kernel/rcutiny_plugin.h

   1 /*
   2  * Read-Copy Update mechanism for mutual exclusion, the Bloatwatch edition
   3  * Internal non-public definitions that provide either classic
   4  * or preemptible semantics.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  19  *
  20  * Copyright (c) 2010 Linaro
  21  *
  22  * Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  23  */
  24
  25 #include <linux/kthread.h>
  26 #include <linux/debugfs.h>
  27 #include <linux/seq_file.h>
  28
  29 #ifdef CONFIG_RCU_TRACE
  30 #define RCU_TRACE(stmt) stmt
  31 #else /* #ifdef CONFIG_RCU_TRACE */
  32 #define RCU_TRACE(stmt)
  33 #endif /* #else #ifdef CONFIG_RCU_TRACE */
  34
  35 /* Global control variables for rcupdate callback mechanism. */
  36 struct rcu_ctrlblk {
  37         struct rcu_head *rcucblist;     /* List of pending callbacks (CBs). */
  38         struct rcu_head **donetail;     /* ->next pointer of last "done" CB. */
  39         struct rcu_head **curtail;      /* ->next pointer of last CB. */
  40         RCU_TRACE(long qlen);           /* Number of pending CBs. */
  41 };
  42
  43 /* Definition for rcupdate control block. */
  44 static struct rcu_ctrlblk rcu_sched_ctrlblk = {
  45         .donetail       = &rcu_sched_ctrlblk.rcucblist,
  46         .curtail        = &rcu_sched_ctrlblk.rcucblist,
  47 };
  48
  49 static struct rcu_ctrlblk rcu_bh_ctrlblk = {
  50         .donetail       = &rcu_bh_ctrlblk.rcucblist,
  51         .curtail        = &rcu_bh_ctrlblk.rcucblist,
  52 };
  53
  54 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  55 int rcu_scheduler_active __read_mostly;
  56 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
  57 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
  58
  59 #ifdef CONFIG_TINY_PREEMPT_RCU
  60
  61 #include <linux/delay.h>
  62
  63 /* Global control variables for preemptible RCU. */
  64 struct rcu_preempt_ctrlblk {
  65         struct rcu_ctrlblk rcb; /* curtail: ->next ptr of last CB for GP. */
  66         struct rcu_head **nexttail;
  67                                 /* Tasks blocked in a preemptible RCU */
  68                                 /*  read-side critical section while an */
  69                                 /*  preemptible-RCU grace period is in */
  70                                 /*  progress must wait for a later grace */
  71                                 /*  period.  This pointer points to the */
  72                                 /*  ->next pointer of the last task that */
  73                                 /*  must wait for a later grace period, or */
  74                                 /*  to &->rcb.rcucblist if there is no */
  75                                 /*  such task. */
  76         struct list_head blkd_tasks;
  77                                 /* Tasks blocked in RCU read-side critical */
  78                                 /*  section.  Tasks are placed at the head */
  79                                 /*  of this list and age towards the tail. */
  80         struct list_head *gp_tasks;
  81                                 /* Pointer to the first task blocking the */
  82                                 /*  current grace period, or NULL if there */
  83                                 /*  is no such task. */
  84         struct list_head *exp_tasks;
  85                                 /* Pointer to first task blocking the */
  86                                 /*  current expedited grace period, or NULL */
  87                                 /*  if there is no such task.  If there */
  88                                 /*  is no current expedited grace period, */
  89                                 /*  then there cannot be any such task. */
  90 #ifdef CONFIG_RCU_BOOST
  91         struct list_head *boost_tasks;
  92                                 /* Pointer to first task that needs to be */
  93                                 /*  priority-boosted, or NULL if no priority */
  94                                 /*  boosting is needed.  If there is no */
  95                                 /*  current or expedited grace period, there */
  96                                 /*  can be no such task. */
  97 #endif /* #ifdef CONFIG_RCU_BOOST */
  98         u8 gpnum;               /* Current grace period. */
  99         u8 gpcpu;               /* Last grace period blocked by the CPU. */
 100         u8 completed;           /* Last grace period completed. */
 101                                 /*  If all three are equal, RCU is idle. */
 102 #ifdef CONFIG_RCU_BOOST
 103         s8 boosted_this_gp;     /* Has boosting already happened? */
 104         unsigned long boost_time; /* When to start boosting (jiffies) */
 105 #endif /* #ifdef CONFIG_RCU_BOOST */
 106 #ifdef CONFIG_RCU_TRACE
 107         unsigned long n_grace_periods;
 108 #ifdef CONFIG_RCU_BOOST
 109         unsigned long n_tasks_boosted;
 110         unsigned long n_exp_boosts;
 111         unsigned long n_normal_boosts;
 112         unsigned long n_normal_balk_blkd_tasks;
 113         unsigned long n_normal_balk_gp_tasks;
 114         unsigned long n_normal_balk_boost_tasks;
 115         unsigned long n_normal_balk_boosted;
 116         unsigned long n_normal_balk_notyet;
 117         unsigned long n_normal_balk_nos;
 118         unsigned long n_exp_balk_blkd_tasks;
 119         unsigned long n_exp_balk_nos;
 120 #endif /* #ifdef CONFIG_RCU_BOOST */
 121 #endif /* #ifdef CONFIG_RCU_TRACE */
 122 };
 123
 124 static struct rcu_preempt_ctrlblk rcu_preempt_ctrlblk = {
 125         .rcb.donetail = &rcu_preempt_ctrlblk.rcb.rcucblist,
 126         .rcb.curtail = &rcu_preempt_ctrlblk.rcb.rcucblist,
 127         .nexttail = &rcu_preempt_ctrlblk.rcb.rcucblist,
 128         .blkd_tasks = LIST_HEAD_INIT(rcu_preempt_ctrlblk.blkd_tasks),
 129 };
 130
 131 static int rcu_preempted_readers_exp(void);
 132 static void rcu_report_exp_done(void);
 133
 134 /*
 135  * Return true if the CPU has not yet responded to the current grace period.
 136  */
 137 static int rcu_cpu_blocking_cur_gp(void)
 138 {
 139         return rcu_preempt_ctrlblk.gpcpu != rcu_preempt_ctrlblk.gpnum;
 140 }
 141
 142 /*
 143  * Check for a running RCU reader.  Because there is only one CPU,
 144  * there can be but one running RCU reader at a time.  ;-)
 145  */
 146 static int rcu_preempt_running_reader(void)
 147 {
 148         return current->rcu_read_lock_nesting;
 149 }
 150
 151 /*
 152  * Check for preempted RCU readers blocking any grace period.
 153  * If the caller needs a reliable answer, it must disable hard irqs.
 154  */
 155 static int rcu_preempt_blocked_readers_any(void)
 156 {
 157         return !list_empty(&rcu_preempt_ctrlblk.blkd_tasks);
 158 }
 159
 160 /*
 161  * Check for preempted RCU readers blocking the current grace period.
 162  * If the caller needs a reliable answer, it must disable hard irqs.
 163  */
 164 static int rcu_preempt_blocked_readers_cgp(void)
 165 {
 166         return rcu_preempt_ctrlblk.gp_tasks != NULL;
 167 }
 168
 169 /*
 170  * Return true if another preemptible-RCU grace period is needed.
 171  */
 172 static int rcu_preempt_needs_another_gp(void)
 173 {
 174         return *rcu_preempt_ctrlblk.rcb.curtail != NULL;
 175 }
 176
 177 /*
 178  * Return true if a preemptible-RCU grace period is in progress.
 179  * The caller must disable hardirqs.
 180  */
 181 static int rcu_preempt_gp_in_progress(void)
 182 {
 183         return rcu_preempt_ctrlblk.completed != rcu_preempt_ctrlblk.gpnum;
 184 }
 185
 186 /*
 187  * Advance a ->blkd_tasks-list pointer to the next entry, instead
 188  * returning NULL if at the end of the list.
 189  */
 190 static struct list_head *rcu_next_node_entry(struct task_struct *t)
 191 {
 192         struct list_head *np;
 193
 194         np = t->rcu_node_entry.next;
 195         if (np == &rcu_preempt_ctrlblk.blkd_tasks)
 196                 np = NULL;
 197         return np;
 198 }
 199
 200 #ifdef CONFIG_RCU_TRACE
 201
 202 #ifdef CONFIG_RCU_BOOST
 203 static void rcu_initiate_boost_trace(void);
 204 static void rcu_initiate_exp_boost_trace(void);
 205 #endif /* #ifdef CONFIG_RCU_BOOST */
 206
 207 /*
 208  * Dump additional statistice for TINY_PREEMPT_RCU.
 209  */
 210 static void show_tiny_preempt_stats(struct seq_file *m)
 211 {
 212         seq_printf(m, "rcu_preempt: qlen=%ld gp=%lu g%u/p%u/c%u tasks=%c%c%c\n",
 213                    rcu_preempt_ctrlblk.rcb.qlen,
 214                    rcu_preempt_ctrlblk.n_grace_periods,
 215                    rcu_preempt_ctrlblk.gpnum,
 216                    rcu_preempt_ctrlblk.gpcpu,
 217                    rcu_preempt_ctrlblk.completed,
 218                    "T."[list_empty(&rcu_preempt_ctrlblk.blkd_tasks)],
 219                    "N."[!rcu_preempt_ctrlblk.gp_tasks],
 220                    "E."[!rcu_preempt_ctrlblk.exp_tasks]);
 221 #ifdef CONFIG_RCU_BOOST
 222         seq_printf(m, "             ttb=%c btg=",
 223                    "B."[!rcu_preempt_ctrlblk.boost_tasks]);
 224         switch (rcu_preempt_ctrlblk.boosted_this_gp) {
 225         case -1:
 226                 seq_puts(m, "exp");
 227                 break;
 228         case 0:
 229                 seq_puts(m, "no");
 230                 break;
 231         case 1:
 232                 seq_puts(m, "begun");
 233                 break;
 234         case 2:
 235                 seq_puts(m, "done");
 236                 break;
 237         default:
 238                 seq_printf(m, "?%d?", rcu_preempt_ctrlblk.boosted_this_gp);
 239         }
 240         seq_printf(m, " ntb=%lu neb=%lu nnb=%lu j=%04x bt=%04x\n",
 241                    rcu_preempt_ctrlblk.n_tasks_boosted,
 242                    rcu_preempt_ctrlblk.n_exp_boosts,
 243                    rcu_preempt_ctrlblk.n_normal_boosts,
 244                    (int)(jiffies & 0xffff),
 245                    (int)(rcu_preempt_ctrlblk.boost_time & 0xffff));
 246         seq_printf(m, "             %s: nt=%lu gt=%lu bt=%lu b=%lu ny=%lu nos=%lu\n",
 247                    "normal balk",
 248                    rcu_preempt_ctrlblk.n_normal_balk_blkd_tasks,
 249                    rcu_preempt_ctrlblk.n_normal_balk_gp_tasks,
 250                    rcu_preempt_ctrlblk.n_normal_balk_boost_tasks,
 251                    rcu_preempt_ctrlblk.n_normal_balk_boosted,
 252                    rcu_preempt_ctrlblk.n_normal_balk_notyet,
 253                    rcu_preempt_ctrlblk.n_normal_balk_nos);
 254         seq_printf(m, "             exp balk: bt=%lu nos=%lu\n",
 255                    rcu_preempt_ctrlblk.n_exp_balk_blkd_tasks,
 256                    rcu_preempt_ctrlblk.n_exp_balk_nos);
 257 #endif /* #ifdef CONFIG_RCU_BOOST */
 258 }
 259
 260 #endif /* #ifdef CONFIG_RCU_TRACE */
 261
 262 #ifdef CONFIG_RCU_BOOST
 263
 264 #include "rtmutex_common.h"
 265
 266 /*
 267  * Carry out RCU priority boosting on the task indicated by ->boost_tasks,
 268  * and advance ->boost_tasks to the next task in the ->blkd_tasks list.
 269  */
 270 static int rcu_boost(void)
 271 {
 272         unsigned long flags;
 273         struct rt_mutex mtx;
 274         struct list_head *np;
 275         struct task_struct *t;
 276
 277         if (rcu_preempt_ctrlblk.boost_tasks == NULL)
 278                 return 0;  /* Nothing to boost. */
 279         raw_local_irq_save(flags);
 280         rcu_preempt_ctrlblk.boosted_this_gp++;
 281         t = container_of(rcu_preempt_ctrlblk.boost_tasks, struct task_struct,
 282                          rcu_node_entry);
 283         np = rcu_next_node_entry(t);
 284         rt_mutex_init_proxy_locked(&mtx, t);
 285         t->rcu_boost_mutex = &mtx;
 286         t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BOOSTED;
 287         raw_local_irq_restore(flags);
 288         rt_mutex_lock(&mtx);
 289         RCU_TRACE(rcu_preempt_ctrlblk.n_tasks_boosted++);
 290         rcu_preempt_ctrlblk.boosted_this_gp++;
 291         rt_mutex_unlock(&mtx);
 292         return rcu_preempt_ctrlblk.boost_tasks != NULL;
 293 }
 294
 295 /*
 296  * Check to see if it is now time to start boosting RCU readers blocking
 297  * the current grace period, and, if so, tell the rcu_kthread_task to
 298  * start boosting them.  If there is an expedited boost in progress,
 299  * we wait for it to complete.
 300  *
 301  * If there are no blocked readers blocking the current grace period,
 302  * return 0 to let the caller know, otherwise return 1.  Note that this
 303  * return value is independent of whether or not boosting was done.
 304  */
 305 static int rcu_initiate_boost(void)
 306 {
 307         if (!rcu_preempt_blocked_readers_cgp()) {
 308                 RCU_TRACE(rcu_preempt_ctrlblk.n_normal_balk_blkd_tasks++);
 309                 return 0;
 310         }
 311         if (rcu_preempt_ctrlblk.gp_tasks != NULL &&
 312             rcu_preempt_ctrlblk.boost_tasks == NULL &&
 313             rcu_preempt_ctrlblk.boosted_this_gp == 0 &&
 314             ULONG_CMP_GE(jiffies, rcu_preempt_ctrlblk.boost_time)) {
 315                 rcu_preempt_ctrlblk.boost_tasks = rcu_preempt_ctrlblk.gp_tasks;
 316                 invoke_rcu_kthread();
 317                 RCU_TRACE(rcu_preempt_ctrlblk.n_normal_boosts++);
 318         } else
 319                 RCU_TRACE(rcu_initiate_boost_trace());
 320         return 1;
 321 }
 322
 323 /*
 324  * Initiate boosting for an expedited grace period.
 325  */
 326 static void rcu_initiate_expedited_boost(void)
 327 {
 328         unsigned long flags;
 329
 330         raw_local_irq_save(flags);
 331         if (!list_empty(&rcu_preempt_ctrlblk.blkd_tasks)) {
 332                 rcu_preempt_ctrlblk.boost_tasks =
 333                         rcu_preempt_ctrlblk.blkd_tasks.next;
 334                 rcu_preempt_ctrlblk.boosted_this_gp = -1;
 335                 invoke_rcu_kthread();
 336                 RCU_TRACE(rcu_preempt_ctrlblk.n_exp_boosts++);
 337         } else
 338                 RCU_TRACE(rcu_initiate_exp_boost_trace());
 339         raw_local_irq_restore(flags);
 340 }
 341
 342 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000);
 343
 344 /*
 345  * Do priority-boost accounting for the start of a new grace period.
 346  */
 347 static void rcu_preempt_boost_start_gp(void)
 348 {
 349         rcu_preempt_ctrlblk.boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
 350         if (rcu_preempt_ctrlblk.boosted_this_gp > 0)
 351                 rcu_preempt_ctrlblk.boosted_this_gp = 0;
 352 }
 353
 354 #else /* #ifdef CONFIG_RCU_BOOST */
 355
 356 /*
 357  * If there is no RCU priority boosting, we don't boost.
 358  */
 359 static int rcu_boost(void)
 360 {
 361         return 0;
 362 }
 363
 364 /*
 365  * If there is no RCU priority boosting, we don't initiate boosting,
 366  * but we do indicate whether there are blocked readers blocking the
 367  * current grace period.
 368  */
 369 static int rcu_initiate_boost(void)
 370 {
 371         return rcu_preempt_blocked_readers_cgp();
 372 }
 373
 374 /*
 375  * If there is no RCU priority boosting, we don't initiate expedited boosting.
 376  */
 377 static void rcu_initiate_expedited_boost(void)
 378 {
 379 }
 380
 381 /*
 382  * If there is no RCU priority boosting, nothing to do at grace-period start.
 383  */
 384 static void rcu_preempt_boost_start_gp(void)
 385 {
 386 }
 387
 388 #endif /* else #ifdef CONFIG_RCU_BOOST */
 389
 390 /*
 391  * Record a preemptible-RCU quiescent state for the specified CPU.  Note
 392  * that this just means that the task currently running on the CPU is
 393  * in a quiescent state.  There might be any number of tasks blocked
 394  * while in an RCU read-side critical section.
 395  *
 396  * Unlike the other rcu_*_qs() functions, callers to this function
 397  * must disable irqs in order to protect the assignment to
 398  * ->rcu_read_unlock_special.
 399  *
 400  * Because this is a single-CPU implementation, the only way a grace
 401  * period can end is if the CPU is in a quiescent state.  The reason is
 402  * that a blocked preemptible-RCU reader can exit its critical section
 403  * only if the CPU is running it at the time.  Therefore, when the
 404  * last task blocking the current grace period exits its RCU read-side
 405  * critical section, neither the CPU nor blocked tasks will be stopping
 406  * the current grace period.  (In contrast, SMP implementations
 407  * might have CPUs running in RCU read-side critical sections that
 408  * block later grace periods -- but this is not possible given only
 409  * one CPU.)
 410  */
 411 static void rcu_preempt_cpu_qs(void)
 412 {
 413         /* Record both CPU and task as having responded to current GP. */
 414         rcu_preempt_ctrlblk.gpcpu = rcu_preempt_ctrlblk.gpnum;
 415         current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
 416
 417         /* If there is no GP then there is nothing more to do.  */
 418         if (!rcu_preempt_gp_in_progress())
 419                 return;
 420         /*
 421          * Check up on boosting.  If there are no readers blocking the
 422          * current grace period, leave.
 423          */
 424         if (rcu_initiate_boost())
 425                 return;
 426
 427         /* Advance callbacks. */
 428         rcu_preempt_ctrlblk.completed = rcu_preempt_ctrlblk.gpnum;
 429         rcu_preempt_ctrlblk.rcb.donetail = rcu_preempt_ctrlblk.rcb.curtail;
 430         rcu_preempt_ctrlblk.rcb.curtail = rcu_preempt_ctrlblk.nexttail;
 431
 432         /* If there are no blocked readers, next GP is done instantly. */
 433         if (!rcu_preempt_blocked_readers_any())
 434                 rcu_preempt_ctrlblk.rcb.donetail = rcu_preempt_ctrlblk.nexttail;
 435
 436         /* If there are done callbacks, cause them to be invoked. */
 437         if (*rcu_preempt_ctrlblk.rcb.donetail != NULL)
 438                 invoke_rcu_kthread();
 439 }
 440
 441 /*
 442  * Start a new RCU grace period if warranted.  Hard irqs must be disabled.
 443  */
 444 static void rcu_preempt_start_gp(void)
 445 {
 446         if (!rcu_preempt_gp_in_progress() && rcu_preempt_needs_another_gp()) {
 447
 448                 /* Official start of GP. */
 449                 rcu_preempt_ctrlblk.gpnum++;
 450                 RCU_TRACE(rcu_preempt_ctrlblk.n_grace_periods++);
 451
 452                 /* Any blocked RCU readers block new GP. */
 453                 if (rcu_preempt_blocked_readers_any())
 454                         rcu_preempt_ctrlblk.gp_tasks =
 455                                 rcu_preempt_ctrlblk.blkd_tasks.next;
 456
 457                 /* Set up for RCU priority boosting. */
 458                 rcu_preempt_boost_start_gp();
 459
 460                 /* If there is no running reader, CPU is done with GP. */
 461                 if (!rcu_preempt_running_reader())
 462                         rcu_preempt_cpu_qs();
 463         }
 464 }
 465
 466 /*
 467  * We have entered the scheduler, and the current task might soon be
 468  * context-switched away from.  If this task is in an RCU read-side
 469  * critical section, we will no longer be able to rely on the CPU to
 470  * record that fact, so we enqueue the task on the blkd_tasks list.
 471  * If the task started after the current grace period began, as recorded
 472  * by ->gpcpu, we enqueue at the beginning of the list.  Otherwise
 473  * before the element referenced by ->gp_tasks (or at the tail if
 474  * ->gp_tasks is NULL) and point ->gp_tasks at the newly added element.
 475  * The task will dequeue itself when it exits the outermost enclosing
 476  * RCU read-side critical section.  Therefore, the current grace period
 477  * cannot be permitted to complete until the ->gp_tasks pointer becomes
 478  * NULL.
 479  *
 480  * Caller must disable preemption.
 481  */
 482 void rcu_preempt_note_context_switch(void)
 483 {
 484         struct task_struct *t = current;
 485         unsigned long flags;
 486
 487         local_irq_save(flags); /* must exclude scheduler_tick(). */
 488         if (rcu_preempt_running_reader() &&
 489             (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
 490
 491                 /* Possibly blocking in an RCU read-side critical section. */
 492                 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
 493
 494                 /*
 495                  * If this CPU has already checked in, then this task
 496                  * will hold up the next grace period rather than the
 497                  * current grace period.  Queue the task accordingly.
 498                  * If the task is queued for the current grace period
 499                  * (i.e., this CPU has not yet passed through a quiescent
 500                  * state for the current grace period), then as long
 501                  * as that task remains queued, the current grace period
 502                  * cannot end.
 503                  */
 504                 list_add(&t->rcu_node_entry, &rcu_preempt_ctrlblk.blkd_tasks);
 505                 if (rcu_cpu_blocking_cur_gp())
 506                         rcu_preempt_ctrlblk.gp_tasks = &t->rcu_node_entry;
 507         }
 508
 509         /*
 510          * Either we were not in an RCU read-side critical section to
 511          * begin with, or we have now recorded that critical section
 512          * globally.  Either way, we can now note a quiescent state
 513          * for this CPU.  Again, if we were in an RCU read-side critical
 514          * section, and if that critical section was blocking the current
 515          * grace period, then the fact that the task has been enqueued
 516          * means that current grace period continues to be blocked.
 517          */
 518         rcu_preempt_cpu_qs();
 519         local_irq_restore(flags);
 520 }
 521
 522 /*
 523  * Tiny-preemptible RCU implementation for rcu_read_lock().
 524  * Just increment ->rcu_read_lock_nesting, shared state will be updated
 525  * if we block.
 526  */
 527 void __rcu_read_lock(void)
 528 {
 529         current->rcu_read_lock_nesting++;
 530         barrier();  /* needed if we ever invoke rcu_read_lock in rcutiny.c */
 531 }
 532 EXPORT_SYMBOL_GPL(__rcu_read_lock);
 533
 534 /*
 535  * Handle special cases during rcu_read_unlock(), such as needing to
 536  * notify RCU core processing or task having blocked during the RCU
 537  * read-side critical section.
 538  */
 539 static void rcu_read_unlock_special(struct task_struct *t)
 540 {
 541         int empty;
 542         int empty_exp;
 543         unsigned long flags;
 544         struct list_head *np;
 545         int special;
 546
 547         /*
 548          * NMI handlers cannot block and cannot safely manipulate state.
 549          * They therefore cannot possibly be special, so just leave.
 550          */
 551         if (in_nmi())
 552                 return;
 553
 554         local_irq_save(flags);
 555
 556         /*
 557          * If RCU core is waiting for this CPU to exit critical section,
 558          * let it know that we have done so.
 559          */
 560         special = t->rcu_read_unlock_special;
 561         if (special & RCU_READ_UNLOCK_NEED_QS)
 562                 rcu_preempt_cpu_qs();
 563
 564         /* Hardware IRQ handlers cannot block. */
 565         if (in_irq()) {
 566                 local_irq_restore(flags);
 567                 return;
 568         }
 569
 570         /* Clean up if blocked during RCU read-side critical section. */
 571         if (special & RCU_READ_UNLOCK_BLOCKED) {
 572                 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
 573
 574                 /*
 575                  * Remove this task from the ->blkd_tasks list and adjust
 576                  * any pointers that might have been referencing it.
 577                  */
 578                 empty = !rcu_preempt_blocked_readers_cgp();
 579                 empty_exp = rcu_preempt_ctrlblk.exp_tasks == NULL;
 580                 np = rcu_next_node_entry(t);
 581                 list_del(&t->rcu_node_entry);
 582                 if (&t->rcu_node_entry == rcu_preempt_ctrlblk.gp_tasks)
 583                         rcu_preempt_ctrlblk.gp_tasks = np;
 584                 if (&t->rcu_node_entry == rcu_preempt_ctrlblk.exp_tasks)
 585                         rcu_preempt_ctrlblk.exp_tasks = np;
 586 #ifdef CONFIG_RCU_BOOST
 587                 if (&t->rcu_node_entry == rcu_preempt_ctrlblk.boost_tasks)
 588                         rcu_preempt_ctrlblk.boost_tasks = np;
 589 #endif /* #ifdef CONFIG_RCU_BOOST */
 590                 INIT_LIST_HEAD(&t->rcu_node_entry);
 591
 592                 /*
 593                  * If this was the last task on the current list, and if
 594                  * we aren't waiting on the CPU, report the quiescent state
 595                  * and start a new grace period if needed.
 596                  */
 597                 if (!empty && !rcu_preempt_blocked_readers_cgp()) {
 598                         rcu_preempt_cpu_qs();
 599                         rcu_preempt_start_gp();
 600                 }
 601
 602                 /*
 603                  * If this was the last task on the expedited lists,
 604                  * then we need wake up the waiting task.
 605                  */
 606                 if (!empty_exp && rcu_preempt_ctrlblk.exp_tasks == NULL)
 607                         rcu_report_exp_done();
 608         }
 609 #ifdef CONFIG_RCU_BOOST
 610         /* Unboost self if was boosted. */
 611         if (special & RCU_READ_UNLOCK_BOOSTED) {
 612                 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BOOSTED;
 613                 rt_mutex_unlock(t->rcu_boost_mutex);
 614                 t->rcu_boost_mutex = NULL;
 615         }
 616 #endif /* #ifdef CONFIG_RCU_BOOST */
 617         local_irq_restore(flags);
 618 }
 619
 620 /*
 621  * Tiny-preemptible RCU implementation for rcu_read_unlock().
 622  * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
 623  * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
 624  * invoke rcu_read_unlock_special() to clean up after a context switch
 625  * in an RCU read-side critical section and other special cases.
 626  */
 627 void __rcu_read_unlock(void)
 628 {
 629         struct task_struct *t = current;
 630
 631         barrier();  /* needed if we ever invoke rcu_read_unlock in rcutiny.c */
 632         --t->rcu_read_lock_nesting;
 633         barrier();  /* decrement before load of ->rcu_read_unlock_special */
 634         if (t->rcu_read_lock_nesting == 0 &&
 635             unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
 636                 rcu_read_unlock_special(t);
 637 #ifdef CONFIG_PROVE_LOCKING
 638         WARN_ON_ONCE(t->rcu_read_lock_nesting < 0);
 639 #endif /* #ifdef CONFIG_PROVE_LOCKING */
 640 }
 641 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
 642
 643 /*
 644  * Check for a quiescent state from the current CPU.  When a task blocks,
 645  * the task is recorded in the rcu_preempt_ctrlblk structure, which is
 646  * checked elsewhere.  This is called from the scheduling-clock interrupt.
 647  *
 648  * Caller must disable hard irqs.
 649  */
 650 static void rcu_preempt_check_callbacks(void)
 651 {
 652         struct task_struct *t = current;
 653
 654         if (rcu_preempt_gp_in_progress() &&
 655             (!rcu_preempt_running_reader() ||
 656              !rcu_cpu_blocking_cur_gp()))
 657                 rcu_preempt_cpu_qs();
 658         if (&rcu_preempt_ctrlblk.rcb.rcucblist !=
 659             rcu_preempt_ctrlblk.rcb.donetail)
 660                 invoke_rcu_kthread();
 661         if (rcu_preempt_gp_in_progress() &&
 662             rcu_cpu_blocking_cur_gp() &&
 663             rcu_preempt_running_reader())
 664                 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
 665 }
 666
 667 /*
 668  * TINY_PREEMPT_RCU has an extra callback-list tail pointer to
 669  * update, so this is invoked from rcu_process_callbacks() to
 670  * handle that case.  Of course, it is invoked for all flavors of
 671  * RCU, but RCU callbacks can appear only on one of the lists, and
 672  * neither ->nexttail nor ->donetail can possibly be NULL, so there
 673  * is no need for an explicit check.
 674  */
 675 static void rcu_preempt_remove_callbacks(struct rcu_ctrlblk *rcp)
 676 {
 677         if (rcu_preempt_ctrlblk.nexttail == rcp->donetail)
 678                 rcu_preempt_ctrlblk.nexttail = &rcp->rcucblist;
 679 }
 680
 681 /*
 682  * Process callbacks for preemptible RCU.
 683  */
 684 static void rcu_preempt_process_callbacks(void)
 685 {
 686         rcu_process_callbacks(&rcu_preempt_ctrlblk.rcb);
 687 }
 688
 689 /*
 690  * Queue a preemptible -RCU callback for invocation after a grace period.
 691  */
 692 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 693 {
 694         unsigned long flags;
 695
 696         debug_rcu_head_queue(head);
 697         head->func = func;
 698         head->next = NULL;
 699
 700         local_irq_save(flags);
 701         *rcu_preempt_ctrlblk.nexttail = head;
 702         rcu_preempt_ctrlblk.nexttail = &head->next;
 703         RCU_TRACE(rcu_preempt_ctrlblk.rcb.qlen++);
 704         rcu_preempt_start_gp();  /* checks to see if GP needed. */
 705         local_irq_restore(flags);
 706 }
 707 EXPORT_SYMBOL_GPL(call_rcu);
 708
 709 void rcu_barrier(void)
 710 {
 711         struct rcu_synchronize rcu;
 712
 713         init_rcu_head_on_stack(&rcu.head);
 714         init_completion(&rcu.completion);
 715         /* Will wake me after RCU finished. */
 716         call_rcu(&rcu.head, wakeme_after_rcu);
 717         /* Wait for it. */
 718         wait_for_completion(&rcu.completion);
 719         destroy_rcu_head_on_stack(&rcu.head);
 720 }
 721 EXPORT_SYMBOL_GPL(rcu_barrier);
 722
 723 /*
 724  * synchronize_rcu - wait until a grace period has elapsed.
 725  *
 726  * Control will return to the caller some time after a full grace
 727  * period has elapsed, in other words after all currently executing RCU
 728  * read-side critical sections have completed.  RCU read-side critical
 729  * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
 730  * and may be nested.
 731  */
 732 void synchronize_rcu(void)
 733 {
 734 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 735         if (!rcu_scheduler_active)
 736                 return;
 737 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 738
 739         WARN_ON_ONCE(rcu_preempt_running_reader());
 740         if (!rcu_preempt_blocked_readers_any())
 741                 return;
 742
 743         /* Once we get past the fastpath checks, same code as rcu_barrier(). */
 744         rcu_barrier();
 745 }
 746 EXPORT_SYMBOL_GPL(synchronize_rcu);
 747
 748 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
 749 static unsigned long sync_rcu_preempt_exp_count;
 750 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
 751
 752 /*
 753  * Return non-zero if there are any tasks in RCU read-side critical
 754  * sections blocking the current preemptible-RCU expedited grace period.
 755  * If there is no preemptible-RCU expedited grace period currently in
 756  * progress, returns zero unconditionally.
 757  */
 758 static int rcu_preempted_readers_exp(void)
 759 {
 760         return rcu_preempt_ctrlblk.exp_tasks != NULL;
 761 }
 762
 763 /*
 764  * Report the exit from RCU read-side critical section for the last task
 765  * that queued itself during or before the current expedited preemptible-RCU
 766  * grace period.
 767  */
 768 static void rcu_report_exp_done(void)
 769 {
 770         wake_up(&sync_rcu_preempt_exp_wq);
 771 }
 772
 773 /*
 774  * Wait for an rcu-preempt grace period, but expedite it.  The basic idea
 775  * is to rely in the fact that there is but one CPU, and that it is
 776  * illegal for a task to invoke synchronize_rcu_expedited() while in a
 777  * preemptible-RCU read-side critical section.  Therefore, any such
 778  * critical sections must correspond to blocked tasks, which must therefore
 779  * be on the ->blkd_tasks list.  So just record the current head of the
 780  * list in the ->exp_tasks pointer, and wait for all tasks including and
 781  * after the task pointed to by ->exp_tasks to drain.
 782  */
 783 void synchronize_rcu_expedited(void)
 784 {
 785         unsigned long flags;
 786         struct rcu_preempt_ctrlblk *rpcp = &rcu_preempt_ctrlblk;
 787         unsigned long snap;
 788
 789         barrier(); /* ensure prior action seen before grace period. */
 790
 791         WARN_ON_ONCE(rcu_preempt_running_reader());
 792
 793         /*
 794          * Acquire lock so that there is only one preemptible RCU grace
 795          * period in flight.  Of course, if someone does the expedited
 796          * grace period for us while we are acquiring the lock, just leave.
 797          */
 798         snap = sync_rcu_preempt_exp_count + 1;
 799         mutex_lock(&sync_rcu_preempt_exp_mutex);
 800         if (ULONG_CMP_LT(snap, sync_rcu_preempt_exp_count))
 801                 goto unlock_mb_ret; /* Others did our work for us. */
 802
 803         local_irq_save(flags);
 804
 805         /*
 806          * All RCU readers have to already be on blkd_tasks because
 807          * we cannot legally be executing in an RCU read-side critical
 808          * section.
 809          */
 810
 811         /* Snapshot current head of ->blkd_tasks list. */
 812         rpcp->exp_tasks = rpcp->blkd_tasks.next;
 813         if (rpcp->exp_tasks == &rpcp->blkd_tasks)
 814                 rpcp->exp_tasks = NULL;
 815         local_irq_restore(flags);
 816
 817         /* Wait for tail of ->blkd_tasks list to drain. */
 818         if (rcu_preempted_readers_exp())
 819                 rcu_initiate_expedited_boost();
 820                 wait_event(sync_rcu_preempt_exp_wq,
 821                            !rcu_preempted_readers_exp());
 822
 823         /* Clean up and exit. */
 824         barrier(); /* ensure expedited GP seen before counter increment. */
 825         sync_rcu_preempt_exp_count++;
 826 unlock_mb_ret:
 827         mutex_unlock(&sync_rcu_preempt_exp_mutex);
 828         barrier(); /* ensure subsequent action seen after grace period. */
 829 }
 830 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 831
 832 /*
 833  * Does preemptible RCU need the CPU to stay out of dynticks mode?
 834  */
 835 int rcu_preempt_needs_cpu(void)
 836 {
 837         if (!rcu_preempt_running_reader())
 838                 rcu_preempt_cpu_qs();
 839         return rcu_preempt_ctrlblk.rcb.rcucblist != NULL;
 840 }
 841
 842 /*
 843  * Check for a task exiting while in a preemptible -RCU read-side
 844  * critical section, clean up if so.  No need to issue warnings,
 845  * as debug_check_no_locks_held() already does this if lockdep
 846  * is enabled.
 847  */
 848 void exit_rcu(void)
 849 {
 850         struct task_struct *t = current;
 851
 852         if (t->rcu_read_lock_nesting == 0)
 853                 return;
 854         t->rcu_read_lock_nesting = 1;
 855         rcu_read_unlock();
 856 }
 857
 858 #else /* #ifdef CONFIG_TINY_PREEMPT_RCU */
 859
 860 #ifdef CONFIG_RCU_TRACE
 861
 862 /*
 863  * Because preemptible RCU does not exist, it is not necessary to
 864  * dump out its statistics.
 865  */
 866 static void show_tiny_preempt_stats(struct seq_file *m)
 867 {
 868 }
 869
 870 #endif /* #ifdef CONFIG_RCU_TRACE */
 871
 872 /*
 873  * Because preemptible RCU does not exist, it is never necessary to
 874  * boost preempted RCU readers.
 875  */
 876 static int rcu_boost(void)
 877 {
 878         return 0;
 879 }
 880
 881 /*
 882  * Because preemptible RCU does not exist, it never has any callbacks
 883  * to check.
 884  */
 885 static void rcu_preempt_check_callbacks(void)
 886 {
 887 }
 888
 889 /*
 890  * Because preemptible RCU does not exist, it never has any callbacks
 891  * to remove.
 892  */
 893 static void rcu_preempt_remove_callbacks(struct rcu_ctrlblk *rcp)
 894 {
 895 }
 896
 897 /*
 898  * Because preemptible RCU does not exist, it never has any callbacks
 899  * to process.
 900  */
 901 static void rcu_preempt_process_callbacks(void)
 902 {
 903 }
 904
 905 #endif /* #else #ifdef CONFIG_TINY_PREEMPT_RCU */
 906
 907 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 908 #include <linux/kernel_stat.h>
 909
 910 /*
 911  * During boot, we forgive RCU lockdep issues.  After this function is
 912  * invoked, we start taking RCU lockdep issues seriously.
 913  */
 914 void __init rcu_scheduler_starting(void)
 915 {
 916         WARN_ON(nr_context_switches() > 0);
 917         rcu_scheduler_active = 1;
 918 }
 919
 920 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 921
 922 #ifdef CONFIG_RCU_BOOST
 923 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
 924 #else /* #ifdef CONFIG_RCU_BOOST */
 925 #define RCU_BOOST_PRIO 1
 926 #endif /* #else #ifdef CONFIG_RCU_BOOST */
 927
 928 #ifdef CONFIG_RCU_TRACE
 929
 930 #ifdef CONFIG_RCU_BOOST
 931
 932 static void rcu_initiate_boost_trace(void)
 933 {
 934         if (rcu_preempt_ctrlblk.gp_tasks == NULL)
 935                 rcu_preempt_ctrlblk.n_normal_balk_gp_tasks++;
 936         else if (rcu_preempt_ctrlblk.boost_tasks != NULL)
 937                 rcu_preempt_ctrlblk.n_normal_balk_boost_tasks++;
 938         else if (rcu_preempt_ctrlblk.boosted_this_gp != 0)
 939                 rcu_preempt_ctrlblk.n_normal_balk_boosted++;
 940         else if (!ULONG_CMP_GE(jiffies, rcu_preempt_ctrlblk.boost_time))
 941                 rcu_preempt_ctrlblk.n_normal_balk_notyet++;
 942         else
 943                 rcu_preempt_ctrlblk.n_normal_balk_nos++;
 944 }
 945
 946 static void rcu_initiate_exp_boost_trace(void)
 947 {
 948         if (list_empty(&rcu_preempt_ctrlblk.blkd_tasks))
 949                 rcu_preempt_ctrlblk.n_exp_balk_blkd_tasks++;
 950         else
 951                 rcu_preempt_ctrlblk.n_exp_balk_nos++;
 952 }
 953
 954 #endif /* #ifdef CONFIG_RCU_BOOST */
 955
 956 static void rcu_trace_sub_qlen(struct rcu_ctrlblk *rcp, int n)
 957 {
 958         unsigned long flags;
 959
 960         raw_local_irq_save(flags);
 961         rcp->qlen -= n;
 962         raw_local_irq_restore(flags);
 963 }
 964
 965 /*
 966  * Dump statistics for TINY_RCU, such as they are.
 967  */
 968 static int show_tiny_stats(struct seq_file *m, void *unused)
 969 {
 970         show_tiny_preempt_stats(m);
 971         seq_printf(m, "rcu_sched: qlen: %ld\n", rcu_sched_ctrlblk.qlen);
 972         seq_printf(m, "rcu_bh: qlen: %ld\n", rcu_bh_ctrlblk.qlen);
 973         return 0;
 974 }
 975
 976 static int show_tiny_stats_open(struct inode *inode, struct file *file)
 977 {
 978         return single_open(file, show_tiny_stats, NULL);
 979 }
 980
 981 static const struct file_operations show_tiny_stats_fops = {
 982         .owner = THIS_MODULE,
 983         .open = show_tiny_stats_open,
 984         .read = seq_read,
 985         .llseek = seq_lseek,
 986         .release = single_release,
 987 };
 988
 989 static struct dentry *rcudir;
 990
 991 static int __init rcutiny_trace_init(void)
 992 {
 993         struct dentry *retval;
 994
 995         rcudir = debugfs_create_dir("rcu", NULL);
 996         if (!rcudir)
 997                 goto free_out;
 998         retval = debugfs_create_file("rcudata", 0444, rcudir,
 999                                      NULL, &show_tiny_stats_fops);
1000         if (!retval)
1001                 goto free_out;
1002         return 0;
1003 free_out:
1004         debugfs_remove_recursive(rcudir);
1005         return 1;
1006 }
1007
1008 static void __exit rcutiny_trace_cleanup(void)
1009 {
1010         debugfs_remove_recursive(rcudir);
1011 }
1012
1013 module_init(rcutiny_trace_init);
1014 module_exit(rcutiny_trace_cleanup);
1015
1016 MODULE_AUTHOR("Paul E. McKenney");
1017 MODULE_DESCRIPTION("Read-Copy Update tracing for tiny implementation");
1018 MODULE_LICENSE("GPL");
1019
1020 #endif /* #ifdef CONFIG_RCU_TRACE */