kernel/rcutiny.c

   1 /*
   2  * Read-Copy Update mechanism for mutual exclusion, the Bloatwatch edition.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17  *
  18  * Copyright IBM Corporation, 2008
  19  *
  20  * Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  21  *
  22  * For detailed explanation of Read-Copy Update mechanism see -
  23  *              Documentation/RCU
  24  */
  25 #include <linux/completion.h>
  26 #include <linux/interrupt.h>
  27 #include <linux/notifier.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/kernel.h>
  30 #include <linux/export.h>
  31 #include <linux/mutex.h>
  32 #include <linux/sched.h>
  33 #include <linux/types.h>
  34 #include <linux/init.h>
  35 #include <linux/time.h>
  36 #include <linux/cpu.h>
  37 #include <linux/prefetch.h>
  38
  39 #ifdef CONFIG_RCU_TRACE
  40 #include <trace/events/rcu.h>
  41 #endif /* #else #ifdef CONFIG_RCU_TRACE */
  42
  43 #include "rcu.h"
  44
  45 /* Forward declarations for rcutiny_plugin.h. */
  46 struct rcu_ctrlblk;
  47 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp);
  48 static void rcu_process_callbacks(struct softirq_action *unused);
  49 static void __call_rcu(struct rcu_head *head,
  50                        void (*func)(struct rcu_head *rcu),
  51                        struct rcu_ctrlblk *rcp);
  52
  53 static long long rcu_dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
  54
  55 #include "rcutiny_plugin.h"
  56
  57 /* Common code for rcu_idle_enter() and rcu_irq_exit(), see kernel/rcutree.c. */
  58 static void rcu_idle_enter_common(long long newval)
  59 {
  60         if (newval) {
  61                 RCU_TRACE(trace_rcu_dyntick("--=",
  62                                             rcu_dynticks_nesting, newval));
  63                 rcu_dynticks_nesting = newval;
  64                 return;
  65         }
  66         RCU_TRACE(trace_rcu_dyntick("Start", rcu_dynticks_nesting, newval));
  67         if (!is_idle_task(current)) {
  68                 struct task_struct *idle = idle_task(smp_processor_id());
  69
  70                 RCU_TRACE(trace_rcu_dyntick("Error on entry: not idle task",
  71                                             rcu_dynticks_nesting, newval));
  72                 ftrace_dump(DUMP_ALL);
  73                 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
  74                           current->pid, current->comm,
  75                           idle->pid, idle->comm); /* must be idle task! */
  76         }
  77         rcu_sched_qs(0); /* implies rcu_bh_qsctr_inc(0) */
  78         barrier();
  79         rcu_dynticks_nesting = newval;
  80 }
  81
  82 /*
  83  * Enter idle, which is an extended quiescent state if we have fully
  84  * entered that mode (i.e., if the new value of dynticks_nesting is zero).
  85  */
  86 void rcu_idle_enter(void)
  87 {
  88         unsigned long flags;
  89         long long newval;
  90
  91         local_irq_save(flags);
  92         WARN_ON_ONCE((rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK) == 0);
  93         if ((rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK) ==
  94             DYNTICK_TASK_NEST_VALUE)
  95                 newval = 0;
  96         else
  97                 newval = rcu_dynticks_nesting - DYNTICK_TASK_NEST_VALUE;
  98         rcu_idle_enter_common(newval);
  99         local_irq_restore(flags);
 100 }
 101 EXPORT_SYMBOL_GPL(rcu_idle_enter);
 102
 103 /*
 104  * Exit an interrupt handler towards idle.
 105  */
 106 void rcu_irq_exit(void)
 107 {
 108         unsigned long flags;
 109         long long newval;
 110
 111         local_irq_save(flags);
 112         newval = rcu_dynticks_nesting - 1;
 113         WARN_ON_ONCE(newval < 0);
 114         rcu_idle_enter_common(newval);
 115         local_irq_restore(flags);
 116 }
 117 EXPORT_SYMBOL_GPL(rcu_irq_exit);
 118
 119 /* Common code for rcu_idle_exit() and rcu_irq_enter(), see kernel/rcutree.c. */
 120 static void rcu_idle_exit_common(long long oldval)
 121 {
 122         if (oldval) {
 123                 RCU_TRACE(trace_rcu_dyntick("++=",
 124                                             oldval, rcu_dynticks_nesting));
 125                 return;
 126         }
 127         RCU_TRACE(trace_rcu_dyntick("End", oldval, rcu_dynticks_nesting));
 128         if (!is_idle_task(current)) {
 129                 struct task_struct *idle = idle_task(smp_processor_id());
 130
 131                 RCU_TRACE(trace_rcu_dyntick("Error on exit: not idle task",
 132                           oldval, rcu_dynticks_nesting));
 133                 ftrace_dump(DUMP_ALL);
 134                 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
 135                           current->pid, current->comm,
 136                           idle->pid, idle->comm); /* must be idle task! */
 137         }
 138 }
 139
 140 /*
 141  * Exit idle, so that we are no longer in an extended quiescent state.
 142  */
 143 void rcu_idle_exit(void)
 144 {
 145         unsigned long flags;
 146         long long oldval;
 147
 148         local_irq_save(flags);
 149         oldval = rcu_dynticks_nesting;
 150         WARN_ON_ONCE(rcu_dynticks_nesting < 0);
 151         if (rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK)
 152                 rcu_dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
 153         else
 154                 rcu_dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
 155         rcu_idle_exit_common(oldval);
 156         local_irq_restore(flags);
 157 }
 158 EXPORT_SYMBOL_GPL(rcu_idle_exit);
 159
 160 /*
 161  * Enter an interrupt handler, moving away from idle.
 162  */
 163 void rcu_irq_enter(void)
 164 {
 165         unsigned long flags;
 166         long long oldval;
 167
 168         local_irq_save(flags);
 169         oldval = rcu_dynticks_nesting;
 170         rcu_dynticks_nesting++;
 171         WARN_ON_ONCE(rcu_dynticks_nesting == 0);
 172         rcu_idle_exit_common(oldval);
 173         local_irq_restore(flags);
 174 }
 175 EXPORT_SYMBOL_GPL(rcu_irq_enter);
 176
 177 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 178
 179 /*
 180  * Test whether RCU thinks that the current CPU is idle.
 181  */
 182 int rcu_is_cpu_idle(void)
 183 {
 184         return !rcu_dynticks_nesting;
 185 }
 186 EXPORT_SYMBOL(rcu_is_cpu_idle);
 187
 188 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 189
 190 /*
 191  * Test whether the current CPU was interrupted from idle.  Nested
 192  * interrupts don't count, we must be running at the first interrupt
 193  * level.
 194  */
 195 static int rcu_is_cpu_rrupt_from_idle(void)
 196 {
 197         return rcu_dynticks_nesting <= 1;
 198 }
 199
 200 /*
 201  * Helper function for rcu_sched_qs() and rcu_bh_qs().
 202  * Also irqs are disabled to avoid confusion due to interrupt handlers
 203  * invoking call_rcu().
 204  */
 205 static int rcu_qsctr_help(struct rcu_ctrlblk *rcp)
 206 {
 207         RCU_TRACE(reset_cpu_stall_ticks(rcp));
 208         if (rcp->rcucblist != NULL &&
 209             rcp->donetail != rcp->curtail) {
 210                 rcp->donetail = rcp->curtail;
 211                 return 1;
 212         }
 213
 214         return 0;
 215 }
 216
 217 /*
 218  * Record an rcu quiescent state.  And an rcu_bh quiescent state while we
 219  * are at it, given that any rcu quiescent state is also an rcu_bh
 220  * quiescent state.  Use "+" instead of "||" to defeat short circuiting.
 221  */
 222 void rcu_sched_qs(int cpu)
 223 {
 224         unsigned long flags;
 225
 226         local_irq_save(flags);
 227         if (rcu_qsctr_help(&rcu_sched_ctrlblk) +
 228             rcu_qsctr_help(&rcu_bh_ctrlblk))
 229                 raise_softirq(RCU_SOFTIRQ);
 230         local_irq_restore(flags);
 231 }
 232
 233 /*
 234  * Record an rcu_bh quiescent state.
 235  */
 236 void rcu_bh_qs(int cpu)
 237 {
 238         unsigned long flags;
 239
 240         local_irq_save(flags);
 241         if (rcu_qsctr_help(&rcu_bh_ctrlblk))
 242                 raise_softirq(RCU_SOFTIRQ);
 243         local_irq_restore(flags);
 244 }
 245
 246 /*
 247  * Check to see if the scheduling-clock interrupt came from an extended
 248  * quiescent state, and, if so, tell RCU about it.  This function must
 249  * be called from hardirq context.  It is normally called from the
 250  * scheduling-clock interrupt.
 251  */
 252 void rcu_check_callbacks(int cpu, int user)
 253 {
 254         RCU_TRACE(check_cpu_stalls());
 255         if (user || rcu_is_cpu_rrupt_from_idle())
 256                 rcu_sched_qs(cpu);
 257         else if (!in_softirq())
 258                 rcu_bh_qs(cpu);
 259 }
 260
 261 /*
 262  * Invoke the RCU callbacks on the specified rcu_ctrlkblk structure
 263  * whose grace period has elapsed.
 264  */
 265 static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp)
 266 {
 267         char *rn = NULL;
 268         struct rcu_head *next, *list;
 269         unsigned long flags;
 270         RCU_TRACE(int cb_count = 0);
 271
 272         /* If no RCU callbacks ready to invoke, just return. */
 273         if (&rcp->rcucblist == rcp->donetail) {
 274                 RCU_TRACE(trace_rcu_batch_start(rcp->name, 0, 0, -1));
 275                 RCU_TRACE(trace_rcu_batch_end(rcp->name, 0,
 276                                               ACCESS_ONCE(rcp->rcucblist),
 277                                               need_resched(),
 278                                               is_idle_task(current),
 279                                               false));
 280                 return;
 281         }
 282
 283         /* Move the ready-to-invoke callbacks to a local list. */
 284         local_irq_save(flags);
 285         RCU_TRACE(trace_rcu_batch_start(rcp->name, 0, rcp->qlen, -1));
 286         list = rcp->rcucblist;
 287         rcp->rcucblist = *rcp->donetail;
 288         *rcp->donetail = NULL;
 289         if (rcp->curtail == rcp->donetail)
 290                 rcp->curtail = &rcp->rcucblist;
 291         rcp->donetail = &rcp->rcucblist;
 292         local_irq_restore(flags);
 293
 294         /* Invoke the callbacks on the local list. */
 295         RCU_TRACE(rn = rcp->name);
 296         while (list) {
 297                 next = list->next;
 298                 prefetch(next);
 299                 debug_rcu_head_unqueue(list);
 300                 local_bh_disable();
 301                 __rcu_reclaim(rn, list);
 302                 local_bh_enable();
 303                 list = next;
 304                 RCU_TRACE(cb_count++);
 305         }
 306         RCU_TRACE(rcu_trace_sub_qlen(rcp, cb_count));
 307         RCU_TRACE(trace_rcu_batch_end(rcp->name, cb_count, 0, need_resched(),
 308                                       is_idle_task(current),
 309                                       false));
 310 }
 311
 312 static void rcu_process_callbacks(struct softirq_action *unused)
 313 {
 314         __rcu_process_callbacks(&rcu_sched_ctrlblk);
 315         __rcu_process_callbacks(&rcu_bh_ctrlblk);
 316 }
 317
 318 /*
 319  * Wait for a grace period to elapse.  But it is illegal to invoke
 320  * synchronize_sched() from within an RCU read-side critical section.
 321  * Therefore, any legal call to synchronize_sched() is a quiescent
 322  * state, and so on a UP system, synchronize_sched() need do nothing.
 323  * Ditto for synchronize_rcu_bh().  (But Lai Jiangshan points out the
 324  * benefits of doing might_sleep() to reduce latency.)
 325  *
 326  * Cool, huh?  (Due to Josh Triplett.)
 327  *
 328  * But we want to make this a static inline later.  The cond_resched()
 329  * currently makes this problematic.
 330  */
 331 void synchronize_sched(void)
 332 {
 333         rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
 334                            !lock_is_held(&rcu_lock_map) &&
 335                            !lock_is_held(&rcu_sched_lock_map),
 336                            "Illegal synchronize_sched() in RCU read-side critical section");
 337         cond_resched();
 338 }
 339 EXPORT_SYMBOL_GPL(synchronize_sched);
 340
 341 /*
 342  * Helper function for call_rcu() and call_rcu_bh().
 343  */
 344 static void __call_rcu(struct rcu_head *head,
 345                        void (*func)(struct rcu_head *rcu),
 346                        struct rcu_ctrlblk *rcp)
 347 {
 348         unsigned long flags;
 349
 350         debug_rcu_head_queue(head);
 351         head->func = func;
 352         head->next = NULL;
 353
 354         local_irq_save(flags);
 355         *rcp->curtail = head;
 356         rcp->curtail = &head->next;
 357         RCU_TRACE(rcp->qlen++);
 358         local_irq_restore(flags);
 359 }
 360
 361 /*
 362  * Post an RCU callback to be invoked after the end of an RCU-sched grace
 363  * period.  But since we have but one CPU, that would be after any
 364  * quiescent state.
 365  */
 366 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 367 {
 368         __call_rcu(head, func, &rcu_sched_ctrlblk);
 369 }
 370 EXPORT_SYMBOL_GPL(call_rcu_sched);
 371
 372 /*
 373  * Post an RCU bottom-half callback to be invoked after any subsequent
 374  * quiescent state.
 375  */
 376 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 377 {
 378         __call_rcu(head, func, &rcu_bh_ctrlblk);
 379 }
 380 EXPORT_SYMBOL_GPL(call_rcu_bh);
 381
 382 void rcu_init(void)
 383 {
 384         open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
 385 }