Make the new mutexes faster:

- Eliminate mutexattr_private and just set a bit in ptm_owner if the mutex
  is recursive. This forces the slow path to be taken for recursive mutexes.
  Overload an unused field in pthread_mutex_t to record whether or not it's
  an errorcheck mutex.
- Streamline pthread_mutex_lock / pthread_mutex_unlock a bit more. As a
  side effect makes it possible to have assembly stubs for them.
This commit is contained in:
ad 2007-09-11 18:11:29 +00:00
parent 4042b7d22a
commit b0efccf4cd
2 changed files with 158 additions and 177 deletions

View File

@ -1,4 +1,4 @@
/* $NetBSD: pthread_int.h,v 1.53 2007/09/10 11:34:05 skrll Exp $ */
/* $NetBSD: pthread_int.h,v 1.54 2007/09/11 18:11:29 ad Exp $ */
/*-
* Copyright (c) 2001, 2002, 2003, 2006, 2007 The NetBSD Foundation, Inc.
@ -257,6 +257,7 @@ void pthread__membar_producer(void);
void pthread__membar_consumer(void);
int pthread__mutex_owned(pthread_t, pthread_mutex_t *);
int pthread__mutex_catchup(pthread_mutex_t *);
#ifndef pthread__smt_pause
#define pthread__smt_pause() /* nothing */

View File

@ -1,4 +1,4 @@
/* $NetBSD: pthread_mutex2.c,v 1.5 2007/09/11 11:30:15 ad Exp $ */
/* $NetBSD: pthread_mutex2.c,v 1.6 2007/09/11 18:11:29 ad Exp $ */
/*-
* Copyright (c) 2001, 2003, 2006, 2007 The NetBSD Foundation, Inc.
@ -37,7 +37,7 @@
*/
#include <sys/cdefs.h>
__RCSID("$NetBSD: pthread_mutex2.c,v 1.5 2007/09/11 11:30:15 ad Exp $");
__RCSID("$NetBSD: pthread_mutex2.c,v 1.6 2007/09/11 18:11:29 ad Exp $");
#include <errno.h>
#include <limits.h>
@ -55,16 +55,31 @@ __RCSID("$NetBSD: pthread_mutex2.c,v 1.5 2007/09/11 11:30:15 ad Exp $");
* the list head in order for the hint arguments to _lwp_park /
* _lwp_unpark_all to match.
*/
#define pt_nextwaiter pt_sleep.ptqe_next
#define ptm_waiters ptm_blocked.ptqh_first
#define pt_nextwaiter pt_sleep.ptqe_next
#define ptm_waiters ptm_blocked.ptqh_first
#define ptm_errorcheck ptm_blocked.ptqh_last
#define MUTEX_WAITERS_BIT (1UL)
#define MUTEX_WAITERS_BIT (0x01UL)
#define MUTEX_RECURSIVE_BIT (0x02UL)
#define MUTEX_THREAD (-16L)
#define MUTEX_HAS_WAITERS(x) ((uintptr_t)(x) & MUTEX_WAITERS_BIT)
#define MUTEX_OWNER(x) ((uintptr_t)(x) & ~MUTEX_WAITERS_BIT)
#define MUTEX_HAS_WAITERS(x) ((uintptr_t)(x) & MUTEX_WAITERS_BIT)
#define MUTEX_RECURSIVE(x) ((uintptr_t)(x) & MUTEX_RECURSIVE_BIT)
#define MUTEX_OWNER(x) ((uintptr_t)(x) & MUTEX_THREAD)
#define MUTEX_GET_RECURSE(ptm) ((intptr_t)(ptm)->ptm_private)
#define MUTEX_SET_RECURSE(ptm, delta) \
((ptm)->ptm_private = (void *)((intptr_t)(ptm)->ptm_private + delta))
#if __GNUC_PREREQ__(3, 0)
#define NOINLINE __attribute ((noinline))
#else
#define NOINLINE /* nothing */
#endif
static void pthread__mutex_wakeup(pthread_t, pthread_mutex_t *);
static int pthread__mutex_lock_slow(pthread_mutex_t *, void *);
static int pthread__mutex_lock_slow(pthread_mutex_t *);
static int pthread__mutex_unlock_slow(pthread_mutex_t *);
static void pthread__mutex_pause(void);
__strong_alias(__libc_mutex_init,pthread_mutex_init)
__strong_alias(__libc_mutex_lock,pthread_mutex_lock)
@ -85,51 +100,34 @@ mutex_cas(volatile void *ptr, void **old, void *new)
return pthread__atomic_cas_ptr(ptr, old, new);
}
struct mutex_private {
int type;
int recursecount;
};
static const struct mutex_private mutex_private_default = {
PTHREAD_MUTEX_DEFAULT,
0,
};
struct mutexattr_private {
int type;
};
static const struct mutexattr_private mutexattr_private_default = {
PTHREAD_MUTEX_DEFAULT,
};
int
pthread_mutex_init(pthread_mutex_t *ptm, const pthread_mutexattr_t *attr)
{
struct mutexattr_private *map;
struct mutex_private *mp;
intptr_t type;
pthread__error(EINVAL, "Invalid mutex attribute",
(attr == NULL) || (attr->ptma_magic == _PT_MUTEXATTR_MAGIC));
if (attr == NULL)
type = PTHREAD_MUTEX_NORMAL;
else
type = (intptr_t)attr->ptma_private;
if (attr != NULL && (map = attr->ptma_private) != NULL &&
memcmp(map, &mutexattr_private_default, sizeof(*map)) != 0) {
mp = malloc(sizeof(*mp));
if (mp == NULL)
return ENOMEM;
mp->type = map->type;
mp->recursecount = 0;
} else {
/* LINTED cast away const */
mp = (struct mutex_private *) &mutex_private_default;
switch (type) {
case PTHREAD_MUTEX_ERRORCHECK:
ptm->ptm_errorcheck = (void *)1;
ptm->ptm_owner = NULL;
break;
case PTHREAD_MUTEX_RECURSIVE:
ptm->ptm_errorcheck = NULL;
ptm->ptm_owner = (void *)MUTEX_RECURSIVE_BIT;
break;
default:
ptm->ptm_errorcheck = NULL;
ptm->ptm_owner = NULL;
break;
}
ptm->ptm_magic = _PT_MUTEX_MAGIC;
ptm->ptm_owner = NULL;
ptm->ptm_waiters = NULL;
pthread_lockinit(&ptm->ptm_lock);
ptm->ptm_private = mp;
ptm->ptm_private = NULL;
return 0;
}
@ -142,13 +140,9 @@ pthread_mutex_destroy(pthread_mutex_t *ptm)
pthread__error(EINVAL, "Invalid mutex",
ptm->ptm_magic == _PT_MUTEX_MAGIC);
pthread__error(EBUSY, "Destroying locked mutex",
ptm->ptm_owner == NULL);
MUTEX_OWNER(ptm->ptm_owner) == 0);
ptm->ptm_magic = _PT_MUTEX_DEAD;
if (ptm->ptm_private != NULL &&
ptm->ptm_private != (const void *)&mutex_private_default)
free(ptm->ptm_private);
return 0;
}
@ -173,72 +167,65 @@ pthread_mutex_lock(pthread_mutex_t *ptm)
owner = NULL;
self = pthread__self();
if (__predict_true(mutex_cas(&ptm->ptm_owner, &owner, self)))
return 0;
return pthread__mutex_lock_slow(ptm, owner);
return pthread__mutex_lock_slow(ptm);
}
#if __GNUC_PREREQ__(3, 0)
__attribute ((noinline))
#endif
static int
pthread__mutex_lock_slow(pthread_mutex_t *ptm, void *owner)
/* We want function call overhead. */
NOINLINE static void
pthread__mutex_pause(void)
{
extern int pthread__started;
struct mutex_private *mp;
void *waiters, *new;
pthread__smt_pause();
}
NOINLINE static int
pthread__mutex_lock_slow(pthread_mutex_t *ptm)
{
void *waiters, *new, *owner;
pthread_t self;
sigset_t ss;
int count;
pthread__error(EINVAL, "Invalid mutex",
ptm->ptm_magic == _PT_MUTEX_MAGIC);
/* Recursive or errorcheck? */
mp = ptm->ptm_private;
owner = ptm->ptm_owner;
self = pthread__self();
if (MUTEX_OWNER(owner) == (uintptr_t)self && mp != NULL) {
switch (mp->type) {
case PTHREAD_MUTEX_ERRORCHECK:
return EDEADLK;
case PTHREAD_MUTEX_RECURSIVE:
if (mp->recursecount == INT_MAX)
/* Recursive or errorcheck? */
if (MUTEX_OWNER(owner) == (uintptr_t)self) {
if (MUTEX_RECURSIVE(owner)) {
if (MUTEX_GET_RECURSE(ptm) == INT_MAX)
return EAGAIN;
mp->recursecount++;
MUTEX_SET_RECURSE(ptm, +1);
return 0;
}
if (ptm->ptm_errorcheck)
return EDEADLK;
}
/* Spin for a while. */
count = pthread__nspins;
while (owner != NULL && --count > 0) {
pthread__smt_pause();
while (MUTEX_OWNER(owner) == 0 && --count > 0) {
pthread__mutex_pause();
owner = ptm->ptm_owner;
}
for (;; owner = ptm->ptm_owner) {
/* If it has become free, try to acquire it again. */
while (owner == NULL) {
if (mutex_cas(&ptm->ptm_owner, &owner, self))
while (MUTEX_OWNER(owner) == 0) {
new = (void *)((uintptr_t)self | (uintptr_t)owner);
if (mutex_cas(&ptm->ptm_owner, &owner, new))
return 0;
}
/* Nope, still held. */
if (pthread__started == 0) {
/* The spec says we must deadlock, so... */
mp = ptm->ptm_private;
pthread__assert(mp->type == PTHREAD_MUTEX_NORMAL);
(void) sigprocmask(SIG_SETMASK, NULL, &ss);
for (;;) {
sigsuspend(&ss);
}
/*NOTREACHED*/
}
/*
* Add thread to the list of waiters. Issue a memory
* barrier to ensure sleeponq/nextwaiter are visible
* before we enter the waiters list.
* Nope, still held. Add thread to the list of waiters.
* Issue a memory barrier to ensure sleeponq/nextwaiter
* are visible before we enter the waiters list.
*/
self->pt_sleeponq = 1;
for (waiters = ptm->ptm_waiters;;) {
@ -268,7 +255,7 @@ pthread__mutex_lock_slow(pthread_mutex_t *ptm, void *owner)
for (owner = ptm->ptm_owner;;) {
if (MUTEX_HAS_WAITERS(owner))
break;
if (owner == NULL) {
if (MUTEX_OWNER(owner) == 0) {
pthread__mutex_wakeup(self, ptm);
break;
}
@ -293,109 +280,126 @@ pthread__mutex_lock_slow(pthread_mutex_t *ptm, void *owner)
int
pthread_mutex_trylock(pthread_mutex_t *ptm)
{
struct mutex_private *mp;
pthread_t self;
void *value;
pthread__error(EINVAL, "Invalid mutex",
ptm->ptm_magic == _PT_MUTEX_MAGIC);
self = pthread__self();
value = NULL;
if (mutex_cas(&ptm->ptm_owner, &value, self))
return 0;
/*
* These tests can be performed without holding the
* interlock because these fields are only modified
* if we know we own the mutex.
*/
mp = ptm->ptm_private;
if (mp != NULL && mp->type == PTHREAD_MUTEX_RECURSIVE &&
MUTEX_OWNER(value) == (uintptr_t)self) {
if (mp->recursecount == INT_MAX)
if (MUTEX_OWNER(value) == (uintptr_t)self && MUTEX_RECURSIVE(value)) {
if (MUTEX_GET_RECURSE(ptm) == INT_MAX)
return EAGAIN;
mp->recursecount++;
MUTEX_SET_RECURSE(ptm, +1);
return 0;
}
return EBUSY;
}
NOINLINE int
pthread__mutex_catchup(pthread_mutex_t *ptm)
{
pthread_t self;
self = pthread__self();
if (self->pt_nwaiters == 1) {
/*
* If the calling thread is about to block, defer
* unparking the target until _lwp_park() is called.
*/
if (self->pt_willpark && self->pt_unpark == 0) {
self->pt_unpark = self->pt_waiters[0];
self->pt_unparkhint = &ptm->ptm_waiters;
} else {
(void)_lwp_unpark(self->pt_waiters[0],
&ptm->ptm_waiters);
}
} else {
(void)_lwp_unpark_all(self->pt_waiters, self->pt_nwaiters,
&ptm->ptm_waiters);
}
self->pt_nwaiters = 0;
return 0;
}
int
pthread_mutex_unlock(pthread_mutex_t *ptm)
{
struct mutex_private *mp;
pthread_t self, owner;
void *owner;
pthread_t self;
self = pthread__self();
owner = self;
if (__predict_false(!mutex_cas(&ptm->ptm_owner, &owner, NULL)))
return pthread__mutex_unlock_slow(ptm);
/*
* There were no waiters, but we may have deferred waking
* other threads until mutex unlock - we must wake them now.
*/
if (self->pt_nwaiters != 0)
return pthread__mutex_catchup(ptm);
return 0;
}
NOINLINE static int
pthread__mutex_unlock_slow(pthread_mutex_t *ptm)
{
pthread_t self, owner, new;
int weown;
pthread__error(EINVAL, "Invalid mutex",
ptm->ptm_magic == _PT_MUTEX_MAGIC);
/*
* These tests can be performed without holding the
* interlock because these fields are only modified
* if we know we own the mutex.
*/
self = pthread_self();
owner = ptm->ptm_owner;
weown = (MUTEX_OWNER(owner) == (uintptr_t)self);
mp = ptm->ptm_private;
if (mp == NULL) {
if (__predict_false(!weown)) {
pthread__error(EPERM, "Unlocking unlocked mutex",
(owner != NULL));
pthread__error(EPERM,
"Unlocking mutex owned by another thread", weown);
}
} else if (mp->type == PTHREAD_MUTEX_RECURSIVE) {
if (ptm->ptm_errorcheck) {
if (!weown)
return EPERM;
if (mp->recursecount != 0) {
mp->recursecount--;
/*
* Wake any threads we deferred waking until
* mutex unlock.
*/
if (self->pt_nwaiters != 0) {
(void)_lwp_unpark_all(self->pt_waiters,
self->pt_nwaiters, &ptm->ptm_waiters);
self->pt_nwaiters = 0;
}
return 0;
}
} else if (mp->type == PTHREAD_MUTEX_ERRORCHECK) {
new = NULL;
} else if (MUTEX_RECURSIVE(owner)) {
if (!weown)
return EPERM;
if (__predict_false(!weown)) {
pthread__error(EPERM, "Unlocking unlocked mutex",
(owner != NULL));
pthread__error(EPERM,
"Unlocking mutex owned by another thread", weown);
}
if (MUTEX_GET_RECURSE(ptm) != 0) {
MUTEX_SET_RECURSE(ptm, -1);
new = owner;
} else
new = (pthread_t)MUTEX_RECURSIVE_BIT;
} else {
pthread__error(EPERM,
"Unlocking unlocked mutex", (owner != NULL));
pthread__error(EPERM,
"Unlocking mutex owned by another thread", weown);
new = NULL;
}
/*
* Release the mutex. If there appear to be waiters, then
* wake them up.
*/
owner = pthread__atomic_swap_ptr(&ptm->ptm_owner, NULL);
if (MUTEX_HAS_WAITERS(owner) != 0) {
pthread__mutex_wakeup(self, ptm);
return 0;
if (new != owner) {
owner = pthread__atomic_swap_ptr(&ptm->ptm_owner, new);
if (MUTEX_HAS_WAITERS(owner) != 0) {
pthread__mutex_wakeup(self, ptm);
return 0;
}
}
/*
* There were no waiters, but we may have deferred waking
* other threads until mutex unlock - we must wake them now.
*/
if (self->pt_nwaiters != 0) {
(void)_lwp_unpark_all(self->pt_waiters, self->pt_nwaiters,
&ptm->ptm_waiters);
self->pt_nwaiters = 0;
}
if (self->pt_nwaiters != 0)
return pthread__mutex_catchup(ptm);
return 0;
}
@ -462,25 +466,15 @@ pthread__mutex_wakeup(pthread_t self, pthread_mutex_t *ptm)
}
}
}
int
pthread_mutexattr_init(pthread_mutexattr_t *attr)
{
struct mutexattr_private *map;
map = malloc(sizeof(*map));
if (map == NULL)
return ENOMEM;
*map = mutexattr_private_default;
attr->ptma_magic = _PT_MUTEXATTR_MAGIC;
attr->ptma_private = map;
attr->ptma_private = (void *)PTHREAD_MUTEX_DEFAULT;
return 0;
}
int
pthread_mutexattr_destroy(pthread_mutexattr_t *attr)
{
@ -488,10 +482,6 @@ pthread_mutexattr_destroy(pthread_mutexattr_t *attr)
pthread__error(EINVAL, "Invalid mutex attribute",
attr->ptma_magic == _PT_MUTEXATTR_MAGIC);
attr->ptma_magic = _PT_MUTEXATTR_DEAD;
if (attr->ptma_private != NULL)
free(attr->ptma_private);
return 0;
}
@ -499,15 +489,11 @@ pthread_mutexattr_destroy(pthread_mutexattr_t *attr)
int
pthread_mutexattr_gettype(const pthread_mutexattr_t *attr, int *typep)
{
struct mutexattr_private *map;
pthread__error(EINVAL, "Invalid mutex attribute",
attr->ptma_magic == _PT_MUTEXATTR_MAGIC);
map = attr->ptma_private;
*typep = map->type;
*typep = (int)(intptr_t)attr->ptma_private;
return 0;
}
@ -515,25 +501,19 @@ pthread_mutexattr_gettype(const pthread_mutexattr_t *attr, int *typep)
int
pthread_mutexattr_settype(pthread_mutexattr_t *attr, int type)
{
struct mutexattr_private *map;
pthread__error(EINVAL, "Invalid mutex attribute",
attr->ptma_magic == _PT_MUTEXATTR_MAGIC);
map = attr->ptma_private;
switch (type) {
case PTHREAD_MUTEX_NORMAL:
case PTHREAD_MUTEX_ERRORCHECK:
case PTHREAD_MUTEX_RECURSIVE:
map->type = type;
break;
attr->ptma_private = (void *)(intptr_t)type;
return 0;
default:
return EINVAL;
}
return 0;
}