On Sun, Jun 06, 2004 at 02:25:15PM -0700, Justin Erenkrantz wrote:
> --On Sunday, June 6, 2004 1:02 AM -0700 Paul Querna <[EMAIL PROTECTED]>
> wrote:
>
> >>Other than that, your patch looks sane. +1 on the concept... -- justin
> >
> >sweet.
>
> Just tested on FreeBSD 5.2.1 - looks fine. +1. I'll give others a
> day or so to look at it - if no one complains, I'll commit it.
>
> FWIW, I fixed up a bunch of style nits - see attached. -- justin
Excellent... Paul, is there a reason you test for kevent() in configure
rather than kqueue()? (if the latter is sufficient, simply adding kqueue
to an AC_CHECK_FUNCS somewhere is easier)
Here's another iteration with proper configure code for safe epoll
detection and more style nit fixes.
Ah, also pollset_create will need to register a cleanup now there is
something to clean up now I notice...
joe
? epoll2.diff
? epoll3.diff
? radish:
Index: CHANGES
===================================================================
RCS file: /home/cvs/apr/CHANGES,v
retrieving revision 1.471
diff -u -r1.471 CHANGES
--- CHANGES 1 Jun 2004 10:03:47 -0000 1.471
+++ CHANGES 7 Jun 2004 20:14:34 -0000
@@ -7,6 +7,8 @@
Changes with APR 1.0
+ *) Add support for KQueue and sys_epoll to apr_pollset. [Paul Querna]
+
*) Add apr_shm_remove() function for removing a named shared
memory segment. [Amit Athavale <amit_athavale persistent.co.in>]
Index: configure.in
===================================================================
RCS file: /home/cvs/apr/configure.in,v
retrieving revision 1.585
diff -u -r1.585 configure.in
--- configure.in 5 Jun 2004 11:52:43 -0000 1.585
+++ configure.in 7 Jun 2004 20:14:34 -0000
@@ -637,6 +637,25 @@
AC_CHECK_FUNCS(poll)
+# Checks for the FreeBSD KQueue and Linux epoll interfaces:
+AC_CHECK_FUNC(kevent,
+ [AC_DEFINE([HAVE_KQUEUE], 1, [Define if the KQueue interface is supported])])
+
+# epoll* may be available in libc but return ENOSYS on a pre-2.6 kernel.
+AC_CACHE_CHECK([for epoll support], [apr_cv_epoll],
+[AC_TRY_RUN([
+#include <sys/epoll.h>
+#include <unistd.h>
+
+int main()
+{
+ return epoll_create(5) == -1;
+}], [apr_cv_epoll=yes], [apr_cv_epoll=no], [apr_cv_epoll=no])])
+
+if test "$apr_cv_epoll" = "yes"; then
+ AC_DEFINE([HAVE_EPOLL], 1, [Define if the epoll interface is supported])
+fi
+
dnl ----------------------------- Checking for missing POSIX thread functions
AC_CHECK_FUNCS([getpwnam_r getpwuid_r getgrnam_r getgrgid_r])
Index: poll/unix/poll.c
===================================================================
RCS file: /home/cvs/apr/poll/unix/poll.c,v
retrieving revision 1.44
diff -u -r1.44 poll.c
--- poll/unix/poll.c 13 Feb 2004 09:38:33 -0000 1.44
+++ poll/unix/poll.c 7 Jun 2004 20:14:34 -0000
@@ -26,12 +26,77 @@
#include <sys/poll.h>
#endif
+#ifdef HAVE_KQUEUE
+#include <sys/types.h>
+#include <sys/event.h>
+#include <sys/time.h>
+#endif
+
+#ifdef HAVE_EPOLL
+#include <sys/epoll.h>
+#endif
#ifdef NETWARE
#define HAS_SOCKETS(dt) (dt == APR_POLL_SOCKET) ? 1 : 0
#define HAS_PIPES(dt) (dt == APR_POLL_FILE) ? 1 : 0
#endif
+#ifdef HAVE_KQUEUE
+static apr_int16_t get_kqueue_revent(apr_int16_t event, apr_int16_t flags)
+{
+ apr_int16_t rv = 0;
+
+ if (event & EVFILT_READ)
+ rv |= APR_POLLIN;
+ if (event & EVFILT_WRITE)
+ rv |= APR_POLLOUT;
+ if (flags & EV_ERROR || flags & EV_EOF)
+ rv |= APR_POLLERR;
+
+ return rv;
+}
+
+#endif
+
+#ifdef HAVE_EPOLL
+static apr_int16_t get_epoll_event(apr_int16_t event)
+{
+ apr_int16_t rv = 0;
+
+ if (event & APR_POLLIN)
+ rv |= EPOLLIN;
+ if (event & APR_POLLPRI)
+ rv |= EPOLLPRI;
+ if (event & APR_POLLOUT)
+ rv |= EPOLLOUT;
+ if (event & APR_POLLERR)
+ rv |= EPOLLERR;
+ if (event & APR_POLLHUP)
+ rv |= EPOLLHUP;
+ /* APR_POLLNVAL is not handled by epoll. */
+
+ return rv;
+}
+
+static apr_int16_t get_epoll_revent(apr_int16_t event)
+{
+ apr_int16_t rv = 0;
+
+ if (event & EPOLLIN)
+ rv |= APR_POLLIN;
+ if (event & EPOLLPRI)
+ rv |= APR_POLLPRI;
+ if (event & EPOLLOUT)
+ rv |= APR_POLLOUT;
+ if (event & EPOLLERR)
+ rv |= APR_POLLERR;
+ if (event & EPOLLHUP)
+ rv |= APR_POLLHUP;
+ /* APR_POLLNVAL is not handled by epoll. */
+
+ return rv;
+}
+#endif
#ifdef HAVE_POLL /* We can just use poll to do our socket polling. */
@@ -286,7 +351,14 @@
struct apr_pollset_t {
apr_uint32_t nelts;
apr_uint32_t nalloc;
-#ifdef HAVE_POLL
+#ifdef HAVE_KQUEUE
+ int kqueue_fd;
+ struct kevent kevent;
+ struct kevent *ke_set;
+#elif defined(HAVE_EPOLL)
+ int epoll_fd;
+ struct epoll_event *pollset;
+#elif defined(HAVE_POLL)
struct pollfd *pollset;
#else
fd_set readset, writeset, exceptset;
@@ -305,7 +377,7 @@
apr_pool_t *p,
apr_uint32_t flags)
{
-#if !defined(HAVE_POLL) && defined(FD_SETSIZE)
+#if !defined(HAVE_KQUEUE) && !defined(HAVE_EPOLL) && !defined(HAVE_POLL) &&
defined(FD_SETSIZE)
if (size > FD_SETSIZE) {
*pollset = NULL;
return APR_EINVAL;
@@ -314,7 +386,17 @@
*pollset = apr_palloc(p, sizeof(**pollset));
(*pollset)->nelts = 0;
(*pollset)->nalloc = size;
-#ifdef HAVE_POLL
+#ifdef HAVE_KQUEUE
+ (*pollset)->ke_set = (struct kevent*)apr_palloc(p, size * sizeof(struct
kevent));
+ memset((*pollset)->ke_set, 0, size * sizeof(struct kevent));
+ (*pollset)->kqueue_fd = kqueue();
+ if ((*pollset)->kqueue_fd == -1) {
+ return APR_ENOMEM;
+ }
+#elif defined(HAVE_EPOLL)
+ (*pollset)->epoll_fd = epoll_create(size);
+ (*pollset)->pollset = apr_palloc(p, size * sizeof(struct epoll_event));
+#elif defined(HAVE_POLL)
(*pollset)->pollset = apr_palloc(p, size * sizeof(struct pollfd));
#else
FD_ZERO(&((*pollset)->readset));
@@ -333,17 +415,26 @@
APR_DECLARE(apr_status_t) apr_pollset_destroy(apr_pollset_t *pollset)
{
- /* A no-op function for now. If we later implement /dev/poll
- * support, we'll need to close the /dev/poll fd here
- */
+#ifdef HAVE_KQUEUE
+ close(pollset->kqueue_fd);
+#elif defined(HAVE_EPOLL)
+ close(pollset->epoll_fd);
+#endif
return APR_SUCCESS;
}
APR_DECLARE(apr_status_t) apr_pollset_add(apr_pollset_t *pollset,
const apr_pollfd_t *descriptor)
{
-#ifndef HAVE_POLL
+#ifdef HAVE_KQUEUE
apr_os_sock_t fd;
+#elif defined(HAVE_EPOLL)
+ struct epoll_event ev;
+ int ret = -1;
+#else
+#if !defined(HAVE_POLL)
+ apr_os_sock_t fd;
+#endif
#endif
if (pollset->nelts == pollset->nalloc) {
@@ -351,7 +442,49 @@
}
pollset->query_set[pollset->nelts] = *descriptor;
-#ifdef HAVE_POLL
+
+#ifdef HAVE_KQUEUE
+ if (descriptor->desc_type == APR_POLL_SOCKET) {
+ fd = descriptor->desc.s->socketdes;
+ }
+ else {
+ fd = descriptor->desc.f->filedes;
+ }
+
+ if (descriptor->reqevents & APR_POLLIN) {
+ EV_SET(&pollset->kevent, fd, EVFILT_READ, EV_ADD, 0, 0, NULL);
+
+ if (kevent(pollset->kqueue_fd, &pollset->kevent, 1, NULL, 0,
+ NULL) == -1) {
+ return APR_ENOMEM;
+ }
+ }
+
+ if (descriptor->reqevents & APR_POLLOUT) {
+ EV_SET(&pollset->kevent, fd, EVFILT_WRITE, EV_ADD, 0, 0, NULL);
+
+ if (kevent(pollset->kqueue_fd, &pollset->kevent, 1, NULL, 0,
+ NULL) == -1) {
+ return APR_ENOMEM;
+ }
+ }
+
+#elif defined(HAVE_EPOLL)
+ ev.events = get_epoll_event(descriptor->reqevents);
+ if (descriptor->desc_type == APR_POLL_SOCKET) {
+ ev.data.fd = descriptor->desc.s->socketdes;
+ ret = epoll_ctl(pollset->epoll_fd, EPOLL_CTL_ADD,
+ descriptor->desc.s->socketdes, &ev);
+ }
+ else {
+ ev.data.fd = descriptor->desc.f->filedes;
+ ret = epoll_ctl(pollset->epoll_fd, EPOLL_CTL_ADD,
+ descriptor->desc.f->filedes, &ev);
+ }
+ if (0 != ret) {
+ return APR_EBADF;
+ }
+#elif defined(HAVE_POLL)
if (descriptor->desc_type == APR_POLL_SOCKET) {
pollset->pollset[pollset->nelts].fd = descriptor->desc.s->socketdes;
@@ -420,11 +553,97 @@
const apr_pollfd_t *descriptor)
{
apr_uint32_t i;
-#ifndef HAVE_POLL
+#ifdef HAVE_KQUEUE
+ apr_os_sock_t fd;
+#elif defined(HAVE_EPOLL)
+ struct epoll_event ev;
+ int ret = -1;
+#elif defined(HAVE_POLL)
apr_os_sock_t fd;
#endif
-#ifdef HAVE_POLL
+#ifdef HAVE_KQUEUE
+ for (i = 0; i < pollset->nelts; i++) {
+ if (descriptor->desc.s == pollset->query_set[i].desc.s) {
+ /* Found an instance of the fd: remove this and any other copies
*/
+ apr_uint32_t dst = i;
+ apr_uint32_t old_nelts = pollset->nelts;
+ pollset->nelts--;
+ for (i++; i < old_nelts; i++) {
+ if (descriptor->desc.s == pollset->query_set[i].desc.s) {
+ pollset->nelts--;
+ }
+ else {
+ pollset->query_set[dst] = pollset->query_set[i];
+ dst++;
+ }
+ }
+
+ if (descriptor->desc_type == APR_POLL_SOCKET) {
+ fd = descriptor->desc.s->socketdes;
+ }
+ else {
+ fd = descriptor->desc.f->filedes;
+ }
+
+ if (descriptor->reqevents & APR_POLLIN) {
+ EV_SET(&pollset->kevent, fd,
+ EVFILT_READ, EV_DELETE, 0, 0, NULL);
+
+ if (kevent(pollset->kqueue_fd, &pollset->kevent, 1, NULL, 0,
+ NULL) == -1) {
+ return APR_EBADF;
+ }
+ }
+
+ if (descriptor->reqevents & APR_POLLOUT) {
+ EV_SET(&pollset->kevent, fd,
+ EVFILT_WRITE, EV_DELETE, 0, 0, NULL);
+
+ if (kevent(pollset->kqueue_fd, &pollset->kevent, 1, NULL, 0,
+ NULL) == -1) {
+ return APR_EBADF;
+ }
+ }
+
+ return APR_SUCCESS;
+ }
+ }
+#elif defined(HAVE_EPOLL)
+ for (i = 0; i < pollset->nelts; i++) {
+ if (descriptor->desc.s == pollset->query_set[i].desc.s) {
+ /* Found an instance of the fd: remove this and any other copies
*/
+ apr_uint32_t dst = i;
+ apr_uint32_t old_nelts = pollset->nelts;
+ pollset->nelts--;
+ for (i++; i < old_nelts; i++) {
+ if (descriptor->desc.s == pollset->query_set[i].desc.s) {
+ pollset->nelts--;
+ }
+ else {
+ pollset->query_set[dst] = pollset->query_set[i];
+ dst++;
+ }
+ }
+ ev.events = get_epoll_event(descriptor->reqevents);
+ if (descriptor->desc_type == APR_POLL_SOCKET) {
+ ev.data.fd = descriptor->desc.s->socketdes;
+ ret = epoll_ctl(pollset->epoll_fd, EPOLL_CTL_DEL,
+ descriptor->desc.s->socketdes, &ev);
+ }
+ else {
+ ev.data.fd = descriptor->desc.f->filedes;
+ ret = epoll_ctl(pollset->epoll_fd, EPOLL_CTL_DEL,
+ descriptor->desc.f->filedes, &ev);
+ }
+ if (ret < 0) {
+ return APR_EBADF;
+ }
+
+ return APR_SUCCESS;
+ }
+ }
+#elif defined(HAVE_POLL)
for (i = 0; i < pollset->nelts; i++) {
if (descriptor->desc.s == pollset->query_set[i].desc.s) {
/* Found an instance of the fd: remove this and any other copies */
@@ -485,8 +704,119 @@
return APR_NOTFOUND;
}
+#ifdef HAVE_KQUEUE
+APR_DECLARE(apr_status_t) apr_pollset_poll(apr_pollset_t *pollset,
+ apr_interval_time_t timeout,
+ apr_int32_t *num,
+ const apr_pollfd_t **descriptors)
+{
+ int rv;
+ apr_uint32_t i, j, r = 0;
+ struct timespec tv, *tvptr;
+
+ if (timeout < 0) {
+ tvptr = NULL;
+ }
+ else {
+ tv.tv_sec = (long)apr_time_sec(timeout);
+ tv.tv_nsec = (long)apr_time_msec(timeout);
+ tvptr = &tv;
+ }
+
+ rv = kevent(pollset->kqueue_fd, NULL, 0, pollset->ke_set, pollset->nelts,
+ tvptr);
+ (*num) = rv;
+ if (rv < 0) {
+ return apr_get_netos_error();
+ }
+ if (rv == 0) {
+ return APR_TIMEUP;
+ }
+
+ /* TODO: Is there a better way to re-associate our data? */
+ for (i = 0; i < pollset->nelts; i++) {
+ apr_os_sock_t fd;
+ if (pollset->query_set[i].desc_type == APR_POLL_SOCKET) {
+ fd = pollset->query_set[i].desc.s->socketdes;
+ }
+ else {
+ fd = pollset->query_set[i].desc.f->filedes;
+ }
+ for (j = 0; j < rv; j++) {
+ if (pollset->ke_set[j].ident == fd ) {
+ pollset->result_set[r] = pollset->query_set[i];
+ pollset->result_set[r].rtnevents =
+ get_kqueue_revent(pollset->ke_set[j].filter,
+ pollset->ke_set[j].flags);
+ r++;
+ }
+ }
+ }
+
+ (*num) = r;
+
+ if (descriptors) {
+ *descriptors = pollset->result_set;
+ }
+
+ return APR_SUCCESS;
+}
+
+#elif defined(HAVE_EPOLL)
-#ifdef HAVE_POLL
+APR_DECLARE(apr_status_t) apr_pollset_poll(apr_pollset_t *pollset,
+ apr_interval_time_t timeout,
+ apr_int32_t *num,
+ const apr_pollfd_t **descriptors)
+{
+ int rv;
+ apr_uint32_t i, j, k;
+
+ if (timeout > 0) {
+ timeout /= 1000;
+ }
+
+ rv = epoll_wait(pollset->epoll_fd, pollset->pollset, pollset->nelts,
+ timeout);
+ (*num) = rv;
+ if (rv < 0) {
+ return apr_get_netos_error();
+ }
+ if (rv == 0) {
+ return APR_TIMEUP;
+ }
+ j = 0;
+ for (i = 0; i < pollset->nelts; i++) {
+ if (pollset->pollset[i].events != 0) {
+ /* TODO: Is there a better way to re-associate our data? */
+ for (k = 0; k < pollset->nelts; k++) {
+ if (pollset->query_set[k].desc_type == APR_POLL_SOCKET &&
+ pollset->query_set[k].desc.s->socketdes ==
+ pollset->pollset[i].data.fd) {
+ pollset->result_set[j] = pollset->query_set[k];
+ pollset->result_set[j].rtnevents =
+ get_epoll_revent(pollset->pollset[i].events);
+ j++;
+ break;
+ }
+ else if (pollset->query_set[k].desc_type == APR_POLL_FILE
+ && pollset->query_set[k].desc.f->filedes ==
+ pollset->pollset[i].data.fd) {
+ pollset->result_set[j] = pollset->query_set[k];
+ pollset->result_set[j].rtnevents =
+ get_epoll_revent(pollset->pollset[i].events);
+ j++;
+ break;
+ }
+ }
+ }
+ }
+ if (descriptors) {
+ *descriptors = pollset->result_set;
+ }
+ return APR_SUCCESS;
+}
+#elif defined(HAVE_POLL)
APR_DECLARE(apr_status_t) apr_pollset_poll(apr_pollset_t *pollset,
apr_interval_time_t timeout,
apr_int32_t *num,
Index: test/testpoll.c
===================================================================
RCS file: /home/cvs/apr/test/testpoll.c,v
retrieving revision 1.33
diff -u -r1.33 testpoll.c
--- test/testpoll.c 26 May 2004 14:50:27 -0000 1.33
+++ test/testpoll.c 7 Jun 2004 20:14:34 -0000
@@ -489,10 +489,14 @@
rv = apr_pollset_poll(pollset, 1000, &num, &hot_files);
ABTS_INT_EQUAL(tc, APR_SUCCESS, rv);
ABTS_INT_EQUAL(tc, 2, num);
- ABTS_PTR_EQUAL(tc, (void *)1, hot_files[0].client_data);
- ABTS_PTR_EQUAL(tc, s[0], hot_files[0].desc.s);
- ABTS_PTR_EQUAL(tc, (void *)4, hot_files[1].client_data);
- ABTS_PTR_EQUAL(tc, s[3], hot_files[1].desc.s);
+ ABTS_ASSERT(tc, "Incorrect socket in result set",
+ ((hot_files[0].desc.s == s[0]) && (hot_files[1].desc.s == s[3]))
||
+ ((hot_files[0].desc.s == s[3]) && (hot_files[1].desc.s == s[0])));
+ ABTS_ASSERT(tc, "Incorrect client data in result set",
+ ((hot_files[0].client_data == (void *)1) &&
+ (hot_files[1].client_data == (void *)4)) ||
+ ((hot_files[0].client_data == (void *)4) &&
+ (hot_files[1].client_data == (void *)1)));
}
abts_suite *testpoll(abts_suite *suite)
