kernel/rcu/tree.h

   1 /*
   2  * Read-Copy Update mechanism for mutual exclusion (tree-based version)
   3  * Internal non-public definitions.
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, you can access it online at
  17  * http://www.gnu.org/licenses/gpl-2.0.html.
  18  *
  19  * Copyright IBM Corporation, 2008
  20  *
  21  * Author: Ingo Molnar <mingo@elte.hu>
  22  *         Paul E. McKenney <paulmck@linux.vnet.ibm.com>
  23  */
  24
  25 #include <linux/cache.h>
  26 #include <linux/spinlock.h>
  27 #include <linux/rtmutex.h>
  28 #include <linux/threads.h>
  29 #include <linux/cpumask.h>
  30 #include <linux/seqlock.h>
  31 #include <linux/swait.h>
  32 #include <linux/stop_machine.h>
  33 #include <linux/rcu_node_tree.h>
  34
  35 #include "rcu_segcblist.h"
  36
  37 /*
  38  * Dynticks per-CPU state.
  39  */
  40 struct rcu_dynticks {
  41         long long dynticks_nesting; /* Track irq/process nesting level. */
  42                                     /* Process level is worth LLONG_MAX/2. */
  43         int dynticks_nmi_nesting;   /* Track NMI nesting level. */
  44         atomic_t dynticks;          /* Even value for idle, else odd. */
  45         bool rcu_need_heavy_qs;     /* GP old, need heavy quiescent state. */
  46         unsigned long rcu_qs_ctr;   /* Light universal quiescent state ctr. */
  47         bool rcu_urgent_qs;         /* GP old need light quiescent state. */
  48 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
  49         long long dynticks_idle_nesting;
  50                                     /* irq/process nesting level from idle. */
  51         atomic_t dynticks_idle;     /* Even value for idle, else odd. */
  52                                     /*  "Idle" excludes userspace execution. */
  53         unsigned long dynticks_idle_jiffies;
  54                                     /* End of last non-NMI non-idle period. */
  55 #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
  56 #ifdef CONFIG_RCU_FAST_NO_HZ
  57         bool all_lazy;              /* Are all CPU's CBs lazy? */
  58         unsigned long nonlazy_posted;
  59                                     /* # times non-lazy CBs posted to CPU. */
  60         unsigned long nonlazy_posted_snap;
  61                                     /* idle-period nonlazy_posted snapshot. */
  62         unsigned long last_accelerate;
  63                                     /* Last jiffy CBs were accelerated. */
  64         unsigned long last_advance_all;
  65                                     /* Last jiffy CBs were all advanced. */
  66         int tick_nohz_enabled_snap; /* Previously seen value from sysfs. */
  67 #endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
  68 };
  69
  70 /* RCU's kthread states for tracing. */
  71 #define RCU_KTHREAD_STOPPED  0
  72 #define RCU_KTHREAD_RUNNING  1
  73 #define RCU_KTHREAD_WAITING  2
  74 #define RCU_KTHREAD_OFFCPU   3
  75 #define RCU_KTHREAD_YIELDING 4
  76 #define RCU_KTHREAD_MAX      4
  77
  78 /*
  79  * Definition for node within the RCU grace-period-detection hierarchy.
  80  */
  81 struct rcu_node {
  82         raw_spinlock_t __private lock;  /* Root rcu_node's lock protects */
  83                                         /*  some rcu_state fields as well as */
  84                                         /*  following. */
  85         unsigned long gpnum;    /* Current grace period for this node. */
  86                                 /*  This will either be equal to or one */
  87                                 /*  behind the root rcu_node's gpnum. */
  88         unsigned long completed; /* Last GP completed for this node. */
  89                                 /*  This will either be equal to or one */
  90                                 /*  behind the root rcu_node's gpnum. */
  91         unsigned long qsmask;   /* CPUs or groups that need to switch in */
  92                                 /*  order for current grace period to proceed.*/
  93                                 /*  In leaf rcu_node, each bit corresponds to */
  94                                 /*  an rcu_data structure, otherwise, each */
  95                                 /*  bit corresponds to a child rcu_node */
  96                                 /*  structure. */
  97         unsigned long qsmaskinit;
  98                                 /* Per-GP initial value for qsmask. */
  99                                 /*  Initialized from ->qsmaskinitnext at the */
 100                                 /*  beginning of each grace period. */
 101         unsigned long qsmaskinitnext;
 102                                 /* Online CPUs for next grace period. */
 103         unsigned long expmask;  /* CPUs or groups that need to check in */
 104                                 /*  to allow the current expedited GP */
 105                                 /*  to complete. */
 106         unsigned long expmaskinit;
 107                                 /* Per-GP initial values for expmask. */
 108                                 /*  Initialized from ->expmaskinitnext at the */
 109                                 /*  beginning of each expedited GP. */
 110         unsigned long expmaskinitnext;
 111                                 /* Online CPUs for next expedited GP. */
 112                                 /*  Any CPU that has ever been online will */
 113                                 /*  have its bit set. */
 114         unsigned long grpmask;  /* Mask to apply to parent qsmask. */
 115                                 /*  Only one bit will be set in this mask. */
 116         int     grplo;          /* lowest-numbered CPU or group here. */
 117         int     grphi;          /* highest-numbered CPU or group here. */
 118         u8      grpnum;         /* CPU/group number for next level up. */
 119         u8      level;          /* root is at level 0. */
 120         bool    wait_blkd_tasks;/* Necessary to wait for blocked tasks to */
 121                                 /*  exit RCU read-side critical sections */
 122                                 /*  before propagating offline up the */
 123                                 /*  rcu_node tree? */
 124         struct rcu_node *parent;
 125         struct list_head blkd_tasks;
 126                                 /* Tasks blocked in RCU read-side critical */
 127                                 /*  section.  Tasks are placed at the head */
 128                                 /*  of this list and age towards the tail. */
 129         struct list_head *gp_tasks;
 130                                 /* Pointer to the first task blocking the */
 131                                 /*  current grace period, or NULL if there */
 132                                 /*  is no such task. */
 133         struct list_head *exp_tasks;
 134                                 /* Pointer to the first task blocking the */
 135                                 /*  current expedited grace period, or NULL */
 136                                 /*  if there is no such task.  If there */
 137                                 /*  is no current expedited grace period, */
 138                                 /*  then there can cannot be any such task. */
 139         struct list_head *boost_tasks;
 140                                 /* Pointer to first task that needs to be */
 141                                 /*  priority boosted, or NULL if no priority */
 142                                 /*  boosting is needed for this rcu_node */
 143                                 /*  structure.  If there are no tasks */
 144                                 /*  queued on this rcu_node structure that */
 145                                 /*  are blocking the current grace period, */
 146                                 /*  there can be no such task. */
 147         struct rt_mutex boost_mtx;
 148                                 /* Used only for the priority-boosting */
 149                                 /*  side effect, not as a lock. */
 150         unsigned long boost_time;
 151                                 /* When to start boosting (jiffies). */
 152         struct task_struct *boost_kthread_task;
 153                                 /* kthread that takes care of priority */
 154                                 /*  boosting for this rcu_node structure. */
 155         unsigned int boost_kthread_status;
 156                                 /* State of boost_kthread_task for tracing. */
 157         unsigned long n_tasks_boosted;
 158                                 /* Total number of tasks boosted. */
 159         unsigned long n_exp_boosts;
 160                                 /* Number of tasks boosted for expedited GP. */
 161         unsigned long n_normal_boosts;
 162                                 /* Number of tasks boosted for normal GP. */
 163         unsigned long n_balk_blkd_tasks;
 164                                 /* Refused to boost: no blocked tasks. */
 165         unsigned long n_balk_exp_gp_tasks;
 166                                 /* Refused to boost: nothing blocking GP. */
 167         unsigned long n_balk_boost_tasks;
 168                                 /* Refused to boost: already boosting. */
 169         unsigned long n_balk_notblocked;
 170                                 /* Refused to boost: RCU RS CS still running. */
 171         unsigned long n_balk_notyet;
 172                                 /* Refused to boost: not yet time. */
 173         unsigned long n_balk_nos;
 174                                 /* Refused to boost: not sure why, though. */
 175                                 /*  This can happen due to race conditions. */
 176 #ifdef CONFIG_RCU_NOCB_CPU
 177         struct swait_queue_head nocb_gp_wq[2];
 178                                 /* Place for rcu_nocb_kthread() to wait GP. */
 179 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
 180         int need_future_gp[2];
 181                                 /* Counts of upcoming no-CB GP requests. */
 182         raw_spinlock_t fqslock ____cacheline_internodealigned_in_smp;
 183
 184         spinlock_t exp_lock ____cacheline_internodealigned_in_smp;
 185         unsigned long exp_seq_rq;
 186         wait_queue_head_t exp_wq[4];
 187 } ____cacheline_internodealigned_in_smp;
 188
 189 /*
 190  * Bitmasks in an rcu_node cover the interval [grplo, grphi] of CPU IDs, and
 191  * are indexed relative to this interval rather than the global CPU ID space.
 192  * This generates the bit for a CPU in node-local masks.
 193  */
 194 #define leaf_node_cpu_bit(rnp, cpu) (1UL << ((cpu) - (rnp)->grplo))
 195
 196 /*
 197  * Union to allow "aggregate OR" operation on the need for a quiescent
 198  * state by the normal and expedited grace periods.
 199  */
 200 union rcu_noqs {
 201         struct {
 202                 u8 norm;
 203                 u8 exp;
 204         } b; /* Bits. */
 205         u16 s; /* Set of bits, aggregate OR here. */
 206 };
 207
 208 /* Index values for nxttail array in struct rcu_data. */
 209 #define RCU_DONE_TAIL           0       /* Also RCU_WAIT head. */
 210 #define RCU_WAIT_TAIL           1       /* Also RCU_NEXT_READY head. */
 211 #define RCU_NEXT_READY_TAIL     2       /* Also RCU_NEXT head. */
 212 #define RCU_NEXT_TAIL           3
 213 #define RCU_NEXT_SIZE           4
 214
 215 /* Per-CPU data for read-copy update. */
 216 struct rcu_data {
 217         /* 1) quiescent-state and grace-period handling : */
 218         unsigned long   completed;      /* Track rsp->completed gp number */
 219                                         /*  in order to detect GP end. */
 220         unsigned long   gpnum;          /* Highest gp number that this CPU */
 221                                         /*  is aware of having started. */
 222         unsigned long   rcu_qs_ctr_snap;/* Snapshot of rcu_qs_ctr to check */
 223                                         /*  for rcu_all_qs() invocations. */
 224         union rcu_noqs  cpu_no_qs;      /* No QSes yet for this CPU. */
 225         bool            core_needs_qs;  /* Core waits for quiesc state. */
 226         bool            beenonline;     /* CPU online at least once. */
 227         bool            gpwrap;         /* Possible gpnum/completed wrap. */
 228         struct rcu_node *mynode;        /* This CPU's leaf of hierarchy */
 229         unsigned long grpmask;          /* Mask to apply to leaf qsmask. */
 230         unsigned long   ticks_this_gp;  /* The number of scheduling-clock */
 231                                         /*  ticks this CPU has handled */
 232                                         /*  during and after the last grace */
 233                                         /* period it is aware of. */
 234
 235         /* 2) batch handling */
 236         struct rcu_segcblist cblist;    /* Segmented callback list, with */
 237                                         /* different callbacks waiting for */
 238                                         /* different grace periods. */
 239         long            qlen_last_fqs_check;
 240                                         /* qlen at last check for QS forcing */
 241         unsigned long   n_cbs_invoked;  /* count of RCU cbs invoked. */
 242         unsigned long   n_nocbs_invoked; /* count of no-CBs RCU cbs invoked. */
 243         unsigned long   n_cbs_orphaned; /* RCU cbs orphaned by dying CPU */
 244         unsigned long   n_cbs_adopted;  /* RCU cbs adopted from dying CPU */
 245         unsigned long   n_force_qs_snap;
 246                                         /* did other CPU force QS recently? */
 247         long            blimit;         /* Upper limit on a processed batch */
 248
 249         /* 3) dynticks interface. */
 250         struct rcu_dynticks *dynticks;  /* Shared per-CPU dynticks state. */
 251         int dynticks_snap;              /* Per-GP tracking for dynticks. */
 252
 253         /* 4) reasons this CPU needed to be kicked by force_quiescent_state */
 254         unsigned long dynticks_fqs;     /* Kicked due to dynticks idle. */
 255         unsigned long offline_fqs;      /* Kicked due to being offline. */
 256         unsigned long cond_resched_completed;
 257                                         /* Grace period that needs help */
 258                                         /*  from cond_resched(). */
 259
 260         /* 5) __rcu_pending() statistics. */
 261         unsigned long n_rcu_pending;    /* rcu_pending() calls since boot. */
 262         unsigned long n_rp_core_needs_qs;
 263         unsigned long n_rp_report_qs;
 264         unsigned long n_rp_cb_ready;
 265         unsigned long n_rp_cpu_needs_gp;
 266         unsigned long n_rp_gp_completed;
 267         unsigned long n_rp_gp_started;
 268         unsigned long n_rp_nocb_defer_wakeup;
 269         unsigned long n_rp_need_nothing;
 270
 271         /* 6) _rcu_barrier(), OOM callbacks, and expediting. */
 272         struct rcu_head barrier_head;
 273 #ifdef CONFIG_RCU_FAST_NO_HZ
 274         struct rcu_head oom_head;
 275 #endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
 276         atomic_long_t exp_workdone0;    /* # done by workqueue. */
 277         atomic_long_t exp_workdone1;    /* # done by others #1. */
 278         atomic_long_t exp_workdone2;    /* # done by others #2. */
 279         atomic_long_t exp_workdone3;    /* # done by others #3. */
 280         int exp_dynticks_snap;          /* Double-check need for IPI. */
 281
 282         /* 7) Callback offloading. */
 283 #ifdef CONFIG_RCU_NOCB_CPU
 284         struct rcu_head *nocb_head;     /* CBs waiting for kthread. */
 285         struct rcu_head **nocb_tail;
 286         atomic_long_t nocb_q_count;     /* # CBs waiting for nocb */
 287         atomic_long_t nocb_q_count_lazy; /*  invocation (all stages). */
 288         struct rcu_head *nocb_follower_head; /* CBs ready to invoke. */
 289         struct rcu_head **nocb_follower_tail;
 290         struct swait_queue_head nocb_wq; /* For nocb kthreads to sleep on. */
 291         struct task_struct *nocb_kthread;
 292         int nocb_defer_wakeup;          /* Defer wakeup of nocb_kthread. */
 293
 294         /* The following fields are used by the leader, hence own cacheline. */
 295         struct rcu_head *nocb_gp_head ____cacheline_internodealigned_in_smp;
 296                                         /* CBs waiting for GP. */
 297         struct rcu_head **nocb_gp_tail;
 298         bool nocb_leader_sleep;         /* Is the nocb leader thread asleep? */
 299         struct rcu_data *nocb_next_follower;
 300                                         /* Next follower in wakeup chain. */
 301
 302         /* The following fields are used by the follower, hence new cachline. */
 303         struct rcu_data *nocb_leader ____cacheline_internodealigned_in_smp;
 304                                         /* Leader CPU takes GP-end wakeups. */
 305 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
 306
 307         /* 8) RCU CPU stall data. */
 308         unsigned int softirq_snap;      /* Snapshot of softirq activity. */
 309
 310         int cpu;
 311         struct rcu_state *rsp;
 312 };
 313
 314 /* Values for nocb_defer_wakeup field in struct rcu_data. */
 315 #define RCU_NOGP_WAKE_NOT       0
 316 #define RCU_NOGP_WAKE           1
 317 #define RCU_NOGP_WAKE_FORCE     2
 318
 319 #define RCU_JIFFIES_TILL_FORCE_QS (1 + (HZ > 250) + (HZ > 500))
 320                                         /* For jiffies_till_first_fqs and */
 321                                         /*  and jiffies_till_next_fqs. */
 322
 323 #define RCU_JIFFIES_FQS_DIV     256     /* Very large systems need more */
 324                                         /*  delay between bouts of */
 325                                         /*  quiescent-state forcing. */
 326
 327 #define RCU_STALL_RAT_DELAY     2       /* Allow other CPUs time to take */
 328                                         /*  at least one scheduling clock */
 329                                         /*  irq before ratting on them. */
 330
 331 #define rcu_wait(cond)                                                  \
 332 do {                                                                    \
 333         for (;;) {                                                      \
 334                 set_current_state(TASK_INTERRUPTIBLE);                  \
 335                 if (cond)                                               \
 336                         break;                                          \
 337                 schedule();                                             \
 338         }                                                               \
 339         __set_current_state(TASK_RUNNING);                              \
 340 } while (0)
 341
 342 /*
 343  * RCU global state, including node hierarchy.  This hierarchy is
 344  * represented in "heap" form in a dense array.  The root (first level)
 345  * of the hierarchy is in ->node[0] (referenced by ->level[0]), the second
 346  * level in ->node[1] through ->node[m] (->node[1] referenced by ->level[1]),
 347  * and the third level in ->node[m+1] and following (->node[m+1] referenced
 348  * by ->level[2]).  The number of levels is determined by the number of
 349  * CPUs and by CONFIG_RCU_FANOUT.  Small systems will have a "hierarchy"
 350  * consisting of a single rcu_node.
 351  */
 352 struct rcu_state {
 353         struct rcu_node node[NUM_RCU_NODES];    /* Hierarchy. */
 354         struct rcu_node *level[RCU_NUM_LVLS + 1];
 355                                                 /* Hierarchy levels (+1 to */
 356                                                 /*  shut bogus gcc warning) */
 357         struct rcu_data __percpu *rda;          /* pointer of percu rcu_data. */
 358         call_rcu_func_t call;                   /* call_rcu() flavor. */
 359         int ncpus;                              /* # CPUs seen so far. */
 360
 361         /* The following fields are guarded by the root rcu_node's lock. */
 362
 363         u8      boost ____cacheline_internodealigned_in_smp;
 364                                                 /* Subject to priority boost. */
 365         unsigned long gpnum;                    /* Current gp number. */
 366         unsigned long completed;                /* # of last completed gp. */
 367         struct task_struct *gp_kthread;         /* Task for grace periods. */
 368         struct swait_queue_head gp_wq;          /* Where GP task waits. */
 369         short gp_flags;                         /* Commands for GP task. */
 370         short gp_state;                         /* GP kthread sleep state. */
 371
 372         /* End of fields guarded by root rcu_node's lock. */
 373
 374         raw_spinlock_t orphan_lock ____cacheline_internodealigned_in_smp;
 375                                                 /* Protect following fields. */
 376         struct rcu_cblist orphan_pend;          /* Orphaned callbacks that */
 377                                                 /*  need a grace period. */
 378         struct rcu_cblist orphan_done;          /* Orphaned callbacks that */
 379                                                 /*  are ready to invoke. */
 380                                                 /* (Contains counts.) */
 381         /* End of fields guarded by orphan_lock. */
 382
 383         struct mutex barrier_mutex;             /* Guards barrier fields. */
 384         atomic_t barrier_cpu_count;             /* # CPUs waiting on. */
 385         struct completion barrier_completion;   /* Wake at barrier end. */
 386         unsigned long barrier_sequence;         /* ++ at start and end of */
 387                                                 /*  _rcu_barrier(). */
 388         /* End of fields guarded by barrier_mutex. */
 389
 390         struct mutex exp_mutex;                 /* Serialize expedited GP. */
 391         struct mutex exp_wake_mutex;            /* Serialize wakeup. */
 392         unsigned long expedited_sequence;       /* Take a ticket. */
 393         atomic_t expedited_need_qs;             /* # CPUs left to check in. */
 394         struct swait_queue_head expedited_wq;   /* Wait for check-ins. */
 395         int ncpus_snap;                         /* # CPUs seen last time. */
 396
 397         unsigned long jiffies_force_qs;         /* Time at which to invoke */
 398                                                 /*  force_quiescent_state(). */
 399         unsigned long jiffies_kick_kthreads;    /* Time at which to kick */
 400                                                 /*  kthreads, if configured. */
 401         unsigned long n_force_qs;               /* Number of calls to */
 402                                                 /*  force_quiescent_state(). */
 403         unsigned long n_force_qs_lh;            /* ~Number of calls leaving */
 404                                                 /*  due to lock unavailable. */
 405         unsigned long n_force_qs_ngp;           /* Number of calls leaving */
 406                                                 /*  due to no GP active. */
 407         unsigned long gp_start;                 /* Time at which GP started, */
 408                                                 /*  but in jiffies. */
 409         unsigned long gp_activity;              /* Time of last GP kthread */
 410                                                 /*  activity in jiffies. */
 411         unsigned long jiffies_stall;            /* Time at which to check */
 412                                                 /*  for CPU stalls. */
 413         unsigned long jiffies_resched;          /* Time at which to resched */
 414                                                 /*  a reluctant CPU. */
 415         unsigned long n_force_qs_gpstart;       /* Snapshot of n_force_qs at */
 416                                                 /*  GP start. */
 417         unsigned long gp_max;                   /* Maximum GP duration in */
 418                                                 /*  jiffies. */
 419         const char *name;                       /* Name of structure. */
 420         char abbr;                              /* Abbreviated name. */
 421         struct list_head flavors;               /* List of RCU flavors. */
 422 };
 423
 424 /* Values for rcu_state structure's gp_flags field. */
 425 #define RCU_GP_FLAG_INIT 0x1    /* Need grace-period initialization. */
 426 #define RCU_GP_FLAG_FQS  0x2    /* Need grace-period quiescent-state forcing. */
 427
 428 /* Values for rcu_state structure's gp_state field. */
 429 #define RCU_GP_IDLE      0      /* Initial state and no GP in progress. */
 430 #define RCU_GP_WAIT_GPS  1      /* Wait for grace-period start. */
 431 #define RCU_GP_DONE_GPS  2      /* Wait done for grace-period start. */
 432 #define RCU_GP_WAIT_FQS  3      /* Wait for force-quiescent-state time. */
 433 #define RCU_GP_DOING_FQS 4      /* Wait done for force-quiescent-state time. */
 434 #define RCU_GP_CLEANUP   5      /* Grace-period cleanup started. */
 435 #define RCU_GP_CLEANED   6      /* Grace-period cleanup complete. */
 436
 437 #ifndef RCU_TREE_NONCORE
 438 static const char * const gp_state_names[] = {
 439         "RCU_GP_IDLE",
 440         "RCU_GP_WAIT_GPS",
 441         "RCU_GP_DONE_GPS",
 442         "RCU_GP_WAIT_FQS",
 443         "RCU_GP_DOING_FQS",
 444         "RCU_GP_CLEANUP",
 445         "RCU_GP_CLEANED",
 446 };
 447 #endif /* #ifndef RCU_TREE_NONCORE */
 448
 449 extern struct list_head rcu_struct_flavors;
 450
 451 /* Sequence through rcu_state structures for each RCU flavor. */
 452 #define for_each_rcu_flavor(rsp) \
 453         list_for_each_entry((rsp), &rcu_struct_flavors, flavors)
 454
 455 /*
 456  * RCU implementation internal declarations:
 457  */
 458 extern struct rcu_state rcu_sched_state;
 459
 460 extern struct rcu_state rcu_bh_state;
 461
 462 #ifdef CONFIG_PREEMPT_RCU
 463 extern struct rcu_state rcu_preempt_state;
 464 #endif /* #ifdef CONFIG_PREEMPT_RCU */
 465
 466 int rcu_dynticks_snap(struct rcu_dynticks *rdtp);
 467 bool rcu_eqs_special_set(int cpu);
 468
 469 #ifdef CONFIG_RCU_BOOST
 470 DECLARE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
 471 DECLARE_PER_CPU(int, rcu_cpu_kthread_cpu);
 472 DECLARE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
 473 DECLARE_PER_CPU(char, rcu_cpu_has_work);
 474 #endif /* #ifdef CONFIG_RCU_BOOST */
 475
 476 #ifndef RCU_TREE_NONCORE
 477
 478 /* Forward declarations for rcutree_plugin.h */
 479 static void rcu_bootup_announce(void);
 480 static void rcu_preempt_note_context_switch(void);
 481 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp);
 482 #ifdef CONFIG_HOTPLUG_CPU
 483 static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
 484 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
 485 static void rcu_print_detail_task_stall(struct rcu_state *rsp);
 486 static int rcu_print_task_stall(struct rcu_node *rnp);
 487 static int rcu_print_task_exp_stall(struct rcu_node *rnp);
 488 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
 489 static void rcu_preempt_check_callbacks(void);
 490 void call_rcu(struct rcu_head *head, rcu_callback_t func);
 491 static void __init __rcu_init_preempt(void);
 492 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags);
 493 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
 494 static void invoke_rcu_callbacks_kthread(void);
 495 static bool rcu_is_callbacks_kthread(void);
 496 #ifdef CONFIG_RCU_BOOST
 497 static void rcu_preempt_do_callbacks(void);
 498 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
 499                                                  struct rcu_node *rnp);
 500 #endif /* #ifdef CONFIG_RCU_BOOST */
 501 static void __init rcu_spawn_boost_kthreads(void);
 502 static void rcu_prepare_kthreads(int cpu);
 503 static void rcu_cleanup_after_idle(void);
 504 static void rcu_prepare_for_idle(void);
 505 static void rcu_idle_count_callbacks_posted(void);
 506 static bool rcu_preempt_has_tasks(struct rcu_node *rnp);
 507 static void print_cpu_stall_info_begin(void);
 508 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu);
 509 static void print_cpu_stall_info_end(void);
 510 static void zero_cpu_stall_ticks(struct rcu_data *rdp);
 511 static void increment_cpu_stall_ticks(void);
 512 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu);
 513 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq);
 514 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp);
 515 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq);
 516 static void rcu_init_one_nocb(struct rcu_node *rnp);
 517 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
 518                             bool lazy, unsigned long flags);
 519 static bool rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
 520                                       struct rcu_data *rdp,
 521                                       unsigned long flags);
 522 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp);
 523 static void do_nocb_deferred_wakeup(struct rcu_data *rdp);
 524 static void rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp);
 525 static void rcu_spawn_all_nocb_kthreads(int cpu);
 526 static void __init rcu_spawn_nocb_kthreads(void);
 527 #ifdef CONFIG_RCU_NOCB_CPU
 528 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp);
 529 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
 530 static void __maybe_unused rcu_kick_nohz_cpu(int cpu);
 531 static bool init_nocb_callback_list(struct rcu_data *rdp);
 532 static void rcu_sysidle_enter(int irq);
 533 static void rcu_sysidle_exit(int irq);
 534 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
 535                                   unsigned long *maxj);
 536 static bool is_sysidle_rcu_state(struct rcu_state *rsp);
 537 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
 538                                   unsigned long maxj);
 539 static void rcu_bind_gp_kthread(void);
 540 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp);
 541 static bool rcu_nohz_full_cpu(struct rcu_state *rsp);
 542 static void rcu_dynticks_task_enter(void);
 543 static void rcu_dynticks_task_exit(void);
 544
 545 #ifdef CONFIG_SRCU
 546 void srcu_online_cpu(unsigned int cpu);
 547 void srcu_offline_cpu(unsigned int cpu);
 548 #else /* #ifdef CONFIG_SRCU */
 549 void srcu_online_cpu(unsigned int cpu) { }
 550 void srcu_offline_cpu(unsigned int cpu) { }
 551 #endif /* #else #ifdef CONFIG_SRCU */
 552
 553 #endif /* #ifndef RCU_TREE_NONCORE */
 554
 555 #ifdef CONFIG_RCU_TRACE
 556 /* Read out queue lengths for tracing. */
 557 static inline void rcu_nocb_q_lengths(struct rcu_data *rdp, long *ql, long *qll)
 558 {
 559 #ifdef CONFIG_RCU_NOCB_CPU
 560         *ql = atomic_long_read(&rdp->nocb_q_count);
 561         *qll = atomic_long_read(&rdp->nocb_q_count_lazy);
 562 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
 563         *ql = 0;
 564         *qll = 0;
 565 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
 566 }
 567 #endif /* #ifdef CONFIG_RCU_TRACE */
 568
 569 /*
 570  * Wrappers for the rcu_node::lock acquire and release.
 571  *
 572  * Because the rcu_nodes form a tree, the tree traversal locking will observe
 573  * different lock values, this in turn means that an UNLOCK of one level
 574  * followed by a LOCK of another level does not imply a full memory barrier;
 575  * and most importantly transitivity is lost.
 576  *
 577  * In order to restore full ordering between tree levels, augment the regular
 578  * lock acquire functions with smp_mb__after_unlock_lock().
 579  *
 580  * As ->lock of struct rcu_node is a __private field, therefore one should use
 581  * these wrappers rather than directly call raw_spin_{lock,unlock}* on ->lock.
 582  */
 583 static inline void raw_spin_lock_rcu_node(struct rcu_node *rnp)
 584 {
 585         raw_spin_lock(&ACCESS_PRIVATE(rnp, lock));
 586         smp_mb__after_unlock_lock();
 587 }
 588
 589 static inline void raw_spin_unlock_rcu_node(struct rcu_node *rnp)
 590 {
 591         raw_spin_unlock(&ACCESS_PRIVATE(rnp, lock));
 592 }
 593
 594 static inline void raw_spin_lock_irq_rcu_node(struct rcu_node *rnp)
 595 {
 596         raw_spin_lock_irq(&ACCESS_PRIVATE(rnp, lock));
 597         smp_mb__after_unlock_lock();
 598 }
 599
 600 static inline void raw_spin_unlock_irq_rcu_node(struct rcu_node *rnp)
 601 {
 602         raw_spin_unlock_irq(&ACCESS_PRIVATE(rnp, lock));
 603 }
 604
 605 #define raw_spin_lock_irqsave_rcu_node(rnp, flags)                      \
 606 do {                                                                    \
 607         typecheck(unsigned long, flags);                                \
 608         raw_spin_lock_irqsave(&ACCESS_PRIVATE(rnp, lock), flags);       \
 609         smp_mb__after_unlock_lock();                                    \
 610 } while (0)
 611
 612 #define raw_spin_unlock_irqrestore_rcu_node(rnp, flags)                 \
 613 do {                                                                    \
 614         typecheck(unsigned long, flags);                                \
 615         raw_spin_unlock_irqrestore(&ACCESS_PRIVATE(rnp, lock), flags);  \
 616 } while (0)
 617
 618 static inline bool raw_spin_trylock_rcu_node(struct rcu_node *rnp)
 619 {
 620         bool locked = raw_spin_trylock(&ACCESS_PRIVATE(rnp, lock));
 621
 622         if (locked)
 623                 smp_mb__after_unlock_lock();
 624         return locked;
 625 }