#include "portab.h"
-static char UNUSED id[] = "$Id: io.c,v 1.15 2006/07/12 19:27:12 fw Exp $";
+static char UNUSED id[] = "$Id: io.c,v 1.21 2006/09/19 18:21:30 fw Exp $";
#include <assert.h>
#include <stdlib.h>
# ifdef HAVE_KQUEUE
#define IO_USE_KQUEUE 1
# else
+# ifdef HAVE_SYS_DEVPOLL_H
+#define IO_USE_DEVPOLL 1
+# else
+# ifdef HAVE_POLL
+#define IO_USE_POLL 1
+# else
#define IO_USE_SELECT 1
-#endif
-#endif
+# endif /* HAVE_POLL */
+# endif /* HAVE_SYS_DEVPOLL_H */
+# endif /* HAVE_KQUEUE */
+#endif /* HAVE_EPOLL_CREATE */
static bool library_initialized;
static bool io_event_change_kqueue(int, short, const int action);
#endif
+#ifdef IO_USE_POLL
+#include <poll.h>
+
+static array pollfds;
+static int poll_maxfd;
+
+static bool io_event_change_poll(int fd, short what);
+#endif
+
+#ifdef IO_USE_DEVPOLL
+#include <sys/devpoll.h>
+static int io_masterfd;
+
+static bool io_event_change_devpoll(int fd, short what);
+#endif
+
#ifdef IO_USE_SELECT
#include "defines.h" /* for conn.h */
#include "conn.h" /* for CONN_IDX (needed by resolve.h) */
}
-bool
-io_library_init(unsigned int eventsize)
+#ifdef IO_USE_DEVPOLL
+static void
+io_library_init_devpoll(unsigned int eventsize)
{
- bool ret;
-#ifdef IO_USE_EPOLL
- int ecreate_hint = (int)eventsize;
- if (ecreate_hint <= 0)
- ecreate_hint = 128;
+ io_masterfd = open("/dev/poll", O_RDWR);
+ if (io_masterfd >= 0)
+ library_initialized = true;
+ Log(LOG_INFO, "IO subsystem: /dev/poll (initial maxfd %u, masterfd %d).",
+ eventsize, io_masterfd);
+}
#endif
- if (library_initialized)
- return true;
-#ifdef IO_USE_SELECT
-#ifdef FD_SETSIZE
- if (eventsize >= FD_SETSIZE)
- eventsize = FD_SETSIZE - 1;
-#endif
+
+#ifdef IO_USE_POLL
+static void
+io_library_init_poll(unsigned int eventsize)
+{
+ struct pollfd *p;
+ array_init(&pollfds);
+ poll_maxfd = 0;
+ Log(LOG_INFO, "IO subsystem: poll (initial maxfd %u).",
+ eventsize);
+ p = array_alloc(&pollfds, sizeof(struct pollfd), eventsize);
+ if (p) {
+ unsigned i;
+ p = array_start(&pollfds);
+ for (i = 0; i < eventsize; i++)
+ p[i].fd = -1;
+
+ library_initialized = true;
+ }
+}
#endif
- if ((eventsize > 0) && !array_alloc(&io_events, sizeof(io_event), (size_t)eventsize))
- eventsize = 0;
-#ifdef IO_USE_EPOLL
- io_masterfd = epoll_create(ecreate_hint);
- Log(LOG_INFO,
- "IO subsystem: epoll (hint size %d, initial maxfd %u, masterfd %d).",
- ecreate_hint, eventsize, io_masterfd);
- ret = io_masterfd >= 0;
- if (ret) library_initialized = true;
- return ret;
-#endif
#ifdef IO_USE_SELECT
+static void
+io_library_init_select(unsigned int eventsize)
+{
Log(LOG_INFO, "IO subsystem: select (initial maxfd %u).",
eventsize);
FD_ZERO(&readers);
Conf_MaxConnections = FD_SETSIZE - 1;
}
-#else
- Log(LOG_WARNING,
- "FD_SETSIZE undefined, don't know how many descriptors select() can handle on your platform ...");
#endif /* FD_SETSIZE */
library_initialized = true;
- return true;
+}
#endif /* SELECT */
+
+
+#ifdef IO_USE_EPOLL
+static void
+io_library_init_epoll(unsigned int eventsize)
+{
+ int ecreate_hint = (int)eventsize;
+ if (ecreate_hint <= 0)
+ ecreate_hint = 128;
+ io_masterfd = epoll_create(ecreate_hint);
+ Log(LOG_INFO,
+ "IO subsystem: epoll (hint size %d, initial maxfd %u, masterfd %d).",
+ ecreate_hint, eventsize, io_masterfd);
+ if (io_masterfd >= 0)
+ library_initialized = true;
+}
+#endif
+
+
#ifdef IO_USE_KQUEUE
+static void
+io_library_init_kqueue(unsigned int eventsize)
+{
io_masterfd = kqueue();
Log(LOG_INFO,
"IO subsystem: kqueue (initial maxfd %u, masterfd %d)",
eventsize, io_masterfd);
- ret = io_masterfd >= 0;
- if (ret) library_initialized = true;
+ if (io_masterfd >= 0)
+ library_initialized = true;
+}
+#endif
- return ret;
+
+bool
+io_library_init(unsigned int eventsize)
+{
+ if (library_initialized)
+ return true;
+#ifdef IO_USE_SELECT
+#ifndef FD_SETSIZE
+ Log(LOG_WARNING,
+ "FD_SETSIZE undefined, don't know how many descriptors select() can handle on your platform ...");
+#else
+ if (eventsize >= FD_SETSIZE)
+ eventsize = FD_SETSIZE - 1;
+#endif /* FD_SETSIZE */
+#endif /* IO_USE_SELECT */
+ if ((eventsize > 0) && !array_alloc(&io_events, sizeof(io_event), (size_t)eventsize))
+ eventsize = 0;
+#ifdef IO_USE_EPOLL
+ io_library_init_epoll(eventsize);
#endif
+#ifdef IO_USE_KQUEUE
+ io_library_init_kqueue(eventsize);
+#endif
+#ifdef IO_USE_DEVPOLL
+ io_library_init_devpoll(eventsize);
+#endif
+#ifdef IO_USE_POLL
+ io_library_init_poll(eventsize);
+#endif
+#ifdef IO_USE_SELECT
+ io_library_init_select(eventsize);
+#endif
+ return library_initialized;
}
#ifdef IO_USE_SELECT
FD_ZERO(&readers);
FD_ZERO(&writers);
-#else
- close(io_masterfd); /* kqueue, epoll */
+#endif
+#ifdef IO_USE_EPOLL
+ close(io_masterfd);
io_masterfd = -1;
#endif
#ifdef IO_USE_KQUEUE
+ close(io_masterfd);
+ io_masterfd = -1;
array_free(&io_evcache);
#endif
library_initialized = false;
io_event *i;
assert(fd >= 0);
-
-#ifdef IO_USE_SELECT
-#ifdef FD_SETSIZE
+#if defined(IO_USE_SELECT) || defined(FD_SETSIZE)
if (fd >= FD_SETSIZE) {
Log(LOG_ERR,
"fd %d exceeds FD_SETSIZE (%u) (select can't handle more file descriptors)",
fd, FD_SETSIZE);
return false;
}
-#endif /* FD_SETSIZE */
-#endif /* IO_USE_SELECT */
-
+#endif
i = (io_event *) array_alloc(&io_events, sizeof(io_event), (size_t) fd);
if (!i) {
Log(LOG_WARNING,
i->callback = cbfunc;
i->what = 0;
+#ifdef IO_USE_DEVPOLL
+ ret = io_event_change_devpoll(fd, what);
+#endif
+#ifdef IO_USE_POLL
+ ret = io_event_change_poll(fd, what);
+#endif
#ifdef IO_USE_EPOLL
ret = io_event_change_epoll(fd, what, EPOLL_CTL_ADD);
#endif
}
+#ifdef IO_USE_DEVPOLL
+static bool
+io_event_change_devpoll(int fd, short what)
+{
+ struct pollfd p;
+
+ p.events = 0;
+
+ if (what & IO_WANTREAD)
+ p.events = POLLIN | POLLPRI;
+ if (what & IO_WANTWRITE)
+ p.events |= POLLOUT;
+
+ p.fd = fd;
+ return write(io_masterfd, &p, sizeof p) == (ssize_t)sizeof p;
+}
+#endif
+
+
+
+#ifdef IO_USE_POLL
+static bool
+io_event_change_poll(int fd, short what)
+{
+ struct pollfd *p;
+ short events = 0;
+
+ if (what & IO_WANTREAD)
+ events = POLLIN | POLLPRI;
+ if (what & IO_WANTWRITE)
+ events |= POLLOUT;
+
+ p = array_alloc(&pollfds, sizeof *p, fd);
+ if (p) {
+ p->events = events;
+ p->fd = fd;
+ if (fd > poll_maxfd)
+ poll_maxfd = fd;
+ }
+ return p != NULL;
+}
+#endif
+
#ifdef IO_USE_EPOLL
static bool
io_event_change_epoll(int fd, short what, const int action)
struct kevent *events;
bool ret;
int len = (int) array_length(&io_evcache, sizeof (struct kevent));
-
+
if (!len) /* nothing to do */
return true;
ret = array_catb(&io_evcache, (char*) &kev, sizeof (kev));
if (!ret)
ret = kevent(io_masterfd, &kev,1, NULL, 0, NULL) == 0;
- }
+ }
if (ret && (what & IO_WANTWRITE)) {
EV_SET(&kev, fd, EVFILT_WRITE, action, 0, 0, 0);
#ifdef IO_USE_KQUEUE
return io_event_change_kqueue(fd, what, EV_ADD | EV_ENABLE);
#endif
-
+#ifdef IO_USE_DEVPOLL
+ return io_event_change_devpoll(fd, i->what);
+#endif
+#ifdef IO_USE_POLL
+ return io_event_change_poll(fd, i->what);
+#endif
#ifdef IO_USE_SELECT
if (fd > select_maxfd)
select_maxfd = fd;
}
-bool
-io_close(int fd)
+#ifdef IO_USE_DEVPOLL
+static void
+io_close_devpoll(int fd)
{
- io_event *i;
+ struct pollfd p;
+ p.events = POLLREMOVE;
+ p.fd = fd;
+ write(io_masterfd, &p, sizeof p);
+}
+#else
+static inline void io_close_devpoll(int UNUSED x) { /* NOTHING */ }
+#endif
+
+
+
+#ifdef IO_USE_POLL
+static void
+io_close_poll(int fd)
+{
+ struct pollfd *p;
+ p = array_get(&pollfds, sizeof *p, fd);
+ if (!p) return;
+
+ p->fd = -1;
+ if (fd == poll_maxfd) {
+ while (poll_maxfd > 0) {
+ --poll_maxfd;
+ p = array_get(&pollfds, sizeof *p, poll_maxfd);
+ if (p && p->fd >= 0)
+ break;
+ }
+ }
+}
+#else
+static inline void io_close_poll(int UNUSED x) { /* NOTHING */ }
+#endif
+
+
#ifdef IO_USE_SELECT
+static void
+io_close_select(int fd)
+{
+ io_event *i;
FD_CLR(fd, &writers);
FD_CLR(fd, &readers);
+ i = io_event_get(fd);
+ if (!i) return;
+
if (fd == select_maxfd) {
while (select_maxfd>0) {
- --select_maxfd; /* find largest fd */
+ --select_maxfd; /* find largest fd */
i = io_event_get(select_maxfd);
if (i && i->callback) break;
- }
- }
+ }
+ }
+}
+#else
+static inline void io_close_select(int UNUSED x) { /* NOTHING */ }
#endif
+
+
+bool
+io_close(int fd)
+{
+ io_event *i;
+
i = io_event_get(fd);
#ifdef IO_USE_KQUEUE
if (array_length(&io_evcache, sizeof (struct kevent))) /* pending data in cache? */
if (i) {
io_event_change_kqueue(fd, i->what, EV_DELETE);
io_event_kqueue_commit_cache();
- }
+ }
#endif
+
+ io_close_devpoll(fd);
+ io_close_poll(fd);
+ io_close_select(fd);
+
#ifdef IO_USE_EPOLL
io_event_change_epoll(fd, 0, EPOLL_CTL_DEL);
#endif
i->what &= ~what;
+#ifdef IO_USE_DEVPOLL
+ return io_event_change_devpoll(fd, i->what);
+#endif
+#ifdef IO_USE_POLL
+ return io_event_change_poll(fd, i->what);
+#endif
#ifdef IO_USE_EPOLL
return io_event_change_epoll(fd, i->what, EPOLL_CTL_MOD);
#endif
#endif
+#ifdef IO_USE_DEVPOLL
+static int
+io_dispatch_devpoll(struct timeval *tv)
+{
+ struct dvpoll dvp;
+ time_t sec = tv->tv_sec * 1000;
+ int i, total, ret, timeout = tv->tv_usec + sec;
+ short what;
+ struct pollfd p[100];
+
+ if (timeout < 0)
+ timeout = 1000;
+
+ total = 0;
+ do {
+ dvp.dp_timeout = timeout;
+ dvp.dp_nfds = 100;
+ dvp.dp_fds = p;
+ ret = ioctl(io_masterfd, DP_POLL, &dvp);
+ total += ret;
+ if (ret <= 0)
+ return total;
+ for (i=0; i < ret ; i++) {
+ what = 0;
+ if (p[i].revents & (POLLIN|POLLPRI))
+ what = IO_WANTREAD;
+
+ if (p[i].revents & POLLOUT)
+ what |= IO_WANTWRITE;
+
+ if (p[i].revents && !what) {
+ /* other flag is set, probably POLLERR */
+ what = IO_ERROR;
+ }
+ io_docallback(p[i].fd, what);
+ }
+ } while (ret == 100);
+
+ return total;
+}
+#endif
+
+
+#ifdef IO_USE_POLL
+static int
+io_dispatch_poll(struct timeval *tv)
+{
+ time_t sec = tv->tv_sec * 1000;
+ int i, ret, timeout = tv->tv_usec + sec;
+ int fds_ready;
+ short what;
+ struct pollfd *p = array_start(&pollfds);
+
+ if (timeout < 0)
+ timeout = 1000;
+
+ ret = poll(p, poll_maxfd + 1, timeout);
+ if (ret <= 0)
+ return ret;
+
+ fds_ready = ret;
+ for (i=0; i <= poll_maxfd; i++) {
+ what = 0;
+ if (p[i].revents & (POLLIN|POLLPRI))
+ what = IO_WANTREAD;
+
+ if (p[i].revents & POLLOUT)
+ what |= IO_WANTWRITE;
+
+ if (p[i].revents && !what) {
+ /* other flag is set, probably POLLERR */
+ what = IO_ERROR;
+ }
+ if (what) {
+ fds_ready--;
+ io_docallback(i, what);
+ }
+ if (fds_ready <= 0)
+ break;
+ }
+
+ return ret;
+}
+#endif
+
+
#ifdef IO_USE_EPOLL
static int
io_dispatch_epoll(struct timeval *tv)
int newevents_len;
ts.tv_sec = tv->tv_sec;
ts.tv_nsec = tv->tv_usec * 1000;
-
+
do {
newevents_len = (int) array_length(&io_evcache, sizeof (struct kevent));
newevents = (newevents_len > 0) ? array_start(&io_evcache) : NULL;
#ifdef DEBUG
LogDebug("kev.flag has EV_EOF set, setting IO_ERROR",
kev[i].filter, kev[i].ident);
-#endif
+#endif
io_docallback((int)kev[i].ident, IO_ERROR);
continue;
- }
+ }
switch (kev[i].filter) {
case EVFILT_READ:
#ifdef IO_USE_KQUEUE
return io_dispatch_kqueue(tv);
#endif
+#ifdef IO_USE_DEVPOLL
+ return io_dispatch_devpoll(tv);
+#endif
+#ifdef IO_USE_POLL
+ return io_dispatch_poll(tv);
+#endif
#ifdef IO_USE_EPOLL
return io_dispatch_epoll(tv);
#endif
#endif
i = io_event_get(fd);
- if (i->callback) { /* callback might be NULL if a previous callback function
+ if (i->callback) { /* callback might be NULL if a previous callback function
called io_close on this fd */
i->callback(fd, (what & IO_ERROR) ? i->what : what);
- }
+ }
/* if error indicator is set, we return the event(s) that were registered */
}