; |
| 1882 | mod_timer_pinned(tp, rdtp->idle_gp_timer_expires); |
| 1883 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; |
Paul E. McKenney | f23f7fa | 2011-11-30 15:41:14 -0800 | [diff] [blame] | 1884 | return; /* Nothing more to do immediately. */ |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1885 | } else if (--(rdtp->dyntick_drain) <= 0) { |
Paul E. McKenney | a47cd88 | 2010-02-26 16:38:56 -0800 | [diff] [blame] | 1886 | /* We have hit the limit, so time to give up. */ |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1887 | rdtp->dyntick_holdoff = jiffies; |
Paul E. McKenney | 433cddd | 2011-11-22 14:58:03 -0800 | [diff] [blame] | 1888 | trace_rcu_prep_idle("Begin holdoff"); |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1889 | invoke_rcu_core(); /* Force the CPU out of dyntick-idle. */ |
| 1890 | return; |
Paul E. McKenney | a47cd88 | 2010-02-26 16:38:56 -0800 | [diff] [blame] | 1891 | } |
| 1892 | |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1893 | /* |
| 1894 | * Do one step of pushing the remaining RCU callbacks through |
| 1895 | * the RCU core state machine. |
| 1896 | */ |
| 1897 | #ifdef CONFIG_TREE_PREEMPT_RCU |
| 1898 | if (per_cpu(rcu_preempt_data, cpu).nxtlist) { |
| 1899 | rcu_preempt_qs(cpu); |
Paul E. McKenney | 4cdfc17 | 2012-06-22 17:06:26 -0700 | [diff] [blame] | 1900 | force_quiescent_state(&rcu_preempt_state); |
Paul E. McKenney | aea1b35 | 2011-11-02 06:54:54 -0700 | [diff] [blame] | 1901 | } |
| 1902 | #endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */ |
Paul E. McKenney | a47cd88 | 2010-02-26 16:38:56 -0800 | [diff] [blame] | 1903 | if (per_cpu(rcu_sched_data, cpu).nxtlist) { |
| 1904 | rcu_sched_qs(cpu); |
Paul E. McKenney | 4cdfc17 | 2012-06-22 17:06:26 -0700 | [diff] [blame] | 1905 | force_quiescent_state(&rcu_sched_state); |
Paul E. McKenney | a47cd88 | 2010-02-26 16:38:56 -0800 | [diff] [blame] | 1906 | } |
| 1907 | if (per_cpu(rcu_bh_data, cpu).nxtlist) { |
| 1908 | rcu_bh_qs(cpu); |
Paul E. McKenney | 4cdfc17 | 2012-06-22 17:06:26 -0700 | [diff] [blame] | 1909 | force_quiescent_state(&rcu_bh_state); |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1910 | } |
| 1911 | |
Paul E. McKenney | 433cddd | 2011-11-22 14:58:03 -0800 | [diff] [blame] | 1912 | /* |
| 1913 | * If RCU callbacks are still pending, RCU still needs this CPU. |
| 1914 | * So try forcing the callbacks through the grace period. |
| 1915 | */ |
Paul E. McKenney | 3ad0dec | 2011-11-22 21:08:13 -0800 | [diff] [blame] | 1916 | if (rcu_cpu_has_callbacks(cpu)) { |
Paul E. McKenney | 433cddd | 2011-11-22 14:58:03 -0800 | [diff] [blame] | 1917 | trace_rcu_prep_idle("More callbacks"); |
Paul E. McKenney | a46e089 | 2011-06-15 15:47:09 -0700 | [diff] [blame] | 1918 | invoke_rcu_core(); |
Paul E. McKenney | c701d5d | 2012-06-28 08:08:25 -0700 | [diff] [blame] | 1919 | } else { |
Paul E. McKenney | 433cddd | 2011-11-22 14:58:03 -0800 | [diff] [blame] | 1920 | trace_rcu_prep_idle("Callbacks drained"); |
Paul E. McKenney | c701d5d | 2012-06-28 08:08:25 -0700 | [diff] [blame] | 1921 | } |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 1922 | } |
| 1923 | |
Paul E. McKenney | c57afe8 | 2012-02-28 11:02:21 -0800 | [diff] [blame] | 1924 | /* |
Paul E. McKenney | 98248a0 | 2012-05-03 15:38:10 -0700 | [diff] [blame] | 1925 | * Keep a running count of the number of non-lazy callbacks posted |
| 1926 | * on this CPU. This running counter (which is never decremented) allows |
| 1927 | * rcu_prepare_for_idle() to detect when something out of the idle loop |
| 1928 | * posts a callback, even if an equal number of callbacks are invoked. |
| 1929 | * Of course, callbacks should only be posted from within a trace event |
| 1930 | * designed to be called from idle or from within RCU_NONIDLE(). |
Paul E. McKenney | c57afe8 | 2012-02-28 11:02:21 -0800 | [diff] [blame] | 1931 | */ |
| 1932 | static void rcu_idle_count_callbacks_posted(void) |
| 1933 | { |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 1934 | __this_cpu_add(rcu_dynticks.nonlazy_posted, 1); |
Paul E. McKenney | c57afe8 | 2012-02-28 11:02:21 -0800 | [diff] [blame] | 1935 | } |
| 1936 | |
Paul E. McKenney | b626c1b | 2012-06-11 17:39:43 -0700 | [diff] [blame] | 1937 | /* |
| 1938 | * Data for flushing lazy RCU callbacks at OOM time. |
| 1939 | */ |
| 1940 | static atomic_t oom_callback_count; |
| 1941 | static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq); |
| 1942 | |
| 1943 | /* |
| 1944 | * RCU OOM callback -- decrement the outstanding count and deliver the |
| 1945 | * wake-up if we are the last one. |
| 1946 | */ |
| 1947 | static void rcu_oom_callback(struct rcu_head *rhp) |
| 1948 | { |
| 1949 | if (atomic_dec_and_test(&oom_callback_count)) |
| 1950 | wake_up(&oom_callback_wq); |
| 1951 | } |
| 1952 | |
| 1953 | /* |
| 1954 | * Post an rcu_oom_notify callback on the current CPU if it has at |
| 1955 | * least one lazy callback. This will unnecessarily post callbacks |
| 1956 | * to CPUs that already have a non-lazy callback at the end of their |
| 1957 | * callback list, but this is an infrequent operation, so accept some |
| 1958 | * extra overhead to keep things simple. |
| 1959 | */ |
| 1960 | static void rcu_oom_notify_cpu(void *unused) |
| 1961 | { |
| 1962 | struct rcu_state *rsp; |
| 1963 | struct rcu_data *rdp; |
| 1964 | |
| 1965 | for_each_rcu_flavor(rsp) { |
| 1966 | rdp = __this_cpu_ptr(rsp->rda); |
| 1967 | if (rdp->qlen_lazy != 0) { |
| 1968 | atomic_inc(&oom_callback_count); |
| 1969 | rsp->call(&rdp->oom_head, rcu_oom_callback); |
| 1970 | } |
| 1971 | } |
| 1972 | } |
| 1973 | |
| 1974 | /* |
| 1975 | * If low on memory, ensure that each CPU has a non-lazy callback. |
| 1976 | * This will wake up CPUs that have only lazy callbacks, in turn |
| 1977 | * ensuring that they free up the corresponding memory in a timely manner. |
| 1978 | * Because an uncertain amount of memory will be freed in some uncertain |
| 1979 | * timeframe, we do not claim to have freed anything. |
| 1980 | */ |
| 1981 | static int rcu_oom_notify(struct notifier_block *self, |
| 1982 | unsigned long notused, void *nfreed) |
| 1983 | { |
| 1984 | int cpu; |
| 1985 | |
| 1986 | /* Wait for callbacks from earlier instance to complete. */ |
| 1987 | wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0); |
| 1988 | |
| 1989 | /* |
| 1990 | * Prevent premature wakeup: ensure that all increments happen |
| 1991 | * before there is a chance of the counter reaching zero. |
| 1992 | */ |
| 1993 | atomic_set(&oom_callback_count, 1); |
| 1994 | |
| 1995 | get_online_cpus(); |
| 1996 | for_each_online_cpu(cpu) { |
| 1997 | smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1); |
| 1998 | cond_resched(); |
| 1999 | } |
| 2000 | put_online_cpus(); |
| 2001 | |
| 2002 | /* Unconditionally decrement: no need to wake ourselves up. */ |
| 2003 | atomic_dec(&oom_callback_count); |
| 2004 | |
| 2005 | return NOTIFY_OK; |
| 2006 | } |
| 2007 | |
| 2008 | static struct notifier_block rcu_oom_nb = { |
| 2009 | .notifier_call = rcu_oom_notify |
| 2010 | }; |
| 2011 | |
| 2012 | static int __init rcu_register_oom_notifier(void) |
| 2013 | { |
| 2014 | register_oom_notifier(&rcu_oom_nb); |
| 2015 | return 0; |
| 2016 | } |
| 2017 | early_initcall(rcu_register_oom_notifier); |
| 2018 | |
Paul E. McKenney | 8bd93a2 | 2010-02-22 17:04:59 -0800 | [diff] [blame] | 2019 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 2020 | |
| 2021 | #ifdef CONFIG_RCU_CPU_STALL_INFO |
| 2022 | |
| 2023 | #ifdef CONFIG_RCU_FAST_NO_HZ |
| 2024 | |
| 2025 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) |
| 2026 | { |
Paul E. McKenney | 5955f7e | 2012-05-09 12:07:05 -0700 | [diff] [blame] | 2027 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
| 2028 | struct timer_list *tltp = &rdtp->idle_gp_timer; |
Paul E. McKenney | 86f343b | 2012-09-21 10:41:50 -0700 | [diff] [blame] | 2029 | char c; |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 2030 | |
Paul E. McKenney | 86f343b | 2012-09-21 10:41:50 -0700 | [diff] [blame] | 2031 | c = rdtp->dyntick_holdoff == jiffies ? 'H' : '.'; |
| 2032 | if (timer_pending(tltp)) |
| 2033 | sprintf(cp, "drain=%d %c timer=%lu", |
| 2034 | rdtp->dyntick_drain, c, tltp->expires - jiffies); |
| 2035 | else |
| 2036 | sprintf(cp, "drain=%d %c timer not pending", |
| 2037 | rdtp->dyntick_drain, c); |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 2038 | } |
| 2039 | |
| 2040 | #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */ |
| 2041 | |
| 2042 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) |
| 2043 | { |
Carsten Emde | 1c17e4d | 2012-06-19 10:43:16 +0200 | [diff] [blame] | 2044 | *cp = '\0'; |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 2045 | } |
| 2046 | |
| 2047 | #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */ |
| 2048 | |
| 2049 | /* Initiate the stall-info list. */ |
| 2050 | static void print_cpu_stall_info_begin(void) |
| 2051 | { |
| 2052 | printk(KERN_CONT "\n"); |
| 2053 | } |
| 2054 | |
| 2055 | /* |
| 2056 | * Print out diagnostic information for the specified stalled CPU. |
| 2057 | * |
| 2058 | * If the specified CPU is aware of the current RCU grace period |
| 2059 | * (flavor specified by rsp), then print the number of scheduling |
| 2060 | * clock interrupts the CPU has taken during the time that it has |
| 2061 | * been aware. Otherwise, print the number of RCU grace periods |
| 2062 | * that this CPU is ignorant of, for example, "1" if the CPU was |
| 2063 | * aware of the previous grace period. |
| 2064 | * |
| 2065 | * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info. |
| 2066 | */ |
| 2067 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) |
| 2068 | { |
| 2069 | char fast_no_hz[72]; |
| 2070 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); |
| 2071 | struct rcu_dynticks *rdtp = rdp->dynticks; |
| 2072 | char *ticks_title; |
| 2073 | unsigned long ticks_value; |
| 2074 | |
| 2075 | if (rsp->gpnum == rdp->gpnum) { |
| 2076 | ticks_title = "ticks this GP"; |
| 2077 | ticks_value = rdp->ticks_this_gp; |
| 2078 | } else { |
| 2079 | ticks_title = "GPs behind"; |
| 2080 | ticks_value = rsp->gpnum - rdp->gpnum; |
| 2081 | } |
| 2082 | print_cpu_stall_fast_no_hz(fast_no_hz, cpu); |
| 2083 | printk(KERN_ERR "\t%d: (%lu %s) idle=%03x/%llx/%d %s\n", |
| 2084 | cpu, ticks_value, ticks_title, |
| 2085 | atomic_read(&rdtp->dynticks) & 0xfff, |
| 2086 | rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting, |
| 2087 | fast_no_hz); |
| 2088 | } |
| 2089 | |
| 2090 | /* Terminate the stall-info list. */ |
| 2091 | static void print_cpu_stall_info_end(void) |
| 2092 | { |
| 2093 | printk(KERN_ERR "\t"); |
| 2094 | } |
| 2095 | |
| 2096 | /* Zero ->ticks_this_gp for all flavors of RCU. */ |
| 2097 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) |
| 2098 | { |
| 2099 | rdp->ticks_this_gp = 0; |
| 2100 | } |
| 2101 | |
| 2102 | /* Increment ->ticks_this_gp for all flavors of RCU. */ |
| 2103 | static void increment_cpu_stall_ticks(void) |
| 2104 | { |
Paul E. McKenney | 115f7a7 | 2012-08-10 13:55:03 -0700 | [diff] [blame] | 2105 | struct rcu_state *rsp; |
| 2106 | |
| 2107 | for_each_rcu_flavor(rsp) |
| 2108 | __this_cpu_ptr(rsp->rda)->ticks_this_gp++; |
Paul E. McKenney | a858af2 | 2012-01-16 13:29:10 -0800 | [diff] [blame] | 2109 | } |
| 2110 | |
| 2111 | #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */ |
| 2112 | |
| 2113 | static void print_cpu_stall_info_begin(void) |
| 2114 | { |
| 2115 | printk(KERN_CONT " {"); |
| 2116 | } |
| 2117 | |
| 2118 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) |
| 2119 | { |
| 2120 | printk(KERN_CONT " %d", cpu); |
| 2121 | } |
| 2122 | |
| 2123 | static void print_cpu_stall_info_end(void) |
| 2124 | { |
| 2125 | printk(KERN_CONT "} "); |
| 2126 | } |
| 2127 | |
| 2128 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) |
| 2129 | { |
| 2130 | } |
| 2131 | |
| 2132 | static void increment_cpu_stall_ticks(void) |
| 2133 | { |
| 2134 | } |
| 2135 | |
| 2136 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */ |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2137 | |
| 2138 | #ifdef CONFIG_RCU_NOCB_CPU |
| 2139 | |
| 2140 | /* |
| 2141 | * Offload callback processing from the boot-time-specified set of CPUs |
| 2142 | * specified by rcu_nocb_mask. For each CPU in the set, there is a |
| 2143 | * kthread created that pulls the callbacks from the corresponding CPU, |
| 2144 | * waits for a grace period to elapse, and invokes the callbacks. |
| 2145 | * The no-CBs CPUs do a wake_up() on their kthread when they insert |
| 2146 | * a callback into any empty list, unless the rcu_nocb_poll boot parameter |
| 2147 | * has been specified, in which case each kthread actively polls its |
| 2148 | * CPU. (Which isn't so great for energy efficiency, but which does |
| 2149 | * reduce RCU's overhead on that CPU.) |
| 2150 | * |
| 2151 | * This is intended to be used in conjunction with Frederic Weisbecker's |
| 2152 | * adaptive-idle work, which would seriously reduce OS jitter on CPUs |
| 2153 | * running CPU-bound user-mode computations. |
| 2154 | * |
| 2155 | * Offloading of callback processing could also in theory be used as |
| 2156 | * an energy-efficiency measure because CPUs with no RCU callbacks |
| 2157 | * queued are more aggressive about entering dyntick-idle mode. |
| 2158 | */ |
| 2159 | |
| 2160 | |
| 2161 | /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */ |
| 2162 | static int __init rcu_nocb_setup(char *str) |
| 2163 | { |
| 2164 | alloc_bootmem_cpumask_var(&rcu_nocb_mask); |
| 2165 | have_rcu_nocb_mask = true; |
| 2166 | cpulist_parse(str, rcu_nocb_mask); |
| 2167 | return 1; |
| 2168 | } |
| 2169 | __setup("rcu_nocbs=", rcu_nocb_setup); |
| 2170 | |
Paul Gortmaker | 1b0048a | 2012-12-20 13:19:22 -0800 | [diff] [blame] | 2171 | static int __init parse_rcu_nocb_poll(char *arg) |
| 2172 | { |
| 2173 | rcu_nocb_poll = 1; |
| 2174 | return 0; |
| 2175 | } |
| 2176 | early_param("rcu_nocb_poll", parse_rcu_nocb_poll); |
| 2177 | |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2178 | /* |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2179 | * Do any no-CBs CPUs need another grace period? |
| 2180 | * |
| 2181 | * Interrupts must be disabled. If the caller does not hold the root |
| 2182 | * rnp_node structure's ->lock, the results are advisory only. |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2183 | */ |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2184 | static int rcu_nocb_needs_gp(struct rcu_state *rsp) |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2185 | { |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2186 | struct rcu_node *rnp = rcu_get_root(rsp); |
| 2187 | |
| 2188 | return rnp->n_nocb_gp_requests[(ACCESS_ONCE(rnp->completed) + 1) & 0x1]; |
| 2189 | } |
| 2190 | |
| 2191 | /* |
| 2192 | * Clean up this rcu_node structure's no-CBs state at the end of |
| 2193 | * a grace period, and also return whether any no-CBs CPU associated |
| 2194 | * with this rcu_node structure needs another grace period. |
| 2195 | */ |
| 2196 | static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
| 2197 | { |
| 2198 | int c = rnp->completed; |
| 2199 | int needmore; |
| 2200 | |
| 2201 | wake_up_all(&rnp->nocb_gp_wq[c & 0x1]); |
| 2202 | rnp->n_nocb_gp_requests[c & 0x1] = 0; |
| 2203 | needmore = rnp->n_nocb_gp_requests[(c + 1) & 0x1]; |
Paul E. McKenney | 09c7b89 | 2013-02-08 15:55:02 -0800 | [diff] [blame] | 2204 | trace_rcu_nocb_grace_period(rsp->name, rnp->gpnum, rnp->completed, |
| 2205 | c, rnp->level, rnp->grplo, rnp->grphi, |
| 2206 | needmore ? "CleanupMore" : "Cleanup"); |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2207 | return needmore; |
| 2208 | } |
| 2209 | |
| 2210 | /* |
| 2211 | * Set the root rcu_node structure's ->n_nocb_gp_requests field |
| 2212 | * based on the sum of those of all rcu_node structures. This does |
| 2213 | * double-count the root rcu_node structure's requests, but this |
| 2214 | * is necessary to handle the possibility of a rcu_nocb_kthread() |
| 2215 | * having awakened during the time that the rcu_node structures |
| 2216 | * were being updated for the end of the previous grace period. |
| 2217 | */ |
| 2218 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
| 2219 | { |
| 2220 | rnp->n_nocb_gp_requests[(rnp->completed + 1) & 0x1] += nrq; |
| 2221 | } |
| 2222 | |
| 2223 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
| 2224 | { |
| 2225 | init_waitqueue_head(&rnp->nocb_gp_wq[0]); |
| 2226 | init_waitqueue_head(&rnp->nocb_gp_wq[1]); |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2227 | } |
| 2228 | |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2229 | /* Is the specified CPU a no-CPUs CPU? */ |
| 2230 | static bool is_nocb_cpu(int cpu) |
| 2231 | { |
| 2232 | if (have_rcu_nocb_mask) |
| 2233 | return cpumask_test_cpu(cpu, rcu_nocb_mask); |
| 2234 | return false; |
| 2235 | } |
| 2236 | |
| 2237 | /* |
| 2238 | * Enqueue the specified string of rcu_head structures onto the specified |
| 2239 | * CPU's no-CBs lists. The CPU is specified by rdp, the head of the |
| 2240 | * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy |
| 2241 | * counts are supplied by rhcount and rhcount_lazy. |
| 2242 | * |
| 2243 | * If warranted, also wake up the kthread servicing this CPUs queues. |
| 2244 | */ |
| 2245 | static void __call_rcu_nocb_enqueue(struct rcu_data *rdp, |
| 2246 | struct rcu_head *rhp, |
| 2247 | struct rcu_head **rhtp, |
| 2248 | int rhcount, int rhcount_lazy) |
| 2249 | { |
| 2250 | int len; |
| 2251 | struct rcu_head **old_rhpp; |
| 2252 | struct task_struct *t; |
| 2253 | |
| 2254 | /* Enqueue the callback on the nocb list and update counts. */ |
| 2255 | old_rhpp = xchg(&rdp->nocb_tail, rhtp); |
| 2256 | ACCESS_ONCE(*old_rhpp) = rhp; |
| 2257 | atomic_long_add(rhcount, &rdp->nocb_q_count); |
| 2258 | atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy); |
| 2259 | |
| 2260 | /* If we are not being polled and there is a kthread, awaken it ... */ |
| 2261 | t = ACCESS_ONCE(rdp->nocb_kthread); |
| 2262 | if (rcu_nocb_poll | !t) |
| 2263 | return; |
| 2264 | len = atomic_long_read(&rdp->nocb_q_count); |
| 2265 | if (old_rhpp == &rdp->nocb_head) { |
| 2266 | wake_up(&rdp->nocb_wq); /* ... only if queue was empty ... */ |
| 2267 | rdp->qlen_last_fqs_check = 0; |
| 2268 | } else if (len > rdp->qlen_last_fqs_check + qhimark) { |
| 2269 | wake_up_process(t); /* ... or if many callbacks queued. */ |
| 2270 | rdp->qlen_last_fqs_check = LONG_MAX / 2; |
| 2271 | } |
| 2272 | return; |
| 2273 | } |
| 2274 | |
| 2275 | /* |
| 2276 | * This is a helper for __call_rcu(), which invokes this when the normal |
| 2277 | * callback queue is inoperable. If this is not a no-CBs CPU, this |
| 2278 | * function returns failure back to __call_rcu(), which can complain |
| 2279 | * appropriately. |
| 2280 | * |
| 2281 | * Otherwise, this function queues the callback where the corresponding |
| 2282 | * "rcuo" kthread can find it. |
| 2283 | */ |
| 2284 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
| 2285 | bool lazy) |
| 2286 | { |
| 2287 | |
| 2288 | if (!is_nocb_cpu(rdp->cpu)) |
| 2289 | return 0; |
| 2290 | __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy); |
Paul E. McKenney | 21e7a60 | 2013-02-09 17:42:16 -0800 | [diff] [blame] | 2291 | if (__is_kfree_rcu_offset((unsigned long)rhp->func)) |
| 2292 | trace_rcu_kfree_callback(rdp->rsp->name, rhp, |
| 2293 | (unsigned long)rhp->func, |
| 2294 | rdp->qlen_lazy, rdp->qlen); |
| 2295 | else |
| 2296 | trace_rcu_callback(rdp->rsp->name, rhp, |
| 2297 | rdp->qlen_lazy, rdp->qlen); |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2298 | return 1; |
| 2299 | } |
| 2300 | |
| 2301 | /* |
| 2302 | * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is |
| 2303 | * not a no-CBs CPU. |
| 2304 | */ |
| 2305 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, |
| 2306 | struct rcu_data *rdp) |
| 2307 | { |
| 2308 | long ql = rsp->qlen; |
| 2309 | long qll = rsp->qlen_lazy; |
| 2310 | |
| 2311 | /* If this is not a no-CBs CPU, tell the caller to do it the old way. */ |
| 2312 | if (!is_nocb_cpu(smp_processor_id())) |
| 2313 | return 0; |
| 2314 | rsp->qlen = 0; |
| 2315 | rsp->qlen_lazy = 0; |
| 2316 | |
| 2317 | /* First, enqueue the donelist, if any. This preserves CB ordering. */ |
| 2318 | if (rsp->orphan_donelist != NULL) { |
| 2319 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist, |
| 2320 | rsp->orphan_donetail, ql, qll); |
| 2321 | ql = qll = 0; |
| 2322 | rsp->orphan_donelist = NULL; |
| 2323 | rsp->orphan_donetail = &rsp->orphan_donelist; |
| 2324 | } |
| 2325 | if (rsp->orphan_nxtlist != NULL) { |
| 2326 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist, |
| 2327 | rsp->orphan_nxttail, ql, qll); |
| 2328 | ql = qll = 0; |
| 2329 | rsp->orphan_nxtlist = NULL; |
| 2330 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; |
| 2331 | } |
| 2332 | return 1; |
| 2333 | } |
| 2334 | |
| 2335 | /* |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2336 | * If necessary, kick off a new grace period, and either way wait |
| 2337 | * for a subsequent grace period to complete. |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2338 | */ |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2339 | static void rcu_nocb_wait_gp(struct rcu_data *rdp) |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2340 | { |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2341 | unsigned long c; |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2342 | bool d; |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2343 | unsigned long flags; |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2344 | unsigned long flags1; |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2345 | struct rcu_node *rnp = rdp->mynode; |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2346 | struct rcu_node *rnp_root = rcu_get_root(rdp->rsp); |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2347 | |
| 2348 | raw_spin_lock_irqsave(&rnp->lock, flags); |
| 2349 | c = rnp->completed + 2; |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2350 | |
| 2351 | /* Count our request for a grace period. */ |
| 2352 | rnp->n_nocb_gp_requests[c & 0x1]++; |
Paul E. McKenney | 09c7b89 | 2013-02-08 15:55:02 -0800 | [diff] [blame] | 2353 | trace_rcu_nocb_grace_period(rdp->rsp->name, rnp->gpnum, rnp->completed, |
| 2354 | c, rnp->level, rnp->grplo, rnp->grphi, |
| 2355 | "Startleaf"); |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2356 | |
| 2357 | if (rnp->gpnum != rnp->completed) { |
| 2358 | |
| 2359 | /* |
| 2360 | * This rcu_node structure believes that a grace period |
| 2361 | * is in progress, so we are done. When this grace |
| 2362 | * period ends, our request will be acted upon. |
| 2363 | */ |
Paul E. McKenney | 09c7b89 | 2013-02-08 15:55:02 -0800 | [diff] [blame] | 2364 | trace_rcu_nocb_grace_period(rdp->rsp->name, |
| 2365 | rnp->gpnum, rnp->completed, c, |
| 2366 | rnp->level, rnp->grplo, rnp->grphi, |
| 2367 | "Startedleaf"); |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2368 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 2369 | |
| 2370 | } else { |
| 2371 | |
| 2372 | /* |
| 2373 | * Might not be a grace period, check root rcu_node |
| 2374 | * structure to see if we must start one. |
| 2375 | */ |
| 2376 | if (rnp != rnp_root) |
| 2377 | raw_spin_lock(&rnp_root->lock); /* irqs disabled. */ |
| 2378 | if (rnp_root->gpnum != rnp_root->completed) { |
Paul E. McKenney | 09c7b89 | 2013-02-08 15:55:02 -0800 | [diff] [blame] | 2379 | trace_rcu_nocb_grace_period(rdp->rsp->name, |
| 2380 | rnp->gpnum, rnp->completed, |
| 2381 | c, rnp->level, |
| 2382 | rnp->grplo, rnp->grphi, |
| 2383 | "Startedleafroot"); |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2384 | raw_spin_unlock(&rnp_root->lock); /* irqs disabled. */ |
| 2385 | } else { |
| 2386 | |
| 2387 | /* |
| 2388 | * No grace period, so we need to start one. |
| 2389 | * The good news is that we can wait for exactly |
| 2390 | * one grace period instead of part of the current |
| 2391 | * grace period and all of the next grace period. |
| 2392 | * Adjust counters accordingly and start the |
| 2393 | * needed grace period. |
| 2394 | */ |
| 2395 | rnp->n_nocb_gp_requests[c & 0x1]--; |
| 2396 | c = rnp_root->completed + 1; |
| 2397 | rnp->n_nocb_gp_requests[c & 0x1]++; |
| 2398 | rnp_root->n_nocb_gp_requests[c & 0x1]++; |
Paul E. McKenney | 09c7b89 | 2013-02-08 15:55:02 -0800 | [diff] [blame] | 2399 | trace_rcu_nocb_grace_period(rdp->rsp->name, |
| 2400 | rnp->gpnum, rnp->completed, |
| 2401 | c, rnp->level, |
| 2402 | rnp->grplo, rnp->grphi, |
| 2403 | "Startedroot"); |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2404 | local_save_flags(flags1); |
| 2405 | rcu_start_gp(rdp->rsp, flags1); /* Rlses ->lock. */ |
| 2406 | } |
| 2407 | |
| 2408 | /* Clean up locking and irq state. */ |
| 2409 | if (rnp != rnp_root) |
| 2410 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
| 2411 | else |
| 2412 | local_irq_restore(flags); |
| 2413 | } |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2414 | |
| 2415 | /* |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2416 | * Wait for the grace period. Do so interruptibly to avoid messing |
| 2417 | * up the load average. |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2418 | */ |
Paul E. McKenney | 09c7b89 | 2013-02-08 15:55:02 -0800 | [diff] [blame] | 2419 | trace_rcu_nocb_grace_period(rdp->rsp->name, rnp->gpnum, rnp->completed, |
| 2420 | c, rnp->level, rnp->grplo, rnp->grphi, |
| 2421 | "StartWait"); |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2422 | for (;;) { |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2423 | wait_event_interruptible( |
| 2424 | rnp->nocb_gp_wq[c & 0x1], |
| 2425 | (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c))); |
| 2426 | if (likely(d)) |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2427 | break; |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2428 | flush_signals(current); |
Paul E. McKenney | 09c7b89 | 2013-02-08 15:55:02 -0800 | [diff] [blame] | 2429 | trace_rcu_nocb_grace_period(rdp->rsp->name, |
| 2430 | rnp->gpnum, rnp->completed, c, |
| 2431 | rnp->level, rnp->grplo, rnp->grphi, |
| 2432 | "ResumeWait"); |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2433 | } |
Paul E. McKenney | 09c7b89 | 2013-02-08 15:55:02 -0800 | [diff] [blame] | 2434 | trace_rcu_nocb_grace_period(rdp->rsp->name, rnp->gpnum, rnp->completed, |
| 2435 | c, rnp->level, rnp->grplo, rnp->grphi, |
| 2436 | "EndWait"); |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2437 | smp_mb(); /* Ensure that CB invocation happens after GP end. */ |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2438 | } |
| 2439 | |
| 2440 | /* |
| 2441 | * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes |
| 2442 | * callbacks queued by the corresponding no-CBs CPU. |
| 2443 | */ |
| 2444 | static int rcu_nocb_kthread(void *arg) |
| 2445 | { |
| 2446 | int c, cl; |
| 2447 | struct rcu_head *list; |
| 2448 | struct rcu_head *next; |
| 2449 | struct rcu_head **tail; |
| 2450 | struct rcu_data *rdp = arg; |
| 2451 | |
| 2452 | /* Each pass through this loop invokes one batch of callbacks */ |
| 2453 | for (;;) { |
| 2454 | /* If not polling, wait for next batch of callbacks. */ |
| 2455 | if (!rcu_nocb_poll) |
Paul Gortmaker | 353af9c | 2012-12-20 09:35:02 -0800 | [diff] [blame] | 2456 | wait_event_interruptible(rdp->nocb_wq, rdp->nocb_head); |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2457 | list = ACCESS_ONCE(rdp->nocb_head); |
| 2458 | if (!list) { |
| 2459 | schedule_timeout_interruptible(1); |
Paul Gortmaker | 353af9c | 2012-12-20 09:35:02 -0800 | [diff] [blame] | 2460 | flush_signals(current); |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2461 | continue; |
| 2462 | } |
| 2463 | |
| 2464 | /* |
| 2465 | * Extract queued callbacks, update counts, and wait |
| 2466 | * for a grace period to elapse. |
| 2467 | */ |
| 2468 | ACCESS_ONCE(rdp->nocb_head) = NULL; |
| 2469 | tail = xchg(&rdp->nocb_tail, &rdp->nocb_head); |
| 2470 | c = atomic_long_xchg(&rdp->nocb_q_count, 0); |
| 2471 | cl = atomic_long_xchg(&rdp->nocb_q_count_lazy, 0); |
| 2472 | ACCESS_ONCE(rdp->nocb_p_count) += c; |
| 2473 | ACCESS_ONCE(rdp->nocb_p_count_lazy) += cl; |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2474 | rcu_nocb_wait_gp(rdp); |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2475 | |
| 2476 | /* Each pass through the following loop invokes a callback. */ |
| 2477 | trace_rcu_batch_start(rdp->rsp->name, cl, c, -1); |
| 2478 | c = cl = 0; |
| 2479 | while (list) { |
| 2480 | next = list->next; |
| 2481 | /* Wait for enqueuing to complete, if needed. */ |
| 2482 | while (next == NULL && &list->next != tail) { |
| 2483 | schedule_timeout_interruptible(1); |
| 2484 | next = list->next; |
| 2485 | } |
| 2486 | debug_rcu_head_unqueue(list); |
| 2487 | local_bh_disable(); |
| 2488 | if (__rcu_reclaim(rdp->rsp->name, list)) |
| 2489 | cl++; |
| 2490 | c++; |
| 2491 | local_bh_enable(); |
| 2492 | list = next; |
| 2493 | } |
| 2494 | trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); |
| 2495 | ACCESS_ONCE(rdp->nocb_p_count) -= c; |
| 2496 | ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl; |
Paul E. McKenney | c635a4e1 | 2012-10-29 07:29:20 -0700 | [diff] [blame] | 2497 | rdp->n_nocbs_invoked += c; |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2498 | } |
| 2499 | return 0; |
| 2500 | } |
| 2501 | |
| 2502 | /* Initialize per-rcu_data variables for no-CBs CPUs. */ |
| 2503 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
| 2504 | { |
| 2505 | rdp->nocb_tail = &rdp->nocb_head; |
| 2506 | init_waitqueue_head(&rdp->nocb_wq); |
| 2507 | } |
| 2508 | |
| 2509 | /* Create a kthread for each RCU flavor for each no-CBs CPU. */ |
| 2510 | static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp) |
| 2511 | { |
| 2512 | int cpu; |
| 2513 | struct rcu_data *rdp; |
| 2514 | struct task_struct *t; |
| 2515 | |
| 2516 | if (rcu_nocb_mask == NULL) |
| 2517 | return; |
| 2518 | for_each_cpu(cpu, rcu_nocb_mask) { |
| 2519 | rdp = per_cpu_ptr(rsp->rda, cpu); |
Paul E. McKenney | a488985 | 2012-12-03 08:16:28 -0800 | [diff] [blame^] | 2520 | t = kthread_run(rcu_nocb_kthread, rdp, |
| 2521 | "rcuo%c/%d", rsp->abbr, cpu); |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2522 | BUG_ON(IS_ERR(t)); |
| 2523 | ACCESS_ONCE(rdp->nocb_kthread) = t; |
| 2524 | } |
| 2525 | } |
| 2526 | |
| 2527 | /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */ |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2528 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2529 | { |
| 2530 | if (rcu_nocb_mask == NULL || |
| 2531 | !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask)) |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2532 | return false; |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2533 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2534 | return true; |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2535 | } |
| 2536 | |
| 2537 | #else /* #ifdef CONFIG_RCU_NOCB_CPU */ |
| 2538 | |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2539 | static int rcu_nocb_needs_gp(struct rcu_state *rsp) |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2540 | { |
| 2541 | return 0; |
| 2542 | } |
| 2543 | |
Paul E. McKenney | dae6e64 | 2013-02-10 20:48:58 -0800 | [diff] [blame] | 2544 | static int rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
| 2545 | { |
| 2546 | return 0; |
| 2547 | } |
| 2548 | |
| 2549 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
| 2550 | { |
| 2551 | } |
| 2552 | |
| 2553 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
| 2554 | { |
| 2555 | } |
| 2556 | |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2557 | static bool is_nocb_cpu(int cpu) |
| 2558 | { |
| 2559 | return false; |
| 2560 | } |
| 2561 | |
| 2562 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
| 2563 | bool lazy) |
| 2564 | { |
| 2565 | return 0; |
| 2566 | } |
| 2567 | |
| 2568 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, |
| 2569 | struct rcu_data *rdp) |
| 2570 | { |
| 2571 | return 0; |
| 2572 | } |
| 2573 | |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2574 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
| 2575 | { |
| 2576 | } |
| 2577 | |
| 2578 | static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp) |
| 2579 | { |
| 2580 | } |
| 2581 | |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2582 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2583 | { |
Paul E. McKenney | 34ed6246 | 2013-01-07 13:37:42 -0800 | [diff] [blame] | 2584 | return false; |
Paul E. McKenney | 3fbfbf7 | 2012-08-19 21:35:53 -0700 | [diff] [blame] | 2585 | } |
| 2586 | |
| 2587 | #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */ |