| 1 |
737 |
jeremybenn |
/* Copyright (C) 2009, 2011 Free Software Foundation, Inc.
|
| 2 |
|
|
Contributed by Richard Henderson <rth@redhat.com>.
|
| 3 |
|
|
|
| 4 |
|
|
This file is part of the GNU Transactional Memory Library (libitm).
|
| 5 |
|
|
|
| 6 |
|
|
Libitm is free software; you can redistribute it and/or modify it
|
| 7 |
|
|
under the terms of the GNU General Public License as published by
|
| 8 |
|
|
the Free Software Foundation; either version 3 of the License, or
|
| 9 |
|
|
(at your option) any later version.
|
| 10 |
|
|
|
| 11 |
|
|
Libitm is distributed in the hope that it will be useful, but WITHOUT ANY
|
| 12 |
|
|
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
|
| 13 |
|
|
FOR A PARTICULAR PURPOSE. See the GNU General Public License for
|
| 14 |
|
|
more details.
|
| 15 |
|
|
|
| 16 |
|
|
Under Section 7 of GPL version 3, you are granted additional
|
| 17 |
|
|
permissions described in the GCC Runtime Library Exception, version
|
| 18 |
|
|
3.1, as published by the Free Software Foundation.
|
| 19 |
|
|
|
| 20 |
|
|
You should have received a copy of the GNU General Public License and
|
| 21 |
|
|
a copy of the GCC Runtime Library Exception along with this program;
|
| 22 |
|
|
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
| 23 |
|
|
<http://www.gnu.org/licenses/>. */
|
| 24 |
|
|
|
| 25 |
|
|
#ifndef LIBITM_STMLOCK_H
|
| 26 |
|
|
#define LIBITM_STMLOCK_H 1
|
| 27 |
|
|
|
| 28 |
|
|
namespace GTM HIDDEN {
|
| 29 |
|
|
|
| 30 |
|
|
/* A versioned write lock on a cacheline. This must be wide enough to
|
| 31 |
|
|
store a pointer, and preferably wide enough to avoid overflowing the
|
| 32 |
|
|
version counter. Thus we use a "word", which should be 64-bits on
|
| 33 |
|
|
64-bit systems even when their pointer size is forced smaller. */
|
| 34 |
|
|
typedef gtm_word gtm_stmlock;
|
| 35 |
|
|
|
| 36 |
|
|
/* This has to be the same size as gtm_stmlock, we just use this name
|
| 37 |
|
|
for documentation purposes. */
|
| 38 |
|
|
typedef gtm_word gtm_version;
|
| 39 |
|
|
|
| 40 |
|
|
/* The maximum value a version number can have. This is a consequence
|
| 41 |
|
|
of having the low bit of gtm_stmlock reserved for the owned bit. */
|
| 42 |
|
|
#define GTM_VERSION_MAX (~(gtm_version)0 >> 1)
|
| 43 |
|
|
|
| 44 |
|
|
/* A value that may be used to indicate "uninitialized" for a version. */
|
| 45 |
|
|
#define GTM_VERSION_INVALID (~(gtm_version)0)
|
| 46 |
|
|
|
| 47 |
|
|
/* This bit is set when the write lock is held. When set, the balance of
|
| 48 |
|
|
the bits in the lock is a pointer that references STM backend specific
|
| 49 |
|
|
data; it is up to the STM backend to determine if this thread holds the
|
| 50 |
|
|
lock. If this bit is clear, the balance of the bits are the last
|
| 51 |
|
|
version number committed to the cacheline. */
|
| 52 |
|
|
static inline bool
|
| 53 |
|
|
gtm_stmlock_owned_p (gtm_stmlock lock)
|
| 54 |
|
|
{
|
| 55 |
|
|
return lock & 1;
|
| 56 |
|
|
}
|
| 57 |
|
|
|
| 58 |
|
|
static inline gtm_stmlock
|
| 59 |
|
|
gtm_stmlock_set_owned (void *data)
|
| 60 |
|
|
{
|
| 61 |
|
|
return (gtm_stmlock)(uintptr_t)data | 1;
|
| 62 |
|
|
}
|
| 63 |
|
|
|
| 64 |
|
|
static inline void *
|
| 65 |
|
|
gtm_stmlock_get_addr (gtm_stmlock lock)
|
| 66 |
|
|
{
|
| 67 |
|
|
return (void *)((uintptr_t)lock & ~(uintptr_t)1);
|
| 68 |
|
|
}
|
| 69 |
|
|
|
| 70 |
|
|
static inline gtm_version
|
| 71 |
|
|
gtm_stmlock_get_version (gtm_stmlock lock)
|
| 72 |
|
|
{
|
| 73 |
|
|
return lock >> 1;
|
| 74 |
|
|
}
|
| 75 |
|
|
|
| 76 |
|
|
static inline gtm_stmlock
|
| 77 |
|
|
gtm_stmlock_set_version (gtm_version ver)
|
| 78 |
|
|
{
|
| 79 |
|
|
return ver << 1;
|
| 80 |
|
|
}
|
| 81 |
|
|
|
| 82 |
|
|
/* We use a fixed set of locks for all memory, hashed into the
|
| 83 |
|
|
following table. */
|
| 84 |
|
|
#define LOCK_ARRAY_SIZE (1024 * 1024)
|
| 85 |
|
|
extern gtm_stmlock gtm_stmlock_array[LOCK_ARRAY_SIZE];
|
| 86 |
|
|
|
| 87 |
|
|
static inline gtm_stmlock *
|
| 88 |
|
|
gtm_get_stmlock (const gtm_cacheline *addr)
|
| 89 |
|
|
{
|
| 90 |
|
|
size_t idx = ((uintptr_t) addr / CACHELINE_SIZE) % LOCK_ARRAY_SIZE;
|
| 91 |
|
|
return gtm_stmlock_array + idx;
|
| 92 |
|
|
}
|
| 93 |
|
|
|
| 94 |
|
|
/* The current global version number. */
|
| 95 |
|
|
extern atomic<gtm_version> gtm_clock;
|
| 96 |
|
|
|
| 97 |
|
|
static inline gtm_version
|
| 98 |
|
|
gtm_get_clock (void)
|
| 99 |
|
|
{
|
| 100 |
|
|
atomic_thread_fence(memory_order_release);
|
| 101 |
|
|
return gtm_clock.load(memory_order_acquire);
|
| 102 |
|
|
}
|
| 103 |
|
|
|
| 104 |
|
|
static inline gtm_version
|
| 105 |
|
|
gtm_inc_clock (void)
|
| 106 |
|
|
{
|
| 107 |
|
|
/* ??? Here we have a choice, the pre-inc operator mapping to
|
| 108 |
|
|
__atomic_add_fetch with memory_order_seq_cst, or fetch_add
|
| 109 |
|
|
with memory_order_acq_rel plus another separate increment.
|
| 110 |
|
|
We really ought to recognize and optimize fetch_op(x) op x... */
|
| 111 |
|
|
gtm_version r = ++gtm_clock;
|
| 112 |
|
|
|
| 113 |
|
|
/* ??? Ought to handle wraparound for 32-bit. */
|
| 114 |
|
|
if (sizeof(r) < 8 && r > GTM_VERSION_MAX)
|
| 115 |
|
|
abort ();
|
| 116 |
|
|
|
| 117 |
|
|
return r;
|
| 118 |
|
|
}
|
| 119 |
|
|
|
| 120 |
|
|
} // namespace GTM
|
| 121 |
|
|
|
| 122 |
|
|
#endif // LIBITM_STMLOCK_H
|