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/module.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/seq_file.h>
  29
  30 /* Global control variables for rcupdate callback mechanism. */
  31 struct rcu_ctrlblk {
  32         struct rcu_head *rcucblist;     /* List of pending callbacks (CBs). */
  33         struct rcu_head **donetail;     /* ->next pointer of last "done" CB. */
  34         struct rcu_head **curtail;      /* ->next pointer of last CB. */
  35         RCU_TRACE(long qlen);           /* Number of pending CBs. */
  36         RCU_TRACE(unsigned long gp_start); /* Start time for stalls. */
  37         RCU_TRACE(unsigned long ticks_this_gp); /* Statistic for stalls. */
  38         RCU_TRACE(unsigned long jiffies_stall); /* Jiffies at next stall. */
  39         RCU_TRACE(char *name);          /* Name of RCU type. */
  40 };
  41
  42 /* Definition for rcupdate control block. */
  43 static struct rcu_ctrlblk rcu_sched_ctrlblk = {
  44         .donetail       = &rcu_sched_ctrlblk.rcucblist,
  45         .curtail        = &rcu_sched_ctrlblk.rcucblist,
  46         RCU_TRACE(.name = "rcu_sched")
  47 };
  48
  49 static struct rcu_ctrlblk rcu_bh_ctrlblk = {
  50         .donetail       = &rcu_bh_ctrlblk.rcucblist,
  51         .curtail        = &rcu_bh_ctrlblk.rcucblist,
  52         RCU_TRACE(.name = "rcu_bh")
  53 };
  54
  55 #ifdef CONFIG_DEBUG_LOCK_ALLOC
  56 int rcu_scheduler_active __read_mostly;
  57 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
  58 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
  59
  60 #ifdef CONFIG_RCU_TRACE
  61
  62 static void check_cpu_stall(struct rcu_ctrlblk *rcp)
  63 {
  64         unsigned long j;
  65         unsigned long js;
  66
  67         if (rcu_cpu_stall_suppress)
  68                 return;
  69         rcp->ticks_this_gp++;
  70         j = jiffies;
  71         js = rcp->jiffies_stall;
  72         if (*rcp->curtail && ULONG_CMP_GE(j, js)) {
  73                 pr_err("INFO: %s stall on CPU (%lu ticks this GP) idle=%llx (t=%lu jiffies q=%ld)\n",
  74                        rcp->name, rcp->ticks_this_gp, rcu_dynticks_nesting,
  75                        jiffies - rcp->gp_start, rcp->qlen);
  76                 dump_stack();
  77         }
  78         if (*rcp->curtail && ULONG_CMP_GE(j, js))
  79                 rcp->jiffies_stall = jiffies +
  80                         3 * rcu_jiffies_till_stall_check() + 3;
  81         else if (ULONG_CMP_GE(j, js))
  82                 rcp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
  83 }
  84
  85 static void check_cpu_stall_preempt(void);
  86
  87 #endif /* #ifdef CONFIG_RCU_TRACE */
  88
  89 static void reset_cpu_stall_ticks(struct rcu_ctrlblk *rcp)
  90 {
  91 #ifdef CONFIG_RCU_TRACE
  92         rcp->ticks_this_gp = 0;
  93         rcp->gp_start = jiffies;
  94         rcp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
  95 #endif /* #ifdef CONFIG_RCU_TRACE */
  96 }
  97
  98 static void check_cpu_stalls(void)
  99 {
 100         RCU_TRACE(check_cpu_stall(&rcu_bh_ctrlblk));
 101         RCU_TRACE(check_cpu_stall(&rcu_sched_ctrlblk));
 102         RCU_TRACE(check_cpu_stall_preempt());
 103 }
 104
 105 #ifdef CONFIG_TINY_PREEMPT_RCU
 106
 107 #include <linux/delay.h>
 108
 109 /* Global control variables for preemptible RCU. */
 110 struct rcu_preempt_ctrlblk {
 111         struct rcu_ctrlblk rcb; /* curtail: ->next ptr of last CB for GP. */
 112         struct rcu_head **nexttail;
 113                                 /* Tasks blocked in a preemptible RCU */
 114                                 /*  read-side critical section while an */
 115                                 /*  preemptible-RCU grace period is in */
 116                                 /*  progress must wait for a later grace */
 117                                 /*  period.  This pointer points to the */
 118                                 /*  ->next pointer of the last task that */
 119                                 /*  must wait for a later grace period, or */
 120                                 /*  to &->rcb.rcucblist if there is no */
 121                                 /*  such task. */
 122         struct list_head blkd_tasks;
 123                                 /* Tasks blocked in RCU read-side critical */
 124                                 /*  section.  Tasks are placed at the head */
 125                                 /*  of this list and age towards the tail. */
 126         struct list_head *gp_tasks;
 127                                 /* Pointer to the first task blocking the */
 128                                 /*  current grace period, or NULL if there */
 129                                 /*  is no such task. */
 130         struct list_head *exp_tasks;
 131                                 /* Pointer to first task blocking the */
 132                                 /*  current expedited grace period, or NULL */
 133                                 /*  if there is no such task.  If there */
 134                                 /*  is no current expedited grace period, */
 135                                 /*  then there cannot be any such task. */
 136 #ifdef CONFIG_RCU_BOOST
 137         struct list_head *boost_tasks;
 138                                 /* Pointer to first task that needs to be */
 139                                 /*  priority-boosted, or NULL if no priority */
 140                                 /*  boosting is needed.  If there is no */
 141                                 /*  current or expedited grace period, there */
 142                                 /*  can be no such task. */
 143 #endif /* #ifdef CONFIG_RCU_BOOST */
 144         u8 gpnum;               /* Current grace period. */
 145         u8 gpcpu;               /* Last grace period blocked by the CPU. */
 146         u8 completed;           /* Last grace period completed. */
 147                                 /*  If all three are equal, RCU is idle. */
 148 #ifdef CONFIG_RCU_BOOST
 149         unsigned long boost_time; /* When to start boosting (jiffies) */
 150 #endif /* #ifdef CONFIG_RCU_BOOST */
 151 #ifdef CONFIG_RCU_TRACE
 152         unsigned long n_grace_periods;
 153 #ifdef CONFIG_RCU_BOOST
 154         unsigned long n_tasks_boosted;
 155                                 /* Total number of tasks boosted. */
 156         unsigned long n_exp_boosts;
 157                                 /* Number of tasks boosted for expedited GP. */
 158         unsigned long n_normal_boosts;
 159                                 /* Number of tasks boosted for normal GP. */
 160         unsigned long n_balk_blkd_tasks;
 161                                 /* Refused to boost: no blocked tasks. */
 162         unsigned long n_balk_exp_gp_tasks;
 163                                 /* Refused to boost: nothing blocking GP. */
 164         unsigned long n_balk_boost_tasks;
 165                                 /* Refused to boost: already boosting. */
 166         unsigned long n_balk_notyet;
 167                                 /* Refused to boost: not yet time. */
 168         unsigned long n_balk_nos;
 169                                 /* Refused to boost: not sure why, though. */
 170                                 /*  This can happen due to race conditions. */
 171 #endif /* #ifdef CONFIG_RCU_BOOST */
 172 #endif /* #ifdef CONFIG_RCU_TRACE */
 173 };
 174
 175 static struct rcu_preempt_ctrlblk rcu_preempt_ctrlblk = {
 176         .rcb.donetail = &rcu_preempt_ctrlblk.rcb.rcucblist,
 177         .rcb.curtail = &rcu_preempt_ctrlblk.rcb.rcucblist,
 178         .nexttail = &rcu_preempt_ctrlblk.rcb.rcucblist,
 179         .blkd_tasks = LIST_HEAD_INIT(rcu_preempt_ctrlblk.blkd_tasks),
 180         RCU_TRACE(.rcb.name = "rcu_preempt")
 181 };
 182
 183 static int rcu_preempted_readers_exp(void);
 184 static void rcu_report_exp_done(void);
 185
 186 /*
 187  * Return true if the CPU has not yet responded to the current grace period.
 188  */
 189 static int rcu_cpu_blocking_cur_gp(void)
 190 {
 191         return rcu_preempt_ctrlblk.gpcpu != rcu_preempt_ctrlblk.gpnum;
 192 }
 193
 194 /*
 195  * Check for a running RCU reader.  Because there is only one CPU,
 196  * there can be but one running RCU reader at a time.  ;-)
 197  *
 198  * Returns zero if there are no running readers.  Returns a positive
 199  * number if there is at least one reader within its RCU read-side
 200  * critical section.  Returns a negative number if an outermost reader
 201  * is in the midst of exiting from its RCU read-side critical section
 202  *
 203  * Returns zero if there are no running readers.  Returns a positive
 204  * number if there is at least one reader within its RCU read-side
 205  * critical section.  Returns a negative number if an outermost reader
 206  * is in the midst of exiting from its RCU read-side critical section.
 207  */
 208 static int rcu_preempt_running_reader(void)
 209 {
 210         return current->rcu_read_lock_nesting;
 211 }
 212
 213 /*
 214  * Check for preempted RCU readers blocking any grace period.
 215  * If the caller needs a reliable answer, it must disable hard irqs.
 216  */
 217 static int rcu_preempt_blocked_readers_any(void)
 218 {
 219         return !list_empty(&rcu_preempt_ctrlblk.blkd_tasks);
 220 }
 221
 222 /*
 223  * Check for preempted RCU readers blocking the current grace period.
 224  * If the caller needs a reliable answer, it must disable hard irqs.
 225  */
 226 static int rcu_preempt_blocked_readers_cgp(void)
 227 {
 228         return rcu_preempt_ctrlblk.gp_tasks != NULL;
 229 }
 230
 231 /*
 232  * Return true if another preemptible-RCU grace period is needed.
 233  */
 234 static int rcu_preempt_needs_another_gp(void)
 235 {
 236         return *rcu_preempt_ctrlblk.rcb.curtail != NULL;
 237 }
 238
 239 /*
 240  * Return true if a preemptible-RCU grace period is in progress.
 241  * The caller must disable hardirqs.
 242  */
 243 static int rcu_preempt_gp_in_progress(void)
 244 {
 245         return rcu_preempt_ctrlblk.completed != rcu_preempt_ctrlblk.gpnum;
 246 }
 247
 248 /*
 249  * Advance a ->blkd_tasks-list pointer to the next entry, instead
 250  * returning NULL if at the end of the list.
 251  */
 252 static struct list_head *rcu_next_node_entry(struct task_struct *t)
 253 {
 254         struct list_head *np;
 255
 256         np = t->rcu_node_entry.next;
 257         if (np == &rcu_preempt_ctrlblk.blkd_tasks)
 258                 np = NULL;
 259         return np;
 260 }
 261
 262 #ifdef CONFIG_RCU_TRACE
 263
 264 #ifdef CONFIG_RCU_BOOST
 265 static void rcu_initiate_boost_trace(void);
 266 #endif /* #ifdef CONFIG_RCU_BOOST */
 267
 268 /*
 269  * Dump additional statistice for TINY_PREEMPT_RCU.
 270  */
 271 static void show_tiny_preempt_stats(struct seq_file *m)
 272 {
 273         seq_printf(m, "rcu_preempt: qlen=%ld gp=%lu g%u/p%u/c%u tasks=%c%c%c\n",
 274                    rcu_preempt_ctrlblk.rcb.qlen,
 275                    rcu_preempt_ctrlblk.n_grace_periods,
 276                    rcu_preempt_ctrlblk.gpnum,
 277                    rcu_preempt_ctrlblk.gpcpu,
 278                    rcu_preempt_ctrlblk.completed,
 279                    "T."[list_empty(&rcu_preempt_ctrlblk.blkd_tasks)],
 280                    "N."[!rcu_preempt_ctrlblk.gp_tasks],
 281                    "E."[!rcu_preempt_ctrlblk.exp_tasks]);
 282 #ifdef CONFIG_RCU_BOOST
 283         seq_printf(m, "%sttb=%c ntb=%lu neb=%lu nnb=%lu j=%04x bt=%04x\n",
 284                    "             ",
 285                    "B."[!rcu_preempt_ctrlblk.boost_tasks],
 286                    rcu_preempt_ctrlblk.n_tasks_boosted,
 287                    rcu_preempt_ctrlblk.n_exp_boosts,
 288                    rcu_preempt_ctrlblk.n_normal_boosts,
 289                    (int)(jiffies & 0xffff),
 290                    (int)(rcu_preempt_ctrlblk.boost_time & 0xffff));
 291         seq_printf(m, "%s: nt=%lu egt=%lu bt=%lu ny=%lu nos=%lu\n",
 292                    "             balk",
 293                    rcu_preempt_ctrlblk.n_balk_blkd_tasks,
 294                    rcu_preempt_ctrlblk.n_balk_exp_gp_tasks,
 295                    rcu_preempt_ctrlblk.n_balk_boost_tasks,
 296                    rcu_preempt_ctrlblk.n_balk_notyet,
 297                    rcu_preempt_ctrlblk.n_balk_nos);
 298 #endif /* #ifdef CONFIG_RCU_BOOST */
 299 }
 300
 301 #endif /* #ifdef CONFIG_RCU_TRACE */
 302
 303 #ifdef CONFIG_RCU_BOOST
 304
 305 #include "rtmutex_common.h"
 306
 307 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
 308
 309 /* Controls for rcu_kthread() kthread. */
 310 static struct task_struct *rcu_kthread_task;
 311 static DECLARE_WAIT_QUEUE_HEAD(rcu_kthread_wq);
 312 static unsigned long have_rcu_kthread_work;
 313
 314 /*
 315  * Carry out RCU priority boosting on the task indicated by ->boost_tasks,
 316  * and advance ->boost_tasks to the next task in the ->blkd_tasks list.
 317  */
 318 static int rcu_boost(void)
 319 {
 320         unsigned long flags;
 321         struct rt_mutex mtx;
 322         struct task_struct *t;
 323         struct list_head *tb;
 324
 325         if (rcu_preempt_ctrlblk.boost_tasks == NULL &&
 326             rcu_preempt_ctrlblk.exp_tasks == NULL)
 327                 return 0;  /* Nothing to boost. */
 328
 329         local_irq_save(flags);
 330
 331         /*
 332          * Recheck with irqs disabled: all tasks in need of boosting
 333          * might exit their RCU read-side critical sections on their own
 334          * if we are preempted just before disabling irqs.
 335          */
 336         if (rcu_preempt_ctrlblk.boost_tasks == NULL &&
 337             rcu_preempt_ctrlblk.exp_tasks == NULL) {
 338                 local_irq_restore(flags);
 339                 return 0;
 340         }
 341
 342         /*
 343          * Preferentially boost tasks blocking expedited grace periods.
 344          * This cannot starve the normal grace periods because a second
 345          * expedited grace period must boost all blocked tasks, including
 346          * those blocking the pre-existing normal grace period.
 347          */
 348         if (rcu_preempt_ctrlblk.exp_tasks != NULL) {
 349                 tb = rcu_preempt_ctrlblk.exp_tasks;
 350                 RCU_TRACE(rcu_preempt_ctrlblk.n_exp_boosts++);
 351         } else {
 352                 tb = rcu_preempt_ctrlblk.boost_tasks;
 353                 RCU_TRACE(rcu_preempt_ctrlblk.n_normal_boosts++);
 354         }
 355         RCU_TRACE(rcu_preempt_ctrlblk.n_tasks_boosted++);
 356
 357         /*
 358          * We boost task t by manufacturing an rt_mutex that appears to
 359          * be held by task t.  We leave a pointer to that rt_mutex where
 360          * task t can find it, and task t will release the mutex when it
 361          * exits its outermost RCU read-side critical section.  Then
 362          * simply acquiring this artificial rt_mutex will boost task
 363          * t's priority.  (Thanks to tglx for suggesting this approach!)
 364          */
 365         t = container_of(tb, struct task_struct, rcu_node_entry);
 366         rt_mutex_init_proxy_locked(&mtx, t);
 367         t->rcu_boost_mutex = &mtx;
 368         local_irq_restore(flags);
 369         rt_mutex_lock(&mtx);
 370         rt_mutex_unlock(&mtx);  /* Keep lockdep happy. */
 371
 372         return ACCESS_ONCE(rcu_preempt_ctrlblk.boost_tasks) != NULL ||
 373                ACCESS_ONCE(rcu_preempt_ctrlblk.exp_tasks) != NULL;
 374 }
 375
 376 /*
 377  * Check to see if it is now time to start boosting RCU readers blocking
 378  * the current grace period, and, if so, tell the rcu_kthread_task to
 379  * start boosting them.  If there is an expedited boost in progress,
 380  * we wait for it to complete.
 381  *
 382  * If there are no blocked readers blocking the current grace period,
 383  * return 0 to let the caller know, otherwise return 1.  Note that this
 384  * return value is independent of whether or not boosting was done.
 385  */
 386 static int rcu_initiate_boost(void)
 387 {
 388         if (!rcu_preempt_blocked_readers_cgp() &&
 389             rcu_preempt_ctrlblk.exp_tasks == NULL) {
 390                 RCU_TRACE(rcu_preempt_ctrlblk.n_balk_exp_gp_tasks++);
 391                 return 0;
 392         }
 393         if (rcu_preempt_ctrlblk.exp_tasks != NULL ||
 394             (rcu_preempt_ctrlblk.gp_tasks != NULL &&
 395              rcu_preempt_ctrlblk.boost_tasks == NULL &&
 396              ULONG_CMP_GE(jiffies, rcu_preempt_ctrlblk.boost_time))) {
 397                 if (rcu_preempt_ctrlblk.exp_tasks == NULL)
 398                         rcu_preempt_ctrlblk.boost_tasks =
 399                                 rcu_preempt_ctrlblk.gp_tasks;
 400                 invoke_rcu_callbacks();
 401         } else {
 402                 RCU_TRACE(rcu_initiate_boost_trace());
 403         }
 404         return 1;
 405 }
 406
 407 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
 408
 409 /*
 410  * Do priority-boost accounting for the start of a new grace period.
 411  */
 412 static void rcu_preempt_boost_start_gp(void)
 413 {
 414         rcu_preempt_ctrlblk.boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
 415 }
 416
 417 #else /* #ifdef CONFIG_RCU_BOOST */
 418
 419 /*
 420  * If there is no RCU priority boosting, we don't initiate boosting,
 421  * but we do indicate whether there are blocked readers blocking the
 422  * current grace period.
 423  */
 424 static int rcu_initiate_boost(void)
 425 {
 426         return rcu_preempt_blocked_readers_cgp();
 427 }
 428
 429 /*
 430  * If there is no RCU priority boosting, nothing to do at grace-period start.
 431  */
 432 static void rcu_preempt_boost_start_gp(void)
 433 {
 434 }
 435
 436 #endif /* else #ifdef CONFIG_RCU_BOOST */
 437
 438 /*
 439  * Record a preemptible-RCU quiescent state for the specified CPU.  Note
 440  * that this just means that the task currently running on the CPU is
 441  * in a quiescent state.  There might be any number of tasks blocked
 442  * while in an RCU read-side critical section.
 443  *
 444  * Unlike the other rcu_*_qs() functions, callers to this function
 445  * must disable irqs in order to protect the assignment to
 446  * ->rcu_read_unlock_special.
 447  *
 448  * Because this is a single-CPU implementation, the only way a grace
 449  * period can end is if the CPU is in a quiescent state.  The reason is
 450  * that a blocked preemptible-RCU reader can exit its critical section
 451  * only if the CPU is running it at the time.  Therefore, when the
 452  * last task blocking the current grace period exits its RCU read-side
 453  * critical section, neither the CPU nor blocked tasks will be stopping
 454  * the current grace period.  (In contrast, SMP implementations
 455  * might have CPUs running in RCU read-side critical sections that
 456  * block later grace periods -- but this is not possible given only
 457  * one CPU.)
 458  */
 459 static void rcu_preempt_cpu_qs(void)
 460 {
 461         /* Record both CPU and task as having responded to current GP. */
 462         rcu_preempt_ctrlblk.gpcpu = rcu_preempt_ctrlblk.gpnum;
 463         current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
 464
 465         /* If there is no GP then there is nothing more to do.  */
 466         if (!rcu_preempt_gp_in_progress())
 467                 return;
 468         /*
 469          * Check up on boosting.  If there are readers blocking the
 470          * current grace period, leave.
 471          */
 472         if (rcu_initiate_boost())
 473                 return;
 474
 475         /* Advance callbacks. */
 476         rcu_preempt_ctrlblk.completed = rcu_preempt_ctrlblk.gpnum;
 477         rcu_preempt_ctrlblk.rcb.donetail = rcu_preempt_ctrlblk.rcb.curtail;
 478         rcu_preempt_ctrlblk.rcb.curtail = rcu_preempt_ctrlblk.nexttail;
 479
 480         /* If there are no blocked readers, next GP is done instantly. */
 481         if (!rcu_preempt_blocked_readers_any())
 482                 rcu_preempt_ctrlblk.rcb.donetail = rcu_preempt_ctrlblk.nexttail;
 483
 484         /* If there are done callbacks, cause them to be invoked. */
 485         if (*rcu_preempt_ctrlblk.rcb.donetail != NULL)
 486                 invoke_rcu_callbacks();
 487 }
 488
 489 /*
 490  * Start a new RCU grace period if warranted.  Hard irqs must be disabled.
 491  */
 492 static void rcu_preempt_start_gp(void)
 493 {
 494         if (!rcu_preempt_gp_in_progress() && rcu_preempt_needs_another_gp()) {
 495
 496                 /* Official start of GP. */
 497                 rcu_preempt_ctrlblk.gpnum++;
 498                 RCU_TRACE(rcu_preempt_ctrlblk.n_grace_periods++);
 499                 reset_cpu_stall_ticks(&rcu_preempt_ctrlblk.rcb);
 500
 501                 /* Any blocked RCU readers block new GP. */
 502                 if (rcu_preempt_blocked_readers_any())
 503                         rcu_preempt_ctrlblk.gp_tasks =
 504                                 rcu_preempt_ctrlblk.blkd_tasks.next;
 505
 506                 /* Set up for RCU priority boosting. */
 507                 rcu_preempt_boost_start_gp();
 508
 509                 /* If there is no running reader, CPU is done with GP. */
 510                 if (!rcu_preempt_running_reader())
 511                         rcu_preempt_cpu_qs();
 512         }
 513 }
 514
 515 /*
 516  * We have entered the scheduler, and the current task might soon be
 517  * context-switched away from.  If this task is in an RCU read-side
 518  * critical section, we will no longer be able to rely on the CPU to
 519  * record that fact, so we enqueue the task on the blkd_tasks list.
 520  * If the task started after the current grace period began, as recorded
 521  * by ->gpcpu, we enqueue at the beginning of the list.  Otherwise
 522  * before the element referenced by ->gp_tasks (or at the tail if
 523  * ->gp_tasks is NULL) and point ->gp_tasks at the newly added element.
 524  * The task will dequeue itself when it exits the outermost enclosing
 525  * RCU read-side critical section.  Therefore, the current grace period
 526  * cannot be permitted to complete until the ->gp_tasks pointer becomes
 527  * NULL.
 528  *
 529  * Caller must disable preemption.
 530  */
 531 void rcu_preempt_note_context_switch(void)
 532 {
 533         struct task_struct *t = current;
 534         unsigned long flags;
 535
 536         local_irq_save(flags); /* must exclude scheduler_tick(). */
 537         if (rcu_preempt_running_reader() > 0 &&
 538             (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
 539
 540                 /* Possibly blocking in an RCU read-side critical section. */
 541                 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
 542
 543                 /*
 544                  * If this CPU has already checked in, then this task
 545                  * will hold up the next grace period rather than the
 546                  * current grace period.  Queue the task accordingly.
 547                  * If the task is queued for the current grace period
 548                  * (i.e., this CPU has not yet passed through a quiescent
 549                  * state for the current grace period), then as long
 550                  * as that task remains queued, the current grace period
 551                  * cannot end.
 552                  */
 553                 list_add(&t->rcu_node_entry, &rcu_preempt_ctrlblk.blkd_tasks);
 554                 if (rcu_cpu_blocking_cur_gp())
 555                         rcu_preempt_ctrlblk.gp_tasks = &t->rcu_node_entry;
 556         } else if (rcu_preempt_running_reader() < 0 &&
 557                    t->rcu_read_unlock_special) {
 558                 /*
 559                  * Complete exit from RCU read-side critical section on
 560                  * behalf of preempted instance of __rcu_read_unlock().
 561                  */
 562                 rcu_read_unlock_special(t);
 563         }
 564
 565         /*
 566          * Either we were not in an RCU read-side critical section to
 567          * begin with, or we have now recorded that critical section
 568          * globally.  Either way, we can now note a quiescent state
 569          * for this CPU.  Again, if we were in an RCU read-side critical
 570          * section, and if that critical section was blocking the current
 571          * grace period, then the fact that the task has been enqueued
 572          * means that current grace period continues to be blocked.
 573          */
 574         rcu_preempt_cpu_qs();
 575         local_irq_restore(flags);
 576 }
 577
 578 /*
 579  * Handle special cases during rcu_read_unlock(), such as needing to
 580  * notify RCU core processing or task having blocked during the RCU
 581  * read-side critical section.
 582  */
 583 void rcu_read_unlock_special(struct task_struct *t)
 584 {
 585         int empty;
 586         int empty_exp;
 587         unsigned long flags;
 588         struct list_head *np;
 589 #ifdef CONFIG_RCU_BOOST
 590         struct rt_mutex *rbmp = NULL;
 591 #endif /* #ifdef CONFIG_RCU_BOOST */
 592         int special;
 593
 594         /*
 595          * NMI handlers cannot block and cannot safely manipulate state.
 596          * They therefore cannot possibly be special, so just leave.
 597          */
 598         if (in_nmi())
 599                 return;
 600
 601         local_irq_save(flags);
 602
 603         /*
 604          * If RCU core is waiting for this CPU to exit critical section,
 605          * let it know that we have done so.
 606          */
 607         special = t->rcu_read_unlock_special;
 608         if (special & RCU_READ_UNLOCK_NEED_QS)
 609                 rcu_preempt_cpu_qs();
 610
 611         /* Hardware IRQ handlers cannot block. */
 612         if (in_irq() || in_serving_softirq()) {
 613                 local_irq_restore(flags);
 614                 return;
 615         }
 616
 617         /* Clean up if blocked during RCU read-side critical section. */
 618         if (special & RCU_READ_UNLOCK_BLOCKED) {
 619                 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
 620
 621                 /*
 622                  * Remove this task from the ->blkd_tasks list and adjust
 623                  * any pointers that might have been referencing it.
 624                  */
 625                 empty = !rcu_preempt_blocked_readers_cgp();
 626                 empty_exp = rcu_preempt_ctrlblk.exp_tasks == NULL;
 627                 np = rcu_next_node_entry(t);
 628                 list_del_init(&t->rcu_node_entry);
 629                 if (&t->rcu_node_entry == rcu_preempt_ctrlblk.gp_tasks)
 630                         rcu_preempt_ctrlblk.gp_tasks = np;
 631                 if (&t->rcu_node_entry == rcu_preempt_ctrlblk.exp_tasks)
 632                         rcu_preempt_ctrlblk.exp_tasks = np;
 633 #ifdef CONFIG_RCU_BOOST
 634                 if (&t->rcu_node_entry == rcu_preempt_ctrlblk.boost_tasks)
 635                         rcu_preempt_ctrlblk.boost_tasks = np;
 636 #endif /* #ifdef CONFIG_RCU_BOOST */
 637
 638                 /*
 639                  * If this was the last task on the current list, and if
 640                  * we aren't waiting on the CPU, report the quiescent state
 641                  * and start a new grace period if needed.
 642                  */
 643                 if (!empty && !rcu_preempt_blocked_readers_cgp()) {
 644                         rcu_preempt_cpu_qs();
 645                         rcu_preempt_start_gp();
 646                 }
 647
 648                 /*
 649                  * If this was the last task on the expedited lists,
 650                  * then we need wake up the waiting task.
 651                  */
 652                 if (!empty_exp && rcu_preempt_ctrlblk.exp_tasks == NULL)
 653                         rcu_report_exp_done();
 654         }
 655 #ifdef CONFIG_RCU_BOOST
 656         /* Unboost self if was boosted. */
 657         if (t->rcu_boost_mutex != NULL) {
 658                 rbmp = t->rcu_boost_mutex;
 659                 t->rcu_boost_mutex = NULL;
 660                 rt_mutex_unlock(rbmp);
 661         }
 662 #endif /* #ifdef CONFIG_RCU_BOOST */
 663         local_irq_restore(flags);
 664 }
 665
 666 /*
 667  * Check for a quiescent state from the current CPU.  When a task blocks,
 668  * the task is recorded in the rcu_preempt_ctrlblk structure, which is
 669  * checked elsewhere.  This is called from the scheduling-clock interrupt.
 670  *
 671  * Caller must disable hard irqs.
 672  */
 673 static void rcu_preempt_check_callbacks(void)
 674 {
 675         struct task_struct *t = current;
 676
 677         if (rcu_preempt_gp_in_progress() &&
 678             (!rcu_preempt_running_reader() ||
 679              !rcu_cpu_blocking_cur_gp()))
 680                 rcu_preempt_cpu_qs();
 681         if (&rcu_preempt_ctrlblk.rcb.rcucblist !=
 682             rcu_preempt_ctrlblk.rcb.donetail)
 683                 invoke_rcu_callbacks();
 684         if (rcu_preempt_gp_in_progress() &&
 685             rcu_cpu_blocking_cur_gp() &&
 686             rcu_preempt_running_reader() > 0)
 687                 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
 688 }
 689
 690 /*
 691  * TINY_PREEMPT_RCU has an extra callback-list tail pointer to
 692  * update, so this is invoked from rcu_process_callbacks() to
 693  * handle that case.  Of course, it is invoked for all flavors of
 694  * RCU, but RCU callbacks can appear only on one of the lists, and
 695  * neither ->nexttail nor ->donetail can possibly be NULL, so there
 696  * is no need for an explicit check.
 697  */
 698 static void rcu_preempt_remove_callbacks(struct rcu_ctrlblk *rcp)
 699 {
 700         if (rcu_preempt_ctrlblk.nexttail == rcp->donetail)
 701                 rcu_preempt_ctrlblk.nexttail = &rcp->rcucblist;
 702 }
 703
 704 /*
 705  * Process callbacks for preemptible RCU.
 706  */
 707 static void rcu_preempt_process_callbacks(void)
 708 {
 709         __rcu_process_callbacks(&rcu_preempt_ctrlblk.rcb);
 710 }
 711
 712 /*
 713  * Queue a preemptible -RCU callback for invocation after a grace period.
 714  */
 715 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
 716 {
 717         unsigned long flags;
 718
 719         debug_rcu_head_queue(head);
 720         head->func = func;
 721         head->next = NULL;
 722
 723         local_irq_save(flags);
 724         *rcu_preempt_ctrlblk.nexttail = head;
 725         rcu_preempt_ctrlblk.nexttail = &head->next;
 726         RCU_TRACE(rcu_preempt_ctrlblk.rcb.qlen++);
 727         rcu_preempt_start_gp();  /* checks to see if GP needed. */
 728         local_irq_restore(flags);
 729 }
 730 EXPORT_SYMBOL_GPL(call_rcu);
 731
 732 /*
 733  * synchronize_rcu - wait until a grace period has elapsed.
 734  *
 735  * Control will return to the caller some time after a full grace
 736  * period has elapsed, in other words after all currently executing RCU
 737  * read-side critical sections have completed.  RCU read-side critical
 738  * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
 739  * and may be nested.
 740  */
 741 void synchronize_rcu(void)
 742 {
 743         rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
 744                            !lock_is_held(&rcu_lock_map) &&
 745                            !lock_is_held(&rcu_sched_lock_map),
 746                            "Illegal synchronize_rcu() in RCU read-side critical section");
 747
 748 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 749         if (!rcu_scheduler_active)
 750                 return;
 751 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
 752
 753         WARN_ON_ONCE(rcu_preempt_running_reader());
 754         if (!rcu_preempt_blocked_readers_any())
 755                 return;
 756
 757         /* Once we get past the fastpath checks, same code as rcu_barrier(). */
 758         if (rcu_expedited)
 759                 synchronize_rcu_expedited();
 760         else
 761                 rcu_barrier();
 762 }
 763 EXPORT_SYMBOL_GPL(synchronize_rcu);
 764
 765 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
 766 static unsigned long sync_rcu_preempt_exp_count;
 767 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
 768
 769 /*
 770  * Return non-zero if there are any tasks in RCU read-side critical
 771  * sections blocking the current preemptible-RCU expedited grace period.
 772  * If there is no preemptible-RCU expedited grace period currently in
 773  * progress, returns zero unconditionally.
 774  */
 775 static int rcu_preempted_readers_exp(void)
 776 {
 777         return rcu_preempt_ctrlblk.exp_tasks != NULL;
 778 }
 779
 780 /*
 781  * Report the exit from RCU read-side critical section for the last task
 782  * that queued itself during or before the current expedited preemptible-RCU
 783  * grace period.
 784  */
 785 static void rcu_report_exp_done(void)
 786 {
 787         wake_up(&sync_rcu_preempt_exp_wq);
 788 }
 789
 790 /*
 791  * Wait for an rcu-preempt grace period, but expedite it.  The basic idea
 792  * is to rely in the fact that there is but one CPU, and that it is
 793  * illegal for a task to invoke synchronize_rcu_expedited() while in a
 794  * preemptible-RCU read-side critical section.  Therefore, any such
 795  * critical sections must correspond to blocked tasks, which must therefore
 796  * be on the ->blkd_tasks list.  So just record the current head of the
 797  * list in the ->exp_tasks pointer, and wait for all tasks including and
 798  * after the task pointed to by ->exp_tasks to drain.
 799  */
 800 void synchronize_rcu_expedited(void)
 801 {
 802         unsigned long flags;
 803         struct rcu_preempt_ctrlblk *rpcp = &rcu_preempt_ctrlblk;
 804         unsigned long snap;
 805
 806         barrier(); /* ensure prior action seen before grace period. */
 807
 808         WARN_ON_ONCE(rcu_preempt_running_reader());
 809
 810         /*
 811          * Acquire lock so that there is only one preemptible RCU grace
 812          * period in flight.  Of course, if someone does the expedited
 813          * grace period for us while we are acquiring the lock, just leave.
 814          */
 815         snap = sync_rcu_preempt_exp_count + 1;
 816         mutex_lock(&sync_rcu_preempt_exp_mutex);
 817         if (ULONG_CMP_LT(snap, sync_rcu_preempt_exp_count))
 818                 goto unlock_mb_ret; /* Others did our work for us. */
 819
 820         local_irq_save(flags);
 821
 822         /*
 823          * All RCU readers have to already be on blkd_tasks because
 824          * we cannot legally be executing in an RCU read-side critical
 825          * section.
 826          */
 827
 828         /* Snapshot current head of ->blkd_tasks list. */
 829         rpcp->exp_tasks = rpcp->blkd_tasks.next;
 830         if (rpcp->exp_tasks == &rpcp->blkd_tasks)
 831                 rpcp->exp_tasks = NULL;
 832
 833         /* Wait for tail of ->blkd_tasks list to drain. */
 834         if (!rcu_preempted_readers_exp()) {
 835                 local_irq_restore(flags);
 836         } else {
 837                 rcu_initiate_boost();
 838                 local_irq_restore(flags);
 839                 wait_event(sync_rcu_preempt_exp_wq,
 840                            !rcu_preempted_readers_exp());
 841         }
 842
 843         /* Clean up and exit. */
 844         barrier(); /* ensure expedited GP seen before counter increment. */
 845         sync_rcu_preempt_exp_count++;
 846 unlock_mb_ret:
 847         mutex_unlock(&sync_rcu_preempt_exp_mutex);
 848         barrier(); /* ensure subsequent action seen after grace period. */
 849 }
 850 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
 851
 852 /*
 853  * Does preemptible RCU need the CPU to stay out of dynticks mode?
 854  */
 855 int rcu_preempt_needs_cpu(void)
 856 {
 857         return rcu_preempt_ctrlblk.rcb.rcucblist != NULL;
 858 }
 859
 860 #else /* #ifdef CONFIG_TINY_PREEMPT_RCU */
 861
 862 #ifdef CONFIG_RCU_TRACE
 863
 864 /*
 865  * Because preemptible RCU does not exist, it is not necessary to
 866  * dump out its statistics.
 867  */
 868 static void show_tiny_preempt_stats(struct seq_file *m)
 869 {
 870 }
 871
 872 #endif /* #ifdef CONFIG_RCU_TRACE */
 873
 874 /*
 875  * Because preemptible RCU does not exist, it never has any callbacks
 876  * to check.
 877  */
 878 static void rcu_preempt_check_callbacks(void)
 879 {
 880 }
 881
 882 /*
 883  * Because preemptible RCU does not exist, it never has any callbacks
 884  * to remove.
 885  */
 886 static void rcu_preempt_remove_callbacks(struct rcu_ctrlblk *rcp)
 887 {
 888 }
 889
 890 /*
 891  * Because preemptible RCU does not exist, it never has any callbacks
 892  * to process.
 893  */
 894 static void rcu_preempt_process_callbacks(void)
 895 {
 896 }
 897
 898 #endif /* #else #ifdef CONFIG_TINY_PREEMPT_RCU */
 899
 900 #ifdef CONFIG_RCU_BOOST
 901
 902 /*
 903  * Wake up rcu_kthread() to process callbacks now eligible for invocation
 904  * or to boost readers.
 905  */
 906 static void invoke_rcu_callbacks(void)
 907 {
 908         have_rcu_kthread_work = 1;
 909         if (rcu_kthread_task != NULL)
 910                 wake_up(&rcu_kthread_wq);
 911 }
 912
 913 #ifdef CONFIG_RCU_TRACE
 914
 915 /*
 916  * Is the current CPU running the RCU-callbacks kthread?
 917  * Caller must have preemption disabled.
 918  */
 919 static bool rcu_is_callbacks_kthread(void)
 920 {
 921         return rcu_kthread_task == current;
 922 }
 923
 924 #endif /* #ifdef CONFIG_RCU_TRACE */
 925
 926 /*
 927  * This kthread invokes RCU callbacks whose grace periods have
 928  * elapsed.  It is awakened as needed, and takes the place of the
 929  * RCU_SOFTIRQ that is used for this purpose when boosting is disabled.
 930  * This is a kthread, but it is never stopped, at least not until
 931  * the system goes down.
 932  */
 933 static int rcu_kthread(void *arg)
 934 {
 935         unsigned long work;
 936         unsigned long morework;
 937         unsigned long flags;
 938
 939         for (;;) {
 940                 wait_event_interruptible(rcu_kthread_wq,
 941                                          have_rcu_kthread_work != 0);
 942                 morework = rcu_boost();
 943                 local_irq_save(flags);
 944                 work = have_rcu_kthread_work;
 945                 have_rcu_kthread_work = morework;
 946                 local_irq_restore(flags);
 947                 if (work)
 948                         rcu_process_callbacks(NULL);
 949                 schedule_timeout_interruptible(1); /* Leave CPU for others. */
 950         }
 951
 952         return 0;  /* Not reached, but needed to shut gcc up. */
 953 }
 954
 955 /*
 956  * Spawn the kthread that invokes RCU callbacks.
 957  */
 958 static int __init rcu_spawn_kthreads(void)
 959 {
 960         struct sched_param sp;
 961
 962         rcu_kthread_task = kthread_run(rcu_kthread, NULL, "rcu_kthread");
 963         sp.sched_priority = RCU_BOOST_PRIO;
 964         sched_setscheduler_nocheck(rcu_kthread_task, SCHED_FIFO, &sp);
 965         return 0;
 966 }
 967 early_initcall(rcu_spawn_kthreads);
 968
 969 #else /* #ifdef CONFIG_RCU_BOOST */
 970
 971 /* Hold off callback invocation until early_initcall() time. */
 972 static int rcu_scheduler_fully_active __read_mostly;
 973
 974 /*
 975  * Start up softirq processing of callbacks.
 976  */
 977 void invoke_rcu_callbacks(void)
 978 {
 979         if (rcu_scheduler_fully_active)
 980                 raise_softirq(RCU_SOFTIRQ);
 981 }
 982
 983 #ifdef CONFIG_RCU_TRACE
 984
 985 /*
 986  * There is no callback kthread, so this thread is never it.
 987  */
 988 static bool rcu_is_callbacks_kthread(void)
 989 {
 990         return false;
 991 }
 992
 993 #endif /* #ifdef CONFIG_RCU_TRACE */
 994
 995 static int __init rcu_scheduler_really_started(void)
 996 {
 997         rcu_scheduler_fully_active = 1;
 998         open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
 999         raise_softirq(RCU_SOFTIRQ);  /* Invoke any callbacks from early boot. */
1000         return 0;
1001 }
1002 early_initcall(rcu_scheduler_really_started);
1003
1004 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1005
1006 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1007 #include <linux/kernel_stat.h>
1008
1009 /*
1010  * During boot, we forgive RCU lockdep issues.  After this function is
1011  * invoked, we start taking RCU lockdep issues seriously.
1012  */
1013 void __init rcu_scheduler_starting(void)
1014 {
1015         WARN_ON(nr_context_switches() > 0);
1016         rcu_scheduler_active = 1;
1017 }
1018
1019 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1020
1021 #ifdef CONFIG_RCU_TRACE
1022
1023 #ifdef CONFIG_RCU_BOOST
1024
1025 static void rcu_initiate_boost_trace(void)
1026 {
1027         if (list_empty(&rcu_preempt_ctrlblk.blkd_tasks))
1028                 rcu_preempt_ctrlblk.n_balk_blkd_tasks++;
1029         else if (rcu_preempt_ctrlblk.gp_tasks == NULL &&
1030                  rcu_preempt_ctrlblk.exp_tasks == NULL)
1031                 rcu_preempt_ctrlblk.n_balk_exp_gp_tasks++;
1032         else if (rcu_preempt_ctrlblk.boost_tasks != NULL)
1033                 rcu_preempt_ctrlblk.n_balk_boost_tasks++;
1034         else if (!ULONG_CMP_GE(jiffies, rcu_preempt_ctrlblk.boost_time))
1035                 rcu_preempt_ctrlblk.n_balk_notyet++;
1036         else
1037                 rcu_preempt_ctrlblk.n_balk_nos++;
1038 }
1039
1040 #endif /* #ifdef CONFIG_RCU_BOOST */
1041
1042 static void rcu_trace_sub_qlen(struct rcu_ctrlblk *rcp, int n)
1043 {
1044         unsigned long flags;
1045
1046         local_irq_save(flags);
1047         rcp->qlen -= n;
1048         local_irq_restore(flags);
1049 }
1050
1051 /*
1052  * Dump statistics for TINY_RCU, such as they are.
1053  */
1054 static int show_tiny_stats(struct seq_file *m, void *unused)
1055 {
1056         show_tiny_preempt_stats(m);
1057         seq_printf(m, "rcu_sched: qlen: %ld\n", rcu_sched_ctrlblk.qlen);
1058         seq_printf(m, "rcu_bh: qlen: %ld\n", rcu_bh_ctrlblk.qlen);
1059         return 0;
1060 }
1061
1062 static int show_tiny_stats_open(struct inode *inode, struct file *file)
1063 {
1064         return single_open(file, show_tiny_stats, NULL);
1065 }
1066
1067 static const struct file_operations show_tiny_stats_fops = {
1068         .owner = THIS_MODULE,
1069         .open = show_tiny_stats_open,
1070         .read = seq_read,
1071         .llseek = seq_lseek,
1072         .release = single_release,
1073 };
1074
1075 static struct dentry *rcudir;
1076
1077 static int __init rcutiny_trace_init(void)
1078 {
1079         struct dentry *retval;
1080
1081         rcudir = debugfs_create_dir("rcu", NULL);
1082         if (!rcudir)
1083                 goto free_out;
1084         retval = debugfs_create_file("rcudata", 0444, rcudir,
1085                                      NULL, &show_tiny_stats_fops);
1086         if (!retval)
1087                 goto free_out;
1088         return 0;
1089 free_out:
1090         debugfs_remove_recursive(rcudir);
1091         return 1;
1092 }
1093
1094 static void __exit rcutiny_trace_cleanup(void)
1095 {
1096         debugfs_remove_recursive(rcudir);
1097 }
1098
1099 module_init(rcutiny_trace_init);
1100 module_exit(rcutiny_trace_cleanup);
1101
1102 MODULE_AUTHOR("Paul E. McKenney");
1103 MODULE_DESCRIPTION("Read-Copy Update tracing for tiny implementation");
1104 MODULE_LICENSE("GPL");
1105
1106 static void check_cpu_stall_preempt(void)
1107 {
1108 #ifdef CONFIG_TINY_PREEMPT_RCU
1109         check_cpu_stall(&rcu_preempt_ctrlblk.rcb);
1110 #endif /* #ifdef CONFIG_TINY_PREEMPT_RCU */
1111 }
1112
1113 #endif /* #ifdef CONFIG_RCU_TRACE */