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