Commit 49394512 authored by Evan Hunt's avatar Evan Hunt
Browse files

refactor tcpquota and pipeline refs; allow special-case overrun in isc_quota

- if the TCP quota has been exceeded but there are no clients listening
  for new connections on the interface, we can now force attachment to the
  quota using isc_quota_force(), instead of carrying on with the quota not
  attached.
- the TCP client quota is now referenced via a reference-counted
  'ns_tcpconn' object, one of which is created whenever a client begins
  listening for new connections, and attached to by members of that
  client's pipeline group. when the last reference to the tcpconn
  object is detached, it is freed and the TCP quota slot is released.
- reduce code duplication by adding mark_tcp_active() function.
- convert counters to atomic.

(cherry picked from commit 7e822237)
parent e965d5f1
......@@ -244,8 +244,7 @@ static void client_start(isc_task_t *task, isc_event_t *event);
static void client_request(isc_task_t *task, isc_event_t *event);
static void ns_client_dumpmessage(ns_client_t *client, const char *reason);
static isc_result_t get_client(ns_clientmgr_t *manager, ns_interface_t *ifp,
dns_dispatch_t *disp, ns_client_t *oldclient,
bool tcp);
dns_dispatch_t *disp, bool tcp);
static isc_result_t get_worker(ns_clientmgr_t *manager, ns_interface_t *ifp,
isc_socket_t *sock, ns_client_t *oldclient);
static inline bool
......@@ -301,16 +300,32 @@ ns_client_settimeout(ns_client_t *client, unsigned int seconds) {
}
/*%
* Allocate a reference counter that will track the number of client structures
* using the TCP connection that 'client' called accept() for. This counter
* will be shared between all client structures associated with this TCP
* connection.
* Allocate a reference-counted object that will maintain a single pointer to
* the (also reference-counted) TCP client quota, shared between all the
* clients processing queries on a single TCP connection, so that all
* clients sharing the one socket will together consume only one slot in
* the 'tcp-clients' quota.
*/
static void
pipeline_init(ns_client_t *client) {
isc_refcount_t *refs;
static isc_result_t
tcpconn_init(ns_client_t *client, bool force) {
isc_result_t result;
isc_quota_t *quota = NULL;
ns_tcpconn_t *tconn = NULL;
REQUIRE(client->pipeline_refs == NULL);
REQUIRE(client->tcpconn == NULL);
/*
* Try to attach to the quota first, so we won't pointlessly
* allocate memory for a tcpconn object if we can't get one.
*/
if (force) {
result = isc_quota_force(&ns_g_server->tcpquota, &quota);
} else {
result = isc_quota_attach(&ns_g_server->tcpquota, &quota);
}
if (result != ISC_R_SUCCESS) {
return (result);
}
/*
* A global memory context is used for the allocation as different
......@@ -320,78 +335,80 @@ pipeline_init(ns_client_t *client) {
* contention here is expected to be negligible, given that this code
* is only executed for TCP connections.
*/
refs = isc_mem_allocate(ns_g_mctx, sizeof(*refs));
isc_refcount_init(refs, 1);
client->pipeline_refs = refs;
tconn = isc_mem_allocate(ns_g_mctx, sizeof(*tconn));
isc_refcount_init(&tconn->refs, 1);
tconn->tcpquota = quota;
quota = NULL;
tconn->pipelined = false;
client->tcpconn = tconn;
return (ISC_R_SUCCESS);
}
/*%
* Increase the count of client structures using the TCP connection that
* 'source' is associated with and put a pointer to that count in 'target',
* thus associating it with the same TCP connection.
* Increase the count of client structures sharing the TCP connection
* that 'source' is associated with; add a pointer to the same tcpconn
* to 'target', thus associating it with the same TCP connection.
*/
static void
pipeline_attach(ns_client_t *source, ns_client_t *target) {
tcpconn_attach(ns_client_t *source, ns_client_t *target) {
int refs;
REQUIRE(source->pipeline_refs != NULL);
REQUIRE(target->pipeline_refs == NULL);
REQUIRE(source->tcpconn != NULL);
REQUIRE(target->tcpconn == NULL);
REQUIRE(source->tcpconn->pipelined);
isc_refcount_increment(source->pipeline_refs, &refs);
isc_refcount_increment(&source->tcpconn->refs, &refs);
INSIST(refs > 1);
target->pipeline_refs = source->pipeline_refs;
target->tcpconn = source->tcpconn;
}
/*%
* Decrease the count of client structures using the TCP connection that
* Decrease the count of client structures sharing the TCP connection that
* 'client' is associated with. If this is the last client using this TCP
* connection, free the reference counter and return true; otherwise, return
* false.
* connection, we detach from the TCP quota and free the tcpconn
* object. Either way, client->tcpconn is set to NULL.
*/
static bool
pipeline_detach(ns_client_t *client) {
isc_refcount_t *refcount;
static void
tcpconn_detach(ns_client_t *client) {
ns_tcpconn_t *tconn = NULL;
int refs;
REQUIRE(client->pipeline_refs != NULL);
refcount = client->pipeline_refs;
client->pipeline_refs = NULL;
REQUIRE(client->tcpconn != NULL);
isc_refcount_decrement(refcount, refs);
tconn = client->tcpconn;
client->tcpconn = NULL;
isc_refcount_decrement(&tconn->refs, &refs);
if (refs == 0) {
isc_mem_free(ns_g_mctx, refs);
return (true);
isc_quota_detach(&tconn->tcpquota);
isc_mem_free(ns_g_mctx, tconn);
}
return (false);
}
/*
* Detach a client from the TCP client quota if appropriate, and set
* the quota pointer to NULL.
*
* Sometimes when the TCP client quota is exhausted but there are no other
* clients servicing the interface, a client will be allowed to continue
* running despite not having been attached to the quota. In this event,
* the TCP quota was never attached to the client, so when the client (or
* associated pipeline group) shuts down, the quota must NOT be detached.
/*%
* Mark a client as active and increment the interface's 'ntcpactive'
* counter, as a signal that there is at least one client servicing
* TCP queries for the interface. If we reach the TCP client quota at
* some point, this will be used to determine whether a quota overrun
* should be permitted.
*
* Otherwise, if the quota pointer is set, it should be detached. If not
* set at all, we just return without doing anything.
* Marking the client active with the 'tcpactive' flag ensures proper
* accounting, by preventing us from incrementing or decrementing
* 'ntcpactive' more than once per client.
*/
static void
tcpquota_disconnect(ns_client_t *client) {
if (client->tcpquota == NULL) {
return;
}
if (client->tcpattached) {
isc_quota_detach(&client->tcpquota);
client->tcpattached = false;
} else {
client->tcpquota = NULL;
mark_tcp_active(ns_client_t *client, bool active) {
if (active && !client->tcpactive) {
isc_atomic_xadd(&client->interface->ntcpactive, 1);
client->tcpactive = active;
} else if (!active && client->tcpactive) {
uint32_t old =
isc_atomic_xadd(&client->interface->ntcpactive, -1);
INSIST(old > 0);
client->tcpactive = active;
}
}
......@@ -484,7 +501,8 @@ exit_check(ns_client_t *client) {
INSIST(client->recursionquota == NULL);
if (NS_CLIENTSTATE_READING == client->newstate) {
if (!client->pipelined) {
INSIST(client->tcpconn != NULL);
if (!client->tcpconn->pipelined) {
client_read(client);
client->newstate = NS_CLIENTSTATE_MAX;
return (true); /* We're done. */
......@@ -507,8 +525,8 @@ exit_check(ns_client_t *client) {
dns_tcpmsg_cancelread(&client->tcpmsg);
}
if (client->nreads != 0) {
/* Still waiting for read cancel completion. */
/* Still waiting for read cancel completion. */
if (client->nreads > 0) {
return (true);
}
......@@ -518,43 +536,45 @@ exit_check(ns_client_t *client) {
}
/*
* Detach from pipeline group and from TCP client quota,
* if appropriate.
* Soon the client will be ready to accept a new TCP
* connection or UDP request, but we may have enough
* clients doing that already. Check whether this client
* needs to remain active and allow it go inactive if
* not.
*
* - If no pipeline group is active, attempt to
* detach from the TCP client quota.
* UDP clients always go inactive at this point, but a TCP
* client may need to stay active and return to READY
* state if no other clients are available to listen
* for TCP requests on this interface.
*
* - If a pipeline group is active, detach from it;
* if the return code indicates that there no more
* clients left if this pipeline group, we also detach
* from the TCP client quota.
*
* - Otherwise we don't try to detach, we just set the
* TCP quota pointer to NULL if it wasn't NULL already.
*
* tcpquota_disconnect() will set tcpquota to NULL, either
* by detaching it or by assignment, depending on the
* needs of the client. See the comments on that function
* for further information.
* Regardless, if we're going to FREED state, that means
* the system is shutting down and we don't need to
* retain clients.
*/
if (client->pipeline_refs == NULL || pipeline_detach(client)) {
tcpquota_disconnect(client);
} else {
client->tcpquota = NULL;
client->tcpattached = false;
if (client->mortal && TCP_CLIENT(client) &&
client->newstate != NS_CLIENTSTATE_FREED &&
!ns_g_clienttest &&
isc_atomic_xadd(&client->interface->ntcpaccepting, 0) == 0)
{
/* Nobody else is accepting */
client->mortal = false;
client->newstate = NS_CLIENTSTATE_READY;
}
/*
* Detach from TCP connection and TCP client quota,
* if appropriate. If this is the last reference to
* the TCP connection in our pipeline group, the
* TCP quota slot will be released.
*/
if (client->tcpconn) {
tcpconn_detach(client);
}
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;
}
mark_tcp_active(client, false);
}
if (client->timerset) {
......@@ -567,35 +587,6 @@ exit_check(ns_client_t *client) {
client->peeraddr_valid = false;
client->state = NS_CLIENTSTATE_READY;
INSIST(client->recursionquota == NULL);
/*
* Now the client is ready to accept a new TCP connection
* or UDP request, but we may have enough clients doing
* 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 a TCP client
* may need to remain active and go into ready state if
* no other clients are available to listen for TCP
* requests on this interface or (in the case of pipelined
* clients) to read for additional messages on the current
* connection.
*/
if (client->mortal && TCP_CLIENT(client) && !ns_g_clienttest) {
LOCK(&client->interface->lock);
if ((client->interface->ntcpaccepting == 0 ||
(client->pipelined &&
client->interface->ntcpactive < 2)) &&
client->newstate != NS_CLIENTSTATE_FREED)
{
client->mortal = false;
client->newstate = NS_CLIENTSTATE_READY;
}
UNLOCK(&client->interface->lock);
}
client->pipelined = false;
/*
* We don't need the client; send it to the inactive
......@@ -630,7 +621,7 @@ exit_check(ns_client_t *client) {
}
/* Still waiting for accept cancel completion. */
if (! (client->naccepts == 0)) {
if (client->naccepts > 0) {
return (true);
}
......@@ -641,7 +632,7 @@ exit_check(ns_client_t *client) {
}
/* Still waiting for recv cancel completion. */
if (! (client->nrecvs == 0)) {
if (client->nrecvs > 0) {
return (true);
}
......@@ -654,14 +645,7 @@ exit_check(ns_client_t *client) {
INSIST(client->recursionquota == NULL);
if (client->tcplistener != NULL) {
isc_socket_detach(&client->tcplistener);
if (client->tcpactive) {
LOCK(&client->interface->lock);
INSIST(client->interface->ntcpactive > 0);
client->interface->ntcpactive--;
UNLOCK(&client->interface->lock);
client->tcpactive = false;
}
mark_tcp_active(client, false);
}
if (client->udpsocket != NULL) {
isc_socket_detach(&client->udpsocket);
......@@ -816,7 +800,7 @@ client_start(isc_task_t *task, isc_event_t *event) {
return;
if (TCP_CLIENT(client)) {
if (client->pipelined) {
if (client->tcpconn != NULL) {
client_read(client);
} else {
client_accept(client);
......@@ -2470,6 +2454,7 @@ client_request(isc_task_t *task, isc_event_t *event) {
client->nrecvs--;
} else {
INSIST(TCP_CLIENT(client));
INSIST(client->tcpconn != NULL);
REQUIRE(event->ev_type == DNS_EVENT_TCPMSG);
REQUIRE(event->ev_sender == &client->tcpmsg);
buffer = &client->tcpmsg.buffer;
......@@ -2657,17 +2642,19 @@ client_request(isc_task_t *task, isc_event_t *event) {
/*
* Pipeline TCP query processing.
*/
if (client->message->opcode != dns_opcode_query) {
client->pipelined = false;
if (TCP_CLIENT(client) &&
client->message->opcode != dns_opcode_query)
{
client->tcpconn->pipelined = false;
}
if (TCP_CLIENT(client) && client->pipelined) {
if (TCP_CLIENT(client) && client->tcpconn->pipelined) {
/*
* We're pipelining. Replace the client; the
* the replacement can read the TCP socket looking
* for new messages and this client can process the
* replacement can read the TCP socket looking
* for new messages and this one can process the
* current message asynchronously.
*
* There are now at least three clients using this
* There will now be at least three clients using this
* TCP socket - one accepting new connections,
* one reading an existing connection to get new
* messages, and one answering the message already
......@@ -2675,7 +2662,7 @@ client_request(isc_task_t *task, isc_event_t *event) {
*/
result = ns_client_replace(client);
if (result != ISC_R_SUCCESS) {
client->pipelined = false;
client->tcpconn->pipelined = false;
}
}
......@@ -3231,10 +3218,7 @@ client_create(ns_clientmgr_t *manager, ns_client_t **clientp) {
client->signer = NULL;
dns_name_init(&client->signername, NULL);
client->mortal = false;
client->pipelined = false;
client->pipeline_refs = NULL;
client->tcpquota = NULL;
client->tcpattached = false;
client->tcpconn = NULL;
client->recursionquota = NULL;
client->interface = NULL;
client->peeraddr_valid = false;
......@@ -3339,9 +3323,10 @@ client_read(ns_client_t *client) {
static void
client_newconn(isc_task_t *task, isc_event_t *event) {
isc_result_t result;
ns_client_t *client = event->ev_arg;
isc_socket_newconnev_t *nevent = (isc_socket_newconnev_t *)event;
isc_result_t result;
uint32_t old;
REQUIRE(event->ev_type == ISC_SOCKEVENT_NEWCONN);
REQUIRE(NS_CLIENT_VALID(client));
......@@ -3361,10 +3346,8 @@ client_newconn(isc_task_t *task, isc_event_t *event) {
INSIST(client->naccepts == 1);
client->naccepts--;
LOCK(&client->interface->lock);
INSIST(client->interface->ntcpaccepting > 0);
client->interface->ntcpaccepting--;
UNLOCK(&client->interface->lock);
old = isc_atomic_xadd(&client->interface->ntcpaccepting, -1);
INSIST(old > 0);
/*
* We must take ownership of the new socket before the exit
......@@ -3397,7 +3380,7 @@ client_newconn(isc_task_t *task, isc_event_t *event) {
NS_LOGMODULE_CLIENT, ISC_LOG_DEBUG(3),
"accept failed: %s",
isc_result_totext(nevent->result));
tcpquota_disconnect(client);
tcpconn_detach(client);
}
if (exit_check(client))
......@@ -3435,15 +3418,13 @@ client_newconn(isc_task_t *task, isc_event_t *event) {
* telnetting to port 53 (once per CPU) will
* deny service to legitimate TCP clients.
*/
client->pipelined = false;
result = ns_client_replace(client);
if (result == ISC_R_SUCCESS &&
(ns_g_server->keepresporder == NULL ||
!allowed(&netaddr, NULL, NULL, 0, NULL,
ns_g_server->keepresporder)))
{
pipeline_init(client);
client->pipelined = true;
client->tcpconn->pipelined = true;
}
client_read(client);
......@@ -3460,78 +3441,59 @@ client_accept(ns_client_t *client) {
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.
* Set up a new TCP connection. This means try to attach to the
* TCP client quota (tcp-clients), but fail if we're over quota.
*/
result = isc_quota_attach(&ns_g_server->tcpquota,
&client->tcpquota);
result = tcpconn_init(client, false);
if (result != ISC_R_SUCCESS) {
bool exit;
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 clients
* are already servicing TCP queries on this
* interface (whether accepting, reading, or
* processing). If there are at least two
* (one reading and one processing a request)
* then it's okay *not* to call accept - we
* can let this client go inactive and another
* one will resume accepting when it's done.
*
* If there aren't enough active clients on the
* interface, then we can be a little bit
* flexible about the quota. We'll allow *one*
* extra 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 listening interfaces plus 2.)
*/
LOCK(&client->interface->lock);
exit = (client->interface->ntcpactive > 1);
UNLOCK(&client->interface->lock);
ns_client_log(client, NS_LOGCATEGORY_CLIENT,
NS_LOGMODULE_CLIENT, ISC_LOG_WARNING,
"TCP client quota reached: %s",
isc_result_totext(result));
if (exit) {
client->newstate = NS_CLIENTSTATE_INACTIVE;
(void)exit_check(client);
return;
}
/*
* 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 clients are
* already servicing TCP queries on this interface (whether
* accepting, reading, or processing). If we find at least
* one, then it's okay *not* to call accept - we can let this
* client go inactive and another will take over when it's
* done.
*
* If there aren't enough active clients on the interface,
* then we can be a little bit flexible about the quota.
* We'll allow *one* extra client through to ensure we're
* listening on every interface; we do this by setting the
* 'force' option to tcpconn_init().
*
* (Note: In practice this means that the real TCP client
* quota is tcp-clients plus the number of listening
* interfaces plus 1.)
*/
exit = (isc_atomic_xadd(&client->interface->ntcpactive, 0) > 0);
if (exit) {
client->newstate = NS_CLIENTSTATE_INACTIVE;
(void)exit_check(client);
return;
}
} else {
client->tcpattached = true;
result = tcpconn_init(client, true);
RUNTIME_CHECK(result == ISC_R_SUCCESS);
}
/*
* 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 this client was set up using get_client() or get_worker(),
* then TCP is already marked active. However, if it was restarted
* from exit_check(), it might not be, so we take care of it now.
*/
if (!client->tcpactive) {
LOCK(&client->interface->lock);
client->interface->ntcpactive++;
UNLOCK(&client->interface->lock);
client->tcpactive = true;
}
mark_tcp_active(client, true);
result = isc_socket_accept(client->tcplistener, client->task,
client_newconn, client);
......@@ -3547,15 +3509,8 @@ client_accept(ns_client_t *client) {
"isc_socket_accept() failed: %s",
isc_result_totext(result));
tcpquota_disconnect(client);
if (client->tcpactive) {
LOCK(&client->interface->lock);
client->interface->ntcpactive--;
UNLOCK(&client->interface->lock);
client->tcpactive = false;
}
tcpconn_detach(client);
mark_tcp_active(client, false);
return;
}
......@@ -3580,9 +3535,7 @@ client_accept(ns_client_t *client) {
* listening for connections itself to prevent the interface
* going dead.
*/
LOCK(&client->interface->lock);
client->interface->ntcpaccepting++;
UNLOCK(&client->interface->lock);
isc_atomic_xadd(&client->interface->ntcpaccepting, 1);
}
static void
......@@ -3653,24 +3606,25 @@ ns_client_replace(ns_client_t *client) {
REQUIRE(client->manager != NULL);
tcp = TCP_CLIENT(client);
if (tcp && client->pipelined) {
if (tcp && client->tcpconn != NULL && client->tcpconn->pipelined) {
result = get_worker(client->manager, client->interface,
client->tcpsocket, client);
} else {
result = get_client(client->manager, client->interface,
client->dispatch, client, tcp);
client->dispatch, tcp);
/*
* The responsibility for listening for new requests is hereby
* transferred to the new client. Therefore, the old client
* should refrain from listening for any more requests.
*/
client->mortal = true;
}
if (result != ISC_R_SUCCESS) {
return (result);
}
/*
* The responsibility for listening for new requests is hereby
* transferred to the new client. Therefore, the old client