kernel/rcu/sync.c

   1 /*
   2  * RCU-based infrastructure for lightweight reader-writer locking
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, you can access it online at
  16  * http://www.gnu.org/licenses/gpl-2.0.html.
  17  *
  18  * Copyright (c) 2015, Red Hat, Inc.
  19  *
  20  * Author: Oleg Nesterov <oleg@redhat.com>
  21  */
  22
  23 #include <linux/rcu_sync.h>
  24 #include <linux/sched.h>
  25
  26 #ifdef CONFIG_PROVE_RCU
  27 #define __INIT_HELD(func)       .held = func,
  28 #else
  29 #define __INIT_HELD(func)
  30 #endif
  31
  32 static const struct {
  33         void (*sync)(void);
  34         void (*call)(struct rcu_head *, void (*)(struct rcu_head *));
  35         void (*wait)(void);
  36 #ifdef CONFIG_PROVE_RCU
  37         int  (*held)(void);
  38 #endif
  39 } gp_ops[] = {
  40         [RCU_SYNC] = {
  41                 .sync = synchronize_rcu,
  42                 .call = call_rcu,
  43                 .wait = rcu_barrier,
  44                 __INIT_HELD(rcu_read_lock_held)
  45         },
  46         [RCU_SCHED_SYNC] = {
  47                 .sync = synchronize_sched,
  48                 .call = call_rcu_sched,
  49                 .wait = rcu_barrier_sched,
  50                 __INIT_HELD(rcu_read_lock_sched_held)
  51         },
  52         [RCU_BH_SYNC] = {
  53                 .sync = synchronize_rcu_bh,
  54                 .call = call_rcu_bh,
  55                 .wait = rcu_barrier_bh,
  56                 __INIT_HELD(rcu_read_lock_bh_held)
  57         },
  58 };
  59
  60 enum { GP_IDLE = 0, GP_PENDING, GP_PASSED };
  61 enum { CB_IDLE = 0, CB_PENDING, CB_REPLAY };
  62
  63 #define rss_lock        gp_wait.lock
  64
  65 #ifdef CONFIG_PROVE_RCU
  66 void rcu_sync_lockdep_assert(struct rcu_sync *rsp)
  67 {
  68         RCU_LOCKDEP_WARN(!gp_ops[rsp->gp_type].held(),
  69                          "suspicious rcu_sync_is_idle() usage");
  70 }
  71 #endif
  72
  73 /**
  74  * rcu_sync_init() - Initialize an rcu_sync structure
  75  * @rsp: Pointer to rcu_sync structure to be initialized
  76  * @type: Flavor of RCU with which to synchronize rcu_sync structure
  77  */
  78 void rcu_sync_init(struct rcu_sync *rsp, enum rcu_sync_type type)
  79 {
  80         memset(rsp, 0, sizeof(*rsp));
  81         init_waitqueue_head(&rsp->gp_wait);
  82         rsp->gp_type = type;
  83 }
  84
  85 /**
  86  * rcu_sync_enter() - Force readers onto slowpath
  87  * @rsp: Pointer to rcu_sync structure to use for synchronization
  88  *
  89  * This function is used by updaters who need readers to make use of
  90  * a slowpath during the update.  After this function returns, all
  91  * subsequent calls to rcu_sync_is_idle() will return false, which
  92  * tells readers to stay off their fastpaths.  A later call to
  93  * rcu_sync_exit() re-enables reader slowpaths.
  94  *
  95  * When called in isolation, rcu_sync_enter() must wait for a grace
  96  * period, however, closely spaced calls to rcu_sync_enter() can
  97  * optimize away the grace-period wait via a state machine implemented
  98  * by rcu_sync_enter(), rcu_sync_exit(), and rcu_sync_func().
  99  */
 100 void rcu_sync_enter(struct rcu_sync *rsp)
 101 {
 102         bool need_wait, need_sync;
 103
 104         spin_lock_irq(&rsp->rss_lock);
 105         need_wait = rsp->gp_count++;
 106         need_sync = rsp->gp_state == GP_IDLE;
 107         if (need_sync)
 108                 rsp->gp_state = GP_PENDING;
 109         spin_unlock_irq(&rsp->rss_lock);
 110
 111         BUG_ON(need_wait && need_sync);
 112
 113         if (need_sync) {
 114                 gp_ops[rsp->gp_type].sync();
 115                 rsp->gp_state = GP_PASSED;
 116                 wake_up_all(&rsp->gp_wait);
 117         } else if (need_wait) {
 118                 wait_event(rsp->gp_wait, rsp->gp_state == GP_PASSED);
 119         } else {
 120                 /*
 121                  * Possible when there's a pending CB from a rcu_sync_exit().
 122                  * Nobody has yet been allowed the 'fast' path and thus we can
 123                  * avoid doing any sync(). The callback will get 'dropped'.
 124                  */
 125                 BUG_ON(rsp->gp_state != GP_PASSED);
 126         }
 127 }
 128
 129 /**
 130  * rcu_sync_func() - Callback function managing reader access to fastpath
 131  * @rsp: Pointer to rcu_sync structure to use for synchronization
 132  *
 133  * This function is passed to one of the call_rcu() functions by
 134  * rcu_sync_exit(), so that it is invoked after a grace period following the
 135  * that invocation of rcu_sync_exit().  It takes action based on events that
 136  * have taken place in the meantime, so that closely spaced rcu_sync_enter()
 137  * and rcu_sync_exit() pairs need not wait for a grace period.
 138  *
 139  * If another rcu_sync_enter() is invoked before the grace period
 140  * ended, reset state to allow the next rcu_sync_exit() to let the
 141  * readers back onto their fastpaths (after a grace period).  If both
 142  * another rcu_sync_enter() and its matching rcu_sync_exit() are invoked
 143  * before the grace period ended, re-invoke call_rcu() on behalf of that
 144  * rcu_sync_exit().  Otherwise, set all state back to idle so that readers
 145  * can again use their fastpaths.
 146  */
 147 static void rcu_sync_func(struct rcu_head *rcu)
 148 {
 149         struct rcu_sync *rsp = container_of(rcu, struct rcu_sync, cb_head);
 150         unsigned long flags;
 151
 152         BUG_ON(rsp->gp_state != GP_PASSED);
 153         BUG_ON(rsp->cb_state == CB_IDLE);
 154
 155         spin_lock_irqsave(&rsp->rss_lock, flags);
 156         if (rsp->gp_count) {
 157                 /*
 158                  * A new rcu_sync_begin() has happened; drop the callback.
 159                  */
 160                 rsp->cb_state = CB_IDLE;
 161         } else if (rsp->cb_state == CB_REPLAY) {
 162                 /*
 163                  * A new rcu_sync_exit() has happened; requeue the callback
 164                  * to catch a later GP.
 165                  */
 166                 rsp->cb_state = CB_PENDING;
 167                 gp_ops[rsp->gp_type].call(&rsp->cb_head, rcu_sync_func);
 168         } else {
 169                 /*
 170                  * We're at least a GP after rcu_sync_exit(); eveybody will now
 171                  * have observed the write side critical section. Let 'em rip!.
 172                  */
 173                 rsp->cb_state = CB_IDLE;
 174                 rsp->gp_state = GP_IDLE;
 175         }
 176         spin_unlock_irqrestore(&rsp->rss_lock, flags);
 177 }
 178
 179 /**
 180  * rcu_sync_exit() - Allow readers back onto fast patch after grace period
 181  * @rsp: Pointer to rcu_sync structure to use for synchronization
 182  *
 183  * This function is used by updaters who have completed, and can therefore
 184  * now allow readers to make use of their fastpaths after a grace period
 185  * has elapsed.  After this grace period has completed, all subsequent
 186  * calls to rcu_sync_is_idle() will return true, which tells readers that
 187  * they can once again use their fastpaths.
 188  */
 189 void rcu_sync_exit(struct rcu_sync *rsp)
 190 {
 191         spin_lock_irq(&rsp->rss_lock);
 192         if (!--rsp->gp_count) {
 193                 if (rsp->cb_state == CB_IDLE) {
 194                         rsp->cb_state = CB_PENDING;
 195                         gp_ops[rsp->gp_type].call(&rsp->cb_head, rcu_sync_func);
 196                 } else if (rsp->cb_state == CB_PENDING) {
 197                         rsp->cb_state = CB_REPLAY;
 198                 }
 199         }
 200         spin_unlock_irq(&rsp->rss_lock);
 201 }
 202
 203 /**
 204  * rcu_sync_dtor() - Clean up an rcu_sync structure
 205  * @rsp: Pointer to rcu_sync structure to be cleaned up
 206  */
 207 void rcu_sync_dtor(struct rcu_sync *rsp)
 208 {
 209         int cb_state;
 210
 211         BUG_ON(rsp->gp_count);
 212
 213         spin_lock_irq(&rsp->rss_lock);
 214         if (rsp->cb_state == CB_REPLAY)
 215                 rsp->cb_state = CB_PENDING;
 216         cb_state = rsp->cb_state;
 217         spin_unlock_irq(&rsp->rss_lock);
 218
 219         if (cb_state != CB_IDLE) {
 220                 gp_ops[rsp->gp_type].wait();
 221                 BUG_ON(rsp->cb_state != CB_IDLE);
 222         }
 223 }