1/*
2 * Copyright 2010-2012 PathScale, Inc. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions are met:
6 *
7 * 1. Redistributions of source code must retain the above copyright notice,
8 *    this list of conditions and the following disclaimer.
9 *
10 * 2. Redistributions in binary form must reproduce the above copyright notice,
11 *    this list of conditions and the following disclaimer in the documentation
12 *    and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS ``AS
15 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
16 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
17 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR
18 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
19 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
20 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
21 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
22 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
23 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
24 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25 */
26
27/**
28 * guard.cc: Functions for thread-safe static initialisation.
29 *
30 * Static values in C++ can be initialised lazily their first use.  This file
31 * contains functions that are used to ensure that two threads attempting to
32 * initialize the same static do not call the constructor twice.  This is
33 * important because constructors can have side effects, so calling the
34 * constructor twice may be very bad.
35 *
36 * Statics that require initialisation are protected by a 64-bit value.  Any
37 * platform that can do 32-bit atomic test and set operations can use this
38 * value as a low-overhead lock.  Because statics (in most sane code) are
39 * accessed far more times than they are initialised, this lock implementation
40 * is heavily optimised towards the case where the static has already been
41 * initialised.
42 */
43#include <stdint.h>
44#include <stdlib.h>
45#include <stdio.h>
46#include <pthread.h>
47#include <assert.h>
48#include "atomic.h"
49
50// Older GCC doesn't define __LITTLE_ENDIAN__
51#ifndef __LITTLE_ENDIAN__
52	// If __BYTE_ORDER__ is defined, use that instead
53#	ifdef __BYTE_ORDER__
54#		if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
55#			define __LITTLE_ENDIAN__
56#		endif
57	// x86 and ARM are the most common little-endian CPUs, so let's have a
58	// special case for them (ARM is already special cased).  Assume everything
59	// else is big endian.
60#	elif defined(__x86_64) || defined(__i386)
61#		define __LITTLE_ENDIAN__
62#	endif
63#endif
64
65
66/*
67 * The least significant bit of the guard variable indicates that the object
68 * has been initialised, the most significant bit is used for a spinlock.
69 */
70#ifdef __arm__
71// ARM ABI - 32-bit guards.
72typedef uint32_t guard_t;
73static const uint32_t LOCKED = ((guard_t)1) << 31;
74static const uint32_t INITIALISED = 1;
75#else
76typedef uint64_t guard_t;
77#	if defined(__LITTLE_ENDIAN__)
78static const guard_t LOCKED = ((guard_t)1) << 63;
79static const guard_t INITIALISED = 1;
80#	else
81static const guard_t LOCKED = 1;
82static const guard_t INITIALISED = ((guard_t)1) << 56;
83#	endif
84#endif
85
86/**
87 * Acquires a lock on a guard, returning 0 if the object has already been
88 * initialised, and 1 if it has not.  If the object is already constructed then
89 * this function just needs to read a byte from memory and return.
90 */
91extern "C" int __cxa_guard_acquire(volatile guard_t *guard_object)
92{
93	// Not an atomic read, doesn't establish a happens-before relationship, but
94	// if one is already established and we end up seeing an initialised state
95	// then it's a fast path, otherwise we'll do something more expensive than
96	// this test anyway...
97	if ((INITIALISED == *guard_object)) { return 0; }
98	// Spin trying to do the initialisation
99	while (1)
100	{
101		// Loop trying to move the value of the guard from 0 (not
102		// locked, not initialised) to the locked-uninitialised
103		// position.
104		switch (__sync_val_compare_and_swap(guard_object, 0, LOCKED))
105		{
106			// If the old value was 0, we succeeded, so continue
107			// initialising
108			case 0:
109				return 1;
110			// If this was already initialised, return and let the caller skip
111			// initialising it again.
112			case INITIALISED:
113				return 0;
114			// If it is locked by another thread, relinquish the CPU and try
115			// again later.
116			case LOCKED:
117			case LOCKED | INITIALISED:
118				sched_yield();
119				break;
120			// If it is some other value, then something has gone badly wrong.
121			// Give up.
122			default:
123				fprintf(stderr, "Invalid state detected attempting to lock static initialiser.\n");
124				abort();
125		}
126	}
127	//__builtin_unreachable();
128	return 0;
129}
130
131/**
132 * Releases the lock without marking the object as initialised.  This function
133 * is called if initialising a static causes an exception to be thrown.
134 */
135extern "C" void __cxa_guard_abort(volatile guard_t *guard_object)
136{
137	__attribute__((unused))
138	bool reset = __sync_bool_compare_and_swap(guard_object, LOCKED, 0);
139	assert(reset);
140}
141/**
142 * Releases the guard and marks the object as initialised.  This function is
143 * called after successful initialisation of a static.
144 */
145extern "C" void __cxa_guard_release(volatile guard_t *guard_object)
146{
147	__attribute__((unused))
148	bool reset = __sync_bool_compare_and_swap(guard_object, LOCKED, INITIALISED);
149	assert(reset);
150}
151
152
153