kamailio/io_wait.h

/* 
 * $Id$
 * 
 * Copyright (C) 2005 iptelorg GmbH
 *
 * Permission to use, copy, modify, and distribute this software for any
 * purpose with or without fee is hereby granted, provided that the above
 * copyright notice and this permission notice appear in all copies.
 *
 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
 */
/*
 * tcp io wait common stuff used by tcp_main.c & tcp_read.c
 * All the functions are inline because of speed reasons and because they are
 * used only from 2 places.
 * You also have to define:
 *     int handle_io(struct fd_map* fm, short events, int idx) (see below)
 *     (this could be trivially replaced by a callback pointer entry attached
 *      to the io_wait handler if more flexibility rather then performance
 *      is needed)
 *      fd_type - define to some enum of you choice and define also
 *                FD_TYPE_DEFINED (if you don't do it fd_type will be defined
 *                to int). 0 has a special not set/not init. meaning
 *                (a lot of sanity checks and the sigio_rt code are based on
 *                 this assumption)
 *     local_malloc (defaults to pkg_malloc)
 *     local_free   (defaults to pkg_free)
 *  
 */
/* 
 * History:
 * --------
 *  2005-06-13  created by andrei
 *  2005-06-26  added kqueue (andrei)
 *  2005-07-01  added /dev/poll (andrei)
 *  2006-05-30  sigio 64 bit workarround enabled for kernels < 2.6.5 (andrei)
 *  2007-11-22  when handle_io() is called in a loop check & stop if the fd was
 *               removed inside handle_io() (andrei)
 *  2007-11-29  support for write (POLLOUT); added io_watch_chg() (andrei)
 *  2008-02-04  POLLRDHUP & EPOLLRDHUP support (automatically enabled if POLLIN
 *               is set) (andrei)
 */



#ifndef _io_wait_h
#define _io_wait_h

#include <errno.h>
#include <string.h>
#ifdef HAVE_SIGIO_RT
#define __USE_GNU /* or else F_SETSIG won't be included */
#include <sys/types.h> /* recv */
#include <sys/socket.h> /* recv */
#include <signal.h> /* sigprocmask, sigwait a.s.o */
#endif

#define _GNU_SOURCE  /* for POLLRDHUP on linux */
#include <sys/poll.h>
#include <fcntl.h>

#ifdef HAVE_EPOLL
#include <sys/epoll.h>
#endif
#ifdef HAVE_KQUEUE
#include <sys/types.h> /* needed on freebsd */
#include <sys/event.h>
#include <sys/time.h>
#endif
#ifdef HAVE_DEVPOLL
#include <sys/devpoll.h>
#endif
#ifdef HAVE_SELECT
/* needed on openbsd for select*/
#include <sys/time.h> 
#include <sys/types.h> 
#include <unistd.h>
/* needed according to POSIX for select*/
#include <sys/select.h>
#endif

#include "dprint.h"

#include "poll_types.h" /* poll_types*/
#ifdef HAVE_SIGIO_RT
#include "pt.h" /* mypid() */
#endif

#include "compiler_opt.h"


#ifdef HAVE_EPOLL
/* fix defines for EPOLL */
#if defined POLLRDHUP && ! defined EPOLLRDHUP
#define EPOLLRDHUP POLLRDHUP  /* should work on all linuxes */
#endif /* POLLRDHUP && EPOLLRDHUP */
#endif /* HAVE_EPOLL */


extern int _os_ver; /* os version number, needed to select bugs workarrounds */


#if 0
enum fd_types; /* this should be defined from the including file,
				  see tcp_main.c for an example, 
				  0 has a special meaning: not used/empty*/
#endif

#ifndef FD_TYPE_DEFINED
typedef int fd_type;
#define FD_TYPE_DEFINED
#endif

/* maps a fd to some other structure; used in almost all cases
 * except epoll and maybe kqueue or /dev/poll */
struct fd_map{
	int fd;               /* fd no */
	fd_type type;         /* "data" type */
	void* data;           /* pointer to the corresponding structure */
	short events;         /* events we are interested int */
};


#ifdef HAVE_KQUEUE
#ifndef KQ_CHANGES_ARRAY_SIZE
#define KQ_CHANGES_ARRAY_SIZE 256

#ifdef __OS_netbsd
#define KEV_UDATA_CAST (intptr_t)
#else
#define KEV_UDATA_CAST
#endif

#endif
#endif


/* handler structure */
struct io_wait_handler{
	enum poll_types poll_method;
	int flags;
	struct fd_map* fd_hash;
	int fd_no; /*  current index used in fd_array and the passed size for 
				   ep_array & kq_array*/
	int max_fd_no; /* maximum fd no, is also the size of fd_array,
						       fd_hash  and ep_array*/
	/* common stuff for POLL, SIGIO_RT and SELECT
	 * since poll support is always compiled => this will always be compiled */
	struct pollfd* fd_array; /* used also by devpoll as devpoll array */
	int crt_fd_array_idx; /*  crt idx for which handle_io is called
							 (updated also by del -> internal optimization) */
	/* end of common stuff */
#ifdef HAVE_EPOLL
	int epfd; /* epoll ctrl fd */
	struct epoll_event* ep_array;
#endif
#ifdef HAVE_SIGIO_RT
	sigset_t sset; /* signal mask for sigio & sigrtmin */
	int signo;     /* real time signal used */
#endif
#ifdef HAVE_KQUEUE
	int kq_fd;
	struct kevent* kq_array;   /* used for the eventlist*/
	struct kevent* kq_changes; /* used for the changelist */
	size_t kq_nchanges;
	size_t kq_changes_size; /* size of the changes array */
#endif
#ifdef HAVE_DEVPOLL
	int dpoll_fd;
#endif
#ifdef HAVE_SELECT
	fd_set master_rset; /* read set */
	fd_set master_wset; /* write set */
	int max_fd_select; /* maximum select used fd */
#endif
};

typedef struct io_wait_handler io_wait_h;


/* get the corresponding fd_map structure pointer */
#define get_fd_map(h, fd)		(&(h)->fd_hash[(fd)])
/* remove a fd_map structure from the hash; the pointer must be returned
 * by get_fd_map or hash_fd_map*/
#define unhash_fd_map(pfm)	\
	do{ \
		(pfm)->type=0 /*F_NONE */; \
		(pfm)->fd=-1; \
	}while(0)

/* add a fd_map structure to the fd hash */
static inline struct fd_map* hash_fd_map(	io_wait_h* h,
											int fd,
											short events,
											fd_type type,
											void* data)
{
	h->fd_hash[fd].fd=fd;
	h->fd_hash[fd].events=events;
	h->fd_hash[fd].type=type;
	h->fd_hash[fd].data=data;
	return &h->fd_hash[fd];
}



#ifdef HANDLE_IO_INLINE
/* generic handle io routine, this must be defined in the including file
 * (faster then registering a callback pointer)
 *
 * params:  fm     - pointer to a fd hash entry
 *          events - combinations of POLLIN, POLLOUT, POLLERR & POLLHUP
 *          idx    - index in the fd_array (or -1 if not known)
 * return: -1 on error
 *          0 on EAGAIN or when by some other way it is known that no more 
 *            io events are queued on the fd (the receive buffer is empty).
 *            Usefull to detect when there are no more io events queued for
 *            sigio_rt, epoll_et, kqueue.
 *         >0 on successfull read from the fd (when there might be more io
 *            queued -- the receive buffer might still be non-empty)
 */
inline static int handle_io(struct fd_map* fm, short events, int idx);
#else
int handle_io(struct fd_map* fm, short events, int idx);
#endif



#ifdef HAVE_KQUEUE
/*
 * kqueue specific function: register a change
 * (adds a change to the kevent change array, and if full flushes it first)
 *
 * TODO: check if the event already exists in the change list or if it's
 *       complementary to an event in the list (e.g. EVFILT_WRITE, EV_DELETE
 *       and EVFILT_WRITE, EV_ADD for the same fd).
 * returns: -1 on error, 0 on success
 */
static inline int kq_ev_change(io_wait_h* h, int fd, int filter, int flag, 
								void* data)
{
	int n;
	struct timespec tspec;

	if (h->kq_nchanges>=h->kq_changes_size){
		/* changes array full ! */
		LOG(L_WARN, "WARNING: kq_ev_change: kqueue changes array full"
					" trying to flush...\n");
		tspec.tv_sec=0;
		tspec.tv_nsec=0;
again:
		n=kevent(h->kq_fd, h->kq_changes, h->kq_nchanges, 0, 0, &tspec);
		if (n==-1){
			if (errno==EINTR) goto again;
			LOG(L_ERR, "ERROR: io_watch_add: kevent flush changes "
						" failed: %s [%d]\n", strerror(errno), errno);
			return -1;
		}
		h->kq_nchanges=0; /* changes array is empty */
	}
	EV_SET(&h->kq_changes[h->kq_nchanges], fd, filter, flag, 0, 0,
			KEV_UDATA_CAST data);
	h->kq_nchanges++;
	return 0;
}
#endif



/* generic io_watch_add function
 * Params:
 *     h      - pointer to initialized io_wait handle
 *     fd     - fd to watch
 *     events - bitmap with the fd events for which the fd should be watched
 *              (combination of POLLIN and POLLOUT)
 *     type   - fd type (non 0 value, returned in the call to handle_io)
 *     data   - pointer/private data returned in the handle_io call
 * returns 0 on success, -1 on error
 *
 * WARNING: handle_io() can be called immediately (from io_watch_add()) so
 *  make sure that any dependent init. (e.g. data stuff) is made before
 *  calling io_watch_add
 *
 * this version should be faster than pointers to poll_method specific
 * functions (it avoids functions calls, the overhead being only an extra
 *  switch())*/
inline static int io_watch_add(	io_wait_h* h,
								int fd,
								short events,
								fd_type type,
								void* data)
{

	/* helper macros */
#define fd_array_setup(ev) \
	do{ \
		h->fd_array[h->fd_no].fd=fd; \
		h->fd_array[h->fd_no].events=(ev); /* useless for select */ \
		h->fd_array[h->fd_no].revents=0;     /* useless for select */ \
	}while(0)
	
#define set_fd_flags(f) \
	do{ \
			flags=fcntl(fd, F_GETFL); \
			if (flags==-1){ \
				LOG(L_ERR, "ERROR: io_watch_add: fnctl: GETFL failed:" \
						" %s [%d]\n", strerror(errno), errno); \
				goto error; \
			} \
			if (fcntl(fd, F_SETFL, flags|(f))==-1){ \
				LOG(L_ERR, "ERROR: io_watch_add: fnctl: SETFL" \
							" failed: %s [%d]\n", strerror(errno), errno); \
				goto error; \
			} \
	}while(0)
	
	
	struct fd_map* e;
	int flags;
#ifdef HAVE_EPOLL
	struct epoll_event ep_event;
#endif
#ifdef HAVE_DEVPOLL
	struct pollfd pfd;
#endif
#if defined(HAVE_SIGIO_RT) || defined (HAVE_EPOLL)
	int n;
#endif
#if defined(HAVE_SIGIO_RT)
	int idx;
	int check_io;
	struct pollfd pf;
	
	check_io=0; /* set to 1 if we need to check for pre-existing queued
				   io/data on the fd */
	idx=-1;
#endif
	e=0;
	/* sanity checks */
	if (unlikely(fd==-1)){
		LOG(L_CRIT, "BUG: io_watch_add: fd is -1!\n");
		goto error;
	}
	if (unlikely((events&(POLLIN|POLLOUT))==0)){
		LOG(L_CRIT, "BUG: io_watch_add: invalid events: 0x%0x\n", events);
		goto error;
	}
	/* check if not too big */
	if (unlikely(h->fd_no>=h->max_fd_no)){
		LOG(L_CRIT, "ERROR: io_watch_add: maximum fd number exceeded:"
				" %d/%d\n", h->fd_no, h->max_fd_no);
		goto error;
	}
	DBG("DBG: io_watch_add(%p, %d, %d, %p), fd_no=%d\n",
			h, fd, type, data, h->fd_no);
	/*  hash sanity check */
	e=get_fd_map(h, fd);
	if (unlikely(e && (e->type!=0 /*F_NONE*/))){
		LOG(L_ERR, "ERROR: io_watch_add: trying to overwrite entry %d"
				" watched for %x in the hash(%d, %d, %p) with (%d, %d, %p)\n",
				fd, events, e->fd, e->type, e->data, fd, type, data);
		e=0;
		goto error;
	}
	
	if (unlikely((e=hash_fd_map(h, fd, events, type, data))==0)){
		LOG(L_ERR, "ERROR: io_watch_add: failed to hash the fd %d\n", fd);
		goto error;
	}
	switch(h->poll_method){ /* faster then pointer to functions */
		case POLL_POLL:
#ifdef POLLRDHUP
			/* listen to POLLRDHUP by default (if POLLIN) */
			events|=((int)!(events & POLLIN) - 1) & POLLRDHUP;
#endif /* POLLRDHUP */
			fd_array_setup(events);
			set_fd_flags(O_NONBLOCK);
			break;
#ifdef HAVE_SELECT
		case POLL_SELECT:
			fd_array_setup(events);
			if (likely(events & POLLIN))
				FD_SET(fd, &h->master_rset);
			if (unlikely(events & POLLOUT))
				FD_SET(fd, &h->master_wset);
			if (h->max_fd_select<fd) h->max_fd_select=fd;
			break;
#endif
#ifdef HAVE_SIGIO_RT
		case POLL_SIGIO_RT:
			fd_array_setup(events);
			/* re-set O_ASYNC might be needed, if not done from 
			 * io_watch_del (or if somebody wants to add a fd which has
			 * already O_ASYNC/F_SETSIG set on a duplicate)
			 */
			/* set async & signal */
			if (fcntl(fd, F_SETOWN, my_pid())==-1){
				LOG(L_ERR, "ERROR: io_watch_add: fnctl: SETOWN"
				" failed: %s [%d]\n", strerror(errno), errno);
				goto error;
			}
			if (fcntl(fd, F_SETSIG, h->signo)==-1){
				LOG(L_ERR, "ERROR: io_watch_add: fnctl: SETSIG"
					" failed: %s [%d]\n", strerror(errno), errno);
				goto error;
			}
			/* set both non-blocking and async */
			set_fd_flags(O_ASYNC| O_NONBLOCK);
#ifdef EXTRA_DEBUG
			DBG("io_watch_add: sigio_rt on f %d, signal %d to pid %d\n",
					fd,  h->signo, my_pid());
#endif
			/* empty socket receive buffer, if buffer is already full
			 * no more space to put packets
			 * => no more signals are ever generated
			 * also when moving fds, the freshly moved fd might have
			 *  already some bytes queued, we want to get them now
			 *  and not later -- andrei */
			idx=h->fd_no;
			check_io=1;
			break;
#endif
#ifdef HAVE_EPOLL
		case POLL_EPOLL_LT:
			ep_event.events=
#ifdef POLLRDHUP
						/* listen for EPOLLRDHUP too */
						((EPOLLIN|EPOLLRDHUP) & ((int)!(events & POLLIN)-1) ) |
#else /* POLLRDHUP */
						(EPOLLIN & ((int)!(events & POLLIN)-1) ) |
#endif /* POLLRDHUP */
						(EPOLLOUT & ((int)!(events & POLLOUT)-1) );
			ep_event.data.ptr=e;
again1:
			n=epoll_ctl(h->epfd, EPOLL_CTL_ADD, fd, &ep_event);
			if (unlikely(n==-1)){
				if (errno==EAGAIN) goto again1;
				LOG(L_ERR, "ERROR: io_watch_add: epoll_ctl failed: %s [%d]\n",
					strerror(errno), errno);
				goto error;
			}
			break;
		case POLL_EPOLL_ET:
			set_fd_flags(O_NONBLOCK);
			ep_event.events=
#ifdef POLLRDHUP
						/* listen for EPOLLRDHUP too */
						((EPOLLIN|EPOLLRDHUP) & ((int)!(events & POLLIN)-1) ) |
#else /* POLLRDHUP */
						(EPOLLIN & ((int)!(events & POLLIN)-1) ) |
#endif /* POLLRDHUP */
						(EPOLLOUT & ((int)!(events & POLLOUT)-1) ) |
						EPOLLET;
			ep_event.data.ptr=e;
again2:
			n=epoll_ctl(h->epfd, EPOLL_CTL_ADD, fd, &ep_event);
			if (unlikely(n==-1)){
				if (errno==EAGAIN) goto again2;
				LOG(L_ERR, "ERROR: io_watch_add: epoll_ctl failed: %s [%d]\n",
					strerror(errno), errno);
				goto error;
			}
			break;
#endif
#ifdef HAVE_KQUEUE
		case POLL_KQUEUE:
			if (likely( events & POLLIN)){
				if (unlikely(kq_ev_change(h, fd, EVFILT_READ, EV_ADD, e)==-1))
				goto error;
			}
			if (unlikely( events & POLLOUT)){
				if (unlikely(kq_ev_change(h, fd, EVFILT_WRITE, EV_ADD, e)==-1))
				{
					if (likely(events & POLLIN)){
						kq_ev_change(h, fd, EVFILT_READ, EV_DELETE, 0);
					}
				}
				goto error;
			}
			break;
#endif
#ifdef HAVE_DEVPOLL
		case POLL_DEVPOLL:
			pfd.fd=fd;
			pfd.events=events;
			pfd.revents=0;
again_devpoll:
			if (write(h->dpoll_fd, &pfd, sizeof(pfd))==-1){
				if (errno==EAGAIN) goto again_devpoll;
				LOG(L_ERR, "ERROR: io_watch_add: /dev/poll write failed:"
							"%s [%d]\n", strerror(errno), errno);
				goto error;
			}
			break;
#endif
			
		default:
			LOG(L_CRIT, "BUG: io_watch_add: no support for poll method "
					" %s (%d)\n", poll_method_str[h->poll_method],
					h->poll_method);
			goto error;
	}
	
	h->fd_no++; /* "activate" changes, for epoll/kqueue/devpoll it
				   has only informative value */
#if defined(HAVE_SIGIO_RT)
	if (check_io){
		/* handle possible pre-existing events */
		pf.fd=fd;
		pf.events=events;
check_io_again:
		n=0;
		while(e->type && ((n=poll(&pf, 1, 0))>0) && 
				(handle_io(e, pf.revents, idx)>0) &&
				(pf.revents & (e->events|POLLERR|POLLHUP)));
		if (unlikely(e->type && (n==-1))){
			if (errno==EINTR) goto check_io_again;
			LOG(L_ERR, "ERROR: io_watch_add: check_io poll: %s [%d]\n",
						strerror(errno), errno);
		}
	}
#endif
	return 0;
error:
	if (e) unhash_fd_map(e);
	return -1;
#undef fd_array_setup
#undef set_fd_flags 
}



#define IO_FD_CLOSING 16
/* parameters:    h - handler 
 *               fd - file descriptor
 *            index - index in the fd_array if known, -1 if not
 *                    (if index==-1 fd_array will be searched for the
 *                     corresponding fd* entry -- slower but unavoidable in 
 *                     some cases). index is not used (no fd_array) for epoll,
 *                     /dev/poll and kqueue
 *            flags - optimization flags, e.g. IO_FD_CLOSING, the fd was 
 *                    or will shortly be closed, in some cases we can avoid
 *                    extra remove operations (e.g.: epoll, kqueue, sigio)
 * returns 0 if ok, -1 on error */
inline static int io_watch_del(io_wait_h* h, int fd, int idx, int flags)
{
	
#define fix_fd_array \
	do{\
			if (unlikely(idx==-1)){ \
				/* fix idx if -1 and needed */ \
				for (idx=0; (idx<h->fd_no) && \
							(h->fd_array[idx].fd!=fd); idx++); \
			} \
			if (likely(idx<h->fd_no)){ \
				memmove(&h->fd_array[idx], &h->fd_array[idx+1], \
					(h->fd_no-(idx+1))*sizeof(*(h->fd_array))); \
				if ((idx<=h->crt_fd_array_idx) && (h->crt_fd_array_idx>=0)) \
					h->crt_fd_array_idx--; \
			} \
	}while(0)
	
	struct fd_map* e;
	int events;
#ifdef HAVE_EPOLL
	int n;
	struct epoll_event ep_event;
#endif
#ifdef HAVE_DEVPOLL
	struct pollfd pfd;
#endif
#ifdef HAVE_SIGIO_RT
	int fd_flags;
#endif
	
	if (unlikely((fd<0) || (fd>=h->max_fd_no))){
		LOG(L_CRIT, "BUG: io_watch_del: invalid fd %d, not in [0, %d) \n",
						fd, h->fd_no);
		goto error;
	}
	DBG("DBG: io_watch_del (%p, %d, %d, 0x%x) fd_no=%d called\n",
			h, fd, idx, flags, h->fd_no);
	e=get_fd_map(h, fd);
	/* more sanity checks */
	if (unlikely(e==0)){
		LOG(L_CRIT, "BUG: io_watch_del: no corresponding hash entry for %d\n",
					fd);
		goto error;
	}
	if (unlikely(e->type==0 /*F_NONE*/)){
		LOG(L_ERR, "ERROR: io_watch_del: trying to delete already erased"
				" entry %d in the hash(%d, %d, %p) flags %x)\n",
				fd, e->fd, e->type, e->data, flags);
		goto error;
	}
	events=e->events;
	unhash_fd_map(e);
	
	switch(h->poll_method){
		case POLL_POLL:
			fix_fd_array;
			break;
#ifdef HAVE_SELECT
		case POLL_SELECT:
			if (likely(events & POLLIN))
				FD_CLR(fd, &h->master_rset);
			if (unlikely(events & POLLOUT))
				FD_CLR(fd, &h->master_wset);
			if (unlikely(h->max_fd_select && (h->max_fd_select==fd)))
				/* we don't know the prev. max, so we just decrement it */
				h->max_fd_select--; 
			fix_fd_array;
			break;
#endif
#ifdef HAVE_SIGIO_RT
		case POLL_SIGIO_RT:
			fix_fd_array;
			/* the O_ASYNC flag must be reset all the time, the fd
			 *  can be changed only if  O_ASYNC is reset (if not and
			 *  the fd is a duplicate, you will get signals from the dup. fd
			 *  and not from the original, even if the dup. fd was closed
			 *  and the signals re-set on the original) -- andrei
			 */
			/*if (!(flags & IO_FD_CLOSING)){*/
				/* reset ASYNC */
				fd_flags=fcntl(fd, F_GETFL); 
				if (unlikely(fd_flags==-1)){ 
					LOG(L_ERR, "ERROR: io_watch_del: fnctl: GETFL failed:" 
							" %s [%d]\n", strerror(errno), errno); 
					goto error; 
				} 
				if (unlikely(fcntl(fd, F_SETFL, fd_flags&(~O_ASYNC))==-1)){ 
					LOG(L_ERR, "ERROR: io_watch_del: fnctl: SETFL" 
								" failed: %s [%d]\n", strerror(errno), errno); 
					goto error; 
				} 
			break;
#endif
#ifdef HAVE_EPOLL
		case POLL_EPOLL_LT:
		case POLL_EPOLL_ET:
			/* epoll doesn't seem to automatically remove sockets,
			 * if the socket is a duplicate/moved and the original
			 * is still open. The fd is removed from the epoll set
			 * only when the original (and all the  copies?) is/are 
			 * closed. This is probably a bug in epoll. --andrei */
#ifdef EPOLL_NO_CLOSE_BUG
			if (!(flags & IO_FD_CLOSING)){
#endif
again_epoll:
				n=epoll_ctl(h->epfd, EPOLL_CTL_DEL, fd, &ep_event);
				if (unlikely(n==-1)){
					if (errno==EAGAIN) goto again_epoll;
					LOG(L_ERR, "ERROR: io_watch_del: removing fd from epoll "
							"list failed: %s [%d]\n", strerror(errno), errno);
					goto error;
				}
#ifdef EPOLL_NO_CLOSE_BUG
			}
#endif
			break;
#endif
#ifdef HAVE_KQUEUE
		case POLL_KQUEUE:
			if (!(flags & IO_FD_CLOSING)){
				if (likely(events & POLLIN)){
					if (unlikely(kq_ev_change(h, fd, EVFILT_READ,
													EV_DELETE, 0) ==-1)){
						/* try to delete the write filter anyway */
						if (events & POLLOUT){
							kq_ev_change(h, fd, EVFILT_WRITE, EV_DELETE, 0);
						}
						goto error;
					}
				}
				if (unlikely(events & POLLOUT)){
					if (unlikely(kq_ev_change(h, fd, EVFILT_WRITE,
													EV_DELETE, 0) ==-1))
						goto error;
				}
			}
			break;
#endif
#ifdef HAVE_DEVPOLL
		case POLL_DEVPOLL:
				/* for /dev/poll the closed fds _must_ be removed
				   (they are not removed automatically on close()) */
				pfd.fd=fd;
				pfd.events=POLLREMOVE;
				pfd.revents=0;
again_devpoll:
				if (write(h->dpoll_fd, &pfd, sizeof(pfd))==-1){
					if (errno==EINTR) goto again_devpoll;
					LOG(L_ERR, "ERROR: io_watch_del: removing fd from "
								"/dev/poll failed: %s [%d]\n", 
								strerror(errno), errno);
					goto error;
				}
				break;
#endif
		default:
			LOG(L_CRIT, "BUG: io_watch_del: no support for poll method "
					" %s (%d)\n", poll_method_str[h->poll_method], 
					h->poll_method);
			goto error;
	}
	h->fd_no--;
	return 0;
error:
	return -1;
#undef fix_fd_array
}



/* parameters:    h - handler 
 *               fd - file descriptor
 *           events - new events to watch for
 *              idx - index in the fd_array if known, -1 if not
 *                    (if index==-1 fd_array will be searched for the
 *                     corresponding fd* entry -- slower but unavoidable in 
 *                     some cases). index is not used (no fd_array) for epoll,
 *                     /dev/poll and kqueue
 * returns 0 if ok, -1 on error */
inline static int io_watch_chg(io_wait_h* h, int fd, short events, int idx )
{
	
#define fd_array_chg(ev) \
	do{\
			if (unlikely(idx==-1)){ \
				/* fix idx if -1 and needed */ \
				for (idx=0; (idx<h->fd_no) && \
							(h->fd_array[idx].fd!=fd); idx++); \
			} \
			if (likely(idx<h->fd_no)){ \
				h->fd_array[idx].events=(ev); \
			} \
	}while(0)
	
	struct fd_map* e;
	int add_events;
	int del_events;
#ifdef HAVE_DEVPOLL
	struct pollfd pfd;
#endif
#ifdef HAVE_EPOLL
	int n;
	struct epoll_event ep_event;
#endif
	
	if (unlikely((fd<0) || (fd>=h->max_fd_no))){
		LOG(L_CRIT, "BUG: io_watch_chg: invalid fd %d, not in [0, %d) \n",
						fd, h->fd_no);
		goto error;
	}
	if (unlikely((events&(POLLIN|POLLOUT))==0)){
		LOG(L_CRIT, "BUG: io_watch_chg: invalid events: 0x%0x\n", events);
		goto error;
	}
	DBG("DBG: io_watch_chg (%p, %d, 0x%x, 0x%x) fd_no=%d called\n",
			h, fd, events, idx, h->fd_no);
	e=get_fd_map(h, fd);
	/* more sanity checks */
	if (unlikely(e==0)){
		LOG(L_CRIT, "BUG: io_watch_chg: no corresponding hash entry for %d\n",
					fd);
		goto error;
	}
	if (unlikely(e->type==0 /*F_NONE*/)){
		LOG(L_ERR, "ERROR: io_watch_chg: trying to change an already erased"
				" entry %d in the hash(%d, %d, %p) )\n",
				fd, e->fd, e->type, e->data);
		goto error;
	}
	
	add_events=events & ~e->events;
	del_events=e->events & ~events;
	e->events=events;
	switch(h->poll_method){
		case POLL_POLL:
#ifdef POLLRDHUP
			/* listen to POLLRDHUP by default (if POLLIN) */
			events|=((int)!(events & POLLIN) - 1) & POLLRDHUP;
#endif /* POLLRDHUP */
			fd_array_chg(events);
			break;
#ifdef HAVE_SELECT
		case POLL_SELECT:
			fd_array_chg(events);
			if (unlikely(del_events & POLLIN))
				FD_CLR(fd, &h->master_rset);
			else if (unlikely(add_events & POLLIN))
				FD_SET(fd, &h->master_rset);
			if (likely(del_events & POLLOUT))
				FD_CLR(fd, &h->master_wset);
			else if (likely(add_events & POLLOUT))
				FD_SET(fd, &h->master_wset);
			break;
#endif
#ifdef HAVE_SIGIO_RT
		case POLL_SIGIO_RT:
			fd_array_chg(events);
			/* no need for check_io, since SIGIO_RT listens by default for all
			 * the events */
			break;
#endif
#ifdef HAVE_EPOLL
		case POLL_EPOLL_LT:
				ep_event.events=
#ifdef POLLRDHUP
						/* listen for EPOLLRDHUP too */
						((EPOLLIN|EPOLLRDHUP) & ((int)!(events & POLLIN)-1) ) |
#else /* POLLRDHUP */
						(EPOLLIN & ((int)!(events & POLLIN)-1) ) |
#endif /* POLLRDHUP */
						(EPOLLOUT & ((int)!(events & POLLOUT)-1) );
				ep_event.data.ptr=e;
again_epoll_lt:
				n=epoll_ctl(h->epfd, EPOLL_CTL_MOD, fd, &ep_event);
				if (unlikely(n==-1)){
					if (errno==EAGAIN) goto again_epoll_lt;
					LOG(L_ERR, "ERROR: io_watch_chg: modifying epoll events"
							" failed: %s [%d]\n", strerror(errno), errno);
					goto error;
				}
			break;
		case POLL_EPOLL_ET:
				ep_event.events=
#ifdef POLLRDHUP
						/* listen for EPOLLRDHUP too */
						((EPOLLIN|EPOLLRDHUP) & ((int)!(events & POLLIN)-1) ) |
#else /* POLLRDHUP */
						(EPOLLIN & ((int)!(events & POLLIN)-1) ) |
#endif /* POLLRDHUP */
						(EPOLLOUT & ((int)!(events & POLLOUT)-1) ) |
						EPOLLET;
				ep_event.data.ptr=e;
again_epoll_et:
				n=epoll_ctl(h->epfd, EPOLL_CTL_MOD, fd, &ep_event);
				if (unlikely(n==-1)){
					if (errno==EAGAIN) goto again_epoll_et;
					LOG(L_ERR, "ERROR: io_watch_chg: modifying epoll events"
							" failed: %s [%d]\n", strerror(errno), errno);
					goto error;
				}
			break;
#endif
#ifdef HAVE_KQUEUE
		case POLL_KQUEUE:
			if (unlikely(del_events & POLLIN)){
				if (unlikely(kq_ev_change(h, fd, EVFILT_READ,
														EV_DELETE, 0) ==-1))
						goto error;
			}else if (unlikely(add_events & POLLIN)){
				if (unlikely(kq_ev_change(h, fd, EVFILT_READ, EV_ADD, e) ==-1))
					goto error;
			}
			if (likely(del_events & POLLOUT)){
				if (unlikely(kq_ev_change(h, fd, EVFILT_WRITE,
														EV_DELETE, 0) ==-1))
						goto error;
			}else if (likely(add_events & POLLOUT)){
				if (unlikely(kq_ev_change(h, fd, EVFILT_WRITE, EV_ADD, e)==-1))
					goto error;
			}
			break;
#endif
#ifdef HAVE_DEVPOLL
		case POLL_DEVPOLL:
				/* for /dev/poll the closed fds _must_ be removed
				   (they are not removed automatically on close()) */
				pfd.fd=fd;
				pfd.events=POLLREMOVE;
				pfd.revents=0;
again_devpoll1:
				if (unlikely(write(h->dpoll_fd, &pfd, sizeof(pfd))==-1)){
					if (errno==EINTR) goto again_devpoll1;
					LOG(L_ERR, "ERROR: io_watch_chg: removing fd from "
								"/dev/poll failed: %s [%d]\n", 
								strerror(errno), errno);
					goto error;
				}
again_devpoll2:
				pfd.events=events;
				pfd.revents=0;
				if (unlikely(write(h->dpoll_fd, &pfd, sizeof(pfd))==-1)){
					if (errno==EINTR) goto again_devpoll2;
					LOG(L_ERR, "ERROR: io_watch_chg: re-adding fd to "
								"/dev/poll failed: %s [%d]\n", 
								strerror(errno), errno);
					goto error;
				}
				break;
#endif
		default:
			LOG(L_CRIT, "BUG: io_watch_chg: no support for poll method "
					" %s (%d)\n", poll_method_str[h->poll_method], 
					h->poll_method);
			goto error;
	}
	return 0;
error:
	return -1;
#undef fix_fd_array
}



/* io_wait_loop_x style function 
 * wait for io using poll()
 * params: h      - io_wait handle
 *         t      - timeout in s
 *         repeat - if !=0 handle_io will be called until it returns <=0
 * returns: number of IO events handled on success (can be 0), -1 on error
 */
inline static int io_wait_loop_poll(io_wait_h* h, int t, int repeat)
{
	int n, r;
	int ret;
	struct fd_map* fm;
	
again:
		ret=n=poll(h->fd_array, h->fd_no, t*1000);
		if (n==-1){
			if (errno==EINTR) goto again; /* signal, ignore it */
			else{
				LOG(L_ERR, "ERROR:io_wait_loop_poll: poll: %s [%d]\n",
						strerror(errno), errno);
				goto error;
			}
		}
		for (r=0; (r<h->fd_no) && n; r++){
			fm=get_fd_map(h, h->fd_array[r].fd);
			if (h->fd_array[r].revents & (fm->events|POLLERR|POLLHUP)){
				n--;
				/* sanity checks */
				if (unlikely((h->fd_array[r].fd >= h->max_fd_no)||
								(h->fd_array[r].fd < 0))){
					LOG(L_CRIT, "BUG: io_wait_loop_poll: bad fd %d "
							"(no in the 0 - %d range)\n",
							h->fd_array[r].fd, h->max_fd_no);
					/* try to continue anyway */
					h->fd_array[r].events=0; /* clear the events */
					continue;
				}
				h->crt_fd_array_idx=r;
				/* repeat handle_io if repeat, fd still watched (not deleted
				 *  inside handle_io), handle_io returns that there's still
				 *  IO and the fd is still watched for the triggering event */
				while(fm->type && 
						(handle_io(fm, h->fd_array[r].revents, r) > 0) &&
						repeat && ((fm->events|POLLERR|POLLHUP) &
													h->fd_array[r].revents));
				r=h->crt_fd_array_idx; /* can change due to io_watch_del(fd) 
										  array shifting */
			}
		}
error:
	return ret;
}



#ifdef HAVE_SELECT
/* wait for io using select */
inline static int io_wait_loop_select(io_wait_h* h, int t, int repeat)
{
	fd_set sel_rset;
	fd_set sel_wset;
	int n, ret;
	struct timeval timeout;
	int r;
	struct fd_map* fm;
	int revents;
	
again:
		sel_rset=h->master_rset;
		sel_wset=h->master_wset;
		timeout.tv_sec=t;
		timeout.tv_usec=0;
		ret=n=select(h->max_fd_select+1, &sel_rset, &sel_wset, 0, &timeout);
		if (n<0){
			if (errno==EINTR) goto again; /* just a signal */
			LOG(L_ERR, "ERROR: io_wait_loop_select: select: %s [%d]\n",
					strerror(errno), errno);
			n=0;
			/* continue */
		}
		/* use poll fd array */
		for(r=0; (r<h->fd_no) && n; r++){
			revents=0;
			if (likely(FD_ISSET(h->fd_array[r].fd, &sel_rset)))
				revents|=POLLIN;
			if (unlikely(FD_ISSET(h->fd_array[r].fd, &sel_wset)))
				revents|=POLLOUT;
			if (unlikely(revents)){
				h->crt_fd_array_idx=r;
				fm=get_fd_map(h, h->fd_array[r].fd);
				while(fm->type && (fm->events & revents) && 
						(handle_io(fm, revents, r)>0) && repeat);
				r=h->crt_fd_array_idx; /* can change due to io_watch_del(fd) 
										  array shifting */
				n--;
			}
		};
	return ret;
}
#endif



#ifdef HAVE_EPOLL
inline static int io_wait_loop_epoll(io_wait_h* h, int t, int repeat)
{
	int n, r;
	struct fd_map* fm;
	int revents;
	
again:
		n=epoll_wait(h->epfd, h->ep_array, h->fd_no, t*1000);
		if (unlikely(n==-1)){
			if (errno==EINTR) goto again; /* signal, ignore it */
			else{
				LOG(L_ERR, "ERROR:io_wait_loop_epoll: "
						"epoll_wait(%d, %p, %d, %d): %s [%d]\n", 
						h->epfd, h->ep_array, h->fd_no, t*1000,
						strerror(errno), errno);
				goto error;
			}
		}
#if 0
		if (n>1){
			for(r=0; r<n; r++){
				LOG(L_ERR, "WARNING: ep_array[%d]= %x, %p\n",
						r, h->ep_array[r].events, h->ep_array[r].data.ptr);
			}
		}
#endif
		for (r=0; r<n; r++){
			revents= (POLLIN & (!(h->ep_array[r].events & (EPOLLIN|EPOLLPRI))
						-1)) |
					 (POLLOUT & (!(h->ep_array[r].events & EPOLLOUT)-1)) |
					 (POLLERR & (!(h->ep_array[r].events & EPOLLERR)-1)) |
					 (POLLHUP & (!(h->ep_array[r].events & EPOLLHUP)-1))
#ifdef POLLRDHUP
					| (POLLRDHUP & (!(h->ep_array[r].events & EPOLLRDHUP)-1))
#endif
					;
			if (likely(revents)){
				fm=(struct fd_map*)h->ep_array[r].data.ptr;
				while(fm->type && ((fm->events|POLLERR|POLLHUP) & revents) && 
						(handle_io(fm, revents, -1)>0) && repeat);
			}else{
				LOG(L_ERR, "ERROR:io_wait_loop_epoll: unexpected event %x"
							" on %d/%d, data=%p\n", h->ep_array[r].events,
							r+1, n, h->ep_array[r].data.ptr);
			}
		}
error:
	return n;
}
#endif



#ifdef HAVE_KQUEUE
inline static int io_wait_loop_kqueue(io_wait_h* h, int t, int repeat)
{
	int n, r;
	struct timespec tspec;
	struct fd_map* fm;
	int revents;
	
	tspec.tv_sec=t;
	tspec.tv_nsec=0;
again:
		n=kevent(h->kq_fd, h->kq_changes, h->kq_nchanges,  h->kq_array,
					h->fd_no, &tspec);
		if (unlikely(n==-1)){
			if (errno==EINTR) goto again; /* signal, ignore it */
			else{
				LOG(L_ERR, "ERROR: io_wait_loop_kqueue: kevent:"
						" %s [%d]\n", strerror(errno), errno);
				goto error;
			}
		}
		h->kq_nchanges=0; /* reset changes array */
		for (r=0; r<n; r++){
#ifdef EXTRA_DEBUG
			DBG("DBG: kqueue: event %d/%d: fd=%d, udata=%lx, flags=0x%x\n",
					r, n, h->kq_array[r].ident, (long)h->kq_array[r].udata,
					h->kq_array[r].flags);
#endif
#if 0
			if (unlikely(h->kq_array[r].flags & EV_ERROR)){
				/* error in changes: we ignore it, it can be caused by
				   trying to remove an already closed fd: race between
				   adding something to the changes array, close() and
				   applying the changes */
				LOG(L_INFO, "INFO: io_wait_loop_kqueue: kevent error on "
							"fd %ld: %s [%ld]\n", h->kq_array[r].ident,
							strerror(h->kq_array[r].data),
							(long)h->kq_array[r].data);
			}else{ 
#endif
				fm=(struct fd_map*)h->kq_array[r].udata;
				if (likely(h->kq_array[r].filter==EVFILT_READ)){
					revents=POLLIN | 
						(((int)!(h->kq_array[r].flags & EV_EOF)-1)&POLLHUP);
					while(fm->type && (fm->events & revents) && 
							(handle_io(fm, revents, -1)>0) && repeat);
				}else if (h->kq_array[r].filter==EVFILT_WRITE){
					revents=POLLOUT | 
						(((int)!(h->kq_array[r].flags & EV_EOF)-1)&POLLHUP);
					while(fm->type && (fm->events & revents) && 
							(handle_io(fm, revents, -1)>0) && repeat);
				}
			/*} */
		}
error:
	return n;
}
#endif



#ifdef HAVE_SIGIO_RT
/* sigio rt version has no repeat (it doesn't make sense)*/
inline static int io_wait_loop_sigio_rt(io_wait_h* h, int t)
{
	int n;
	int ret;
	struct timespec ts;
	siginfo_t siginfo;
	int sigio_band;
	int sigio_fd;
	struct fd_map* fm;
	int revents;
#ifdef SIGINFO64_WORKARROUND
	int* pi;
#endif
	
	
	ret=1; /* 1 event per call normally */
	ts.tv_sec=t;
	ts.tv_nsec=0;
	if (unlikely(!sigismember(&h->sset, h->signo) ||
					!sigismember(&h->sset, SIGIO))) {
		LOG(L_CRIT, "BUG: io_wait_loop_sigio_rt: the signal mask"
				" is not properly set!\n");
		goto error;
	}
again:
	n=sigtimedwait(&h->sset, &siginfo, &ts);
	if (unlikely(n==-1)){
		if (errno==EINTR) goto again; /* some other signal, ignore it */
		else if (errno==EAGAIN){ /* timeout */
			ret=0;
			goto end;
		}else{
			LOG(L_ERR, "ERROR: io_wait_loop_sigio_rt: sigtimed_wait"
					" %s [%d]\n", strerror(errno), errno);
			goto error;
		}
	}
	if (likely(n!=SIGIO)){
#ifdef SIGINFO64_WORKARROUND
		/* on linux siginfo.si_band is defined as long in userspace
		 * and as int in kernel (< 2.6.5) => on 64 bits things will break!
		 * (si_band will include si_fd, and si_fd will contain
		 *  garbage).
		 *  see /usr/src/linux/include/asm-generic/siginfo.h and
		 *      /usr/include/bits/siginfo.h
		 *  On newer kernels this is fixed (si_band is long in the kernel too).
		 * -- andrei */
		if  ((_os_ver<0x020605) && (sizeof(siginfo.si_band)>sizeof(int))){
			pi=(int*)(void*)&siginfo.si_band; /* avoid type punning warnings */
			sigio_band=*pi;
			sigio_fd=*(pi+1);
		}else
#endif
		{
			sigio_band=siginfo.si_band;
			sigio_fd=siginfo.si_fd;
		}
		if (unlikely(siginfo.si_code==SI_SIGIO)){
			/* old style, we don't know the event (linux 2.2.?) */
			LOG(L_WARN, "WARNING: io_wait_loop_sigio_rt: old style sigio"
					" interface\n");
			fm=get_fd_map(h, sigio_fd);
			/* we can have queued signals generated by fds not watched
			 * any more, or by fds in transition, to a child => ignore them*/
			if (fm->type)
				handle_io(fm, POLLIN|POLLOUT, -1);
		}else{
			/* si_code contains the SIGPOLL reason: POLL_IN, POLL_OUT,
			 *  POLL_MSG, POLL_ERR, POLL_PRI or POLL_HUP
			 * and si_band the translated poll event bitmap:
			 *  POLLIN|POLLRDNORM  (=POLL_IN),
			 *  POLLOUT|POLLWRNORM|POLLWRBAND (=POLL_OUT),
			 *  POLLIN|POLLRDNORM|POLLMSG (=POLL_MSG),
			 *  POLLERR (=POLL_ERR),
			 *  POLLPRI|POLLRDBAND (=POLL_PRI),
			 *  POLLHUP|POLLERR (=POLL_HUP) 
			 *  [linux 2.6.22 fs/fcntl.c:447]
			 */
#ifdef EXTRA_DEBUG
			DBG("io_wait_loop_sigio_rt: siginfo: signal=%d (%d),"
					" si_code=%d, si_band=0x%x,"
					" si_fd=%d\n",
					siginfo.si_signo, n, siginfo.si_code, 
					(unsigned)sigio_band,
					sigio_fd);
#endif
			/* on some errors (e.g. when receving TCP RST), sigio_band will
			 * be set to 0x08 (POLLERR) or 0x18 (POLLERR|POLLHUP - on stream
			 *  unix socket close) , so better catch all events --andrei */
			if (likely(sigio_band)){
				fm=get_fd_map(h, sigio_fd);
				revents=sigio_band;
				/* fix revents==POLLPRI case */
				revents |= (!(revents & POLLPRI)-1) & POLLIN;
				/* we can have queued signals generated by fds not watched
			 	 * any more, or by fds in transition, to a child 
				 * => ignore them */
				if (fm->type && ((fm->events|POLLERR|POLLHUP) & revents))
					handle_io(fm, revents, -1);
				else
					DBG("WARNING: io_wait_loop_sigio_rt: ignoring event"
							" %x on fd %d, watching for %x, si_code=%x "
							"(fm->type=%d, fm->fd=%d, fm->data=%p)\n",
							sigio_band, sigio_fd, fm->events, siginfo.si_code,
							fm->type, fm->fd, fm->data);
			}else{
				LOG(L_ERR, "ERROR: io_wait_loop_sigio_rt: unexpected event"
							" on fd %d: %x\n", sigio_fd, sigio_band);
			}
		}
	}else{
		/* signal queue overflow 
		 * TODO: increase signal queue size: 2.4x /proc/.., 2.6x -rlimits */
		LOG(L_WARN, "WARNING: io_wait_loop_sigio_rt: signal queue overflowed"
					"- falling back to poll\n");
		/* clear real-time signal queue
		 * both SIG_IGN and SIG_DFL are needed , it doesn't work
		 * only with SIG_DFL  */
		if (signal(h->signo, SIG_IGN)==SIG_ERR){
			LOG(L_CRIT, "BUG: do_poll: couldn't reset signal to IGN\n");
		}
		
		if (signal(h->signo, SIG_DFL)==SIG_ERR){
			LOG(L_CRIT, "BUG: do_poll: couldn't reset signal to DFL\n");
		}
		/* falling back to normal poll */
		ret=io_wait_loop_poll(h, -1, 1);
	}
end:
	return ret;
error:
	return -1;
}
#endif



#ifdef HAVE_DEVPOLL
inline static int io_wait_loop_devpoll(io_wait_h* h, int t, int repeat)
{
	int n, r;
	int ret;
	struct dvpoll dpoll;
	struct fd_map* fm;

		dpoll.dp_timeout=t*1000;
		dpoll.dp_nfds=h->fd_no;
		dpoll.dp_fds=h->fd_array;
again:
		ret=n=ioctl(h->dpoll_fd, DP_POLL, &dpoll);
		if (unlikely(n==-1)){
			if (errno==EINTR) goto again; /* signal, ignore it */
			else{
				LOG(L_ERR, "ERROR:io_wait_loop_devpoll: ioctl: %s [%d]\n",
						strerror(errno), errno);
				goto error;
			}
		}
		for (r=0; r< n; r++){
			if (h->fd_array[r].revents & (POLLNVAL|POLLERR)){
				LOG(L_ERR, "ERROR: io_wait_loop_devpoll: pollinval returned"
							" for fd %d, revents=%x\n",
							h->fd_array[r].fd, h->fd_array[r].revents);
			}
			/* POLLIN|POLLHUP just go through */
			fm=get_fd_map(h, h->fd_array[r].fd);
			while(fm->type && (fm->events & h->fd_array[r].revents) &&
					(handle_io(fm, h->fd_array[r].revents, r) > 0) && repeat);
		}
error:
	return ret;
}
#endif



/* init */


/* initializes the static vars/arrays
 * params:      h - pointer to the io_wait_h that will be initialized
 *         max_fd - maximum allowed fd number
 *         poll_m - poll method (0 for automatic best fit)
 */
int init_io_wait(io_wait_h* h, int max_fd, enum poll_types poll_method);

/* destroys everything init_io_wait allocated */
void destroy_io_wait(io_wait_h* h);


#endif