rx: Use _IsLast to check for last call in queue

[openafs.git] / src / rx / rx.c

/*
 * Copyright 2000, International Business Machines Corporation and others.
 * All Rights Reserved.
 *
 * This software has been released under the terms of the IBM Public
 * License.  For details, see the LICENSE file in the top-level source
 * directory or online at http://www.openafs.org/dl/license10.html
 */

/* RX:  Extended Remote Procedure Call */

#include <afsconfig.h>
#include <afs/param.h>

#ifdef KERNEL
# include "afs/sysincludes.h"
# include "afsincludes.h"
# ifndef UKERNEL
#  include "h/types.h"
#  include "h/time.h"
#  include "h/stat.h"
#  ifdef AFS_LINUX20_ENV
#   include "h/socket.h"
#  endif
#  include "netinet/in.h"
#  ifdef AFS_SUN5_ENV
#   include "netinet/ip6.h"
#   include "inet/common.h"
#   include "inet/ip.h"
#   include "inet/ip_ire.h"
#  endif
#  include "afs/afs_args.h"
#  include "afs/afs_osi.h"
#  ifdef RX_KERNEL_TRACE
#   include "rx_kcommon.h"
#  endif
#  if   defined(AFS_AIX_ENV)
#   include "h/systm.h"
#  endif
#  ifdef RXDEBUG
#   undef RXDEBUG                       /* turn off debugging */
#  endif /* RXDEBUG */
#  if defined(AFS_SGI_ENV)
#   include "sys/debug.h"
#  endif
# else /* !UKERNEL */
#  include "afs/sysincludes.h"
#  include "afsincludes.h"
# endif /* !UKERNEL */
# include "afs/lock.h"
# include "rx_kmutex.h"
# include "rx_kernel.h"
# define        AFSOP_STOP_RXCALLBACK   210     /* Stop CALLBACK process */
# define        AFSOP_STOP_AFS          211     /* Stop AFS process */
# define        AFSOP_STOP_BKG          212     /* Stop BKG process */
extern afs_int32 afs_termState;
# ifdef AFS_AIX41_ENV
#  include "sys/lockl.h"
#  include "sys/lock_def.h"
# endif /* AFS_AIX41_ENV */
# include "afs/rxgen_consts.h"
#else /* KERNEL */
# include <roken.h>

# ifdef AFS_NT40_ENV
#  include <afs/afsutil.h>
#  include <WINNT\afsreg.h>
# endif

# include <afs/opr.h>

# include "rx_user.h"
#endif /* KERNEL */

#include <opr/queue.h>
#include <hcrypto/rand.h>

#include "rx.h"
#include "rx_clock.h"
#include "rx_atomic.h"
#include "rx_globals.h"
#include "rx_trace.h"
#include "rx_internal.h"
#include "rx_stats.h"
#include "rx_event.h"

#include "rx_peer.h"
#include "rx_conn.h"
#include "rx_call.h"
#include "rx_packet.h"
#include "rx_server.h"

#include <afs/rxgen_consts.h>

#ifndef KERNEL
#ifdef AFS_PTHREAD_ENV
#ifndef AFS_NT40_ENV
int (*registerProgram) (pid_t, char *) = 0;
int (*swapNameProgram) (pid_t, const char *, char *) = 0;
#endif
#else
int (*registerProgram) (PROCESS, char *) = 0;
int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
#endif
#endif

/* Local static routines */
static void rxi_DestroyConnectionNoLock(struct rx_connection *conn);
static void rxi_ComputeRoundTripTime(struct rx_packet *, struct rx_ackPacket *,
                                     struct rx_call *, struct rx_peer *,
                                     struct clock *);
static void rxi_Resend(struct rxevent *event, void *arg0, void *arg1,
                       int istack);
static void rxi_SendDelayedAck(struct rxevent *event, void *call,
                               void *dummy, int dummy2);
static void rxi_SendDelayedCallAbort(struct rxevent *event, void *arg1,
                                     void *dummy, int dummy2);
static void rxi_SendDelayedConnAbort(struct rxevent *event, void *arg1,
                                     void *unused, int unused2);
static void rxi_ReapConnections(struct rxevent *unused, void *unused1,
                                void *unused2, int unused3);
static struct rx_packet *rxi_SendCallAbort(struct rx_call *call,
                                           struct rx_packet *packet,
                                           int istack, int force);
static void rxi_AckAll(struct rx_call *call);
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,
                            int *unknownService);
static struct rx_packet
        *rxi_ReceiveDataPacket(struct rx_call *call, struct rx_packet *np,
                               int istack, osi_socket socket,
                               afs_uint32 host, u_short port, int *tnop,
                               struct rx_call **newcallp);
static struct rx_packet
        *rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
                              int istack);
static struct rx_packet
        *rxi_ReceiveResponsePacket(struct rx_connection *conn,
                                   struct rx_packet *np, int istack);
static struct rx_packet
        *rxi_ReceiveChallengePacket(struct rx_connection *conn,
                                    struct rx_packet *np, int istack);
static void rxi_AttachServerProc(struct rx_call *call, osi_socket socket,
                                 int *tnop, struct rx_call **newcallp);
static void rxi_ClearTransmitQueue(struct rx_call *call, int force);
static void rxi_ClearReceiveQueue(struct rx_call *call);
static void rxi_ResetCall(struct rx_call *call, int newcall);
static void rxi_ScheduleKeepAliveEvent(struct rx_call *call);
static void rxi_ScheduleNatKeepAliveEvent(struct rx_connection *conn);
static void rxi_ScheduleGrowMTUEvent(struct rx_call *call, int secs);
static void rxi_KeepAliveOn(struct rx_call *call);
static void rxi_GrowMTUOn(struct rx_call *call);
static int 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 {
    rx_atomic_t rxi_start_aborted; /* rxi_start awoke after rxi_Send in error.*/
    rx_atomic_t rxi_start_in_error;
} rx_tq_debug;
#endif /* RX_ENABLE_LOCKS */

/* Constant delay time before sending an acknowledge of the last packet
 * received.  This is to avoid sending an extra acknowledge when the
 * client is about to make another call, anyway, or the server is
 * about to respond.
 *
 * The lastAckDelay may not exceeed 400ms without causing peers to
 * unecessarily timeout.
 */
struct clock rx_lastAckDelay = {0, 400000};

/* Constant delay time before sending a soft ack when none was requested.
 * This is to make sure we send soft acks before the sender times out,
 * Normally we wait and send a hard ack when the receiver consumes the packet
 *
 * This value has been 100ms in all shipping versions of OpenAFS. Changing it
 * will require changes to the peer's RTT calculations.
 */
struct clock rx_softAckDelay = {0, 100000};

/*
 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
 * currently allocated within rx.  This number is used to allocate the
 * memory required to return the statistics when queried.
 * Protected by the rx_rpc_stats mutex.
 */

static unsigned int rxi_rpc_peer_stat_cnt;

/*
 * rxi_rpc_process_stat_cnt counts the total number of local process stat
 * structures currently allocated within rx.  The number is used to allocate
 * the memory required to return the statistics when queried.
 * Protected by the rx_rpc_stats mutex.
 */

static unsigned int rxi_rpc_process_stat_cnt;

rx_atomic_t rx_nWaiting = RX_ATOMIC_INIT(0);
rx_atomic_t rx_nWaited = RX_ATOMIC_INIT(0);

/* Incoming calls wait on this queue when there are no available
 * server processes */
struct opr_queue rx_incomingCallQueue;

/* Server processes wait on this queue when there are no appropriate
 * calls to process */
struct opr_queue rx_idleServerQueue;

/* List of free rx_serverQueueEntry structs */
struct opr_queue rx_freeServerQueue;

#if !defined(offsetof)
#include <stddef.h>             /* for definition of offsetof() */
#endif

#ifdef RX_ENABLE_LOCKS
afs_kmutex_t rx_atomic_mutex;
static afs_kmutex_t freeSQEList_lock;
#endif

/* Forward prototypes */
static struct rx_call * rxi_NewCall(struct rx_connection *, int);

static_inline void
putConnection (struct rx_connection *conn) {
    MUTEX_ENTER(&rx_refcnt_mutex);
    conn->refCount--;
    MUTEX_EXIT(&rx_refcnt_mutex);
}

#ifdef AFS_PTHREAD_ENV

/*
 * Use procedural initialization of mutexes/condition variables
 * to ease NT porting
 */

extern afs_kmutex_t rx_quota_mutex;
extern afs_kmutex_t rx_pthread_mutex;
extern afs_kmutex_t rx_packets_mutex;
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_kmutex_t listener_mutex;
extern afs_kmutex_t rx_if_init_mutex;
extern afs_kmutex_t rx_if_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 afs_kmutex_t rx_debug_mutex;
static afs_kmutex_t rx_rpc_stats;

static void
rxi_InitPthread(void)
{
    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);
#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();
#endif /* RX_LOCKS_DB */
    MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
               0);
    CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
            0);
    MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
               0);
    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 */
}

pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
#define INIT_PTHREAD_LOCKS osi_Assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
/*
 * The rx_stats_mutex mutex protects the following global variables:
 * rxi_lowConnRefCount
 * rxi_lowPeerRefCount
 * rxi_nCalls
 * rxi_Alloccnt
 * rxi_Allocsize
 * rx_tq_debug
 * rx_stats
 */

/*
 * The rx_quota_mutex mutex protects the following global variables:
 * rxi_dataQuota
 * rxi_minDeficit
 * rxi_availProcs
 * rxi_totalMin
 */

/*
 * The rx_freePktQ_lock protects the following global variables:
 * rx_nFreePackets
 */

/*
 * The rx_packets_mutex mutex protects the following global variables:
 * rx_nPackets
 * rx_TSFPQLocalMax
 * rx_TSFPQGlobSize
 * rx_TSFPQMaxProcs
 */

/*
 * The rx_pthread_mutex mutex protects the following global variables:
 * rxi_fcfs_thread_num
 */
#else
#define INIT_PTHREAD_LOCKS
#endif


/* Variables for handling the minProcs implementation.  availProcs gives the
 * number of threads available in the pool at this moment (not counting dudes
 * executing right now).  totalMin gives the total number of procs required
 * for handling all minProcs requests.  minDeficit is a dynamic variable
 * tracking the # of procs required to satisfy all of the remaining minProcs
 * demands.
 * For fine grain locking to work, the quota check and the reservation of
 * a server thread has to come while rxi_availProcs and rxi_minDeficit
 * are locked. To this end, the code has been modified under #ifdef
 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
 * same time. A new function, ReturnToServerPool() returns the allocation.
 *
 * A call can be on several queue's (but only one at a time). When
 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
 * that no one else is touching the queue. To this end, we store the address
 * of the queue lock in the call structure (under the call lock) when we
 * put the call on a queue, and we clear the call_queue_lock when the
 * call is removed from a queue (once the call lock has been obtained).
 * This allows rxi_ResetCall to safely synchronize with others wishing
 * to manipulate the queue.
 */

#if defined(RX_ENABLE_LOCKS)
static afs_kmutex_t rx_rpc_stats;
#endif

/* We keep a "last conn pointer" in rxi_FindConnection. The odds are
** pretty good that the next packet coming in is from the same connection
** as the last packet, since we're send multiple packets in a transmit window.
*/
struct rx_connection *rxLastConn = 0;

#ifdef RX_ENABLE_LOCKS
/* The locking hierarchy for rx fine grain locking is composed of these
 * 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:
 *      rx_freeCallQueue_lock
 *      rxi_keyCreate_lock
 * rx_serverPool_lock
 * freeSQEList_lock
 *
 * serverQueueEntry->lock
 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
 * rx_rpc_stats
 * peer->lock - locks peer data fields.
 * conn_data_lock - that more than one thread is not updating a conn data
 *                  field at the same time.
 * rx_freePktQ_lock
 *
 * lowest level:
 *      multi_handle->lock
 *      rxevent_lock
 *      rx_packets_mutex
 *      rx_stats_mutex
 *      rx_refcnt_mutex
 *      rx_atomic_mutex
 *
 * Do we need a lock to protect the peer field in the conn structure?
 *      conn->peer was previously a constant for all intents and so has no
 *      lock protecting this field. The multihomed client delta introduced
 *      a RX code change : change the peer field in the connection structure
 *      to that remote interface from which the last packet for this
 *      connection was sent out. This may become an issue if further changes
 *      are made.
 */
#define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
#define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
#ifdef RX_LOCKS_DB
/* rxdb_fileID is used to identify the lock location, along with line#. */
static int rxdb_fileID = RXDB_FILE_RX;
#endif /* RX_LOCKS_DB */
#else /* RX_ENABLE_LOCKS */
#define SET_CALL_QUEUE_LOCK(C, L)
#define CLEAR_CALL_QUEUE_LOCK(C)
#endif /* RX_ENABLE_LOCKS */
struct rx_serverQueueEntry *rx_waitForPacket = 0;

/*
 * 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.
 */
#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_RX_INIT
# define UNLOCK_RX_INIT
#endif

/* ------------Exported Interfaces------------- */

static rx_atomic_t rxi_running = RX_ATOMIC_INIT(0);
int
rxi_IsRunning(void)
{
    return rx_atomic_read(&rxi_running);
}

/* Initialize rx.  A port number may be mentioned, in which case this
 * becomes the default port number for any service installed later.
 * If 0 is provided for the port number, a random port will be chosen
 * by the kernel.  Whether this will ever overlap anything in
 * /etc/services is anybody's guess...  Returns 0 on success, -1 on
 * error. */
int
rx_InitHost(u_int host, u_int port)
{
#ifdef KERNEL
    osi_timeval_t tv;
#else /* KERNEL */
    struct timeval tv;
#endif /* KERNEL */
    char *htable, *ptable;

    SPLVAR;

    INIT_PTHREAD_LOCKS;
    LOCK_RX_INIT;
    if (rxi_IsRunning()) {
        UNLOCK_RX_INIT;
        return 0; /* already started */
    }
#ifdef RXDEBUG
    rxi_DebugInit();
#endif
#ifdef AFS_NT40_ENV
    if (afs_winsockInit() < 0)
        goto error;
#endif

#ifndef KERNEL
    /*
     * Initialize anything necessary to provide a non-premptive threading
     * environment.
     */
    rxi_InitializeThreadSupport();
#endif

    /* Allocate and initialize a socket for client and perhaps server
     * connections. */

    rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
    if (rx_socket == OSI_NULLSOCKET) {
        goto addrinuse;
    }
#if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
#ifdef RX_LOCKS_DB
    rxdb_init();
#endif /* RX_LOCKS_DB */
    MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_quota_mutex, "rx_quota_mutex", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_atomic_mutex, "rx_atomic_mutex", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_pthread_mutex, "rx_pthread_mutex", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_packets_mutex, "rx_packets_mutex", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_refcnt_mutex, "rx_refcnt_mutex", MUTEX_DEFAULT, 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(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
               0);
    CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
            0);
    MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
               0);
    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");
#endif /* AFS_HPUX110_ENV */
#endif /* RX_ENABLE_LOCKS && KERNEL */

    rxi_nCalls = 0;
    rx_connDeadTime = 12;
    rx_tranquil = 0;            /* reset flag */
    rxi_ResetStatistics();
    htable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
    PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *));     /* XXXXX */
    memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
    ptable = osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
    PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *));   /* XXXXX */
    memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));

    /* Malloc up a bunch of packets & buffers */
    rx_nFreePackets = 0;
    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)
        rx_extraPackets = rxi_nSendFrags * rx_maxSendWindow;

    /* allocate the initial free packet pool */
#ifdef RX_ENABLE_TSFPQ
    rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
#else /* RX_ENABLE_TSFPQ */
    rxi_MorePackets(rx_extraPackets + RX_MAX_QUOTA + 2);        /* fudge */
#endif /* RX_ENABLE_TSFPQ */
    rx_CheckPackets();

    NETPRI;

    clock_Init();

#if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
    tv.tv_sec = clock_now.sec;
    tv.tv_usec = clock_now.usec;
    srand((unsigned int)tv.tv_usec);
#else
    osi_GetTime(&tv);
#endif
    if (port) {
        rx_port = port;
    } else {
#if defined(KERNEL) && !defined(UKERNEL)
        /* Really, this should never happen in a real kernel */
        rx_port = 0;
#else
        struct sockaddr_in addr;
#ifdef AFS_NT40_ENV
        int addrlen = sizeof(addr);
#else
        socklen_t addrlen = sizeof(addr);
#endif
        if (getsockname((intptr_t)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
            rxi_Finalize_locked();
            osi_Free(htable, rx_hashTableSize * sizeof(struct rx_connection *));
            goto error;
        }
        rx_port = addr.sin_port;
#endif
    }
    rx_stats.minRtt.sec = 9999999;
    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);
    /* *Slightly* random start time for the cid.  This is just to help
     * out with the hashing function at the peer */
    rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
    rx_connHashTable = (struct rx_connection **)htable;
    rx_peerHashTable = (struct rx_peer **)ptable;

    rx_hardAckDelay.sec = 0;
    rx_hardAckDelay.usec = 100000;      /* 100 milliseconds */

    rxevent_Init(20, rxi_ReScheduleEvents);

    /* Initialize various global queues */
    opr_queue_Init(&rx_idleServerQueue);
    opr_queue_Init(&rx_freeServerQueue);
    opr_queue_Init(&rx_incomingCallQueue);
    opr_queue_Init(&rx_freeCallQueue);

#if defined(AFS_NT40_ENV) && !defined(KERNEL)
    /* Initialize our list of usable IP addresses. */
    rx_GetIFInfo();
#endif

    /* Start listener process (exact function is dependent on the
     * implementation environment--kernel or user space) */
    rxi_StartListener();

    USERPRI;

    rx_atomic_set(&rxi_running, 1);
    UNLOCK_RX_INIT;

    return 0;

 addrinuse:
    UNLOCK_RX_INIT;
    return RX_ADDRINUSE;

 error:
    UNLOCK_RX_INIT;
    return -1;
}

int
rx_Init(u_int port)
{
    return rx_InitHost(htonl(INADDR_ANY), port);
}

/* RTT Timer
 * ---------
 *
 * The rxi_rto functions implement a TCP (RFC2988) style algorithm for
 * maintaing the round trip timer.
 *
 */

/*!
 * Start a new RTT timer for a given call and packet.
 *
 * There must be no resendEvent already listed for this call, otherwise this
 * will leak events - intended for internal use within the RTO code only
 *
 * @param[in] call
 *      the RX call to start the timer for
 * @param[in] lastPacket
 *      a flag indicating whether the last packet has been sent or not
 *
 * @pre call must be locked before calling this function
 *
 */
static_inline void
rxi_rto_startTimer(struct rx_call *call, int lastPacket, int istack)
{
    struct clock now, retryTime;

    MUTEX_ASSERT(&call->lock);
    clock_GetTime(&now);
    retryTime = now;

    clock_Add(&retryTime, &call->rto);

    /* If we're sending the last packet, and we're the client, then the server
     * may wait for an additional 400ms before returning the ACK, wait for it
     * rather than hitting a timeout */
    if (lastPacket && call->conn->type == RX_CLIENT_CONNECTION)
        clock_Addmsec(&retryTime, 400);

    CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
    call->resendEvent = rxevent_Post(&retryTime, &now, rxi_Resend,
                                     call, NULL, istack);
}

/*!
 * Cancel an RTT timer for a given call.
 *
 *
 * @param[in] call
 *      the RX call to cancel the timer for
 *
 * @pre call must be locked before calling this function
 *
 */

static_inline void
rxi_rto_cancel(struct rx_call *call)
{
    MUTEX_ASSERT(&call->lock);
    if (rxevent_Cancel(&call->resendEvent))
        CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
}

/*!
 * Tell the RTO timer that we have sent a packet.
 *
 * If the timer isn't already running, then start it. If the timer is running,
 * then do nothing.
 *
 * @param[in] call
 *      the RX call that the packet has been sent on
 * @param[in] lastPacket
 *      A flag which is true if this is the last packet for the call
 *
 * @pre The call must be locked before calling this function
 *
 */

static_inline void
rxi_rto_packet_sent(struct rx_call *call, int lastPacket, int istack)
{
    if (call->resendEvent)
        return;

    rxi_rto_startTimer(call, lastPacket, istack);
}

/*!
 * Tell the RTO timer that we have received an new ACK message
 *
 * This function should be called whenever a call receives an ACK that
 * acknowledges new packets. Whatever happens, we stop the current timer.
 * If there are unacked packets in the queue which have been sent, then
 * we restart the timer from now. Otherwise, we leave it stopped.
 *
 * @param[in] call
 *      the RX call that the ACK has been received on
 */

static_inline void
rxi_rto_packet_acked(struct rx_call *call, int istack)
{
    struct opr_queue *cursor;

    rxi_rto_cancel(call);

    if (opr_queue_IsEmpty(&call->tq))
        return;

    for (opr_queue_Scan(&call->tq, cursor)) {
        struct rx_packet *p = opr_queue_Entry(cursor, struct rx_packet, entry);
        if (p->header.seq > call->tfirst + call->twind)
            return;

        if (!(p->flags & RX_PKTFLAG_ACKED) && p->flags & RX_PKTFLAG_SENT) {
            rxi_rto_startTimer(call, p->header.flags & RX_LAST_PACKET, istack);
            return;
        }
    }
}


/**
 * Set an initial round trip timeout for a peer connection
 *
 * @param[in] secs The timeout to set in seconds
 */

void
rx_rto_setPeerTimeoutSecs(struct rx_peer *peer, int secs) {
    peer->rtt = secs * 8000;
}

/**
 * Set a delayed ack event on the specified call for the given time
 *
 * @param[in] call - the call on which to set the event
 * @param[in] offset - the delay from now after which the event fires
 */
void
rxi_PostDelayedAckEvent(struct rx_call *call, struct clock *offset)
{
    struct clock now, when;

    MUTEX_ASSERT(&call->lock);
    clock_GetTime(&now);
    when = now;
    clock_Add(&when, offset);

    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);

        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);
        call->delayedAckTime = when;
    }
}

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.
 */
#ifdef RX_ENABLE_LOCKS
/* This verion of QuotaOK reserves quota if it's ok while the
 * rx_serverPool_lock is held.  Return quota using ReturnToServerPool().
 */
static int
QuotaOK(struct rx_service *aservice)
{
    /* check if over max quota */
    if (aservice->nRequestsRunning >= aservice->maxProcs) {
        return 0;
    }

    /* under min quota, we're OK */
    /* otherwise, can use only if there are enough to allow everyone
     * to go to their min quota after this guy starts.
     */

    MUTEX_ENTER(&rx_quota_mutex);
    if ((aservice->nRequestsRunning < aservice->minProcs)
        || (rxi_availProcs > rxi_minDeficit)) {
        aservice->nRequestsRunning++;
        /* just started call in minProcs pool, need fewer to maintain
         * guarantee */
        if (aservice->nRequestsRunning <= aservice->minProcs)
            rxi_minDeficit--;
        rxi_availProcs--;
        MUTEX_EXIT(&rx_quota_mutex);
        return 1;
    }
    MUTEX_EXIT(&rx_quota_mutex);

    return 0;
}

static void
ReturnToServerPool(struct rx_service *aservice)
{
    aservice->nRequestsRunning--;
    MUTEX_ENTER(&rx_quota_mutex);
    if (aservice->nRequestsRunning < aservice->minProcs)
        rxi_minDeficit++;
    rxi_availProcs++;
    MUTEX_EXIT(&rx_quota_mutex);
}

#else /* RX_ENABLE_LOCKS */
static int
QuotaOK(struct rx_service *aservice)
{
    int rc = 0;
    /* under min quota, we're OK */
    if (aservice->nRequestsRunning < aservice->minProcs)
        return 1;

    /* check if over max quota */
    if (aservice->nRequestsRunning >= aservice->maxProcs)
        return 0;

    /* otherwise, can use only if there are enough to allow everyone
     * to go to their min quota after this guy starts.
     */
    MUTEX_ENTER(&rx_quota_mutex);
    if (rxi_availProcs > rxi_minDeficit)
        rc = 1;
    MUTEX_EXIT(&rx_quota_mutex);
    return rc;
}
#endif /* RX_ENABLE_LOCKS */

#ifndef KERNEL
/* Called by rx_StartServer to start up lwp's to service calls.
   NExistingProcs gives the number of procs already existing, and which
   therefore needn't be created. */
static void
rxi_StartServerProcs(int nExistingProcs)
{
    struct rx_service *service;
    int i;
    int maxdiff = 0;
    int nProcs = 0;

    /* For each service, reserve N processes, where N is the "minimum"
     * number of processes that MUST be able to execute a request in parallel,
     * at any time, for that process.  Also compute the maximum difference
     * between any service's maximum number of processes that can run
     * (i.e. the maximum number that ever will be run, and a guarantee
     * that this number will run if other services aren't running), and its
     * minimum number.  The result is the extra number of processes that
     * we need in order to provide the latter guarantee */
    for (i = 0; i < RX_MAX_SERVICES; i++) {
        int diff;
        service = rx_services[i];
        if (service == (struct rx_service *)0)
            break;
        nProcs += service->minProcs;
        diff = service->maxProcs - service->minProcs;
        if (diff > maxdiff)
            maxdiff = diff;
    }
    nProcs += maxdiff;          /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
    nProcs -= nExistingProcs;   /* Subtract the number of procs that were previously created for use as server procs */
    for (i = 0; i < nProcs; i++) {
        rxi_StartServerProc(rx_ServerProc, rx_stackSize);
    }
}
#endif /* KERNEL */

#ifdef AFS_NT40_ENV
/* This routine is only required on Windows */
void
rx_StartClientThread(void)
{
#ifdef AFS_PTHREAD_ENV
    pthread_t pid;
    pid = pthread_self();
#endif /* AFS_PTHREAD_ENV */
}
#endif /* AFS_NT40_ENV */

/* This routine must be called if any services are exported.  If the
 * donateMe flag is set, the calling process is donated to the server
 * process pool */
void
rx_StartServer(int donateMe)
{
    struct rx_service *service;
    int i;
    SPLVAR;
    clock_NewTime();

    NETPRI;
    /* Start server processes, if necessary (exact function is dependent
     * on the implementation environment--kernel or user space).  DonateMe
     * will be 1 if there is 1 pre-existing proc, i.e. this one.  In this
     * case, one less new proc will be created rx_StartServerProcs.
     */
    rxi_StartServerProcs(donateMe);

    /* count up the # of threads in minProcs, and add set the min deficit to
     * be that value, too.
     */
    for (i = 0; i < RX_MAX_SERVICES; i++) {
        service = rx_services[i];
        if (service == (struct rx_service *)0)
            break;
        MUTEX_ENTER(&rx_quota_mutex);
        rxi_totalMin += service->minProcs;
        /* below works even if a thread is running, since minDeficit would
         * still have been decremented and later re-incremented.
         */
        rxi_minDeficit += service->minProcs;
        MUTEX_EXIT(&rx_quota_mutex);
    }

    /* Turn on reaping of idle server connections */
    rxi_ReapConnections(NULL, NULL, NULL, 0);

    USERPRI;

    if (donateMe) {
#ifndef AFS_NT40_ENV
#ifndef KERNEL
        char name[32];
        static int nProcs;
#ifdef AFS_PTHREAD_ENV
        pid_t pid;
        pid = afs_pointer_to_int(pthread_self());
#else /* AFS_PTHREAD_ENV */
        PROCESS pid;
        LWP_CurrentProcess(&pid);
#endif /* AFS_PTHREAD_ENV */

        sprintf(name, "srv_%d", ++nProcs);
        if (registerProgram)
            (*registerProgram) (pid, name);
#endif /* KERNEL */
#endif /* AFS_NT40_ENV */
        rx_ServerProc(NULL);    /* Never returns */
    }
#ifdef RX_ENABLE_TSFPQ
    /* no use leaving packets around in this thread's local queue if
     * it isn't getting donated to the server thread pool.
     */
    rxi_FlushLocalPacketsTSFPQ();
#endif /* RX_ENABLE_TSFPQ */
    return;
}

/* Create a new client connection to the specified service, using the
 * specified security object to implement the security model for this
 * connection. */
struct rx_connection *
rx_NewConnection(afs_uint32 shost, u_short sport, u_short sservice,
                 struct rx_securityClass *securityObject,
                 int serviceSecurityIndex)
{
    int hashindex, i;
    struct rx_connection *conn;
    int code;

    SPLVAR;

    clock_NewTime();
    dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %p, "
         "serviceSecurityIndex %d)\n",
         ntohl(shost), ntohs(sport), sservice, securityObject,
         serviceSecurityIndex));

    /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
     * the case of kmem_alloc? */
    conn = rxi_AllocConnection();
#ifdef  RX_ENABLE_LOCKS
    MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
    CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
#endif
    NETPRI;
    MUTEX_ENTER(&rx_connHashTable_lock);
    conn->type = RX_CLIENT_CONNECTION;
    conn->epoch = rx_epoch;
    conn->cid = rx_nextCid;
    update_nextCid();
    conn->peer = rxi_FindPeer(shost, sport, 1);
    conn->serviceId = sservice;
    conn->securityObject = securityObject;
    conn->securityData = (void *) 0;
    conn->securityIndex = serviceSecurityIndex;
    rx_SetConnDeadTime(conn, rx_connDeadTime);
    rx_SetConnSecondsUntilNatPing(conn, 0);
    conn->ackRate = RX_FAST_ACK_RATE;
    conn->nSpecific = 0;
    conn->specific = NULL;
    conn->challengeEvent = NULL;
    conn->delayedAbortEvent = NULL;
    conn->abortCount = 0;
    conn->error = 0;
    for (i = 0; i < RX_MAXCALLS; i++) {
        conn->twind[i] = rx_initSendWindow;
        conn->rwind[i] = rx_initReceiveWindow;
        conn->lastBusy[i] = 0;
    }

    code = RXS_NewConnection(securityObject, conn);
    hashindex =
        CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);

    conn->refCount++;           /* no lock required since only this thread knows... */
    conn->next = rx_connHashTable[hashindex];
    rx_connHashTable[hashindex] = conn;
    if (rx_stats_active)
        rx_atomic_inc(&rx_stats.nClientConns);
    MUTEX_EXIT(&rx_connHashTable_lock);
    USERPRI;
    if (code) {
        rxi_ConnectionError(conn, code);
    }
    return conn;
}

/**
 * Ensure a connection's timeout values are valid.
 *
 * @param[in] conn The connection to check
 *
 * @post conn->secondUntilDead <= conn->idleDeadTime <= conn->hardDeadTime,
 *       unless idleDeadTime and/or hardDeadTime are not set
 * @internal
 */
static void
rxi_CheckConnTimeouts(struct rx_connection *conn)
{
    /* a connection's timeouts must have the relationship
     * deadTime <= idleDeadTime <= hardDeadTime. Otherwise, for example, a
     * total loss of network to a peer may cause an idle timeout instead of a
     * dead timeout, simply because the idle timeout gets hit first. Also set
     * a minimum deadTime of 6, just to ensure it doesn't get set too low. */
    /* this logic is slightly complicated by the fact that
     * idleDeadTime/hardDeadTime may not be set at all, but it's not too bad.
     */
    conn->secondsUntilDead = MAX(conn->secondsUntilDead, 6);
    if (conn->idleDeadTime) {
        conn->idleDeadTime = MAX(conn->idleDeadTime, conn->secondsUntilDead);
    }
    if (conn->hardDeadTime) {
        if (conn->idleDeadTime) {
            conn->hardDeadTime = MAX(conn->idleDeadTime, conn->hardDeadTime);
        } else {
            conn->hardDeadTime = MAX(conn->secondsUntilDead, conn->hardDeadTime);
        }
    }
}

void
rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
{
    /* The idea is to set the dead time to a value that allows several
     * keepalives to be dropped without timing out the connection. */
    conn->secondsUntilDead = seconds;
    rxi_CheckConnTimeouts(conn);
    conn->secondsUntilPing = conn->secondsUntilDead / 6;
}

void
rx_SetConnHardDeadTime(struct rx_connection *conn, int seconds)
{
    conn->hardDeadTime = seconds;
    rxi_CheckConnTimeouts(conn);
}

void
rx_SetConnIdleDeadTime(struct rx_connection *conn, int seconds)
{
    conn->idleDeadTime = seconds;
    rxi_CheckConnTimeouts(conn);
}

int rxi_lowPeerRefCount = 0;
int rxi_lowConnRefCount = 0;

/*
 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
 * NOTE: must not be called with rx_connHashTable_lock held.
 */
static void
rxi_CleanupConnection(struct rx_connection *conn)
{
    /* Notify the service exporter, if requested, that this connection
     * is being destroyed */
    if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
        (*conn->service->destroyConnProc) (conn);

    /* Notify the security module that this connection is being destroyed */
    RXS_DestroyConnection(conn->securityObject, conn);

    /* If this is the last connection using the rx_peer struct, set its
     * idle time to now. rxi_ReapConnections will reap it if it's still
     * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
     */
    MUTEX_ENTER(&rx_peerHashTable_lock);
    if (conn->peer->refCount < 2) {
        conn->peer->idleWhen = clock_Sec();
        if (conn->peer->refCount < 1) {
            conn->peer->refCount = 1;
            if (rx_stats_active) {
                MUTEX_ENTER(&rx_stats_mutex);
                rxi_lowPeerRefCount++;
                MUTEX_EXIT(&rx_stats_mutex);
            }
        }
    }
    conn->peer->refCount--;
    MUTEX_EXIT(&rx_peerHashTable_lock);

    if (rx_stats_active)
    {
        if (conn->type == RX_SERVER_CONNECTION)
            rx_atomic_dec(&rx_stats.nServerConns);
        else
            rx_atomic_dec(&rx_stats.nClientConns);
    }
#ifndef KERNEL
    if (conn->specific) {
        int i;
        for (i = 0; i < conn->nSpecific; i++) {
            if (conn->specific[i] && rxi_keyCreate_destructor[i])
                (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
            conn->specific[i] = NULL;
        }
        free(conn->specific);
    }
    conn->specific = NULL;
    conn->nSpecific = 0;
#endif /* !KERNEL */

    MUTEX_DESTROY(&conn->conn_call_lock);
    MUTEX_DESTROY(&conn->conn_data_lock);
    CV_DESTROY(&conn->conn_call_cv);

    rxi_FreeConnection(conn);
}

/* Destroy the specified connection */
void
rxi_DestroyConnection(struct rx_connection *conn)
{
    MUTEX_ENTER(&rx_connHashTable_lock);
    rxi_DestroyConnectionNoLock(conn);
    /* conn should be at the head of the cleanup list */
    if (conn == rx_connCleanup_list) {
        rx_connCleanup_list = rx_connCleanup_list->next;
        MUTEX_EXIT(&rx_connHashTable_lock);
        rxi_CleanupConnection(conn);
    }
#ifdef RX_ENABLE_LOCKS
    else {
        MUTEX_EXIT(&rx_connHashTable_lock);
    }
#endif /* RX_ENABLE_LOCKS */
}

static void
rxi_DestroyConnectionNoLock(struct rx_connection *conn)
{
    struct rx_connection **conn_ptr;
    int havecalls = 0;
    int i;
    SPLVAR;

    clock_NewTime();

    NETPRI;
    MUTEX_ENTER(&conn->conn_data_lock);
    MUTEX_ENTER(&rx_refcnt_mutex);
    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++;
            MUTEX_EXIT(&rx_stats_mutex);
        }
    }

    if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
        /* Busy; wait till the last guy before proceeding */
        MUTEX_EXIT(&rx_refcnt_mutex);
        MUTEX_EXIT(&conn->conn_data_lock);
        USERPRI;
        return;
    }

    /* If the client previously called rx_NewCall, but it is still
     * waiting, treat this as a running call, and wait to destroy the
     * connection later when the call completes. */
    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;
    }
    MUTEX_EXIT(&rx_refcnt_mutex);
    MUTEX_EXIT(&conn->conn_data_lock);

    /* Check for extant references to this connection */
    MUTEX_ENTER(&conn->conn_call_lock);
    for (i = 0; i < RX_MAXCALLS; i++) {
        struct rx_call *call = conn->call[i];
        if (call) {
            havecalls = 1;
            if (conn->type == RX_CLIENT_CONNECTION) {
                MUTEX_ENTER(&call->lock);
                if (call->delayedAckEvent) {
                    /* Push the final acknowledgment out now--there
                     * won't be a subsequent call to acknowledge the
                     * last reply packets */
                    rxi_CancelDelayedAckEvent(call);
                    if (call->state == RX_STATE_PRECALL
                        || call->state == RX_STATE_ACTIVE) {
                        rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
                    } else {
                        rxi_AckAll(call);
                    }
                }
                MUTEX_EXIT(&call->lock);
            }
        }
    }
    MUTEX_EXIT(&conn->conn_call_lock);

#ifdef RX_ENABLE_LOCKS
    if (!havecalls) {
        if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
            MUTEX_EXIT(&conn->conn_data_lock);
        } else {
            /* Someone is accessing a packet right now. */
            havecalls = 1;
        }
    }
#endif /* RX_ENABLE_LOCKS */

    if (havecalls) {
        /* Don't destroy the connection if there are any call
         * structures still in use */
        MUTEX_ENTER(&conn->conn_data_lock);
        conn->flags |= RX_CONN_DESTROY_ME;
        MUTEX_EXIT(&conn->conn_data_lock);
        USERPRI;
        return;
    }

    /* Remove from connection hash table before proceeding */
    conn_ptr =
        &rx_connHashTable[CONN_HASH
                          (peer->host, peer->port, conn->cid, conn->epoch,
                           conn->type)];
    for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
        if (*conn_ptr == conn) {
            *conn_ptr = conn->next;
            break;
        }
    }
    /* if the conn that we are destroying was the last connection, then we
     * clear rxLastConn as well */
    if (rxLastConn == conn)
        rxLastConn = 0;

    /* Make sure the connection is completely reset before deleting it. */
    /*
     * 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
     * in the routines we call to inform others that this connection is
     * being destroyed. */
    conn->next = rx_connCleanup_list;
    rx_connCleanup_list = conn;
}

/* Externally available version */
void
rx_DestroyConnection(struct rx_connection *conn)
{
    SPLVAR;

    NETPRI;
    rxi_DestroyConnection(conn);
    USERPRI;
}

void
rx_GetConnection(struct rx_connection *conn)
{
    SPLVAR;

    NETPRI;
    MUTEX_ENTER(&rx_refcnt_mutex);
    conn->refCount++;
    MUTEX_EXIT(&rx_refcnt_mutex);
    USERPRI;
}

#ifdef RX_ENABLE_LOCKS
/* Wait for the transmit queue to no longer be busy.
 * requires the call->lock to be held */
void
rxi_WaitforTQBusy(struct rx_call *call) {
    while (!call->error && (call->flags & RX_CALL_TQ_BUSY)) {
        call->flags |= RX_CALL_TQ_WAIT;
        call->tqWaiters++;
        MUTEX_ASSERT(&call->lock);
        CV_WAIT(&call->cv_tq, &call->lock);
        call->tqWaiters--;
        if (call->tqWaiters == 0) {
            call->flags &= ~RX_CALL_TQ_WAIT;
        }
    }
}
#endif

static void
rxi_WakeUpTransmitQueue(struct rx_call *call)
{
    if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
        dpf(("call %p has %d waiters and flags %d\n",
             call, call->tqWaiters, call->flags));
#ifdef RX_ENABLE_LOCKS
        MUTEX_ASSERT(&call->lock);
        CV_BROADCAST(&call->cv_tq);
#else /* RX_ENABLE_LOCKS */
        osi_rxWakeup(&call->tq);
#endif /* RX_ENABLE_LOCKS */
    }
}

/* Start a new rx remote procedure call, on the specified connection.
 * If wait is set to 1, wait for a free call channel; otherwise return
 * 0.  Maxtime gives the maximum number of seconds this call may take,
 * after rx_NewCall returns.  After this time interval, a call to any
 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
 * For fine grain locking, we hold the conn_call_lock in order to
 * to ensure that we don't get signalle after we found a call in an active
 * state and before we go to sleep.
 */
struct rx_call *
rx_NewCall(struct rx_connection *conn)
{
    int i, wait, ignoreBusy = 1;
    struct rx_call *call;
    struct clock queueTime;
    afs_uint32 leastBusy = 0;
    SPLVAR;

    clock_NewTime();
    dpf(("rx_NewCall(conn %p)\n", conn));

    NETPRI;
    clock_GetTime(&queueTime);
    /*
     * Check if there are others waiting for a new call.
     * If so, let them go first to avoid starving them.
     * This is a fairly simple scheme, and might not be
     * a complete solution for large numbers of waiters.
     *
     * makeCallWaiters keeps track of the number of
     * threads waiting to make calls and the
     * RX_CONN_MAKECALL_WAITING flag bit is used to
     * indicate that there are indeed calls waiting.
     * The flag is set when the waiter is incremented.
     * It is only cleared when makeCallWaiters is 0.
     * This prevents us from accidently destroying the
     * connection while it is potentially about to be used.
     */
    MUTEX_ENTER(&conn->conn_call_lock);
    MUTEX_ENTER(&conn->conn_data_lock);
    while (conn->flags & RX_CONN_MAKECALL_ACTIVE) {
        conn->flags |= RX_CONN_MAKECALL_WAITING;
        conn->makeCallWaiters++;
        MUTEX_EXIT(&conn->conn_data_lock);

#ifdef  RX_ENABLE_LOCKS
        CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
#else
        osi_rxSleep(conn);
#endif
        MUTEX_ENTER(&conn->conn_data_lock);
        conn->makeCallWaiters--;
        if (conn->makeCallWaiters == 0)
            conn->flags &= ~RX_CONN_MAKECALL_WAITING;
    }

    /* We are now the active thread in rx_NewCall */
    conn->flags |= RX_CONN_MAKECALL_ACTIVE;
    MUTEX_EXIT(&conn->conn_data_lock);

    for (;;) {
        wait = 1;

        for (i = 0; i < RX_MAXCALLS; i++) {
            call = conn->call[i];
            if (call) {
                if (!ignoreBusy && conn->lastBusy[i] != leastBusy) {
                    /* we're not ignoring busy call slots; only look at the
                     * call slot that is the "least" busy */
                    continue;
                }

                if (call->state == RX_STATE_DALLY) {
                    MUTEX_ENTER(&call->lock);
                    if (call->state == RX_STATE_DALLY) {
                        if (ignoreBusy && conn->lastBusy[i]) {
                            /* if we're ignoring busy call slots, skip any ones that
                             * have lastBusy set */
                            if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
                                leastBusy = conn->lastBusy[i];
                            }
                            MUTEX_EXIT(&call->lock);
                            continue;
                        }

                        /*
                         * We are setting the state to RX_STATE_RESET to
                         * ensure that no one else will attempt to use this
                         * call once we drop the conn->conn_call_lock and
                         * call->lock.  We must drop the conn->conn_call_lock
                         * before calling rxi_ResetCall because the process
                         * of clearing the transmit queue can block for an
                         * extended period of time.  If we block while holding
                         * the conn->conn_call_lock, then all rx_EndCall
                         * processing will block as well.  This has a detrimental
                         * effect on overall system performance.
                         */
                        call->state = RX_STATE_RESET;
                        (*call->callNumber)++;
                        MUTEX_EXIT(&conn->conn_call_lock);
                        CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
                        rxi_ResetCall(call, 0);
                        if (MUTEX_TRYENTER(&conn->conn_call_lock))
                            break;

                        /*
                         * If we failed to be able to safely obtain the
                         * conn->conn_call_lock we will have to drop the
                         * call->lock to avoid a deadlock.  When the call->lock
                         * is released the state of the call can change.  If it
                         * is no longer RX_STATE_RESET then some other thread is
                         * using the call.
                         */
                        MUTEX_EXIT(&call->lock);
                        MUTEX_ENTER(&conn->conn_call_lock);
                        MUTEX_ENTER(&call->lock);

                        if (call->state == RX_STATE_RESET)
                            break;

                        /*
                         * If we get here it means that after dropping
                         * the conn->conn_call_lock and call->lock that
                         * the call is no longer ours.  If we can't find
                         * a free call in the remaining slots we should
                         * not go immediately to RX_CONN_MAKECALL_WAITING
                         * because by dropping the conn->conn_call_lock
                         * we have given up synchronization with rx_EndCall.
                         * Instead, cycle through one more time to see if
                         * we can find a call that can call our own.
                         */
                        CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
                        wait = 0;
                    }
                    MUTEX_EXIT(&call->lock);
                }
            } else {
                if (ignoreBusy && conn->lastBusy[i]) {
                    /* if we're ignoring busy call slots, skip any ones that
                     * have lastBusy set */
                    if (leastBusy == 0 || conn->lastBusy[i] < leastBusy) {
                        leastBusy = conn->lastBusy[i];
                    }
                    continue;
                }

                /* rxi_NewCall returns with mutex locked */
                call = rxi_NewCall(conn, i);
                CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
                break;
            }
        }
        if (i < RX_MAXCALLS) {
            conn->lastBusy[i] = 0;
            break;
        }
        if (!wait)
            continue;
        if (leastBusy && ignoreBusy) {
            /* we didn't find a useable call slot, but we did see at least one
             * 'busy' slot; look again and only use a slot with the 'least
             * busy time */
            ignoreBusy = 0;
            continue;
        }

        MUTEX_ENTER(&conn->conn_data_lock);
        conn->flags |= RX_CONN_MAKECALL_WAITING;
        conn->makeCallWaiters++;
        MUTEX_EXIT(&conn->conn_data_lock);

#ifdef  RX_ENABLE_LOCKS
        CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
#else
        osi_rxSleep(conn);
#endif
        MUTEX_ENTER(&conn->conn_data_lock);
        conn->makeCallWaiters--;
        if (conn->makeCallWaiters == 0)
            conn->flags &= ~RX_CONN_MAKECALL_WAITING;
        MUTEX_EXIT(&conn->conn_data_lock);
    }
    /* Client is initially in send mode */
    call->state = RX_STATE_ACTIVE;
    call->error = conn->error;
    if (call->error)
        call->app.mode = RX_MODE_ERROR;
    else
        call->app.mode = RX_MODE_SENDING;

#ifdef AFS_RXERRQ_ENV
    /* remember how many network errors the peer has when we started, so if
     * more errors are encountered after the call starts, we know the other endpoint won't be
     * responding to us */
    call->neterr_gen = rx_atomic_read(&conn->peer->neterrs);
#endif

    /* remember start time for call in case we have hard dead time limit */
    call->queueTime = queueTime;
    clock_GetTime(&call->startTime);
    call->app.bytesSent = 0;
    call->app.bytesRcvd = 0;

    /* Turn on busy protocol. */
    rxi_KeepAliveOn(call);

    /* Attempt MTU discovery */
    rxi_GrowMTUOn(call);

    /*
     * We are no longer the active thread in rx_NewCall
     */
    MUTEX_ENTER(&conn->conn_data_lock);
    conn->flags &= ~RX_CONN_MAKECALL_ACTIVE;
    MUTEX_EXIT(&conn->conn_data_lock);

    /*
     * Wake up anyone else who might be giving us a chance to
     * run (see code above that avoids resource starvation).
     */
#ifdef  RX_ENABLE_LOCKS
    if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
        osi_Panic("rx_NewCall call about to be used without an empty tq");
    }

    CV_BROADCAST(&conn->conn_call_cv);
#else
    osi_rxWakeup(conn);
#endif
    MUTEX_EXIT(&conn->conn_call_lock);
    MUTEX_EXIT(&call->lock);
    USERPRI;

    dpf(("rx_NewCall(call %p)\n", call));
    return call;
}

static int
rxi_HasActiveCalls(struct rx_connection *aconn)
{
    int i;
    struct rx_call *tcall;
    SPLVAR;

    NETPRI;
    for (i = 0; i < RX_MAXCALLS; i++) {
        if ((tcall = aconn->call[i])) {
            if ((tcall->state == RX_STATE_ACTIVE)
                || (tcall->state == RX_STATE_PRECALL)) {
                USERPRI;
                return 1;
            }
        }
    }
    USERPRI;
    return 0;
}

int
rxi_GetCallNumberVector(struct rx_connection *aconn,
                        afs_int32 * aint32s)
{
    int i;
    struct rx_call *tcall;
    SPLVAR;

    NETPRI;
    MUTEX_ENTER(&aconn->conn_call_lock);
    for (i = 0; i < RX_MAXCALLS; i++) {
        if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
            aint32s[i] = aconn->callNumber[i] + 1;
        else
            aint32s[i] = aconn->callNumber[i];
    }
    MUTEX_EXIT(&aconn->conn_call_lock);
    USERPRI;
    return 0;
}

int
rxi_SetCallNumberVector(struct rx_connection *aconn,
                        afs_int32 * aint32s)
{
    int i;
    struct rx_call *tcall;
    SPLVAR;

    NETPRI;
    MUTEX_ENTER(&aconn->conn_call_lock);
    for (i = 0; i < RX_MAXCALLS; i++) {
        if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
            aconn->callNumber[i] = aint32s[i] - 1;
        else
            aconn->callNumber[i] = aint32s[i];
    }
    MUTEX_EXIT(&aconn->conn_call_lock);
    USERPRI;
    return 0;
}

/* Advertise a new service.  A service is named locally by a UDP port
 * number plus a 16-bit service id.  Returns (struct rx_service *) 0
 * on a failure.
 *
     char *serviceName;  Name for identification purposes (e.g. the
                         service name might be used for probing for
                         statistics) */
struct rx_service *
rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
                  char *serviceName, struct rx_securityClass **securityObjects,
                  int nSecurityObjects,
                  afs_int32(*serviceProc) (struct rx_call * acall))
{
    osi_socket socket = OSI_NULLSOCKET;
    struct rx_service *tservice;
    int i;
    SPLVAR;

    clock_NewTime();

    if (serviceId == 0) {
        (osi_Msg
         "rx_NewService:  service id for service %s is not non-zero.\n",
         serviceName);
        return 0;
    }
    if (port == 0) {
        if (rx_port == 0) {
            (osi_Msg
             "rx_NewService: A non-zero port must be specified on this call if a non-zero port was not provided at Rx initialization (service %s).\n",
             serviceName);
            return 0;
        }
        port = rx_port;
        socket = rx_socket;
    }

    tservice = rxi_AllocService();
    NETPRI;

    MUTEX_INIT(&tservice->svc_data_lock, "svc data lock", MUTEX_DEFAULT, 0);

    for (i = 0; i < RX_MAX_SERVICES; i++) {
        struct rx_service *service = rx_services[i];
        if (service) {
            if (port == service->servicePort && host == service->serviceHost) {
                if (service->serviceId == serviceId) {
                    /* The identical service has already been
                     * installed; if the caller was intending to
                     * change the security classes used by this
                     * service, he/she loses. */
                    (osi_Msg
                     "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
                     serviceName, serviceId, service->serviceName);
                    USERPRI;
                    rxi_FreeService(tservice);
                    return service;
                }
                /* Different service, same port: re-use the socket
                 * which is bound to the same port */
                socket = service->socket;
            }
        } else {
            if (socket == OSI_NULLSOCKET) {
                /* If we don't already have a socket (from another
                 * service on same port) get a new one */
                socket = rxi_GetHostUDPSocket(host, port);
                if (socket == OSI_NULLSOCKET) {
                    USERPRI;
                    rxi_FreeService(tservice);
                    return 0;
                }
            }
            service = tservice;
            service->socket = socket;
            service->serviceHost = host;
            service->servicePort = port;
            service->serviceId = serviceId;
            service->serviceName = serviceName;
            service->nSecurityObjects = nSecurityObjects;
            service->securityObjects = securityObjects;
            service->minProcs = 0;
            service->maxProcs = 1;
            service->idleDeadTime = 60;
            service->connDeadTime = rx_connDeadTime;
            service->executeRequestProc = serviceProc;
            service->checkReach = 0;
            service->nSpecific = 0;
            service->specific = NULL;
            rx_services[i] = service;   /* not visible until now */
            USERPRI;
            return service;
        }
    }
    USERPRI;
    rxi_FreeService(tservice);
    (osi_Msg "rx_NewService: cannot support > %d services\n",
     RX_MAX_SERVICES);
    return 0;
}

/* Set configuration options for all of a service's security objects */

afs_int32
rx_SetSecurityConfiguration(struct rx_service *service,
                            rx_securityConfigVariables type,
                            void *value)
{
    int i;
    int code;
    for (i = 0; i<service->nSecurityObjects; i++) {
        if (service->securityObjects[i]) {
            code = RXS_SetConfiguration(service->securityObjects[i], NULL, type,
                                        value, NULL);
            if (code) {
                return code;
            }
        }
    }
    return 0;
}

struct rx_service *
rx_NewService(u_short port, u_short serviceId, char *serviceName,
              struct rx_securityClass **securityObjects, int nSecurityObjects,
              afs_int32(*serviceProc) (struct rx_call * acall))
{
    return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
}

/* Generic request processing loop. This routine should be called
 * by the implementation dependent rx_ServerProc. If socketp is
 * non-null, it will be set to the file descriptor that this thread
 * is now listening on. If socketp is null, this routine will never
 * returns. */
void
rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
{
    struct rx_call *call;
    afs_int32 code;
    struct rx_service *tservice = NULL;

    for (;;) {
        if (newcall) {
            call = newcall;
            newcall = NULL;
        } else {
            call = rx_GetCall(threadID, tservice, socketp);
            if (socketp && *socketp != OSI_NULLSOCKET) {
                /* We are now a listener thread */
                return;
            }
        }

#ifdef  KERNEL
        if (afs_termState == AFSOP_STOP_RXCALLBACK) {
#ifdef RX_ENABLE_LOCKS
            AFS_GLOCK();
#endif /* RX_ENABLE_LOCKS */
            afs_termState = AFSOP_STOP_AFS;
            afs_osi_Wakeup(&afs_termState);
#ifdef RX_ENABLE_LOCKS
            AFS_GUNLOCK();
#endif /* RX_ENABLE_LOCKS */
            return;
        }
#endif

        /* if server is restarting( typically smooth shutdown) then do not
         * allow any new calls.
         */

        if (rx_tranquil && (call != NULL)) {
            SPLVAR;

            NETPRI;
            MUTEX_ENTER(&call->lock);

            rxi_CallError(call, RX_RESTARTING);
            rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);

            MUTEX_EXIT(&call->lock);
            USERPRI;
            continue;
        }

        tservice = call->conn->service;

        if (tservice->beforeProc)
            (*tservice->beforeProc) (call);

        code = tservice->executeRequestProc(call);

        if (tservice->afterProc)
            (*tservice->afterProc) (call, code);

        rx_EndCall(call, code);

        if (tservice->postProc)
            (*tservice->postProc) (code);

        if (rx_stats_active) {
            MUTEX_ENTER(&rx_stats_mutex);
            rxi_nCalls++;
            MUTEX_EXIT(&rx_stats_mutex);
        }
    }
}


void
rx_WakeupServerProcs(void)
{
    struct rx_serverQueueEntry *np;
    struct opr_queue *cursor;
    SPLVAR;

    NETPRI;
    MUTEX_ENTER(&rx_serverPool_lock);

#ifdef RX_ENABLE_LOCKS
    if (rx_waitForPacket)
        CV_BROADCAST(&rx_waitForPacket->cv);
#else /* RX_ENABLE_LOCKS */
    if (rx_waitForPacket)
        osi_rxWakeup(rx_waitForPacket);
#endif /* RX_ENABLE_LOCKS */
    MUTEX_ENTER(&freeSQEList_lock);
    for (opr_queue_Scan(&rx_freeServerQueue, cursor)) {
        np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
#ifdef RX_ENABLE_LOCKS
        CV_BROADCAST(&np->cv);
#else /* RX_ENABLE_LOCKS */
        osi_rxWakeup(np);
#endif /* RX_ENABLE_LOCKS */
    }
    MUTEX_EXIT(&freeSQEList_lock);
    for (opr_queue_Scan(&rx_idleServerQueue, cursor)) {
        np = opr_queue_Entry(cursor, struct rx_serverQueueEntry, entry);
#ifdef RX_ENABLE_LOCKS
        CV_BROADCAST(&np->cv);
#else /* RX_ENABLE_LOCKS */
        osi_rxWakeup(np);
#endif /* RX_ENABLE_LOCKS */
    }
    MUTEX_EXIT(&rx_serverPool_lock);
    USERPRI;
}

/* meltdown:
 * One thing that seems to happen is that all the server threads get
 * tied up on some empty or slow call, and then a whole bunch of calls
 * arrive at once, using up the packet pool, so now there are more
 * empty calls.  The most critical resources here are server threads
 * and the free packet pool.  The "doreclaim" code seems to help in
 * general.  I think that eventually we arrive in this state: there
 * are lots of pending calls which do have all their packets present,
 * so they won't be reclaimed, are multi-packet calls, so they won't
 * be scheduled until later, and thus are tying up most of the free
 * packet pool for a very long time.
 * future options:
 * 1.  schedule multi-packet calls if all the packets are present.
 * Probably CPU-bound operation, useful to return packets to pool.
 * Do what if there is a full window, but the last packet isn't here?
 * 3.  preserve one thread which *only* runs "best" calls, otherwise
 * it sleeps and waits for that type of call.
 * 4.  Don't necessarily reserve a whole window for each thread.  In fact,
 * the current dataquota business is badly broken.  The quota isn't adjusted
 * to reflect how many packets are presently queued for a running call.
 * So, when we schedule a queued call with a full window of packets queued
 * up for it, that *should* free up a window full of packets for other 2d-class
 * calls to be able to use from the packet pool.  But it doesn't.
 *
 * NB.  Most of the time, this code doesn't run -- since idle server threads
 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
 * as a new call arrives.
 */
/* Sleep until a call arrives.  Returns a pointer to the call, ready
 * for an rx_Read. */
#ifdef RX_ENABLE_LOCKS
struct rx_call *
rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
{
    struct rx_serverQueueEntry *sq;
    struct rx_call *call = (struct rx_call *)0;
    struct rx_service *service = NULL;

    MUTEX_ENTER(&freeSQEList_lock);

    if (!opr_queue_IsEmpty(&rx_freeServerQueue)) {
        sq = opr_queue_First(&rx_freeServerQueue, struct rx_serverQueueEntry,
                             entry);
        opr_queue_Remove(&sq->entry);
        MUTEX_EXIT(&freeSQEList_lock);
    } else {                    /* otherwise allocate a new one and return that */
        MUTEX_EXIT(&freeSQEList_lock);
        sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
        MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
        CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
    }

    MUTEX_ENTER(&rx_serverPool_lock);
    if (cur_service != NULL) {
        ReturnToServerPool(cur_service);
    }
    while (1) {
        if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
            struct rx_call *tcall, *choice2 = NULL;
            struct opr_queue *cursor;

            /* Scan for eligible incoming calls.  A call is not eligible
             * if the maximum number of calls for its service type are
             * already executing */
            /* One thread will process calls FCFS (to prevent starvation),
             * while the other threads may run ahead looking for calls which
             * have all their input data available immediately.  This helps
             * keep threads from blocking, waiting for data from the client. */
            for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
                tcall = opr_queue_Entry(cursor, struct rx_call, entry);

                service = tcall->conn->service;
                if (!QuotaOK(service)) {
                    continue;
                }
                MUTEX_ENTER(&rx_pthread_mutex);
                if (tno == rxi_fcfs_thread_num
                        || opr_queue_IsLast(&rx_incomingCallQueue, cursor)) {
                    MUTEX_EXIT(&rx_pthread_mutex);
                    /* If we're the fcfs thread , then  we'll just use
                     * this call. If we haven't been able to find an optimal
                     * choice, and we're at the end of the list, then use a
                     * 2d choice if one has been identified.  Otherwise... */
                    call = (choice2 ? choice2 : tcall);
                    service = call->conn->service;
                } else {
                    MUTEX_EXIT(&rx_pthread_mutex);
                    if (!opr_queue_IsEmpty(&tcall->rq)) {
                        struct rx_packet *rp;
                        rp = opr_queue_First(&tcall->rq, struct rx_packet,
                                            entry);
                        if (rp->header.seq == 1) {
                            if (!meltdown_1pkt
                                || (rp->header.flags & RX_LAST_PACKET)) {
                                call = tcall;
                            } else if (rxi_2dchoice && !choice2
                                       && !(tcall->flags & RX_CALL_CLEARED)
                                       && (tcall->rprev > rxi_HardAckRate)) {
                                choice2 = tcall;
                            } else
                                rxi_md2cnt++;
                        }
                    }
                }
                if (call) {
                    break;
                } else {
                    ReturnToServerPool(service);
                }
            }
        }

        if (call) {
            opr_queue_Remove(&call->entry);
            MUTEX_EXIT(&rx_serverPool_lock);
            MUTEX_ENTER(&call->lock);
            CLEAR_CALL_QUEUE_LOCK(call);

            if (call->flags & RX_CALL_WAIT_PROC) {
                call->flags &= ~RX_CALL_WAIT_PROC;
                rx_atomic_dec(&rx_nWaiting);
            }

            if (call->state != RX_STATE_PRECALL || call->error) {
                MUTEX_EXIT(&call->lock);
                MUTEX_ENTER(&rx_serverPool_lock);
                ReturnToServerPool(service);
                call = NULL;
                continue;
            }

            if (opr_queue_IsEmpty(&call->rq)
                || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1)
                rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);

            break;
        } else {
            /* If there are no eligible incoming calls, add this process
             * to the idle server queue, to wait for one */
            sq->newcall = 0;
            sq->tno = tno;
            if (socketp) {
                *socketp = OSI_NULLSOCKET;
            }
            sq->socketp = socketp;
            opr_queue_Append(&rx_idleServerQueue, &sq->entry);
#ifndef AFS_AIX41_ENV
            rx_waitForPacket = sq;
#endif /* AFS_AIX41_ENV */
            do {
                CV_WAIT(&sq->cv, &rx_serverPool_lock);
#ifdef  KERNEL
                if (afs_termState == AFSOP_STOP_RXCALLBACK) {
                    break;
                }
#endif
            } while (!(call = sq->newcall)
                     && !(socketp && *socketp != OSI_NULLSOCKET));
            if (opr_queue_IsOnQueue(&sq->entry)) {
                opr_queue_Remove(&sq->entry);
            }
            MUTEX_EXIT(&rx_serverPool_lock);
            if (call) {
                MUTEX_ENTER(&call->lock);
            }
            break;
        }
    }

    MUTEX_ENTER(&freeSQEList_lock);
    opr_queue_Prepend(&rx_freeServerQueue, &sq->entry);
    MUTEX_EXIT(&freeSQEList_lock);

    if (call) {
        clock_GetTime(&call->startTime);
        call->state = RX_STATE_ACTIVE;
        call->app.mode = RX_MODE_RECEIVING;
#ifdef RX_KERNEL_TRACE
        if (ICL_SETACTIVE(afs_iclSetp)) {
            int glockOwner = ISAFS_GLOCK();
            if (!glockOwner)
                AFS_GLOCK();
            afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
                       __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
                       call);
            if (!glockOwner)
                AFS_GUNLOCK();
        }
#endif

        rxi_calltrace(RX_CALL_START, call);
        dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
             call->conn->service->servicePort, call->conn->service->serviceId,
             call));

        MUTEX_EXIT(&call->lock);
        CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
    } else {
        dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
    }

    return call;
}
#else /* RX_ENABLE_LOCKS */
struct rx_call *
rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
{
    struct rx_serverQueueEntry *sq;
    struct rx_call *call = (struct rx_call *)0, *choice2;
    struct rx_service *service = NULL;
    SPLVAR;

    NETPRI;
    MUTEX_ENTER(&freeSQEList_lock);

    if (!opr_queue_IsEmpty(&rx_freeServerQueue)) {
        sq = opr_queue_First(&rx_freeServerQueue, struct rx_serverQueueEntry,
                             entry);
        opr_queue_Remove(&sq->entry);
        MUTEX_EXIT(&freeSQEList_lock);
    } else {                    /* otherwise allocate a new one and return that */
        MUTEX_EXIT(&freeSQEList_lock);
        sq = rxi_Alloc(sizeof(struct rx_serverQueueEntry));
        MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
        CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
    }
    MUTEX_ENTER(&sq->lock);

    if (cur_service != NULL) {
        cur_service->nRequestsRunning--;
        MUTEX_ENTER(&rx_quota_mutex);
        if (cur_service->nRequestsRunning < cur_service->minProcs)
            rxi_minDeficit++;
        rxi_availProcs++;
        MUTEX_EXIT(&rx_quota_mutex);
    }
    if (!opr_queue_IsEmpty(&rx_incomingCallQueue)) {
        struct rx_call *tcall;
        struct opr_queue *cursor;
        /* Scan for eligible incoming calls.  A call is not eligible
         * if the maximum number of calls for its service type are
         * already executing */
        /* One thread will process calls FCFS (to prevent starvation),
         * while the other threads may run ahead looking for calls which
         * have all their input data available immediately.  This helps
         * keep threads from blocking, waiting for data from the client. */
        choice2 = (struct rx_call *)0;
        for (opr_queue_Scan(&rx_incomingCallQueue, cursor)) {
            tcall = opr_queue_Entry(cursor, struct rx_call, entry);
            service = tcall->conn->service;
            if (QuotaOK(service)) {
                MUTEX_ENTER(&rx_pthread_mutex);
                if (tno == rxi_fcfs_thread_num
                        || opr_queue_IsLast(&rx_incomingCallQueue, cursor)) {
                    MUTEX_EXIT(&rx_pthread_mutex);
                    /* If we're the fcfs thread, then  we'll just use
                     * this call. If we haven't been able to find an optimal
                     * choice, and we're at the end of the list, then use a
                     * 2d choice if one has been identified.  Otherwise... */
                    call = (choice2 ? choice2 : tcall);
                    service = call->conn->service;
                } else {
                    MUTEX_EXIT(&rx_pthread_mutex);
                    if (!opr_queue_IsEmpty(&tcall->rq)) {
                        struct rx_packet *rp;
                        rp = opr_queue_First(&tcall->rq, struct rx_packet,
                                            entry);
                        if (rp->header.seq == 1
                            && (!meltdown_1pkt
                                || (rp->header.flags & RX_LAST_PACKET))) {
                            call = tcall;
                        } else if (rxi_2dchoice && !choice2
                                   && !(tcall->flags & RX_CALL_CLEARED)
                                   && (tcall->rprev > rxi_HardAckRate)) {
                            choice2 = tcall;
                        } else
                            rxi_md2cnt++;
                    }
                }
            }
            if (call)
                break;
        }
    }

    if (call) {
        opr_queue_Remove(&call->entry);
        CLEAR_CALL_QUEUE_LOCK(call);
        /* we can't schedule a call if there's no data!!! */
        /* send an ack if there's no data, if we're missing the
         * first packet, or we're missing something between first
         * and last -- there's a "hole" in the incoming data. */
        if (opr_queue_IsEmpty(&call->rq)
            || opr_queue_First(&call->rq, struct rx_packet, entry)->header.seq != 1
            || call->rprev != opr_queue_Last(&call->rq, struct rx_packet, entry)->header.seq)
            rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);

        call->flags &= (~RX_CALL_WAIT_PROC);
        service->nRequestsRunning++;
        /* just started call in minProcs pool, need fewer to maintain
         * guarantee */
        MUTEX_ENTER(&rx_quota_mutex);
        if (service->nRequestsRunning <= service->minProcs)
            rxi_minDeficit--;
        rxi_availProcs--;
        MUTEX_EXIT(&rx_quota_mutex);
        rx_atomic_dec(&rx_nWaiting);
        /* MUTEX_EXIT(&call->lock); */
    } else {
        /* If there are no eligible incoming calls, add this process
         * to the idle server queue, to wait for one */
        sq->newcall = 0;
        if (socketp) {
            *socketp = OSI_NULLSOCKET;
        }
        sq->socketp = socketp;
        opr_queue_Append(&rx_idleServerQueue, &sq->entry);
        do {
            osi_rxSleep(sq);
#ifdef  KERNEL
            if (afs_termState == AFSOP_STOP_RXCALLBACK) {
                USERPRI;
                rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
                return (struct rx_call *)0;
            }
#endif
        } while (!(call = sq->newcall)
                 && !(socketp && *socketp != OSI_NULLSOCKET));
    }
    MUTEX_EXIT(&sq->lock);

    MUTEX_ENTER(&freeSQEList_lock);
    opr_queue_Prepend(&rx_freeServerQueue, &sq->entry);
    MUTEX_EXIT(&freeSQEList_lock);

    if (call) {
        clock_GetTime(&call->startTime);
        call->state = RX_STATE_ACTIVE;
        call->app.mode = RX_MODE_RECEIVING;
#ifdef RX_KERNEL_TRACE
        if (ICL_SETACTIVE(afs_iclSetp)) {
            int glockOwner = ISAFS_GLOCK();
            if (!glockOwner)
                AFS_GLOCK();
            afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
                       __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
                       call);
            if (!glockOwner)
                AFS_GUNLOCK();
        }
#endif

        rxi_calltrace(RX_CALL_START, call);
        dpf(("rx_GetCall(port=%d, service=%d) ==> call %p\n",
             call->conn->service->servicePort, call->conn->service->serviceId,
             call));
    } else {
        dpf(("rx_GetCall(socketp=%p, *socketp=0x%x)\n", socketp, *socketp));
    }

    USERPRI;

    return call;
}
#endif /* RX_ENABLE_LOCKS */



/* Establish a procedure to be called when a packet arrives for a
 * call.  This routine will be called at most once after each call,
 * and will also be called if there is an error condition on the or
 * the call is complete.  Used by multi rx to build a selection
 * function which determines which of several calls is likely to be a
 * good one to read from.
 * NOTE: the way this is currently implemented it is probably only a
 * good idea to (1) use it immediately after a newcall (clients only)
 * and (2) only use it once.  Other uses currently void your warranty
 */
void
rx_SetArrivalProc(struct rx_call *call,
                  void (*proc) (struct rx_call * call,
                                        void * mh,
                                        int index),
                  void * handle, int arg)
{
    call->arrivalProc = proc;
    call->arrivalProcHandle = handle;
    call->arrivalProcArg = arg;
}

/* Call is finished (possibly prematurely).  Return rc to the peer, if
 * appropriate, and return the final error code from the conversation
 * to the caller */

afs_int32
rx_EndCall(struct rx_call *call, afs_int32 rc)
{
    struct rx_connection *conn = call->conn;
    afs_int32 error;
    SPLVAR;

    dpf(("rx_EndCall(call %p rc %d error %d abortCode %d)\n",
          call, rc, call->error, call->abortCode));

    NETPRI;
    MUTEX_ENTER(&call->lock);

    if (rc == 0 && call->error == 0) {
        call->abortCode = 0;
        call->abortCount = 0;
    }

    call->arrivalProc = NULL;
    if (rc && call->error == 0) {
        rxi_CallError(call, rc);
        call->app.mode = RX_MODE_ERROR;
        /* Send an abort message to the peer if this error code has
         * only just been set.  If it was set previously, assume the
         * peer has already been sent the error code or will request it
         */
        rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
    }
    if (conn->type == RX_SERVER_CONNECTION) {
        /* Make sure reply or at least dummy reply is sent */
        if (call->app.mode == RX_MODE_RECEIVING) {
            MUTEX_EXIT(&call->lock);
            rxi_WriteProc(call, 0, 0);
            MUTEX_ENTER(&call->lock);
        }
        if (call->app.mode == RX_MODE_SENDING) {
            rxi_FlushWriteLocked(call);
        }
        rxi_calltrace(RX_CALL_END, call);
        /* Call goes to hold state until reply packets are acknowledged */
        if (call->tfirst + call->nSoftAcked < call->tnext) {
            call->state = RX_STATE_HOLD;
        } else {
            call->state = RX_STATE_DALLY;
            rxi_ClearTransmitQueue(call, 0);
            rxi_rto_cancel(call);
            rxi_CancelKeepAliveEvent(call);
        }
    } else {                    /* Client connection */
        char dummy;
        /* Make sure server receives input packets, in the case where
         * no reply arguments are expected */

        if ((call->app.mode == RX_MODE_SENDING)
            || (call->app.mode == RX_MODE_RECEIVING && call->rnext == 1)) {
            MUTEX_EXIT(&call->lock);
            (void)rxi_ReadProc(call, &dummy, 1);
            MUTEX_ENTER(&call->lock);
        }

        /* If we had an outstanding delayed ack, be nice to the server
         * and force-send it now.
         */
        if (call->delayedAckEvent) {
            rxi_CancelDelayedAckEvent(call);
            rxi_SendDelayedAck(NULL, call, NULL, 0);
        }

        /* We need to release the call lock since it's lower than the
         * conn_call_lock and we don't want to hold the conn_call_lock
         * over the rx_ReadProc call. The conn_call_lock needs to be held
         * here for the case where rx_NewCall is perusing the calls on
         * the connection structure. We don't want to signal until
         * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
         * have checked this call, found it active and by the time it
         * goes to sleep, will have missed the signal.
         */
        MUTEX_EXIT(&call->lock);
        MUTEX_ENTER(&conn->conn_call_lock);
        MUTEX_ENTER(&call->lock);

        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);
        conn->flags |= RX_CONN_BUSY;
        if (conn->flags & RX_CONN_MAKECALL_WAITING) {
            MUTEX_EXIT(&conn->conn_data_lock);
#ifdef  RX_ENABLE_LOCKS
            CV_BROADCAST(&conn->conn_call_cv);
#else
            osi_rxWakeup(conn);
#endif
        }
#ifdef RX_ENABLE_LOCKS
        else {
            MUTEX_EXIT(&conn->conn_data_lock);
        }
#endif /* RX_ENABLE_LOCKS */
        call->state = RX_STATE_DALLY;
    }
    error = call->error;

    /* currentPacket, nLeft, and NFree must be zeroed here, because
     * ResetCall cannot: ResetCall may be called at splnet(), in the
     * kernel version, and may interrupt the macros rx_Read or
     * rx_Write, which run at normal priority for efficiency. */
    if (call->app.currentPacket) {
#ifdef RX_TRACK_PACKETS
        call->app.currentPacket->flags &= ~RX_PKTFLAG_CP;
#endif
        rxi_FreePacket(call->app.currentPacket);
        call->app.currentPacket = (struct rx_packet *)0;
    }

    call->app.nLeft = call->app.nFree = call->app.curlen = 0;

    /* Free any packets from the last call to ReadvProc/WritevProc */
#ifdef RXDEBUG_PACKET
    call->iovqc -=
#endif /* RXDEBUG_PACKET */
        rxi_FreePackets(0, &call->app.iovq);
    MUTEX_EXIT(&call->lock);

    CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
    if (conn->type == RX_CLIENT_CONNECTION) {
        MUTEX_ENTER(&conn->conn_data_lock);
        conn->flags &= ~RX_CONN_BUSY;
        MUTEX_EXIT(&conn->conn_data_lock);
        MUTEX_EXIT(&conn->conn_call_lock);
    }
    USERPRI;
    /*
     * 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;
}

#if !defined(KERNEL)

/* Call this routine when shutting down a server or client (especially
 * clients).  This will allow Rx to gracefully garbage collect server
 * connections, and reduce the number of retries that a server might
 * make to a dead client.
 * This is not quite right, since some calls may still be ongoing and
 * we can't lock them to destroy them. */
void
rx_Finalize(void)
{
    INIT_PTHREAD_LOCKS;
    LOCK_RX_INIT;
    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_ptr = &rx_connHashTable[0], conn_end =
             &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
             conn_ptr++) {
            struct rx_connection *conn, *next;
            for (conn = *conn_ptr; conn; conn = next) {
                next = conn->next;
                if (conn->type == RX_CLIENT_CONNECTION) {
                    rx_GetConnection(conn);
#ifdef RX_ENABLE_LOCKS
                    rxi_DestroyConnectionNoLock(conn);
#else /* RX_ENABLE_LOCKS */
                    rxi_DestroyConnection(conn);
#endif /* RX_ENABLE_LOCKS */
                }
            }
        }
#ifdef RX_ENABLE_LOCKS
        while (rx_connCleanup_list) {
            struct rx_connection *conn;
            conn = rx_connCleanup_list;
            rx_connCleanup_list = rx_connCleanup_list->next;
            MUTEX_EXIT(&rx_connHashTable_lock);
            rxi_CleanupConnection(conn);
            MUTEX_ENTER(&rx_connHashTable_lock);
        }
        MUTEX_EXIT(&rx_connHashTable_lock);
#endif /* RX_ENABLE_LOCKS */
    }
    rxi_flushtrace();

#ifdef AFS_NT40_ENV
    afs_winsockCleanup();
#endif
}
#endif

/* if we wakeup packet waiter too often, can get in loop with two
    AllocSendPackets each waking each other up (from ReclaimPacket calls) */
void
rxi_PacketsUnWait(void)
{
    if (!rx_waitingForPackets) {
        return;
    }
#ifdef KERNEL
    if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
        return;                 /* still over quota */
    }
#endif /* KERNEL */
    rx_waitingForPackets = 0;
#ifdef  RX_ENABLE_LOCKS
    CV_BROADCAST(&rx_waitingForPackets_cv);
#else
    osi_rxWakeup(&rx_waitingForPackets);
#endif
    return;
}


/* ------------------Internal interfaces------------------------- */

/* Return this process's service structure for the
 * specified socket and service */
static struct rx_service *
rxi_FindService(osi_socket socket, u_short serviceId)
{
    struct rx_service **sp;
    for (sp = &rx_services[0]; *sp; sp++) {
        if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
            return *sp;
    }
    return 0;
}

#ifdef RXDEBUG_PACKET
#ifdef KDUMP_RX_LOCK
static struct rx_call_rx_lock *rx_allCallsp = 0;
#else
static struct rx_call *rx_allCallsp = 0;
#endif
#endif /* RXDEBUG_PACKET */

/* Allocate a call structure, for the indicated channel of the
 * supplied connection.  The mode and state of the call must be set by
 * the caller. Returns the call with mutex locked. */
static struct rx_call *
rxi_NewCall(struct rx_connection *conn, int channel)
{
    struct rx_call *call;
#ifdef RX_ENABLE_LOCKS
    struct rx_call *cp; /* Call pointer temp */
    struct opr_queue *cursor;
#endif

    dpf(("rxi_NewCall(conn %p, channel %d)\n", conn, channel));

    /* Grab an existing call structure, or allocate a new one.
     * Existing call structures are assumed to have been left reset by
     * rxi_FreeCall */
    MUTEX_ENTER(&rx_freeCallQueue_lock);

#ifdef RX_ENABLE_LOCKS
    /*
     * EXCEPT that the TQ might not yet be cleared out.
     * Skip over those with in-use TQs.
     */
    call = NULL;
    for (opr_queue_Scan(&rx_freeCallQueue, cursor)) {
        cp = opr_queue_Entry(cursor, struct rx_call, entry);
        if (!(cp->flags & RX_CALL_TQ_BUSY)) {
            call = cp;
            break;
        }
    }
    if (call) {
#else /* RX_ENABLE_LOCKS */
    if (!opr_queue_IsEmpty(&rx_freeCallQueue)) {
        call = opr_queue_First(&rx_freeCallQueue, struct rx_call, entry);
#endif /* RX_ENABLE_LOCKS */
        opr_queue_Remove(&call->entry);
        if (rx_stats_active)
            rx_atomic_dec(&rx_stats.nFreeCallStructs);
        MUTEX_EXIT(&rx_freeCallQueue_lock);
        MUTEX_ENTER(&call->lock);
        CLEAR_CALL_QUEUE_LOCK(call);
#ifdef RX_ENABLE_LOCKS
        /* Now, if TQ wasn't cleared earlier, do it now. */
        rxi_WaitforTQBusy(call);
        if (call->flags & RX_CALL_TQ_CLEARME) {
            rxi_ClearTransmitQueue(call, 1);
            /*queue_Init(&call->tq);*/
        }
#endif /* RX_ENABLE_LOCKS */
        /* Bind the call to its connection structure */
        call->conn = conn;
        rxi_ResetCall(call, 1);
    } else {

        call = rxi_Alloc(sizeof(struct rx_call));
#ifdef RXDEBUG_PACKET
        call->allNextp = rx_allCallsp;
        rx_allCallsp = call;
        call->call_id =
            rx_atomic_inc_and_read(&rx_stats.nCallStructs);
#else /* RXDEBUG_PACKET */
        rx_atomic_inc(&rx_stats.nCallStructs);
#endif /* RXDEBUG_PACKET */

        MUTEX_EXIT(&rx_freeCallQueue_lock);
        MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
        MUTEX_ENTER(&call->lock);
        CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
        CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
        CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);

        /* Initialize once-only items */
        opr_queue_Init(&call->tq);
        opr_queue_Init(&call->rq);
        opr_queue_Init(&call->app.iovq);
#ifdef RXDEBUG_PACKET
        call->rqc = call->tqc = call->iovqc = 0;
#endif /* RXDEBUG_PACKET */
        /* Bind the call to its connection structure (prereq for reset) */
        call->conn = conn;
        rxi_ResetCall(call, 1);
    }
    call->channel = channel;
    call->callNumber = &conn->callNumber[channel];
    call->rwind = conn->rwind[channel];
    call->twind = conn->twind[channel];
    /* Note that the next expected call number is retained (in
     * conn->callNumber[i]), even if we reallocate the call structure
     */
    conn->call[channel] = call;
    /* if the channel's never been used (== 0), we should start at 1, otherwise
     * the call number is valid from the last time this channel was used */
    if (*call->callNumber == 0)
        *call->callNumber = 1;

    return call;
}

/* A call has been inactive long enough that so we can throw away
 * state, including the call structure, which is placed on the call
 * free list.
 *
 * call->lock amd rx_refcnt_mutex are held upon entry.
 * haveCTLock is set when called from rxi_ReapConnections.
 *
 * return 1 if the call is freed, 0 if not.
 */
static int
rxi_FreeCall(struct rx_call *call, int haveCTLock)
{
    int channel = call->channel;
    struct rx_connection *conn = call->conn;
    u_char state = call->state;

    /*
     * We are setting the state to RX_STATE_RESET to
     * ensure that no one else will attempt to use this
     * call once we drop the refcnt lock. We must drop
     * the refcnt lock before calling rxi_ResetCall
     * because it cannot be held across acquiring the
     * freepktQ lock. NewCall does the same.
     */
    call->state = RX_STATE_RESET;
    MUTEX_EXIT(&rx_refcnt_mutex);
    rxi_ResetCall(call, 0);

    if (MUTEX_TRYENTER(&conn->conn_call_lock))
    {
        if (state == RX_STATE_DALLY || state == RX_STATE_HOLD)
            (*call->callNumber)++;

        if (call->conn->call[channel] == call)
            call->conn->call[channel] = 0;
        MUTEX_EXIT(&conn->conn_call_lock);
    } else {
        /*
         * We couldn't obtain the conn_call_lock so we can't
         * disconnect the call from the connection.  Set the
         * call state to dally so that the call can be reused.
         */
        MUTEX_ENTER(&rx_refcnt_mutex);
        call->state = RX_STATE_DALLY;
        return 0;
    }

    MUTEX_ENTER(&rx_freeCallQueue_lock);
    SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
#ifdef RX_ENABLE_LOCKS
    /* A call may be free even though its transmit queue is still in use.
     * Since we search the call list from head to tail, put busy calls at
     * the head of the list, and idle calls at the tail.
     */
    if (call->flags & RX_CALL_TQ_BUSY)
        opr_queue_Prepend(&rx_freeCallQueue, &call->entry);
    else
        opr_queue_Append(&rx_freeCallQueue, &call->entry);
#else /* RX_ENABLE_LOCKS */
    opr_queue_Append(&rx_freeCallQueue, &call->entry);
#endif /* RX_ENABLE_LOCKS */
    if (rx_stats_active)
        rx_atomic_inc(&rx_stats.nFreeCallStructs);
    MUTEX_EXIT(&rx_freeCallQueue_lock);

    /* Destroy the connection if it was previously slated for
     * destruction, i.e. the Rx client code previously called
     * rx_DestroyConnection (client connections), or
     * rxi_ReapConnections called the same routine (server
     * connections).  Only do this, however, if there are no
     * outstanding calls. Note that for fine grain locking, there appears
     * to be a deadlock in that rxi_FreeCall has a call locked and
     * DestroyConnectionNoLock locks each call in the conn. But note a
     * few lines up where we have removed this call from the conn.
     * If someone else destroys a connection, they either have no
     * call lock held or are going through this section of code.
     */
    MUTEX_ENTER(&conn->conn_data_lock);
    if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
        rx_GetConnection(conn);
        MUTEX_EXIT(&conn->conn_data_lock);
#ifdef RX_ENABLE_LOCKS
        if (haveCTLock)
            rxi_DestroyConnectionNoLock(conn);
        else
            rxi_DestroyConnection(conn);
#else /* RX_ENABLE_LOCKS */
        rxi_DestroyConnection(conn);
#endif /* RX_ENABLE_LOCKS */
    } else {
        MUTEX_EXIT(&conn->conn_data_lock);
    }
    MUTEX_ENTER(&rx_refcnt_mutex);
    return 1;
}

rx_atomic_t rxi_Allocsize = RX_ATOMIC_INIT(0);
rx_atomic_t rxi_Alloccnt = RX_ATOMIC_INIT(0);

void *
rxi_Alloc(size_t size)
{
    char *p;

    if (rx_stats_active) {
        rx_atomic_add(&rxi_Allocsize, (int) size);
        rx_atomic_inc(&rxi_Alloccnt);
    }

p = (char *)
#if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD_ENV)
  afs_osi_Alloc_NoSleep(size);
#else
  osi_Alloc(size);
#endif
    if (!p)
        osi_Panic("rxi_Alloc error");
    memset(p, 0, size);
    return p;
}

void
rxi_Free(void *addr, size_t size)
{
    if (!addr) {
        return;
    }
    if (rx_stats_active) {
        rx_atomic_sub(&rxi_Allocsize, (int) size);
        rx_atomic_dec(&rxi_Alloccnt);
    }
    osi_Free(addr, size);
}

void
rxi_SetPeerMtu(struct rx_peer *peer, afs_uint32 host, afs_uint32 port, int mtu)
{
    struct rx_peer **peer_ptr = NULL, **peer_end = NULL;
    struct rx_peer *next = NULL;
    int hashIndex;

    if (!peer) {
        MUTEX_ENTER(&rx_peerHashTable_lock);
        if (port == 0) {
            peer_ptr = &rx_peerHashTable[0];
            peer_end = &rx_peerHashTable[rx_hashTableSize];
            next = NULL;
        resume:
            for ( ; peer_ptr < peer_end; peer_ptr++) {
                if (!peer)
                    peer = *peer_ptr;
                for ( ; peer; peer = next) {
                    next = peer->next;
                    if (host == peer->host)
                        break;
                }
            }
        } else {
            hashIndex = PEER_HASH(host, port);
            for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
                if ((peer->host == host) && (peer->port == port))
                    break;
            }
        }
    } else {
        MUTEX_ENTER(&rx_peerHashTable_lock);
    }

    if (peer) {
        peer->refCount++;
        MUTEX_EXIT(&rx_peerHashTable_lock);

        MUTEX_ENTER(&peer->peer_lock);
        /* We don't handle dropping below min, so don't */
        mtu = MAX(mtu, RX_MIN_PACKET_SIZE);
        peer->ifMTU=MIN(mtu, peer->ifMTU);
        peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
        /* if we tweaked this down, need to tune our peer MTU too */
        peer->MTU = MIN(peer->MTU, peer->natMTU);
        /* if we discovered a sub-1500 mtu, degrade */
        if (peer->ifMTU < OLD_MAX_PACKET_SIZE)
            peer->maxDgramPackets = 1;
        /* We no longer have valid peer packet information */
        if (peer->maxPacketSize + RX_HEADER_SIZE > peer->ifMTU)
            peer->maxPacketSize = 0;
        MUTEX_EXIT(&peer->peer_lock);

        MUTEX_ENTER(&rx_peerHashTable_lock);
        peer->refCount--;
        if (host && !port) {
            peer = next;
            /* pick up where we left off */
            goto resume;
        }
    }
    MUTEX_EXIT(&rx_peerHashTable_lock);
}

#ifdef AFS_RXERRQ_ENV
static void
rxi_SetPeerDead(struct sock_extended_err *err, afs_uint32 host, afs_uint16 port)
{
    int hashIndex = PEER_HASH(host, port);
    struct rx_peer *peer;

    MUTEX_ENTER(&rx_peerHashTable_lock);

    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_ENTER(&rx_peerHashTable_lock);
        peer->refCount--;
        MUTEX_EXIT(&rx_peerHashTable_lock);
    }
}

void
rxi_ProcessNetError(struct sock_extended_err *err, afs_uint32 addr, afs_uint16 port)
{
# ifdef AFS_ADAPT_PMTU
    if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
        rxi_SetPeerMtu(NULL, addr, port, err->ee_info - RX_IPUDP_SIZE);
        return;
    }
# endif
    if (err->ee_origin == SO_EE_ORIGIN_ICMP && err->ee_type == ICMP_DEST_UNREACH) {
        switch (err->ee_code) {
        case ICMP_NET_UNREACH:
        case ICMP_HOST_UNREACH:
        case ICMP_PORT_UNREACH:
        case ICMP_NET_ANO:
        case ICMP_HOST_ANO:
            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
 */
struct rx_peer *
rxi_FindPeer(afs_uint32 host, u_short port, int create)
{
    struct rx_peer *pp;
    int hashIndex;
    hashIndex = PEER_HASH(host, port);
    MUTEX_ENTER(&rx_peerHashTable_lock);
    for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
        if ((pp->host == host) && (pp->port == port))
            break;
    }
    if (!pp) {
        if (create) {
            pp = rxi_AllocPeer();       /* This bzero's *pp */
            pp->host = host;    /* set here or in InitPeerParams is zero */
            pp->port = port;
#ifdef AFS_RXERRQ_ENV
            rx_atomic_set(&pp->neterrs, 0);
#endif
            MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
            opr_queue_Init(&pp->rpcStats);
            pp->next = rx_peerHashTable[hashIndex];
            rx_peerHashTable[hashIndex] = pp;
            rxi_InitPeerParams(pp);
            if (rx_stats_active)
                rx_atomic_inc(&rx_stats.nPeerStructs);
        }
    }
    if (pp && create) {
        pp->refCount++;
    }
    MUTEX_EXIT(&rx_peerHashTable_lock);
    return pp;
}


/* Find the connection at (host, port) started at epoch, and with the
 * given connection id.  Creates the server connection if necessary.
 * The type specifies whether a client connection or a server
 * connection is desired.  In both cases, (host, port) specify the
 * peer's (host, pair) pair.  Client connections are not made
 * automatically by this routine.  The parameter socket gives the
 * socket descriptor on which the packet was received.  This is used,
 * in the case of server connections, to check that *new* connections
 * come via a valid (port, serviceId).  Finally, the securityIndex
 * parameter must match the existing index for the connection.  If a
 * server connection is created, it will be created using the supplied
 * index, if the index is valid for this service */
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,
                   int *unknownService)
{
    int hashindex, flag, i;
    int code = 0;
    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 =
                                                  rx_connHashTable[hashindex],
                                                  flag = 1);
    for (; conn;) {
        if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
            && (epoch == conn->epoch)) {
            struct rx_peer *pp = conn->peer;
            if (securityIndex != conn->securityIndex) {
                /* this isn't supposed to happen, but someone could forge a packet
                 * like this, and there seems to be some CM bug that makes this
                 * happen from time to time -- in which case, the fileserver
                 * asserts. */
                MUTEX_EXIT(&rx_connHashTable_lock);
                return (struct rx_connection *)0;
            }
            if (pp->host == host && pp->port == port)
                break;
            if (type == RX_CLIENT_CONNECTION && pp->port == port)
                break;
            /* So what happens when it's a callback connection? */
            if (                /*type == RX_CLIENT_CONNECTION && */
                   (conn->epoch & 0x80000000))
                break;
        }
        if (!flag) {
            /* the connection rxLastConn that was used the last time is not the
             ** one we are looking for now. Hence, start searching in the hash */
            flag = 1;
            conn = rx_connHashTable[hashindex];
        } else
            conn = conn->next;
    }
    if (!conn) {
        struct rx_service *service;
        if (type == RX_CLIENT_CONNECTION) {
            MUTEX_EXIT(&rx_connHashTable_lock);
            return (struct rx_connection *)0;
        }
        service = rxi_FindService(socket, serviceId);
        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 */
        MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
        MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
        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, 1);
        conn->type = RX_SERVER_CONNECTION;
        conn->lastSendTime = clock_Sec();       /* don't GC immediately */
        conn->epoch = epoch;
        conn->cid = cid & RX_CIDMASK;
        conn->ackRate = RX_FAST_ACK_RATE;
        conn->service = service;
        conn->serviceId = serviceId;
        conn->securityIndex = securityIndex;
        conn->securityObject = service->securityObjects[securityIndex];
        conn->nSpecific = 0;
        conn->specific = NULL;
        rx_SetConnDeadTime(conn, service->connDeadTime);
        rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
        for (i = 0; i < RX_MAXCALLS; i++) {
            conn->twind[i] = rx_initSendWindow;
            conn->rwind[i] = rx_initReceiveWindow;
        }
        /* Notify security object of the new connection */
        code = RXS_NewConnection(conn->securityObject, conn);
        /* XXXX Connection timeout? */
        if (service->newConnProc)
            (*service->newConnProc) (conn);
        if (rx_stats_active)
            rx_atomic_inc(&rx_stats.nServerConns);
    }

    rx_GetConnection(conn);

    rxLastConn = conn;          /* store this connection as the last conn used */
    MUTEX_EXIT(&rx_connHashTable_lock);
    if (code) {
        rxi_ConnectionError(conn, code);
    }
    return conn;
}

/*!
 * Abort the call if the server is over the busy threshold. This
 * can be used without requiring a call structure be initialised,
 * or connected to a particular channel
 */
static_inline int
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,
                         serial, rx_BusyError, np, 0);
        if (rx_stats_active)
            rx_atomic_inc(&rx_stats.nBusies);
        return 1;
    }

    return 0;
}

static_inline struct rx_call *
rxi_ReceiveClientCall(struct rx_packet *np, struct rx_connection *conn)
{
    int channel;
    struct rx_call *call;

    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)
            rx_atomic_inc(&rx_stats.spuriousPacketsRead);
        return NULL;
    }

    MUTEX_ENTER(&call->lock);
    MUTEX_EXIT(&conn->conn_call_lock);

    if ((call->state == RX_STATE_DALLY)
        && np->header.type == RX_PACKET_TYPE_ACK) {
        if (rx_stats_active)
            rx_atomic_inc(&rx_stats.ignorePacketDally);
        MUTEX_EXIT(&call->lock);
        return NULL;
    }

    return call;
}

static_inline struct rx_call *
rxi_ReceiveServerCall(osi_socket socket, struct rx_packet *np,
                      struct rx_connection *conn)
{
    int channel;
    struct rx_call *call;

    channel = np->header.cid & RX_CHANNELMASK;
    MUTEX_ENTER(&conn->conn_call_lock);
    call = conn->call[channel];

    if (!call) {
        if (rxi_AbortIfServerBusy(socket, conn, np)) {
            MUTEX_EXIT(&conn->conn_call_lock);
            return NULL;
        }

        call = rxi_NewCall(conn, channel);  /* returns locked call */
        *call->callNumber = np->header.callNumber;
        MUTEX_EXIT(&conn->conn_call_lock);

        call->state = RX_STATE_PRECALL;
        clock_GetTime(&call->queueTime);
        call->app.bytesSent = 0;
        call->app.bytesRcvd = 0;
        rxi_KeepAliveOn(call);

        return call;
    }

    if (np->header.callNumber == conn->callNumber[channel]) {
        MUTEX_ENTER(&call->lock);
        MUTEX_EXIT(&conn->conn_call_lock);
        return call;
    }

    if (np->header.callNumber < conn->callNumber[channel]) {
        MUTEX_EXIT(&conn->conn_call_lock);
        if (rx_stats_active)
            rx_atomic_inc(&rx_stats.spuriousPacketsRead);
        return NULL;
    }

    MUTEX_ENTER(&call->lock);
    MUTEX_EXIT(&conn->conn_call_lock);

    /* Wait until the transmit queue is idle before deciding
     * whether to reset the current call. Chances are that the
     * call will be in ether DALLY or HOLD state once the TQ_BUSY
     * flag is cleared.
     */
#ifdef RX_ENABLE_LOCKS
    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
         * to processed when the tqWaiter count hits zero.
         */
        if (call->error) {
            rxi_CallError(call, call->error);
            MUTEX_EXIT(&call->lock);
            return NULL;
        }
    }
#endif /* RX_ENABLE_LOCKS */
    /* If the new call cannot be taken right now send a busy and set
     * the error condition in this call, so that it terminates as
     * quickly as possible */
    if (call->state == RX_STATE_ACTIVE) {
        rxi_CallError(call, RX_CALL_DEAD);
        rxi_SendSpecial(call, conn, NULL, RX_PACKET_TYPE_BUSY,
                        NULL, 0, 1);
        MUTEX_EXIT(&call->lock);
        return NULL;
    }

    if (rxi_AbortIfServerBusy(socket, conn, np)) {
        MUTEX_EXIT(&call->lock);
        return NULL;
    }

    rxi_ResetCall(call, 0);
    /* The conn_call_lock is not held but no one else should be
     * using this call channel while we are processing this incoming
     * packet.  This assignment should be safe.
     */
    *call->callNumber = np->header.callNumber;
    call->state = RX_STATE_PRECALL;
    clock_GetTime(&call->queueTime);
    call->app.bytesSent = 0;
    call->app.bytesRcvd = 0;
    rxi_KeepAliveOn(call);

    return call;
}


/* There are two packet tracing routines available for testing and monitoring
 * Rx.  One is called just after every packet is received and the other is
 * called just before every packet is sent.  Received packets, have had their
 * headers decoded, and packets to be sent have not yet had their headers
 * encoded.  Both take two parameters: a pointer to the packet and a sockaddr
 * containing the network address.  Both can be modified.  The return value, if
 * non-zero, indicates that the packet should be dropped.  */

int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;

/* A packet has been received off the interface.  Np is the packet, socket is
 * the socket number it was received from (useful in determining which service
 * this packet corresponds to), and (host, port) reflect the host,port of the
 * sender.  This call returns the packet to the caller if it is finished with
 * it, rather than de-allocating it, just as a small performance hack */

struct rx_packet *
rxi_ReceivePacket(struct rx_packet *np, osi_socket socket,
                  afs_uint32 host, u_short port, int *tnop,
                  struct rx_call **newcallp)
{
    struct rx_call *call;
    struct rx_connection *conn;
    int type;
    int unknownService = 0;
#ifdef RXDEBUG
    char *packetType;
#endif
    struct rx_packet *tnp;

#ifdef RXDEBUG
/* We don't print out the packet until now because (1) the time may not be
 * accurate enough until now in the lwp implementation (rx_Listener only gets
 * the time after the packet is read) and (2) from a protocol point of view,
 * this is the first time the packet has been seen */
    packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
        ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
    dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %p\n",
         np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
         np->header.epoch, np->header.cid, np->header.callNumber,
         np->header.seq, np->header.flags, np));
#endif

    /* Account for connectionless packets */
    if (rx_stats_active &&
        ((np->header.type == RX_PACKET_TYPE_VERSION) ||
         (np->header.type == RX_PACKET_TYPE_DEBUG))) {
        struct rx_peer *peer;

        /* Try to look up the peer structure, but don't create one */
        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
         */

        if (peer && (peer->refCount > 0)) {
#ifdef AFS_RXERRQ_ENV
            if (rx_atomic_read(&peer->neterrs)) {
                rx_atomic_set(&peer->neterrs, 0);
            }
#endif
            MUTEX_ENTER(&peer->peer_lock);
            peer->bytesReceived += np->length;
            MUTEX_EXIT(&peer->peer_lock);
        }
    }

    if (np->header.type == RX_PACKET_TYPE_VERSION) {
        return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
    }

    if (np->header.type == RX_PACKET_TYPE_DEBUG) {
        return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
    }
#ifdef RXDEBUG
    /* If an input tracer function is defined, call it with the packet and
     * network address.  Note this function may modify its arguments. */
    if (rx_justReceived) {
        struct sockaddr_in addr;
        int drop;
        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
        drop = (*rx_justReceived) (np, &addr);
        /* drop packet if return value is non-zero */
        if (drop)
            return np;
        port = addr.sin_port;   /* in case fcn changed addr */
        host = addr.sin_addr.s_addr;
    }
#endif

    /* If packet was not sent by the client, then *we* must be the client */
    type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
        ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;

    /* Find the connection (or fabricate one, if we're the server & if
     * necessary) associated with this packet */
    conn =
        rxi_FindConnection(socket, host, port, np->header.serviceId,
                           np->header.cid, np->header.epoch, type,
                           np->header.securityIndex, &unknownService);

    /* To avoid having 2 connections just abort at each other,
       don't abort an abort. */
    if (!conn) {
        if (unknownService && (np->header.type != RX_PACKET_TYPE_ABORT))
            rxi_SendRawAbort(socket, host, port, 0, RX_INVALID_OPERATION,
                             np, 0);
        return np;
    }

#ifdef AFS_RXERRQ_ENV
    if (rx_atomic_read(&conn->peer->neterrs)) {
        rx_atomic_set(&conn->peer->neterrs, 0);
    }
#endif

    /* If we're doing statistics, then account for the incoming packet */
    if (rx_stats_active) {
        MUTEX_ENTER(&conn->peer->peer_lock);
        conn->peer->bytesReceived += np->length;
        MUTEX_EXIT(&conn->peer->peer_lock);
    }

    /* If the connection is in an error state, send an abort packet and ignore
     * the incoming packet */
    if (conn->error) {
        /* Don't respond to an abort packet--we don't want loops! */
        MUTEX_ENTER(&conn->conn_data_lock);
        if (np->header.type != RX_PACKET_TYPE_ABORT)
            np = rxi_SendConnectionAbort(conn, np, 1, 0);
        putConnection(conn);
        MUTEX_EXIT(&conn->conn_data_lock);
        return np;
    }

    /* Check for connection-only requests (i.e. not call specific). */
    if (np->header.callNumber == 0) {
        switch (np->header.type) {
        case RX_PACKET_TYPE_ABORT: {
            /* What if the supplied error is zero? */
            afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
            dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d\n", errcode));
            rxi_ConnectionError(conn, errcode);
            putConnection(conn);
            return np;
        }
        case RX_PACKET_TYPE_CHALLENGE:
            tnp = rxi_ReceiveChallengePacket(conn, np, 1);
            putConnection(conn);
            return tnp;
        case RX_PACKET_TYPE_RESPONSE:
            tnp = rxi_ReceiveResponsePacket(conn, np, 1);
            putConnection(conn);
            return tnp;
        case RX_PACKET_TYPE_PARAMS:
        case RX_PACKET_TYPE_PARAMS + 1:
        case RX_PACKET_TYPE_PARAMS + 2:
            /* ignore these packet types for now */
            putConnection(conn);
            return np;

        default:
            /* Should not reach here, unless the peer is broken: send an
             * abort packet */
            rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
            MUTEX_ENTER(&conn->conn_data_lock);
            tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
            putConnection(conn);
            MUTEX_EXIT(&conn->conn_data_lock);
            return tnp;
        }
    }

    if (type == RX_SERVER_CONNECTION)
        call = rxi_ReceiveServerCall(socket, np, conn);
    else
        call = rxi_ReceiveClientCall(np, conn);

    if (call == NULL) {
        putConnection(conn);
        return np;
    }

    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;
    case RX_PACKET_TYPE_ACK:
        /* Respond immediately to ack packets requesting acknowledgement
         * (ping packets) */
        if (np->header.flags & RX_REQUEST_ACK) {
            if (call->error)
                (void)rxi_SendCallAbort(call, 0, 1, 0);
            else
                (void)rxi_SendAck(call, 0, np->header.serial,
                                  RX_ACK_PING_RESPONSE, 1);
        }
        np = rxi_ReceiveAckPacket(call, np, 1);
        break;
    case RX_PACKET_TYPE_ABORT: {
        /* An abort packet: reset the call, passing the error up to the user. */
        /* What if error is zero? */
        /* What if the error is -1? the application will treat it as a timeout. */
        afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
        dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d\n", errdata));
        rxi_CallError(call, errdata);
        MUTEX_EXIT(&call->lock);
        putConnection(conn);
        return np;              /* xmitting; drop packet */
    }
    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
         * readied for sending */
        rxi_AckAllInTransmitQueue(call);
        break;
    default:
        /* Should not reach here, unless the peer is broken: send an abort
         * packet */
        rxi_CallError(call, RX_PROTOCOL_ERROR);
        np = rxi_SendCallAbort(call, np, 1, 0);
        break;
    };
    /* Note when this last legitimate packet was received, for keep-alive
     * processing.  Note, we delay getting the time until now in the hope that
     * 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();
    MUTEX_EXIT(&call->lock);
    putConnection(conn);
    return np;
}

/* return true if this is an "interesting" connection from the point of view
    of someone trying to debug the system */
int
rxi_IsConnInteresting(struct rx_connection *aconn)
{
    int i;
    struct rx_call *tcall;

    if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
        return 1;

    for (i = 0; i < RX_MAXCALLS; i++) {
        tcall = aconn->call[i];
        if (tcall) {
            if ((tcall->state == RX_STATE_PRECALL)
                || (tcall->state == RX_STATE_ACTIVE))
                return 1;
            if ((tcall->app.mode == RX_MODE_SENDING)
                || (tcall->app.mode == RX_MODE_RECEIVING))
                return 1;
        }
    }
    return 0;
}

#ifdef KERNEL
/* if this is one of the last few packets AND it wouldn't be used by the
   receiving call to immediately satisfy a read request, then drop it on
   the floor, since accepting it might prevent a lock-holding thread from
   making progress in its reading. If a call has been cleared while in
   the precall state then ignore all subsequent packets until the call
   is assigned to a thread. */

static int
TooLow(struct rx_packet *ap, struct rx_call *acall)
{
    int rc = 0;

    MUTEX_ENTER(&rx_quota_mutex);
    if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
         && (acall->state == RX_STATE_PRECALL))
        || ((rx_nFreePackets < rxi_dataQuota + 2)
            && !((ap->header.seq < acall->rnext + rx_initSendWindow)
                 && (acall->flags & RX_CALL_READER_WAIT)))) {
        rc = 1;
    }
    MUTEX_EXIT(&rx_quota_mutex);
    return rc;
}
#endif /* KERNEL */

/*!
 * Clear the attach wait flag on a connection and proceed.
 *
 * Any processing waiting for a connection to be attached should be
 * unblocked. We clear the flag and do any other needed tasks.
 *
 * @param[in] conn
 *      the conn to unmark waiting for attach
 *
 * @pre conn's conn_data_lock must be locked before calling this function
 *
 */
static void
rxi_ConnClearAttachWait(struct rx_connection *conn)
{
    /* Indicate that rxi_CheckReachEvent is no longer running by
     * clearing the flag.  Must be atomic under conn_data_lock to
     * avoid a new call slipping by: rxi_CheckConnReach holds
     * conn_data_lock while checking RX_CONN_ATTACHWAIT.
     */
    conn->flags &= ~RX_CONN_ATTACHWAIT;
    if (conn->flags & RX_CONN_NAT_PING) {
        conn->flags &= ~RX_CONN_NAT_PING;
        rxi_ScheduleNatKeepAliveEvent(conn);
    }
}

/*
 * 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 rx_connection *conn = arg1;
    struct rx_call *acall = arg2;
    struct rx_call *call = acall;
    struct clock when, now;
    int i, waiting;

    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;
    MUTEX_EXIT(&conn->conn_data_lock);

    if (waiting) {
        if (!call) {
            MUTEX_ENTER(&conn->conn_call_lock);
            MUTEX_ENTER(&conn->conn_data_lock);
            for (i = 0; i < RX_MAXCALLS; i++) {
                struct rx_call *tc = conn->call[i];
                if (tc && tc->state == RX_STATE_PRECALL) {
                    call = tc;
                    break;
                }
            }
            if (!call)
                rxi_ConnClearAttachWait(conn);
            MUTEX_EXIT(&conn->conn_data_lock);
            MUTEX_EXIT(&conn->conn_call_lock);
        }

        if (call) {
            if (call != acall)
                MUTEX_ENTER(&call->lock);
            rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
            if (call != acall)
                MUTEX_EXIT(&call->lock);

            clock_GetTime(&now);
            when = now;
            when.sec += RX_CHECKREACH_TIMEOUT;
            MUTEX_ENTER(&conn->conn_data_lock);
            if (!conn->checkReachEvent) {
                rx_GetConnection(conn);
                conn->checkReachEvent = rxevent_Post(&when, &now,
                                                     rxi_CheckReachEvent, conn,
                                                     NULL, 0);
            }