#endif /* KERNEL */
#include <opr/queue.h>
+#include <hcrypto/rand.h>
#include "rx.h"
#include "rx_clock.h"
static struct rx_connection
*rxi_FindConnection(osi_socket socket, afs_uint32 host, u_short port,
u_short serviceId, afs_uint32 cid,
- afs_uint32 epoch, int type, u_int securityIndex);
+ afs_uint32 epoch, int type, u_int securityIndex,
+ int *unknownService);
static struct rx_packet
*rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
int istack, osi_socket socket,
static void rxi_ChallengeOn(struct rx_connection *conn);
static int rxi_CheckCall(struct rx_call *call, int haveCTLock);
static void rxi_AckAllInTransmitQueue(struct rx_call *call);
+static void rxi_CancelKeepAliveEvent(struct rx_call *call);
+static void rxi_CancelDelayedAbortEvent(struct rx_call *call);
+static void rxi_CancelGrowMTUEvent(struct rx_call *call);
+static void update_nextCid(void);
+
+#ifndef KERNEL
+static void rxi_Finalize_locked(void);
+#elif defined(UKERNEL)
+# define rxi_Finalize_locked() do { } while (0)
+#endif
#ifdef RX_ENABLE_LOCKS
struct rx_tq_debug {
static unsigned int rxi_rpc_process_stat_cnt;
-/*
- * rxi_busyChannelError is a boolean. It indicates whether or not RX_CALL_BUSY
- * errors should be reported to the application when a call channel appears busy
- * (inferred from the receipt of RX_PACKET_TYPE_BUSY packets on the channel),
- * and there are other call channels in the connection that are not busy.
- * If 0, we do not return errors upon receiving busy packets; we just keep
- * trying on the same call channel until we hit a timeout.
- */
-static afs_int32 rxi_busyChannelError = 0;
-
rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);
extern afs_kmutex_t rx_refcnt_mutex;
extern afs_kmutex_t des_init_mutex;
extern afs_kmutex_t des_random_mutex;
+#ifndef KERNEL
extern afs_kmutex_t rx_clock_mutex;
extern afs_kmutex_t rxi_connCacheMutex;
extern afs_kmutex_t event_handler_mutex;
extern afs_kcondvar_t rx_event_handler_cond;
extern afs_kcondvar_t rx_listener_cond;
+#endif /* !KERNEL */
static afs_kmutex_t epoch_mutex;
static afs_kmutex_t rx_init_mutex;
static void
rxi_InitPthread(void)
{
- MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
- MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
- MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
- MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
- MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
- MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
+#ifndef KERNEL
+ MUTEX_INIT(&rx_clock_mutex, "clock", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rxi_connCacheMutex, "conn cache", MUTEX_DEFAULT, 0);
+ MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
MUTEX_INIT(&listener_mutex, "listener", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rx_if_init_mutex, "if init", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rx_if_mutex, "if", MUTEX_DEFAULT, 0);
+#endif
+ MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
+ MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
+ MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
+ MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);
+#ifndef KERNEL
CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);
+#endif
osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
+ MUTEX_INIT(&rx_mallocedPktQ_lock, "rx_mallocedPktQ_lock", MUTEX_DEFAULT,
+ 0);
+
#ifdef RX_ENABLE_LOCKS
#ifdef RX_LOCKS_DB
rxdb_init();
MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
0);
MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
+#ifndef KERNEL
MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
+#endif
#endif /* RX_ENABLE_LOCKS */
}
* tiers:
*
* rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
+ * also protects updates to rx_nextCid
* conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
* call->lock - locks call data fields.
* These are independent of each other:
#define CLEAR_CALL_QUEUE_LOCK(C)
#endif /* RX_ENABLE_LOCKS */
struct rx_serverQueueEntry *rx_waitForPacket = 0;
-struct rx_serverQueueEntry *rx_waitingForPacket = 0;
-/* ------------Exported Interfaces------------- */
-
-/* This function allows rxkad to set the epoch to a suitably random number
- * which rx_NewConnection will use in the future. The principle purpose is to
- * get rxnull connections to use the same epoch as the rxkad connections do, at
- * least once the first rxkad connection is established. This is important now
- * that the host/port addresses aren't used in FindConnection: the uniqueness
- * of epoch/cid matters and the start time won't do. */
-
-#ifdef AFS_PTHREAD_ENV
/*
- * This mutex protects the following global variables:
- * rx_epoch
+ * This mutex serializes calls to our initialization and shutdown routines
+ * (rx_InitHost, rx_Finalize and shutdown_rx). Only one thread can be running
+ * these at any time; all other threads must wait for it to finish running, and
+ * then examine the value of rxi_running afterwards.
*/
-
-#define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
-#define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
+#ifdef AFS_PTHREAD_ENV
+# define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
+# define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
#else
-#define LOCK_EPOCH
-#define UNLOCK_EPOCH
-#endif /* AFS_PTHREAD_ENV */
+# define LOCK_RX_INIT
+# define UNLOCK_RX_INIT
+#endif
-void
-rx_SetEpoch(afs_uint32 epoch)
+/* ------------Exported Interfaces------------- */
+
+static rx_atomic_t rxi_running = RX_ATOMIC_INIT(0);
+int
+rxi_IsRunning(void)
{
- LOCK_EPOCH;
- rx_epoch = epoch;
- UNLOCK_EPOCH;
+ return rx_atomic_read(&rxi_running);
}
/* Initialize rx. A port number may be mentioned, in which case this
* by the kernel. Whether this will ever overlap anything in
* /etc/services is anybody's guess... Returns 0 on success, -1 on
* error. */
-#ifndef AFS_NT40_ENV
-static
-#endif
-int rxinit_status = 1;
-#ifdef AFS_PTHREAD_ENV
-/*
- * This mutex protects the following global variables:
- * rxinit_status
- */
-
-#define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
-#define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
-#else
-#define LOCK_RX_INIT
-#define UNLOCK_RX_INIT
-#endif
-
int
rx_InitHost(u_int host, u_int port)
{
struct timeval tv;
#endif /* KERNEL */
char *htable, *ptable;
- int tmp_status;
SPLVAR;
INIT_PTHREAD_LOCKS;
LOCK_RX_INIT;
- if (rxinit_status == 0) {
- tmp_status = rxinit_status;
+ if (rxi_IsRunning()) {
UNLOCK_RX_INIT;
- return tmp_status; /* Already started; return previous error code. */
+ return 0; /* already started */
}
#ifdef RXDEBUG
rxi_DebugInit();
#endif
#ifdef AFS_NT40_ENV
if (afs_winsockInit() < 0)
- return -1;
+ goto error;
#endif
#ifndef KERNEL
rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
if (rx_socket == OSI_NULLSOCKET) {
- UNLOCK_RX_INIT;
- return RX_ADDRINUSE;
+ goto addrinuse;
}
#if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
#ifdef RX_LOCKS_DB
MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
0);
MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
+ MUTEX_INIT(&rx_mallocedPktQ_lock, "rx_mallocedPktQ_lock", MUTEX_DEFAULT,
+ 0);
+
#if defined(AFS_HPUX110_ENV)
if (!uniprocessor)
rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
opr_queue_Init(&rx_freePacketQueue);
rxi_NeedMorePackets = FALSE;
rx_nPackets = 0; /* rx_nPackets is managed by rxi_MorePackets* */
+ opr_queue_Init(&rx_mallocedPacketQueue);
/* enforce a minimum number of allocated packets */
if (rx_extraPackets < rxi_nSendFrags * rx_maxSendWindow)
socklen_t addrlen = sizeof(addr);
#endif
if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
- rx_Finalize();
+ rxi_Finalize_locked();
osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
- return -1;
+ goto error;
}
rx_port = addr.sin_port;
#endif
}
rx_stats.minRtt.sec = 9999999;
-#ifdef KERNEL
- rx_SetEpoch(tv.tv_sec | 0x80000000);
-#else
- rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
- * will provide a randomer value. */
-#endif
+ if (RAND_bytes(&rx_epoch, sizeof(rx_epoch)) != 1)
+ goto error;
+ rx_epoch = (rx_epoch & ~0x40000000) | 0x80000000;
+ if (RAND_bytes(&rx_nextCid, sizeof(rx_nextCid)) != 1)
+ goto error;
+ rx_nextCid &= RX_CIDMASK;
MUTEX_ENTER(&rx_quota_mutex);
rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
MUTEX_EXIT(&rx_quota_mutex);
rx_GetIFInfo();
#endif
-#if defined(RXK_LISTENER_ENV) || !defined(KERNEL)
/* Start listener process (exact function is dependent on the
* implementation environment--kernel or user space) */
rxi_StartListener();
-#endif
USERPRI;
- tmp_status = rxinit_status = 0;
+
+ rx_atomic_set(&rxi_running, 1);
+ UNLOCK_RX_INIT;
+
+ return 0;
+
+ addrinuse:
UNLOCK_RX_INIT;
- return tmp_status;
+ return RX_ADDRINUSE;
+
+ error:
+ UNLOCK_RX_INIT;
+ return -1;
}
int
{
struct clock now, retryTime;
+ MUTEX_ASSERT(&call->lock);
clock_GetTime(&now);
retryTime = now;
static_inline void
rxi_rto_cancel(struct rx_call *call)
{
- rxevent_Cancel(&call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
+ MUTEX_ASSERT(&call->lock);
+ if (rxevent_Cancel(&call->resendEvent))
+ CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
}
/*!
}
/**
- * Enables or disables the busy call channel error (RX_CALL_BUSY).
- *
- * @param[in] onoff Non-zero to enable busy call channel errors.
- *
- * @pre Neither rx_Init nor rx_InitHost have been called yet
- */
-void
-rx_SetBusyChannelError(afs_int32 onoff)
-{
- osi_Assert(rxinit_status != 0);
- rxi_busyChannelError = onoff ? 1 : 0;
-}
-
-/**
* Set a delayed ack event on the specified call for the given time
*
* @param[in] call - the call on which to set the event
{
struct clock now, when;
+ MUTEX_ASSERT(&call->lock);
clock_GetTime(&now);
when = now;
clock_Add(&when, offset);
- if (!call->delayedAckEvent
- || clock_Gt(&call->delayedAckTime, &when)) {
+ if (clock_Gt(&call->delayedAckTime, &when) &&
+ rxevent_Cancel(&call->delayedAckEvent)) {
+ /* We successfully cancelled an event too far in the future to install
+ * our new one; we can reuse the reference on the call. */
+ call->delayedAckEvent = rxevent_Post(&when, &now, rxi_SendDelayedAck,
+ call, NULL, 0);
- rxevent_Cancel(&call->delayedAckEvent, call,
- RX_CALL_REFCOUNT_DELAY);
+ call->delayedAckTime = when;
+ } else if (call->delayedAckEvent == NULL) {
CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
-
call->delayedAckEvent = rxevent_Post(&when, &now,
rxi_SendDelayedAck,
call, NULL, 0);
}
}
+void
+rxi_CancelDelayedAckEvent(struct rx_call *call)
+{
+ MUTEX_ASSERT(&call->lock);
+ /* Only drop the ref if we cancelled it before it could run. */
+ if (rxevent_Cancel(&call->delayedAckEvent))
+ CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
+}
+
/* called with unincremented nRequestsRunning to see if it is OK to start
* a new thread in this service. Could be "no" for two reasons: over the
* max quota, or would prevent others from reaching their min quota.
int serviceSecurityIndex)
{
int hashindex, i;
- afs_int32 cid;
struct rx_connection *conn;
SPLVAR;
#endif
NETPRI;
MUTEX_ENTER(&rx_connHashTable_lock);
- cid = (rx_nextCid += RX_MAXCALLS);
conn->type = RX_CLIENT_CONNECTION;
- conn->cid = cid;
conn->epoch = rx_epoch;
- conn->peer = rxi_FindPeer(shost, sport, 0, 1);
+ conn->cid = rx_nextCid;
+ update_nextCid();
+ conn->peer = rxi_FindPeer(shost, sport, 1);
conn->serviceId = sservice;
conn->securityObject = securityObject;
conn->securityData = (void *) 0;
rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
{
conn->idleDeadTime = seconds;
- conn->idleDeadDetection = (seconds ? 1 : 0);
rxi_CheckConnTimeouts(conn);
}
{
struct rx_connection **conn_ptr;
int havecalls = 0;
- struct rx_packet *packet;
int i;
SPLVAR;
if (conn->refCount > 0)
conn->refCount--;
else {
+#ifdef RX_REFCOUNT_CHECK
+ osi_Assert(conn->refCount == 0);
+#endif
if (rx_stats_active) {
MUTEX_ENTER(&rx_stats_mutex);
rxi_lowConnRefCount++;
if ((conn->type == RX_CLIENT_CONNECTION)
&& (conn->flags & (RX_CONN_MAKECALL_WAITING|RX_CONN_MAKECALL_ACTIVE))) {
conn->flags |= RX_CONN_DESTROY_ME;
+ MUTEX_EXIT(&rx_refcnt_mutex);
MUTEX_EXIT(&conn->conn_data_lock);
USERPRI;
return;
/* Push the final acknowledgment out now--there
* won't be a subsequent call to acknowledge the
* last reply packets */
- rxevent_Cancel(&call->delayedAckEvent, call,
- RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
if (call->state == RX_STATE_PRECALL
|| call->state == RX_STATE_ACTIVE) {
rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
return;
}
- if (conn->natKeepAliveEvent) {
- rxi_NatKeepAliveOff(conn);
- }
-
- if (conn->delayedAbortEvent) {
- rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
- packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
- if (packet) {
- MUTEX_ENTER(&conn->conn_data_lock);
- rxi_SendConnectionAbort(conn, packet, 0, 1);
- MUTEX_EXIT(&conn->conn_data_lock);
- rxi_FreePacket(packet);
- }
- }
-
/* Remove from connection hash table before proceeding */
conn_ptr =
&rx_connHashTable[CONN_HASH
rxLastConn = 0;
/* Make sure the connection is completely reset before deleting it. */
- /* get rid of pending events that could zap us later */
- rxevent_Cancel(&conn->challengeEvent, NULL, 0);
- rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
- rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
+ /*
+ * Pending events hold a refcount, so we can't get here if they are
+ * non-NULL. */
+ osi_Assert(conn->challengeEvent == NULL);
+ osi_Assert(conn->delayedAbortEvent == NULL);
+ osi_Assert(conn->natKeepAliveEvent == NULL);
+ osi_Assert(conn->checkReachEvent == NULL);
/* Add the connection to the list of destroyed connections that
* need to be cleaned up. This is necessary to avoid deadlocks
while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
call->flags |= RX_CALL_TQ_WAIT;
call->tqWaiters++;
- osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
+ MUTEX_ASSERT(&call->lock);
CV_WAIT(&call->cv_tq, &call->lock);
call->tqWaiters--;
if (call->tqWaiters == 0) {
dpf(("call %"AFS_PTR_FMT" has %d waiters and flags %d\n",
call, call->tqWaiters, call->flags));
#ifdef RX_ENABLE_LOCKS
- osirx_AssertMine(&call->lock, "rxi_Start start");
+ MUTEX_ASSERT(&call->lock);
CV_BROADCAST(&call->cv_tq);
#else /* RX_ENABLE_LOCKS */
osi_rxWakeup(&call->tq);
}
if (i < RX_MAXCALLS) {
conn->lastBusy[i] = 0;
- call->flags &= ~RX_CALL_PEER_BUSY;
break;
}
if (!wait)
service->minProcs = 0;
service->maxProcs = 1;
service->idleDeadTime = 60;
- service->idleDeadErr = 0;
service->connDeadTime = rx_connDeadTime;
service->executeRequestProc = serviceProc;
service->checkReach = 0;
opr_queue_Append(&rx_idleServerQueue, &sq->entry);
#ifndef AFS_AIX41_ENV
rx_waitForPacket = sq;
-#else
- rx_waitingForPacket = sq;
#endif /* AFS_AIX41_ENV */
do {
CV_WAIT(&sq->cv, &rx_serverPool_lock);
call->abortCount = 0;
}
- call->arrivalProc = (void (*)())0;
+ call->arrivalProc = NULL;
if (rc && call->error == 0) {
rxi_CallError(call, rc);
call->app.mode = RX_MODE_ERROR;
MUTEX_ENTER(&call->lock);
}
if (call->app.mode == RX_MODE_SENDING) {
- MUTEX_EXIT(&call->lock);
- rxi_FlushWrite(call);
- MUTEX_ENTER(&call->lock);
+ rxi_FlushWriteLocked(call);
}
rxi_calltrace(RX_CALL_END, call);
/* Call goes to hold state until reply packets are acknowledged */
call->state = RX_STATE_DALLY;
rxi_ClearTransmitQueue(call, 0);
rxi_rto_cancel(call);
- rxevent_Cancel(&call->keepAliveEvent, call,
- RX_CALL_REFCOUNT_ALIVE);
+ rxi_CancelKeepAliveEvent(call);
}
} else { /* Client connection */
char dummy;
* and force-send it now.
*/
if (call->delayedAckEvent) {
- rxevent_Cancel(&call->delayedAckEvent, call,
- RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
rxi_SendDelayedAck(NULL, call, NULL, 0);
}
MUTEX_ENTER(&conn->conn_call_lock);
MUTEX_ENTER(&call->lock);
- if (!(call->flags & RX_CALL_PEER_BUSY)) {
+ if (!call->error) {
+ /* While there are some circumstances where a call with an error is
+ * obviously not on a "busy" channel, be conservative (clearing
+ * lastBusy is just best-effort to possibly speed up rx_NewCall).
+ * The call channel is definitely not busy if we just successfully
+ * completed a call on it. */
conn->lastBusy[call->channel] = 0;
+
+ } else if (call->error == RX_CALL_TIMEOUT) {
+ /* The call is still probably running on the server side, so try to
+ * avoid this call channel in the future. */
+ conn->lastBusy[call->channel] = clock_Sec();
}
MUTEX_ENTER(&conn->conn_data_lock);
* Map errors to the local host's errno.h format.
*/
error = ntoh_syserr_conv(error);
+
+ /* If the caller said the call failed with some error, we had better
+ * return an error code. */
+ osi_Assert(!rc || error);
return error;
}
void
rx_Finalize(void)
{
- struct rx_connection **conn_ptr, **conn_end;
-
INIT_PTHREAD_LOCKS;
LOCK_RX_INIT;
- if (rxinit_status == 1) {
+ if (!rxi_IsRunning()) {
UNLOCK_RX_INIT;
return; /* Already shutdown. */
}
+ rxi_Finalize_locked();
+ UNLOCK_RX_INIT;
+}
+
+static void
+rxi_Finalize_locked(void)
+{
+ struct rx_connection **conn_ptr, **conn_end;
+ rx_atomic_set(&rxi_running, 0);
rxi_DeleteCachedConnections();
if (rx_connHashTable) {
MUTEX_ENTER(&rx_connHashTable_lock);
for (conn = *conn_ptr; conn; conn = next) {
next = conn->next;
if (conn->type == RX_CLIENT_CONNECTION) {
- MUTEX_ENTER(&rx_refcnt_mutex);
- conn->refCount++;
- MUTEX_EXIT(&rx_refcnt_mutex);
+ rx_GetConnection(conn);
#ifdef RX_ENABLE_LOCKS
rxi_DestroyConnectionNoLock(conn);
#else /* RX_ENABLE_LOCKS */
#ifdef AFS_NT40_ENV
afs_winsockCleanup();
#endif
-
- rxinit_status = 1;
- UNLOCK_RX_INIT;
}
#endif
*/
MUTEX_ENTER(&conn->conn_data_lock);
if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
- MUTEX_ENTER(&rx_refcnt_mutex);
- conn->refCount++;
- MUTEX_EXIT(&rx_refcnt_mutex);
+ rx_GetConnection(conn);
MUTEX_EXIT(&conn->conn_data_lock);
#ifdef RX_ENABLE_LOCKS
if (haveCTLock)
if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
peer->maxDgramPackets = 1;
/* We no longer have valid peer packet information */
- if (peer->maxPacketSize-RX_IPUDP_SIZE > peer->ifMTU)
+ if (peer->maxPacketSize + RX_HEADER_SIZE > peer->ifMTU)
peer->maxPacketSize = 0;
MUTEX_EXIT(&peer->peer_lock);
#ifdef AFS_RXERRQ_ENV
static void
-rxi_SetPeerDead(afs_uint32 host, afs_uint16 port)
+rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
{
int hashIndex = PEER_HASH(host, port);
struct rx_peer *peer;
for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
if (peer->host == host && peer->port == port) {
+ peer->refCount++;
break;
}
}
+ MUTEX_EXIT(&rx_peerHashTable_lock);
+
if (peer) {
rx_atomic_inc(&peer->neterrs);
- }
+ MUTEX_ENTER(&peer->peer_lock);
+ peer->last_err_origin = RX_NETWORK_ERROR_ORIGIN_ICMP;
+ peer->last_err_type = err->ee_type;
+ peer->last_err_code = err->ee_code;
+ MUTEX_EXIT(&peer->peer_lock);
- MUTEX_EXIT(&rx_peerHashTable_lock);
+ MUTEX_ENTER(&rx_peerHashTable_lock);
+ peer->refCount--;
+ MUTEX_EXIT(&rx_peerHashTable_lock);
+ }
}
void
case ICMP_PORT_UNREACH:
case ICMP_NET_ANO:
case ICMP_HOST_ANO:
- rxi_SetPeerDead(addr, port);
+ rxi_SetPeerDead(err, addr, port);
break;
}
}
}
+
+static const char *
+rxi_TranslateICMP(int type, int code)
+{
+ switch (type) {
+ case ICMP_DEST_UNREACH:
+ switch (code) {
+ case ICMP_NET_UNREACH:
+ return "Destination Net Unreachable";
+ case ICMP_HOST_UNREACH:
+ return "Destination Host Unreachable";
+ case ICMP_PROT_UNREACH:
+ return "Destination Protocol Unreachable";
+ case ICMP_PORT_UNREACH:
+ return "Destination Port Unreachable";
+ case ICMP_NET_ANO:
+ return "Destination Net Prohibited";
+ case ICMP_HOST_ANO:
+ return "Destination Host Prohibited";
+ }
+ break;
+ }
+ return NULL;
+}
#endif /* AFS_RXERRQ_ENV */
+/**
+ * Get the last network error for a connection
+ *
+ * A "network error" here means an error retrieved from ICMP, or some other
+ * mechanism outside of Rx that informs us of errors in network reachability.
+ *
+ * If a peer associated with the given Rx connection has received a network
+ * error recently, this function allows the caller to know what error
+ * specifically occurred. This can be useful to know, since e.g. ICMP errors
+ * can cause calls to that peer to be quickly aborted. So, this function can
+ * help see why a call was aborted due to network errors.
+ *
+ * If we have received traffic from a peer since the last network error, we
+ * treat that peer as if we had not received an network error for it.
+ *
+ * @param[in] conn The Rx connection to examine
+ * @param[out] err_origin The origin of the last network error (e.g. ICMP);
+ * one of the RX_NETWORK_ERROR_ORIGIN_* constants
+ * @param[out] err_type The type of the last error
+ * @param[out] err_code The code of the last error
+ * @param[out] msg Human-readable error message, if applicable; NULL otherwise
+ *
+ * @return If we have an error
+ * @retval -1 No error to get; 'out' params are undefined
+ * @retval 0 We have an error; 'out' params contain the last error
+ */
+int
+rx_GetNetworkError(struct rx_connection *conn, int *err_origin, int *err_type,
+ int *err_code, const char **msg)
+{
+#ifdef AFS_RXERRQ_ENV
+ struct rx_peer *peer = conn->peer;
+ if (rx_atomic_read(&peer->neterrs)) {
+ MUTEX_ENTER(&peer->peer_lock);
+ *err_origin = peer->last_err_origin;
+ *err_type = peer->last_err_type;
+ *err_code = peer->last_err_code;
+ MUTEX_EXIT(&peer->peer_lock);
+
+ *msg = NULL;
+ if (*err_origin == RX_NETWORK_ERROR_ORIGIN_ICMP) {
+ *msg = rxi_TranslateICMP(*err_type, *err_code);
+ }
+
+ return 0;
+ }
+#endif
+ return -1;
+}
+
/* Find the peer process represented by the supplied (host,port)
* combination. If there is no appropriate active peer structure, a
* new one will be allocated and initialized
- * The origPeer, if set, is a pointer to a peer structure on which the
- * refcount will be be decremented. This is used to replace the peer
- * structure hanging off a connection structure */
+ */
struct rx_peer *
-rxi_FindPeer(afs_uint32 host, u_short port,
- struct rx_peer *origPeer, int create)
+rxi_FindPeer(afs_uint32 host, u_short port, int create)
{
struct rx_peer *pp;
int hashIndex;
if (pp && create) {
pp->refCount++;
}
- if (origPeer)
- origPeer->refCount--;
MUTEX_EXIT(&rx_peerHashTable_lock);
return pp;
}
static struct rx_connection *
rxi_FindConnection(osi_socket socket, afs_uint32 host,
u_short port, u_short serviceId, afs_uint32 cid,
- afs_uint32 epoch, int type, u_int securityIndex)
+ afs_uint32 epoch, int type, u_int securityIndex,
+ int *unknownService)
{
int hashindex, flag, i;
struct rx_connection *conn;
+ *unknownService = 0;
hashindex = CONN_HASH(host, port, cid, epoch, type);
MUTEX_ENTER(&rx_connHashTable_lock);
rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
if (!service || (securityIndex >= service->nSecurityObjects)
|| (service->securityObjects[securityIndex] == 0)) {
MUTEX_EXIT(&rx_connHashTable_lock);
+ *unknownService = 1;
return (struct rx_connection *)0;
}
conn = rxi_AllocConnection(); /* This bzero's the connection */
CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
conn->next = rx_connHashTable[hashindex];
rx_connHashTable[hashindex] = conn;
- conn->peer = rxi_FindPeer(host, port, 0, 1);
+ conn->peer = rxi_FindPeer(host, port, 1);
conn->type = RX_SERVER_CONNECTION;
conn->lastSendTime = clock_Sec(); /* don't GC immediately */
conn->epoch = epoch;
conn->nSpecific = 0;
conn->specific = NULL;
rx_SetConnDeadTime(conn, service->connDeadTime);
- conn->idleDeadTime = service->idleDeadTime;
- conn->idleDeadDetection = service->idleDeadErr ? 1 : 0;
+ rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
for (i = 0; i < RX_MAXCALLS; i++) {
conn->twind[i] = rx_initSendWindow;
conn->rwind[i] = rx_initReceiveWindow;
rx_atomic_inc(&rx_stats.nServerConns);
}
- MUTEX_ENTER(&rx_refcnt_mutex);
- conn->refCount++;
- MUTEX_EXIT(&rx_refcnt_mutex);
+ rx_GetConnection(conn);
rxLastConn = conn; /* store this connection as the last conn used */
MUTEX_EXIT(&rx_connHashTable_lock);
return conn;
}
-/**
- * Timeout a call on a busy call channel if appropriate.
- *
- * @param[in] call The busy call.
- *
- * @pre 'call' is marked as busy (namely,
- * call->conn->lastBusy[call->channel] != 0)
- *
- * @pre call->lock is held
- * @pre rxi_busyChannelError is nonzero
- *
- * @note call->lock is dropped and reacquired
- */
-static void
-rxi_CheckBusy(struct rx_call *call)
-{
- struct rx_connection *conn = call->conn;
- int channel = call->channel;
- int freechannel = 0;
- int i;
- afs_uint32 callNumber;
-
- MUTEX_EXIT(&call->lock);
-
- MUTEX_ENTER(&conn->conn_call_lock);
- callNumber = *call->callNumber;
-
- /* Are there any other call slots on this conn that we should try? Look for
- * slots that are empty and are either non-busy, or were marked as busy
- * longer than conn->secondsUntilDead seconds before this call started. */
-
- for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
- if (i == channel) {
- /* only look at channels that aren't us */
- continue;
- }
-
- if (conn->lastBusy[i]) {
- /* if this channel looked busy too recently, don't look at it */
- if (conn->lastBusy[i] >= call->startTime.sec) {
- continue;
- }
- if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
- continue;
- }
- }
-
- if (conn->call[i]) {
- struct rx_call *tcall = conn->call[i];
- MUTEX_ENTER(&tcall->lock);
- if (tcall->state == RX_STATE_DALLY) {
- freechannel = 1;
- }
- MUTEX_EXIT(&tcall->lock);
- } else {
- freechannel = 1;
- }
- }
-
- MUTEX_ENTER(&call->lock);
-
- /* Since the call->lock and conn->conn_call_lock have been released it is
- * possible that (1) the call may no longer be busy and/or (2) the call may
- * have been reused by another waiting thread. Therefore, we must confirm
- * that the call state has not changed when deciding whether or not to
- * force this application thread to retry by forcing a Timeout error. */
-
- if (freechannel && *call->callNumber == callNumber &&
- (call->flags & RX_CALL_PEER_BUSY)) {
- /* Since 'freechannel' is set, there exists another channel in this
- * rx_conn that the application thread might be able to use. We know
- * that we have the correct call since callNumber is unchanged, and we
- * know that the call is still busy. So, set the call error state to
- * rxi_busyChannelError so the application can retry the request,
- * presumably on a less-busy call channel. */
-
- rxi_CallError(call, RX_CALL_BUSY);
- }
- MUTEX_EXIT(&conn->conn_call_lock);
-}
-
/*!
* Abort the call if the server is over the busy threshold. This
* can be used without requiring a call structure be initialised,
rxi_AbortIfServerBusy(osi_socket socket, struct rx_connection *conn,
struct rx_packet *np)
{
+ afs_uint32 serial;
+
if ((rx_BusyThreshold > 0) &&
(rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
+ MUTEX_ENTER(&conn->conn_data_lock);
+ serial = ++conn->serial;
+ MUTEX_EXIT(&conn->conn_data_lock);
rxi_SendRawAbort(socket, conn->peer->host, conn->peer->port,
- rx_BusyError, np, 0);
+ serial, rx_BusyError, np, 0);
if (rx_stats_active)
rx_atomic_inc(&rx_stats.nBusies);
return 1;
channel = np->header.cid & RX_CHANNELMASK;
MUTEX_ENTER(&conn->conn_call_lock);
call = conn->call[channel];
+ if (np->header.type == RX_PACKET_TYPE_BUSY) {
+ conn->lastBusy[channel] = clock_Sec();
+ }
if (!call || conn->callNumber[channel] != np->header.callNumber) {
MUTEX_EXIT(&conn->conn_call_lock);
if (rx_stats_active)
* flag is cleared.
*/
#ifdef RX_ENABLE_LOCKS
- if (call->state == RX_STATE_ACTIVE) {
+ if (call->state == RX_STATE_ACTIVE && !call->error) {
rxi_WaitforTQBusy(call);
/* If we entered error state while waiting,
* must call rxi_CallError to permit rxi_ResetCall
struct rx_call *call;
struct rx_connection *conn;
int type;
+ int unknownService = 0;
#ifdef RXDEBUG
char *packetType;
#endif
struct rx_peer *peer;
/* Try to look up the peer structure, but don't create one */
- peer = rxi_FindPeer(host, port, 0, 0);
+ peer = rxi_FindPeer(host, port, 0);
/* Since this may not be associated with a connection, it may have
* no refCount, meaning we could race with ReapConnections
addr.sin_family = AF_INET;
addr.sin_port = port;
addr.sin_addr.s_addr = host;
+ memset(&addr.sin_zero, 0, sizeof(addr.sin_zero));
#ifdef STRUCT_SOCKADDR_HAS_SA_LEN
addr.sin_len = sizeof(addr);
-#endif /* AFS_OSF_ENV */
+#endif
drop = (*rx_justReceived) (np, &addr);
/* drop packet if return value is non-zero */
if (drop)
conn =
rxi_FindConnection(socket, host, port, np->header.serviceId,
np->header.cid, np->header.epoch, type,
- np->header.securityIndex);
+ np->header.securityIndex, &unknownService);
/* To avoid having 2 connections just abort at each other,
don't abort an abort. */
if (!conn) {
- if (np->header.type != RX_PACKET_TYPE_ABORT)
- rxi_SendRawAbort(socket, host, port, RX_INVALID_OPERATION,
+ if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
+ rxi_SendRawAbort(socket, host, port, 0, RX_INVALID_OPERATION,
np, 0);
return np;
}
}
}
- if (type == RX_SERVER_CONNECTION) {
+ if (type == RX_SERVER_CONNECTION)
call = rxi_ReceiveServerCall(socket, np, conn);
- if (call == NULL) {
- putConnection(conn);
- return np;
- }
- } else {
+ else
call = rxi_ReceiveClientCall(np, conn);
- if (call == NULL) {
- putConnection(conn);
- return np;
- }
- /* If we're receiving the response, then all transmit packets are
- * implicitly acknowledged. Get rid of them. */
- if (np->header.type == RX_PACKET_TYPE_DATA)
- rxi_AckAllInTransmitQueue(call);
+ if (call == NULL) {
+ putConnection(conn);
+ return np;
}
- osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
+ MUTEX_ASSERT(&call->lock);
/* Set remote user defined status from packet */
call->remoteStatus = np->header.userStatus;
/* Now do packet type-specific processing */
switch (np->header.type) {
case RX_PACKET_TYPE_DATA:
+ /* If we're a client, and receiving a response, then all the packets
+ * we transmitted packets are implicitly acknowledged. */
+ if (type == RX_CLIENT_CONNECTION && !opr_queue_IsEmpty(&call->tq))
+ rxi_AckAllInTransmitQueue(call);
+
np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
newcallp);
break;
putConnection(conn);
return np; /* xmitting; drop packet */
}
- case RX_PACKET_TYPE_BUSY: {
- struct clock busyTime;
- clock_NewTime();
- clock_GetTime(&busyTime);
-
- MUTEX_EXIT(&call->lock);
-
- MUTEX_ENTER(&conn->conn_call_lock);
- MUTEX_ENTER(&call->lock);
- conn->lastBusy[call->channel] = busyTime.sec;
- call->flags |= RX_CALL_PEER_BUSY;
- MUTEX_EXIT(&call->lock);
- MUTEX_EXIT(&conn->conn_call_lock);
-
- putConnection(conn);
- return np;
- }
+ case RX_PACKET_TYPE_BUSY:
+ /* Mostly ignore BUSY packets. We will update lastReceiveTime below,
+ * so we don't think the endpoint is completely dead, but otherwise
+ * just act as if we never saw anything. If all we get are BUSY packets
+ * back, then we will eventually error out with RX_CALL_TIMEOUT if the
+ * connection is configured with idle/hard timeouts. */
+ break;
case RX_PACKET_TYPE_ACKALL:
/* All packets acknowledged, so we can drop all packets previously
* the packet will be delivered to the user before any get time is required
* (if not, then the time won't actually be re-evaluated here). */
call->lastReceiveTime = clock_Sec();
- /* we've received a legit packet, so the channel is not busy */
- call->flags &= ~RX_CALL_PEER_BUSY;
MUTEX_EXIT(&call->lock);
putConnection(conn);
return np;
}
}
+/*
+ * Event handler function for connection-specific events for checking
+ * reachability. Also called directly from main code with |event| == NULL
+ * in order to trigger the initial reachability check.
+ *
+ * When |event| == NULL, must be called with the connection data lock held,
+ * but returns with the lock unlocked.
+ */
static void
rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2, int dummy)
{
struct clock when, now;
int i, waiting;
- MUTEX_ENTER(&conn->conn_data_lock);
-
- if (event) {
- rxevent_Put(conn->checkReachEvent);
- conn->checkReachEvent = NULL;
- }
+ if (event != NULL)
+ MUTEX_ENTER(&conn->conn_data_lock);
+ else
+ MUTEX_ASSERT(&conn->conn_data_lock);
+ if (event != NULL && event == conn->checkReachEvent)
+ rxevent_Put(&conn->checkReachEvent);
waiting = conn->flags & RX_CONN_ATTACHWAIT;
- if (event) {
- putConnection(conn);
- }
MUTEX_EXIT(&conn->conn_data_lock);
if (waiting) {
when.sec += RX_CHECKREACH_TIMEOUT;
MUTEX_ENTER(&conn->conn_data_lock);
if (!conn->checkReachEvent) {
- MUTEX_ENTER(&rx_refcnt_mutex);
- conn->refCount++;
- MUTEX_EXIT(&rx_refcnt_mutex);
+ rx_GetConnection(conn);
conn->checkReachEvent = rxevent_Post(&when, &now,
rxi_CheckReachEvent, conn,
NULL, 0);
MUTEX_EXIT(&conn->conn_data_lock);
}
}
+ /* If fired as an event handler, drop our refcount on the connection. */
+ if (event != NULL)
+ putConnection(conn);
}
static int
return 1;
}
conn->flags |= RX_CONN_ATTACHWAIT;
- MUTEX_EXIT(&conn->conn_data_lock);
- if (!conn->checkReachEvent)
+ if (conn->checkReachEvent == NULL) {
+ /* rxi_CheckReachEvent(NULL, ...) will drop the lock. */
rxi_CheckReachEvent(NULL, conn, call, 0);
+ } else {
+ MUTEX_EXIT(&conn->conn_data_lock);
+ }
return 1;
}
if (rx_stats_active)
rx_atomic_inc(&rx_stats.dupPacketsRead);
dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
- rxevent_Cancel(&call->delayedAckEvent, call,
- RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
ackNeeded = 0;
call->rprev = seq;
if (call->arrivalProc) {
(*call->arrivalProc) (call, call->arrivalProcHandle,
call->arrivalProcArg);
- call->arrivalProc = (void (*)())0;
+ call->arrivalProc = NULL;
}
/* Update last packet received */
if (seq < call->rnext) {
if (rx_stats_active)
rx_atomic_inc(&rx_stats.dupPacketsRead);
- rxevent_Cancel(&call->delayedAckEvent, call,
- RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
ackNeeded = 0;
call->rprev = seq;
* accomodated by the current window, then send a negative
* acknowledge and drop the packet */
if ((call->rnext + call->rwind) <= seq) {
- rxevent_Cancel(&call->delayedAckEvent, call,
- RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
istack);
ackNeeded = 0;
if (seq == tp->header.seq) {
if (rx_stats_active)
rx_atomic_inc(&rx_stats.dupPacketsRead);
- rxevent_Cancel(&call->delayedAckEvent, call,
- RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
istack);
ackNeeded = 0;
* received. Always send a soft ack for the last packet in
* the server's reply. */
if (ackNeeded) {
- rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
np = rxi_SendAck(call, np, serial, ackNeeded, istack);
} else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
- rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
} else if (call->nSoftAcks) {
if (haveLast && !(flags & RX_CLIENT_INITIATED))
else
rxi_PostDelayedAckEvent(call, &rx_softAckDelay);
} else if (call->flags & RX_CALL_RECEIVE_DONE) {
- rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
}
return np;
* but we are clearly receiving.
*/
if (!peer->maxPacketSize)
- peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;
+ peer->maxPacketSize = RX_MIN_PACKET_SIZE - RX_HEADER_SIZE;
if (pktsize > peer->maxPacketSize) {
peer->maxPacketSize = pktsize;
- if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
- peer->ifMTU=pktsize-RX_IPUDP_SIZE;
+ if ((pktsize + RX_HEADER_SIZE > peer->ifMTU)) {
+ peer->ifMTU = pktsize + RX_HEADER_SIZE;
peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
rxi_ScheduleGrowMTUEvent(call, 1);
}
* queue - they're not addressed by the contents of this ACK packet.
*/
- /* If the window has been extended by this acknowledge packet,
- * then wakeup a sender waiting in alloc for window space, or try
- * sending packets now, if he's been sitting on packets due to
- * lack of window space */
- if (call->tnext < (call->tfirst + call->twind)) {
-#ifdef RX_ENABLE_LOCKS
- CV_SIGNAL(&call->cv_twind);
-#else
- if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
- call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
- osi_rxWakeup(&call->twind);
- }
-#endif
- if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
- call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
- }
- }
-
/* if the ack packet has a receivelen field hanging off it,
* update our state */
if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
rx_packetread(np, rx_AckDataSize(ap->nAcks) + (int)sizeof(afs_int32),
(int)sizeof(afs_int32), &tSize);
tSize = (afs_uint32) ntohl(tSize);
+ if (tSize > RX_MAX_PACKET_SIZE)
+ tSize = RX_MAX_PACKET_SIZE;
+ if (tSize < RX_MIN_PACKET_SIZE)
+ tSize = RX_MIN_PACKET_SIZE;
peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
/* Get the maximum packet size to send to this peer */
rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
&tSize);
tSize = (afs_uint32) ntohl(tSize);
+ if (tSize > RX_MAX_PACKET_SIZE)
+ tSize = RX_MAX_PACKET_SIZE;
+ if (tSize < RX_MIN_PACKET_SIZE)
+ tSize = RX_MIN_PACKET_SIZE;
tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
(int)sizeof(afs_int32), &tSize);
tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
+ if (tSize == 0)
+ tSize = 1;
+ if (tSize >= rx_maxSendWindow)
+ tSize = rx_maxSendWindow;
if (tSize < call->twind) { /* smaller than our send */
call->twind = tSize; /* window, we must send less... */
call->ssthresh = MIN(call->twind, call->ssthresh);
rx_AckDataSize(ap->nAcks) + 2 * (int)sizeof(afs_int32),
sizeof(afs_int32), &tSize);
tSize = (afs_uint32) ntohl(tSize);
+ if (tSize == 0)
+ tSize = 1;
+ if (tSize >= rx_maxSendWindow)
+ tSize = rx_maxSendWindow;
/*
* As of AFS 3.5 we set the send window to match the receive window.
*/
call->MTU = OLD_MAX_PACKET_SIZE;
}
+ /* If the window has been extended by this acknowledge packet,
+ * then wakeup a sender waiting in alloc for window space, or try
+ * sending packets now, if he's been sitting on packets due to
+ * lack of window space */
+ if (call->tnext < (call->tfirst + call->twind)) {
+#ifdef RX_ENABLE_LOCKS
+ CV_SIGNAL(&call->cv_twind);
+#else
+ if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
+ call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
+ osi_rxWakeup(&call->twind);
+ }
+#endif
+ if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
+ call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
+ }
+ }
+
if (nNacked) {
/*
* Calculate how many datagrams were successfully received after
&& call->tfirst + call->nSoftAcked >= call->tnext) {
call->state = RX_STATE_DALLY;
rxi_ClearTransmitQueue(call, 0);
- rxevent_Cancel(&call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
+ rxi_CancelKeepAliveEvent(call);
} else if (!opr_queue_IsEmpty(&call->tq)) {
rxi_Start(call, istack);
}
return np;
}
+/**
+ * Schedule a connection abort to be sent after some delay.
+ *
+ * @param[in] conn The connection to send the abort on.
+ * @param[in] msec The number of milliseconds to wait before sending.
+ *
+ * @pre conn_data_lock must be held
+ */
+static void
+rxi_SendConnectionAbortLater(struct rx_connection *conn, int msec)
+{
+ struct clock when, now;
+
+ MUTEX_ASSERT(&conn->conn_data_lock);
+ if (!conn->error) {
+ return;
+ }
+ if (!conn->delayedAbortEvent) {
+ clock_GetTime(&now);
+ when = now;
+ clock_Addmsec(&when, msec);
+ rx_GetConnection(conn);
+ conn->delayedAbortEvent =
+ rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
+ }
+}
+
/* Received a response to a challenge packet */
static struct rx_packet *
rxi_ReceiveResponsePacket(struct rx_connection *conn,
if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
return np;
+ if (!conn->securityChallengeSent) {
+ /* We've never sent out a challenge for this connection, so this
+ * response cannot possibly be correct; ignore it. This can happen
+ * if we sent a challenge to the client, then we were restarted, and
+ * then the client sent us a response. If we ignore the response, the
+ * client will eventually resend a data packet, causing us to send a
+ * new challenge and the client to send a new response. */
+ return np;
+ }
+
/* Otherwise, have the security object evaluate the response packet */
error = RXS_CheckResponse(conn->securityObject, conn, np);
if (error) {
/* If the response is invalid, reset the connection, sending
- * an abort to the peer */
-#ifndef KERNEL
- rxi_Delay(1);
-#endif
+ * an abort to the peer. Send the abort with a 1 second delay,
+ * to avoid a peer hammering us by constantly recreating a
+ * connection with bad credentials. */
rxi_ConnectionError(conn, error);
MUTEX_ENTER(&conn->conn_data_lock);
- np = rxi_SendConnectionAbort(conn, np, istack, 0);
+ rxi_SendConnectionAbortLater(conn, 1000);
MUTEX_EXIT(&conn->conn_data_lock);
return np;
} else {
call->flags |= RX_CALL_ACKALL_SENT;
}
+/*
+ * Event handler for per-call delayed acks.
+ * Also called synchronously, with |event| == NULL, to send a "delayed" ack
+ * immediately.
+ */
static void
rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused1,
int unused2)
#ifdef RX_ENABLE_LOCKS
if (event) {
MUTEX_ENTER(&call->lock);
- if (event == call->delayedAckEvent) {
- rxevent_Put(call->delayedAckEvent);
- call->delayedAckEvent = NULL;
- }
- CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
+ if (event == call->delayedAckEvent)
+ rxevent_Put(&call->delayedAckEvent);
}
(void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
if (event)
MUTEX_EXIT(&call->lock);
#else /* RX_ENABLE_LOCKS */
- if (event) {
- rxevent_Put(call->delayedAckEvent);
- call->delayedAckEvent = NULL;
- }
+ if (event)
+ rxevent_Put(&call->delayedAckEvent);
(void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
#endif /* RX_ENABLE_LOCKS */
+ /* Release the call reference for the event that fired. */
+ if (event)
+ CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
}
#ifdef RX_ENABLE_LOCKS
rxi_SendCallAbort(struct rx_call *call, struct rx_packet *packet,
int istack, int force)
{
- afs_int32 error, cerror;
+ afs_int32 error;
struct clock when, now;
if (!call->error)
return packet;
- switch (call->error) {
- case RX_CALL_IDLE:
- case RX_CALL_BUSY:
- cerror = RX_CALL_TIMEOUT;
- break;
- default:
- cerror = call->error;
- }
-
/* Clients should never delay abort messages */
if (rx_IsClientConn(call->conn))
force = 1;
- if (call->abortCode != cerror) {
- call->abortCode = cerror;
+ /*
+ * An opcode that has been deprecated or has yet to be implemented is not
+ * a misbehavior of the client. Do not punish the client by introducing
+ * delays.
+ */
+ if (call->error == RXGEN_OPCODE) {
+ force = 1;
+ } else if (call->abortCode != call->error) {
+ call->abortCode = call->error;
call->abortCount = 0;
}
if (force || rxi_callAbortThreshhold == 0
|| call->abortCount < rxi_callAbortThreshhold) {
- if (call->delayedAbortEvent) {
- rxevent_Cancel(&call->delayedAbortEvent, call,
- RX_CALL_REFCOUNT_ABORT);
- }
- error = htonl(cerror);
- call->abortCount++;
+ rxi_CancelDelayedAbortEvent(call);
+ error = htonl(call->error);
+ if (!force)
+ call->abortCount++;
packet =
rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
(char *)&error, sizeof(error), istack);
return packet;
}
+static void
+rxi_CancelDelayedAbortEvent(struct rx_call *call)
+{
+ MUTEX_ASSERT(&call->lock);
+ if (rxevent_Cancel(&call->delayedAbortEvent))
+ CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
+}
+
/* Send an abort packet for the specified connection. Packet is an
* optional pointer to a packet that can be used to send the abort.
* Once the number of abort messages reaches the threshhold, an
struct rx_packet *packet, int istack, int force)
{
afs_int32 error;
- struct clock when, now;
if (!conn->error)
return packet;
if (force || rxi_connAbortThreshhold == 0
|| conn->abortCount < rxi_connAbortThreshhold) {
- rxevent_Cancel(&conn->delayedAbortEvent, NULL, 0);
+ if (rxevent_Cancel(&conn->delayedAbortEvent))
+ putConnection(conn);
error = htonl(conn->error);
conn->abortCount++;
MUTEX_EXIT(&conn->conn_data_lock);
RX_PACKET_TYPE_ABORT, (char *)&error,
sizeof(error), istack);
MUTEX_ENTER(&conn->conn_data_lock);
- } else if (!conn->delayedAbortEvent) {
- clock_GetTime(&now);
- when = now;
- clock_Addmsec(&when, rxi_connAbortDelay);
- conn->delayedAbortEvent =
- rxevent_Post(&when, &now, rxi_SendDelayedConnAbort, conn, NULL, 0);
+ } else {
+ rxi_SendConnectionAbortLater(conn, rxi_connAbortDelay);
}
return packet;
}
dpf(("rxi_ConnectionError conn %"AFS_PTR_FMT" error %d\n", conn, error));
MUTEX_ENTER(&conn->conn_data_lock);
- rxevent_Cancel(&conn->challengeEvent, NULL, 0);
- rxevent_Cancel(&conn->natKeepAliveEvent, NULL, 0);
- if (conn->checkReachEvent) {
- rxevent_Cancel(&conn->checkReachEvent, NULL, 0);
+ if (rxevent_Cancel(&conn->challengeEvent))
+ putConnection(conn);
+ if (rxevent_Cancel(&conn->natKeepAliveEvent))
+ putConnection(conn);
+ if (rxevent_Cancel(&conn->checkReachEvent)) {
conn->flags &= ~(RX_CONN_ATTACHWAIT|RX_CONN_NAT_PING);
putConnection(conn);
}
void
rxi_CallError(struct rx_call *call, afs_int32 error)
{
-#ifdef DEBUG
- osirx_AssertMine(&call->lock, "rxi_CallError");
-#endif
+ MUTEX_ASSERT(&call->lock);
dpf(("rxi_CallError call %"AFS_PTR_FMT" error %d call->error %d\n", call, error, call->error));
if (call->error)
error = call->error;
int flags;
struct rx_peer *peer;
struct rx_packet *packet;
-#ifdef DEBUG
- osirx_AssertMine(&call->lock, "rxi_ResetCall");
-#endif
+
+ MUTEX_ASSERT(&call->lock);
dpf(("rxi_ResetCall(call %"AFS_PTR_FMT", newcall %d)\n", call, newcall));
/* Notify anyone who is waiting for asynchronous packet arrival */
if (call->arrivalProc) {
(*call->arrivalProc) (call, call->arrivalProcHandle,
call->arrivalProcArg);
- call->arrivalProc = (void (*)())0;
+ call->arrivalProc = NULL;
}
- rxevent_Cancel(&call->growMTUEvent, call, RX_CALL_REFCOUNT_MTU);
+ rxi_CancelGrowMTUEvent(call);
if (call->delayedAbortEvent) {
- rxevent_Cancel(&call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
+ rxi_CancelDelayedAbortEvent(call);
packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
if (packet) {
rxi_SendCallAbort(call, packet, 0, 1);
}
call->flags = 0;
- if (!newcall && (flags & RX_CALL_PEER_BUSY)) {
- /* The call channel is still busy; resetting the call doesn't change
- * that. However, if 'newcall' is set, we are processing a call
- * structure that has either been recycled from the free list, or has
- * been newly allocated. So, RX_CALL_PEER_BUSY is not relevant if
- * 'newcall' is set, since it describes a completely different call
- * channel which we do not care about. */
- call->flags |= RX_CALL_PEER_BUSY;
- }
-
rxi_ClearReceiveQueue(call);
/* why init the queue if you just emptied it? queue_Init(&call->rq); */
}
#endif /* RX_ENABLE_LOCKS */
- rxi_KeepAliveOff(call);
- rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelKeepAliveEvent(call);
+ rxi_CancelDelayedAckEvent(call);
}
/* Send an acknowledge for the indicated packet (seq,serial) of the
int reason; Reason an acknowledge was prompted
*/
+#define RX_ZEROS 1024
+static char rx_zeros[RX_ZEROS];
+
struct rx_packet *
rxi_SendAck(struct rx_call *call,
struct rx_packet *optionalPacket, int serial, int reason,
*/
if (call->conn->peer->maxPacketSize &&
(call->conn->peer->maxPacketSize < OLD_MAX_PACKET_SIZE
- +RX_IPUDP_SIZE))
+ - RX_HEADER_SIZE))
padbytes = call->conn->peer->maxPacketSize+16;
else
padbytes = call->conn->peer->maxMTU + 128;
ap->nAcks = offset;
p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
+ /* Must zero the 3 octets that rx_AckDataSize skips at the end of the
+ * ACK list.
+ */
+ rx_packetwrite(p, rx_AckDataSize(offset) - 3, 3, rx_zeros);
+
/* these are new for AFS 3.3 */
templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
templ = htonl(templ);
rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
sizeof(afs_int32), &templ);
+ p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
+
p->header.serviceId = call->conn->serviceId;
p->header.cid = (call->conn->cid | call->channel);
p->header.callNumber = *call->callNumber;
p->header.epoch = call->conn->epoch;
p->header.type = RX_PACKET_TYPE_ACK;
p->header.flags = RX_SLOW_START_OK;
- if (reason == RX_ACK_PING) {
+ if (reason == RX_ACK_PING)
p->header.flags |= RX_REQUEST_ACK;
- if (padbytes) {
- p->length = padbytes +
- rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32);
- while (padbytes--)
- /* not fast but we can potentially use this if truncated
- * fragments are delivered to figure out the mtu.
- */
- rx_packetwrite(p, rx_AckDataSize(offset) + 4 *
- sizeof(afs_int32), sizeof(afs_int32),
- &padbytes);
+ while (padbytes > 0) {
+ if (padbytes > RX_ZEROS) {
+ rx_packetwrite(p, p->length, RX_ZEROS, rx_zeros);
+ p->length += RX_ZEROS;
+ padbytes -= RX_ZEROS;
+ } else {
+ rx_packetwrite(p, p->length, padbytes, rx_zeros);
+ p->length += padbytes;
+ padbytes = 0;
}
}
+
if (call->conn->type == RX_CLIENT_CONNECTION)
p->header.flags |= RX_CLIENT_INITIATED;
/* Since we're about to send a data packet to the peer, it's
* safe to nuke any scheduled end-of-packets ack */
- rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
MUTEX_EXIT(&call->lock);
CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
* processing), and for the connection (so that we can discover
* idle connections) */
conn->lastSendTime = call->lastSendTime = clock_Sec();
- /* Let a set of retransmits trigger an idle timeout */
- if (!xmit->resending)
- call->lastSendData = call->lastSendTime;
}
/* When sending packets we need to follow these rules:
/* Make sure that the event pointer is removed from the call
* structure, since there is no longer a per-call retransmission
* event pending. */
- if (event == call->resendEvent) {
- CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
- rxevent_Put(call->resendEvent);
- call->resendEvent = NULL;
- }
+ if (event == call->resendEvent)
+ rxevent_Put(&call->resendEvent);
rxi_CheckPeerDead(call);
- if (rxi_busyChannelError && (call->flags & RX_CALL_PEER_BUSY)) {
- rxi_CheckBusy(call);
- }
-
if (opr_queue_IsEmpty(&call->tq)) {
/* Nothing to do. This means that we've been raced, and that an
* ACK has come in between when we were triggered, and when we
rxi_Start(call, istack);
out:
+ CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
MUTEX_EXIT(&call->lock);
}
/* Since we're about to send SOME sort of packet to the peer, it's
* safe to nuke any scheduled end-of-packets ack */
- rxevent_Cancel(&call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
/* Actually send the packet, filling in more connection-specific fields */
MUTEX_EXIT(&call->lock);
(p->length <= (rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32))))
{
conn->lastSendTime = call->lastSendTime = clock_Sec();
- /* Don't count keepalive ping/acks here, so idleness can be tracked. */
- if ((p->header.type != RX_PACKET_TYPE_ACK) ||
- ((((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING) &&
- (((struct rx_ackPacket *)rx_DataOf(p))->reason !=
- RX_ACK_PING_RESPONSE)))
- call->lastSendData = call->lastSendTime;
}
}
* number of seconds. */
if (now > (call->lastReceiveTime + deadTime)) {
if (call->state == RX_STATE_ACTIVE) {
-#ifdef AFS_ADAPT_PMTU
-# if defined(KERNEL) && defined(AFS_SUN5_ENV)
- ire_t *ire;
-# if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
- netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
- ip_stack_t *ipst = ns->netstack_ip;
-# endif
- ire = ire_cache_lookup(conn->peer->host
-# if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
- , ALL_ZONES
-# if defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID)
- , NULL
-# if defined(GLOBAL_NETSTACKID)
- , ipst
-# endif
-# endif
-# endif
- );
-
- if (ire && ire->ire_max_frag > 0)
- rxi_SetPeerMtu(NULL, conn->peer->host, 0,
- ire->ire_max_frag);
-# if defined(GLOBAL_NETSTACKID)
- netstack_rele(ns);
-# endif
-# endif
-#endif /* AFS_ADAPT_PMTU */
cerror = RX_CALL_DEAD;
goto mtuout;
} else {
#ifdef RX_ENABLE_LOCKS
/* Cancel pending events */
- rxevent_Cancel(&call->delayedAckEvent, call,
- RX_CALL_REFCOUNT_DELAY);
+ rxi_CancelDelayedAckEvent(call);
rxi_rto_cancel(call);
- rxevent_Cancel(&call->keepAliveEvent, call,
- RX_CALL_REFCOUNT_ALIVE);
- rxevent_Cancel(&call->growMTUEvent, call,
- RX_CALL_REFCOUNT_MTU);
+ rxi_CancelKeepAliveEvent(call);
+ rxi_CancelGrowMTUEvent(call);
MUTEX_ENTER(&rx_refcnt_mutex);
/* if rxi_FreeCall returns 1 it has freed the call */
if (call->refCount == 0 &&
* attached process can die reasonably gracefully. */
}
- if (conn->idleDeadDetection) {
- if (conn->idleDeadTime) {
- idleDeadTime = conn->idleDeadTime + fudgeFactor;
- }
-
- if (idleDeadTime) {
- /* see if we have a non-activity timeout */
- if (call->startWait && ((call->startWait + idleDeadTime) < now) &&
- (call->flags & RX_CALL_READER_WAIT)) {
- if (call->state == RX_STATE_ACTIVE) {
- cerror = RX_CALL_TIMEOUT;
- goto mtuout;
- }
- }
+ if (conn->idleDeadTime) {
+ idleDeadTime = conn->idleDeadTime + fudgeFactor;
+ }
- if (call->lastSendData && ((call->lastSendData + idleDeadTime) < now)) {
- if (call->state == RX_STATE_ACTIVE) {
- cerror = conn->service ? conn->service->idleDeadErr : RX_CALL_IDLE;
- idle_timeout = 1;
- goto mtuout;
- }
- }
- }
+ if (idleDeadTime) {
+ /* see if we have a non-activity timeout */
+ if (call->startWait && ((call->startWait + idleDeadTime) < now)) {
+ if (call->state == RX_STATE_ACTIVE) {
+ cerror = RX_CALL_TIMEOUT;
+ goto mtuout;
+ }
+ }
}
if (conn->hardDeadTime) {
call->lastReceiveTime) {
int oldMTU = conn->peer->ifMTU;
- /* if we thought we could send more, perhaps things got worse */
- if (conn->peer->maxPacketSize > conn->lastPacketSize)
- /* maxpacketsize will be cleared in rxi_SetPeerMtu */
- newmtu = MAX(conn->peer->maxPacketSize-RX_IPUDP_SIZE,
- conn->lastPacketSize-(128+RX_IPUDP_SIZE));
+ /* If we thought we could send more, perhaps things got worse.
+ * Shrink by 128 bytes and try again. */
+ if (conn->peer->maxPacketSize < conn->lastPacketSize)
+ /* maxPacketSize will be cleared in rxi_SetPeerMtu */
+ newmtu = MAX(conn->peer->maxPacketSize + RX_HEADER_SIZE,
+ conn->lastPacketSize - 128 + RX_HEADER_SIZE);
else
- newmtu = conn->lastPacketSize-(128+RX_IPUDP_SIZE);
+ newmtu = conn->lastPacketSize - 128 + RX_HEADER_SIZE;
/* minimum capped in SetPeerMtu */
rxi_SetPeerMtu(conn->peer, 0, 0, newmtu);
/* clean up */
- conn->lastPacketSize = 0;
+ conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
/* needed so ResetCall doesn't clobber us. */
call->MTU = conn->peer->ifMTU;
struct sockaddr_in taddr;
char *tp;
char a[1] = { 0 };
+ int resched = 0;
struct iovec tmpiov[2];
osi_socket socket =
(conn->type ==
taddr.sin_family = AF_INET;
taddr.sin_port = rx_PortOf(rx_PeerOf(conn));
taddr.sin_addr.s_addr = rx_HostOf(rx_PeerOf(conn));
+ memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
#ifdef STRUCT_SOCKADDR_HAS_SA_LEN
taddr.sin_len = sizeof(struct sockaddr_in);
#endif
osi_NetSend(socket, &taddr, tmpiov, 1, 1 + sizeof(struct rx_header), 1);
MUTEX_ENTER(&conn->conn_data_lock);
+ /* We ran, so the handle is no longer needed to try to cancel ourselves. */
+ if (event == conn->natKeepAliveEvent)
+ rxevent_Put(&conn->natKeepAliveEvent);
MUTEX_ENTER(&rx_refcnt_mutex);
/* Only reschedule ourselves if the connection would not be destroyed */
- if (conn->refCount <= 1) {
- rxevent_Put(conn->natKeepAliveEvent);
- conn->natKeepAliveEvent = NULL;
- MUTEX_EXIT(&rx_refcnt_mutex);
- MUTEX_EXIT(&conn->conn_data_lock);
- rx_DestroyConnection(conn); /* drop the reference for this */
- } else {
- conn->refCount--; /* drop the reference for this */
- MUTEX_EXIT(&rx_refcnt_mutex);
- rxevent_Put(conn->natKeepAliveEvent);
- conn->natKeepAliveEvent = NULL;
- rxi_ScheduleNatKeepAliveEvent(conn);
- MUTEX_EXIT(&conn->conn_data_lock);
+ if (conn->refCount > 1)
+ resched = 1;
+ if (conn->refCount <= 0) {
+#ifdef RX_REFCOUNT_CHECK
+ osi_Assert(conn->refCount == 0);
+#endif
+ if (rx_stats_active) {
+ MUTEX_ENTER(&rx_stats_mutex);
+ rxi_lowConnRefCount++;
+ MUTEX_EXIT(&rx_stats_mutex);
+ }
}
+ MUTEX_EXIT(&rx_refcnt_mutex);
+ if (resched)
+ rxi_ScheduleNatKeepAliveEvent(conn);
+ MUTEX_EXIT(&conn->conn_data_lock);
+ putConnection(conn);
}
static void
rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn)
{
+ MUTEX_ASSERT(&conn->conn_data_lock);
if (!conn->natKeepAliveEvent && conn->secondsUntilNatPing) {
struct clock when, now;
clock_GetTime(&now);
when = now;
when.sec += conn->secondsUntilNatPing;
- MUTEX_ENTER(&rx_refcnt_mutex);
- conn->refCount++; /* hold a reference for this */
- MUTEX_EXIT(&rx_refcnt_mutex);
+ rx_GetConnection(conn);
conn->natKeepAliveEvent =
rxevent_Post(&when, &now, rxi_NatKeepAliveEvent, conn, NULL, 0);
}
struct rx_connection *conn;
afs_uint32 now;
- CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
MUTEX_ENTER(&call->lock);
- if (event == call->keepAliveEvent) {
- rxevent_Put(call->keepAliveEvent);
- call->keepAliveEvent = NULL;
- }
+ if (event == call->keepAliveEvent)
+ rxevent_Put(&call->keepAliveEvent);
now = clock_Sec();
if (rxi_CheckCall(call, 0)) {
MUTEX_EXIT(&call->lock);
+ CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
return;
}
/* Don't try to keep alive dallying calls */
if (call->state == RX_STATE_DALLY) {
MUTEX_EXIT(&call->lock);
+ CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
return;
}
}
rxi_ScheduleKeepAliveEvent(call);
MUTEX_EXIT(&call->lock);
+ CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
}
/* Does what's on the nameplate. */
struct rx_call *call = arg1;
struct rx_connection *conn;
- CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
MUTEX_ENTER(&call->lock);
- if (event == call->growMTUEvent) {
- rxevent_Put(call->growMTUEvent);
- call->growMTUEvent = NULL;
- }
+ if (event == call->growMTUEvent)
+ rxevent_Put(&call->growMTUEvent);
- if (rxi_CheckCall(call, 0)) {
- MUTEX_EXIT(&call->lock);
- return;
- }
+ if (rxi_CheckCall(call, 0))
+ goto out;
/* Don't bother with dallying calls */
- if (call->state == RX_STATE_DALLY) {
- MUTEX_EXIT(&call->lock);
- return;
- }
+ if (call->state == RX_STATE_DALLY)
+ goto out;
conn = call->conn;
*/
if ((conn->peer->maxPacketSize != 0) &&
(conn->peer->natMTU < RX_MAX_PACKET_SIZE) &&
- conn->idleDeadDetection)
+ conn->idleDeadTime)
(void)rxi_SendAck(call, NULL, 0, RX_ACK_MTU, 0);
rxi_ScheduleGrowMTUEvent(call, 0);
+out:
MUTEX_EXIT(&call->lock);
+ CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
}
static void
rxi_ScheduleKeepAliveEvent(struct rx_call *call)
{
+ MUTEX_ASSERT(&call->lock);
if (!call->keepAliveEvent) {
struct clock when, now;
clock_GetTime(&now);
}
static void
+rxi_CancelKeepAliveEvent(struct rx_call *call) {
+ MUTEX_ASSERT(&call->lock);
+ if (rxevent_Cancel(&call->keepAliveEvent))
+ CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
+}
+
+static void
rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs)
{
+ MUTEX_ASSERT(&call->lock);
if (!call->growMTUEvent) {
struct clock when, now;
}
}
-/* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
+static void
+rxi_CancelGrowMTUEvent(struct rx_call *call)
+{
+ MUTEX_ASSERT(&call->lock);
+ if (rxevent_Cancel(&call->growMTUEvent))
+ CALL_RELE(call, RX_CALL_REFCOUNT_MTU);
+}
+
+/*
+ * Increment the counter for the next connection ID, handling overflow.
+ */
+static void
+update_nextCid(void)
+{
+ /* Overflow is technically undefined behavior; avoid it. */
+ if (rx_nextCid > MAX_AFS_INT32 - (1 << RX_CIDSHIFT))
+ rx_nextCid = -1 * ((MAX_AFS_INT32 / RX_CIDSHIFT) * RX_CIDSHIFT);
+ else
+ rx_nextCid += 1 << RX_CIDSHIFT;
+}
+
static void
rxi_KeepAliveOn(struct rx_call *call)
{
rxi_ScheduleKeepAliveEvent(call);
}
-/*
- * Solely in order that callers not need to include rx_call.h
- */
-void
-rx_KeepAliveOff(struct rx_call *call)
-{
- rxi_KeepAliveOff(call);
-}
-void
-rx_KeepAliveOn(struct rx_call *call)
-{
- rxi_KeepAliveOn(call);
-}
-
static void
rxi_GrowMTUOn(struct rx_call *call)
{
struct rx_packet *packet;
MUTEX_ENTER(&conn->conn_data_lock);
- rxevent_Put(conn->delayedAbortEvent);
- conn->delayedAbortEvent = NULL;
+ if (event == conn->delayedAbortEvent)
+ rxevent_Put(&conn->delayedAbortEvent);
error = htonl(conn->error);
conn->abortCount++;
MUTEX_EXIT(&conn->conn_data_lock);
sizeof(error), 0);
rxi_FreePacket(packet);
}
+ putConnection(conn);
}
/* This routine is called to send call abort messages
struct rx_packet *packet;
MUTEX_ENTER(&call->lock);
- rxevent_Put(call->delayedAbortEvent);
- call->delayedAbortEvent = NULL;
+ if (event == call->delayedAbortEvent)
+ rxevent_Put(&call->delayedAbortEvent);
error = htonl(call->error);
call->abortCount++;
packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
}
-/* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
+/*
+ * This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
* seconds) to ask the client to authenticate itself. The routine
* issues a challenge to the client, which is obtained from the
- * security object associated with the connection */
+ * security object associated with the connection
+ *
+ * This routine is both an event handler and a function called directly;
+ * when called directly the passed |event| is NULL and the
+ * conn->conn->data>lock must must not be held. Also, when called as an
+ * an event handler, we must putConnection before we exit; but when called
+ * directly (the first challenge), we must NOT putConnection.
+ */
static void
rxi_ChallengeEvent(struct rxevent *event,
void *arg0, void *arg1, int tries)
{
struct rx_connection *conn = arg0;
+ int event_raised = 0; /* assume we were called directly */
- if (event) {
- rxevent_Put(conn->challengeEvent);
- conn->challengeEvent = NULL;
+ MUTEX_ENTER(&conn->conn_data_lock);
+ if (event != NULL && event == conn->challengeEvent) {
+ event_raised = 1; /* called as an event */
+ rxevent_Put(&conn->challengeEvent);
}
+ MUTEX_EXIT(&conn->conn_data_lock);
+
+ /* If there are no active calls it is not worth re-issuing the
+ * challenge. If the client issues another call on this connection
+ * the challenge can be requested at that time.
+ */
+ if (!rxi_HasActiveCalls(conn))
+ goto done;
if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
struct rx_packet *packet;
}
}
MUTEX_EXIT(&conn->conn_call_lock);
- return;
+ goto done;
}
packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
rxi_SendSpecial((struct rx_call *)0, conn, packet,
RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
rxi_FreePacket(packet);
+ conn->securityChallengeSent = 1;
}
clock_GetTime(&now);
when = now;
when.sec += RX_CHALLENGE_TIMEOUT;
- conn->challengeEvent =
- rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
- (tries - 1));
+ MUTEX_ENTER(&conn->conn_data_lock);
+ /* Only reschedule ourselves if not already pending. */
+ if (conn->challengeEvent == NULL) {
+ rx_GetConnection(conn);
+ conn->challengeEvent =
+ rxevent_Post(&when, &now, rxi_ChallengeEvent, conn, 0,
+ (tries - 1));
+ }
+ MUTEX_EXIT(&conn->conn_data_lock);
}
+ done:
+ if (event_raised)
+ putConnection(conn);
}
/* Call this routine to start requesting the client to authenticate
* itself. This will continue until authentication is established,
* the call times out, or an invalid response is returned. The
* security object associated with the connection is asked to create
- * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
- * defined earlier. */
+ * the challenge at this time. */
static void
rxi_ChallengeOn(struct rx_connection *conn)
{
- if (!conn->challengeEvent) {
+ int start = 0;
+ MUTEX_ENTER(&conn->conn_data_lock);
+ if (!conn->challengeEvent)
+ start = 1;
+ MUTEX_EXIT(&conn->conn_data_lock);
+ if (start) {
RXS_CreateChallenge(conn->securityObject, conn);
rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
};
int unused3)
{
struct clock now, when;
+ struct rxevent *event;
clock_GetTime(&now);
/* Find server connection structures that haven't been used for
when = now;
when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
- rxevent_Put(rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0));
+ event = rxevent_Post(&when, &now, rxi_ReapConnections, 0, NULL, 0);
+ rxevent_Put(&event);
}
taddr.sin_family = AF_INET;
taddr.sin_port = remotePort;
taddr.sin_addr.s_addr = remoteAddr;
+ memset(&taddr.sin_zero, 0, sizeof(taddr.sin_zero));
#ifdef STRUCT_SOCKADDR_HAS_SA_LEN
taddr.sin_len = sizeof(struct sockaddr_in);
#endif
#endif /* KERNEL */
LOCK_RX_INIT;
- if (rxinit_status == 1) {
+ if (!rxi_IsRunning()) {
UNLOCK_RX_INIT;
return; /* Already shutdown. */
}
+ rx_atomic_set(&rxi_running, 0);
#ifndef KERNEL
rx_port = 0;
#ifndef AFS_PTHREAD_ENV
rxi_StopListener();
#endif /* AFS_PTHREAD_ENV */
shutdown_rxevent();
- rx_SetEpoch(0);
+ rx_epoch = 0;
#ifndef AFS_PTHREAD_ENV
#ifndef AFS_USE_GETTIMEOFDAY
clock_UnInit();
rx_hashTableSize * sizeof(struct rx_connection *));
UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
- rxi_FreeAllPackets();
-
MUTEX_ENTER(&rx_quota_mutex);
rxi_dataQuota = RX_MAX_QUOTA;
rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
MUTEX_EXIT(&rx_quota_mutex);
- rxinit_status = 1;
UNLOCK_RX_INIT;
}
-#ifdef RX_ENABLE_LOCKS
-void
-osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
-{
- if (!MUTEX_ISMINE(lockaddr))
- osi_Panic("Lock not held: %s", msg);
-}
-#endif /* RX_ENABLE_LOCKS */
-
#ifndef KERNEL
/*
if (rxInterface == -1)
return;
- peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
+ peer = rxi_FindPeer(peerHost, peerPort, 0);
if (!peer)
return;
if (rpcop_stat == NULL)
return NULL;
- peer = rxi_FindPeer(peerHost, peerPort, 0, 0);
+ peer = rxi_FindPeer(peerHost, peerPort, 0);
if (!peer)
return NULL;
for (opr_queue_Scan(&processStats, cursor)) {
unsigned int num_funcs = 0, i;
struct rx_interface_stat *rpc_stat
- = opr_queue_Entry(rpc_stat, struct rx_interface_stat, entry);
+ = opr_queue_Entry(cursor, struct rx_interface_stat, entry);
num_funcs = rpc_stat->stats[0].func_total;
for (i = 0; i < num_funcs; i++) {
RXDPRINTF(RXDPRINTOUT, "%s - call=0x%p, id=%u, state=%u, mode=%u, conn=%p, epoch=%u, cid=%u, callNum=%u, connFlags=0x%x, flags=0x%x, "
"rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
"lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
- "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
- "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
+ "resendEvent=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
+ "lastSendTime=%u, lastRecvTime=%u"
#ifdef RX_ENABLE_LOCKS
", refCount=%u"
#endif
c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
(afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
(afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
- c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
- c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
+ c->resendEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
+ c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime
#ifdef RX_ENABLE_LOCKS
, (afs_uint32)c->refCount
#endif
git://git.openafs.org / openafs.git / blobdiff