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