/* signal handling */
static int
kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p)
{
struct kqop *kqop = base->evbase;
struct kevent kev;
struct timespec timeout = { 0, 0 };
(void)p;
EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
memset(&kev, 0, sizeof(kev));
kev.ident = nsignal;
kev.filter = EVFILT_SIGNAL;
kev.flags = EV_ADD;
/* Be ready for the signal if it is sent any
* time between now and the next call to
* kq_dispatch. */
if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
return (-1);
/* We can set the handler for most signals to SIG_IGN and
* still have them reported to us in the queue. However,
* if the handler for SIGCHLD is SIG_IGN, the system reaps
* zombie processes for us, and we don't get any notification.
* This appears to be the only signal with this quirk. */
if (evsig_set_handler_(base, nsignal,
nsignal == SIGCHLD ? SIG_DFL : SIG_IGN) == -1)
return (-1);
return (0);
}
static void
debug_lock_free(void *lock_, unsigned locktype)
{
struct debug_lock *lock = lock_;
EVUTIL_ASSERT(lock->count == 0);
EVUTIL_ASSERT(locktype == lock->locktype);
EVUTIL_ASSERT(DEBUG_LOCK_SIG == lock->signature);
if (original_lock_fns_.free) {
original_lock_fns_.free(lock->lock,
lock->locktype|EVTHREAD_LOCKTYPE_RECURSIVE);
}
lock->lock = NULL;
lock->count = -100;
lock->signature = 0x12300fda;
mm_free(lock);
}
static int
debug_lock_lock(unsigned mode, void *lock_)
{
struct debug_lock *lock = lock_;
int res = 0;
if (lock->locktype & EVTHREAD_LOCKTYPE_READWRITE)
EVUTIL_ASSERT(mode & (EVTHREAD_READ|EVTHREAD_WRITE));
else
EVUTIL_ASSERT((mode & (EVTHREAD_READ|EVTHREAD_WRITE)) == 0);
if (original_lock_fns_.lock)
res = original_lock_fns_.lock(mode, lock->lock);
if (!res) {
evthread_debug_lock_mark_locked(mode, lock);
}
return res;
}
/* Called to write data info the BIO */
static int
bio_bufferevent_write(BIO *b, const char *in, int inlen)
{
struct bufferevent *bufev = b->ptr;
struct evbuffer *output;
size_t outlen;
BIO_clear_retry_flags(b);
if (!b->ptr)
return -1;
output = bufferevent_get_output(bufev);
outlen = evbuffer_get_length(output);
/* Copy only as much data onto the output buffer as can fit under the
* high-water mark. */
if (bufev->wm_write.high && bufev->wm_write.high <= (outlen+inlen)) {
if (bufev->wm_write.high <= outlen) {
/* If no data can fit, we'll need to retry later. */
BIO_set_retry_write(b);
return -1;
}
inlen = bufev->wm_write.high - outlen;
}
EVUTIL_ASSERT(inlen > 0);
evbuffer_add(output, in, inlen);
return inlen;
}
static int
kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p)
{
struct kqop *kqop = base->evbase;
struct kevent kev;
struct timespec timeout = { 0, 0 };
(void)p;
EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
memset(&kev, 0, sizeof(kev));
kev.ident = nsignal;
kev.filter = EVFILT_SIGNAL;
kev.flags = EV_DELETE;
/* Because we insert signal events
* immediately, we need to delete them
* immediately, too */
if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
return (-1);
if (_evsig_restore_handler(base, nsignal) == -1)
return (-1);
return (0);
}
static void
connect_complete(struct event_overlapped *eo, ev_uintptr_t key,
ev_ssize_t nbytes, int ok)
{
struct bufferevent_async *bev_a = upcast_connect(eo);
struct bufferevent *bev = &bev_a->bev.bev;
evutil_socket_t sock;
BEV_LOCK(bev);
EVUTIL_ASSERT(bev_a->bev.connecting);
bev_a->bev.connecting = 0;
sock = _evbuffer_overlapped_get_fd(bev_a->bev.bev.input);
/* XXXX Handle error? */
setsockopt(sock, SOL_SOCKET, SO_UPDATE_CONNECT_CONTEXT, NULL, 0);
if (ok)
bufferevent_async_set_connected(bev);
else
bev_async_set_wsa_error(bev, eo);
_bufferevent_run_eventcb(bev,
ok? BEV_EVENT_CONNECTED : BEV_EVENT_ERROR);
event_base_del_virtual(bev->ev_base);
_bufferevent_decref_and_unlock(bev);
}
/* signal handling */
static int
kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p)
{
struct kqop *kqop = base->evbase;
struct kevent kev;
struct timespec timeout = { 0, 0 };
(void)p;
EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG);
memset(&kev, 0, sizeof(kev));
kev.ident = nsignal;
kev.filter = EVFILT_SIGNAL;
kev.flags = EV_ADD;
/* Be ready for the signal if it is sent any
* time between now and the next call to
* kq_dispatch. */
if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1)
return (-1);
/* XXXX The manpage suggest we could use SIG_IGN instead of a
* do-nothing handler */
if (_evsig_set_handler(base, nsignal, kq_sighandler) == -1)
return (-1);
return (0);
}
static void
poll_check_ok(struct pollop *pop)
{
int i, idx;
struct event *ev;
for (i = 0; i < pop->fd_count; ++i) {
idx = pop->idxplus1_by_fd[i]-1;
if (idx < 0)
continue;
EVUTIL_ASSERT(pop->event_set[idx].fd == i);
}
for (i = 0; i < pop->nfds; ++i) {
struct pollfd *pfd = &pop->event_set[i];
EVUTIL_ASSERT(pop->idxplus1_by_fd[pfd->fd] == i+1);
}
}
static inline struct bufferevent_async *
upcast_write(struct event_overlapped *eo)
{
struct bufferevent_async *bev_a;
bev_a = EVUTIL_UPCAST(eo, struct bufferevent_async, write_overlapped);
EVUTIL_ASSERT(BEV_IS_ASYNC(&bev_a->bev.bev));
return bev_a;
}
static inline struct bufferevent_async *
upcast_overlapped(struct event_overlapped *eo)
{
struct bufferevent_async *bev_a;
bev_a = EVUTIL_UPCAST(eo, struct bufferevent_async, connect_overlapped);
EVUTIL_ASSERT(bev_a->bev.bev.be_ops == &bufferevent_ops_async);
return bev_a;
}
static void
listener_read_cb(evutil_socket_t fd, short what, void *p)
{
struct evconnlistener *lev = p;
int err;
evconnlistener_cb cb;
evconnlistener_errorcb errorcb;
void *user_data;
LOCK(lev);
while (1) {
struct sockaddr_storage ss;
#ifdef WIN32
int socklen = sizeof(ss);
#else
socklen_t socklen = sizeof(ss);
#endif
evutil_socket_t new_fd = accept(fd, (struct sockaddr*)&ss, &socklen);
if (new_fd < 0)
break;
if (!(lev->flags & LEV_OPT_LEAVE_SOCKETS_BLOCKING))
evutil_make_socket_nonblocking(new_fd);
if (lev->cb == NULL) {
UNLOCK(lev);
return;
}
++lev->refcnt;
cb = lev->cb;
user_data = lev->user_data;
UNLOCK(lev);
cb(lev, new_fd, (struct sockaddr*)&ss, (int)socklen,
user_data);
LOCK(lev);
if (lev->refcnt == 1) {
int freed = listener_decref_and_unlock(lev);
EVUTIL_ASSERT(freed);
return;
}
--lev->refcnt;
}
err = evutil_socket_geterror(fd);
if (EVUTIL_ERR_ACCEPT_RETRIABLE(err)) {
UNLOCK(lev);
return;
}
if (lev->errorcb != NULL) {
++lev->refcnt;
errorcb = lev->errorcb;
user_data = lev->user_data;
UNLOCK(lev);
errorcb(lev, user_data);
LOCK(lev);
listener_decref_and_unlock(lev);
} else {
event_sock_warn(fd, "Error from accept() call");
}
}
static void
listener_read_cb(evutil_socket_t fd, short what, void *p)
{
struct evconnlistener *lev = (struct evconnlistener *)p;
int err;
evconnlistener_cb cb;
evconnlistener_errorcb errorcb;
void *user_data;
LOCK(lev);
while (1) {
struct sockaddr_storage ss;
ev_socklen_t socklen = sizeof(ss);
evutil_socket_t new_fd = evutil_accept4_(fd, (struct sockaddr*)&ss, &socklen, lev->accept4_flags);
if (new_fd < 0)
break;
if (socklen == 0) {
/* This can happen with some older linux kernels in
* response to nmap. */
evutil_closesocket(new_fd);
continue;
}
if (lev->cb == NULL) {
evutil_closesocket(new_fd);
UNLOCK(lev);
return;
}
++lev->refcnt;
cb = lev->cb;
user_data = lev->user_data;
UNLOCK(lev);
cb(lev, new_fd, (struct sockaddr*)&ss, (int)socklen,
user_data);
LOCK(lev);
if (lev->refcnt == 1) {
int freed = listener_decref_and_unlock(lev);
EVUTIL_ASSERT(freed);
return;
}
--lev->refcnt;
}
err = evutil_socket_geterror(fd);
if (EVUTIL_ERR_ACCEPT_RETRIABLE(err)) {
UNLOCK(lev);
return;
}
if (lev->errorcb != NULL) {
++lev->refcnt;
errorcb = lev->errorcb;
user_data = lev->user_data;
UNLOCK(lev);
errorcb(lev, user_data);
LOCK(lev);
listener_decref_and_unlock(lev);
} else {
event_sock_warn(fd, "Error from accept() call");
}
}
static int
realloc_fd_sets(struct win32op *op, size_t new_size)
{
size_t size;
EVUTIL_ASSERT(new_size >= op->readset_in->fd_count &&
new_size >= op->writeset_in->fd_count);
EVUTIL_ASSERT(new_size >= 1);
size = FD_SET_ALLOC_SIZE(new_size);
if (!(op->readset_in = mm_realloc(op->readset_in, size)))
return (-1);
if (!(op->writeset_in = mm_realloc(op->writeset_in, size)))
return (-1);
op->resize_out_sets = 1;
op->fd_setsz = new_size;
return (0);
}
static int
evsig_del(struct event_base *base, int evsignal, short old, short events, void *p)
{
EVUTIL_ASSERT(evsignal >= 0 && evsignal < NSIG);
event_debug(("%s: %d: restoring signal handler", __func__, evsignal));
return (_evsig_restore_handler(base, evsignal));
}
static void
evthread_debug_lock_mark_unlocked(unsigned mode, struct debug_lock *lock)
{
EVUTIL_ASSERT(DEBUG_LOCK_SIG == lock->signature);
if (lock->locktype & EVTHREAD_LOCKTYPE_READWRITE)
EVUTIL_ASSERT(mode & (EVTHREAD_READ|EVTHREAD_WRITE));
else
EVUTIL_ASSERT((mode & (EVTHREAD_READ|EVTHREAD_WRITE)) == 0);
if (evthread_id_fn_) {
unsigned long me;
me = evthread_id_fn_();
EVUTIL_ASSERT(lock->held_by == me);
if (lock->count == 1)
lock->held_by = 0;
}
--lock->count;
EVUTIL_ASSERT(lock->count >= 0);
}
static int
grow_fd_sets(struct win32op *op, unsigned new_num_fds)
{
size_t size;
EVUTIL_ASSERT(new_num_fds >= op->readset_in->fd_count &&
new_num_fds >= op->writeset_in->fd_count);
EVUTIL_ASSERT(new_num_fds >= 1);
size = FD_SET_ALLOC_SIZE(new_num_fds);
if (!(op->readset_in = mm_realloc(op->readset_in, size)))
return (-1);
if (!(op->writeset_in = mm_realloc(op->writeset_in, size)))
return (-1);
op->resize_out_sets = 1;
op->num_fds_in_fd_sets = new_num_fds;
return (0);
}
static int
poll_add(struct event_base *base, int fd, short old, short events, void *_idx)
{
struct pollop *pop = base->evbase;
struct pollfd *pfd = NULL;
struct pollidx *idx = _idx;
int i;
EVUTIL_ASSERT((events & EV_SIGNAL) == 0);
if (!(events & (EV_READ|EV_WRITE)))
return (0);
poll_check_ok(pop);
if (pop->nfds + 1 >= pop->event_count) {
struct pollfd *tmp_event_set;
int tmp_event_count;
if (pop->event_count < 32)
tmp_event_count = 32;
else
tmp_event_count = pop->event_count * 2;
/* We need more file descriptors */
tmp_event_set = mm_realloc(pop->event_set,
tmp_event_count * sizeof(struct pollfd));
if (tmp_event_set == NULL) {
event_warn("realloc");
return (-1);
}
pop->event_set = tmp_event_set;
pop->event_count = tmp_event_count;
pop->realloc_copy = 1;
}
i = idx->idxplus1 - 1;
if (i >= 0) {
pfd = &pop->event_set[i];
} else {
i = pop->nfds++;
pfd = &pop->event_set[i];
pfd->events = 0;
pfd->fd = fd;
idx->idxplus1 = i + 1;
}
pfd->revents = 0;
if (events & EV_WRITE)
pfd->events |= POLLOUT;
if (events & EV_READ)
pfd->events |= POLLIN;
poll_check_ok(pop);
return (0);
}
static inline struct bufferevent_async *
upcast(struct bufferevent *bev)
{
struct bufferevent_async *bev_a;
if (bev->be_ops != &bufferevent_ops_async)
return NULL;
bev_a = EVUTIL_UPCAST(bev, struct bufferevent_async, bev.bev);
EVUTIL_ASSERT(bev_a->bev.bev.be_ops == &bufferevent_ops_async);
return bev_a;
}
/*
* Verifies very basic integrity of a given port_event.
*/
static void
check_event(port_event_t* pevt)
{
/*
* We've only registered for PORT_SOURCE_FD events. The only
* other thing we can legitimately receive is PORT_SOURCE_ALERT,
* but since we're not using port_alert either, we can assume
* PORT_SOURCE_FD.
*/
EVUTIL_ASSERT(pevt->portev_source == PORT_SOURCE_FD);
}
void evbuffer_commit_write_(struct evbuffer *evbuf, ev_ssize_t nBytes)
{
struct evbuffer_overlapped *buf = upcast_evbuffer(evbuf);
EVUTIL_ASSERT(buf->write_in_progress && !buf->read_in_progress);
evbuffer_unfreeze(evbuf, 1);
evbuffer_drain(evbuf, nBytes);
pin_release(buf,EVBUFFER_MEM_PINNED_W);
buf->write_in_progress = 0;
evbuffer_decref(evbuf);
}
static int
debug_cond_wait(void *_cond, void *_lock, const struct timeval *tv)
{
int r;
struct debug_lock *lock = _lock;
EVUTIL_ASSERT(lock);
EVLOCK_ASSERT_LOCKED(_lock);
evthread_debug_lock_mark_unlocked(0, lock);
r = _original_cond_fns.wait_condition(_cond, lock->lock, tv);
evthread_debug_lock_mark_locked(0, lock);
return r;
}
/** Unpin all the chains noted as pinned in 'eo'. */
static void pin_release(struct evbuffer_overlapped *eo, unsigned flag)
{
int i;
struct evbuffer_chain *next, *chain = eo->first_pinned;
for (i = 0; i < eo->n_buffers; ++i) {
EVUTIL_ASSERT(chain);
next = chain->next;
evbuffer_chain_unpin_(chain, flag);
chain = next;
}
}
void
evbuffer_commit_read(struct evbuffer *evbuf, ev_ssize_t nBytes)
{
struct evbuffer_overlapped *buf = upcast_evbuffer(evbuf);
struct evbuffer_chain **chainp;
size_t remaining, len;
unsigned i;
EVBUFFER_LOCK(evbuf);
EVUTIL_ASSERT(buf->read_in_progress && !buf->write_in_progress);
EVUTIL_ASSERT(nBytes >= 0); /* XXXX Can this be false? */
evbuffer_unfreeze(evbuf, 0);
chainp = evbuf->last_with_datap;
if (!((*chainp)->flags & EVBUFFER_MEM_PINNED_R))
chainp = &(*chainp)->next;
remaining = nBytes;
for (i = 0; remaining > 0 && i < (unsigned)buf->n_buffers; ++i) {
EVUTIL_ASSERT(*chainp);
len = buf->buffers[i].len;
if (remaining < len)
len = remaining;
(*chainp)->off += len;
evbuf->last_with_datap = chainp;
remaining -= len;
chainp = &(*chainp)->next;
}
pin_release(buf, EVBUFFER_MEM_PINNED_R);
buf->read_in_progress = 0;
evbuf->total_len += nBytes;
evbuf->n_add_for_cb += nBytes;
evbuffer_invoke_callbacks(evbuf);
_evbuffer_decref_and_unlock(evbuf);
}
static void
be_socket_destruct(struct bufferevent *bufev)
{
struct bufferevent_private *bufev_p =
EVUTIL_UPCAST(bufev, struct bufferevent_private, bev);
evutil_socket_t fd;
EVUTIL_ASSERT(bufev->be_ops == &bufferevent_ops_socket);
fd = event_get_fd(&bufev->ev_read);
if ((bufev_p->options & BEV_OPT_CLOSE_ON_FREE) && fd >= 0)
EVUTIL_CLOSESOCKET(fd);
}
static int
evsig_del(struct event_base *base, evutil_socket_t evsignal, short old, short events, void *p)
{
EVUTIL_ASSERT(evsignal >= 0 && evsignal < NSIG);
event_debug(("%s: %d: restoring signal handler", __func__, evsignal));
EVSIGBASE_LOCK();
--evsig_base_n_signals_added;
--base->sig.ev_n_signals_added;
EVSIGBASE_UNLOCK();
return (evsig_restore_handler_(base, (int)evsignal));
}
/* strsep replacement for platforms that lack it. Only works if
* del is one character long. */
static char *
strsep(char **s, const char *del)
{
char *d, *tok;
EVUTIL_ASSERT(strlen(del) == 1);
if (!s || !*s)
return NULL;
tok = *s;
d = strstr(tok, del);
if (d) {
*d = '\0';
*s = d + 1;
} else
*s = NULL;
return tok;
}
static void
be_async_destruct(struct bufferevent *bev)
{
struct bufferevent_private *bev_p = BEV_UPCAST(bev);
evutil_socket_t fd;
EVUTIL_ASSERT(!upcast(bev)->write_in_progress && !upcast(bev)->read_in_progress);
/* XXX cancel any outstanding I/O operations */
fd = _evbuffer_overlapped_get_fd(bev->input);
/* delete this in case non-blocking connect was used */
event_del(&bev->ev_write);
if (bev_p->options & BEV_OPT_CLOSE_ON_FREE)
EVUTIL_CLOSESOCKET(fd);
_bufferevent_del_generic_timeout_cbs(bev);
}
void *
evthread_setup_global_lock_(void *lock_, unsigned locktype, int enable_locks)
{
/* there are four cases here:
1) we're turning on debugging; locking is not on.
2) we're turning on debugging; locking is on.
3) we're turning on locking; debugging is not on.
4) we're turning on locking; debugging is on. */
if (!enable_locks && _original_lock_fns.alloc == NULL) {
/* Case 1: allocate a debug lock. */
EVUTIL_ASSERT(lock_ == NULL);
return debug_lock_alloc(locktype);
} else if (!enable_locks && _original_lock_fns.alloc != NULL) {
/* Case 2: wrap the lock in a debug lock. */
struct debug_lock *lock;
EVUTIL_ASSERT(lock_ != NULL);
if (!(locktype & EVTHREAD_LOCKTYPE_RECURSIVE)) {
/* We can't wrap it: We need a recursive lock */
_original_lock_fns.free(lock_, locktype);
return debug_lock_alloc(locktype);
}
lock = mm_malloc(sizeof(struct debug_lock));
if (!lock) {
_original_lock_fns.free(lock_, locktype);
return NULL;
}
lock->lock = lock_;
lock->locktype = locktype;
lock->count = 0;
lock->held_by = 0;
return lock;
} else if (enable_locks && ! ompi__evthread_lock_debugging_enabled) {
/* Case 3: allocate a regular lock */
EVUTIL_ASSERT(lock_ == NULL);
return ompi__evthread_lock_fns.alloc(locktype);
} else {
/* Case 4: Fill in a debug lock with a real lock */
struct debug_lock *lock = lock_;
EVUTIL_ASSERT(enable_locks &&
ompi__evthread_lock_debugging_enabled);
EVUTIL_ASSERT(lock->locktype == locktype);
EVUTIL_ASSERT(lock->lock == NULL);
lock->lock = _original_lock_fns.alloc(
locktype|EVTHREAD_LOCKTYPE_RECURSIVE);
if (!lock->lock) {
lock->count = -200;
mm_free(lock);
return NULL;
}
return lock;
}
}
static void
connect_complete(struct event_overlapped *eo, uintptr_t key,
ev_ssize_t nbytes, int ok)
{
struct bufferevent_async *bev_a = upcast_overlapped(eo);
struct bufferevent *bev = &bev_a->bev.bev; /* XXX locking issue ? */
_bufferevent_incref_and_lock(bev);
EVUTIL_ASSERT(bev_a->bev.connecting);
bev_a->bev.connecting = 0;
_bufferevent_run_eventcb(bev,
ok? BEV_EVENT_CONNECTED : BEV_EVENT_ERROR);
_bufferevent_decref_and_unlock(bev);
}
int
bufferevent_async_connect(struct bufferevent *bev, evutil_socket_t fd,
const struct sockaddr *sa, int socklen)
{
BOOL rc;
struct bufferevent_async *bev_async = upcast(bev);
struct sockaddr_storage ss;
const struct win32_extension_fns *ext =
event_get_win32_extension_fns();
EVUTIL_ASSERT(ext && ext->ConnectEx && fd >= 0 && sa != NULL);
/* ConnectEx() requires that the socket be bound to an address
* with bind() before using, otherwise it will fail. We attempt
* to issue a bind() here, taking into account that the error
* code is set to WSAEINVAL when the socket is already bound. */
memset(&ss, 0, sizeof(ss));
if (sa->sa_family == AF_INET) {
struct sockaddr_in *sin = (struct sockaddr_in *)&ss;
sin->sin_family = AF_INET;
sin->sin_addr.s_addr = INADDR_ANY;
} else if (sa->sa_family == AF_INET6) {
struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&ss;
sin6->sin6_family = AF_INET6;
sin6->sin6_addr = in6addr_any;
} else {
/* Well, the user will have to bind() */
return -1;
}
if (bind(fd, (struct sockaddr *)&ss, sizeof(ss)) < 0 &&
WSAGetLastError() != WSAEINVAL)
return -1;
event_base_add_virtual(bev->ev_base);
bufferevent_incref(bev);
rc = ext->ConnectEx(fd, sa, socklen, NULL, 0, NULL,
&bev_async->connect_overlapped.overlapped);
if (rc || WSAGetLastError() == ERROR_IO_PENDING)
return 0;
event_base_del_virtual(bev->ev_base);
bufferevent_decref(bev);
return -1;
}
