jni/pjproject-android/pjlib/src/pj/ioqueue_epoll.c

/* $Id: ioqueue_epoll.c 4528 2013-05-30 07:01:11Z ming $ */
/* 
 * Copyright (C) 2008-2011 Teluu Inc. (http://www.teluu.com)
 * Copyright (C) 2003-2008 Benny Prijono <benny@prijono.org>
 *
 * 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.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA 
 */
/*
 * ioqueue_epoll.c
 *
 * This is the implementation of IOQueue framework using /dev/epoll
 * API in _both_ Linux user-mode and kernel-mode.
 */

#include <pj/ioqueue.h>
#include <pj/os.h>
#include <pj/lock.h>
#include <pj/log.h>
#include <pj/list.h>
#include <pj/pool.h>
#include <pj/string.h>
#include <pj/assert.h>
#include <pj/errno.h>
#include <pj/sock.h>
#include <pj/compat/socket.h>
#include <pj/rand.h>

#if !defined(PJ_LINUX_KERNEL) || PJ_LINUX_KERNEL==0
    /*
     * Linux user mode
     */
#   include <sys/epoll.h>
#   include <errno.h>
#   include <unistd.h>

#   define epoll_data		data.ptr
#   define epoll_data_type	void*
#   define ioctl_val_type	unsigned long
#   define getsockopt_val_ptr	int*
#   define os_getsockopt	getsockopt
#   define os_ioctl		ioctl
#   define os_read		read
#   define os_close		close
#   define os_epoll_create	epoll_create
#   define os_epoll_ctl		epoll_ctl
#   define os_epoll_wait	epoll_wait
#else
    /*
     * Linux kernel mode.
     */
#   include <linux/config.h>
#   include <linux/version.h>
#   if defined(MODVERSIONS)
#	include <linux/modversions.h>
#   endif
#   include <linux/kernel.h>
#   include <linux/poll.h>
#   include <linux/eventpoll.h>
#   include <linux/syscalls.h>
#   include <linux/errno.h>
#   include <linux/unistd.h>
#   include <asm/ioctls.h>
    enum EPOLL_EVENTS
    {
	EPOLLIN = 0x001,
	EPOLLOUT = 0x004,
	EPOLLERR = 0x008,
    };
#   define os_epoll_create		sys_epoll_create
    static int os_epoll_ctl(int epfd, int op, int fd, struct epoll_event *event)
    {
	long rc;
	mm_segment_t oldfs = get_fs();
	set_fs(KERNEL_DS);
	rc = sys_epoll_ctl(epfd, op, fd, event);
	set_fs(oldfs);
	if (rc) {
	    errno = -rc;
	    return -1;
	} else {
	    return 0;
	}
    }
    static int os_epoll_wait(int epfd, struct epoll_event *events,
			  int maxevents, int timeout)
    {
	int count;
	mm_segment_t oldfs = get_fs();
	set_fs(KERNEL_DS);
	count = sys_epoll_wait(epfd, events, maxevents, timeout);
	set_fs(oldfs);
	return count;
    }
#   define os_close		sys_close
#   define os_getsockopt	pj_sock_getsockopt
    static int os_read(int fd, void *buf, size_t len)
    {
	long rc;
	mm_segment_t oldfs = get_fs();
	set_fs(KERNEL_DS);
	rc = sys_read(fd, buf, len);
	set_fs(oldfs);
	if (rc) {
	    errno = -rc;
	    return -1;
	} else {
	    return 0;
	}
    }
#   define socklen_t		unsigned
#   define ioctl_val_type	unsigned long
    int ioctl(int fd, int opt, ioctl_val_type value);
    static int os_ioctl(int fd, int opt, ioctl_val_type value)
    {
	int rc;
        mm_segment_t oldfs = get_fs();
	set_fs(KERNEL_DS);
	rc = ioctl(fd, opt, value);
	set_fs(oldfs);
	if (rc < 0) {
	    errno = -rc;
	    return rc;
	} else
	    return rc;
    }
#   define getsockopt_val_ptr	char*

#   define epoll_data		data
#   define epoll_data_type	__u32
#endif

#define THIS_FILE   "ioq_epoll"

//#define TRACE_(expr) PJ_LOG(3,expr)
#define TRACE_(expr)

/*
 * Include common ioqueue abstraction.
 */
#include "ioqueue_common_abs.h"

/*
 * This describes each key.
 */
struct pj_ioqueue_key_t
{
    DECLARE_COMMON_KEY
};

struct queue
{
    pj_ioqueue_key_t	    *key;
    enum ioqueue_event_type  event_type;
};

/*
 * This describes the I/O queue.
 */
struct pj_ioqueue_t
{
    DECLARE_COMMON_IOQUEUE

    unsigned		max, count;
    //pj_ioqueue_key_t	hlist;
    pj_ioqueue_key_t	active_list;    
    int			epfd;
    //struct epoll_event *events;
    //struct queue       *queue;

#if PJ_IOQUEUE_HAS_SAFE_UNREG
    pj_mutex_t	       *ref_cnt_mutex;
    pj_ioqueue_key_t	closing_list;
    pj_ioqueue_key_t	free_list;
#endif
};

/* Include implementation for common abstraction after we declare
 * pj_ioqueue_key_t and pj_ioqueue_t.
 */
#include "ioqueue_common_abs.c"

#if PJ_IOQUEUE_HAS_SAFE_UNREG
/* Scan closing keys to be put to free list again */
static void scan_closing_keys(pj_ioqueue_t *ioqueue);
#endif

/*
 * pj_ioqueue_name()
 */
PJ_DEF(const char*) pj_ioqueue_name(void)
{
#if defined(PJ_LINUX_KERNEL) && PJ_LINUX_KERNEL!=0
	return "epoll-kernel";
#else
	return "epoll";
#endif
}

/*
 * pj_ioqueue_create()
 *
 * Create select ioqueue.
 */
PJ_DEF(pj_status_t) pj_ioqueue_create( pj_pool_t *pool, 
                                       pj_size_t max_fd,
                                       pj_ioqueue_t **p_ioqueue)
{
    pj_ioqueue_t *ioqueue;
    pj_status_t rc;
    pj_lock_t *lock;
    int i;

    /* Check that arguments are valid. */
    PJ_ASSERT_RETURN(pool != NULL && p_ioqueue != NULL && 
                     max_fd > 0, PJ_EINVAL);

    /* Check that size of pj_ioqueue_op_key_t is sufficient */
    PJ_ASSERT_RETURN(sizeof(pj_ioqueue_op_key_t)-sizeof(void*) >=
                     sizeof(union operation_key), PJ_EBUG);

    ioqueue = pj_pool_alloc(pool, sizeof(pj_ioqueue_t));

    ioqueue_init(ioqueue);

    ioqueue->max = max_fd;
    ioqueue->count = 0;
    pj_list_init(&ioqueue->active_list);

#if PJ_IOQUEUE_HAS_SAFE_UNREG
    /* When safe unregistration is used (the default), we pre-create
     * all keys and put them in the free list.
     */

    /* Mutex to protect key's reference counter 
     * We don't want to use key's mutex or ioqueue's mutex because
     * that would create deadlock situation in some cases.
     */
    rc = pj_mutex_create_simple(pool, NULL, &ioqueue->ref_cnt_mutex);
    if (rc != PJ_SUCCESS)
	return rc;


    /* Init key list */
    pj_list_init(&ioqueue->free_list);
    pj_list_init(&ioqueue->closing_list);


    /* Pre-create all keys according to max_fd */
    for ( i=0; i<max_fd; ++i) {
	pj_ioqueue_key_t *key;

	key = PJ_POOL_ALLOC_T(pool, pj_ioqueue_key_t);
	key->ref_count = 0;
	rc = pj_lock_create_recursive_mutex(pool, NULL, &key->lock);
	if (rc != PJ_SUCCESS) {
	    key = ioqueue->free_list.next;
	    while (key != &ioqueue->free_list) {
		pj_lock_destroy(key->lock);
		key = key->next;
	    }
	    pj_mutex_destroy(ioqueue->ref_cnt_mutex);
	    return rc;
	}

	pj_list_push_back(&ioqueue->free_list, key);
    }
#endif

    rc = pj_lock_create_simple_mutex(pool, "ioq%p", &lock);
    if (rc != PJ_SUCCESS)
	return rc;

    rc = pj_ioqueue_set_lock(ioqueue, lock, PJ_TRUE);
    if (rc != PJ_SUCCESS)
        return rc;

    ioqueue->epfd = os_epoll_create(max_fd);
    if (ioqueue->epfd < 0) {
	ioqueue_destroy(ioqueue);
	return PJ_RETURN_OS_ERROR(pj_get_native_os_error());
    }
    
    /*ioqueue->events = pj_pool_calloc(pool, max_fd, sizeof(struct epoll_event));
    PJ_ASSERT_RETURN(ioqueue->events != NULL, PJ_ENOMEM);

    ioqueue->queue = pj_pool_calloc(pool, max_fd, sizeof(struct queue));
    PJ_ASSERT_RETURN(ioqueue->queue != NULL, PJ_ENOMEM);
   */
    PJ_LOG(4, ("pjlib", "epoll I/O Queue created (%p)", ioqueue));

    *p_ioqueue = ioqueue;
    return PJ_SUCCESS;
}

/*
 * pj_ioqueue_destroy()
 *
 * Destroy ioqueue.
 */
PJ_DEF(pj_status_t) pj_ioqueue_destroy(pj_ioqueue_t *ioqueue)
{
    pj_ioqueue_key_t *key;

    PJ_ASSERT_RETURN(ioqueue, PJ_EINVAL);
    PJ_ASSERT_RETURN(ioqueue->epfd > 0, PJ_EINVALIDOP);

    pj_lock_acquire(ioqueue->lock);
    os_close(ioqueue->epfd);
    ioqueue->epfd = 0;

#if PJ_IOQUEUE_HAS_SAFE_UNREG
    /* Destroy reference counters */
    key = ioqueue->active_list.next;
    while (key != &ioqueue->active_list) {
	pj_lock_destroy(key->lock);
	key = key->next;
    }

    key = ioqueue->closing_list.next;
    while (key != &ioqueue->closing_list) {
	pj_lock_destroy(key->lock);
	key = key->next;
    }

    key = ioqueue->free_list.next;
    while (key != &ioqueue->free_list) {
	pj_lock_destroy(key->lock);
	key = key->next;
    }

    pj_mutex_destroy(ioqueue->ref_cnt_mutex);
#endif
    return ioqueue_destroy(ioqueue);
}

/*
 * pj_ioqueue_register_sock()
 *
 * Register a socket to ioqueue.
 */
PJ_DEF(pj_status_t) pj_ioqueue_register_sock2(pj_pool_t *pool,
					      pj_ioqueue_t *ioqueue,
					      pj_sock_t sock,
					      pj_grp_lock_t *grp_lock,
					      void *user_data,
					      const pj_ioqueue_callback *cb,
                                              pj_ioqueue_key_t **p_key)
{
    pj_ioqueue_key_t *key = NULL;
    pj_uint32_t value;
    struct epoll_event ev;
    int status;
    pj_status_t rc = PJ_SUCCESS;
    
    PJ_ASSERT_RETURN(pool && ioqueue && sock != PJ_INVALID_SOCKET &&
                     cb && p_key, PJ_EINVAL);

    pj_lock_acquire(ioqueue->lock);

    if (ioqueue->count >= ioqueue->max) {
        rc = PJ_ETOOMANY;
	TRACE_((THIS_FILE, "pj_ioqueue_register_sock error: too many files"));
	goto on_return;
    }

    /* Set socket to nonblocking. */
    value = 1;
    if ((rc=os_ioctl(sock, FIONBIO, (ioctl_val_type)&value))) {
	TRACE_((THIS_FILE, "pj_ioqueue_register_sock error: ioctl rc=%d", 
                rc));
        rc = pj_get_netos_error();
	goto on_return;
    }

    /* If safe unregistration (PJ_IOQUEUE_HAS_SAFE_UNREG) is used, get
     * the key from the free list. Otherwise allocate a new one. 
     */
#if PJ_IOQUEUE_HAS_SAFE_UNREG

    /* Scan closing_keys first to let them come back to free_list */
    scan_closing_keys(ioqueue);

    pj_assert(!pj_list_empty(&ioqueue->free_list));
    if (pj_list_empty(&ioqueue->free_list)) {
	rc = PJ_ETOOMANY;
	goto on_return;
    }

    key = ioqueue->free_list.next;
    pj_list_erase(key);
#else
    /* Create key. */
    key = (pj_ioqueue_key_t*)pj_pool_zalloc(pool, sizeof(pj_ioqueue_key_t));
#endif

    rc = ioqueue_init_key(pool, ioqueue, key, sock, grp_lock, user_data, cb);
    if (rc != PJ_SUCCESS) {
	key = NULL;
	goto on_return;
    }

    /* Create key's mutex */
 /*   rc = pj_mutex_create_recursive(pool, NULL, &key->mutex);
    if (rc != PJ_SUCCESS) {
	key = NULL;
	goto on_return;
    }
*/
    /* os_epoll_ctl. */
    ev.events = EPOLLIN | EPOLLERR;
    ev.epoll_data = (epoll_data_type)key;
    status = os_epoll_ctl(ioqueue->epfd, EPOLL_CTL_ADD, sock, &ev);
    if (status < 0) {
	rc = pj_get_os_error();
	pj_lock_destroy(key->lock);
	key = NULL;
	TRACE_((THIS_FILE, 
                "pj_ioqueue_register_sock error: os_epoll_ctl rc=%d", 
                status));
	goto on_return;
    }
    
    /* Register */
    pj_list_insert_before(&ioqueue->active_list, key);
    ++ioqueue->count;

    //TRACE_((THIS_FILE, "socket registered, count=%d", ioqueue->count));

on_return:
    if (rc != PJ_SUCCESS) {
	if (key && key->grp_lock)
	    pj_grp_lock_dec_ref_dbg(key->grp_lock, "ioqueue", 0);
    }
    *p_key = key;
    pj_lock_release(ioqueue->lock);
    
    return rc;
}

PJ_DEF(pj_status_t) pj_ioqueue_register_sock( pj_pool_t *pool,
					      pj_ioqueue_t *ioqueue,
					      pj_sock_t sock,
					      void *user_data,
					      const pj_ioqueue_callback *cb,
					      pj_ioqueue_key_t **p_key)
{
    return pj_ioqueue_register_sock2(pool, ioqueue, sock, NULL, user_data,
                                     cb, p_key);
}

#if PJ_IOQUEUE_HAS_SAFE_UNREG
/* Increment key's reference counter */
static void increment_counter(pj_ioqueue_key_t *key)
{
    pj_mutex_lock(key->ioqueue->ref_cnt_mutex);
    ++key->ref_count;
    pj_mutex_unlock(key->ioqueue->ref_cnt_mutex);
}

/* Decrement the key's reference counter, and when the counter reach zero,
 * destroy the key.
 *
 * Note: MUST NOT CALL THIS FUNCTION WHILE HOLDING ioqueue's LOCK.
 */
static void decrement_counter(pj_ioqueue_key_t *key)
{
    pj_lock_acquire(key->ioqueue->lock);
    pj_mutex_lock(key->ioqueue->ref_cnt_mutex);
    --key->ref_count;
    if (key->ref_count == 0) {

	pj_assert(key->closing == 1);
	pj_gettickcount(&key->free_time);
	key->free_time.msec += PJ_IOQUEUE_KEY_FREE_DELAY;
	pj_time_val_normalize(&key->free_time);

	pj_list_erase(key);
	pj_list_push_back(&key->ioqueue->closing_list, key);

    }
    pj_mutex_unlock(key->ioqueue->ref_cnt_mutex);
    pj_lock_release(key->ioqueue->lock);
}
#endif

/*
 * pj_ioqueue_unregister()
 *
 * Unregister handle from ioqueue.
 */
PJ_DEF(pj_status_t) pj_ioqueue_unregister( pj_ioqueue_key_t *key)
{
    pj_ioqueue_t *ioqueue;
    struct epoll_event ev;
    int status;
    
    PJ_ASSERT_RETURN(key != NULL, PJ_EINVAL);

    ioqueue = key->ioqueue;

    /* Lock the key to make sure no callback is simultaneously modifying
     * the key. We need to lock the key before ioqueue here to prevent
     * deadlock.
     */
    pj_ioqueue_lock_key(key);

    /* Also lock ioqueue */
    pj_lock_acquire(ioqueue->lock);

    pj_assert(ioqueue->count > 0);
    --ioqueue->count;
#if !PJ_IOQUEUE_HAS_SAFE_UNREG
    pj_list_erase(key);
#endif

    ev.events = 0;
    ev.epoll_data = (epoll_data_type)key;
    status = os_epoll_ctl( ioqueue->epfd, EPOLL_CTL_DEL, key->fd, &ev);
    if (status != 0) {
	pj_status_t rc = pj_get_os_error();
	pj_lock_release(ioqueue->lock);
	return rc;
    }

    /* Destroy the key. */
    pj_sock_close(key->fd);

    pj_lock_release(ioqueue->lock);


#if PJ_IOQUEUE_HAS_SAFE_UNREG
    /* Mark key is closing. */
    key->closing = 1;

    /* Decrement counter. */
    decrement_counter(key);

    /* Done. */
    if (key->grp_lock) {
	/* just dec_ref and unlock. we will set grp_lock to NULL
	 * elsewhere */
	pj_grp_lock_t *grp_lock = key->grp_lock;
	// Don't set grp_lock to NULL otherwise the other thread
	// will crash. Just leave it as dangling pointer, but this
	// should be safe
	//key->grp_lock = NULL;
	pj_grp_lock_dec_ref_dbg(grp_lock, "ioqueue", 0);
	pj_grp_lock_release(grp_lock);
    } else {
	pj_ioqueue_unlock_key(key);
    }
#else
    if (key->grp_lock) {
	/* set grp_lock to NULL and unlock */
	pj_grp_lock_t *grp_lock = key->grp_lock;
	// Don't set grp_lock to NULL otherwise the other thread
	// will crash. Just leave it as dangling pointer, but this
	// should be safe
	//key->grp_lock = NULL;
	pj_grp_lock_dec_ref_dbg(grp_lock, "ioqueue", 0);
	pj_grp_lock_release(grp_lock);
    } else {
	pj_ioqueue_unlock_key(key);
    }

    pj_lock_destroy(key->lock);
#endif

    return PJ_SUCCESS;
}

/* ioqueue_remove_from_set()
 * This function is called from ioqueue_dispatch_event() to instruct
 * the ioqueue to remove the specified descriptor from ioqueue's descriptor
 * set for the specified event.
 */
static void ioqueue_remove_from_set( pj_ioqueue_t *ioqueue,
                                     pj_ioqueue_key_t *key, 
                                     enum ioqueue_event_type event_type)
{
    if (event_type == WRITEABLE_EVENT) {
	struct epoll_event ev;

	ev.events = EPOLLIN | EPOLLERR;
	ev.epoll_data = (epoll_data_type)key;
	os_epoll_ctl( ioqueue->epfd, EPOLL_CTL_MOD, key->fd, &ev);
    }	
}

/*
 * ioqueue_add_to_set()
 * This function is called from pj_ioqueue_recv(), pj_ioqueue_send() etc
 * to instruct the ioqueue to add the specified handle to ioqueue's descriptor
 * set for the specified event.
 */
static void ioqueue_add_to_set( pj_ioqueue_t *ioqueue,
                                pj_ioqueue_key_t *key,
                                enum ioqueue_event_type event_type )
{
    if (event_type == WRITEABLE_EVENT) {
	struct epoll_event ev;

	ev.events = EPOLLIN | EPOLLOUT | EPOLLERR;
	ev.epoll_data = (epoll_data_type)key;
	os_epoll_ctl( ioqueue->epfd, EPOLL_CTL_MOD, key->fd, &ev);
    }	
}

#if PJ_IOQUEUE_HAS_SAFE_UNREG
/* Scan closing keys to be put to free list again */
static void scan_closing_keys(pj_ioqueue_t *ioqueue)
{
    pj_time_val now;
    pj_ioqueue_key_t *h;

    pj_gettickcount(&now);
    h = ioqueue->closing_list.next;
    while (h != &ioqueue->closing_list) {
	pj_ioqueue_key_t *next = h->next;

	pj_assert(h->closing != 0);

	if (PJ_TIME_VAL_GTE(now, h->free_time)) {
	    pj_list_erase(h);
	    // Don't set grp_lock to NULL otherwise the other thread
	    // will crash. Just leave it as dangling pointer, but this
	    // should be safe
	    //h->grp_lock = NULL;
	    pj_list_push_back(&ioqueue->free_list, h);
	}
	h = next;
    }
}
#endif

/*
 * pj_ioqueue_poll()
 *
 */
PJ_DEF(int) pj_ioqueue_poll( pj_ioqueue_t *ioqueue, const pj_time_val *timeout)
{
    int i, count, processed;
    int msec;
    //struct epoll_event *events = ioqueue->events;
    //struct queue *queue = ioqueue->queue;
    struct epoll_event events[PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL];
    struct queue queue[PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL];
    pj_timestamp t1, t2;
    
    PJ_CHECK_STACK();

    msec = timeout ? PJ_TIME_VAL_MSEC(*timeout) : 9000;

    TRACE_((THIS_FILE, "start os_epoll_wait, msec=%d", msec));
    pj_get_timestamp(&t1);
 
    //count = os_epoll_wait( ioqueue->epfd, events, ioqueue->max, msec);
    count = os_epoll_wait( ioqueue->epfd, events, PJ_IOQUEUE_MAX_EVENTS_IN_SINGLE_POLL, msec);
    if (count == 0) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
    /* Check the closing keys only when there's no activity and when there are
     * pending closing keys.
     */
    if (count == 0 && !pj_list_empty(&ioqueue->closing_list)) {
	pj_lock_acquire(ioqueue->lock);
	scan_closing_keys(ioqueue);
	pj_lock_release(ioqueue->lock);
    }
#endif
	TRACE_((THIS_FILE, "os_epoll_wait timed out"));
	return count;
    }
    else if (count < 0) {
	TRACE_((THIS_FILE, "os_epoll_wait error"));
	return -pj_get_netos_error();
    }

    pj_get_timestamp(&t2);
    TRACE_((THIS_FILE, "os_epoll_wait returns %d, time=%d usec",
		       count, pj_elapsed_usec(&t1, &t2)));

    /* Lock ioqueue. */
    pj_lock_acquire(ioqueue->lock);

    for (processed=0, i=0; i<count; ++i) {
	pj_ioqueue_key_t *h = (pj_ioqueue_key_t*)(epoll_data_type)
				events[i].epoll_data;

	TRACE_((THIS_FILE, "event %d: events=%d", i, events[i].events));

	/*
	 * Check readability.
	 */
	if ((events[i].events & EPOLLIN) && 
	    (key_has_pending_read(h) || key_has_pending_accept(h)) && !IS_CLOSING(h) ) {

#if PJ_IOQUEUE_HAS_SAFE_UNREG
	    increment_counter(h);
#endif
	    queue[processed].key = h;
	    queue[processed].event_type = READABLE_EVENT;
	    ++processed;
	    continue;
	}

	/*
	 * Check for writeability.
	 */
	if ((events[i].events & EPOLLOUT) && key_has_pending_write(h) && !IS_CLOSING(h)) {

#if PJ_IOQUEUE_HAS_SAFE_UNREG
	    increment_counter(h);
#endif
	    queue[processed].key = h;
	    queue[processed].event_type = WRITEABLE_EVENT;
	    ++processed;
	    continue;
	}

#if PJ_HAS_TCP
	/*
	 * Check for completion of connect() operation.
	 */
	if ((events[i].events & EPOLLOUT) && (h->connecting) && !IS_CLOSING(h)) {

#if PJ_IOQUEUE_HAS_SAFE_UNREG
	    increment_counter(h);
#endif
	    queue[processed].key = h;
	    queue[processed].event_type = WRITEABLE_EVENT;
	    ++processed;
	    continue;
	}
#endif /* PJ_HAS_TCP */

	/*
	 * Check for error condition.
	 */
	if ((events[i].events & EPOLLERR) && !IS_CLOSING(h)) {
	    /*
	     * We need to handle this exception event.  If it's related to us
	     * connecting, report it as such.  If not, just report it as a
	     * read event and the higher layers will handle it.
	     */
	    if (h->connecting) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
		increment_counter(h);
#endif
		queue[processed].key = h;
		queue[processed].event_type = EXCEPTION_EVENT;
		++processed;
	    } else if (key_has_pending_read(h) || key_has_pending_accept(h)) {
#if PJ_IOQUEUE_HAS_SAFE_UNREG
		increment_counter(h);
#endif
		queue[processed].key = h;
		queue[processed].event_type = READABLE_EVENT;
		++processed;
	    }
	    continue;
	}
    }
    for (i=0; i<processed; ++i) {
	if (queue[i].key->grp_lock)
	    pj_grp_lock_add_ref_dbg(queue[i].key->grp_lock, "ioqueue", 0);
    }

    PJ_RACE_ME(5);

    pj_lock_release(ioqueue->lock);

    PJ_RACE_ME(5);

    /* Now process the events. */
    for (i=0; i<processed; ++i) {
	switch (queue[i].event_type) {
        case READABLE_EVENT:
            ioqueue_dispatch_read_event(ioqueue, queue[i].key);
            break;
        case WRITEABLE_EVENT:
            ioqueue_dispatch_write_event(ioqueue, queue[i].key);
            break;
        case EXCEPTION_EVENT:
            ioqueue_dispatch_exception_event(ioqueue, queue[i].key);
            break;
        case NO_EVENT:
            pj_assert(!"Invalid event!");
            break;
        }

#if PJ_IOQUEUE_HAS_SAFE_UNREG
	decrement_counter(queue[i].key);
#endif

	if (queue[i].key->grp_lock)
	    pj_grp_lock_dec_ref_dbg(queue[i].key->grp_lock,
	                            "ioqueue", 0);
    }

    /* Special case:
     * When epoll returns > 0 but no descriptors are actually set!
     */
    if (count > 0 && !processed && msec > 0) {
	pj_thread_sleep(msec);
    }

    pj_get_timestamp(&t1);
    TRACE_((THIS_FILE, "ioqueue_poll() returns %d, time=%d usec",
		       processed, pj_elapsed_usec(&t2, &t1)));

    return processed;
}