static pj_bool_t on_rx_msg(pjsip_rx_data* rdata)
{
// Before we start, get a timestamp. This will track the time from
// receiving a message to forwarding it on (or rejecting it).
struct rx_msg_qe qe;
if (clock_gettime(CLOCK_MONOTONIC, &qe.rx_time) != 0)
{
LOG_ERROR("Failed to get receive timestamp: %s", strerror(errno));
return PJ_TRUE;
}
// Do logging.
local_log_rx_msg(rdata);
sas_log_rx_msg(rdata);
// Clone the message and queue it to a scheduler thread.
pjsip_rx_data* clone_rdata;
pj_status_t status = pjsip_rx_data_clone(rdata, 0, &clone_rdata);
if (status != PJ_SUCCESS)
{
// Failed to clone the message, so drop it.
LOG_ERROR("Failed to clone incoming message (%s)", PJUtils::pj_status_to_string(status).c_str());
return PJ_TRUE;
}
// Make sure the trail identifier is passed across.
set_trail(clone_rdata, get_trail(rdata));
// @TODO - need to think about back-pressure mechanisms. For example,
// should we have a maximum depth of queue and drop messages after that?
// May be better to hold on to the message until the queue has space - this
// will force back pressure on the particular TCP connection. Or should we
// have a queue per transport and round-robin them?
LOG_DEBUG("Queuing cloned received message %p for worker threads", clone_rdata);
qe.rdata = clone_rdata;
rx_msg_q.push(qe);
// return TRUE to flag that we have absorbed the incoming message.
return PJ_TRUE;
}
static struct rx_task_data *rx_task_data_create(pjsip_rx_data *rdata,
struct ast_sip_endpoint *endpoint,
struct ast_sip_aor *aor)
{
struct rx_task_data *task_data = ao2_alloc(
sizeof(*task_data), rx_task_data_destroy);
if (!task_data) {
return NULL;
}
pjsip_rx_data_clone(rdata, 0, &task_data->rdata);
task_data->endpoint = endpoint;
ao2_ref(task_data->endpoint, +1);
task_data->aor = aor;
ao2_ref(task_data->aor, +1);
return task_data;
}
static pj_bool_t distributor(pjsip_rx_data *rdata)
{
pjsip_dialog *dlg;
struct distributor_dialog_data *dist = NULL;
struct ast_taskprocessor *serializer = NULL;
pjsip_rx_data *clone;
if (!ast_test_flag(&ast_options, AST_OPT_FLAG_FULLY_BOOTED)) {
/*
* Ignore everything until we are fully booted. Let the
* peer retransmit messages until we are ready.
*/
return PJ_TRUE;
}
dlg = find_dialog(rdata);
if (dlg) {
ast_debug(3, "Searching for serializer associated with dialog %s for %s\n",
dlg->obj_name, pjsip_rx_data_get_info(rdata));
dist = ao2_find(dialog_associations, dlg, OBJ_SEARCH_KEY);
if (dist) {
ao2_lock(dist);
serializer = ao2_bump(dist->serializer);
ao2_unlock(dist);
if (serializer) {
ast_debug(3, "Found serializer %s associated with dialog %s\n",
ast_taskprocessor_name(serializer), dlg->obj_name);
}
}
}
if (serializer) {
/* We have a serializer so we know where to send the message. */
} else if (rdata->msg_info.msg->type == PJSIP_RESPONSE_MSG) {
ast_debug(3, "No dialog serializer for %s. Using request transaction as basis.\n",
pjsip_rx_data_get_info(rdata));
serializer = find_request_serializer(rdata);
if (!serializer) {
/*
* Pick a serializer for the unmatched response.
* We couldn't determine what serializer originally
* sent the request or the serializer is gone.
*/
serializer = ast_sip_get_distributor_serializer(rdata);
}
} else if (!pjsip_method_cmp(&rdata->msg_info.msg->line.req.method, &pjsip_cancel_method)
|| !pjsip_method_cmp(&rdata->msg_info.msg->line.req.method, &pjsip_bye_method)) {
/* We have a BYE or CANCEL request without a serializer. */
pjsip_endpt_respond_stateless(ast_sip_get_pjsip_endpoint(), rdata,
PJSIP_SC_CALL_TSX_DOES_NOT_EXIST, NULL, NULL, NULL);
ao2_cleanup(dist);
return PJ_TRUE;
} else {
if (ast_taskprocessor_alert_get()) {
/*
* When taskprocessors get backed up, there is a good chance that
* we are being overloaded and need to defer adding new work to
* the system. To defer the work we will ignore the request and
* rely on the peer's transport layer to retransmit the message.
* We usually work off the overload within a few seconds. The
* alternative is to send back a 503 response to these requests
* and be done with it.
*/
ast_debug(3, "Taskprocessor overload alert: Ignoring '%s'.\n",
pjsip_rx_data_get_info(rdata));
ao2_cleanup(dist);
return PJ_TRUE;
}
/* Pick a serializer for the out-of-dialog request. */
serializer = ast_sip_get_distributor_serializer(rdata);
}
if (pjsip_rx_data_clone(rdata, 0, &clone) != PJ_SUCCESS) {
ast_taskprocessor_unreference(serializer);
ao2_cleanup(dist);
return PJ_TRUE;
}
if (dist) {
ao2_lock(dist);
clone->endpt_info.mod_data[endpoint_mod.id] = ao2_bump(dist->endpoint);
ao2_unlock(dist);
ao2_cleanup(dist);
}
if (ast_sip_push_task(serializer, distribute, clone)) {
ao2_cleanup(clone->endpt_info.mod_data[endpoint_mod.id]);
pjsip_rx_data_free_cloned(clone);
}
ast_taskprocessor_unreference(serializer);
return PJ_TRUE;
}
static pj_bool_t on_rx_msg(pjsip_rx_data* rdata)
{
// Do logging.
local_log_rx_msg(rdata);
sas_log_rx_msg(rdata);
requests_counter->increment();
// Check whether the request should be processed
if (!(load_monitor->admit_request()) &&
(rdata->msg_info.msg->type == PJSIP_REQUEST_MSG) &&
(rdata->msg_info.msg->line.req.method.id != PJSIP_ACK_METHOD))
{
// Discard non-ACK requests if there are no available tokens.
// Respond statelessly with a 503 Service Unavailable, including a
// Retry-After header with a zero length timeout.
LOG_DEBUG("Rejected request due to overload");
pjsip_cid_hdr* cid = (pjsip_cid_hdr*)rdata->msg_info.cid;
SAS::TrailId trail = get_trail(rdata);
SAS::Marker start_marker(trail, MARKER_ID_START, 1u);
SAS::report_marker(start_marker);
SAS::Event event(trail, SASEvent::SIP_OVERLOAD, 0);
event.add_static_param(load_monitor->get_target_latency());
event.add_static_param(load_monitor->get_current_latency());
event.add_static_param(load_monitor->get_rate_limit());
SAS::report_event(event);
PJUtils::report_sas_to_from_markers(trail, rdata->msg_info.msg);
if ((rdata->msg_info.msg->line.req.method.id == PJSIP_REGISTER_METHOD) ||
((pjsip_method_cmp(&rdata->msg_info.msg->line.req.method, pjsip_get_subscribe_method())) == 0) ||
((pjsip_method_cmp(&rdata->msg_info.msg->line.req.method, pjsip_get_notify_method())) == 0))
{
// Omit the Call-ID for these requests, as the same Call-ID can be
// reused over a long period of time and produce huge SAS trails.
PJUtils::mark_sas_call_branch_ids(trail, NULL, rdata->msg_info.msg);
}
else
{
PJUtils::mark_sas_call_branch_ids(trail, cid, rdata->msg_info.msg);
}
SAS::Marker end_marker(trail, MARKER_ID_END, 1u);
SAS::report_marker(end_marker);
pjsip_retry_after_hdr* retry_after = pjsip_retry_after_hdr_create(rdata->tp_info.pool, 0);
PJUtils::respond_stateless(stack_data.endpt,
rdata,
PJSIP_SC_SERVICE_UNAVAILABLE,
NULL,
(pjsip_hdr*)retry_after,
NULL);
// We no longer terminate TCP connections on overload as the shutdown has
// to wait for existing transactions to end and therefore it takes too
// long to get feedback to the downstream node. We expect downstream nodes
// to rebalance load if possible triggered by receipt of the 503 responses.
overload_counter->increment();
return PJ_TRUE;
}
// Check that the worker threads are not all deadlocked.
if (rx_msg_q.is_deadlocked())
{
// The queue has not been serviced for sufficiently long to imply that
// all the worker threads are deadlock, so exit the process so it will be
// restarted.
LOG_ERROR("Detected worker thread deadlock - exiting");
abort();
}
// Before we start, get a timestamp. This will track the time from
// receiving a message to forwarding it on (or rejecting it).
struct rx_msg_qe qe;
qe.stop_watch.start();
// Notify the connection tracker that the transport is active.
connection_tracker->connection_active(rdata->tp_info.transport);
// Clone the message and queue it to a scheduler thread.
pjsip_rx_data* clone_rdata;
pj_status_t status = pjsip_rx_data_clone(rdata, 0, &clone_rdata);
if (status != PJ_SUCCESS)
{
// Failed to clone the message, so drop it.
LOG_ERROR("Failed to clone incoming message (%s)", PJUtils::pj_status_to_string(status).c_str());
return PJ_TRUE;
}
// Make sure the trail identifier is passed across.
set_trail(clone_rdata, get_trail(rdata));
// @TODO - need to think about back-pressure mechanisms. For example,
// should we have a maximum depth of queue and drop messages after that?
// May be better to hold on to the message until the queue has space - this
// will force back pressure on the particular TCP connection. Or should we
// have a queue per transport and round-robin them?
LOG_DEBUG("Queuing cloned received message %p for worker threads", clone_rdata);
qe.rdata = clone_rdata;
// Track the current queue size
queue_size_accumulator->accumulate(rx_msg_q.size());
rx_msg_q.push(qe);
// return TRUE to flag that we have absorbed the incoming message.
return PJ_TRUE;
}