kernel/rcutree.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, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  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/threads.h>
  28 #include <linux/cpumask.h>
  29 #include <linux/seqlock.h>
  30
  31 /*
  32  * Define shape of hierarchy based on NR_CPUS and CONFIG_RCU_FANOUT.
  33  * In theory, it should be possible to add more levels straightforwardly.
  34  * In practice, this has not been tested, so there is probably some
  35  * bug somewhere.
  36  */
  37 #define MAX_RCU_LVLS 4
  38 #define RCU_FANOUT            (CONFIG_RCU_FANOUT)
  39 #define RCU_FANOUT_SQ         (RCU_FANOUT * RCU_FANOUT)
  40 #define RCU_FANOUT_CUBE       (RCU_FANOUT_SQ * RCU_FANOUT)
  41 #define RCU_FANOUT_FOURTH     (RCU_FANOUT_CUBE * RCU_FANOUT)
  42
  43 #if NR_CPUS <= RCU_FANOUT
  44 #  define NUM_RCU_LVLS        1
  45 #  define NUM_RCU_LVL_0       1
  46 #  define NUM_RCU_LVL_1       (NR_CPUS)
  47 #  define NUM_RCU_LVL_2       0
  48 #  define NUM_RCU_LVL_3       0
  49 #  define NUM_RCU_LVL_4       0
  50 #elif NR_CPUS <= RCU_FANOUT_SQ
  51 #  define NUM_RCU_LVLS        2
  52 #  define NUM_RCU_LVL_0       1
  53 #  define NUM_RCU_LVL_1       DIV_ROUND_UP(NR_CPUS, RCU_FANOUT)
  54 #  define NUM_RCU_LVL_2       (NR_CPUS)
  55 #  define NUM_RCU_LVL_3       0
  56 #  define NUM_RCU_LVL_4       0
  57 #elif NR_CPUS <= RCU_FANOUT_CUBE
  58 #  define NUM_RCU_LVLS        3
  59 #  define NUM_RCU_LVL_0       1
  60 #  define NUM_RCU_LVL_1       DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_SQ)
  61 #  define NUM_RCU_LVL_2       DIV_ROUND_UP(NR_CPUS, RCU_FANOUT)
  62 #  define NUM_RCU_LVL_3       NR_CPUS
  63 #  define NUM_RCU_LVL_4       0
  64 #elif NR_CPUS <= RCU_FANOUT_FOURTH
  65 #  define NUM_RCU_LVLS        4
  66 #  define NUM_RCU_LVL_0       1
  67 #  define NUM_RCU_LVL_1       DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_CUBE)
  68 #  define NUM_RCU_LVL_2       DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_SQ)
  69 #  define NUM_RCU_LVL_3       DIV_ROUND_UP(NR_CPUS, RCU_FANOUT)
  70 #  define NUM_RCU_LVL_4       NR_CPUS
  71 #else
  72 # error "CONFIG_RCU_FANOUT insufficient for NR_CPUS"
  73 #endif /* #if (NR_CPUS) <= RCU_FANOUT */
  74
  75 #define RCU_SUM (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2 + NUM_RCU_LVL_3 + NUM_RCU_LVL_4)
  76 #define NUM_RCU_NODES (RCU_SUM - NR_CPUS)
  77
  78 /*
  79  * Dynticks per-CPU state.
  80  */
  81 struct rcu_dynticks {
  82         int dynticks_nesting;   /* Track nesting level, sort of. */
  83         int dynticks;           /* Even value for dynticks-idle, else odd. */
  84         int dynticks_nmi;       /* Even value for either dynticks-idle or */
  85                                 /*  not in nmi handler, else odd.  So this */
  86                                 /*  remains even for nmi from irq handler. */
  87 };
  88
  89 /*
  90  * Definition for node within the RCU grace-period-detection hierarchy.
  91  */
  92 struct rcu_node {
  93         raw_spinlock_t lock;    /* Root rcu_node's lock protects some */
  94                                 /*  rcu_state fields as well as following. */
  95         unsigned long gpnum;    /* Current grace period for this node. */
  96                                 /*  This will either be equal to or one */
  97                                 /*  behind the root rcu_node's gpnum. */
  98         unsigned long completed; /* Last GP completed for this node. */
  99                                 /*  This will either be equal to or one */
 100                                 /*  behind the root rcu_node's gpnum. */
 101         unsigned long qsmask;   /* CPUs or groups that need to switch in */
 102                                 /*  order for current grace period to proceed.*/
 103                                 /*  In leaf rcu_node, each bit corresponds to */
 104                                 /*  an rcu_data structure, otherwise, each */
 105                                 /*  bit corresponds to a child rcu_node */
 106                                 /*  structure. */
 107         unsigned long expmask;  /* Groups that have ->blocked_tasks[] */
 108                                 /*  elements that need to drain to allow the */
 109                                 /*  current expedited grace period to */
 110                                 /*  complete (only for TREE_PREEMPT_RCU). */
 111         unsigned long qsmaskinit;
 112                                 /* Per-GP initial value for qsmask & expmask. */
 113         unsigned long grpmask;  /* Mask to apply to parent qsmask. */
 114                                 /*  Only one bit will be set in this mask. */
 115         int     grplo;          /* lowest-numbered CPU or group here. */
 116         int     grphi;          /* highest-numbered CPU or group here. */
 117         u8      grpnum;         /* CPU/group number for next level up. */
 118         u8      level;          /* root is at level 0. */
 119         struct rcu_node *parent;
 120         struct list_head blocked_tasks[4];
 121                                 /* Tasks blocked in RCU read-side critsect. */
 122                                 /*  Grace period number (->gpnum) x blocked */
 123                                 /*  by tasks on the (x & 0x1) element of the */
 124                                 /*  blocked_tasks[] array. */
 125 } ____cacheline_internodealigned_in_smp;
 126
 127 /*
 128  * Do a full breadth-first scan of the rcu_node structures for the
 129  * specified rcu_state structure.
 130  */
 131 #define rcu_for_each_node_breadth_first(rsp, rnp) \
 132         for ((rnp) = &(rsp)->node[0]; \
 133              (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
 134
 135 /*
 136  * Do a breadth-first scan of the non-leaf rcu_node structures for the
 137  * specified rcu_state structure.  Note that if there is a singleton
 138  * rcu_node tree with but one rcu_node structure, this loop is a no-op.
 139  */
 140 #define rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) \
 141         for ((rnp) = &(rsp)->node[0]; \
 142              (rnp) < (rsp)->level[NUM_RCU_LVLS - 1]; (rnp)++)
 143
 144 /*
 145  * Scan the leaves of the rcu_node hierarchy for the specified rcu_state
 146  * structure.  Note that if there is a singleton rcu_node tree with but
 147  * one rcu_node structure, this loop -will- visit the rcu_node structure.
 148  * It is still a leaf node, even if it is also the root node.
 149  */
 150 #define rcu_for_each_leaf_node(rsp, rnp) \
 151         for ((rnp) = (rsp)->level[NUM_RCU_LVLS - 1]; \
 152              (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
 153
 154 /* Index values for nxttail array in struct rcu_data. */
 155 #define RCU_DONE_TAIL           0       /* Also RCU_WAIT head. */
 156 #define RCU_WAIT_TAIL           1       /* Also RCU_NEXT_READY head. */
 157 #define RCU_NEXT_READY_TAIL     2       /* Also RCU_NEXT head. */
 158 #define RCU_NEXT_TAIL           3
 159 #define RCU_NEXT_SIZE           4
 160
 161 /* Per-CPU data for read-copy update. */
 162 struct rcu_data {
 163         /* 1) quiescent-state and grace-period handling : */
 164         unsigned long   completed;      /* Track rsp->completed gp number */
 165                                         /*  in order to detect GP end. */
 166         unsigned long   gpnum;          /* Highest gp number that this CPU */
 167                                         /*  is aware of having started. */
 168         unsigned long   passed_quiesc_completed;
 169                                         /* Value of completed at time of qs. */
 170         bool            passed_quiesc;  /* User-mode/idle loop etc. */
 171         bool            qs_pending;     /* Core waits for quiesc state. */
 172         bool            beenonline;     /* CPU online at least once. */
 173         bool            preemptable;    /* Preemptable RCU? */
 174         struct rcu_node *mynode;        /* This CPU's leaf of hierarchy */
 175         unsigned long grpmask;          /* Mask to apply to leaf qsmask. */
 176
 177         /* 2) batch handling */
 178         /*
 179          * If nxtlist is not NULL, it is partitioned as follows.
 180          * Any of the partitions might be empty, in which case the
 181          * pointer to that partition will be equal to the pointer for
 182          * the following partition.  When the list is empty, all of
 183          * the nxttail elements point to the ->nxtlist pointer itself,
 184          * which in that case is NULL.
 185          *
 186          * [nxtlist, *nxttail[RCU_DONE_TAIL]):
 187          *      Entries that batch # <= ->completed
 188          *      The grace period for these entries has completed, and
 189          *      the other grace-period-completed entries may be moved
 190          *      here temporarily in rcu_process_callbacks().
 191          * [*nxttail[RCU_DONE_TAIL], *nxttail[RCU_WAIT_TAIL]):
 192          *      Entries that batch # <= ->completed - 1: waiting for current GP
 193          * [*nxttail[RCU_WAIT_TAIL], *nxttail[RCU_NEXT_READY_TAIL]):
 194          *      Entries known to have arrived before current GP ended
 195          * [*nxttail[RCU_NEXT_READY_TAIL], *nxttail[RCU_NEXT_TAIL]):
 196          *      Entries that might have arrived after current GP ended
 197          *      Note that the value of *nxttail[RCU_NEXT_TAIL] will
 198          *      always be NULL, as this is the end of the list.
 199          */
 200         struct rcu_head *nxtlist;
 201         struct rcu_head **nxttail[RCU_NEXT_SIZE];
 202         long            qlen;           /* # of queued callbacks */
 203         long            qlen_last_fqs_check;
 204                                         /* qlen at last check for QS forcing */
 205         unsigned long   n_force_qs_snap;
 206                                         /* did other CPU force QS recently? */
 207         long            blimit;         /* Upper limit on a processed batch */
 208
 209 #ifdef CONFIG_NO_HZ
 210         /* 3) dynticks interface. */
 211         struct rcu_dynticks *dynticks;  /* Shared per-CPU dynticks state. */
 212         int dynticks_snap;              /* Per-GP tracking for dynticks. */
 213         int dynticks_nmi_snap;          /* Per-GP tracking for dynticks_nmi. */
 214 #endif /* #ifdef CONFIG_NO_HZ */
 215
 216         /* 4) reasons this CPU needed to be kicked by force_quiescent_state */
 217 #ifdef CONFIG_NO_HZ
 218         unsigned long dynticks_fqs;     /* Kicked due to dynticks idle. */
 219 #endif /* #ifdef CONFIG_NO_HZ */
 220         unsigned long offline_fqs;      /* Kicked due to being offline. */
 221         unsigned long resched_ipi;      /* Sent a resched IPI. */
 222
 223         /* 5) __rcu_pending() statistics. */
 224         unsigned long n_rcu_pending;    /* rcu_pending() calls since boot. */
 225         unsigned long n_rp_qs_pending;
 226         unsigned long n_rp_report_qs;
 227         unsigned long n_rp_cb_ready;
 228         unsigned long n_rp_cpu_needs_gp;
 229         unsigned long n_rp_gp_completed;
 230         unsigned long n_rp_gp_started;
 231         unsigned long n_rp_need_fqs;
 232         unsigned long n_rp_need_nothing;
 233
 234         int cpu;
 235 };
 236
 237 /* Values for signaled field in struct rcu_state. */
 238 #define RCU_GP_IDLE             0       /* No grace period in progress. */
 239 #define RCU_GP_INIT             1       /* Grace period being initialized. */
 240 #define RCU_SAVE_DYNTICK        2       /* Need to scan dyntick state. */
 241 #define RCU_FORCE_QS            3       /* Need to force quiescent state. */
 242 #ifdef CONFIG_NO_HZ
 243 #define RCU_SIGNAL_INIT         RCU_SAVE_DYNTICK
 244 #else /* #ifdef CONFIG_NO_HZ */
 245 #define RCU_SIGNAL_INIT         RCU_FORCE_QS
 246 #endif /* #else #ifdef CONFIG_NO_HZ */
 247
 248 #define RCU_JIFFIES_TILL_FORCE_QS        3      /* for rsp->jiffies_force_qs */
 249 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
 250
 251 #ifdef CONFIG_PROVE_RCU
 252 #define RCU_STALL_DELAY_DELTA          (5 * HZ)
 253 #else
 254 #define RCU_STALL_DELAY_DELTA          0
 255 #endif
 256
 257 #define RCU_SECONDS_TILL_STALL_CHECK   (10 * HZ + RCU_STALL_DELAY_DELTA)
 258                                                 /* for rsp->jiffies_stall */
 259 #define RCU_SECONDS_TILL_STALL_RECHECK (30 * HZ + RCU_STALL_DELAY_DELTA)
 260                                                 /* for rsp->jiffies_stall */
 261 #define RCU_STALL_RAT_DELAY             2       /* Allow other CPUs time */
 262                                                 /*  to take at least one */
 263                                                 /*  scheduling clock irq */
 264                                                 /*  before ratting on them. */
 265
 266 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
 267
 268 #define ULONG_CMP_GE(a, b)      (ULONG_MAX / 2 >= (a) - (b))
 269 #define ULONG_CMP_LT(a, b)      (ULONG_MAX / 2 < (a) - (b))
 270
 271 /*
 272  * RCU global state, including node hierarchy.  This hierarchy is
 273  * represented in "heap" form in a dense array.  The root (first level)
 274  * of the hierarchy is in ->node[0] (referenced by ->level[0]), the second
 275  * level in ->node[1] through ->node[m] (->node[1] referenced by ->level[1]),
 276  * and the third level in ->node[m+1] and following (->node[m+1] referenced
 277  * by ->level[2]).  The number of levels is determined by the number of
 278  * CPUs and by CONFIG_RCU_FANOUT.  Small systems will have a "hierarchy"
 279  * consisting of a single rcu_node.
 280  */
 281 struct rcu_state {
 282         struct rcu_node node[NUM_RCU_NODES];    /* Hierarchy. */
 283         struct rcu_node *level[NUM_RCU_LVLS];   /* Hierarchy levels. */
 284         u32 levelcnt[MAX_RCU_LVLS + 1];         /* # nodes in each level. */
 285         u8 levelspread[NUM_RCU_LVLS];           /* kids/node in each level. */
 286         struct rcu_data *rda[NR_CPUS];          /* array of rdp pointers. */
 287
 288         /* The following fields are guarded by the root rcu_node's lock. */
 289
 290         u8      signaled ____cacheline_internodealigned_in_smp;
 291                                                 /* Force QS state. */
 292         u8      fqs_active;                     /* force_quiescent_state() */
 293                                                 /*  is running. */
 294         u8      fqs_need_gp;                    /* A CPU was prevented from */
 295                                                 /*  starting a new grace */
 296                                                 /*  period because */
 297                                                 /*  force_quiescent_state() */
 298                                                 /*  was running. */
 299         unsigned long gpnum;                    /* Current gp number. */
 300         unsigned long completed;                /* # of last completed gp. */
 301
 302         /* End of fields guarded by root rcu_node's lock. */
 303
 304         raw_spinlock_t onofflock;               /* exclude on/offline and */
 305                                                 /*  starting new GP.  Also */
 306                                                 /*  protects the following */
 307                                                 /*  orphan_cbs fields. */
 308         struct rcu_head *orphan_cbs_list;       /* list of rcu_head structs */
 309                                                 /*  orphaned by all CPUs in */
 310                                                 /*  a given leaf rcu_node */
 311                                                 /*  going offline. */
 312         struct rcu_head **orphan_cbs_tail;      /* And tail pointer. */
 313         long orphan_qlen;                       /* Number of orphaned cbs. */
 314         raw_spinlock_t fqslock;                 /* Only one task forcing */
 315                                                 /*  quiescent states. */
 316         unsigned long jiffies_force_qs;         /* Time at which to invoke */
 317                                                 /*  force_quiescent_state(). */
 318         unsigned long n_force_qs;               /* Number of calls to */
 319                                                 /*  force_quiescent_state(). */
 320         unsigned long n_force_qs_lh;            /* ~Number of calls leaving */
 321                                                 /*  due to lock unavailable. */
 322         unsigned long n_force_qs_ngp;           /* Number of calls leaving */
 323                                                 /*  due to no GP active. */
 324 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
 325         unsigned long gp_start;                 /* Time at which GP started, */
 326                                                 /*  but in jiffies. */
 327         unsigned long jiffies_stall;            /* Time at which to check */
 328                                                 /*  for CPU stalls. */
 329 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
 330         char *name;                             /* Name of structure. */
 331 };
 332
 333 /* Return values for rcu_preempt_offline_tasks(). */
 334
 335 #define RCU_OFL_TASKS_NORM_GP   0x1             /* Tasks blocking normal */
 336                                                 /*  GP were moved to root. */
 337 #define RCU_OFL_TASKS_EXP_GP    0x2             /* Tasks blocking expedited */
 338                                                 /*  GP were moved to root. */
 339
 340 /*
 341  * RCU implementation internal declarations:
 342  */
 343 extern struct rcu_state rcu_sched_state;
 344 DECLARE_PER_CPU(struct rcu_data, rcu_sched_data);
 345
 346 extern struct rcu_state rcu_bh_state;
 347 DECLARE_PER_CPU(struct rcu_data, rcu_bh_data);
 348
 349 #ifdef CONFIG_TREE_PREEMPT_RCU
 350 extern struct rcu_state rcu_preempt_state;
 351 DECLARE_PER_CPU(struct rcu_data, rcu_preempt_data);
 352 #endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
 353
 354 #ifndef RCU_TREE_NONCORE
 355
 356 /* Forward declarations for rcutree_plugin.h */
 357 static void rcu_bootup_announce(void);
 358 long rcu_batches_completed(void);
 359 static void rcu_preempt_note_context_switch(int cpu);
 360 static int rcu_preempted_readers(struct rcu_node *rnp);
 361 #ifdef CONFIG_HOTPLUG_CPU
 362 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp,
 363                                       unsigned long flags);
 364 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
 365 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
 366 static void rcu_print_detail_task_stall(struct rcu_state *rsp);
 367 static void rcu_print_task_stall(struct rcu_node *rnp);
 368 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
 369 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
 370 #ifdef CONFIG_HOTPLUG_CPU
 371 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
 372                                      struct rcu_node *rnp,
 373                                      struct rcu_data *rdp);
 374 static void rcu_preempt_offline_cpu(int cpu);
 375 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
 376 static void rcu_preempt_check_callbacks(int cpu);
 377 static void rcu_preempt_process_callbacks(void);
 378 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu));
 379 #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_TREE_PREEMPT_RCU)
 380 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp);
 381 #endif /* #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_TREE_PREEMPT_RCU) */
 382 static int rcu_preempt_pending(int cpu);
 383 static int rcu_preempt_needs_cpu(int cpu);
 384 static void __cpuinit rcu_preempt_init_percpu_data(int cpu);
 385 static void rcu_preempt_send_cbs_to_orphanage(void);
 386 static void __init __rcu_init_preempt(void);
 387 static void rcu_needs_cpu_flush(void);
 388
 389 #endif /* #ifndef RCU_TREE_NONCORE */