]> git.ipfire.org Git - thirdparty/haproxy.git/commitdiff
[MAJOR] implemented support for FreeBSD's kqueue() polling mechanism
authorWilly Tarreau <w@1wt.eu>
Mon, 9 Apr 2007 10:03:06 +0000 (12:03 +0200)
committerWilly Tarreau <w@1wt.eu>
Mon, 9 Apr 2007 10:03:06 +0000 (12:03 +0200)
It has not been tested yet, but at least it builds.

Makefile.bsd
include/types/polling.h
src/ev_kqueue.c [new file with mode: 0644]
src/haproxy.c

index 043ff04dce9bb15ce955df5f4e764f9654d82084..0050701eef64329cbb94753a51f2507bb3d38144 100644 (file)
@@ -29,7 +29,7 @@ PCREDIR!= pcre-config --prefix 2>/dev/null || :
 #PCREDIR=/usr/local
 
 # This is for OpenBSD 3.0 and above
-COPTS.openbsd = -DENABLE_POLL
+COPTS.openbsd = -DENABLE_POLL -DENABLE_KQUEUE
 LIBS.openbsd =
 
 # CPU dependant optimizations
@@ -87,7 +87,8 @@ OBJS = src/haproxy.o src/list.o src/chtbl.o src/hashpjw.o src/base64.o \
        src/time.o src/fd.o src/regex.o src/cfgparse.o src/server.o \
        src/checks.o src/queue.o src/capture.o src/client.o src/proxy.o \
        src/proto_http.o src/stream_sock.o src/appsession.o src/backend.o \
-       src/session.o src/hdr_idx.o src/rbtree.o src/ev_select.o src/ev_poll.o
+       src/session.o src/hdr_idx.o src/rbtree.o src/ev_select.o src/ev_poll.o \
+       src/ev_kqueue.o
 
 all: haproxy
 
index ed3cf64dc015859b465d6a6c091420ce705dff85..81921ee4ab158f6ad6a52040b1f09afc872bd3b1 100644 (file)
@@ -52,8 +52,9 @@
 #define POLL_USE_SELECT         (1<<0)
 #define POLL_USE_POLL           (1<<1)
 #define POLL_USE_EPOLL          (1<<2)
+#define POLL_USE_KQUEUE         (1<<3)
 
-extern int cfg_polling_mechanism;       /* POLL_USE_{SELECT|POLL|EPOLL} */
+extern int cfg_polling_mechanism;       /* POLL_USE_{SELECT|POLL|EPOLL|KQUEUE} */
 
 
 #endif /* _TYPES_POLLING_H */
diff --git a/src/ev_kqueue.c b/src/ev_kqueue.c
new file mode 100644 (file)
index 0000000..98115fa
--- /dev/null
@@ -0,0 +1,238 @@
+/*
+ * FD polling functions for FreeBSD kqueue()
+ *
+ * Copyright 2000-2007 Willy Tarreau <w@1wt.eu>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ *
+ * Note: not knowing much about kqueue, I had to rely on OpenBSD's detailed man
+ * page and to check how it was implemented in lighttpd to understand it better.
+ * But it is possible that I got things wrong.
+ *
+ */
+
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/time.h>
+#include <sys/types.h>
+
+#include <sys/event.h>
+#include <sys/time.h>
+
+#include <common/compat.h>
+#include <common/config.h>
+#include <common/time.h>
+
+#include <types/fd.h>
+#include <types/global.h>
+
+#include <proto/fd.h>
+#include <proto/polling.h>
+#include <proto/task.h>
+
+/* private data */
+static fd_set *fd_evts[2];
+static int kqueue_fd;
+static struct kevent *kev = NULL;
+
+/* speeds up conversion of DIR_RD/DIR_WR to EVFILT* */
+static const int dir2filt[2] = { EVFILT_READ, EVFILT_WRITE };
+
+/* completes a change list for deletion */
+REGPRM3 static int kqev_del(struct kevent *kev, const int fd, const int dir)
+{
+       if (FD_ISSET(fd, fd_evts[dir])) {
+               FD_CLR(fd, fd_evts[dir]);
+               EV_SET(kev, fd, dir2filt[dir], EV_DELETE, 0, 0, NULL);
+               return 1;
+       }
+       return 0;
+}
+
+/*
+ * Returns non-zero if direction <dir> is already set for <fd>.
+ */
+REGPRM2 static int __fd_isset(const int fd, int dir)
+{
+       return FD_ISSET(fd, fd_evts[dir]);
+}
+
+REGPRM2 static int __fd_set(const int fd, int dir)
+{
+       /* if the value was set, do nothing */
+       if (FD_ISSET(fd, fd_evts[dir]))
+               return 0;
+
+       FD_SET(fd, fd_evts[dir]);
+       EV_SET(kev, fd, dir2filt[dir], EV_ADD|EV_CLEAR, 0, 0, NULL);
+       kevent(kqueue_fd, kev, 1, NULL, 0, NULL);
+       return 1;
+}
+
+REGPRM2 static int __fd_clr(const int fd, int dir)
+{
+       /* if the value was not set, do nothing */
+       //if (!FD_ISSET(fd, fd_evts[dir]))
+       //      return 0;
+       //
+       //FD_CLR(fd, fd_evts[dir]);
+       //EV_SET(&kev, fd, dir2filt[dir], EV_DELETE, 0, 0, NULL);
+       //kevent(kqueue_fd, &kev, 1, NULL, 0, NULL);
+       //return 1;
+       if (!kqev_del(kev, fd, dir))
+               return 0;
+       kevent(kqueue_fd, kev, 1, NULL, 0, NULL);
+       return 1;
+}
+
+REGPRM1 static void __fd_rem(int fd)
+{
+       int changes = 0;
+
+       //if (FD_ISSET(fd, fd_evts[DIR_RD])) {
+       //      FD_CLR(fd, fd_evts[DIR_RD]);
+       //      EV_SET(&kev[changes], fd, dir2filt[DIR_RD], EV_DELETE, 0, 0, NULL);
+       //      changes++;
+       //}
+       //if (FD_ISSET(fd, fd_evts[DIR_WR])) {
+       //      FD_CLR(fd, fd_evts[DIR_WR]);
+       //      EV_SET(&kev[changes], fd, dir2filt[DIR_WR], EV_DELETE, 0, 0, NULL);
+       //      changes++;
+       //}
+       changes += kqev_del(&kev[changes], fd, DIR_RD);
+       changes += kqev_del(&kev[changes], fd, DIR_WR);
+
+       if (changes)
+               kevent(kqueue_fd, kev, changes, NULL, 0, NULL);
+}
+
+/*
+ * kqueue() poller
+ */
+REGPRM2 static void kqueue_poll(struct poller *p, int wait_time)
+{
+       int status;
+       int count, fd;
+       struct timespec timeout;
+
+       timeout.tv_sec  =  wait_time / 1000;
+       timeout.tv_nsec = (wait_time % 1000) * 1000000;
+
+       status = kevent(kqueue_fd, // int kq
+                       NULL,      // const struct kevent *changelist
+                       0,         // int nchanges
+                       kev,       // struct kevent *eventlist
+                       maxfd,     // int nevents
+                       &timeout); // const struct timespec *timeout
+
+       for (count = 0; count < status; count++) {
+               fd = kev[count].ident;
+               if (kev[count].filter ==  EVFILT_READ) {
+                       if (FD_ISSET(fd, fd_evts[DIR_RD])) {
+                               if (fdtab[fd].state == FD_STCLOSE)
+                                       continue;
+                               fdtab[fd].cb[DIR_RD].f(fd);
+                       }
+               } else if (kev[count].filter ==  EVFILT_WRITE) {
+                       if (FD_ISSET(fd, fd_evts[DIR_WR])) {
+                               if (fdtab[fd].state == FD_STCLOSE)
+                                       continue;
+                               fdtab[fd].cb[DIR_WR].f(fd);
+                       }
+               }
+       }
+}
+
+/*
+ * Initialization of the kqueue() poller.
+ * Returns 0 in case of failure, non-zero in case of success. If it fails, it
+ * disables the poller by setting its pref to 0.
+ */
+REGPRM1 static int kqueue_init(struct poller *p)
+{
+       __label__ fail_wevt, fail_revt, fail_fd;
+       int fd_set_bytes;
+
+       p->private = NULL;
+       fd_set_bytes = sizeof(fd_set) * (global.maxsock + FD_SETSIZE - 1) / FD_SETSIZE;
+
+       kqueue_fd = kqueue();
+       if (kqueue_fd < 0)
+               goto fail_fd;
+
+       kev = (struct kevent*)calloc(1, sizeof(struct kevent) * global.maxsock);
+
+       if (kev == NULL)
+               goto fail_kev;
+               
+       if ((fd_evts[DIR_RD] = (fd_set *)calloc(1, fd_set_bytes)) == NULL)
+               goto fail_revt;
+
+       if ((fd_evts[DIR_WR] = (fd_set *)calloc(1, fd_set_bytes)) == NULL)
+               goto fail_wevt;
+
+       return 1;
+
+ fail_wevt:
+       free(fd_evts[DIR_RD]);
+ fail_revt:
+       free(kev);
+ fail_kev:
+       close(kqueue_fd);
+       kqueue_fd = 0;
+ fail_fd:
+       p->pref = 0;
+       return 0;
+}
+
+/*
+ * Termination of the kqueue() poller.
+ * Memory is released and the poller is marked as unselectable.
+ */
+REGPRM1 static void kqueue_term(struct poller *p)
+{
+       if (fd_evts[DIR_WR])
+               free(fd_evts[DIR_WR]);
+       if (fd_evts[DIR_RD])
+               free(fd_evts[DIR_RD]);
+       if (kev)
+               free(kev);
+       close(kqueue_fd);
+       kqueue_fd = 0;
+
+       p->private = NULL;
+       p->pref = 0;
+}
+
+/*
+ * The only exported function. Returns 1.
+ */
+//int kqueue_native_register(struct poller *p)
+int kqueue_register(struct poller *p)
+{
+       p->name = "kqueue";
+       p->pref = 300;
+       p->private = NULL;
+
+       p->init = kqueue_init;
+       p->term = kqueue_term;
+       p->poll = kqueue_poll;
+
+       p->isset  = __fd_isset;
+       p->cond_s = p->set = __fd_set;
+       p->cond_c = p->clr = __fd_clr;
+       p->clo    = p->rem = __fd_rem;
+       
+       return 1;
+}
+
+
+/*
+ * Local variables:
+ *  c-indent-level: 8
+ *  c-basic-offset: 8
+ * End:
+ */
index bafff4bfb1ed61f555192136e4579c5c70adafa9..0ab069994a1a0ab3fc48588684116bd1981fb6c5 100644 (file)
@@ -175,6 +175,9 @@ void usage(char *name)
 #if defined(ENABLE_EPOLL)
                "        -de disables epoll() usage even when available\n"
 #endif
+#if defined(ENABLE_KQUEUE)
+               "        -dk disables kqueue() usage even when available\n"
+#endif
 #if defined(ENABLE_POLL)
                "        -dp disables poll() usage even when available\n"
 #endif
@@ -371,6 +374,9 @@ void init(int argc, char **argv)
 #if defined(ENABLE_EPOLL)
        cfg_polling_mechanism |= POLL_USE_EPOLL;
 #endif
+#if defined(ENABLE_KQUEUE)
+       cfg_polling_mechanism |= POLL_USE_KQUEUE;
+#endif
 
        pid = getpid();
        progname = *argv;
@@ -396,6 +402,10 @@ void init(int argc, char **argv)
 #if defined(ENABLE_POLL)
                        else if (*flag == 'd' && flag[1] == 'p')
                                cfg_polling_mechanism &= ~POLL_USE_POLL;
+#endif
+#if defined(ENABLE_POLL)
+                       else if (*flag == 'd' && flag[1] == 'k')
+                               cfg_polling_mechanism &= ~POLL_USE_KQUEUE;
 #endif
                        else if (*flag == 'V')
                                arg_mode |= MODE_VERBOSE;
@@ -516,6 +526,9 @@ void init(int argc, char **argv)
        register_pollers();
        /* Note: we could register external pollers here */
 
+       if (!(cfg_polling_mechanism & POLL_USE_KQUEUE))
+               disable_poller("kqueue");
+
        if (!(cfg_polling_mechanism & POLL_USE_EPOLL))
                disable_poller("epoll");