Line data    Source code

       1             : // SPDX-License-Identifier: GPL-2.0+
       2             : /*
       3             :  * RCU-based infrastructure for lightweight reader-writer locking
       4             :  *
       5             :  * Copyright (c) 2015, Red Hat, Inc.
       6             :  *
       7             :  * Author: Oleg Nesterov <oleg@redhat.com>
       8             :  */
       9             : 
      10             : #include <linux/rcu_sync.h>
      11             : #include <linux/sched.h>
      12             : 
      13             : enum { GP_IDLE = 0, GP_ENTER, GP_PASSED, GP_EXIT, GP_REPLAY };
      14             : 
      15             : #define rss_lock        gp_wait.lock
      16             : 
      17             : /**
      18             :  * rcu_sync_init() - Initialize an rcu_sync structure
      19             :  * @rsp: Pointer to rcu_sync structure to be initialized
      20             :  */
      21          45 : void rcu_sync_init(struct rcu_sync *rsp)
      22             : {
      23          90 :         memset(rsp, 0, sizeof(*rsp));
      24          45 :         init_waitqueue_head(&rsp->gp_wait);
      25          45 : }
      26             : 
      27             : /**
      28             :  * rcu_sync_enter_start - Force readers onto slow path for multiple updates
      29             :  * @rsp: Pointer to rcu_sync structure to use for synchronization
      30             :  *
      31             :  * Must be called after rcu_sync_init() and before first use.
      32             :  *
      33             :  * Ensures rcu_sync_is_idle() returns false and rcu_sync_{enter,exit}()
      34             :  * pairs turn into NO-OPs.
      35             :  */
      36           0 : void rcu_sync_enter_start(struct rcu_sync *rsp)
      37             : {
      38           0 :         rsp->gp_count++;
      39           0 :         rsp->gp_state = GP_PASSED;
      40           0 : }
      41             : 
      42             : 
      43             : static void rcu_sync_func(struct rcu_head *rhp);
      44             : 
      45             : static void rcu_sync_call(struct rcu_sync *rsp)
      46             : {
      47           0 :         call_rcu_hurry(&rsp->cb_head, rcu_sync_func);
      48             : }
      49             : 
      50             : /**
      51             :  * rcu_sync_func() - Callback function managing reader access to fastpath
      52             :  * @rhp: Pointer to rcu_head in rcu_sync structure to use for synchronization
      53             :  *
      54             :  * This function is passed to call_rcu() function by rcu_sync_enter() and
      55             :  * rcu_sync_exit(), so that it is invoked after a grace period following the
      56             :  * that invocation of enter/exit.
      57             :  *
      58             :  * If it is called by rcu_sync_enter() it signals that all the readers were
      59             :  * switched onto slow path.
      60             :  *
      61             :  * If it is called by rcu_sync_exit() it takes action based on events that
      62             :  * have taken place in the meantime, so that closely spaced rcu_sync_enter()
      63             :  * and rcu_sync_exit() pairs need not wait for a grace period.
      64             :  *
      65             :  * If another rcu_sync_enter() is invoked before the grace period
      66             :  * ended, reset state to allow the next rcu_sync_exit() to let the
      67             :  * readers back onto their fastpaths (after a grace period).  If both
      68             :  * another rcu_sync_enter() and its matching rcu_sync_exit() are invoked
      69             :  * before the grace period ended, re-invoke call_rcu() on behalf of that
      70             :  * rcu_sync_exit().  Otherwise, set all state back to idle so that readers
      71             :  * can again use their fastpaths.
      72             :  */
      73           0 : static void rcu_sync_func(struct rcu_head *rhp)
      74             : {
      75           0 :         struct rcu_sync *rsp = container_of(rhp, struct rcu_sync, cb_head);
      76             :         unsigned long flags;
      77             : 
      78           0 :         WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_IDLE);
      79           0 :         WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_PASSED);
      80             : 
      81           0 :         spin_lock_irqsave(&rsp->rss_lock, flags);
      82           0 :         if (rsp->gp_count) {
      83             :                 /*
      84             :                  * We're at least a GP after the GP_IDLE->GP_ENTER transition.
      85             :                  */
      86           0 :                 WRITE_ONCE(rsp->gp_state, GP_PASSED);
      87           0 :                 wake_up_locked(&rsp->gp_wait);
      88           0 :         } else if (rsp->gp_state == GP_REPLAY) {
      89             :                 /*
      90             :                  * A new rcu_sync_exit() has happened; requeue the callback to
      91             :                  * catch a later GP.
      92             :                  */
      93           0 :                 WRITE_ONCE(rsp->gp_state, GP_EXIT);
      94             :                 rcu_sync_call(rsp);
      95             :         } else {
      96             :                 /*
      97             :                  * We're at least a GP after the last rcu_sync_exit(); everybody
      98             :                  * will now have observed the write side critical section.
      99             :                  * Let 'em rip!
     100             :                  */
     101           0 :                 WRITE_ONCE(rsp->gp_state, GP_IDLE);
     102             :         }
     103           0 :         spin_unlock_irqrestore(&rsp->rss_lock, flags);
     104           0 : }
     105             : 
     106             : /**
     107             :  * rcu_sync_enter() - Force readers onto slowpath
     108             :  * @rsp: Pointer to rcu_sync structure to use for synchronization
     109             :  *
     110             :  * This function is used by updaters who need readers to make use of
     111             :  * a slowpath during the update.  After this function returns, all
     112             :  * subsequent calls to rcu_sync_is_idle() will return false, which
     113             :  * tells readers to stay off their fastpaths.  A later call to
     114             :  * rcu_sync_exit() re-enables reader fastpaths.
     115             :  *
     116             :  * When called in isolation, rcu_sync_enter() must wait for a grace
     117             :  * period, however, closely spaced calls to rcu_sync_enter() can
     118             :  * optimize away the grace-period wait via a state machine implemented
     119             :  * by rcu_sync_enter(), rcu_sync_exit(), and rcu_sync_func().
     120             :  */
     121           0 : void rcu_sync_enter(struct rcu_sync *rsp)
     122             : {
     123             :         int gp_state;
     124             : 
     125           0 :         spin_lock_irq(&rsp->rss_lock);
     126           0 :         gp_state = rsp->gp_state;
     127           0 :         if (gp_state == GP_IDLE) {
     128           0 :                 WRITE_ONCE(rsp->gp_state, GP_ENTER);
     129           0 :                 WARN_ON_ONCE(rsp->gp_count);
     130             :                 /*
     131             :                  * Note that we could simply do rcu_sync_call(rsp) here and
     132             :                  * avoid the "if (gp_state == GP_IDLE)" block below.
     133             :                  *
     134             :                  * However, synchronize_rcu() can be faster if rcu_expedited
     135             :                  * or rcu_blocking_is_gp() is true.
     136             :                  *
     137             :                  * Another reason is that we can't wait for rcu callback if
     138             :                  * we are called at early boot time but this shouldn't happen.
     139             :                  */
     140             :         }
     141           0 :         rsp->gp_count++;
     142           0 :         spin_unlock_irq(&rsp->rss_lock);
     143             : 
     144           0 :         if (gp_state == GP_IDLE) {
     145             :                 /*
     146             :                  * See the comment above, this simply does the "synchronous"
     147             :                  * call_rcu(rcu_sync_func) which does GP_ENTER -> GP_PASSED.
     148             :                  */
     149           0 :                 synchronize_rcu();
     150           0 :                 rcu_sync_func(&rsp->cb_head);
     151             :                 /* Not really needed, wait_event() would see GP_PASSED. */
     152           0 :                 return;
     153             :         }
     154             : 
     155           0 :         wait_event(rsp->gp_wait, READ_ONCE(rsp->gp_state) >= GP_PASSED);
     156             : }
     157             : 
     158             : /**
     159             :  * rcu_sync_exit() - Allow readers back onto fast path after grace period
     160             :  * @rsp: Pointer to rcu_sync structure to use for synchronization
     161             :  *
     162             :  * This function is used by updaters who have completed, and can therefore
     163             :  * now allow readers to make use of their fastpaths after a grace period
     164             :  * has elapsed.  After this grace period has completed, all subsequent
     165             :  * calls to rcu_sync_is_idle() will return true, which tells readers that
     166             :  * they can once again use their fastpaths.
     167             :  */
     168           0 : void rcu_sync_exit(struct rcu_sync *rsp)
     169             : {
     170           0 :         WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_IDLE);
     171           0 :         WARN_ON_ONCE(READ_ONCE(rsp->gp_count) == 0);
     172             : 
     173           0 :         spin_lock_irq(&rsp->rss_lock);
     174           0 :         if (!--rsp->gp_count) {
     175           0 :                 if (rsp->gp_state == GP_PASSED) {
     176           0 :                         WRITE_ONCE(rsp->gp_state, GP_EXIT);
     177             :                         rcu_sync_call(rsp);
     178           0 :                 } else if (rsp->gp_state == GP_EXIT) {
     179           0 :                         WRITE_ONCE(rsp->gp_state, GP_REPLAY);
     180             :                 }
     181             :         }
     182           0 :         spin_unlock_irq(&rsp->rss_lock);
     183           0 : }
     184             : 
     185             : /**
     186             :  * rcu_sync_dtor() - Clean up an rcu_sync structure
     187             :  * @rsp: Pointer to rcu_sync structure to be cleaned up
     188             :  */
     189          15 : void rcu_sync_dtor(struct rcu_sync *rsp)
     190             : {
     191             :         int gp_state;
     192             : 
     193          15 :         WARN_ON_ONCE(READ_ONCE(rsp->gp_count));
     194          15 :         WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_PASSED);
     195             : 
     196          30 :         spin_lock_irq(&rsp->rss_lock);
     197          15 :         if (rsp->gp_state == GP_REPLAY)
     198           0 :                 WRITE_ONCE(rsp->gp_state, GP_EXIT);
     199          15 :         gp_state = rsp->gp_state;
     200          30 :         spin_unlock_irq(&rsp->rss_lock);
     201             : 
     202          15 :         if (gp_state != GP_IDLE) {
     203           0 :                 rcu_barrier();
     204           0 :                 WARN_ON_ONCE(rsp->gp_state != GP_IDLE);
     205             :         }
     206          15 : }