int poll_loop(struct npool *nsp, int msec_timeout) {
int results_left = 0;
int event_msecs; /* msecs before an event goes off */
int combined_msecs;
int sock_err = 0;
struct poll_engine_info *pinfo = (struct poll_engine_info *)nsp->engine_data;
assert(msec_timeout >= -1);
if (nsp->events_pending == 0)
return 0; /* No need to wait on 0 events ... */
do {
struct nevent *nse;
nsock_log_debug_all(nsp, "wait for events");
nse = next_expirable_event(nsp);
if (!nse)
event_msecs = -1; /* None of the events specified a timeout */
else
event_msecs = MAX(0, TIMEVAL_MSEC_SUBTRACT(nse->timeout, nsock_tod));
#if HAVE_PCAP
#ifndef PCAP_CAN_DO_SELECT
/* Force a low timeout when capturing packets on systems where
* the pcap descriptor is not select()able. */
if (gh_list_count(&nsp->pcap_read_events) > 0)
if (event_msecs > PCAP_POLL_INTERVAL)
event_msecs = PCAP_POLL_INTERVAL;
#endif
#endif
/* We cast to unsigned because we want -1 to be very high (since it means no
* timeout) */
combined_msecs = MIN((unsigned)event_msecs, (unsigned)msec_timeout);
#if HAVE_PCAP
#ifndef PCAP_CAN_DO_SELECT
/* do non-blocking read on pcap devices that doesn't support select()
* If there is anything read, just leave this loop. */
if (pcap_read_on_nonselect(nsp)) {
/* okay, something was read. */
} else
#endif
#endif
{
results_left = Poll(pinfo->events, pinfo->max_fd + 1, combined_msecs);
if (results_left == -1)
sock_err = socket_errno();
}
gettimeofday(&nsock_tod, NULL); /* Due to poll delay */
} while (results_left == -1 && sock_err == EINTR); /* repeat only if signal occurred */
if (results_left == -1 && sock_err != EINTR) {
nsock_log_error(nsp, "nsock_loop error %d: %s", sock_err, socket_strerror(sock_err));
nsp->errnum = sock_err;
return -1;
}
iterate_through_event_lists(nsp);
return 1;
}
/* Adds an event to the appropriate nsp event list, handles housekeeping such as
* adjusting the descriptor select/poll lists, registering the timeout value,
* etc. */
void nsock_pool_add_event(struct npool *nsp, struct nevent *nse) {
nsock_log_debug("NSE #%lu: Adding event (timeout in %ldms)",
nse->id,
(long)TIMEVAL_MSEC_SUBTRACT(nse->timeout, nsock_tod));
nsp->events_pending++;
if (!nse->event_done && nse->timeout.tv_sec) {
/* This event is expirable, add it to the queue */
gh_heap_push(&nsp->expirables, &nse->expire);
}
/* Now we do the event type specific actions */
switch (nse->type) {
case NSE_TYPE_CONNECT:
case NSE_TYPE_CONNECT_SSL:
if (!nse->event_done) {
assert(nse->iod->sd >= 0);
socket_count_read_inc(nse->iod);
socket_count_write_inc(nse->iod);
update_events(nse->iod, nsp, EV_READ|EV_WRITE|EV_EXCEPT, EV_NONE);
}
iod_add_event(nse->iod, nse);
break;
case NSE_TYPE_READ:
if (!nse->event_done) {
assert(nse->iod->sd >= 0);
socket_count_read_inc(nse->iod);
update_events(nse->iod, nsp, EV_READ, EV_NONE);
#if HAVE_OPENSSL
if (nse->iod->ssl)
nse->sslinfo.ssl_desire = SSL_ERROR_WANT_READ;
#endif
}
iod_add_event(nse->iod, nse);
break;
case NSE_TYPE_WRITE:
if (!nse->event_done) {
assert(nse->iod->sd >= 0);
socket_count_write_inc(nse->iod);
update_events(nse->iod, nsp, EV_WRITE, EV_NONE);
#if HAVE_OPENSSL
if (nse->iod->ssl)
nse->sslinfo.ssl_desire = SSL_ERROR_WANT_WRITE;
#endif
}
iod_add_event(nse->iod, nse);
break;
case NSE_TYPE_TIMER:
/* nothing to do */
break;
#if HAVE_PCAP
case NSE_TYPE_PCAP_READ: {
mspcap *mp = (mspcap *)nse->iod->pcap;
assert(mp);
if (mp->pcap_desc >= 0) { /* pcap descriptor present */
if (!nse->event_done) {
socket_count_readpcap_inc(nse->iod);
update_events(nse->iod, nsp, EV_READ, EV_NONE);
}
nsock_log_debug_all("PCAP NSE #%lu: Adding event to READ_EVENTS", nse->id);
#if PCAP_BSD_SELECT_HACK
/* when using BSD hack we must do pcap_next() after select().
* Let's insert this pcap to bot queues, to selectable and nonselectable.
* This will result in doing pcap_next_ex() just before select() */
nsock_log_debug_all("PCAP NSE #%lu: Adding event to PCAP_READ_EVENTS", nse->id);
#endif
} else {
/* pcap isn't selectable. Add it to pcap-specific queue. */
nsock_log_debug_all("PCAP NSE #%lu: Adding event to PCAP_READ_EVENTS", nse->id);
}
iod_add_event(nse->iod, nse);
break;
}
#endif
default:
fatal("Unknown nsock event type (%d)", nse->type);
}
/* It can happen that the event already completed. In which case we can
* already deliver it, even though we're probably not inside nsock_loop(). */
if (nse->event_done) {
event_dispatch_and_delete(nsp, nse, 1);
update_first_events(nse);
nevent_unref(nsp, nse);
}
}
void process_event(struct npool *nsp, gh_list_t *evlist, struct nevent *nse, int ev) {
int match_r = 0, match_w = 0;
#if HAVE_OPENSSL
int desire_r = 0, desire_w = 0;
#endif
nsock_log_debug_all("Processing event %lu (timeout in %ldms, done=%d)",
nse->id,
(long)TIMEVAL_MSEC_SUBTRACT(nse->timeout, nsock_tod),
nse->event_done);
if (!nse->event_done) {
switch (nse->type) {
case NSE_TYPE_CONNECT:
case NSE_TYPE_CONNECT_SSL:
if (ev != EV_NONE)
handle_connect_result(nsp, nse, NSE_STATUS_SUCCESS);
if (event_timedout(nse))
handle_connect_result(nsp, nse, NSE_STATUS_TIMEOUT);
break;
case NSE_TYPE_READ:
match_r = ev & EV_READ;
match_w = ev & EV_WRITE;
#if HAVE_OPENSSL
desire_r = nse->sslinfo.ssl_desire == SSL_ERROR_WANT_READ;
desire_w = nse->sslinfo.ssl_desire == SSL_ERROR_WANT_WRITE;
if (nse->iod->ssl && ((desire_r && match_r) || (desire_w && match_w)))
handle_read_result(nsp, nse, NSE_STATUS_SUCCESS);
else
#endif
if (!nse->iod->ssl && match_r)
handle_read_result(nsp, nse, NSE_STATUS_SUCCESS);
if (event_timedout(nse))
handle_read_result(nsp, nse, NSE_STATUS_TIMEOUT);
break;
case NSE_TYPE_WRITE:
match_r = ev & EV_READ;
match_w = ev & EV_WRITE;
#if HAVE_OPENSSL
desire_r = nse->sslinfo.ssl_desire == SSL_ERROR_WANT_READ;
desire_w = nse->sslinfo.ssl_desire == SSL_ERROR_WANT_WRITE;
if (nse->iod->ssl && ((desire_r && match_r) || (desire_w && match_w)))
handle_write_result(nsp, nse, NSE_STATUS_SUCCESS);
else
#endif
if (!nse->iod->ssl && match_w)
handle_write_result(nsp, nse, NSE_STATUS_SUCCESS);
if (event_timedout(nse))
handle_write_result(nsp, nse, NSE_STATUS_TIMEOUT);
break;
case NSE_TYPE_TIMER:
if (event_timedout(nse))
handle_timer_result(nsp, nse, NSE_STATUS_SUCCESS);
break;
#if HAVE_PCAP
case NSE_TYPE_PCAP_READ:{
nsock_log_debug_all("PCAP iterating %lu", nse->id);
if (ev & EV_READ) {
/* buffer empty? check it! */
if (fs_length(&(nse->iobuf)) == 0)
do_actual_pcap_read(nse);
}
/* if already received something */
if (fs_length(&(nse->iobuf)) > 0)
handle_pcap_read_result(nsp, nse, NSE_STATUS_SUCCESS);
if (event_timedout(nse))
handle_pcap_read_result(nsp, nse, NSE_STATUS_TIMEOUT);
#if PCAP_BSD_SELECT_HACK
/* If event occurred, and we're in BSD_HACK mode, then this event was added
* to two queues. read_event and pcap_read_event
* Of course we should destroy it only once.
* I assume we're now in read_event, so just unlink this event from
* pcap_read_event */
if (((mspcap *)nse->iod->pcap)->pcap_desc >= 0
&& nse->event_done
&& evlist == &nsp->read_events) {
/* event is done, list is read_events and we're in BSD_HACK mode.
* So unlink event from pcap_read_events */
update_first_events(nse);
gh_list_remove(&nsp->pcap_read_events, &nse->nodeq_pcap);
nsock_log_debug_all("PCAP NSE #%lu: Removing event from PCAP_READ_EVENTS",
nse->id);
}
if (((mspcap *)nse->iod->pcap)->pcap_desc >= 0
&& nse->event_done
&& evlist == &nsp->pcap_read_events) {
update_first_events(nse);
gh_list_remove(&nsp->read_events, &nse->nodeq_io);
nsock_log_debug_all("PCAP NSE #%lu: Removing event from READ_EVENTS",
nse->id);
}
#endif
break;
}
#endif
default:
fatal("Event has unknown type (%d)", nse->type);
}
}
if (nse->event_done) {
/* Security sanity check: don't return a functional SSL iod without
* setting an SSL data structure. */
if (nse->type == NSE_TYPE_CONNECT_SSL && nse->status == NSE_STATUS_SUCCESS)
assert(nse->iod->ssl != NULL);
nsock_log_debug_all("NSE #%lu: Sending event", nse->id);
/* WooHoo! The event is ready to be sent */
event_dispatch_and_delete(nsp, nse, 1);
}
}
void process_iod_events(struct npool *nsp, struct niod *nsi, int ev) {
int i = 0;
/* store addresses of the pointers to the first elements of each kind instead
* of storing the values, as a connect can add a read for instance */
gh_lnode_t **start_elems[] = {
&nsi->first_connect,
&nsi->first_read,
&nsi->first_write,
#if HAVE_PCAP
&nsi->first_pcap_read,
#endif
NULL
};
gh_list_t *evlists[] = {
&nsp->connect_events,
&nsp->read_events,
&nsp->write_events,
#if HAVE_PCAP
&nsp->pcap_read_events,
#endif
NULL
};
assert(nsp == nsi->nsp);
nsock_log_debug_all("Processing events on IOD %lu (ev=%d)", nsi->id, ev);
/* We keep the events separate because we want to handle them in the
* order: connect => read => write => timer for several reasons:
*
* 1) Makes sure we have gone through all the net i/o events before
* a timer expires (would be a shame to timeout after the data was
* available but before we delivered the events
*
* 2) The connect() results often lead to a read or write that can be
* processed in the same cycle. In the same way, read() often
* leads to write().
*/
for (i = 0; evlists[i] != NULL; i++) {
gh_lnode_t *current, *next, *last;
/* for each list, get the last event and don't look past it as an event
* could add another event in the same list and so on... */
last = gh_list_last_elem(evlists[i]);
for (current = *start_elems[i];
current != NULL && gh_lnode_prev(current) != last;
current = next) {
struct nevent *nse;
#if HAVE_PCAP
if (evlists[i] == &nsi->nsp->pcap_read_events)
nse = lnode_nevent2(current);
else
#endif
nse = lnode_nevent(current);
/* events are grouped by IOD. Break if we're done with the events for the
* current IOD */
if (nse->iod != nsi)
break;
process_event(nsp, evlists[i], nse, ev);
next = gh_lnode_next(current);
if (nse->event_done) {
/* event is done, remove it from the event list and update IOD pointers
* to the first events of each kind */
update_first_events(nse);
gh_list_remove(evlists[i], current);
gh_list_append(&nsp->free_events, &nse->nodeq_io);
if (nse->timeout.tv_sec)
gh_heap_remove(&nsp->expirables, &nse->expire);
}
}
}
}
int kqueue_loop(struct npool *nsp, int msec_timeout) {
int results_left = 0;
int event_msecs; /* msecs before an event goes off */
int combined_msecs;
struct timespec ts, *ts_p;
int sock_err = 0;
struct kqueue_engine_info *kinfo = (struct kqueue_engine_info *)nsp->engine_data;
assert(msec_timeout >= -1);
if (nsp->events_pending == 0)
return 0; /* No need to wait on 0 events ... */
if (gh_list_count(&nsp->active_iods) > kinfo->evlen) {
kinfo->evlen = gh_list_count(&nsp->active_iods) * 2;
kinfo->events = (struct kevent *)safe_realloc(kinfo->events, kinfo->evlen * sizeof(struct kevent));
}
do {
struct nevent *nse;
nsock_log_debug_all(nsp, "wait for events");
nse = next_expirable_event(nsp);
if (!nse)
event_msecs = -1; /* None of the events specified a timeout */
else
event_msecs = MAX(0, TIMEVAL_MSEC_SUBTRACT(nse->timeout, nsock_tod));
#if HAVE_PCAP
#ifndef PCAP_CAN_DO_SELECT
/* Force a low timeout when capturing packets on systems where
* the pcap descriptor is not select()able. */
if (gh_list_count(&nsp->pcap_read_events) > 0)
if (event_msecs > PCAP_POLL_INTERVAL)
event_msecs = PCAP_POLL_INTERVAL;
#endif
#endif
/* We cast to unsigned because we want -1 to be very high (since it means no
* timeout) */
combined_msecs = MIN((unsigned)event_msecs, (unsigned)msec_timeout);
/* Set up the timeval pointer we will give to kevent() */
memset(&ts, 0, sizeof(struct timespec));
if (combined_msecs >= 0) {
ts.tv_sec = combined_msecs / 1000;
ts.tv_nsec = (combined_msecs % 1000) * 1000000L;
ts_p = &ts;
} else {
ts_p = NULL;
}
#if HAVE_PCAP
#ifndef PCAP_CAN_DO_SELECT
/* do non-blocking read on pcap devices that doesn't support select()
* If there is anything read, just leave this loop. */
if (pcap_read_on_nonselect(nsp)) {
/* okay, something was read. */
} else
#endif
#endif
{
results_left = kevent(kinfo->kqfd, NULL, 0, kinfo->events, kinfo->evlen, ts_p);
if (results_left == -1)
sock_err = socket_errno();
}
gettimeofday(&nsock_tod, NULL); /* Due to kevent delay */
} while (results_left == -1 && sock_err == EINTR); /* repeat only if signal occurred */
if (results_left == -1 && sock_err != EINTR) {
nsock_log_error(nsp, "nsock_loop error %d: %s", sock_err, socket_strerror(sock_err));
nsp->errnum = sock_err;
return -1;
}
iterate_through_event_lists(nsp, results_left);
return 1;
}
/* This does the actual logistics of requesting a connection. It is shared
* by nsock_connect_tcp and nsock_connect_ssl, among others */
void nsock_connect_internal(struct npool *ms, struct nevent *nse, int type, int proto, struct sockaddr_storage *ss, size_t sslen,
unsigned short port) {
struct sockaddr_in *sin;
#if HAVE_IPV6
struct sockaddr_in6 *sin6;
#endif
struct niod *iod = nse->iod;
if (iod->px_ctx /* proxy enabled */
&& proto == IPPROTO_TCP /* restrict proxying to TCP connections */
&& (nse->handler != nsock_proxy_ev_dispatch)) { /* for reentrancy */
struct proxy_node *current;
nsock_log_debug_all("TCP connection request (EID %lu) redirected through proxy chain",
(long)nse->id);
current = iod->px_ctx->px_current;
assert(current != NULL);
memcpy(&iod->px_ctx->target_ss, ss, sslen);
iod->px_ctx->target_sslen = sslen;
iod->px_ctx->target_port = port;
ss = ¤t->ss;
sslen = current->sslen;
port = current->port;
iod->px_ctx->target_handler = nse->handler;
nse->handler = nsock_proxy_ev_dispatch;
iod->px_ctx->target_ev_type = nse->type;
nse->type = NSE_TYPE_CONNECT;
}
sin = (struct sockaddr_in *)ss;
#if HAVE_IPV6
sin6 = (struct sockaddr_in6 *)ss;
#endif
/* Now it is time to actually attempt the connection */
if (nsock_make_socket(ms, iod, ss->ss_family, type, proto) == -1) {
nse->event_done = 1;
nse->status = NSE_STATUS_ERROR;
nse->errnum = socket_errno();
} else {
if (ss->ss_family == AF_INET) {
sin->sin_port = htons(port);
}
#if HAVE_IPV6
else if (ss->ss_family == AF_INET6) {
sin6->sin6_port = htons(port);
}
#endif
#if HAVE_SYS_UN_H
else if (ss->ss_family == AF_UNIX) {
}
#endif
else {
fatal("Unknown address family %d\n", ss->ss_family);
}
assert(sslen <= sizeof(iod->peer));
if (&iod->peer != ss)
memcpy(&iod->peer, ss, sslen);
iod->peerlen = sslen;
if (ms->engine->io_operations->iod_connect(ms, iod->sd, (struct sockaddr *)ss, sslen) == -1) {
int err = socket_errno();
if ((proto == IPPROTO_UDP) || (err != EINPROGRESS && err != EAGAIN)) {
nse->event_done = 1;
nse->status = NSE_STATUS_ERROR;
nse->errnum = err;
}
}
/* The callback handle_connect_result handles the connection once it completes. */
}
}
int select_loop(struct npool *nsp, int msec_timeout) {
int results_left = 0;
int event_msecs; /* msecs before an event goes off */
int combined_msecs;
int sock_err = 0;
struct timeval select_tv;
struct timeval *select_tv_p;
struct select_engine_info *sinfo = (struct select_engine_info *)nsp->engine_data;
assert(msec_timeout >= -1);
if (nsp->events_pending == 0)
return 0; /* No need to wait on 0 events ... */
do {
struct nevent *nse;
nsock_log_debug_all(nsp, "wait for events");
nse = next_expirable_event(nsp);
if (!nse)
event_msecs = -1; /* None of the events specified a timeout */
else
event_msecs = MAX(0, TIMEVAL_MSEC_SUBTRACT(nse->timeout, nsock_tod));
#if HAVE_PCAP
#ifndef PCAP_CAN_DO_SELECT
/* Force a low timeout when capturing packets on systems where
* the pcap descriptor is not select()able. */
if (gh_list_count(&nsp->pcap_read_events))
if (event_msecs > PCAP_POLL_INTERVAL)
event_msecs = PCAP_POLL_INTERVAL;
#endif
#endif
/* We cast to unsigned because we want -1 to be very high (since it means no
* timeout) */
combined_msecs = MIN((unsigned)event_msecs, (unsigned)msec_timeout);
/* Set up the timeval pointer we will give to select() */
memset(&select_tv, 0, sizeof(select_tv));
if (combined_msecs > 0) {
select_tv.tv_sec = combined_msecs / 1000;
select_tv.tv_usec = (combined_msecs % 1000) * 1000;
select_tv_p = &select_tv;
} else if (combined_msecs == 0) {
/* we want the tv_sec and tv_usec to be zero but they already are from bzero */
select_tv_p = &select_tv;
} else {
assert(combined_msecs == -1);
select_tv_p = NULL;
}
#if HAVE_PCAP
#ifndef PCAP_CAN_DO_SELECT
/* do non-blocking read on pcap devices that doesn't support select()
* If there is anything read, just leave this loop. */
if (pcap_read_on_nonselect(nsp)) {
/* okay, something was read. */
} else
#endif
#endif
{
/* Set up the descriptors for select */
sinfo->fds_results_r = sinfo->fds_master_r;
sinfo->fds_results_w = sinfo->fds_master_w;
sinfo->fds_results_x = sinfo->fds_master_x;
results_left = fselect(sinfo->max_sd + 1, &sinfo->fds_results_r,
&sinfo->fds_results_w, &sinfo->fds_results_x, select_tv_p);
if (results_left == -1)
sock_err = socket_errno();
}
gettimeofday(&nsock_tod, NULL); /* Due to select delay */
} while (results_left == -1 && sock_err == EINTR); /* repeat only if signal occurred */
if (results_left == -1 && sock_err != EINTR) {
nsock_log_error(nsp, "nsock_loop error %d: %s", sock_err, socket_strerror(sock_err));
nsp->errnum = sock_err;
return -1;
}
iterate_through_event_lists(nsp);
return 1;
}
/* An internal function for cancelling an event when you already have a pointer
* to the struct nevent (use nsock_event_cancel if you just have an ID). The
* event_list passed in should correspond to the type of the event. For example,
* with NSE_TYPE_READ, you would pass in &nsp->read_events;. elem is the list
* element in event_list which holds the event. Pass a nonzero for notify if
* you want the program owning the event to be notified that it has been
* cancelled */
int nevent_delete(struct npool *nsp, struct nevent *nse, gh_list_t *event_list,
gh_lnode_t *elem, int notify) {
if (nse->event_done) {
/* This event has already been marked for death somewhere else -- it will be
* gone soon (and if we try to kill it now all hell will break loose due to
* reentrancy. */
return 0;
}
nsock_log_info("%s on event #%li (type %s)", __func__, nse->id,
nse_type2str(nse->type));
/* Now that we found the event... we go through the motions of cleanly
* cancelling it */
switch (nse->type) {
case NSE_TYPE_CONNECT:
case NSE_TYPE_CONNECT_SSL:
handle_connect_result(nsp, nse, NSE_STATUS_CANCELLED);
break;
case NSE_TYPE_READ:
handle_read_result(nsp, nse, NSE_STATUS_CANCELLED);
break;
case NSE_TYPE_WRITE:
handle_write_result(nsp, nse, NSE_STATUS_CANCELLED);
break;
case NSE_TYPE_TIMER:
handle_timer_result(nsp, nse, NSE_STATUS_CANCELLED);
break;
#if HAVE_PCAP
case NSE_TYPE_PCAP_READ:
handle_pcap_read_result(nsp, nse, NSE_STATUS_CANCELLED);
break;
#endif
default:
fatal("Invalid nsock event type (%d)", nse->type);
}
assert(nse->event_done);
if (nse->timeout.tv_sec)
gh_heap_remove(&nsp->expirables, &nse->expire);
if (event_list) {
update_first_events(nse);
gh_list_remove(event_list, elem);
}
gh_list_append(&nsp->free_events, &nse->nodeq_io);
nsock_log_debug_all("NSE #%lu: Removing event from list", nse->id);
#if HAVE_PCAP
#if PCAP_BSD_SELECT_HACK
if (nse->type == NSE_TYPE_PCAP_READ) {
nsock_log_debug_all("PCAP NSE #%lu: CANCEL TEST pcap=%p read=%p curr=%p sd=%i",
nse->id, &nsp->pcap_read_events, &nsp->read_events,
event_list,((mspcap *)nse->iod->pcap)->pcap_desc);
/* If event occurred, and we're in BSD_HACK mode, then this event was added to
* two queues. read_event and pcap_read_event Of course we should
* destroy it only once. I assume we're now in read_event, so just unlink
* this event from pcap_read_event */
if (((mspcap *)nse->iod->pcap)->pcap_desc >= 0 && event_list == &nsp->read_events) {
/* event is done, list is read_events and we're in BSD_HACK mode. So unlink
* event from pcap_read_events */
gh_list_remove(&nsp->pcap_read_events, &nse->nodeq_pcap);
nsock_log_debug_all("PCAP NSE #%lu: Removing event from PCAP_READ_EVENTS", nse->id);
}
if (((mspcap *)nse->iod->pcap)->pcap_desc >= 0 && event_list == &nsp->pcap_read_events) {
/* event is done, list is read_events and we're in BSD_HACK mode.
* So unlink event from read_events */
gh_list_remove(&nsp->read_events, &nse->nodeq_io);
nsock_log_debug_all("PCAP NSE #%lu: Removing event from READ_EVENTS", nse->id);
}
}
#endif
#endif
event_dispatch_and_delete(nsp, nse, notify);
return 1;
}
