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
 32static 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
 60enum { GP_IDLE = 0, GP_PENDING, GP_PASSED };
 61enum { CB_IDLE = 0, CB_PENDING, CB_REPLAY };
 62
 63#define	rss_lock	gp_wait.lock
 64
 65#ifdef CONFIG_PROVE_RCU
 66void 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
 72EXPORT_SYMBOL_GPL(rcu_sync_lockdep_assert);
 73#endif
 74
 75/**
 76 * rcu_sync_init() - Initialize an rcu_sync structure
 77 * @rsp: Pointer to rcu_sync structure to be initialized
 78 * @type: Flavor of RCU with which to synchronize rcu_sync structure
 79 */
 80void rcu_sync_init(struct rcu_sync *rsp, enum rcu_sync_type type)
 81{
 82	memset(rsp, 0, sizeof(*rsp));
 83	init_waitqueue_head(&rsp->gp_wait);
 84	rsp->gp_type = type;
 85}
 86
 87/**
 88 * Must be called after rcu_sync_init() and before first use.
 89 *
 90 * Ensures rcu_sync_is_idle() returns false and rcu_sync_{enter,exit}()
 91 * pairs turn into NO-OPs.
 92 */
 93void rcu_sync_enter_start(struct rcu_sync *rsp)
 94{
 95	rsp->gp_count++;
 96	rsp->gp_state = GP_PASSED;
 97}
 98
 99/**
100 * rcu_sync_enter() - Force readers onto slowpath
101 * @rsp: Pointer to rcu_sync structure to use for synchronization
102 *
103 * This function is used by updaters who need readers to make use of
104 * a slowpath during the update.  After this function returns, all
105 * subsequent calls to rcu_sync_is_idle() will return false, which
106 * tells readers to stay off their fastpaths.  A later call to
107 * rcu_sync_exit() re-enables reader slowpaths.
108 *
109 * When called in isolation, rcu_sync_enter() must wait for a grace
110 * period, however, closely spaced calls to rcu_sync_enter() can
111 * optimize away the grace-period wait via a state machine implemented
112 * by rcu_sync_enter(), rcu_sync_exit(), and rcu_sync_func().
113 */
114void rcu_sync_enter(struct rcu_sync *rsp)
115{
116	bool need_wait, need_sync;
117
118	spin_lock_irq(&rsp->rss_lock);
119	need_wait = rsp->gp_count++;
120	need_sync = rsp->gp_state == GP_IDLE;
121	if (need_sync)
122		rsp->gp_state = GP_PENDING;
123	spin_unlock_irq(&rsp->rss_lock);
124
125	BUG_ON(need_wait && need_sync);
126
127	if (need_sync) {
128		gp_ops[rsp->gp_type].sync();
129		rsp->gp_state = GP_PASSED;
130		wake_up_all(&rsp->gp_wait);
131	} else if (need_wait) {
132		wait_event(rsp->gp_wait, rsp->gp_state == GP_PASSED);
133	} else {
134		/*
135		 * Possible when there's a pending CB from a rcu_sync_exit().
136		 * Nobody has yet been allowed the 'fast' path and thus we can
137		 * avoid doing any sync(). The callback will get 'dropped'.
138		 */
139		BUG_ON(rsp->gp_state != GP_PASSED);
140	}
141}
142
143/**
144 * rcu_sync_func() - Callback function managing reader access to fastpath
145 * @rsp: Pointer to rcu_sync structure to use for synchronization
146 *
147 * This function is passed to one of the call_rcu() functions by
148 * rcu_sync_exit(), so that it is invoked after a grace period following the
149 * that invocation of rcu_sync_exit().  It takes action based on events that
150 * have taken place in the meantime, so that closely spaced rcu_sync_enter()
151 * and rcu_sync_exit() pairs need not wait for a grace period.
152 *
153 * If another rcu_sync_enter() is invoked before the grace period
154 * ended, reset state to allow the next rcu_sync_exit() to let the
155 * readers back onto their fastpaths (after a grace period).  If both
156 * another rcu_sync_enter() and its matching rcu_sync_exit() are invoked
157 * before the grace period ended, re-invoke call_rcu() on behalf of that
158 * rcu_sync_exit().  Otherwise, set all state back to idle so that readers
159 * can again use their fastpaths.
160 */
161static void rcu_sync_func(struct rcu_head *rcu)
162{
163	struct rcu_sync *rsp = container_of(rcu, struct rcu_sync, cb_head);
164	unsigned long flags;
165
166	BUG_ON(rsp->gp_state != GP_PASSED);
167	BUG_ON(rsp->cb_state == CB_IDLE);
168
169	spin_lock_irqsave(&rsp->rss_lock, flags);
170	if (rsp->gp_count) {
171		/*
172		 * A new rcu_sync_begin() has happened; drop the callback.
173		 */
174		rsp->cb_state = CB_IDLE;
175	} else if (rsp->cb_state == CB_REPLAY) {
176		/*
177		 * A new rcu_sync_exit() has happened; requeue the callback
178		 * to catch a later GP.
179		 */
180		rsp->cb_state = CB_PENDING;
181		gp_ops[rsp->gp_type].call(&rsp->cb_head, rcu_sync_func);
182	} else {
183		/*
184		 * We're at least a GP after rcu_sync_exit(); eveybody will now
185		 * have observed the write side critical section. Let 'em rip!.
186		 */
187		rsp->cb_state = CB_IDLE;
188		rsp->gp_state = GP_IDLE;
189	}
190	spin_unlock_irqrestore(&rsp->rss_lock, flags);
191}
192
193/**
194 * rcu_sync_exit() - Allow readers back onto fast patch after grace period
195 * @rsp: Pointer to rcu_sync structure to use for synchronization
196 *
197 * This function is used by updaters who have completed, and can therefore
198 * now allow readers to make use of their fastpaths after a grace period
199 * has elapsed.  After this grace period has completed, all subsequent
200 * calls to rcu_sync_is_idle() will return true, which tells readers that
201 * they can once again use their fastpaths.
202 */
203void rcu_sync_exit(struct rcu_sync *rsp)
204{
205	spin_lock_irq(&rsp->rss_lock);
206	if (!--rsp->gp_count) {
207		if (rsp->cb_state == CB_IDLE) {
208			rsp->cb_state = CB_PENDING;
209			gp_ops[rsp->gp_type].call(&rsp->cb_head, rcu_sync_func);
210		} else if (rsp->cb_state == CB_PENDING) {
211			rsp->cb_state = CB_REPLAY;
212		}
213	}
214	spin_unlock_irq(&rsp->rss_lock);
215}
216
217/**
218 * rcu_sync_dtor() - Clean up an rcu_sync structure
219 * @rsp: Pointer to rcu_sync structure to be cleaned up
220 */
221void rcu_sync_dtor(struct rcu_sync *rsp)
222{
223	int cb_state;
224
225	BUG_ON(rsp->gp_count);
226
227	spin_lock_irq(&rsp->rss_lock);
228	if (rsp->cb_state == CB_REPLAY)
229		rsp->cb_state = CB_PENDING;
230	cb_state = rsp->cb_state;
231	spin_unlock_irq(&rsp->rss_lock);
232
233	if (cb_state != CB_IDLE) {
234		gp_ops[rsp->gp_type].wait();
235		BUG_ON(rsp->cb_state != CB_IDLE);
236	}
237}