/* internally to wait for one object. Also used as a shortcut to wait
on events and semaphores */
static int _DkObjectWaitOne (PAL_HANDLE handle, PAL_NUM timeout)
{
/* only for all these handle which has a file descriptor, or
a eventfd. events and semaphores will skip this part */
if (HANDLE_HDR(handle)->flags & HAS_FDS) {
struct pollfd fds[MAX_FDS];
int off[MAX_FDS];
int nfds = 0;
for (int i = 0 ; i < MAX_FDS ; i++) {
int events = 0;
if ((HANDLE_HDR(handle)->flags & RFD(i)) &&
!(HANDLE_HDR(handle)->flags & ERROR(i)))
events |= POLLIN;
if ((HANDLE_HDR(handle)->flags & WFD(i)) &&
!(HANDLE_HDR(handle)->flags & WRITEABLE(i)) &&
!(HANDLE_HDR(handle)->flags & ERROR(i)))
events |= POLLOUT;
if (events) {
fds[nfds].fd = handle->generic.fds[i];
fds[nfds].events = events|POLLHUP|POLLERR;
fds[nfds].revents = 0;
off[nfds] = i;
nfds++;
}
}
if (!nfds)
return -PAL_ERROR_TRYAGAIN;
int64_t waittime = timeout;
int ret = ocall_poll(fds, nfds, timeout != NO_TIMEOUT ? &waittime : NULL);
if (IS_ERR(ret))
return unix_to_pal_error(ERRNO(ret));
if (!ret)
return -PAL_ERROR_TRYAGAIN;
for (int i = 0 ; i < nfds ; i++) {
if (!fds[i].revents)
continue;
if (fds[i].revents & POLLOUT)
HANDLE_HDR(handle)->flags |= WRITEABLE(off[i]);
if (fds[i].revents & (POLLHUP|POLLERR))
HANDLE_HDR(handle)->flags |= ERROR(off[i]);
}
return 0;
}
const struct handle_ops * ops = HANDLE_OPS(handle);
if (!ops || !ops->wait)
return -PAL_ERROR_NOTSUPPORT;
return ops->wait(handle, timeout);
}
int _DkMutexLock (struct mutex_handle * mut)
{
int i, c = 0;
int ret;
struct atomic_int * m = &mut->value;
/* Spin and try to take lock */
for (i = 0; i < MUTEX_SPINLOCK_TIMES; i++) {
c = atomic_dec_and_test(m);
if (c)
goto success;
cpu_relax();
}
/* The lock is now contended */
while (!c) {
int val = atomic_read(m);
if (val == 1)
goto again;
ret = INLINE_SYSCALL(futex, 6, m, FUTEX_WAIT, val, NULL, NULL, 0);
if (IS_ERR(ret) &&
ERRNO(ret) != EWOULDBLOCK &&
ERRNO(ret) != EINTR) {
ret = unix_to_pal_error(ERRNO(ret));
goto out;
}
#ifdef DEBUG_MUTEX
if (IS_ERR(ret))
printf("mutex held by thread %d\n", mut->owner);
#endif
again:
/* Upon wakeup, we still need to check whether mutex is unlocked or
* someone else took it.
* If c==0 upon return from xchg (i.e., the older value of m==0), we
* will exit the loop. Else, we sleep again (through a futex call).
*/
c = atomic_dec_and_test(m);
}
success:
#ifdef DEBUG_MUTEX
mut->owner = INLINE_SYSCALL(gettid, 0);
#endif
ret = 0;
out:
return ret;
}
int _DkVirtualMemoryAlloc (void ** paddr, int size, int alloc_type,
int prot)
{
void * addr = *paddr, * mem = addr;
int flags = HOST_FLAGS(alloc_type, prot|PAL_PROT_WRITECOPY);
prot = HOST_PROT(prot);
/* The memory should have MAP_PRIVATE and MAP_ANONYMOUS */
flags |= MAP_ANONYMOUS|(addr ? MAP_FIXED : 0);
mem = (void *) ARCH_MMAP(addr, size, prot, flags, -1, 0);
if (IS_ERR_P(mem))
return unix_to_pal_error(ERRNO_P(mem));
*paddr = mem;
return 0;
}
int _DkMutexLockTimeout (struct mutex_handle * m, PAL_NUM timeout)
{
int i, ret = 0;
#ifdef DEBUG_MUTEX
int tid = INLINE_SYSCALL(gettid, 0);
#endif
/* If this is a trylock-style call, break more quickly. */
int iterations = (timeout == 0) ? 1 : MUTEX_SPINLOCK_TIMES;
/* Spin and try to take lock. Ignore any contribution this makes toward
* the timeout.*/
for (i = 0; i < iterations; i++) {
if (MUTEX_UNLOCKED == cmpxchg(&m->locked, MUTEX_UNLOCKED, MUTEX_LOCKED))
goto success;
CPU_RELAX();
}
if (timeout == 0) {
ret = -PAL_ERROR_TRYAGAIN;
goto out;
}
// Bump up the waiters count; we are probably going to block
atomic_inc(&m->nwaiters);
while (MUTEX_LOCKED == cmpxchg(&m->locked, MUTEX_UNLOCKED, MUTEX_LOCKED)) {
struct timespec waittime, *waittimep = NULL;
if (timeout != NO_TIMEOUT) {
long sec = timeout / 1000000;
long microsec = timeout - (sec * 1000000);
waittime.tv_sec = sec;
waittime.tv_nsec = microsec * 1000;
waittimep = &waittime;
}
ret = INLINE_SYSCALL(futex, 6, m, FUTEX_WAIT, MUTEX_LOCKED, waittimep, NULL, 0);
if (IS_ERR(ret)) {
if (ERRNO(ret) == EWOULDBLOCK) {
if (timeout != NO_TIMEOUT) {
ret = -PAL_ERROR_TRYAGAIN;
atomic_dec(&m->nwaiters);
goto out;
}
} else {
#ifdef DEBUG_MUTEX
printf("futex failed (err = %d)\n", ERRNO(ret));
#endif
ret = unix_to_pal_error(ERRNO(ret));
atomic_dec(&m->nwaiters);
goto out;
}
}
}
atomic_dec(&m->nwaiters);
success:
#ifdef DEBUG_MUTEX
m->owner = tid;
#endif
ret = 0;
out:
#ifdef DEBUG_MUTEX
if (ret < 0)
printf("mutex failed (%s, tid = %d)\n", PAL_STRERROR(ret), tid);
#endif
return ret;
}
int _DkVirtualMemoryProtect (void * addr, int size, int prot)
{
int ret = INLINE_SYSCALL(mprotect, 3, addr, size, HOST_PROT(prot));
return IS_ERR(ret) ? unix_to_pal_error(ERRNO(ret)) : 0;
}
int _DkVirtualMemoryFree (void * addr, int size)
{
int ret = INLINE_SYSCALL(munmap, 2, addr, size);
return IS_ERR(ret) ? unix_to_pal_error(ERRNO(ret)) : 0;
}
int _DkMutexLockTimeout (struct mutex_handle * mut, int timeout)
{
int i, c = 0;
if (timeout == -1)
return -_DkMutexLock(mut);
struct atomic_int * m = &mut->value;
/* Spin and try to take lock */
for (i = 0 ; i < MUTEX_SPINLOCK_TIMES ; i++)
{
c = atomic_dec_and_test(m);
if (c)
goto success;
CPU_RELAX();
}
/* The lock is now contended */
int ret;
if (timeout == 0) {
ret = c ? 0 : -PAL_ERROR_TRYAGAIN;
goto out;
}
while (!c) {
int val = atomic_read(m);
if (val == 1)
goto again;
struct timespec waittime;
long sec = timeout / 1000000;
long microsec = timeout - (sec * 1000000);
waittime.tv_sec = sec;
waittime.tv_nsec = microsec * 1000;
ret = INLINE_SYSCALL(_umtx_op, 5, m, UMTX_OP_WAIT_UINT, val,
NULL, &waittime);
if (IS_ERR(ret) && ERRNO(ret) != EWOULDBLOCK) {
ret = unix_to_pal_error(ERRNO(ret));
goto out;
}
#ifdef DEBUG_MUTEX
if (IS_ERR(ret))
printf("mutex held by thread %d\n", mut->owner);
#endif
again:
/* Upon wakeup, we still need to check whether mutex is unlocked or
* someone else took it.
* If c==0 upon return from xchg (i.e., the older value of m==0), we
* will exit the loop. Else, we sleep again (through a futex call).
*/
c = atomic_dec_and_test(m);
}
success:
#ifdef DEBUG_MUTEX
mut->owner = INLINE_SYSCALL(gettid, 0);
#endif
ret = 0;
out:
return ret;
}
/* internally to wait for one object. Also used as a shortcut to wait
on events and semaphores */
static int _DkObjectWaitOne (PAL_HANDLE handle, int timeout)
{
/* only for all these handle which has a file descriptor, or
a eventfd. events and semaphores will skip this part */
if (handle->__in.flags & HAS_FDS) {
struct pollfd fds[MAX_FDS];
int off[MAX_FDS];
int nfds = 0;
for (int i = 0 ; i < MAX_FDS ; i++) {
int events = 0;
if ((handle->__in.flags & RFD(i)) &&
!(handle->__in.flags & ERROR(i)))
events |= POLLIN;
if ((handle->__in.flags & WFD(i)) &&
!(handle->__in.flags & WRITEABLE(i)) &&
!(handle->__in.flags & ERROR(i)))
events |= POLLOUT;
if (events) {
fds[nfds].fd = handle->__in.fds[i];
fds[nfds].events = events|POLLHUP|POLLERR;
fds[nfds].revents = 0;
off[nfds] = i;
nfds++;
}
}
if (!nfds)
return -PAL_ERROR_TRYAGAIN;
int ret = INLINE_SYSCALL(poll, 3, &fds, nfds,
timeout ? timeout : -1);
if (IS_ERR(ret))
switch (ERRNO(ret)) {
case EINTR:
return -PAL_ERROR_INTERRUPTED;
default:
return unix_to_pal_error(ERRNO(ret));
}
if (!ret)
return -PAL_ERROR_TRYAGAIN;
for (int i = 0 ; i < nfds ; i++) {
if (!fds[i].revents)
continue;
if (fds[i].revents & POLLOUT)
handle->__in.flags |= WRITEABLE(off[i]);
if (fds[i].revents & (POLLHUP|POLLERR))
handle->__in.flags |= ERROR(off[i]);
}
return 0;
}
const struct handle_ops * ops = HANDLE_OPS(handle);
if (!ops->wait)
return -PAL_ERROR_NOTSUPPORT;
return ops->wait(handle, timeout);
}
/* _DkObjectsWaitAny for internal use. The function wait for any of the handle
in the handle array. timeout can be set for the wait. */
int _DkObjectsWaitAny (int count, PAL_HANDLE * handleArray, int timeout,
PAL_HANDLE * polled)
{
if (count <= 0)
return 0;
if (count == 1) {
*polled = handleArray[0];
return _DkObjectWaitOne(handleArray[0], timeout);
}
int i, j, ret, maxfds = 0, nfds = 0;
/* we are not gonna to allow any polling on muliple synchronous
objects, doing this is simply violating the division of
labor between PAL and library OS */
for (i = 0 ; i < count ; i++) {
PAL_HANDLE hdl = handleArray[i];
if (!hdl)
continue;
if (!(hdl->__in.flags & HAS_FDS))
return -PAL_ERROR_NOTSUPPORT;
/* eliminate repeated entries */
for (j = 0 ; j < i ; j++)
if (hdl == handleArray[j])
break;
if (j == i) {
for (j = 0 ; j < MAX_FDS ; j++)
if (hdl->__in.flags & (RFD(j)|WFD(j)))
maxfds++;
}
}
struct pollfd * fds = __alloca(sizeof(struct pollfd) * maxfds);
PAL_HANDLE * hdls = __alloca(sizeof(PAL_HANDLE) * maxfds);
for (i = 0 ; i < count ; i++) {
PAL_HANDLE hdl = handleArray[i];
if (!hdl)
continue;
for (j = 0 ; j < i ; j++)
if (hdl == handleArray[j])
break;
if (j < i)
continue;
for (j = 0 ; j < MAX_FDS ; j++) {
int events = 0;
if ((hdl->__in.flags & RFD(j)) &&
!(hdl->__in.flags & ERROR(j)))
events |= POLLIN;
if ((hdl->__in.flags & WFD(j)) &&
!(hdl->__in.flags & WRITEABLE(j)) &&
!(hdl->__in.flags & ERROR(j)))
events |= POLLOUT;
if (events && hdl->__in.fds[j] != PAL_IDX_POISON) {
fds[nfds].fd = hdl->__in.fds[j];
fds[nfds].events = events|POLLHUP|POLLERR;
fds[nfds].revents = 0;
hdls[nfds] = hdl;
nfds++;
}
}
}
if (!nfds)
return -PAL_ERROR_TRYAGAIN;
ret = INLINE_SYSCALL(poll, 3, fds, nfds, timeout ? timeout : -1);
if (IS_ERR(ret))
switch (ERRNO(ret)) {
case EINTR:
return -PAL_ERROR_INTERRUPTED;
default:
return unix_to_pal_error(ERRNO(ret));
}
if (!ret)
return -PAL_ERROR_TRYAGAIN;
PAL_HANDLE polled_hdl = NULL;
for (i = 0 ; i < nfds ; i++) {
if (!fds[i].revents)
continue;
PAL_HANDLE hdl = hdls[i];
if (polled_hdl) {
if (hdl != polled_hdl)
continue;
} else {
polled_hdl = hdl;
}
for (j = 0 ; j < MAX_FDS ; j++)
if ((hdl->__in.flags & (RFD(j)|WFD(j))) &&
hdl->__in.fds[j] == fds[i].fd)
break;
if (j == MAX_FDS)
continue;
if (fds[i].revents & POLLOUT)
hdl->__in.flags |= WRITEABLE(j);
if (fds[i].revents & (POLLHUP|POLLERR))
hdl->__in.flags |= ERROR(j);
}
*polled = polled_hdl;
return polled_hdl ? 0 : -PAL_ERROR_TRYAGAIN;
}
/* _DkObjectsWaitAny for internal use. The function wait for any of the handle
in the handle array. timeout can be set for the wait. */
int _DkObjectsWaitAny (int count, PAL_HANDLE * handleArray, PAL_NUM timeout,
PAL_HANDLE * polled)
{
if (count <= 0)
return 0;
if (count == 1) {
// It is possible to have NULL pointers in the handle array.
// In this case, assume nothing is polled.
if (!handleArray[0])
return -PAL_ERROR_TRYAGAIN;
int rv = _DkObjectWaitOne(handleArray[0], timeout);
if (rv == 0)
*polled = handleArray[0];
return rv;
}
int i, j, ret, maxfds = 0, nfds = 0;
/* we are not gonna to allow any polling on muliple synchronous
objects, doing this is simply violating the division of
labor between PAL and library OS */
for (i = 0 ; i < count ; i++) {
PAL_HANDLE hdl = handleArray[i];
if (!hdl)
continue;
if (!(HANDLE_HDR(hdl)->flags & HAS_FDS))
return -PAL_ERROR_NOTSUPPORT;
/* eliminate repeated entries */
for (j = 0 ; j < i ; j++)
if (hdl == handleArray[j])
break;
if (j == i) {
for (j = 0 ; j < MAX_FDS ; j++)
if (HANDLE_HDR(hdl)->flags & (RFD(j)|WFD(j)))
maxfds++;
}
}
struct pollfd * fds = __alloca(sizeof(struct pollfd) * maxfds);
PAL_HANDLE * hdls = __alloca(sizeof(PAL_HANDLE) * maxfds);
for (i = 0 ; i < count ; i++) {
PAL_HANDLE hdl = handleArray[i];
if (!hdl)
continue;
for (j = 0 ; j < i ; j++)
if (hdl == handleArray[j])
break;
if (j < i)
continue;
for (j = 0 ; j < MAX_FDS ; j++) {
int events = 0;
if ((HANDLE_HDR(hdl)->flags & RFD(j)) &&
!(HANDLE_HDR(hdl)->flags & ERROR(j)))
events |= POLLIN;
if ((HANDLE_HDR(hdl)->flags & WFD(j)) &&
!(HANDLE_HDR(hdl)->flags & WRITEABLE(j)) &&
!(HANDLE_HDR(hdl)->flags & ERROR(j)))
events |= POLLOUT;
if (events && hdl->generic.fds[j] != PAL_IDX_POISON) {
fds[nfds].fd = hdl->generic.fds[j];
fds[nfds].events = events|POLLHUP|POLLERR;
fds[nfds].revents = 0;
hdls[nfds] = hdl;
nfds++;
}
}
}
if (!nfds)
return -PAL_ERROR_TRYAGAIN;
int64_t waittime = timeout;
ret = ocall_poll(fds, nfds, timeout != NO_TIMEOUT ? &waittime : NULL);
if (IS_ERR(ret))
return unix_to_pal_error(ERRNO(ret));
if (!ret)
return -PAL_ERROR_TRYAGAIN;
PAL_HANDLE polled_hdl = NULL;
for (i = 0 ; i < nfds ; i++) {
if (!fds[i].revents)
continue;
PAL_HANDLE hdl = hdls[i];
if (polled_hdl) {
if (hdl != polled_hdl)
continue;
} else {
polled_hdl = hdl;
}
for (j = 0 ; j < MAX_FDS ; j++)
if ((HANDLE_HDR(hdl)->flags & (RFD(j)|WFD(j))) &&
hdl->generic.fds[j] == (PAL_IDX)fds[i].fd)
break;
if (j == MAX_FDS)
continue;
if (fds[i].revents & POLLOUT)
HANDLE_HDR(hdl)->flags |= WRITEABLE(j);
if (fds[i].revents & (POLLHUP|POLLERR))
HANDLE_HDR(hdl)->flags |= ERROR(j);
}
*polled = polled_hdl;
return polled_hdl ? 0 : -PAL_ERROR_TRYAGAIN;
}