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