Reimplement futexes.

Old implemention used Giant to protect the kernel data structures,
but at the same time called malloc(M_WAITOK), that could cause the
calling thread to sleep and lost Giant protection. User-visible
result was the missed wakeup.

New implementation uses one sx lock per futex. The sx protects
the futex structures and allows to sleep while copyin or copyout
are performed.

Unlike linux, we return EINVAL when FUTEX_CMP_REQUEUE operation
is requested and either caller specified futexes are equial or
second futex already exists. This is acceptable since the situation
can only occur from the application error, and glibc falls back to
old FUTEX_WAKE operation when FUTEX_CMP_REQUEUE returns an error.

Approved by:	kib (mentor)
MFC after:	1 month
This commit is contained in:
Dmitry Chagin 2009-05-01 15:36:02 +00:00
parent a3d3b633a9
commit 79262bf1f0
3 changed files with 478 additions and 387 deletions

View File

@ -128,7 +128,7 @@ static void exec_linux_setregs(struct thread *td, u_long entry,
static void linux32_fixlimit(struct rlimit *rl, int which);
extern LIST_HEAD(futex_list, futex) futex_list;
extern struct sx futex_sx;
extern struct mtx futex_mtx;
static eventhandler_tag linux_exit_tag;
static eventhandler_tag linux_schedtail_tag;
@ -1117,7 +1117,7 @@ linux_elf_modevent(module_t mod, int type, void *data)
mtx_init(&emul_lock, "emuldata lock", NULL, MTX_DEF);
sx_init(&emul_shared_lock, "emuldata->shared lock");
LIST_INIT(&futex_list);
sx_init(&futex_sx, "futex protection lock");
mtx_init(&futex_mtx, "ftllk", NULL, MTX_DEF);
linux_exit_tag = EVENTHANDLER_REGISTER(process_exit,
linux_proc_exit, NULL, 1000);
linux_schedtail_tag = EVENTHANDLER_REGISTER(schedtail,
@ -1149,7 +1149,7 @@ linux_elf_modevent(module_t mod, int type, void *data)
linux_device_unregister_handler(*ldhp);
mtx_destroy(&emul_lock);
sx_destroy(&emul_shared_lock);
sx_destroy(&futex_sx);
mtx_destroy(&futex_mtx);
EVENTHANDLER_DEREGISTER(process_exit, linux_exit_tag);
EVENTHANDLER_DEREGISTER(schedtail, linux_schedtail_tag);
EVENTHANDLER_DEREGISTER(process_exec, linux_exec_tag);

View File

@ -62,419 +62,284 @@ __KERNEL_RCSID(1, "$NetBSD: linux_futex.c,v 1.7 2006/07/24 19:01:49 manu Exp $")
#include <compat/linux/linux_futex.h>
#include <compat/linux/linux_emul.h>
MALLOC_DEFINE(M_FUTEX, "futex", "Linux futexes");
MALLOC_DEFINE(M_FUTEX_WP, "futex wp", "Linux futexes wp");
struct futex;
struct waiting_proc {
struct thread *wp_t;
struct futex *wp_new_futex;
uint32_t wp_flags;
struct futex *wp_futex;
TAILQ_ENTRY(waiting_proc) wp_list;
};
struct futex {
void *f_uaddr;
int f_refcount;
struct sx f_lck;
uint32_t *f_uaddr;
uint32_t f_refcount;
LIST_ENTRY(futex) f_list;
TAILQ_HEAD(lf_waiting_proc, waiting_proc) f_waiting_proc;
};
LIST_HEAD(futex_list, futex) futex_list;
struct sx futex_sx; /* this protects the LIST of futexes */
#define FUTEX_LOCK sx_xlock(&futex_sx)
#define FUTEX_UNLOCK sx_xunlock(&futex_sx)
#define FUTEX_LOCK(f) sx_xlock(&(f)->f_lck)
#define FUTEX_UNLOCK(f) sx_xunlock(&(f)->f_lck)
#define FUTEX_INIT(f) sx_init_flags(&(f)->f_lck, "ftlk", 0)
#define FUTEX_DESTROY(f) sx_destroy(&(f)->f_lck)
#define FUTEX_ASSERT_LOCKED(f) sx_assert(&(f)->f_lck, SA_XLOCKED)
#define FUTEX_LOCKED 1
#define FUTEX_UNLOCKED 0
struct mtx futex_mtx; /* protects the futex list */
#define FUTEXES_LOCK mtx_lock(&futex_mtx)
#define FUTEXES_UNLOCK mtx_unlock(&futex_mtx)
#define FUTEX_SYSTEM_LOCK mtx_lock(&Giant)
#define FUTEX_SYSTEM_UNLOCK mtx_unlock(&Giant)
/* flags for futex_get() */
#define FUTEX_CREATE_WP 0x1 /* create waiting_proc */
#define FUTEX_DONTCREATE 0x2 /* don't create futex if not exists */
#define FUTEX_DONTEXISTS 0x4 /* return EINVAL if futex exists */
static struct futex *futex_get(void *, int);
static void futex_put(struct futex *);
static int futex_sleep(struct futex *, struct thread *, unsigned long);
static int futex_wake(struct futex *, int, struct futex *, int);
static int futex_atomic_op(struct thread *td, int encoded_op, caddr_t uaddr);
/* wp_flags */
#define FUTEX_WP_REQUEUED 0x1 /* wp requeued - wp moved from wp_list
* of futex where thread sleep to wp_list
* of another futex.
*/
#define FUTEX_WP_REMOVED 0x2 /* wp is woken up and removed from futex
* wp_list to prevent double wakeup.
*/
/* support.s */
int futex_xchgl(int oparg, caddr_t uaddr, int *oldval);
int futex_addl(int oparg, caddr_t uaddr, int *oldval);
int futex_orl(int oparg, caddr_t uaddr, int *oldval);
int futex_andl(int oparg, caddr_t uaddr, int *oldval);
int futex_xorl(int oparg, caddr_t uaddr, int *oldval);
int futex_xchgl(int oparg, uint32_t *uaddr, int *oldval);
int futex_addl(int oparg, uint32_t *uaddr, int *oldval);
int futex_orl(int oparg, uint32_t *uaddr, int *oldval);
int futex_andl(int oparg, uint32_t *uaddr, int *oldval);
int futex_xorl(int oparg, uint32_t *uaddr, int *oldval);
int
linux_sys_futex(struct thread *td, struct linux_sys_futex_args *args)
static void
futex_put(struct futex *f, struct waiting_proc *wp)
{
int val;
int ret;
struct l_timespec timeout = {0, 0};
int error = 0;
struct futex *f;
struct futex *newf;
int timeout_hz;
struct timeval tv = {0, 0};
struct futex *f2;
int op_ret;
struct linux_emuldata *em;
#ifdef DEBUG
if (ldebug(sys_futex))
printf(ARGS(futex, "%p, %i, %i, *, %p, %i"), args->uaddr, args->op,
args->val, args->uaddr2, args->val3);
#endif
/*
* Our implementation provides only privates futexes. Most of the apps
* should use private futexes but don't claim so. Therefore we treat
* all futexes as private by clearing the FUTEX_PRIVATE_FLAG. It works
* in most cases (ie. when futexes are not shared on file descriptor
* or between different processes.).
*/
args->op = (args->op & ~LINUX_FUTEX_PRIVATE_FLAG);
switch (args->op) {
case LINUX_FUTEX_WAIT:
FUTEX_SYSTEM_LOCK;
if ((error = copyin(args->uaddr,
&val, sizeof(val))) != 0) {
FUTEX_SYSTEM_UNLOCK;
return error;
}
if (val != args->val) {
FUTEX_SYSTEM_UNLOCK;
return EWOULDBLOCK;
}
if (args->timeout != NULL) {
if ((error = copyin(args->timeout,
&timeout, sizeof(timeout))) != 0) {
FUTEX_SYSTEM_UNLOCK;
return error;
}
}
#ifdef DEBUG
if (ldebug(sys_futex))
printf("FUTEX_WAIT %d: val = %d, uaddr = %p, "
"*uaddr = %d, timeout = %d.%09lu\n",
td->td_proc->p_pid, args->val,
args->uaddr, val, timeout.tv_sec,
(unsigned long)timeout.tv_nsec);
#endif
tv.tv_usec = timeout.tv_sec * 1000000 + timeout.tv_nsec / 1000;
timeout_hz = tvtohz(&tv);
if (timeout.tv_sec == 0 && timeout.tv_nsec == 0)
timeout_hz = 0;
/*
* If the user process requests a non null timeout,
* make sure we do not turn it into an infinite
* timeout because timeout_hz gets null.
*
* We use a minimal timeout of 1/hz. Maybe it would
* make sense to just return ETIMEDOUT without sleeping.
*/
if (((timeout.tv_sec != 0) || (timeout.tv_nsec != 0)) &&
(timeout_hz == 0))
timeout_hz = 1;
f = futex_get(args->uaddr, FUTEX_UNLOCKED);
ret = futex_sleep(f, td, timeout_hz);
futex_put(f);
#ifdef DEBUG
if (ldebug(sys_futex))
printf("FUTEX_WAIT %d: uaddr = %p, "
"ret = %d\n", td->td_proc->p_pid, args->uaddr, ret);
#endif
FUTEX_SYSTEM_UNLOCK;
switch (ret) {
case EWOULDBLOCK: /* timeout */
return ETIMEDOUT;
break;
case EINTR: /* signal */
return EINTR;
break;
case 0: /* FUTEX_WAKE received */
#ifdef DEBUG
if (ldebug(sys_futex))
printf("FUTEX_WAIT %d: uaddr = %p, "
"got FUTEX_WAKE\n",
td->td_proc->p_pid, args->uaddr);
#endif
return 0;
break;
default:
#ifdef DEBUG
if (ldebug(sys_futex))
printf("FUTEX_WAIT: unexpected ret = %d\n",
ret);
#endif
break;
}
/* NOTREACHED */
break;
case LINUX_FUTEX_WAKE:
FUTEX_SYSTEM_LOCK;
/*
* XXX: Linux is able to cope with different addresses
* corresponding to the same mapped memory in the sleeping
* and waker process(es).
*/
#ifdef DEBUG
if (ldebug(sys_futex))
printf("FUTEX_WAKE %d: uaddr = %p, val = %d\n",
td->td_proc->p_pid, args->uaddr, args->val);
#endif
f = futex_get(args->uaddr, FUTEX_UNLOCKED);
td->td_retval[0] = futex_wake(f, args->val, NULL, 0);
futex_put(f);
FUTEX_SYSTEM_UNLOCK;
break;
case LINUX_FUTEX_CMP_REQUEUE:
FUTEX_SYSTEM_LOCK;
if ((error = copyin(args->uaddr,
&val, sizeof(val))) != 0) {
FUTEX_SYSTEM_UNLOCK;
return error;
}
if (val != args->val3) {
FUTEX_SYSTEM_UNLOCK;
return EAGAIN;
}
f = futex_get(args->uaddr, FUTEX_UNLOCKED);
newf = futex_get(args->uaddr2, FUTEX_UNLOCKED);
td->td_retval[0] = futex_wake(f, args->val, newf,
(int)(unsigned long)args->timeout);
futex_put(f);
futex_put(newf);
FUTEX_SYSTEM_UNLOCK;
break;
case LINUX_FUTEX_WAKE_OP:
FUTEX_SYSTEM_LOCK;
#ifdef DEBUG
if (ldebug(sys_futex))
printf("FUTEX_WAKE_OP: %d: uaddr = %p, op = %d, "
"val = %x, uaddr2 = %p, val3 = %x\n",
td->td_proc->p_pid, args->uaddr, args->op,
args->val, args->uaddr2, args->val3);
#endif
f = futex_get(args->uaddr, FUTEX_UNLOCKED);
f2 = futex_get(args->uaddr2, FUTEX_UNLOCKED);
/*
* This function returns positive number as results and
* negative as errors
*/
op_ret = futex_atomic_op(td, args->val3, args->uaddr2);
#ifdef DEBUG
if (ldebug(sys_futex))
printf("futex_atomic_op ret %d\n", op_ret);
#endif
if (op_ret < 0) {
/* XXX: We don't handle the EFAULT yet. */
if (op_ret != -EFAULT) {
futex_put(f);
futex_put(f2);
FUTEX_SYSTEM_UNLOCK;
return (-op_ret);
}
futex_put(f);
futex_put(f2);
FUTEX_SYSTEM_UNLOCK;
return (EFAULT);
}
ret = futex_wake(f, args->val, NULL, 0);
futex_put(f);
if (op_ret > 0) {
op_ret = 0;
/*
* Linux abuses the address of the timespec parameter
* as the number of retries.
*/
op_ret += futex_wake(f2,
(int)(unsigned long)args->timeout, NULL, 0);
ret += op_ret;
}
futex_put(f2);
td->td_retval[0] = ret;
FUTEX_SYSTEM_UNLOCK;
break;
case LINUX_FUTEX_LOCK_PI:
/* not yet implemented */
return (ENOSYS);
case LINUX_FUTEX_UNLOCK_PI:
/* not yet implemented */
return (ENOSYS);
case LINUX_FUTEX_TRYLOCK_PI:
/* not yet implemented */
return (ENOSYS);
case LINUX_FUTEX_REQUEUE:
/*
* Glibc does not use this operation since Jun 2004 (2.3.3),
* as it is racy and replaced by FUTEX_CMP_REQUEUE operation.
* Glibc versions prior to 2.3.3 fall back to FUTEX_WAKE when
* FUTEX_REQUEUE returned EINVAL.
*/
em = em_find(td->td_proc, EMUL_DONTLOCK);
if (em->used_requeue == 0) {
printf("linux(%s (%d)) sys_futex: "
"unsupported futex_requeue op\n",
td->td_proc->p_comm, td->td_proc->p_pid);
em->used_requeue = 1;
}
return (EINVAL);
default:
printf("linux_sys_futex: unknown op %d\n",
args->op);
return (ENOSYS);
FUTEX_ASSERT_LOCKED(f);
if (wp != NULL) {
if ((wp->wp_flags & FUTEX_WP_REMOVED) == 0)
TAILQ_REMOVE(&f->f_waiting_proc, wp, wp_list);
free(wp, M_FUTEX_WP);
}
FUTEXES_LOCK;
if (--f->f_refcount == 0) {
LIST_REMOVE(f, f_list);
FUTEXES_UNLOCK;
FUTEX_UNLOCK(f);
FUTEX_DESTROY(f);
free(f, M_FUTEX);
return;
}
FUTEXES_UNLOCK;
FUTEX_UNLOCK(f);
}
static int
futex_get0(uint32_t *uaddr, struct futex **newf, uint32_t flags)
{
struct futex *f, *tmpf;
*newf = tmpf = NULL;
retry:
FUTEXES_LOCK;
LIST_FOREACH(f, &futex_list, f_list) {
if (f->f_uaddr == uaddr) {
if (tmpf != NULL) {
FUTEX_UNLOCK(tmpf);
FUTEX_DESTROY(tmpf);
free(tmpf, M_FUTEX);
}
if (flags & FUTEX_DONTEXISTS) {
FUTEXES_UNLOCK;
return (EINVAL);
}
/*
* Increment refcount of the found futex to
* prevent it from deallocation before FUTEX_LOCK()
*/
++f->f_refcount;
FUTEXES_UNLOCK;
FUTEX_LOCK(f);
*newf = f;
return (0);
}
}
if (flags & FUTEX_DONTCREATE) {
FUTEXES_UNLOCK;
return (0);
}
if (tmpf == NULL) {
FUTEXES_UNLOCK;
tmpf = malloc(sizeof(*tmpf), M_FUTEX, M_WAITOK | M_ZERO);
tmpf->f_uaddr = uaddr;
tmpf->f_refcount = 1;
FUTEX_INIT(tmpf);
TAILQ_INIT(&tmpf->f_waiting_proc);
/*
* Lock the new futex before an insert into the futex_list
* to prevent futex usage by other.
*/
FUTEX_LOCK(tmpf);
goto retry;
}
LIST_INSERT_HEAD(&futex_list, tmpf, f_list);
FUTEXES_UNLOCK;
*newf = tmpf;
return (0);
}
static struct futex *
futex_get(void *uaddr, int locked)
static int
futex_get(uint32_t *uaddr, struct waiting_proc **wp, struct futex **f,
uint32_t flags)
{
struct futex *f;
int error;
if (locked == FUTEX_UNLOCKED)
FUTEX_LOCK;
LIST_FOREACH(f, &futex_list, f_list) {
if (f->f_uaddr == uaddr) {
f->f_refcount++;
if (locked == FUTEX_UNLOCKED)
FUTEX_UNLOCK;
return f;
if (flags & FUTEX_CREATE_WP) {
*wp = malloc(sizeof(struct waiting_proc), M_FUTEX_WP, M_WAITOK);
(*wp)->wp_flags = 0;
}
error = futex_get0(uaddr, f, flags);
if (error) {
if (flags & FUTEX_CREATE_WP)
free(*wp, M_FUTEX_WP);
return (error);
}
if (flags & FUTEX_CREATE_WP) {
TAILQ_INSERT_HEAD(&(*f)->f_waiting_proc, *wp, wp_list);
(*wp)->wp_futex = *f;
}
return (error);
}
static int
futex_sleep(struct futex *f, struct waiting_proc *wp, unsigned long timeout)
{
int error;
FUTEX_ASSERT_LOCKED(f);
error = sx_sleep(wp, &f->f_lck, PCATCH, "futex", timeout);
if (wp->wp_flags & FUTEX_WP_REQUEUED) {
KASSERT(f != wp->wp_futex, ("futex != wp_futex"));
futex_put(f, NULL);
f = wp->wp_futex;
FUTEX_LOCK(f);
}
futex_put(f, wp);
return (error);
}
static int
futex_wake(struct futex *f, int n)
{
struct waiting_proc *wp, *wpt;
int count = 0;
FUTEX_ASSERT_LOCKED(f);
TAILQ_FOREACH_SAFE(wp, &f->f_waiting_proc, wp_list, wpt) {
wp->wp_flags |= FUTEX_WP_REMOVED;
TAILQ_REMOVE(&f->f_waiting_proc, wp, wp_list);
wakeup_one(wp);
if (++count == n)
break;
}
return (count);
}
static int
futex_requeue(struct futex *f, int n, struct futex *f2, int n2)
{
struct waiting_proc *wp, *wpt;
int count = 0;
FUTEX_ASSERT_LOCKED(f);
FUTEX_ASSERT_LOCKED(f2);
TAILQ_FOREACH_SAFE(wp, &f->f_waiting_proc, wp_list, wpt) {
if (++count <= n) {
wp->wp_flags |= FUTEX_WP_REMOVED;
TAILQ_REMOVE(&f->f_waiting_proc, wp, wp_list);
wakeup_one(wp);
} else {
wp->wp_flags |= FUTEX_WP_REQUEUED;
/* Move wp to wp_list of f2 futex */
TAILQ_REMOVE(&f->f_waiting_proc, wp, wp_list);
TAILQ_INSERT_HEAD(&f2->f_waiting_proc, wp, wp_list);
/*
* Thread which sleeps on wp after waking should
* acquire f2 lock, so increment refcount of f2 to
* prevent it from premature deallocation.
*/
wp->wp_futex = f2;
FUTEXES_LOCK;
++f2->f_refcount;
FUTEXES_UNLOCK;
if (count - n >= n2)
break;
}
}
f = malloc(sizeof(*f), M_LINUX, M_WAITOK);
f->f_uaddr = uaddr;
f->f_refcount = 1;
TAILQ_INIT(&f->f_waiting_proc);
LIST_INSERT_HEAD(&futex_list, f, f_list);
if (locked == FUTEX_UNLOCKED)
FUTEX_UNLOCK;
return f;
}
static void
futex_put(f)
struct futex *f;
{
FUTEX_LOCK;
f->f_refcount--;
if (f->f_refcount == 0) {
LIST_REMOVE(f, f_list);
free(f, M_LINUX);
}
FUTEX_UNLOCK;
return;
return (count);
}
static int
futex_sleep(struct futex *f, struct thread *td, unsigned long timeout)
futex_wait(struct futex *f, struct waiting_proc *wp, struct l_timespec *ts)
{
struct waiting_proc *wp;
int ret;
struct l_timespec timeout = {0, 0};
struct timeval tv = {0, 0};
int timeout_hz;
int error;
wp = malloc(sizeof(*wp), M_LINUX, M_WAITOK);
wp->wp_t = td;
wp->wp_new_futex = NULL;
FUTEX_LOCK;
TAILQ_INSERT_TAIL(&f->f_waiting_proc, wp, wp_list);
FUTEX_UNLOCK;
#ifdef DEBUG
if (ldebug(sys_futex))
printf("FUTEX --> %d tlseep timeout = %ld\n",
td->td_proc->p_pid, timeout);
#endif
ret = tsleep(wp, PCATCH | PZERO, "linuxfutex", timeout);
#ifdef DEBUG
if (ldebug(sys_futex))
printf("FUTEX -> %d tsleep returns %d\n",
td->td_proc->p_pid, ret);
#endif
FUTEX_LOCK;
TAILQ_REMOVE(&f->f_waiting_proc, wp, wp_list);
FUTEX_UNLOCK;
/* if we got woken up in futex_wake */
if ((ret == 0) && (wp->wp_new_futex != NULL)) {
/* suspend us on the new futex */
ret = futex_sleep(wp->wp_new_futex, td, timeout);
/* and release the old one */
futex_put(wp->wp_new_futex);
if (ts != NULL) {
error = copyin(ts, &timeout, sizeof(timeout));
if (error)
return (error);
}
free(wp, M_LINUX);
tv.tv_usec = timeout.tv_sec * 1000000 + timeout.tv_nsec / 1000;
timeout_hz = tvtohz(&tv);
return ret;
}
static int
futex_wake(struct futex *f, int n, struct futex *newf, int n2)
{
struct waiting_proc *wp;
int count;
if (timeout.tv_sec == 0 && timeout.tv_nsec == 0)
timeout_hz = 0;
/*
* Linux is very strange it wakes up N threads for
* all operations BUT requeue ones where its N+1
* mimic this.
* If the user process requests a non null timeout,
* make sure we do not turn it into an infinite
* timeout because timeout_hz gets null.
*
* We use a minimal timeout of 1/hz. Maybe it would
* make sense to just return ETIMEDOUT without sleeping.
*/
count = newf ? 0 : 1;
if (((timeout.tv_sec != 0) || (timeout.tv_nsec != 0)) &&
(timeout_hz == 0))
timeout_hz = 1;
FUTEX_LOCK;
TAILQ_FOREACH(wp, &f->f_waiting_proc, wp_list) {
if (count <= n) {
wakeup_one(wp);
count++;
} else {
if (newf != NULL) {
/* futex_put called after tsleep */
wp->wp_new_futex = futex_get(newf->f_uaddr,
FUTEX_LOCKED);
wakeup_one(wp);
if (count - n >= n2)
break;
}
}
}
FUTEX_UNLOCK;
error = futex_sleep(f, wp, timeout_hz);
if (error == EWOULDBLOCK)
error = ETIMEDOUT;
return count;
return (error);
}
static int
futex_atomic_op(struct thread *td, int encoded_op, caddr_t uaddr)
futex_atomic_op(struct thread *td, int encoded_op, uint32_t *uaddr)
{
int op = (encoded_op >> 28) & 7;
int cmp = (encoded_op >> 24) & 15;
@ -535,15 +400,238 @@ futex_atomic_op(struct thread *td, int encoded_op, caddr_t uaddr)
}
}
int
linux_sys_futex(struct thread *td, struct linux_sys_futex_args *args)
{
int op_ret, val, ret, nrwake;
struct linux_emuldata *em;
struct waiting_proc *wp;
struct futex *f, *f2;
int error = 0;
/*
* Our implementation provides only privates futexes. Most of the apps
* should use private futexes but don't claim so. Therefore we treat
* all futexes as private by clearing the FUTEX_PRIVATE_FLAG. It works
* in most cases (ie. when futexes are not shared on file descriptor
* or between different processes.).
*/
args->op = (args->op & ~LINUX_FUTEX_PRIVATE_FLAG);
switch (args->op) {
case LINUX_FUTEX_WAIT:
#ifdef DEBUG
if (ldebug(sys_futex))
printf(ARGS(sys_futex, "futex_wait val %d uaddr %p"),
args->val, args->uaddr);
#endif
error = futex_get(args->uaddr, &wp, &f, FUTEX_CREATE_WP);
if (error)
return (error);
error = copyin(args->uaddr, &val, sizeof(val));
if (error) {
futex_put(f, wp);
return (error);
}
if (val != args->val) {
#ifdef DEBUG
if (ldebug(sys_futex))
printf(ARGS(sys_futex, "futex_wait uaddr %p WHOOPS %d != %d"),
args->uaddr, args->val, val);
#endif
futex_put(f, wp);
return (EWOULDBLOCK);
}
error = futex_wait(f, wp, args->timeout);
break;
case LINUX_FUTEX_WAKE:
/*
* XXX: Linux is able to cope with different addresses
* corresponding to the same mapped memory in the sleeping
* and waker process(es).
*/
#ifdef DEBUG
if (ldebug(sys_futex))
printf(ARGS(sys_futex, "futex_wake val %d uaddr %p"),
args->val, args->uaddr);
#endif
error = futex_get(args->uaddr, NULL, &f, FUTEX_DONTCREATE);
if (error)
return (error);
if (f == NULL) {
td->td_retval[0] = 0;
return (error);;
}
td->td_retval[0] = futex_wake(f, args->val);
futex_put(f, NULL);
break;
case LINUX_FUTEX_CMP_REQUEUE:
#ifdef DEBUG
if (ldebug(sys_futex))
printf(ARGS(sys_futex, "futex_cmp_requeue uaddr %p "
"val %d val3 %d uaddr2 %p val2 %d"),
args->uaddr, args->val, args->val3, args->uaddr2,
(int)(unsigned long)args->timeout);
#endif
/*
* Linux allows this, we would not, it is an incorrect
* usage of declared ABI, so return EINVAL.
*/
if (args->uaddr == args->uaddr2)
return (EINVAL);
error = futex_get0(args->uaddr, &f, 0);
if (error)
return (error);
/*
* To avoid deadlocks return EINVAL if second futex
* exists at this time. Otherwise create the new futex
* and ignore false positive LOR which thus happens.
*
* Glibc fall back to FUTEX_WAKE in case of any error
* returned by FUTEX_CMP_REQUEUE.
*/
error = futex_get0(args->uaddr2, &f2, FUTEX_DONTEXISTS);
if (error) {
futex_put(f, NULL);
return (error);
}
error = copyin(args->uaddr, &val, sizeof(val));
if (error) {
futex_put(f2, NULL);
futex_put(f, NULL);
return (error);
}
if (val != args->val3) {
#ifdef DEBUG
if (ldebug(sys_futex))
printf(ARGS(sys_futex, "futex_cmp_requeue WHOOPS"
" VAL %d != UVAL %d"), args->val, val);
#endif
futex_put(f2, NULL);
futex_put(f, NULL);
return (EAGAIN);
}
nrwake = (int)(unsigned long)args->timeout;
td->td_retval[0] = futex_requeue(f, args->val, f2, nrwake);
futex_put(f2, NULL);
futex_put(f, NULL);
break;
case LINUX_FUTEX_WAKE_OP:
#ifdef DEBUG
if (ldebug(sys_futex))
printf(ARGS(sys_futex, "futex_wake_op "
"uaddr %p op %d val %x uaddr2 %p val3 %x"),
args->uaddr, args->op, args->val,
args->uaddr2, args->val3);
#endif
error = futex_get0(args->uaddr, &f, 0);
if (error)
return (error);
if (args->uaddr != args->uaddr2)
error = futex_get0(args->uaddr2, &f2, 0);
if (error) {
futex_put(f, NULL);
return (error);
}
/*
* This function returns positive number as results and
* negative as errors
*/
op_ret = futex_atomic_op(td, args->val3, args->uaddr2);
if (op_ret < 0) {
/* XXX: We don't handle the EFAULT yet. */
if (op_ret != -EFAULT) {
if (f2 != NULL)
futex_put(f2, NULL);
futex_put(f, NULL);
return (-op_ret);
}
if (f2 != NULL)
futex_put(f2, NULL);
futex_put(f, NULL);
return (EFAULT);
}
ret = futex_wake(f, args->val);
if (op_ret > 0) {
op_ret = 0;
nrwake = (int)(unsigned long)args->timeout;
if (f2 != NULL)
op_ret += futex_wake(f2, nrwake);
else
op_ret += futex_wake(f, nrwake);
ret += op_ret;
}
if (f2 != NULL)
futex_put(f2, NULL);
futex_put(f, NULL);
td->td_retval[0] = ret;
break;
case LINUX_FUTEX_LOCK_PI:
/* not yet implemented */
return (ENOSYS);
case LINUX_FUTEX_UNLOCK_PI:
/* not yet implemented */
return (ENOSYS);
case LINUX_FUTEX_TRYLOCK_PI:
/* not yet implemented */
return (ENOSYS);
case LINUX_FUTEX_REQUEUE:
/*
* Glibc does not use this operation since version 2.3.3,
* as it is racy and replaced by FUTEX_CMP_REQUEUE operation.
* Glibc versions prior to 2.3.3 fall back to FUTEX_WAKE when
* FUTEX_REQUEUE returned EINVAL.
*/
em = em_find(td->td_proc, EMUL_DONTLOCK);
if (em->used_requeue == 0) {
printf("linux(%s (%d)) sys_futex: "
"unsupported futex_requeue op\n",
td->td_proc->p_comm, td->td_proc->p_pid);
em->used_requeue = 1;
}
return (EINVAL);
default:
printf("linux_sys_futex: unknown op %d\n", args->op);
return (ENOSYS);
}
return (error);
}
int
linux_set_robust_list(struct thread *td, struct linux_set_robust_list_args *args)
{
struct linux_emuldata *em;
#ifdef DEBUG
#ifdef DEBUG
if (ldebug(set_robust_list))
printf(ARGS(set_robust_list, ""));
printf(ARGS(set_robust_list, "head %p len %d"),
args->head, args->len);
#endif
if (args->len != sizeof(struct linux_robust_list_head))
return (EINVAL);
@ -598,16 +686,16 @@ linux_get_robust_list(struct thread *td, struct linux_get_robust_list_args *args
}
static int
handle_futex_death(void *uaddr, pid_t pid, int pi)
handle_futex_death(struct proc *p, uint32_t *uaddr, int pi)
{
int uval, nval, mval;
uint32_t uval, nval, mval;
struct futex *f;
int error;
retry:
if (copyin(uaddr, &uval, 4))
return (EFAULT);
if ((uval & FUTEX_TID_MASK) == pid) {
if ((uval & FUTEX_TID_MASK) == p->p_pid) {
mval = (uval & FUTEX_WAITERS) | FUTEX_OWNER_DIED;
nval = casuword32(uaddr, uval, mval);
@ -618,8 +706,14 @@ handle_futex_death(void *uaddr, pid_t pid, int pi)
goto retry;
if (!pi && (uval & FUTEX_WAITERS)) {
f = futex_get(uaddr, FUTEX_UNLOCKED);
futex_wake(f, 1, NULL, 0);
error = futex_get(uaddr, NULL, &f,
FUTEX_DONTCREATE);
if (error)
return (error);
if (f != NULL) {
futex_wake(f, 1);
futex_put(f, NULL);
}
}
}
@ -671,10 +765,8 @@ release_futexes(struct proc *p)
rc = fetch_robust_entry(&next_entry, PTRIN(&entry->next), &next_pi);
if (entry != pending)
if (handle_futex_death((char *)entry + futex_offset,
p->p_pid, pi))
if (handle_futex_death(p, (uint32_t *)entry + futex_offset, pi))
return;
if (rc)
return;
@ -688,6 +780,5 @@ release_futexes(struct proc *p)
}
if (pending)
handle_futex_death((char *) pending + futex_offset,
p->p_pid, pip);
handle_futex_death(p, (uint32_t *)pending + futex_offset, pip);
}

View File

@ -112,7 +112,7 @@ static int linux_szplatform;
const char *linux_platform;
extern LIST_HEAD(futex_list, futex) futex_list;
extern struct sx futex_sx;
extern struct mtx futex_mtx;
static eventhandler_tag linux_exit_tag;
static eventhandler_tag linux_schedtail_tag;
@ -1083,7 +1083,7 @@ linux_elf_modevent(module_t mod, int type, void *data)
mtx_init(&emul_lock, "emuldata lock", NULL, MTX_DEF);
sx_init(&emul_shared_lock, "emuldata->shared lock");
LIST_INIT(&futex_list);
sx_init(&futex_sx, "futex protection lock");
mtx_init(&futex_mtx, "ftllk", NULL, MTX_DEF);
linux_exit_tag = EVENTHANDLER_REGISTER(process_exit, linux_proc_exit,
NULL, 1000);
linux_schedtail_tag = EVENTHANDLER_REGISTER(schedtail, linux_schedtail,
@ -1116,7 +1116,7 @@ linux_elf_modevent(module_t mod, int type, void *data)
linux_device_unregister_handler(*ldhp);
mtx_destroy(&emul_lock);
sx_destroy(&emul_shared_lock);
sx_destroy(&futex_sx);
mtx_destroy(&futex_mtx);
EVENTHANDLER_DEREGISTER(process_exit, linux_exit_tag);
EVENTHANDLER_DEREGISTER(schedtail, linux_schedtail_tag);
EVENTHANDLER_DEREGISTER(process_exec, linux_exec_tag);