[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_SUN58_ENV
#   include "netinet/ip6.h"
#  endif
#  ifdef AFS_SUN57_ENV
#   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>
# include <sys/types.h>
# include <string.h>
# include <stdarg.h>
# include <errno.h>
# ifdef HAVE_STDINT_H
#  include <stdint.h>
# endif
# ifdef AFS_NT40_ENV
#  include <stdlib.h>
#  include <fcntl.h>
#  include <afs/afsutil.h>
#  include <WINNT\afsreg.h>
# else
#  include <sys/socket.h>
#  include <sys/file.h>
#  include <netdb.h>
#  include <sys/stat.h>
#  include <netinet/in.h>
#  include <sys/time.h>
# endif
# include "rx_user.h"
#endif /* KERNEL */

#include "rx.h"
#include "rx_clock.h"
#include "rx_queue.h"
#include "rx_atomic.h"
#include "rx_globals.h"
#include "rx_trace.h"
#include "rx_internal.h"
#include "rx_stats.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_peer *, struct clock *);

#ifdef RX_ENABLE_LOCKS
static void rxi_SetAcksInTransmitQueue(struct rx_call *call);
#endif

#ifdef  AFS_GLOBAL_RXLOCK_KERNEL
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 /* AFS_GLOBAL_RXLOCK_KERNEL */

/*
 * 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);

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

#ifdef RX_ENABLE_LOCKS
afs_kmutex_t rx_atomic_mutex;
#endif

#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;
extern afs_kmutex_t rx_clock_mutex;
extern afs_kmutex_t rxi_connCacheMutex;
extern afs_kmutex_t rx_event_mutex;
extern afs_kmutex_t osi_malloc_mutex;
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_kmutex_t rxkad_client_uid_mutex;
extern afs_kmutex_t rxkad_random_mutex;

extern afs_kcondvar_t rx_event_handler_cond;
extern afs_kcondvar_t rx_listener_cond;

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_clock_mutex, "clock", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_stats_mutex, "stats", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_atomic_mutex, "atomic", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_quota_mutex, "quota", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_pthread_mutex, "pthread", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_packets_mutex, "packets", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_refcnt_mutex, "refcnts", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&epoch_mutex, "epoch", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_init_mutex, "init", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_event_mutex, "event", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&osi_malloc_mutex, "malloc", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&event_handler_mutex, "event handler", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rxi_connCacheMutex, "conn cache", 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);
    MUTEX_INIT(&rxkad_client_uid_mutex, "uid", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rxkad_random_mutex, "rxkad random", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_debug_mutex, "debug", MUTEX_DEFAULT, 0);

    CV_INIT(&rx_event_handler_cond, "evhand", CV_DEFAULT, 0);
    CV_INIT(&rx_listener_cond, "rxlisten", CV_DEFAULT, 0);

    osi_Assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
    osi_Assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);

    rxkad_global_stats_init();

    MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
    MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_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);
    MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
#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) && defined(KERNEL)
static afs_kmutex_t rx_rpc_stats;
void rxi_StartUnlocked(struct rxevent *event, void *call,
                       void *arg1, int istack);
#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
 * 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;
struct rx_serverQueueEntry *rx_waitingForPacket = 0;

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

/* This function allows rxkad to set the epoch to a suitably random number
 * which rx_NewConnection will use in the future.  The principle purpose is to
 * get rxnull connections to use the same epoch as the rxkad connections do, at
 * least once the first rxkad connection is established.  This is important now
 * that the host/port addresses aren't used in FindConnection: the uniqueness
 * of epoch/cid matters and the start time won't do. */

#ifdef AFS_PTHREAD_ENV
/*
 * This mutex protects the following global variables:
 * rx_epoch
 */

#define LOCK_EPOCH MUTEX_ENTER(&epoch_mutex)
#define UNLOCK_EPOCH MUTEX_EXIT(&epoch_mutex)
#else
#define LOCK_EPOCH
#define UNLOCK_EPOCH
#endif /* AFS_PTHREAD_ENV */

void
rx_SetEpoch(afs_uint32 epoch)
{
    LOCK_EPOCH;
    rx_epoch = epoch;
    UNLOCK_EPOCH;
}

/* 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. */
#ifndef AFS_NT40_ENV
static
#endif
int rxinit_status = 1;
#ifdef AFS_PTHREAD_ENV
/*
 * This mutex protects the following global variables:
 * rxinit_status
 */

#define LOCK_RX_INIT MUTEX_ENTER(&rx_init_mutex)
#define UNLOCK_RX_INIT MUTEX_EXIT(&rx_init_mutex)
#else
#define LOCK_RX_INIT
#define UNLOCK_RX_INIT
#endif

int
rx_InitHost(u_int host, u_int port)
{
#ifdef KERNEL
    osi_timeval_t tv;
#else /* KERNEL */
    struct timeval tv;
#endif /* KERNEL */
    char *htable, *ptable;
    int tmp_status;

    SPLVAR;

    INIT_PTHREAD_LOCKS;
    LOCK_RX_INIT;
    if (rxinit_status == 0) {
        tmp_status = rxinit_status;
        UNLOCK_RX_INIT;
        return tmp_status;      /* Already started; return previous error code. */
    }
#ifdef RXDEBUG
    rxi_DebugInit();
#endif
#ifdef AFS_NT40_ENV
    if (afs_winsockInit() < 0)
        return -1;
#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) {
        UNLOCK_RX_INIT;
        return RX_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_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);
#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 = (char *)
        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 = (char *)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;
    queue_Init(&rx_freePacketQueue);
    rxi_NeedMorePackets = FALSE;
    rx_nPackets = 0;    /* rx_nPackets is managed by rxi_MorePackets* */

    /* 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)) {
            rx_Finalize();
            return -1;
        }
        rx_port = addr.sin_port;
#endif
    }
    rx_stats.minRtt.sec = 9999999;
#ifdef  KERNEL
    rx_SetEpoch(tv.tv_sec | 0x80000000);
#else
    rx_SetEpoch(tv.tv_sec);     /* Start time of this package, rxkad
                                 * will provide a randomer value. */
#endif
    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_lastAckDelay.sec = 0;
    rx_lastAckDelay.usec = 400000;      /* 400 milliseconds */
    rx_hardAckDelay.sec = 0;
    rx_hardAckDelay.usec = 100000;      /* 100 milliseconds */
    rx_softAckDelay.sec = 0;
    rx_softAckDelay.usec = 100000;      /* 100 milliseconds */

    rxevent_Init(20, rxi_ReScheduleEvents);

    /* Initialize various global queues */
    queue_Init(&rx_idleServerQueue);
    queue_Init(&rx_incomingCallQueue);
    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;
    tmp_status = rxinit_status = 0;
    UNLOCK_RX_INIT;
    return tmp_status;
}

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

/* 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. */
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);

    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;
    afs_int32 cid;
    struct rx_connection *conn;

    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);
    cid = (rx_nextCid += RX_MAXCALLS);
    conn->type = RX_CLIENT_CONNECTION;
    conn->cid = cid;
    conn->epoch = rx_epoch;
    conn->peer = rxi_FindPeer(shost, sport, 0, 1);
    conn->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;
    }

    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;
    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.
 */
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;
    struct rx_packet *packet;
    int i;
    SPLVAR;

    clock_NewTime();

    NETPRI;
    MUTEX_ENTER(&conn->conn_data_lock);
    MUTEX_ENTER(&rx_refcnt_mutex);
    if (conn->refCount > 0)
        conn->refCount--;
    else {
        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(&conn->conn_data_lock);
        USERPRI;
        return;
    }
    MUTEX_EXIT(&rx_refcnt_mutex);
    MUTEX_EXIT(&conn->conn_data_lock);

    /* Check for extant references to this connection */
    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 */
                    rxevent_Cancel(call->delayedAckEvent, call,
                                   RX_CALL_REFCOUNT_DELAY);
                    if (call->state == RX_STATE_PRECALL
                        || call->state == RX_STATE_ACTIVE) {
                        rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
                    } else {
                        rxi_AckAll(NULL, call, 0);
                    }
                }
                MUTEX_EXIT(&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;
    }

    if (conn->natKeepAliveEvent) {
        rxi_NatKeepAliveOff(conn);
    }

    if (conn->delayedAbortEvent) {
        rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
        packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
        if (packet) {
            MUTEX_ENTER(&conn->conn_data_lock);
            rxi_SendConnectionAbort(conn, packet, 0, 1);
            MUTEX_EXIT(&conn->conn_data_lock);
            rxi_FreePacket(packet);
        }
    }

    /* Remove from connection hash table before proceeding */
    conn_ptr =
        &rx_connHashTable[CONN_HASH
                          (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. */
    /* get rid of pending events that could zap us later */
    if (conn->challengeEvent)
        rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
    if (conn->checkReachEvent)
        rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
    if (conn->natKeepAliveEvent)
        rxevent_Cancel(conn->natKeepAliveEvent, (struct rx_call *)0, 0);

    /* 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  AFS_GLOBAL_RXLOCK_KERNEL
/* 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++;
#ifdef RX_ENABLE_LOCKS
        osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
        CV_WAIT(&call->cv_tq, &call->lock);
#else /* RX_ENABLE_LOCKS */
        osi_rxSleep(&call->tq);
#endif /* RX_ENABLE_LOCKS */
        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 %"AFS_PTR_FMT" has %d waiters and flags %d\n",
             call, call->tqWaiters, call->flags));
#ifdef RX_ENABLE_LOCKS
        osirx_AssertMine(&call->lock, "rxi_Start start");
        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 %"AFS_PTR_FMT")\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;
                        MUTEX_EXIT(&conn->conn_call_lock);
                        MUTEX_ENTER(&rx_refcnt_mutex);
                        CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
                        MUTEX_EXIT(&rx_refcnt_mutex);
                        rxi_ResetCall(call, 0);
                        (*call->callNumber)++;
                        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.
                         */
                        MUTEX_ENTER(&rx_refcnt_mutex);
                        CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
                        MUTEX_EXIT(&rx_refcnt_mutex);
                        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);
                MUTEX_ENTER(&rx_refcnt_mutex);
                CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
                MUTEX_EXIT(&rx_refcnt_mutex);
                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->mode = RX_MODE_ERROR;
    else
        call->mode = RX_MODE_SENDING;

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

    /* 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
    CV_BROADCAST(&conn->conn_call_cv);
#else
    osi_rxWakeup(conn);
#endif
    MUTEX_EXIT(&conn->conn_call_lock);

#ifdef  AFS_GLOBAL_RXLOCK_KERNEL
    if (call->flags & (RX_CALL_TQ_BUSY | RX_CALL_TQ_CLEARME)) {
        osi_Panic("rx_NewCall call about to be used without an empty tq");
    }
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */

    MUTEX_EXIT(&call->lock);
    USERPRI;

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

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;
    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];
    }
    USERPRI;
    return 0;
}

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

    NETPRI;
    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];
    }
    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;

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

    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->idleDeadErr = 0;
            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;
    for (i = 0; i<service->nSecurityObjects; i++) {
        if (service->securityObjects[i]) {
            RXS_SetConfiguration(service->securityObjects[i], NULL, type,
                                 value, NULL);
        }
    }
    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;
            }
        }

        /* 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;
        }
#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

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


void
rx_WakeupServerProcs(void)
{
    struct rx_serverQueueEntry *np, *tqp;
    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 (np = rx_FreeSQEList; np; np = tqp) {
        tqp = *(struct rx_serverQueueEntry **)np;
#ifdef RX_ENABLE_LOCKS
        CV_BROADCAST(&np->cv);
#else /* RX_ENABLE_LOCKS */
        osi_rxWakeup(np);
#endif /* RX_ENABLE_LOCKS */
    }
    MUTEX_EXIT(&freeSQEList_lock);
    for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
#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;
    SPLVAR;

    MUTEX_ENTER(&freeSQEList_lock);

    if ((sq = rx_FreeSQEList)) {
        rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
        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 (queue_IsNotEmpty(&rx_incomingCallQueue)) {
            struct rx_call *tcall, *ncall, *choice2 = NULL;

            /* 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 (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
                service = tcall->conn->service;
                if (!QuotaOK(service)) {
                    continue;
                }
                MUTEX_ENTER(&rx_pthread_mutex);
                if (tno == rxi_fcfs_thread_num
                    || !tcall->queue_item_header.next) {
                    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 (!queue_IsEmpty(&tcall->rq)) {
                        struct rx_packet *rp;
                        rp = queue_First(&tcall->rq, rx_packet);
                        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) {
            queue_Remove(call);
            MUTEX_EXIT(&rx_serverPool_lock);
            MUTEX_ENTER(&call->lock);

            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 (queue_IsEmpty(&call->rq)
                || queue_First(&call->rq, rx_packet)->header.seq != 1)
                rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);

            CLEAR_CALL_QUEUE_LOCK(call);
            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;
            queue_Append(&rx_idleServerQueue, sq);
#ifndef AFS_AIX41_ENV
            rx_waitForPacket = sq;
#else
            rx_waitingForPacket = sq;
#endif /* AFS_AIX41_ENV */
            do {
                CV_WAIT(&sq->cv, &rx_serverPool_lock);
#ifdef  KERNEL
                if (afs_termState == AFSOP_STOP_RXCALLBACK) {
                    MUTEX_EXIT(&rx_serverPool_lock);
                    return (struct rx_call *)0;
                }
#endif
            } while (!(call = sq->newcall)
                     && !(socketp && *socketp != OSI_NULLSOCKET));
            MUTEX_EXIT(&rx_serverPool_lock);
            if (call) {
                MUTEX_ENTER(&call->lock);
            }
            break;
        }
    }

    MUTEX_ENTER(&freeSQEList_lock);
    *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
    rx_FreeSQEList = sq;
    MUTEX_EXIT(&freeSQEList_lock);

    if (call) {
        clock_GetTime(&call->startTime);
        call->state = RX_STATE_ACTIVE;
        call->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 %"AFS_PTR_FMT"\n",
             call->conn->service->servicePort, call->conn->service->serviceId,
             call));

        MUTEX_EXIT(&call->lock);
        MUTEX_ENTER(&rx_refcnt_mutex);
        CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
        MUTEX_EXIT(&rx_refcnt_mutex);
    } 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 ((sq = rx_FreeSQEList)) {
        rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
        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 (queue_IsNotEmpty(&rx_incomingCallQueue)) {
        struct rx_call *tcall, *ncall;
        /* 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 (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
            service = tcall->conn->service;
            if (QuotaOK(service)) {
                MUTEX_ENTER(&rx_pthread_mutex);
                if (tno == rxi_fcfs_thread_num
                    || !tcall->queue_item_header.next) {
                    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 (!queue_IsEmpty(&tcall->rq)) {
                        struct rx_packet *rp;
                        rp = queue_First(&tcall->rq, rx_packet);
                        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) {
        queue_Remove(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 (queue_IsEmpty(&call->rq)
            || queue_First(&call->rq, rx_packet)->header.seq != 1
            || call->rprev != queue_Last(&call->rq, rx_packet)->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;
        queue_Append(&rx_idleServerQueue, sq);
        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);
    *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
    rx_FreeSQEList = sq;
    MUTEX_EXIT(&freeSQEList_lock);

    if (call) {
        clock_GetTime(&call->startTime);
        call->state = RX_STATE_ACTIVE;
        call->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 %"AFS_PTR_FMT" 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 = (void (*)())0;
    if (rc && call->error == 0) {
        rxi_CallError(call, rc);
        call->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->mode == RX_MODE_RECEIVING) {
            MUTEX_EXIT(&call->lock);
            rxi_WriteProc(call, 0, 0);
            MUTEX_ENTER(&call->lock);
        }
        if (call->mode == RX_MODE_SENDING) {
            MUTEX_EXIT(&call->lock);
            rxi_FlushWrite(call);
            MUTEX_ENTER(&call->lock);
        }
        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);
            rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
            rxevent_Cancel(call->keepAliveEvent, call,
                           RX_CALL_REFCOUNT_ALIVE);
        }
    } else {                    /* Client connection */
        char dummy;
        /* Make sure server receives input packets, in the case where
         * no reply arguments are expected */
        if ((call->mode == RX_MODE_SENDING)
            || (call->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) {
            rxevent_Cancel(call->delayedAckEvent, call,
                           RX_CALL_REFCOUNT_DELAY);
            call->delayedAckEvent = NULL;
            rxi_SendDelayedAck(NULL, call, NULL);
        }

        /* 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->flags & RX_CALL_PEER_BUSY)) {
            conn->lastBusy[call->channel] = 0;
        }

        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->currentPacket) {
#ifdef RX_TRACK_PACKETS
        call->currentPacket->flags &= ~RX_PKTFLAG_CP;
#endif
        rxi_FreePacket(call->currentPacket);
        call->currentPacket = (struct rx_packet *)0;
    }

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

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

    MUTEX_ENTER(&rx_refcnt_mutex);
    CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
    MUTEX_EXIT(&rx_refcnt_mutex);
    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);
    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)
{
    struct rx_connection **conn_ptr, **conn_end;

    INIT_PTHREAD_LOCKS;
    LOCK_RX_INIT;
    if (rxinit_status == 1) {
        UNLOCK_RX_INIT;
        return;                 /* Already shutdown. */
    }
    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) {
                    MUTEX_ENTER(&rx_refcnt_mutex);
                    conn->refCount++;
                    MUTEX_EXIT(&rx_refcnt_mutex);
#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

    rxinit_status = 1;
    UNLOCK_RX_INIT;
}
#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 */
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. */
struct rx_call *
rxi_NewCall(struct rx_connection *conn, int channel)
{
    struct rx_call *call;
#ifdef  AFS_GLOBAL_RXLOCK_KERNEL
    struct rx_call *cp; /* Call pointer temp */
    struct rx_call *nxp;        /* Next call pointer, for queue_Scan */
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */

    dpf(("rxi_NewCall(conn %"AFS_PTR_FMT", 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  AFS_GLOBAL_RXLOCK_KERNEL
    /*
     * EXCEPT that the TQ might not yet be cleared out.
     * Skip over those with in-use TQs.
     */
    call = NULL;
    for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
        if (!(cp->flags & RX_CALL_TQ_BUSY)) {
            call = cp;
            break;
        }
    }
    if (call) {
#else /* AFS_GLOBAL_RXLOCK_KERNEL */
    if (queue_IsNotEmpty(&rx_freeCallQueue)) {
        call = queue_First(&rx_freeCallQueue, rx_call);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
        queue_Remove(call);
        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  AFS_GLOBAL_RXLOCK_KERNEL
        /* 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 /* AFS_GLOBAL_RXLOCK_KERNEL */
        /* 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 */
        queue_Init(&call->tq);
        queue_Init(&call->rq);
        queue_Init(&call->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.
 */
void
rxi_FreeCall(struct rx_call *call, int haveCTLock)
{
    int channel = call->channel;
    struct rx_connection *conn = call->conn;


    if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
        (*call->callNumber)++;
    /*
     * 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);
    call->conn->call[channel] = (struct rx_call *)0;

    MUTEX_ENTER(&rx_freeCallQueue_lock);
    SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
#ifdef AFS_GLOBAL_RXLOCK_KERNEL
    /* 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)
        queue_Prepend(&rx_freeCallQueue, call);
    else
        queue_Append(&rx_freeCallQueue, call);
#else /* AFS_GLOBAL_RXLOCK_KERNEL */
    queue_Append(&rx_freeCallQueue, call);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
    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)) {
        MUTEX_ENTER(&rx_refcnt_mutex);
        conn->refCount++;
        MUTEX_EXIT(&rx_refcnt_mutex);
        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);
}

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_FBSD80_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 (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_IPUDP_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);
}

/* Find the peer process represented by the supplied (host,port)
 * combination.  If there is no appropriate active peer structure, a
 * new one will be allocated and initialized
 * The origPeer, if set, is a pointer to a peer structure on which the
 * refcount will be be decremented. This is used to replace the peer
 * structure hanging off a connection structure */
struct rx_peer *
rxi_FindPeer(afs_uint32 host, u_short port,
             struct rx_peer *origPeer, int create)
{
    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;
            MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
            queue_Init(&pp->congestionQueue);
            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++;
    }
    if (origPeer)
        origPeer->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 */
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 hashindex, flag, i;
    struct rx_connection *conn;
    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);
            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, 0, 1);
        conn->type = RX_SERVER_CONNECTION;
        conn->lastSendTime = clock_Sec();       /* don't GC immediately */
        conn->epoch = epoch;
        conn->cid = cid & RX_CIDMASK;
        /* conn->serial = conn->lastSerial = 0; */
        /* conn->timeout = 0; */
        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);
        rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
        for (i = 0; i < RX_MAXCALLS; i++) {
            conn->twind[i] = rx_initSendWindow;
            conn->rwind[i] = rx_initReceiveWindow;
        }
        /* Notify security object of the new connection */
        RXS_NewConnection(conn->securityObject, conn);
        /* XXXX Connection timeout? */
        if (service->newConnProc)
            (*service->newConnProc) (conn);
        if (rx_stats_active)
            rx_atomic_inc(&rx_stats.nServerConns);
    }

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

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

/**
 * Timeout a call on a busy call channel if appropriate.
 *
 * @param[in] call The busy call.
 *
 * @pre 'call' is marked as busy (namely,
 *      call->conn->lastBusy[call->channel] != 0)
 *
 * @pre call->lock is held
 *
 * @note call->lock is dropped and reacquired
 */
static void
rxi_CheckBusy(struct rx_call *call)
{
    struct rx_connection *conn = call->conn;
    int channel = call->channel;
    int freechannel = 0;
    int i;
    afs_uint32 callNumber = *call->callNumber;

    MUTEX_EXIT(&call->lock);

    MUTEX_ENTER(&conn->conn_call_lock);

    /* Are there any other call slots on this conn that we should try? Look for
     * slots that are empty and are either non-busy, or were marked as busy
     * longer than conn->secondsUntilDead seconds before this call started. */

    for (i = 0; i < RX_MAXCALLS && !freechannel; i++) {
        if (i == channel) {
            /* only look at channels that aren't us */
            continue;
        }

        if (conn->lastBusy[i]) {
            /* if this channel looked busy too recently, don't look at it */
            if (conn->lastBusy[i] >= call->startTime.sec) {
                continue;
            }
            if (call->startTime.sec - conn->lastBusy[i] < conn->secondsUntilDead) {
                continue;
            }
        }

        if (conn->call[i]) {
            struct rx_call *tcall = conn->call[i];
            MUTEX_ENTER(&tcall->lock);
            if (tcall->state == RX_STATE_DALLY) {
                freechannel = 1;
            }
            MUTEX_EXIT(&tcall->lock);
        } else {
            freechannel = 1;
        }
    }

    MUTEX_EXIT(&conn->conn_call_lock);

    MUTEX_ENTER(&call->lock);

    /* Since the call->lock and conn->conn_call_lock have been released it is
     * possible that (1) the call may no longer be busy and/or (2) the call may
     * have been reused by another waiting thread. Therefore, we must confirm
     * that the call state has not changed when deciding whether or not to
     * force this application thread to retry by forcing a Timeout error. */

    if (freechannel && *call->callNumber == callNumber &&
        (call->flags & RX_CALL_PEER_BUSY)) {
        /* Since 'freechannel' is set, there exists another channel in this
         * rx_conn that the application thread might be able to use. We know
         * that we have the correct call since callNumber is unchanged, and we
         * know that the call is still busy. So, set the call error state to
         * RX_CALL_TIMEOUT so the application can retry the request, presumably
         * on a less-busy call channel. */

        rxi_CallError(call, RX_CALL_TIMEOUT);
    }
}

/* 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 channel;
    afs_uint32 currentCallNumber;
    int type;
    int skew;
#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 %"AFS_PTR_FMT"\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

    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;
#ifdef STRUCT_SOCKADDR_HAS_SA_LEN
        addr.sin_len = sizeof(addr);
#endif /* AFS_OSF_ENV */
        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);

    if (!conn) {
        /* If no connection found or fabricated, just ignore the packet.
         * (An argument could be made for sending an abort packet for
         * the conn) */
        return np;
    }

    MUTEX_ENTER(&conn->conn_data_lock);
    if (conn->maxSerial < np->header.serial)
        conn->maxSerial = np->header.serial;
    MUTEX_EXIT(&conn->conn_data_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);
        MUTEX_ENTER(&rx_refcnt_mutex);
        conn->refCount--;
        MUTEX_EXIT(&rx_refcnt_mutex);
        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);
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            return np;
        }
        case RX_PACKET_TYPE_CHALLENGE:
            tnp = rxi_ReceiveChallengePacket(conn, np, 1);
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            return tnp;
        case RX_PACKET_TYPE_RESPONSE:
            tnp = rxi_ReceiveResponsePacket(conn, np, 1);
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            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 */
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            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);
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            MUTEX_EXIT(&conn->conn_data_lock);
            return tnp;
        }
    }

    channel = np->header.cid & RX_CHANNELMASK;
    call = conn->call[channel];
#ifdef  RX_ENABLE_LOCKS
    if (call)
        MUTEX_ENTER(&call->lock);
    /* Test to see if call struct is still attached to conn. */
    if (call != conn->call[channel]) {
        if (call)
            MUTEX_EXIT(&call->lock);
        if (type == RX_SERVER_CONNECTION) {
            call = conn->call[channel];
            /* If we started with no call attached and there is one now,
             * another thread is also running this routine and has gotten
             * the connection channel. We should drop this packet in the tests
             * below. If there was a call on this connection and it's now
             * gone, then we'll be making a new call below.
             * If there was previously a call and it's now different then
             * the old call was freed and another thread running this routine
             * has created a call on this channel. One of these two threads
             * has a packet for the old call and the code below handles those
             * cases.
             */
            if (call)
                MUTEX_ENTER(&call->lock);
        } else {
            /* This packet can't be for this call. If the new call address is
             * 0 then no call is running on this channel. If there is a call
             * then, since this is a client connection we're getting data for
             * it must be for the previous call.
             */
            if (rx_stats_active)
                rx_atomic_inc(&rx_stats.spuriousPacketsRead);
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            return np;
        }
    }
#endif
    currentCallNumber = conn->callNumber[channel];

    if (type == RX_SERVER_CONNECTION) { /* We're the server */
        if (np->header.callNumber < currentCallNumber) {
            if (rx_stats_active)
                rx_atomic_inc(&rx_stats.spuriousPacketsRead);
#ifdef  RX_ENABLE_LOCKS
            if (call)
                MUTEX_EXIT(&call->lock);
#endif
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            return np;
        }
        if (!call) {
            MUTEX_ENTER(&conn->conn_call_lock);
            call = rxi_NewCall(conn, channel);
            MUTEX_EXIT(&conn->conn_call_lock);
            *call->callNumber = np->header.callNumber;
#ifdef RXDEBUG
            if (np->header.callNumber == 0)
                dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%.06d len %d\n",
                      np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
                      np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
                      np->header.flags, np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
#endif
            call->state = RX_STATE_PRECALL;
            clock_GetTime(&call->queueTime);
            hzero(call->bytesSent);
            hzero(call->bytesRcvd);
            /*
             * If the number of queued calls exceeds the overload
             * threshold then abort this call.
             */
            if ((rx_BusyThreshold > 0) &&
                (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
                struct rx_packet *tp;

                rxi_CallError(call, rx_BusyError);
                tp = rxi_SendCallAbort(call, np, 1, 0);
                MUTEX_EXIT(&call->lock);
                MUTEX_ENTER(&rx_refcnt_mutex);
                conn->refCount--;
                MUTEX_EXIT(&rx_refcnt_mutex);
                if (rx_stats_active)
                    rx_atomic_inc(&rx_stats.nBusies);
                return tp;
            }
            rxi_KeepAliveOn(call);
        } else if (np->header.callNumber != currentCallNumber) {
            /* 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 AFS_GLOBAL_RXLOCK_KERNEL
            if (call->state == RX_STATE_ACTIVE) {
                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);
                    MUTEX_ENTER(&rx_refcnt_mutex);
                    conn->refCount--;
                    MUTEX_EXIT(&rx_refcnt_mutex);
                    return np;
                }
            }
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
            /* 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) {
                struct rx_packet *tp;

                rxi_CallError(call, RX_CALL_DEAD);
                tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
                                     NULL, 0, 1);
                MUTEX_EXIT(&call->lock);
                MUTEX_ENTER(&rx_refcnt_mutex);
                conn->refCount--;
                MUTEX_EXIT(&rx_refcnt_mutex);
                return tp;
            }
            rxi_ResetCall(call, 0);
            *call->callNumber = np->header.callNumber;
#ifdef RXDEBUG
            if (np->header.callNumber == 0)
                dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %"AFS_PTR_FMT" resend %d.%06d len %d\n",
                      np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port),
                      np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq,
                      np->header.flags, np, np->retryTime.sec, np->retryTime.usec, np->length));
#endif
            call->state = RX_STATE_PRECALL;
            clock_GetTime(&call->queueTime);
            hzero(call->bytesSent);
            hzero(call->bytesRcvd);
            /*
             * If the number of queued calls exceeds the overload
             * threshold then abort this call.
             */
            if ((rx_BusyThreshold > 0) &&
                (rx_atomic_read(&rx_nWaiting) > rx_BusyThreshold)) {
                struct rx_packet *tp;

                rxi_CallError(call, rx_BusyError);
                tp = rxi_SendCallAbort(call, np, 1, 0);
                MUTEX_EXIT(&call->lock);
                MUTEX_ENTER(&rx_refcnt_mutex);
                conn->refCount--;
                MUTEX_EXIT(&rx_refcnt_mutex);
                if (rx_stats_active)
                    rx_atomic_inc(&rx_stats.nBusies);
                return tp;
            }
            rxi_KeepAliveOn(call);
        } else {
            /* Continuing call; do nothing here. */
        }
    } else {                    /* we're the client */
        /* Ignore all incoming acknowledgements for calls in DALLY state */
        if (call && (call->state == RX_STATE_DALLY)
            && (np->header.type == RX_PACKET_TYPE_ACK)) {
            if (rx_stats_active)
                rx_atomic_inc(&rx_stats.ignorePacketDally);
#ifdef  RX_ENABLE_LOCKS
            if (call) {
                MUTEX_EXIT(&call->lock);
            }
#endif
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            return np;
        }

        /* Ignore anything that's not relevant to the current call.  If there
         * isn't a current call, then no packet is relevant. */
        if (!call || (np->header.callNumber != currentCallNumber)) {
            if (rx_stats_active)
                rx_atomic_inc(&rx_stats.spuriousPacketsRead);
#ifdef  RX_ENABLE_LOCKS
            if (call) {
                MUTEX_EXIT(&call->lock);
            }
#endif
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            return np;
        }
        /* If the service security object index stamped in the packet does not
         * match the connection's security index, ignore the packet */
        if (np->header.securityIndex != conn->securityIndex) {
#ifdef  RX_ENABLE_LOCKS
            MUTEX_EXIT(&call->lock);
#endif
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            return np;
        }

        /* If we're receiving the response, then all transmit packets are
         * implicitly acknowledged.  Get rid of them. */
        if (np->header.type == RX_PACKET_TYPE_DATA) {
#ifdef  AFS_GLOBAL_RXLOCK_KERNEL
            /* XXX Hack. Because we must release the global rx lock when
             * sending packets (osi_NetSend) we drop all acks while we're
             * traversing the tq in rxi_Start sending packets out because
             * packets may move to the freePacketQueue as result of being here!
             * So we drop these packets until we're safely out of the
             * traversing. Really ugly!
             * For fine grain RX locking, we set the acked field in the
             * packets and let rxi_Start remove them from the transmit queue.
             */
            if (call->flags & RX_CALL_TQ_BUSY) {
#ifdef  RX_ENABLE_LOCKS
                rxi_SetAcksInTransmitQueue(call);
#else
                MUTEX_ENTER(&rx_refcnt_mutex);
                conn->refCount--;
                MUTEX_EXIT(&rx_refcnt_mutex);
                return np;      /* xmitting; drop packet */
#endif
            } else {
                rxi_ClearTransmitQueue(call, 0);
            }
#else /* AFS_GLOBAL_RXLOCK_KERNEL */
            rxi_ClearTransmitQueue(call, 0);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
        } else {
            if (np->header.type == RX_PACKET_TYPE_ACK) {
                /* now check to see if this is an ack packet acknowledging that the
                 * server actually *lost* some hard-acked data.  If this happens we
                 * ignore this packet, as it may indicate that the server restarted in
                 * the middle of a call.  It is also possible that this is an old ack
                 * packet.  We don't abort the connection in this case, because this
                 * *might* just be an old ack packet.  The right way to detect a server
                 * restart in the midst of a call is to notice that the server epoch
                 * changed, btw.  */
                /* XXX I'm not sure this is exactly right, since tfirst **IS**
                 * XXX unacknowledged.  I think that this is off-by-one, but
                 * XXX I don't dare change it just yet, since it will
                 * XXX interact badly with the server-restart detection
                 * XXX code in receiveackpacket.  */
                if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
                    if (rx_stats_active)
                        rx_atomic_inc(&rx_stats.spuriousPacketsRead);
                    MUTEX_EXIT(&call->lock);
                    MUTEX_ENTER(&rx_refcnt_mutex);
                    conn->refCount--;
                    MUTEX_EXIT(&rx_refcnt_mutex);
                    return np;
                }
            }
        }                       /* else not a data packet */
    }

    osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
    /* Set remote user defined status from packet */
    call->remoteStatus = np->header.userStatus;

    /* Note the gap between the expected next packet and the actual
     * packet that arrived, when the new packet has a smaller serial number
     * than expected.  Rioses frequently reorder packets all by themselves,
     * so this will be quite important with very large window sizes.
     * Skew is checked against 0 here to avoid any dependence on the type of
     * inPacketSkew (which may be unsigned).  In C, -1 > (unsigned) 0 is always
     * true!
     * The inPacketSkew should be a smoothed running value, not just a maximum.  MTUXXX
     * see CalculateRoundTripTime for an example of how to keep smoothed values.
     * I think using a beta of 1/8 is probably appropriate.  93.04.21
     */
    MUTEX_ENTER(&conn->conn_data_lock);
    skew = conn->lastSerial - np->header.serial;
    conn->lastSerial = np->header.serial;
    MUTEX_EXIT(&conn->conn_data_lock);
    if (skew > 0) {
        struct rx_peer *peer;
        peer = conn->peer;
        if (skew > peer->inPacketSkew) {
            dpf(("*** In skew changed from %d to %d\n",
                  peer->inPacketSkew, skew));
            peer->inPacketSkew = skew;
        }
    }

    /* Now do packet type-specific processing */
    switch (np->header.type) {
    case RX_PACKET_TYPE_DATA:
        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);
        MUTEX_ENTER(&rx_refcnt_mutex);
        conn->refCount--;
        MUTEX_EXIT(&rx_refcnt_mutex);
        return np;              /* xmitting; drop packet */
    }
    case RX_PACKET_TYPE_BUSY: {
        struct clock busyTime;
        clock_NewTime();
        clock_GetTime(&busyTime);

        MUTEX_EXIT(&call->lock);

        MUTEX_ENTER(&conn->conn_call_lock);
        MUTEX_ENTER(&call->lock);
        conn->lastBusy[call->channel] = busyTime.sec;
        call->flags |= RX_CALL_PEER_BUSY;
        MUTEX_EXIT(&call->lock);
        MUTEX_EXIT(&conn->conn_call_lock);

        MUTEX_ENTER(&rx_refcnt_mutex);
        conn->refCount--;
        MUTEX_EXIT(&rx_refcnt_mutex);
        return np;
    }

    case RX_PACKET_TYPE_ACKALL:
        /* All packets acknowledged, so we can drop all packets previously
         * readied for sending */
#ifdef  AFS_GLOBAL_RXLOCK_KERNEL
        /* XXX Hack. We because we can't release the global rx lock when
         * sending packets (osi_NetSend) we drop all ack pkts while we're
         * traversing the tq in rxi_Start sending packets out because
         * packets may move to the freePacketQueue as result of being
         * here! So we drop these packets until we're safely out of the
         * traversing. Really ugly!
         * For fine grain RX locking, we set the acked field in the packets
         * and let rxi_Start remove the packets from the transmit queue.
         */
        if (call->flags & RX_CALL_TQ_BUSY) {
#ifdef  RX_ENABLE_LOCKS
            rxi_SetAcksInTransmitQueue(call);
            break;
#else /* RX_ENABLE_LOCKS */
            MUTEX_EXIT(&call->lock);
            MUTEX_ENTER(&rx_refcnt_mutex);
            conn->refCount--;
            MUTEX_EXIT(&rx_refcnt_mutex);
            return np;          /* xmitting; drop packet */
#endif /* RX_ENABLE_LOCKS */
        }
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
        rxi_ClearTransmitQueue(call, 0);
        rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
        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();
    /* we've received a legit packet, so the channel is not busy */
    call->flags &= ~RX_CALL_PEER_BUSY;
    MUTEX_EXIT(&call->lock);
    MUTEX_ENTER(&rx_refcnt_mutex);
    conn->refCount--;
    MUTEX_EXIT(&rx_refcnt_mutex);
    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->mode == RX_MODE_SENDING)
                || (tcall->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 */

static void
rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
{
    struct rx_connection *conn = arg1;
    struct rx_call *acall = arg2;
    struct rx_call *call = acall;
    struct clock when, now;
    int i, waiting;

    MUTEX_ENTER(&conn->conn_data_lock);
    conn->checkReachEvent = NULL;
    waiting = conn->flags & RX_CONN_ATTACHWAIT;
    if (event) {
        MUTEX_ENTER(&rx_refcnt_mutex);
        conn->refCount--;
        MUTEX_EXIT(&rx_refcnt_mutex);
    }
    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)
                /* 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;
            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) {
                MUTEX_ENTER(&rx_refcnt_mutex);
                conn->refCount++;
                MUTEX_EXIT(&rx_refcnt_mutex);
                conn->checkReachEvent =
                    rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
                                    NULL);
            }
            MUTEX_EXIT(&conn->conn_data_lock);
        }
    }
}

static int
rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
{
    struct rx_service *service = conn->service;
    struct rx_peer *peer = conn->peer;
    afs_uint32 now, lastReach;

    if (service->checkReach == 0)
        return 0;

    now = clock_Sec();
    MUTEX_ENTER(&peer->peer_lock);
    lastReach = peer->lastReachTime;
    MUTEX_EXIT(&peer->peer_lock);
    if (now - lastReach < RX_CHECKREACH_TTL)
        return 0;

    MUTEX_ENTER(&conn->conn_data_lock);
    if (conn->flags & RX_CONN_ATTACHWAIT) {
        MUTEX_EXIT(&conn->conn_data_lock);
        return 1;
    }
    conn->flags |= RX_CONN_ATTACHWAIT;
    MUTEX_EXIT(&conn->conn_data_lock);
    if (!conn->checkReachEvent)
        rxi_CheckReachEvent(NULL, conn, call);

    return 1;
}

/* try to attach call, if authentication is complete */
static void
TryAttach(struct rx_call *acall, osi_socket socket,
          int *tnop, struct rx_call **newcallp,
          int reachOverride)
{
    struct rx_connection *conn = acall->conn;

    if (conn->type == RX_SERVER_CONNECTION
        && acall->state == RX_STATE_PRECALL) {
        /* Don't attach until we have any req'd. authentication. */
        if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
            if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
                rxi_AttachServerProc(acall, socket, tnop, newcallp);
            /* Note:  this does not necessarily succeed; there
             * may not any proc available
             */
        } else {
            rxi_ChallengeOn(acall->conn);
        }
    }
}

/* A data packet has been received off the interface.  This packet is
 * appropriate to the call (the call is in the right state, etc.).  This
 * routine can return a packet to the caller, for re-use */

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)
{
    int ackNeeded = 0;          /* 0 means no, otherwise ack_reason */
    int newPackets = 0;
    int didHardAck = 0;
    int haveLast = 0;
    afs_uint32 seq;
    afs_uint32 serial=0, flags=0;
    int isFirst;
    struct rx_packet *tnp;
    struct clock when, now;
    if (rx_stats_active)
        rx_atomic_inc(&rx_stats.dataPacketsRead);

#ifdef KERNEL
    /* If there are no packet buffers, drop this new packet, unless we can find
     * packet buffers from inactive calls */
    if (!call->error
        && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
        MUTEX_ENTER(&rx_freePktQ_lock);
        rxi_NeedMorePackets = TRUE;
        MUTEX_EXIT(&rx_freePktQ_lock);
        if (rx_stats_active)
            rx_atomic_inc(&rx_stats.noPacketBuffersOnRead);
        call->rprev = np->header.serial;
        rxi_calltrace(RX_TRACE_DROP, call);
        dpf(("packet %"AFS_PTR_FMT" dropped on receipt - quota problems\n", np));
        if (rxi_doreclaim)
            rxi_ClearReceiveQueue(call);
        clock_GetTime(&now);
        when = now;
        clock_Add(&when, &rx_softAckDelay);
        if (!call->delayedAckEvent
            || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
            rxevent_Cancel(call->delayedAckEvent, call,
                           RX_CALL_REFCOUNT_DELAY);
            MUTEX_ENTER(&rx_refcnt_mutex);
            CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
            MUTEX_EXIT(&rx_refcnt_mutex);

            call->delayedAckEvent =
                rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
        }
        /* we've damaged this call already, might as well do it in. */
        return np;
    }
#endif /* KERNEL */

    /*
     * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
     * packet is one of several packets transmitted as a single
     * datagram. Do not send any soft or hard acks until all packets
     * in a jumbogram have been processed. Send negative acks right away.
     */
    for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
        /* tnp is non-null when there are more packets in the
         * current jumbo gram */
        if (tnp) {
            if (np)
                rxi_FreePacket(np);
            np = tnp;
        }

        seq = np->header.seq;
        serial = np->header.serial;
        flags = np->header.flags;

        /* If the call is in an error state, send an abort message */
        if (call->error)
            return rxi_SendCallAbort(call, np, istack, 0);

        /* The RX_JUMBO_PACKET is set in all but the last packet in each
         * AFS 3.5 jumbogram. */
        if (flags & RX_JUMBO_PACKET) {
            tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
        } else {
            tnp = NULL;
        }

        if (np->header.spare != 0) {
            MUTEX_ENTER(&call->conn->conn_data_lock);
            call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
            MUTEX_EXIT(&call->conn->conn_data_lock);
        }

        /* The usual case is that this is the expected next packet */
        if (seq == call->rnext) {

            /* Check to make sure it is not a duplicate of one already queued */
            if (queue_IsNotEmpty(&call->rq)
                && queue_First(&call->rq, rx_packet)->header.seq == seq) {
                if (rx_stats_active)
                    rx_atomic_inc(&rx_stats.dupPacketsRead);
                dpf(("packet %"AFS_PTR_FMT" dropped on receipt - duplicate\n", np));
                rxevent_Cancel(call->delayedAckEvent, call,
                               RX_CALL_REFCOUNT_DELAY);
                np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
                ackNeeded = 0;
                call->rprev = seq;
                continue;
            }

            /* It's the next packet. Stick it on the receive queue
             * for this call. Set newPackets to make sure we wake
             * the reader once all packets have been processed */
#ifdef RX_TRACK_PACKETS
            np->flags |= RX_PKTFLAG_RQ;
#endif
            queue_Prepend(&call->rq, np);
#ifdef RXDEBUG_PACKET
            call->rqc++;
#endif /* RXDEBUG_PACKET */
            call->nSoftAcks++;
            np = NULL;          /* We can't use this anymore */
            newPackets = 1;

            /* If an ack is requested then set a flag to make sure we
             * send an acknowledgement for this packet */
            if (flags & RX_REQUEST_ACK) {
                ackNeeded = RX_ACK_REQUESTED;
            }

            /* Keep track of whether we have received the last packet */
            if (flags & RX_LAST_PACKET) {
                call->flags |= RX_CALL_HAVE_LAST;
                haveLast = 1;
            }

            /* Check whether we have all of the packets for this call */
            if (call->flags & RX_CALL_HAVE_LAST) {
                afs_uint32 tseq;        /* temporary sequence number */
                struct rx_packet *tp;   /* Temporary packet pointer */
                struct rx_packet *nxp;  /* Next pointer, for queue_Scan */

                for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
                    if (tseq != tp->header.seq)
                        break;
                    if (tp->header.flags & RX_LAST_PACKET) {
                        call->flags |= RX_CALL_RECEIVE_DONE;
                        break;
                    }
                    tseq++;
                }
            }

            /* Provide asynchronous notification for those who want it
             * (e.g. multi rx) */
            if (call->arrivalProc) {
                (*call->arrivalProc) (call, call->arrivalProcHandle,
                                      call->arrivalProcArg);
                call->arrivalProc = (void (*)())0;
            }

            /* Update last packet received */
            call->rprev = seq;

            /* If there is no server process serving this call, grab
             * one, if available. We only need to do this once. If a
             * server thread is available, this thread becomes a server
             * thread and the server thread becomes a listener thread. */
            if (isFirst) {
                TryAttach(call, socket, tnop, newcallp, 0);
            }
        }
        /* This is not the expected next packet. */
        else {
            /* Determine whether this is a new or old packet, and if it's
             * a new one, whether it fits into the current receive window.
             * Also figure out whether the packet was delivered in sequence.
             * We use the prev variable to determine whether the new packet
             * is the successor of its immediate predecessor in the
             * receive queue, and the missing flag to determine whether
             * any of this packets predecessors are missing.  */

            afs_uint32 prev;    /* "Previous packet" sequence number */
            struct rx_packet *tp;       /* Temporary packet pointer */
            struct rx_packet *nxp;      /* Next pointer, for queue_Scan */
            int missing;        /* Are any predecessors missing? */

            /* If the new packet's sequence number has been sent to the
             * application already, then this is a duplicate */
            if (seq < call->rnext) {
                if (rx_stats_active)
                    rx_atomic_inc(&rx_stats.dupPacketsRead);
                rxevent_Cancel(call->delayedAckEvent, call,
                               RX_CALL_REFCOUNT_DELAY);
                np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
                ackNeeded = 0;
                call->rprev = seq;
                continue;
            }

            /* If the sequence number is greater than what can be
             * accomodated by the current window, then send a negative
             * acknowledge and drop the packet */
            if ((call->rnext + call->rwind) <= seq) {
                rxevent_Cancel(call->delayedAckEvent, call,
                               RX_CALL_REFCOUNT_DELAY);
                np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
                                 istack);
                ackNeeded = 0;
                call->rprev = seq;
                continue;
            }

            /* Look for the packet in the queue of old received packets */
            for (prev = call->rnext - 1, missing =
                 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
                /*Check for duplicate packet */
                if (seq == tp->header.seq) {
                    if (rx_stats_active)
                        rx_atomic_inc(&rx_stats.dupPacketsRead);
                    rxevent_Cancel(call->delayedAckEvent, call,
                                   RX_CALL_REFCOUNT_DELAY);
                    np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
                                     istack);
                    ackNeeded = 0;
                    call->rprev = seq;
                    goto nextloop;
                }
                /* If we find a higher sequence packet, break out and
                 * insert the new packet here. */
                if (seq < tp->header.seq)
                    break;
                /* Check for missing packet */
                if (tp->header.seq != prev + 1) {
                    missing = 1;
                }

                prev = tp->header.seq;
            }

            /* Keep track of whether we have received the last packet. */
            if (flags & RX_LAST_PACKET) {
                call->flags |= RX_CALL_HAVE_LAST;
            }

            /* It's within the window: add it to the the receive queue.
             * tp is left by the previous loop either pointing at the
             * packet before which to insert the new packet, or at the
             * queue head if the queue is empty or the packet should be
             * appended. */
#ifdef RX_TRACK_PACKETS
            np->flags |= RX_PKTFLAG_RQ;
#endif
#ifdef RXDEBUG_PACKET
            call->rqc++;
#endif /* RXDEBUG_PACKET */
            queue_InsertBefore(tp, np);
            call->nSoftAcks++;
            np = NULL;

            /* Check whether we have all of the packets for this call */
            if ((call->flags & RX_CALL_HAVE_LAST)
                && !(call->flags & RX_CALL_RECEIVE_DONE)) {
                afs_uint32 tseq;        /* temporary sequence number */

                for (tseq =
                     call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
                    if (tseq != tp->header.seq)
                        break;
                    if (tp->header.flags & RX_LAST_PACKET) {
                        call->flags |= RX_CALL_RECEIVE_DONE;
                        break;
                    }
                    tseq++;
                }
            }

            /* We need to send an ack of the packet is out of sequence,
             * or if an ack was requested by the peer. */
            if (seq != prev + 1 || missing) {
                ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
            } else if (flags & RX_REQUEST_ACK) {
                ackNeeded = RX_ACK_REQUESTED;
            }

            /* Acknowledge the last packet for each call */
            if (flags & RX_LAST_PACKET) {
                haveLast = 1;
            }

            call->rprev = seq;
        }
      nextloop:;
    }

    if (newPackets) {
        /*
         * If the receiver is waiting for an iovec, fill the iovec
         * using the data from the receive queue */
        if (call->flags & RX_CALL_IOVEC_WAIT) {
            didHardAck = rxi_FillReadVec(call, serial);
            /* the call may have been aborted */
            if (call->error) {
                return NULL;
            }
            if (didHardAck) {
                ackNeeded = 0;
            }
        }

        /* Wakeup the reader if any */
        if ((call->flags & RX_CALL_READER_WAIT)
            && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
                || (call->iovNext >= call->iovMax)
                || (call->flags & RX_CALL_RECEIVE_DONE))) {
            call->flags &= ~RX_CALL_READER_WAIT;
#ifdef  RX_ENABLE_LOCKS
            CV_BROADCAST(&call->cv_rq);
#else
            osi_rxWakeup(&call->rq);
#endif
        }
    }

    /*
     * Send an ack when requested by the peer, or once every
     * rxi_SoftAckRate packets until the last packet has been
     * received. Always send a soft ack for the last packet in
     * the server's reply.
     *
     * If we have received all of the packets for the call
     * immediately send an RX_PACKET_TYPE_ACKALL packet so that
     * the peer can empty its packet queue and cancel all resend
     * events.
     */
    if (call->flags & RX_CALL_RECEIVE_DONE) {
        rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
        rxi_AckAll(NULL, call, 0);
    } else if (ackNeeded) {
        rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
        np = rxi_SendAck(call, np, serial, ackNeeded, istack);
    } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
        rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
        np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
    } else if (call->nSoftAcks) {
        clock_GetTime(&now);
        when = now;
        if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
            clock_Add(&when, &rx_lastAckDelay);
        } else {
            clock_Add(&when, &rx_softAckDelay);
        }
        if (!call->delayedAckEvent
            || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
            rxevent_Cancel(call->delayedAckEvent, call,
                           RX_CALL_REFCOUNT_DELAY);
            MUTEX_ENTER(&rx_refcnt_mutex);
            CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
            MUTEX_EXIT(&rx_refcnt_mutex);
            call->delayedAckEvent =
                rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
        }
    }

    return np;
}

#ifdef  ADAPT_WINDOW
static void rxi_ComputeRate();
#endif

static void
rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
{
    struct rx_peer *peer = conn->peer;

    MUTEX_ENTER(&peer->peer_lock);
    peer->lastReachTime = clock_Sec();
    MUTEX_EXIT(&peer->peer_lock);

    MUTEX_ENTER(&conn->conn_data_lock);
    if (conn->flags & RX_CONN_ATTACHWAIT) {
        int i;

        conn->flags &= ~RX_CONN_ATTACHWAIT;
        MUTEX_EXIT(&conn->conn_data_lock);

        for (i = 0; i < RX_MAXCALLS; i++) {
            struct rx_call *call = conn->call[i];
            if (call) {
                if (call != acall)
                    MUTEX_ENTER(&call->lock);
                /* tnop can be null if newcallp is null */
                TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
                if (call != acall)
                    MUTEX_EXIT(&call->lock);
            }
        }
    } else
        MUTEX_EXIT(&conn->conn_data_lock);
}

#if defined(RXDEBUG) && defined(AFS_NT40_ENV)
static const char *
rx_ack_reason(int reason)
{
    switch (reason) {
    case RX_ACK_REQUESTED:
        return "requested";
    case RX_ACK_DUPLICATE:
        return "duplicate";
    case RX_ACK_OUT_OF_SEQUENCE:
        return "sequence";
    case RX_ACK_EXCEEDS_WINDOW:
        return "window";
    case RX_ACK_NOSPACE:
        return "nospace";
    case RX_ACK_PING:
        return "ping";
    case RX_ACK_PING_RESPONSE:
        return "response";
    case RX_ACK_DELAY:
        return "delay";
    case RX_ACK_IDLE:
        return "idle";
    default:
        return "unknown!!";
    }
}
#endif


/* The real smarts of the whole thing.  */
struct rx_packet *
rxi_ReceiveAckPacket(struct rx_call *call, struct rx_packet *np,
                     int istack)
{
    struct rx_ackPacket *ap;
    int nAcks;
    struct rx_packet *tp;
    struct rx_packet *nxp;      /* Next packet pointer for queue_Scan */
    struct rx_connection *conn = call->conn;
    struct rx_peer *peer = conn->peer;
    struct clock now;           /* Current time, for RTT calculations */
    afs_uint32 first;
    afs_uint32 prev;
    afs_uint32 serial;
    /* because there are CM's that are bogus, sending weird values for this. */
    afs_uint32 skew = 0;
    int nbytes;
    int missing;
    int acked;
    int nNacked = 0;
    int newAckCount = 0;
    int maxDgramPackets = 0;    /* Set if peer supports AFS 3.5 jumbo datagrams */
    int pktsize = 0;            /* Set if we need to update the peer mtu */
    int conn_data_locked = 0;

    if (rx_stats_active)
        rx_atomic_inc(&rx_stats.ackPacketsRead);
    ap = (struct rx_ackPacket *)rx_DataOf(np);
    nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
    if (nbytes < 0)
        return np;              /* truncated ack packet */

    /* depends on ack packet struct */
    nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
    first = ntohl(ap->firstPacket);
    prev = ntohl(ap->previousPacket);
    serial = ntohl(ap->serial);
    /* temporarily disabled -- needs to degrade over time
     * skew = ntohs(ap->maxSkew); */

    /* Ignore ack packets received out of order */
    if (first < call->tfirst ||
        (first == call->tfirst && prev < call->tprev)) {
        return np;
    }

    call->tprev = prev;

    if (np->header.flags & RX_SLOW_START_OK) {
        call->flags |= RX_CALL_SLOW_START_OK;
    }

    if (ap->reason == RX_ACK_PING_RESPONSE)
        rxi_UpdatePeerReach(conn, call);

    if (conn->lastPacketSizeSeq) {
        MUTEX_ENTER(&conn->conn_data_lock);
        conn_data_locked = 1;
        if ((first > conn->lastPacketSizeSeq) && (conn->lastPacketSize)) {
            pktsize = conn->lastPacketSize;
            conn->lastPacketSize = conn->lastPacketSizeSeq = 0;
        }
    }
    if ((ap->reason == RX_ACK_PING_RESPONSE) && (conn->lastPingSizeSer)) {
        if (!conn_data_locked) {
            MUTEX_ENTER(&conn->conn_data_lock);
            conn_data_locked = 1;
        }
        if ((conn->lastPingSizeSer == serial) && (conn->lastPingSize)) {
            /* process mtu ping ack */
            pktsize = conn->lastPingSize;
            conn->lastPingSizeSer = conn->lastPingSize = 0;
        }
    }

    if (conn_data_locked) {
        MUTEX_EXIT(&conn->conn_data_lock);
        conn_data_locked = 0;
    }
#ifdef RXDEBUG
#ifdef AFS_NT40_ENV
    if (rxdebug_active) {
        char msg[512];
        size_t len;

        len = _snprintf(msg, sizeof(msg),
                        "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
                         GetCurrentThreadId(), rx_ack_reason(ap->reason),
                         ntohl(ap->serial), ntohl(ap->previousPacket),
                         (unsigned int)np->header.seq, (unsigned int)skew,
                         ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
        if (nAcks) {
            int offset;

            for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
                msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
        }
        msg[len++]='\n';
        msg[len] = '\0';
        OutputDebugString(msg);
    }
#else /* AFS_NT40_ENV */
    if (rx_Log) {
        fprintf(rx_Log,
                "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
                ap->reason, ntohl(ap->previousPacket),
                (unsigned int)np->header.seq, (unsigned int)serial,
                (unsigned int)skew, ntohl(ap->firstPacket));
        if (nAcks) {
            int offset;
            for (offset = 0; offset < nAcks; offset++)
                putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
                     rx_Log);
        }
        putc('\n', rx_Log);
    }
#endif /* AFS_NT40_ENV */
#endif

    MUTEX_ENTER(&peer->peer_lock);
    if (pktsize) {
        /*
         * Start somewhere. Can't assume we can send what we can receive,
         * but we are clearly receiving.
         */
        if (!peer->maxPacketSize)
            peer->maxPacketSize = RX_MIN_PACKET_SIZE+RX_IPUDP_SIZE;

        if (pktsize > peer->maxPacketSize) {
            peer->maxPacketSize = pktsize;
            if ((pktsize-RX_IPUDP_SIZE > peer->ifMTU)) {
                peer->ifMTU=pktsize-RX_IPUDP_SIZE;
                peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
                rxi_ScheduleGrowMTUEvent(call, 1);
            }
        }
    }

    /* Update the outgoing packet skew value to the latest value of
     * the peer's incoming packet skew value.  The ack packet, of
     * course, could arrive out of order, but that won't affect things
     * much */
    peer->outPacketSkew = skew;

    /* Check for packets that no longer need to be transmitted, and
     * discard them.  This only applies to packets positively
     * acknowledged as having been sent to the peer's upper level.
     * All other packets must be retained.  So only packets with
     * sequence numbers < ap->firstPacket are candidates. */

    clock_GetTime(&now);

    for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
        if (tp->header.seq >= first)
            break;
        call->tfirst = tp->header.seq + 1;

        if (!(tp->flags & RX_PKTFLAG_ACKED)) {
            newAckCount++;

            rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
        }

#ifdef ADAPT_WINDOW
        rxi_ComputeRate(call->conn->peer, call, p, np, ap->reason);
#endif

#ifdef  AFS_GLOBAL_RXLOCK_KERNEL
        /* XXX Hack. Because we have to release the global rx lock when sending
         * packets (osi_NetSend) we drop all acks while we're traversing the tq
         * in rxi_Start sending packets out because packets may move to the
         * freePacketQueue as result of being here! So we drop these packets until
         * we're safely out of the traversing. Really ugly!
         * To make it even uglier, if we're using fine grain locking, we can
         * set the ack bits in the packets and have rxi_Start remove the packets
         * when it's done transmitting.
         */
        if (call->flags & RX_CALL_TQ_BUSY) {
#ifdef RX_ENABLE_LOCKS
            tp->flags |= RX_PKTFLAG_ACKED;
            call->flags |= RX_CALL_TQ_SOME_ACKED;
#else /* RX_ENABLE_LOCKS */
            break;
#endif /* RX_ENABLE_LOCKS */
        } else
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
        {
            queue_Remove(tp);
#ifdef RX_TRACK_PACKETS
            tp->flags &= ~RX_PKTFLAG_TQ;
#endif
#ifdef RXDEBUG_PACKET
            call->tqc--;
#endif /* RXDEBUG_PACKET */
            rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
        }
    }

#ifdef ADAPT_WINDOW
    /* Give rate detector a chance to respond to ping requests */
    if (ap->reason == RX_ACK_PING_RESPONSE) {
        rxi_ComputeRate(peer, call, 0, np, ap->reason);
    }
#endif

    /* N.B. we don't turn off any timers here.  They'll go away by themselves, anyway */

    /* Now go through explicit acks/nacks and record the results in
     * the waiting packets.  These are packets that can't be released
     * yet, even with a positive acknowledge.  This positive
     * acknowledge only means the packet has been received by the
     * peer, not that it will be retained long enough to be sent to
     * the peer's upper level.  In addition, reset the transmit timers
     * of any missing packets (those packets that must be missing
     * because this packet was out of sequence) */

    call->nSoftAcked = 0;
    for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {

        /* Set the acknowledge flag per packet based on the
         * information in the ack packet. An acknowlegded packet can
         * be downgraded when the server has discarded a packet it
         * soacked previously, or when an ack packet is received
         * out of sequence. */
        if (tp->header.seq < first) {
            /* Implicit ack information */
            if (!(tp->flags & RX_PKTFLAG_ACKED)) {
                newAckCount++;
            }
            tp->flags |= RX_PKTFLAG_ACKED;
        } else if (tp->header.seq < first + nAcks) {
            /* Explicit ack information:  set it in the packet appropriately */
            if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
                if (!(tp->flags & RX_PKTFLAG_ACKED)) {
                    newAckCount++;
                    tp->flags |= RX_PKTFLAG_ACKED;

                    rxi_ComputeRoundTripTime(tp, ap, call->conn->peer, &now);
#ifdef ADAPT_WINDOW
                    rxi_ComputeRate(call->conn->peer, call, tp, np,
                                    ap->reason);
#endif
                }
                if (missing) {
                    nNacked++;
                } else {
                    call->nSoftAcked++;
                }
            } else /* RX_ACK_TYPE_NACK */ {
                tp->flags &= ~RX_PKTFLAG_ACKED;
                missing = 1;
            }
        } else {
            if (tp->flags & RX_PKTFLAG_ACKED) {
                tp->flags &= ~RX_PKTFLAG_ACKED;
                missing = 1;
            }
        }

        /*
         * Following the suggestion of Phil Kern, we back off the peer's
         * timeout value for future packets until a successful response
         * is received for an initial transmission.
         */
        if (missing && !peer->backedOff) {
            struct clock c = peer->timeout;
            struct clock max_to = {3, 0};

            clock_Add(&peer->timeout, &c);
            if (clock_Gt(&peer->timeout, &max_to))
                peer->timeout = max_to;
            peer->backedOff = 1;
        }

        /* If packet isn't yet acked, and it has been transmitted at least
         * once, reset retransmit time using latest timeout
         * ie, this should readjust the retransmit timer for all outstanding
         * packets...  So we don't just retransmit when we should know better*/

        if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
            tp->retryTime = tp->timeSent;
            clock_Add(&tp->retryTime, &peer->timeout);
            /* shift by eight because one quarter-sec ~ 256 milliseconds */
            clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
        }
    }

    /* If the window has been extended by this acknowledge packet,
     * then wakeup a sender waiting in alloc for window space, or try
     * sending packets now, if he's been sitting on packets due to
     * lack of window space */
    if (call->tnext < (call->tfirst + call->twind)) {
#ifdef  RX_ENABLE_LOCKS
        CV_SIGNAL(&call->cv_twind);
#else
        if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
            call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
            osi_rxWakeup(&call->twind);
        }
#endif
        if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
            call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
        }
    }

    /* if the ack packet has a receivelen field hanging off it,
     * update our state */
    if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
        afs_uint32 tSize;

        /* If the ack packet has a "recommended" size that is less than
         * what I am using now, reduce my size to match */
        rx_packetread(np, rx_AckDataSize(ap->nAcks) + (in