static isc_result_t get_client(ns_clientmgr_t *manager, ns_interface_t *ifp,
dns_dispatch_t *disp, bool tcp);
static isc_result_t get_worker(ns_clientmgr_t *manager, ns_interface_t *ifp,
- isc_socket_t *sock);
+ isc_socket_t *sock, ns_client_t *oldclient);
static inline bool
-allowed(isc_netaddr_t *addr, dns_name_t *signer, isc_netaddr_t *ecs_addr,
- uint8_t ecs_addrlen, uint8_t *ecs_scope, dns_acl_t *acl);
+allowed(isc_netaddr_t *addr, dns_name_t *signer,
+ isc_netaddr_t *ecs_addr, uint8_t ecs_addrlen,
+ uint8_t *ecs_scope, dns_acl_t *acl)
static void compute_cookie(ns_client_t *client, uint32_t when,
uint32_t nonce, const unsigned char *secret,
isc_buffer_t *buf);
*/
INSIST(client->recursionquota == NULL);
INSIST(client->newstate <= NS_CLIENTSTATE_READY);
- if (client->nreads > 0)
+
+ if (client->nreads > 0) {
dns_tcpmsg_cancelread(&client->tcpmsg);
+ }
+
if (client->nreads != 0) {
/* Still waiting for read cancel completion. */
return (true);
dns_tcpmsg_invalidate(&client->tcpmsg);
client->tcpmsg_valid = false;
}
+
if (client->tcpsocket != NULL) {
CTRACE("closetcp");
isc_socket_detach(&client->tcpsocket);
+
+ if (client->tcpactive) {
+ LOCK(&client->interface->lock);
+ INSIST(client->interface->ntcpactive > 0);
+ client->interface->ntcpactive--;
+ UNLOCK(&client->interface->lock);
+ client->tcpactive = false;
+ }
}
if (client->tcpquota != NULL) {
- isc_quota_detach(&client->tcpquota);
- } else {
/*
- * We went over quota with this client, we don't
- * want to restart listening unless this is the
- * last client on this interface, which is
- * checked later.
+ * If we are not in a pipeline group, or
+ * we are the last client in the group, detach from
+ * tcpquota; otherwise, transfer the quota to
+ * another client in the same group.
*/
- if (TCP_CLIENT(client)) {
- client->mortal = true;
+ if (!ISC_LINK_LINKED(client, glink) ||
+ (client->glink.next == NULL &&
+ client->glink.prev == NULL))
+ {
+ isc_quota_detach(&client->tcpquota);
+ } else if (client->glink.next != NULL) {
+ INSIST(client->glink.next->tcpquota == NULL);
+ client->glink.next->tcpquota = client->tcpquota;
+ client->tcpquota = NULL;
+ } else {
+ INSIST(client->glink.prev->tcpquota == NULL);
+ client->glink.prev->tcpquota = client->tcpquota;
+ client->tcpquota = NULL;
}
}
+ /*
+ * Unlink from pipeline group.
+ */
+ if (ISC_LINK_LINKED(client, glink)) {
+ if (client->glink.next != NULL) {
+ client->glink.next->glink.prev =
+ client->glink.prev;
+ }
+ if (client->glink.prev != NULL) {
+ client->glink.prev->glink.next =
+ client->glink.next;
+ }
+ ISC_LINK_INIT(client, glink);
+ }
+
if (client->timerset) {
(void)isc_timer_reset(client->timer,
isc_timertype_inactive,
* that already. Check whether this client needs to remain
* active and force it to go inactive if not.
*
- * UDP clients go inactive at this point, but TCP clients
- * may remain active if we have fewer active TCP client
- * objects than desired due to an earlier quota exhaustion.
+ * UDP clients go inactive at this point, but a TCP client
+ * will needs to remain active if no other clients are
+ * listening for TCP requests on this interface, to
+ * prevent this interface from going nonresponsive.
*/
if (client->mortal && TCP_CLIENT(client) && !ns_g_clienttest) {
LOCK(&client->interface->lock);
- if (client->interface->ntcpcurrent <
- client->interface->ntcptarget)
+ if (client->interface->ntcpaccepting == 0) {
client->mortal = false;
+ }
UNLOCK(&client->interface->lock);
}
* queue for recycling.
*/
if (client->mortal) {
- if (client->newstate > NS_CLIENTSTATE_INACTIVE)
+ if (client->newstate > NS_CLIENTSTATE_INACTIVE) {
client->newstate = NS_CLIENTSTATE_INACTIVE;
+ }
}
if (NS_CLIENTSTATE_READY == client->newstate) {
if (TCP_CLIENT(client)) {
client_accept(client);
- } else
+ } else {
client_udprecv(client);
+ }
client->newstate = NS_CLIENTSTATE_MAX;
return (true);
}
/*
* We are trying to enter the inactive state.
*/
- if (client->naccepts > 0)
+ if (client->naccepts > 0) {
isc_socket_cancel(client->tcplistener, client->task,
ISC_SOCKCANCEL_ACCEPT);
+ }
/* Still waiting for accept cancel completion. */
- if (! (client->naccepts == 0))
+ if (! (client->naccepts == 0)) {
return (true);
+ }
/* Accept cancel is complete. */
- if (client->nrecvs > 0)
+ if (client->nrecvs > 0) {
isc_socket_cancel(client->udpsocket, client->task,
ISC_SOCKCANCEL_RECV);
+ }
/* Still waiting for recv cancel completion. */
- if (! (client->nrecvs == 0))
+ if (! (client->nrecvs == 0)) {
return (true);
+ }
/* Still waiting for control event to be delivered */
- if (client->nctls > 0)
+ if (client->nctls > 0) {
return (true);
-
- /* Deactivate the client. */
- if (client->interface)
- ns_interface_detach(&client->interface);
+ }
INSIST(client->naccepts == 0);
INSIST(client->recursionquota == NULL);
- if (client->tcplistener != NULL)
+ if (client->tcplistener != NULL) {
isc_socket_detach(&client->tcplistener);
- if (client->udpsocket != NULL)
+ if (client->tcpactive) {
+ LOCK(&client->interface->lock);
+ INSIST(client->interface->ntcpactive > 0);
+ client->interface->ntcpactive--;
+ UNLOCK(&client->interface->lock);
+ client->tcpactive = false;
+ }
+ }
+ if (client->udpsocket != NULL) {
isc_socket_detach(&client->udpsocket);
+ }
- if (client->dispatch != NULL)
+ /* Deactivate the client. */
+ if (client->interface != NULL) {
+ ns_interface_detach(&client->interface);
+ }
+
+ if (client->dispatch != NULL) {
dns_dispatch_detach(&client->dispatch);
+ }
client->attributes = 0;
client->mortal = false;
client->newstate = NS_CLIENTSTATE_MAX;
if (!ns_g_clienttest && manager != NULL &&
!manager->exiting)
+ {
ISC_QUEUE_PUSH(manager->inactive, client,
ilink);
- if (client->needshutdown)
+ }
+ if (client->needshutdown) {
isc_task_shutdown(client->task);
+ }
return (true);
}
}
}
}
-
/*%
* The client's task has received a shutdown event.
*/
/*
* Pipeline TCP query processing.
*/
- if (client->message->opcode != dns_opcode_query)
+ if (client->message->opcode != dns_opcode_query) {
client->pipelined = false;
+ }
if (TCP_CLIENT(client) && client->pipelined) {
- result = isc_quota_reserve(&ns_g_server->tcpquota);
- if (result == ISC_R_SUCCESS)
- result = ns_client_replace(client);
+ result = ns_client_replace(client);
if (result != ISC_R_SUCCESS) {
- ns_client_log(client, NS_LOGCATEGORY_CLIENT,
- NS_LOGMODULE_CLIENT, ISC_LOG_WARNING,
- "no more TCP clients(read): %s",
- isc_result_totext(result));
client->pipelined = false;
}
}
client->filter_aaaa = dns_aaaa_ok;
#endif
client->needshutdown = ns_g_clienttest;
+ client->tcpactive = false;
ISC_EVENT_INIT(&client->ctlevent, sizeof(client->ctlevent), 0, NULL,
NS_EVENT_CLIENTCONTROL, client_start, client, client,
client->formerrcache.id = 0;
ISC_LINK_INIT(client, link);
ISC_LINK_INIT(client, rlink);
+ ISC_LINK_INIT(client, glink);
ISC_QLINK_INIT(client, ilink);
client->keytag = NULL;
client->keytag_len = 0;
INSIST(client->state == NS_CLIENTSTATE_READY);
+ /*
+ * The accept() was successful and we're now establishing a new
+ * connection. We need to make note of it in the client and
+ * interface objects so client objects can do the right thing
+ * when going inactive in exit_check() (see comments in
+ * client_accept() for details).
+ */
INSIST(client->naccepts == 1);
client->naccepts--;
LOCK(&client->interface->lock);
- INSIST(client->interface->ntcpcurrent > 0);
- client->interface->ntcpcurrent--;
+ INSIST(client->interface->ntcpaccepting > 0);
+ client->interface->ntcpaccepting--;
UNLOCK(&client->interface->lock);
/*
NS_LOGMODULE_CLIENT, ISC_LOG_DEBUG(3),
"accept failed: %s",
isc_result_totext(nevent->result));
+ if (client->tcpquota != NULL) {
+ isc_quota_detach(&client->tcpquota);
+ }
}
if (exit_check(client))
* deny service to legitimate TCP clients.
*/
client->pipelined = false;
- result = isc_quota_attach(&ns_g_server->tcpquota,
- &client->tcpquota);
- if (result == ISC_R_SUCCESS)
- result = ns_client_replace(client);
- if (result != ISC_R_SUCCESS) {
- ns_client_log(client, NS_LOGCATEGORY_CLIENT,
- NS_LOGMODULE_CLIENT, ISC_LOG_WARNING,
- "no more TCP clients(accept): %s",
- isc_result_totext(result));
- } else if (ns_g_server->keepresporder == NULL ||
- !allowed(&netaddr, NULL, NULL, 0, NULL,
- ns_g_server->keepresporder)) {
+ result = ns_client_replace(client);
+ if (result == ISC_R_SUCCESS &&
+ (client->sctx->keepresporder == NULL ||
+ !allowed(&netaddr, NULL, NULL, 0, NULL,
+ ns_g_server->keepresporder)))
+ {
client->pipelined = true;
}
CTRACE("accept");
+ /*
+ * The tcpquota object can only be simultaneously referenced a
+ * pre-defined number of times; this is configured by 'tcp-clients'
+ * in named.conf. If we can't attach to it here, that means the TCP
+ * client quota has been exceeded.
+ */
+ result = isc_quota_attach(&client->sctx->tcpquota,
+ &client->tcpquota);
+ if (result != ISC_R_SUCCESS) {
+ bool exit;
+
+ ns_client_log(client, NS_LOGCATEGORY_CLIENT,
+ NS_LOGMODULE_CLIENT, ISC_LOG_DEBUG(1),
+ "no more TCP clients: %s",
+ isc_result_totext(result));
+
+ /*
+ * We have exceeded the system-wide TCP client
+ * quota. But, we can't just block this accept
+ * in all cases, because if we did, a heavy TCP
+ * load on other interfaces might cause this
+ * interface to be starved, with no clients able
+ * to accept new connections.
+ *
+ * So, we check here to see if any other client
+ * is already servicing TCP queries on this
+ * interface (whether accepting, reading, or
+ * processing).
+ *
+ * If so, then it's okay *not* to call
+ * accept - we can let this client to go inactive
+ * and the other one handle the next connection
+ * when it's ready.
+ *
+ * But if not, then we need to be a little bit
+ * flexible about the quota. We allow *one* extra
+ * TCP client through, to ensure we're listening on
+ * every interface.
+ *
+ * (Note: In practice this means that the *real*
+ * TCP client quota is tcp-clients plus the number
+ * of interfaces.)
+ */
+ LOCK(&client->interface->lock);
+ exit = (client->interface->ntcpactive > 0);
+ UNLOCK(&client->interface->lock);
+
+ if (exit) {
+ client->newstate = NS_CLIENTSTATE_INACTIVE;
+ (void)exit_check(client);
+ return;
+ }
+ }
+
+ /*
+ * By incrementing the interface's ntcpactive counter we signal
+ * that there is at least one client servicing TCP queries for the
+ * interface.
+ *
+ * We also make note of the fact in the client itself with the
+ * tcpactive flag. This ensures proper accounting by preventing
+ * us from accidentally incrementing or decrementing ntcpactive
+ * more than once per client object.
+ */
+ if (!client->tcpactive) {
+ LOCK(&client->interface->lock);
+ client->interface->ntcpactive++;
+ UNLOCK(&client->interface->lock);
+ client->tcpactive = true;
+ }
+
result = isc_socket_accept(client->tcplistener, client->task,
client_newconn, client);
if (result != ISC_R_SUCCESS) {
- UNEXPECTED_ERROR(__FILE__, __LINE__,
- "isc_socket_accept() failed: %s",
- isc_result_totext(result));
/*
* XXXRTH What should we do? We're trying to accept but
* it didn't work. If we just give up, then TCP
*
* For now, we just go idle.
*/
+ UNEXPECTED_ERROR(__FILE__, __LINE__,
+ "isc_socket_accept() failed: %s",
+ isc_result_totext(result));
+ if (client->tcpquota != NULL) {
+ isc_quota_detach(&client->tcpquota);
+ }
return;
}
+
+ /*
+ * The client's 'naccepts' counter indicates that this client has
+ * called accept() and is waiting for a new connection. It should
+ * never exceed 1.
+ */
INSIST(client->naccepts == 0);
client->naccepts++;
+
+ /*
+ * The interface's 'ntcpaccepting' counter is incremented when
+ * any client calls accept(), and decremented in client_newconn()
+ * once the connection is established.
+ *
+ * When the client object is shutting down after handling a TCP
+ * request (see exit_check()), it looks to see whether this value is
+ * non-zero. If so, that means another client has already called
+ * accept() and is waiting to establish the next connection, which
+ * means the first client is free to go inactive. Otherwise,
+ * the first client must come back and call accept() again; this
+ * guarantees there will always be at least one client listening
+ * for new TCP connections on each interface.
+ */
LOCK(&client->interface->lock);
- client->interface->ntcpcurrent++;
+ client->interface->ntcpaccepting++;
UNLOCK(&client->interface->lock);
}
tcp = TCP_CLIENT(client);
if (tcp && client->pipelined) {
result = get_worker(client->manager, client->interface,
- client->tcpsocket);
+ client->tcpsocket, client);
} else {
result = get_client(client->manager, client->interface,
client->dispatch, tcp);
}
- if (result != ISC_R_SUCCESS)
+ if (result != ISC_R_SUCCESS) {
return (result);
+ }
/*
* The responsibility for listening for new requests is hereby
client->attributes |= NS_CLIENTATTR_TCP;
isc_socket_attach(ifp->tcpsocket,
&client->tcplistener);
+
} else {
isc_socket_t *sock;
}
static isc_result_t
-get_worker(ns_clientmgr_t *manager, ns_interface_t *ifp, isc_socket_t *sock)
+get_worker(ns_clientmgr_t *manager, ns_interface_t *ifp, isc_socket_t *sock,
+ ns_client_t *oldclient)
{
isc_result_t result = ISC_R_SUCCESS;
isc_event_t *ev;
MTRACE("get worker");
REQUIRE(manager != NULL);
+ REQUIRE(oldclient != NULL);
if (manager->exiting)
return (ISC_R_SHUTTINGDOWN);
ns_interface_attach(ifp, &client->interface);
client->newstate = client->state = NS_CLIENTSTATE_WORKING;
INSIST(client->recursionquota == NULL);
- client->tcpquota = &ns_g_server->tcpquota;
+
+ /*
+ * Transfer TCP quota to the new client.
+ */
+ INSIST(client->tcpquota == NULL);
+ INSIST(oldclient->tcpquota != NULL);
+ client->tcpquota = oldclient->tcpquota;
+ oldclient->tcpquota = NULL;
+
+ /*
+ * Link to a pipeline group, creating it if needed.
+ */
+ if (!ISC_LINK_LINKED(oldclient, glink)) {
+ oldclient->glink.next = NULL;
+ oldclient->glink.prev = NULL;
+ }
+ client->glink.next = oldclient->glink.next;
+ client->glink.prev = oldclient;
+ if (oldclient->glink.next != NULL) {
+ oldclient->glink.next->glink.prev = client;
+ }
+ oldclient->glink.next = client;
client->dscp = ifp->dscp;
(void)isc_socket_getpeername(client->tcpsocket, &client->peeraddr);
client->peeraddr_valid = true;
+ LOCK(&client->interface->lock);
+ client->interface->ntcpactive++;
+ UNLOCK(&client->interface->lock);
+
+ client->tcpactive = true;
+
INSIST(client->tcpmsg_valid == false);
dns_tcpmsg_init(client->mctx, client->tcpsocket, &client->tcpmsg);
client->tcpmsg_valid = true;
projects / thirdparty / bind9.git / commitdiff
