afs: more cache truncation stats

[openafs.git] / src / afs / afs_dcache.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
 */

/*
 * Implements:
 */
#include <afsconfig.h>
#include "afs/param.h"


#include "afs/sysincludes.h"    /*Standard vendor system headers */
#include "afsincludes.h"        /*AFS-based standard headers */
#include "afs/afs_stats.h"      /* statistics */
#include "afs/afs_cbqueue.h"
#include "afs/afs_osidnlc.h"

#include <opr/ffs.h>

/* Forward declarations. */
static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
static int afs_FreeDiscardedDCache(void);
static void afs_DiscardDCache(struct dcache *);
static void afs_FreeDCache(struct dcache *);
/* For split cache */
static afs_int32 afs_DCGetBucket(struct vcache *);
static void afs_DCAdjustSize(struct dcache *, afs_int32, afs_int32);
static void afs_DCMoveBucket(struct dcache *, afs_int32, afs_int32);
static void afs_DCSizeInit(void);
static afs_int32 afs_DCWhichBucket(afs_int32, afs_int32);

/*
 * --------------------- Exported definitions ---------------------
 */
/* For split cache */
afs_int32 afs_blocksUsed_0;    /*!< 1K blocks in cache - in theory is zero */
afs_int32 afs_blocksUsed_1;    /*!< 1K blocks in cache */
afs_int32 afs_blocksUsed_2;    /*!< 1K blocks in cache */
afs_int32 afs_pct1 = -1;
afs_int32 afs_pct2 = -1;
afs_uint32 afs_tpct1 = 0;
afs_uint32 afs_tpct2 = 0;
afs_uint32 splitdcache = 0;

afs_lock_t afs_xdcache;         /*!< Lock: alloc new disk cache entries */
afs_int32 afs_freeDCList;       /*!< Free list for disk cache entries */
afs_int32 afs_freeDCCount;      /*!< Count of elts in freeDCList */
afs_int32 afs_discardDCList;    /*!< Discarded disk cache entries */
afs_int32 afs_discardDCCount;   /*!< Count of elts in discardDCList */
struct dcache *afs_freeDSList;  /*!< Free list for disk slots */
struct dcache *afs_Initial_freeDSList;  /*!< Initial list for above */
afs_dcache_id_t cacheInode;               /*!< Inode for CacheItems file */
struct osi_file *afs_cacheInodep = 0;   /*!< file for CacheItems inode */
struct afs_q afs_DLRU;          /*!< dcache LRU */
afs_int32 afs_dhashsize = 1024;
afs_int32 *afs_dvhashTbl;       /*!< Data cache hash table: hashed by FID + chunk number. */
afs_int32 *afs_dchashTbl;       /*!< Data cache hash table: hashed by FID. */
afs_int32 *afs_dvnextTbl;       /*!< Dcache hash table links */
afs_int32 *afs_dcnextTbl;       /*!< Dcache hash table links */
struct dcache **afs_indexTable; /*!< Pointers to dcache entries */
afs_hyper_t *afs_indexTimes;    /*!< Dcache entry Access times */
afs_int32 *afs_indexUnique;     /*!< dcache entry Fid.Unique */
unsigned char *afs_indexFlags;  /*!< (only one) Is there data there? */
afs_hyper_t afs_indexCounter;   /*!< Fake time for marking index
                                 * entries */
afs_int32 afs_cacheFiles = 0;   /*!< Size of afs_indexTable */
afs_int32 afs_cacheBlocks;      /*!< 1K blocks in cache */
afs_int32 afs_cacheStats;       /*!< Stat entries in cache */
afs_int32 afs_blocksUsed;       /*!< Number of blocks in use */
afs_int32 afs_blocksDiscarded;  /*!<Blocks freed but not truncated */
afs_int32 afs_fsfragsize = AFS_MIN_FRAGSIZE;    /*!< Underlying Filesystem minimum unit
                                         *of disk allocation usually 1K
                                         *this value is (truefrag -1 ) to
                                         *save a bunch of subtracts... */
#ifdef AFS_64BIT_CLIENT
#ifdef AFS_VM_RDWR_ENV
afs_size_t afs_vmMappingEnd;    /* !< For large files (>= 2GB) the VM
                                 * mapping an 32bit addressing machines
                                 * can only be used below the 2 GB
                                 * line. From this point upwards we
                                 * must do direct I/O into the cache
                                 * files. The value should be on a
                                 * chunk boundary. */
#endif /* AFS_VM_RDWR_ENV */
#endif /* AFS_64BIT_CLIENT */

/* The following is used to ensure that new dcache's aren't obtained when
 * the cache is nearly full.
 */
int afs_WaitForCacheDrain = 0;
int afs_TruncateDaemonRunning = 0;
int afs_CacheTooFull = 0;
afs_uint32 afs_CacheTooFullCount = 0;
afs_uint32 afs_WaitForCacheDrainCount = 0;

afs_int32 afs_dcentries;        /*!< In-memory dcache entries */


int dcacheDisabled = 0;

struct afs_cacheOps afs_UfsCacheOps = {
#ifndef HAVE_STRUCT_LABEL_SUPPORT
    osi_UFSOpen,
    osi_UFSTruncate,
    afs_osi_Read,
    afs_osi_Write,
    osi_UFSClose,
    afs_UFSReadUIO,
    afs_UFSWriteUIO,
    afs_UFSGetDSlot,
    afs_UFSGetVolSlot,
    afs_UFSHandleLink,
#else
    .open       = osi_UFSOpen,
    .truncate   = osi_UFSTruncate,
    .fread      = afs_osi_Read,
    .fwrite     = afs_osi_Write,
    .close      = osi_UFSClose,
    .vreadUIO   = afs_UFSReadUIO,
    .vwriteUIO  = afs_UFSWriteUIO,
    .GetDSlot   = afs_UFSGetDSlot,
    .GetVolSlot = afs_UFSGetVolSlot,
    .HandleLink = afs_UFSHandleLink,
#endif
};

struct afs_cacheOps afs_MemCacheOps = {
#ifndef HAVE_STRUCT_LABEL_SUPPORT
    afs_MemCacheOpen,
    afs_MemCacheTruncate,
    afs_MemReadBlk,
    afs_MemWriteBlk,
    afs_MemCacheClose,
    afs_MemReadUIO,
    afs_MemWriteUIO,
    afs_MemGetDSlot,
    afs_MemGetVolSlot,
    afs_MemHandleLink,
#else
    .open       = afs_MemCacheOpen,
    .truncate   = afs_MemCacheTruncate,
    .fread      = afs_MemReadBlk,
    .fwrite     = afs_MemWriteBlk,
    .close      = afs_MemCacheClose,
    .vreadUIO   = afs_MemReadUIO,
    .vwriteUIO  = afs_MemWriteUIO,
    .GetDSlot   = afs_MemGetDSlot,
    .GetVolSlot = afs_MemGetVolSlot,
    .HandleLink = afs_MemHandleLink,
#endif
};

int cacheDiskType;              /*Type of backing disk for cache */
struct afs_cacheOps *afs_cacheType;


/*
 * The PFlush algorithm makes use of the fact that Fid.Unique is not used in
 * below hash algorithms.  Change it if need be so that flushing algorithm
 * doesn't move things from one hash chain to another.
 */
/*Vnode, Chunk -> Hash table index */
int DCHash(struct VenusFid *fid, afs_int32 chunk)
{
    afs_uint32 buf[3];

    buf[0] = fid->Fid.Volume;
    buf[1] = fid->Fid.Vnode;
    buf[2] = chunk;
    return opr_jhash(buf, 3, 0) & (afs_dhashsize - 1);
}
/*Vnode -> Other hash table index */
int DVHash(struct VenusFid *fid)
{
    return opr_jhash_int2(fid->Fid.Volume, fid->Fid.Vnode, 0) &
        (afs_dhashsize - 1);
}

/*!
 * Where is this vcache's entry associated dcache located/
 * \param avc The vcache entry.
 * \return Bucket index:
 *      1 : main
 *      2 : RO
 */
static afs_int32
afs_DCGetBucket(struct vcache *avc)
{
    if (!splitdcache)
        return 1;

    /* This should be replaced with some sort of user configurable function */
    if (avc->f.states & CRO) {
        return 2;
    } else if (avc->f.states & CBackup) {
        return 1;
    } else {
        /* RW */
    }
    /* main bucket */
    return 1;
}

/*!
 * Readjust a dcache's size.
 *
 * \param adc The dcache to be adjusted.
 * \param oldSize Old size for the dcache.
 * \param newSize The new size to be adjusted to.
 *
 */
static void
afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
{
    afs_int32 adjustSize = newSize - oldSize;

    if (!splitdcache)
        return;

    switch (adc->bucket)
    {
    case 0:
        afs_blocksUsed_0 += adjustSize;
        afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
        break;
    case 1:
        afs_blocksUsed_1 += adjustSize;
        afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
        break;
    case 2:
        afs_blocksUsed_2 += adjustSize;
        afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
        break;
    }

    return;
}

/*!
 * Move a dcache from one bucket to another.
 *
 * \param adc Operate on this dcache.
 * \param size Size in bucket (?).
 * \param newBucket Destination bucket.
 *
 */
static void
afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
{
    if (!splitdcache)
        return;

    /* Substract size from old bucket. */
    switch (adc->bucket)
    {
    case 0:
        afs_blocksUsed_0 -= size;
        break;
    case 1:
        afs_blocksUsed_1 -= size;
        break;
    case 2:
        afs_blocksUsed_2 -= size;
        break;
    }

    /* Set new bucket and increase destination bucket size. */
    adc->bucket = newBucket;

    switch (adc->bucket)
    {
    case 0:
        afs_blocksUsed_0 += size;
        break;
    case 1:
        afs_blocksUsed_1 += size;
        break;
    case 2:
        afs_blocksUsed_2 += size;
        break;
    }

    return;
}

/*!
 * Init split caches size.
 */
static void
afs_DCSizeInit(void)
{
    afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
}


/*!
 * \param phase
 * \param bucket
 */
static afs_int32
afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
{
    if (!splitdcache)
        return 0;

    afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
    afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);

    /* Short cut: if we don't know about it, try to kill it */
    if (phase < 2 && afs_blocksUsed_0)
        return 0;

    if (afs_pct1 > afs_tpct1)
        return 1;
    if (afs_pct2 > afs_tpct2)
        return 2;
    return 0; /* unlikely */
}


/*!
 * Warn about failing to store a file.
 *
 * \param acode Associated error code.
 * \param avolume Volume involved.
 * \param aflags How to handle the output:
 *      aflags & 1: Print out on console
 *      aflags & 2: Print out on controlling tty
 *
 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
 */

void
afs_StoreWarn(afs_int32 acode, afs_int32 avolume,
              afs_int32 aflags)
{
    static char problem_fmt[] =
        "afs: failed to store file in volume %d (%s)\n";
    static char problem_fmt_w_error[] =
        "afs: failed to store file in volume %d (error %d)\n";
    static char netproblems[] = "network problems";
    static char partfull[] = "partition full";
    static char overquota[] = "over quota";

    AFS_STATCNT(afs_StoreWarn);
    if (acode < 0) {
        /*
         * Network problems
         */
        if (aflags & 1)
            afs_warn(problem_fmt, avolume, netproblems);
        if (aflags & 2)
            afs_warnuser(problem_fmt, avolume, netproblems);
    } else if (acode == ENOSPC) {
        /*
         * Partition full
         */
        if (aflags & 1)
            afs_warn(problem_fmt, avolume, partfull);
        if (aflags & 2)
            afs_warnuser(problem_fmt, avolume, partfull);
    } else
#ifdef  EDQUOT
        /* EDQUOT doesn't exist on solaris and won't be sent by the server.
         * Instead ENOSPC will be sent...
         */
    if (acode == EDQUOT) {
        /*
         * Quota exceeded
         */
        if (aflags & 1)
            afs_warn(problem_fmt, avolume, overquota);
        if (aflags & 2)
            afs_warnuser(problem_fmt, avolume, overquota);
    } else
#endif
    {
        /*
         * Unknown error
         */
        if (aflags & 1)
            afs_warn(problem_fmt_w_error, avolume, acode);
        if (aflags & 2)
            afs_warnuser(problem_fmt_w_error, avolume, acode);
    }
}                               /*afs_StoreWarn */

/*!
 * Try waking up truncation daemon, if it's worth it.
 */
void
afs_MaybeWakeupTruncateDaemon(void)
{
    if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
        afs_CacheTooFullCount++;
        afs_CacheTooFull = 1;
        if (!afs_TruncateDaemonRunning)
            afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
    } else if (!afs_TruncateDaemonRunning
               && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
        afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
    }
}

/*!
 * /struct CTD_stats
 *
 * Keep statistics on run time for afs_CacheTruncateDaemon. This is a
 * struct so we need only export one symbol for AIX.
 */
static struct CTD_stats {
    osi_timeval32_t CTD_beforeSleep;
    osi_timeval32_t CTD_afterSleep;
    osi_timeval32_t CTD_sleepTime;
    osi_timeval32_t CTD_runTime;
    int CTD_nSleeps;
} CTD_stats;

u_int afs_min_cache = 0;

/*!
 * If there are waiters for the cache to drain, wake them if
 * the number of free or discarded cache blocks reaches the
 * CM_CACHESIZEDDRAINEDPCT limit.
 *
 * \note Environment:
 *      This routine must be called with the afs_xdcache lock held
 *      (in write mode).
 */
static void
afs_WakeCacheWaitersIfDrained(void)
{
    if (afs_WaitForCacheDrain) {
        if ((afs_blocksUsed - afs_blocksDiscarded) <=
            PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
            afs_WaitForCacheDrain = 0;
            afs_osi_Wakeup(&afs_WaitForCacheDrain);
        }
    }
}

/*!
 * Keeps the cache clean and free by truncating uneeded files, when used.
 * \param
 * \return
 */
void
afs_CacheTruncateDaemon(void)
{
    osi_timeval32_t CTD_tmpTime;
    u_int counter;
    u_int cb_lowat;
    u_int dc_hiwat =
        PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
    afs_min_cache =
        (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;

    osi_GetTime(&CTD_stats.CTD_afterSleep);
    afs_TruncateDaemonRunning = 1;
    while (1) {
        cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
        ObtainWriteLock(&afs_xdcache, 266);
        if (afs_CacheTooFull || afs_WaitForCacheDrain) {
            int space_needed, slots_needed;
            /* if we get woken up, we should try to clean something out */
            for (counter = 0; counter < 10; counter++) {
                space_needed =
                    afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
                if (space_needed < 0)
                    space_needed = 0;
                slots_needed =
                    dc_hiwat - afs_freeDCCount - afs_discardDCCount;
                if (slots_needed < 0)
                    slots_needed = 0;
                if (slots_needed || space_needed)
                    afs_GetDownD(slots_needed, &space_needed, 0);
                if ((space_needed <= 0) && (slots_needed <= 0)) {
                    break;
                }
                if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
                    break;
            }
            if (!afs_CacheIsTooFull()) {
                afs_CacheTooFull = 0;
                afs_WakeCacheWaitersIfDrained();
            }
        }       /* end of cache cleanup */
        ReleaseWriteLock(&afs_xdcache);

        /*
         * This is a defensive check to try to avoid starving threads
         * that may need the global lock so thay can help free some
         * cache space. If this thread won't be sleeping or truncating
         * any cache files then give up the global lock so other
         * threads get a chance to run.
         */
        if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
            && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
            afs_osi_Wait(100, 0, 0);    /* 100 milliseconds */
        }

        /*
         * This is where we free the discarded cache elements.
         */
        while (afs_blocksDiscarded && !afs_WaitForCacheDrain
               && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
            int code = afs_FreeDiscardedDCache();
            if (code) {
                /* If we can't free any discarded dcache entries, that's okay.
                 * We're just doing this in the background; if someone needs
                 * discarded entries freed, they will try it themselves and/or
                 * signal us that the cache is too full. In any case, we'll
                 * try doing this again the next time we run through the loop.
                 */
                break;
            }
        }

        /* See if we need to continue to run. Someone may have
         * signalled us while we were executing.
         */
        if (!afs_WaitForCacheDrain && !afs_CacheTooFull
            && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
            /* Collect statistics on truncate daemon. */
            CTD_stats.CTD_nSleeps++;
            osi_GetTime(&CTD_stats.CTD_beforeSleep);
            afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
                              CTD_stats.CTD_beforeSleep);
            afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);

            afs_TruncateDaemonRunning = 0;
            afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
            afs_TruncateDaemonRunning = 1;

            osi_GetTime(&CTD_stats.CTD_afterSleep);
            afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
                              CTD_stats.CTD_afterSleep);
            afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
        }
        if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
            afs_termState = AFSOP_STOP_AFSDB;
            afs_osi_Wakeup(&afs_termState);
            break;
        }
    }
}


/*!
 * Make adjustment for the new size in the disk cache entry
 *
 * \note Major Assumptions Here:
 *      Assumes that frag size is an integral power of two, less one,
 *      and that this is a two's complement machine.  I don't
 *      know of any filesystems which violate this assumption...
 *
 * \param adc Ptr to dcache entry.
 * \param anewsize New size desired.
 *
 */

void
afs_AdjustSize(struct dcache *adc, afs_int32 newSize)
{
    afs_int32 oldSize;

    AFS_STATCNT(afs_AdjustSize);

    if (newSize > afs_OtherCSize && !(adc->f.fid.Fid.Vnode & 1)) {
        /* No non-dir cache files should be larger than the chunk size.
         * (Directory blobs are fetched in a single chunk file, so directories
         * can be larger.) If someone is requesting that a chunk is larger than
         * the chunk size, something strange is happening. Log a message about
         * it, to give a hint to subsequent strange behavior, if any occurs. */
        static int warned;
        if (!warned) {
            warned = 1;
            afs_warn("afs: Warning: dcache %d is very large (%d > %d). This "
                     "should not happen, but trying to continue regardless. If "
                     "AFS starts hanging or behaving strangely, this might be "
                     "why.\n",
                     adc->index, newSize, afs_OtherCSize);
        }
    }

    adc->dflags |= DFEntryMod;
    oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10;    /* round up */
    adc->f.chunkBytes = newSize;
    if (!newSize)
        adc->validPos = 0;
    newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10;      /* round up */
    afs_DCAdjustSize(adc, oldSize, newSize);
    if ((newSize > oldSize) && !AFS_IS_DISCONNECTED) {

        /* We're growing the file, wakeup the daemon */
        afs_MaybeWakeupTruncateDaemon();
    }
    afs_blocksUsed += (newSize - oldSize);
    afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
}


/*!
 * This routine is responsible for moving at least one entry (but up
 * to some number of them) from the LRU queue to the free queue.
 *
 * \param anumber Number of entries that should ideally be moved.
 * \param aneedSpace How much space we need (1K blocks);
 *
 * \note Environment:
 *      The anumber parameter is just a hint; at least one entry MUST be
 *      moved, or we'll panic.  We must be called with afs_xdcache
 *      write-locked.  We should try to satisfy both anumber and aneedspace,
 *      whichever is more demanding - need to do several things:
 *      1.  only grab up to anumber victims if aneedSpace <= 0, not
 *          the whole set of MAXATONCE.
 *      2.  dynamically choose MAXATONCE to reflect severity of
 *          demand: something like (*aneedSpace >> (logChunk - 9))
 *
 *  \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
 *  indicates that the cache is not properly configured/tuned or
 *  something. We should be able to automatically correct that problem.
 */

#define MAXATONCE   16          /* max we can obtain at once */
static void
afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
{

    struct dcache *tdc;
    struct VenusFid *afid;
    afs_int32 i, j;
    afs_hyper_t vtime;
    int skip, phase;
    struct vcache *tvc;
    afs_uint32 victims[MAXATONCE];
    struct dcache *victimDCs[MAXATONCE];
    afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
    afs_uint32 victimPtr;       /* next free item in victim arrays */
    afs_hyper_t maxVictimTime;  /* youngest (largest LRU time) victim */
    afs_uint32 maxVictimPtr;    /* where it is */
    int discard;
    int curbucket;

    AFS_STATCNT(afs_GetDownD);

    if (CheckLock(&afs_xdcache) != -1)
        osi_Panic("getdownd nolock");
    /* decrement anumber first for all dudes in free list */
    /* SHOULD always decrement anumber first, even if aneedSpace >0,
     * because we should try to free space even if anumber <=0 */
    if (!aneedSpace || *aneedSpace <= 0) {
        anumber -= afs_freeDCCount;
        if (anumber <= 0) {
            return;             /* enough already free */
        }
    }

    /* bounds check parameter */
    if (anumber > MAXATONCE)
        anumber = MAXATONCE;    /* all we can do */

    /* rewrite so phases include a better eligiblity for gc test*/
    /*
     * The phase variable manages reclaims.  Set to 0, the first pass,
     * we don't reclaim active entries, or other than target bucket.
     * Set to 1, we reclaim even active ones in target bucket.
     * Set to 2, we reclaim any inactive one.
     * Set to 3, we reclaim even active ones. On Solaris, we also reclaim
     * entries whose corresponding vcache has a nonempty multiPage list, when
     * possible.
     */
    if (splitdcache) {
        phase = 0;
    } else {
        phase = 4;
    }

    for (i = 0; i < afs_cacheFiles; i++)
        /* turn off all flags */
        afs_indexFlags[i] &= ~IFFlag;

    while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
        /* find oldest entries for reclamation */
        maxVictimPtr = victimPtr = 0;
        hzero(maxVictimTime);
        curbucket = afs_DCWhichBucket(phase, buckethint);
        /* select victims from access time array */
        for (i = 0; i < afs_cacheFiles; i++) {
            if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
                /* skip if dirty or already free */
                continue;
            }
            tdc = afs_indexTable[i];
            if (tdc && (curbucket != tdc->bucket) && (phase < 4))
            {
                /* Wrong bucket; can't use it! */
                continue;
            }
            if (tdc && (tdc->refCount != 0)) {
                /* Referenced; can't use it! */
                continue;
            }
            hset(vtime, afs_indexTimes[i]);

            /* if we've already looked at this one, skip it */
            if (afs_indexFlags[i] & IFFlag)
                continue;

            if (victimPtr < MAXATONCE) {
                /* if there's at least one free victim slot left */
                victims[victimPtr] = i;
                hset(victimTimes[victimPtr], vtime);
                if (hcmp(vtime, maxVictimTime) > 0) {
                    hset(maxVictimTime, vtime);
                    maxVictimPtr = victimPtr;
                }
                victimPtr++;
            } else if (hcmp(vtime, maxVictimTime) < 0) {
                /*
                 * We're older than youngest victim, so we replace at
                 * least one victim
                 */
                /* find youngest (largest LRU) victim */
                j = maxVictimPtr;
                if (j == victimPtr)
                    osi_Panic("getdownd local");
                victims[j] = i;
                hset(victimTimes[j], vtime);
                /* recompute maxVictimTime */
                hset(maxVictimTime, vtime);
                for (j = 0; j < victimPtr; j++)
                    if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
                        hset(maxVictimTime, victimTimes[j]);
                        maxVictimPtr = j;
                    }
            }
        }                       /* big for loop */

        /* now really reclaim the victims */
        j = 0;                  /* flag to track if we actually got any of the victims */
        /* first, hold all the victims, since we're going to release the lock
         * during the truncate operation.
         */
        for (i = 0; i < victimPtr; i++) {
            tdc = afs_GetValidDSlot(victims[i]);
            /* We got tdc->tlock(R) here */
            if (tdc && tdc->refCount == 1)
                victimDCs[i] = tdc;
            else
                victimDCs[i] = 0;
            if (tdc) {
                ReleaseReadLock(&tdc->tlock);
                if (!victimDCs[i])
                    afs_PutDCache(tdc);
            }
        }
        for (i = 0; i < victimPtr; i++) {
            /* q is first elt in dcache entry */
            tdc = victimDCs[i];
            /* now, since we're dropping the afs_xdcache lock below, we
             * have to verify, before proceeding, that there are no other
             * references to this dcache entry, even now.  Note that we
             * compare with 1, since we bumped it above when we called
             * afs_GetValidDSlot to preserve the entry's identity.
             */
            if (tdc && tdc->refCount == 1) {
                unsigned char chunkFlags;
                afs_size_t tchunkoffset = 0;
                afid = &tdc->f.fid;
                /* xdcache is lower than the xvcache lock */
                ReleaseWriteLock(&afs_xdcache);
                ObtainReadLock(&afs_xvcache);
                tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
                ReleaseReadLock(&afs_xvcache);
                ObtainWriteLock(&afs_xdcache, 527);
                skip = 0;
                if (tdc->refCount > 1)
                    skip = 1;
                if (tvc) {
                    tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
                    chunkFlags = afs_indexFlags[tdc->index];
                    if (((phase & 1) == 0) && osi_Active(tvc))
                        skip = 1;
                    if (((phase & 1) == 1) && osi_Active(tvc)
                        && (tvc->f.states & CDCLock)
                        && (chunkFlags & IFAnyPages))
                        skip = 1;
                    if (chunkFlags & IFDataMod)
                        skip = 1;
                    afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
                               ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
                               ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
                               ICL_HANDLE_OFFSET(tchunkoffset));

#if     defined(AFS_SUN5_ENV)
                    /*
                     * Now we try to invalidate pages.  We do this only for
                     * Solaris.  For other platforms, it's OK to recycle a
                     * dcache entry out from under a page, because the strategy
                     * function can call afs_GetDCache().
                     */
                    if (!skip && (chunkFlags & IFAnyPages)) {
                        int code;

                        ReleaseWriteLock(&afs_xdcache);
                        ObtainWriteLock(&tvc->vlock, 543);
                        if (!QEmpty(&tvc->multiPage)) {
                            if (phase < 3 || osi_VM_MultiPageConflict(tvc, tdc)) {
                                skip = 1;
                                goto endmultipage;
                            }
                        }
                        /* block locking pages */
                        tvc->vstates |= VPageCleaning;
                        /* block getting new pages */
                        tvc->activeV++;
                        ReleaseWriteLock(&tvc->vlock);
                        /* One last recheck */
                        ObtainWriteLock(&afs_xdcache, 333);
                        chunkFlags = afs_indexFlags[tdc->index];
                        if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
                            || (osi_Active(tvc) && (tvc->f.states & CDCLock)
                                && (chunkFlags & IFAnyPages))) {
                            skip = 1;
                            ReleaseWriteLock(&afs_xdcache);
                            goto endputpage;
                        }
                        ReleaseWriteLock(&afs_xdcache);

                        code = osi_VM_GetDownD(tvc, tdc);

                        ObtainWriteLock(&afs_xdcache, 269);
                        /* we actually removed all pages, clean and dirty */
                        if (code == 0) {
                            afs_indexFlags[tdc->index] &=
                                ~(IFDirtyPages | IFAnyPages);
                        } else
                            skip = 1;
                        ReleaseWriteLock(&afs_xdcache);
                      endputpage:
                        ObtainWriteLock(&tvc->vlock, 544);
                        if (--tvc->activeV == 0
                            && (tvc->vstates & VRevokeWait)) {
                            tvc->vstates &= ~VRevokeWait;
                            afs_osi_Wakeup((char *)&tvc->vstates);

                        }
                        if (tvc->vstates & VPageCleaning) {
                            tvc->vstates &= ~VPageCleaning;
                            afs_osi_Wakeup((char *)&tvc->vstates);
                        }
                      endmultipage:
                        ReleaseWriteLock(&tvc->vlock);
                    } else
#endif /* AFS_SUN5_ENV */
                    {
                        ReleaseWriteLock(&afs_xdcache);
                    }

                    afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
                    ObtainWriteLock(&afs_xdcache, 528);
                    if (afs_indexFlags[tdc->index] &
                        (IFDataMod | IFDirtyPages | IFAnyPages))
                        skip = 1;
                    if (tdc->refCount > 1)
                        skip = 1;
                }
#if     defined(AFS_SUN5_ENV)
                else {
                    /* no vnode, so IFDirtyPages is spurious (we don't
                     * sweep dcaches on vnode recycling, so we can have
                     * DIRTYPAGES set even when all pages are gone).  Just
                     * clear the flag.
                     * Hold vcache lock to prevent vnode from being
                     * created while we're clearing IFDirtyPages.
                     */
                    afs_indexFlags[tdc->index] &=
                        ~(IFDirtyPages | IFAnyPages);
                }
#endif
                if (skip) {
                    /* skip this guy and mark him as recently used */
                    afs_indexFlags[tdc->index] |= IFFlag;
                    afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
                               ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
                               ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
                               ICL_HANDLE_OFFSET(tchunkoffset));
                } else {
                    /* flush this dude from the data cache and reclaim;
                     * first, make sure no one will care that we damage
                     * it, by removing it from all hash tables.  Then,
                     * melt it down for parts.  Note that any concurrent
                     * (new possibility!) calls to GetDownD won't touch
                     * this guy because his reference count is > 0. */
                    afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
                               ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
                               ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
                               ICL_HANDLE_OFFSET(tchunkoffset));
                    AFS_STATCNT(afs_gget);
                    afs_HashOutDCache(tdc, 1);
                    if (tdc->f.chunkBytes != 0) {
                        discard = 1;
                        if (aneedSpace)
                            *aneedSpace -=
                                (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
                    } else {
                        discard = 0;
                    }
                    if (discard) {
                        afs_DiscardDCache(tdc);
                    } else {
                        afs_FreeDCache(tdc);
                    }
                    anumber--;
                    j = 1;      /* we reclaimed at least one victim */
                }
            }
            if (tdc)
                afs_PutDCache(tdc);
        }                       /* end of for victims loop */

        if (phase < 5) {
            /* Phase is 0 and no one was found, so try phase 1 (ignore
             * osi_Active flag) */
            if (j == 0) {
                phase++;
                for (i = 0; i < afs_cacheFiles; i++)
                    /* turn off all flags */
                    afs_indexFlags[i] &= ~IFFlag;
            }
        } else {
            /* found no one in phases 0-5, we're hosed */
            if (victimPtr == 0)
                break;
        }
    }                           /* big while loop */

    return;

}                               /*afs_GetDownD */


/*!
 * Remove adc from any hash tables that would allow it to be located
 * again by afs_FindDCache or afs_GetDCache.
 *
 * \param adc Pointer to dcache entry to remove from hash tables.
 *
 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
 *
 */
int
afs_HashOutDCache(struct dcache *adc, int zap)
{
    int i, us;

    AFS_STATCNT(afs_glink);
    if (zap)
        /* we know this guy's in the LRUQ.  We'll move dude into DCQ below */
        DZap(adc);
    /* if this guy is in the hash table, pull him out */
    if (adc->f.fid.Fid.Volume != 0) {
        /* remove entry from first hash chains */
        i = DCHash(&adc->f.fid, adc->f.chunk);
        us = afs_dchashTbl[i];
        if (us == adc->index) {
            /* first dude in the list */
            afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
        } else {
            /* somewhere on the chain */
            while (us != NULLIDX) {
                if (afs_dcnextTbl[us] == adc->index) {
                    /* found item pointing at the one to delete */
                    afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
                    break;
                }
                us = afs_dcnextTbl[us];
            }
            if (us == NULLIDX)
                osi_Panic("dcache hc");
        }
        /* remove entry from *other* hash chain */
        i = DVHash(&adc->f.fid);
        us = afs_dvhashTbl[i];
        if (us == adc->index) {
            /* first dude in the list */
            afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
        } else {
            /* somewhere on the chain */
            while (us != NULLIDX) {
                if (afs_dvnextTbl[us] == adc->index) {
                    /* found item pointing at the one to delete */
                    afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
                    break;
                }
                us = afs_dvnextTbl[us];
            }
            if (us == NULLIDX)
                osi_Panic("dcache hv");
        }
    }

    if (zap) {
        /* prevent entry from being found on a reboot (it is already out of
         * the hash table, but after a crash, we just look at fid fields of
         * stable (old) entries).
         */
         adc->f.fid.Fid.Volume = 0;     /* invalid */

        /* mark entry as modified */
        adc->dflags |= DFEntryMod;
    }

    /* all done */
    return 0;
}                               /*afs_HashOutDCache */

/*!
 * Flush the given dcache entry, pulling it from hash chains
 * and truncating the associated cache file.
 *
 * \param adc Ptr to dcache entry to flush.
 *
 * \note Environment:
 *      This routine must be called with the afs_xdcache lock held
 *      (in write mode).
 */
void
afs_FlushDCache(struct dcache *adc)
{
    AFS_STATCNT(afs_FlushDCache);
    /*
     * Bump the number of cache files flushed.
     */
    afs_stats_cmperf.cacheFlushes++;

    /* remove from all hash tables */
    afs_HashOutDCache(adc, 1);

    /* Free its space; special case null operation, since truncate operation
     * in UFS is slow even in this case, and this allows us to pre-truncate
     * these files at more convenient times with fewer locks set
     * (see afs_GetDownD).
     */
    if (adc->f.chunkBytes != 0) {
        afs_DiscardDCache(adc);
        afs_MaybeWakeupTruncateDaemon();
    } else {
        afs_FreeDCache(adc);
    }
}                               /*afs_FlushDCache */


/*!
 * Put a dcache entry on the free dcache entry list.
 *
 * \param adc dcache entry to free.
 *
 * \note Environment: called with afs_xdcache lock write-locked.
 */
static void
afs_FreeDCache(struct dcache *adc)
{
    /* Thread on free list, update free list count and mark entry as
     * freed in its indexFlags element.  Also, ensure DCache entry gets
     * written out (set DFEntryMod).
     */

    afs_dvnextTbl[adc->index] = afs_freeDCList;
    afs_freeDCList = adc->index;
    afs_freeDCCount++;
    afs_indexFlags[adc->index] |= IFFree;
    adc->dflags |= DFEntryMod;

    afs_WakeCacheWaitersIfDrained();
}                               /* afs_FreeDCache */

/*!
 * Discard the cache element by moving it to the discardDCList.
 * This puts the cache element into a quasi-freed state, where
 * the space may be reused, but the file has not been truncated.
 *
 * \note Major Assumptions Here:
 *      Assumes that frag size is an integral power of two, less one,
 *      and that this is a two's complement machine.  I don't
 *      know of any filesystems which violate this assumption...
 *
 * \param adr Ptr to dcache entry.
 *
 * \note Environment:
 *      Must be called with afs_xdcache write-locked.
 */

static void
afs_DiscardDCache(struct dcache *adc)
{
    afs_int32 size;

    AFS_STATCNT(afs_DiscardDCache);

    osi_Assert(adc->refCount == 1);

    size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10;       /* round up */
    afs_blocksDiscarded += size;
    afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;

    afs_dvnextTbl[adc->index] = afs_discardDCList;
    afs_discardDCList = adc->index;
    afs_discardDCCount++;

    adc->f.fid.Fid.Volume = 0;
    adc->dflags |= DFEntryMod;
    afs_indexFlags[adc->index] |= IFDiscarded;

    afs_WakeCacheWaitersIfDrained();
}                               /*afs_DiscardDCache */

/**
 * Get a dcache entry from the discard or free list
 *
 * @param[out] adc    On success, a dcache from the given list. Otherwise, NULL.
 * @param[in] indexp  A pointer to the head of the dcache free list or discard
 *                    list (afs_freeDCList, or afs_discardDCList)
 *
 * @return 0 on success. If there are no dcache slots available, return ENOSPC.
 *         If we encountered an error in disk i/o while trying to find a
 *         dcache, return EIO.
 *
 * @pre afs_xdcache is write-locked
 */
static int
afs_GetDSlotFromList(struct dcache **adc, afs_int32 *indexp)
{
    struct dcache *tdc;

    *adc = NULL;

    if (*indexp == NULLIDX) {
        return ENOSPC;
    }

    tdc = afs_GetUnusedDSlot(*indexp);
    if (tdc == NULL) {
        return EIO;
    }

    osi_Assert(tdc->refCount == 1);
    ReleaseReadLock(&tdc->tlock);
    *indexp = afs_dvnextTbl[tdc->index];
    afs_dvnextTbl[tdc->index] = NULLIDX;

    *adc = tdc;
    return 0;
}

/*!
 * Free the next element on the list of discarded cache elements.
 *
 * Returns -1 if we encountered an error preventing us from freeing a
 * discarded dcache, or 0 on success.
 */
static int
afs_FreeDiscardedDCache(void)
{
    struct dcache *tdc;
    struct osi_file *tfile;
    afs_int32 size;

    AFS_STATCNT(afs_FreeDiscardedDCache);

    ObtainWriteLock(&afs_xdcache, 510);
    if (!afs_blocksDiscarded) {
        ReleaseWriteLock(&afs_xdcache);
        return 0;
    }

    /*
     * Get an entry from the list of discarded cache elements
     */
    (void)afs_GetDSlotFromList(&tdc, &afs_discardDCList);
    if (!tdc) {
        ReleaseWriteLock(&afs_xdcache);
        return -1;
    }

    afs_discardDCCount--;
    size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10;       /* round up */
    afs_blocksDiscarded -= size;
    afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
    /* We can lock because we just took it off the free list */
    ObtainWriteLock(&tdc->lock, 626);
    ReleaseWriteLock(&afs_xdcache);

    /*
     * Truncate the element to reclaim its space
     */
    tfile = afs_CFileOpen(&tdc->f.inode);
    osi_Assert(tfile);
    afs_CFileTruncate(tfile, 0);
    afs_CFileClose(tfile);
    afs_AdjustSize(tdc, 0);
    afs_DCMoveBucket(tdc, 0, 0);

    /*
     * Free the element we just truncated
     */
    ObtainWriteLock(&afs_xdcache, 511);
    afs_indexFlags[tdc->index] &= ~IFDiscarded;
    afs_FreeDCache(tdc);
    tdc->f.states &= ~(DRO|DBackup|DRW);
    ReleaseWriteLock(&tdc->lock);
    afs_PutDCache(tdc);
    ReleaseWriteLock(&afs_xdcache);

    return 0;
}

/*!
 * Free as many entries from the list of discarded cache elements
 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
 *
 * \return 0
 */
int
afs_MaybeFreeDiscardedDCache(void)
{

    AFS_STATCNT(afs_MaybeFreeDiscardedDCache);

    while (afs_blocksDiscarded
           && (afs_blocksUsed >
               PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
        int code = afs_FreeDiscardedDCache();
        if (code) {
            /* Callers depend on us to get the afs_blocksDiscarded count down.
             * If we cannot do that, the callers can spin by calling us over
             * and over. Panic for now until we can figure out something
             * better. */
            osi_Panic("Error freeing discarded dcache");
        }
    }
    return 0;
}

/*!
 * Try to free up a certain number of disk slots.
 *
 * \param anumber Targeted number of disk slots to free up.
 *
 * \note Environment:
 *      Must be called with afs_xdcache write-locked.
 *
 */
static void
afs_GetDownDSlot(int anumber)
{
    struct afs_q *tq, *nq;
    struct dcache *tdc;
    int ix;
    unsigned int cnt;

    AFS_STATCNT(afs_GetDownDSlot);
    if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
        osi_Panic("diskless getdowndslot");

    if (CheckLock(&afs_xdcache) != -1)
        osi_Panic("getdowndslot nolock");

    /* decrement anumber first for all dudes in free list */
    for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
        anumber--;
    if (anumber <= 0)
        return;                 /* enough already free */

    for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
         tq = nq, cnt++) {
        tdc = (struct dcache *)tq;      /* q is first elt in dcache entry */
        nq = QPrev(tq);         /* in case we remove it */
        if (tdc->refCount == 0) {
            if ((ix = tdc->index) == NULLIDX)
                osi_Panic("getdowndslot");

            /* write-through if modified */
            if (tdc->dflags & DFEntryMod) {
#if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
                /*
                 * ask proxy to do this for us - we don't have the stack space
                 */
                while (tdc->dflags & DFEntryMod) {
                    int s;
                    AFS_GUNLOCK();
                    s = SPLOCK(afs_sgibklock);
                    if (afs_sgibklist == NULL) {
                        /* if slot is free, grab it. */
                        afs_sgibklist = tdc;
                        SV_SIGNAL(&afs_sgibksync);
                    }
                    /* wait for daemon to (start, then) finish. */
                    SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
                    AFS_GLOCK();
                }
#else
                int code;

                code = afs_WriteDCache(tdc, 1);
                if (code) {
                    /*
                     * We couldn't flush it at this time; return early because
                     * if afs_WriteDCache() failed once it is likely to
                     * continue failing for subsequent dcaches.
                     */
                    return;
                }
                tdc->dflags &= ~DFEntryMod;
#endif
            }

            /* pull the entry out of the lruq and put it on the free list */
            QRemove(&tdc->lruq);
            afs_indexTable[ix] = NULL;
            afs_indexFlags[ix] &= ~IFEverUsed;
            tdc->index = NULLIDX;
            tdc->lruq.next = (struct afs_q *)afs_freeDSList;
            afs_freeDSList = tdc;
            anumber--;
        }
    }
}                               /*afs_GetDownDSlot */


/*
 * afs_RefDCache
 *
 * Description:
 *      Increment the reference count on a disk cache entry,
 *      which already has a non-zero refcount.  In order to
 *      increment the refcount of a zero-reference entry, you
 *      have to hold afs_xdcache.
 *
 * Parameters:
 *      adc : Pointer to the dcache entry to increment.
 *
 * Environment:
 *      Nothing interesting.
 */
int
afs_RefDCache(struct dcache *adc)
{
    ObtainWriteLock(&adc->tlock, 627);
    if (adc->refCount < 0)
        osi_Panic("RefDCache: negative refcount");
    adc->refCount++;
    ReleaseWriteLock(&adc->tlock);
    return 0;
}


/*
 * afs_PutDCache
 *
 * Description:
 *      Decrement the reference count on a disk cache entry.
 *
 * Parameters:
 *      ad : Ptr to the dcache entry to decrement.
 *
 * Environment:
 *      Nothing interesting.
 */
int
afs_PutDCache(struct dcache *adc)
{
    AFS_STATCNT(afs_PutDCache);
    ObtainWriteLock(&adc->tlock, 276);
    if (adc->refCount <= 0)
        osi_Panic("putdcache");
    --adc->refCount;
    ReleaseWriteLock(&adc->tlock);
    return 0;
}


/*
 * afs_TryToSmush
 *
 * Description:
 *      Try to discard all data associated with this file from the
 *      cache.
 *
 * Parameters:
 *      avc : Pointer to the cache info for the file.
 *
 * Environment:
 *      Both pvnLock and lock are write held.
 */
void
afs_TryToSmush(struct vcache *avc, afs_ucred_t *acred, int sync)
{
    struct dcache *tdc;
    int index;
    int i;
    AFS_STATCNT(afs_TryToSmush);
    afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
               ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
    sync = 1;                   /* XX Temp testing XX */

#if     defined(AFS_SUN5_ENV)
    ObtainWriteLock(&avc->vlock, 573);
    avc->activeV++;             /* block new getpages */
    ReleaseWriteLock(&avc->vlock);
#endif

    /* Flush VM pages */
    osi_VM_TryToSmush(avc, acred, sync);

    /*
     * Get the hash chain containing all dce's for this fid
     */
    i = DVHash(&avc->f.fid);
    ObtainWriteLock(&afs_xdcache, 277);
    for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
        i = afs_dvnextTbl[index];       /* next pointer this hash table */
        if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
            int releaseTlock = 1;
            tdc = afs_GetValidDSlot(index);
            if (!tdc) {
                /* afs_TryToSmush is best-effort; we may not actually discard
                 * everything, so failure to discard dcaches due to an i/o
                 * error is okay. */
                break;
            }
            if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
                if (sync) {
                    if ((afs_indexFlags[index] & IFDataMod) == 0
                        && tdc->refCount == 1) {
                        ReleaseReadLock(&tdc->tlock);
                        releaseTlock = 0;
                        afs_FlushDCache(tdc);
                    }
                } else
                    afs_indexTable[index] = 0;
            }
            if (releaseTlock)
                ReleaseReadLock(&tdc->tlock);
            afs_PutDCache(tdc);
        }
    }
#if     defined(AFS_SUN5_ENV)
    ObtainWriteLock(&avc->vlock, 545);
    if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
        avc->vstates &= ~VRevokeWait;
        afs_osi_Wakeup((char *)&avc->vstates);
    }
    ReleaseWriteLock(&avc->vlock);
#endif
    ReleaseWriteLock(&afs_xdcache);
    /*
     * It's treated like a callback so that when we do lookups we'll
     * invalidate the unique bit if any
     * trytoSmush occured during the lookup call
     */
    afs_allCBs++;
}

/*
 * afs_DCacheMissingChunks
 *
 * Description
 *      Given the cached info for a file, return the number of chunks that
 *      are not available from the dcache.
 *
 * Parameters:
 *      avc:    Pointer to the (held) vcache entry to look in.
 *
 * Returns:
 *      The number of chunks which are not currently cached.
 *
 * Environment:
 *      The vcache entry is held upon entry.
 */

int
afs_DCacheMissingChunks(struct vcache *avc)
{
    int i, index;
    afs_size_t totalLength = 0;
    afs_uint32 totalChunks = 0;
    struct dcache *tdc;

    totalLength = avc->f.m.Length;
    if (avc->f.truncPos < totalLength)
        totalLength = avc->f.truncPos;

    /* Length is 0, no chunk missing. */
    if (totalLength == 0)
        return 0;

    /* If totalLength is a multiple of chunksize, the last byte appears
     * as being part of the next chunk, which does not exist.
     * Decrementing totalLength by one fixes that.
     */
    totalLength--;
    totalChunks = (AFS_CHUNK(totalLength) + 1);

    /* If we're a directory, we only ever have one chunk, regardless of
     * the size of the dir.
     */
    if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
        totalChunks = 1;

    /*
     printf("Should have %d chunks for %u bytes\n",
                totalChunks, (totalLength + 1));
    */
    i = DVHash(&avc->f.fid);
    ObtainWriteLock(&afs_xdcache, 1001);
    for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
        i = afs_dvnextTbl[index];
        if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
            tdc = afs_GetValidDSlot(index);
            if (!tdc) {
                break;
            }
            if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
                totalChunks--;
            }
            ReleaseReadLock(&tdc->tlock);
            afs_PutDCache(tdc);
        }
    }
    ReleaseWriteLock(&afs_xdcache);

    /*printf("Missing %d chunks\n", totalChunks);*/

    return (totalChunks);
}

/*
 * afs_FindDCache
 *
 * Description:
 *      Given the cached info for a file and a byte offset into the
 *      file, make sure the dcache entry for that file and containing
 *      the given byte is available, returning it to our caller.
 *
 * Parameters:
 *      avc   : Pointer to the (held) vcache entry to look in.
 *      abyte : Which byte we want to get to.
 *
 * Returns:
 *      Pointer to the dcache entry covering the file & desired byte,
 *      or NULL if not found.
 *
 * Environment:
 *      The vcache entry is held upon entry.
 */

struct dcache *
afs_FindDCache(struct vcache *avc, afs_size_t abyte)
{
    afs_int32 chunk;
    afs_int32 i, index;
    struct dcache *tdc = NULL;

    AFS_STATCNT(afs_FindDCache);
    chunk = AFS_CHUNK(abyte);

    /*
     * Hash on the [fid, chunk] and get the corresponding dcache index
     * after write-locking the dcache.
     */
    i = DCHash(&avc->f.fid, chunk);
    ObtainWriteLock(&afs_xdcache, 278);
    for (index = afs_dchashTbl[i]; index != NULLIDX; index = afs_dcnextTbl[index]) {
        if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
            tdc = afs_GetValidDSlot(index);
            if (!tdc) {
                /* afs_FindDCache is best-effort; we may not find the given
                 * file/offset, so if we cannot find the given dcache due to
                 * i/o errors, that is okay. */
                index = NULLIDX;
                break;
            }
            ReleaseReadLock(&tdc->tlock);
            if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
                break;          /* leaving refCount high for caller */
            }
            afs_PutDCache(tdc);
        }
    }
    if (index != NULLIDX) {
        hset(afs_indexTimes[tdc->index], afs_indexCounter);
        hadd32(afs_indexCounter, 1);
        ReleaseWriteLock(&afs_xdcache);
        return tdc;
    }
    ReleaseWriteLock(&afs_xdcache);
    return NULL;
}                               /*afs_FindDCache */

/* only call these from afs_AllocDCache() */
static int
afs_AllocFreeDSlot(struct dcache **adc)
{
    int code;
    struct dcache *tdc;

    code = afs_GetDSlotFromList(&tdc, &afs_freeDCList);
    if (code) {
        return code;
    }
    afs_indexFlags[tdc->index] &= ~IFFree;
    ObtainWriteLock(&tdc->lock, 604);
    afs_freeDCCount--;

    *adc = tdc;
    return 0;
}
static int
afs_AllocDiscardDSlot(struct dcache **adc, afs_int32 lock)
{
    int code;
    struct dcache *tdc;
    afs_uint32 size = 0;
    struct osi_file *file;

    code = afs_GetDSlotFromList(&tdc, &afs_discardDCList);
    if (code) {
        return code;
    }
    afs_indexFlags[tdc->index] &= ~IFDiscarded;
    ObtainWriteLock(&tdc->lock, 605);
    afs_discardDCCount--;
    size =
        ((tdc->f.chunkBytes +
          afs_fsfragsize) ^ afs_fsfragsize) >> 10;
    tdc->f.states &= ~(DRO|DBackup|DRW);
    afs_DCMoveBucket(tdc, size, 0);
    afs_blocksDiscarded -= size;
    afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
    if ((lock & 2)) {
        /* Truncate the chunk so zeroes get filled properly */
        file = afs_CFileOpen(&tdc->f.inode);
        osi_Assert(file);
        afs_CFileTruncate(file, 0);
        afs_CFileClose(file);
        afs_AdjustSize(tdc, 0);
    }

    *adc = tdc;
    return 0;
}

/*!
 * Get a fresh dcache from the free or discarded list.
 *
 * \param adc Set to the new dcache on success, and NULL on error.
 * \param avc Who's dcache is this going to be?
 * \param chunk The position where it will be placed in.
 * \param lock How are locks held.
 * \param ashFid If this dcache going to be used for a shadow dir,
 *              this is it's fid.
 *
 * \note Required locks:
 *      - afs_xdcache (W)
 *      - avc (R if (lock & 1) set and W otherwise)
 * \note It write locks the new dcache. The caller must unlock it.
 *
 * \return If we're out of dslots, ENOSPC. If we encountered disk errors, EIO.
 *         On success, return 0.
 */
static int
afs_AllocDCache(struct dcache **adc, struct vcache *avc, afs_int32 chunk,
                afs_int32 lock, struct VenusFid *ashFid)
{
    int code;
    struct dcache *tdc = NULL;

    *adc = NULL;

    /* if (lock & 2), prefer 'free' dcaches; otherwise, prefer 'discard'
     * dcaches. In either case, try both if our first choice doesn't work due
     * to ENOSPC. */
    if ((lock & 2)) {
        code = afs_AllocFreeDSlot(&tdc);
        if (code == ENOSPC) {
            code = afs_AllocDiscardDSlot(&tdc, lock);
        }
    } else {
        code = afs_AllocDiscardDSlot(&tdc, lock);
        if (code == ENOSPC) {
            code = afs_AllocFreeDSlot(&tdc);
        }
    }
    if (code) {
        return code;
    }

    /*
     * Locks held:
     * avc->lock(R) if setLocks
     * avc->lock(W) if !setLocks
     * tdc->lock(W)
     * afs_xdcache(W)
     */

    /*
     * Fill in the newly-allocated dcache record.
     */
    afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
    if (ashFid)
        /* Use shadow fid if provided. */
        tdc->f.fid = *ashFid;
    else
        /* Use normal vcache's fid otherwise. */
        tdc->f.fid = avc->f.fid;
    if (avc->f.states & CRO)
        tdc->f.states = DRO;
    else if (avc->f.states & CBackup)
        tdc->f.states = DBackup;
    else
        tdc->f.states = DRW;
    afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
    afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
    if (!ashFid)
        hones(tdc->f.versionNo);        /* invalid value */
    tdc->f.chunk = chunk;
    tdc->validPos = AFS_CHUNKTOBASE(chunk);
    /* XXX */
    if (tdc->lruq.prev == &tdc->lruq)
        osi_Panic("lruq 1");

    *adc = tdc;
    return 0;
}

static int
IsDCacheSizeOK(struct dcache *adc, struct vcache *avc, afs_int32 chunk_bytes,
               afs_size_t file_length, afs_uint32 versionNo, int from_net)
{
    afs_size_t expected_bytes;
    afs_size_t chunk_start = AFS_CHUNKTOBASE(adc->f.chunk);

    if (vType(avc) == VDIR) {
        /*
         * Directory blobs may be constructed locally (see afs_LocalHero), and
         * the size of the blob may differ slightly compared to what's on the
         * fileserver. So, skip size checks for directories.
         */
        return 1;
    }

    if ((avc->f.states & CDirty)) {
        /*
         * Our vcache may have writes that are local to our cache, but not yet
         * written to the fileserver. In such a situation, we may have dcaches
         * for that file that are "short". For example:
         *
         * Say we have a file that is 0 bytes long. A process opens that file,
         * and writes some data to offset 5M (keeping the file open). Another
         * process comes along and reads data from offset 1M. We'll try to
         * fetch data at offset 1M, and the fileserver will respond with 0
         * bytes, since our locally-written data hasn't been written to the
         * fileserver yet (on the fileserver, the file is still 0-bytes long).
         * So our dcache at offset 1M will have 0 bytes.
         *
         * So if CDirty is set, don't do any size/length checks at all, since
         * we have no idea if the avc length is valid.
         */
        return 1;
    }

    if (!from_net && (adc->f.states & DRW)) {
        /*
         * The dcache data we're looking at is from our local cache (not from a
         * fileserver), and it's for data in an RW volume. For cached RW data,
         * there are some edge cases that can cause the below length checks to
         * trigger false positives.
         *
         * For example: if the local client writes 4 bytes to a new file at
         * offset 0, and then 4 bytes at offset 0x400000, the file will be
         * 0x400004 bytes long, but the first dcache chunk will only contain 4
         * bytes. If such a file is fetched from a fileserver, the first chunk
         * will have a full chunk of data (most of it zeroes), but on the
         * client that did the write, the sparse data will not appear in the
         * dcache.
         *
         * Such false positives should only be possible with RW data, since
         * non-RW data is never generated locally. So to avoid the false
         * positives, assume the dcache length is OK for RW data if the dcache
         * came from our local cache (and not directly from a fileserver).
         */
        return 1;
    }

    if (file_length < chunk_start) {
        expected_bytes = 0;

    } else {
        expected_bytes = file_length - chunk_start;

        if (vType(avc) != VDIR && expected_bytes > AFS_CHUNKTOSIZE(adc->f.chunk)) {
            /* A non-dir chunk cannot have more bytes than the chunksize. */
            expected_bytes = AFS_CHUNKTOSIZE(adc->f.chunk);
        }
    }

    if (chunk_bytes != expected_bytes) {
        static const afs_uint32 one_hour = 60 * 60;
        static afs_uint32 last_warn;
        afs_uint32 now = osi_Time();

        if (now < last_warn) {
            /* clock went backwards */
            last_warn = now;
        }

        if (now - last_warn > one_hour) {
            unsigned int mtime = adc->f.modTime;

            last_warn = now;

            if (from_net) {
                /*
                 * The dcache we're looking at didn't come from the cache, but is
                 * being populated from the net. Don't print out its mtime in that
                 * case; that would be misleading since that's the mtime from the
                 * last time this dcache slot was written to.
                 */
                mtime = 0;
            }

            afs_warn("afs: Detected corrupt dcache for file %d.%u.%u.%u: chunk %d "
                     "(offset %lu) has %d bytes, but it should have %lu bytes\n",
                     adc->f.fid.Cell,
                     adc->f.fid.Fid.Volume,
                     adc->f.fid.Fid.Vnode,
                     adc->f.fid.Fid.Unique,
                     adc->f.chunk,
                     (unsigned long)chunk_start,
                     chunk_bytes,
                     (unsigned long)expected_bytes);
            afs_warn("afs: (dcache %p, file length %lu, DV %u, dcache mtime %u, "
                     "index %d, dflags 0x%x, mflags 0x%x, states 0x%x, vcache "
                     "states 0x%x)\n",
                     adc,
                     (unsigned long)file_length,
                     versionNo,
                     mtime,
                     adc->index,
                     (unsigned)adc->dflags,
                     (unsigned)adc->mflags,
                     (unsigned)adc->f.states,
                     avc->f.states);
            afs_warn("afs: Ignoring the dcache for now, but this may indicate "
                     "corruption in the AFS cache, or a bug.\n");
        }
        return 0;
    }
    return 1;
}

/*!
 * Check if a dcache is "fresh". That is, if the dcache's DV matches the DV of
 * the vcache for that file, and the dcache looks "sane" (its length makes
 * sense, when considering the length of the given avc).
 *
 * \param adc The dcache to check
 * \param avc The vcache for adc
 *
 * \return 1 if the dcache is "fresh". 0 otherwise.
 */
int
afs_IsDCacheFresh(struct dcache *adc, struct vcache *avc)
{
    if (!hsame(adc->f.versionNo, avc->f.m.DataVersion)) {
        return 0;
    }

    /*
     * If we've reached here, the DV in adc matches the DV of our avc. Check if
     * the number of bytes in adc agrees with the avc file length, as a sanity
     * check. If they don't match, we'll pretend the DVs don't match, so the
     * bad dcache data will not be used, and we'll probably re-fetch the chunk
     * data, replacing the bad chunk.
     */

    if (!IsDCacheSizeOK(adc, avc, adc->f.chunkBytes, avc->f.m.Length,
                        hgetlo(adc->f.versionNo), 0)) {
        return 0;
    }

    return 1;
}

/*
 * afs_GetDCache
 *
 * Description:
 *      This function is called to obtain a reference to data stored in
 *      the disk cache, locating a chunk of data containing the desired
 *      byte and returning a reference to the disk cache entry, with its
 *      reference count incremented.
 *
 * Parameters:
 * IN:
 *      avc     : Ptr to a vcache entry (unlocked)
 *      abyte   : Byte position in the file desired
 *      areq    : Request structure identifying the requesting user.
 *      aflags  : Settings as follows:
 *                      1 : Set locks
 *                      2 : Return after creating entry.
 *                      4 : called from afs_vnop_write.c
 *                          *alen contains length of data to be written.
 * OUT:
 *      aoffset : Set to the offset within the chunk where the resident
 *                byte is located.
 *      alen    : Set to the number of bytes of data after the desired
 *                byte (including the byte itself) which can be read
 *                from this chunk.
 *
 * Environment:
 *      The vcache entry pointed to by avc is unlocked upon entry.
 */

/*
 * Update the vnode-to-dcache hint if we can get the vnode lock
 * right away.  Assumes dcache entry is at least read-locked.
 */
void
updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
{
    if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
        if (afs_IsDCacheFresh(d, v) && v->callback)
            v->dchint = d;
        if (lockVc)
            ReleaseWriteLock(&v->lock);
    }
}

/* avc - Write-locked unless aflags & 1 */
struct dcache *
afs_GetDCache(struct vcache *avc, afs_size_t abyte,
              struct vrequest *areq, afs_size_t * aoffset,
              afs_size_t * alen, int aflags)
{
    afs_int32 i, code, shortcut;
#if     defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
    afs_int32 adjustsize = 0;
#endif
    int setLocks;
    afs_int32 index;
    afs_int32 us;
    afs_int32 chunk;
    afs_size_t Position = 0;
    afs_int32 size, tlen;       /* size of segment to transfer */
    struct afs_FetchOutput *tsmall = 0;
    struct dcache *tdc;
    struct osi_file *file;
    struct afs_conn *tc;
    int downDCount = 0;
    struct server *newCallback = NULL;
    char setNewCallback;
    char setVcacheStatus;
    char doVcacheUpdate;
    char slowPass = 0;
    int doAdjustSize = 0;
    int doReallyAdjustSize = 0;
    int overWriteWholeChunk = 0;
    struct rx_connection *rxconn;

#ifndef AFS_NOSTATS
    struct afs_stats_AccessInfo *accP;  /*Ptr to access record in stats */
    int fromReplica;            /*Are we reading from a replica? */
    int numFetchLoops;          /*# times around the fetch/analyze loop */
#endif /* AFS_NOSTATS */

    AFS_STATCNT(afs_GetDCache);
    if (dcacheDisabled)
        return NULL;

    setLocks = aflags & 1;

    /*
     * Determine the chunk number and offset within the chunk corresponding
     * to the desired byte.
     */
    if (avc->f.fid.Fid.Vnode & 1) {     /* if (vType(avc) == VDIR) */
        chunk = 0;
    } else {
        chunk = AFS_CHUNK(abyte);
    }

    /* come back to here if we waited for the cache to drain. */
  RetryGetDCache:

    setNewCallback = setVcacheStatus = 0;

    if (setLocks) {
        if (slowPass)
            ObtainWriteLock(&avc->lock, 616);
        else
            ObtainReadLock(&avc->lock);
    }

    /*
     * Locks held:
     * avc->lock(R) if setLocks && !slowPass
     * avc->lock(W) if !setLocks || slowPass
     */

    shortcut = 0;

    /* check hints first! (might could use bcmp or some such...) */
    if ((tdc = avc->dchint)) {
        int dcLocked;

        /*
         * The locking order between afs_xdcache and dcache lock matters.
         * The hint dcache entry could be anywhere, even on the free list.
         * Locking afs_xdcache ensures that noone is trying to pull dcache
         * entries from the free list, and thereby assuming them to be not
         * referenced and not locked.
         */
        ObtainReadLock(&afs_xdcache);
        dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));

        if (dcLocked && (tdc->index != NULLIDX)
            && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
            && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
            /* got the right one.  It might not be the right version, and it
             * might be fetching, but it's the right dcache entry.
             */
            /* All this code should be integrated better with what follows:
             * I can save a good bit more time under a write lock if I do..
             */
            ObtainWriteLock(&tdc->tlock, 603);
            tdc->refCount++;
            ReleaseWriteLock(&tdc->tlock);

            ReleaseReadLock(&afs_xdcache);
            shortcut = 1;

            if (afs_IsDCacheFresh(tdc, avc)
                && !(tdc->dflags & DFFetching)) {

                afs_stats_cmperf.dcacheHits++;
                ObtainWriteLock(&afs_xdcache, 559);
                QRemove(&tdc->lruq);
                QAdd(&afs_DLRU, &tdc->lruq);
                ReleaseWriteLock(&afs_xdcache);

                /* Locks held:
                 * avc->lock(R) if setLocks && !slowPass
                 * avc->lock(W) if !setLocks || slowPass
                 * tdc->lock(S)
                 */
                goto done;
            }
        } else {
            if (dcLocked)
                ReleaseSharedLock(&tdc->lock);
            ReleaseReadLock(&afs_xdcache);
        }

        if (!shortcut)
            tdc = 0;
    }

    /* Locks held:
     * avc->lock(R) if setLocks && !slowPass
     * avc->lock(W) if !setLocks || slowPass
     * tdc->lock(S) if tdc
     */

    if (!tdc) {                 /* If the hint wasn't the right dcache entry */
        int dslot_error = 0;
        /*
         * Hash on the [fid, chunk] and get the corresponding dcache index
         * after write-locking the dcache.
         */
      RetryLookup:

        /* Locks held:
         * avc->lock(R) if setLocks && !slowPass
         * avc->lock(W) if !setLocks || slowPass
         */

        i = DCHash(&avc->f.fid, chunk);
        /* check to make sure our space is fine */
        afs_MaybeWakeupTruncateDaemon();

        ObtainWriteLock(&afs_xdcache, 280);
        us = NULLIDX;
        for (index = afs_dchashTbl[i]; index != NULLIDX; us = index, index = afs_dcnextTbl[index]) {
            if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
                tdc = afs_GetValidDSlot(index);
                if (!tdc) {
                    /* we got an i/o error when trying to get the given dslot.
                     * it's possible the dslot we're looking for is elsewhere,
                     * but most likely the disk cache is currently unusable, so
                     * all afs_GetValidDSlot calls will fail, so just bail out. */
                    dslot_error = 1;
                    index = NULLIDX;
                    break;
                }
                ReleaseReadLock(&tdc->tlock);
                /*
                 * Locks held:
                 * avc->lock(R) if setLocks && !slowPass
                 * avc->lock(W) if !setLocks || slowPass
                 * afs_xdcache(W)
                 */
                if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
                    /* Move it up in the beginning of the list */
                    if (afs_dchashTbl[i] != index) {
                        afs_dcnextTbl[us] = afs_dcnextTbl[index];
                        afs_dcnextTbl[index] = afs_dchashTbl[i];
                        afs_dchashTbl[i] = index;
                    }
                    ReleaseWriteLock(&afs_xdcache);
                    ObtainSharedLock(&tdc->lock, 606);
                    break;      /* leaving refCount high for caller */
                }
                afs_PutDCache(tdc);
                tdc = 0;
            }
        }

        /*
         * If we didn't find the entry, we'll create one.
         */
        if (index == NULLIDX) {
            /*
             * Locks held:
             * avc->lock(R) if setLocks
             * avc->lock(W) if !setLocks
             * afs_xdcache(W)
             */
            afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
                       avc, ICL_TYPE_INT32, chunk);

            if (dslot_error) {
                /* We couldn't find the dcache we want, but we hit some i/o
                 * errors when trying to find it, so we're not sure if the
                 * dcache we want is in the cache or not. Error out, so we
                 * don't try to possibly create 2 separate dcaches for the
                 * same exact data. */
                ReleaseWriteLock(&afs_xdcache);
                goto done;
            }

            if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
                if (!setLocks)
                    avc->f.states |= CDCLock;
                /* just need slots */
                afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
                if (!setLocks)
                    avc->f.states &= ~CDCLock;
            }
            code = afs_AllocDCache(&tdc, avc, chunk, aflags, NULL);
            if (code) {
                ReleaseWriteLock(&afs_xdcache);
                if (code == ENOSPC) {
                    /* It looks like afs_AllocDCache failed because we don't
                     * have any free dslots to use. Maybe if we wait a little
                     * while, we'll be able to free up some slots, so try for 5
                     * minutes, then bail out. */
                    if (++downDCount > 300) {
                        afs_warn("afs: Unable to get free cache space for file "
                                 "%u:%u.%u.%u for 5 minutes; failing with an i/o error\n",
                                 avc->f.fid.Cell,
                                 avc->f.fid.Fid.Volume,
                                 avc->f.fid.Fid.Vnode,
                                 avc->f.fid.Fid.Unique);
                        goto done;
                    }
                    afs_osi_Wait(1000, 0, 0);
                    goto RetryLookup;
                }

                /* afs_AllocDCache failed, but not because we're out of free
                 * dslots. Something must be screwy with the cache, so bail out
                 * immediately without waiting. */
                afs_warn("afs: Error while alloc'ing cache slot for file "
                         "%u:%u.%u.%u; failing with an i/o error\n",
                         avc->f.fid.Cell,
                         avc->f.fid.Fid.Volume,
                         avc->f.fid.Fid.Vnode,
                         avc->f.fid.Fid.Unique);
                goto done;
            }

            /*
             * Locks held:
             * avc->lock(R) if setLocks
             * avc->lock(W) if !setLocks
             * tdc->lock(W)
             * afs_xdcache(W)
             */

            /*
             * Now add to the two hash chains - note that i is still set
             * from the above DCHash call.
             */
            afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
            afs_dchashTbl[i] = tdc->index;
            i = DVHash(&avc->f.fid);
            afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
            afs_dvhashTbl[i] = tdc->index;
            tdc->dflags = DFEntryMod;
            tdc->mflags = 0;
            afs_MaybeWakeupTruncateDaemon();
            ReleaseWriteLock(&afs_xdcache);
            ConvertWToSLock(&tdc->lock);
        }
    }


    /* vcache->dcache hint failed */
    /*
     * Locks held:
     * avc->lock(R) if setLocks && !slowPass
     * avc->lock(W) if !setLocks || slowPass
     * tdc->lock(S)
     */
    afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
               ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
               hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
               hgetlo(avc->f.m.DataVersion));
    /*
     * Here we have the entry in tdc, with its refCount incremented.
     * Note: we don't use the S-lock on avc; it costs concurrency when
     * storing a file back to the server.
     */

    /*
     * Not a newly created file so we need to check the file's length and
     * compare data versions since someone could have changed the data or we're
     * reading a file written elsewhere. We only want to bypass doing no-op
     * read rpcs on newly created files (dv of 0) since only then we guarantee
     * that this chunk's data hasn't been filled by another client.
     */
    size = AFS_CHUNKSIZE(abyte);
    if (aflags & 4)             /* called from write */
        tlen = *alen;
    else                        /* called from read */
        tlen = tdc->validPos - abyte;
    Position = AFS_CHUNKTOBASE(chunk);
    afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
               ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
               ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
               ICL_HANDLE_OFFSET(Position));
    if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
        doAdjustSize = 1;
    if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
         ((aflags & 4) && (abyte == Position) && (tlen >= size)))
        overWriteWholeChunk = 1;
    if (doAdjustSize || overWriteWholeChunk) {
#if     defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
#ifdef  AFS_SGI_ENV
#ifdef AFS_SGI64_ENV
        if (doAdjustSize)
            adjustsize = NBPP;
#else /* AFS_SGI64_ENV */
        if (doAdjustSize)
            adjustsize = 8192;
#endif /* AFS_SGI64_ENV */
#else /* AFS_SGI_ENV */
        if (doAdjustSize)
            adjustsize = 4096;
#endif /* AFS_SGI_ENV */
        if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
#else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
#if     defined(AFS_SUN5_ENV)
        if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
#else
        if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
#endif
#endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
            !afs_IsDCacheFresh(tdc, avc))
            doReallyAdjustSize = 1;

        if (doReallyAdjustSize || overWriteWholeChunk) {
            /* no data in file to read at this position */
            UpgradeSToWLock(&tdc->lock, 607);
            file = afs_CFileOpen(&tdc->f.inode);
            osi_Assert(file);
            afs_CFileTruncate(file, 0);
            afs_CFileClose(file);
            afs_AdjustSize(tdc, 0);
            hset(tdc->f.versionNo, avc->f.m.DataVersion);
            tdc->dflags |= DFEntryMod;

            ConvertWToSLock(&tdc->lock);
        }
    }

    /*
     * We must read in the whole chunk if the version number doesn't
     * match.
     */
    if (aflags & 2) {
        /* don't need data, just a unique dcache entry */
        ObtainWriteLock(&afs_xdcache, 608);
        hset(afs_indexTimes[tdc->index], afs_indexCounter);
        hadd32(afs_indexCounter, 1);
        ReleaseWriteLock(&afs_xdcache);

        updateV2DC(setLocks, avc, tdc, 553);
        if (vType(avc) == VDIR)
            *aoffset = abyte;
        else
            *aoffset = AFS_CHUNKOFFSET(abyte);
        if (tdc->validPos < abyte)
            *alen = (afs_size_t) 0;
        else
            *alen = tdc->validPos - abyte;
        ReleaseSharedLock(&tdc->lock);
        if (setLocks) {
            if (slowPass)
                ReleaseWriteLock(&avc->lock);
            else
                ReleaseReadLock(&avc->lock);
        }
        return tdc;             /* check if we're done */
    }

    /*
     * Locks held:
     * avc->lock(R) if setLocks && !slowPass
     * avc->lock(W) if !setLocks || slowPass
     * tdc->lock(S)
     */
    osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));

    setNewCallback = setVcacheStatus = 0;

    /*
     * Locks held:
     * avc->lock(R) if setLocks && !slowPass
     * avc->lock(W) if !setLocks || slowPass
     * tdc->lock(S)
     */
    if (!afs_IsDCacheFresh(tdc, avc) && !overWriteWholeChunk) {
        /*
         * Version number mismatch.
         */
        /*
         * If we are disconnected, then we can't do much of anything
         * because the data doesn't match the file.
         */
        if (AFS_IS_DISCONNECTED) {
            ReleaseSharedLock(&tdc->lock);
            if (setLocks) {
                if (slowPass)
                    ReleaseWriteLock(&avc->lock);
                else
                    ReleaseReadLock(&avc->lock);
            }
            /* Flush the Dcache */
            afs_PutDCache(tdc);

            return NULL;
        }
        UpgradeSToWLock(&tdc->lock, 609);

        /*
         * If data ever existed for this vnode, and this is a text object,
         * do some clearing.  Now, you'd think you need only do the flush
         * when VTEXT is on, but VTEXT is turned off when the text object
         * is freed, while pages are left lying around in memory marked
         * with this vnode.  If we would reactivate (create a new text
         * object from) this vnode, we could easily stumble upon some of
         * these old pages in pagein.  So, we always flush these guys.
         * Sun has a wonderful lack of useful invariants in this system.
         *
         * avc->flushDV is the data version # of the file at the last text
         * flush.  Clearly, at least, we don't have to flush the file more
         * often than it changes
         */
        if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
            /*
             * By here, the cache entry is always write-locked.  We can
             * deadlock if we call osi_Flush with the cache entry locked...
             * Unlock the dcache too.
             */
            ReleaseWriteLock(&tdc->lock);
            if (setLocks && !slowPass)
                ReleaseReadLock(&avc->lock);
            else
                ReleaseWriteLock(&avc->lock);

            osi_FlushText(avc);
            /*
             * Call osi_FlushPages in open, read/write, and map, since it
             * is too hard here to figure out if we should lock the
             * pvnLock.
             */
            if (setLocks && !slowPass)
                ObtainReadLock(&avc->lock);
            else
                ObtainWriteLock(&avc->lock, 66);
            ObtainWriteLock(&tdc->lock, 610);
        }

        /*
         * Locks held:
         * avc->lock(R) if setLocks && !slowPass
         * avc->lock(W) if !setLocks || slowPass
         * tdc->lock(W)
         */

        /* Watch for standard race condition around osi_FlushText */
        if (afs_IsDCacheFresh(tdc, avc)) {
            updateV2DC(setLocks, avc, tdc, 569);        /* set hint */
            afs_stats_cmperf.dcacheHits++;
            ConvertWToSLock(&tdc->lock);
            goto done;
        }

        /* Sleep here when cache needs to be drained. */
        if (setLocks && !slowPass
            && (afs_blocksUsed >
                PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
            /* Make sure truncate daemon is running */
            afs_MaybeWakeupTruncateDaemon();
            ObtainWriteLock(&tdc->tlock, 614);
            tdc->refCount--;    /* we'll re-obtain the dcache when we re-try. */
            ReleaseWriteLock(&tdc->tlock);
            ReleaseWriteLock(&tdc->lock);
            ReleaseReadLock(&avc->lock);
            while ((afs_blocksUsed - afs_blocksDiscarded) >
                   PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
                if (afs_WaitForCacheDrain == 0)
                    afs_WaitForCacheDrainCount++;
                afs_WaitForCacheDrain = 1;
                afs_osi_Sleep(&afs_WaitForCacheDrain);
            }
            afs_MaybeFreeDiscardedDCache();
            /* need to check if someone else got the chunk first. */
            goto RetryGetDCache;
        }

        Position = AFS_CHUNKBASE(abyte);
        if (vType(avc) == VDIR) {
            size = avc->f.m.Length;
            if (size > tdc->f.chunkBytes) {
                /* pre-reserve space for file */
                afs_AdjustSize(tdc, size);
            }
            size = 999999999;   /* max size for transfer */
        } else {
            afs_size_t maxGoodLength;

            /* estimate how much data we're expecting back from the server,
             * and reserve space in the dcache entry for it */

            maxGoodLength = avc->f.m.Length;
            if (avc->f.truncPos < maxGoodLength)
                maxGoodLength = avc->f.truncPos;

            size = AFS_CHUNKSIZE(abyte);        /* expected max size */
            if (Position > maxGoodLength) { /* If we're beyond EOF */
                size = 0;
            } else if (Position + size > maxGoodLength) {
                size = maxGoodLength - Position;
            }
            osi_Assert(size >= 0);

            if (size > tdc->f.chunkBytes) {
                /* pre-reserve estimated space for file */
                afs_AdjustSize(tdc, size);      /* changes chunkBytes */
            }

            if (size) {
                /* For the actual fetch, do not limit the request to the
                 * length of the file. If this results in a read past EOF on
                 * the server, the server will just reply with less data than
                 * requested. If we limit ourselves to only requesting data up
                 * to the avc file length, we open ourselves up to races if the
                 * file is extended on the server at about the same time.
                 *
                 * However, we must restrict ourselves to the avc->f.truncPos
                 * length, since this represents an outstanding local
                 * truncation of the file that will be committed to the
                 * fileserver when we actually write the fileserver contents.
                 * If we do not restrict the fetch length based on
                 * avc->f.truncPos, a different truncate operation extending
                 * the file length could cause the old data after
                 * avc->f.truncPos to reappear, instead of extending the file
                 * with NUL bytes. */
                size = AFS_CHUNKSIZE(abyte);
                if (Position > avc->f.truncPos) {
                    size = 0;
                } else if (Position + size > avc->f.truncPos) {
                    size = avc->f.truncPos - Position;
                }
                osi_Assert(size >= 0);
            }
        }
        if (afs_mariner && !tdc->f.chunk)
            afs_MarinerLog("fetch$Fetching", avc);      /* , Position, size, afs_indexCounter ); */
        /*
         * Right now, we only have one tool, and it's a hammer.  So, we
         * fetch the whole file.
         */
        DZap(tdc);      /* pages in cache may be old */
        file = afs_CFileOpen(&tdc->f.inode);
        if (!file) {
            /* We can't access the file in the disk cache backing this dcache;
             * bail out. */
            ReleaseWriteLock(&tdc->lock);
            afs_PutDCache(tdc);
            tdc = NULL;
            goto done;
        }
        afs_RemoveVCB(&avc->f.fid);
        tdc->f.states |= DWriting;
        tdc->dflags |= DFFetching;
        tdc->validPos = Position;       /*  which is AFS_CHUNKBASE(abyte) */
        if (tdc->mflags & DFFetchReq) {
            tdc->mflags &= ~DFFetchReq;
            if (afs_osi_Wakeup(&tdc->validPos) == 0)
                afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
                           __FILE__, ICL_TYPE_INT32, __LINE__,
                           ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
                           tdc->dflags);
        }
        tsmall = osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
        setVcacheStatus = 0;
#ifndef AFS_NOSTATS
        /*
         * Remember if we are doing the reading from a replicated volume,
         * and how many times we've zipped around the fetch/analyze loop.
         */
        fromReplica = (avc->f.states & CRO) ? 1 : 0;
        numFetchLoops = 0;
        accP = &(afs_stats_cmfullperf.accessinf);
        if (fromReplica)
            (accP->replicatedRefs)++;
        else
            (accP->unreplicatedRefs)++;
#endif /* AFS_NOSTATS */
        /* this is a cache miss */
        afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
                   ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
                   ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);

        if (size)
            afs_stats_cmperf.dcacheMisses++;
        code = 0;
        /*
         * Dynamic root support:  fetch data from local memory.
         */
        if (afs_IsDynroot(avc)) {
            char *dynrootDir;
            int dynrootLen;

            afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);

            dynrootDir += Position;
            dynrootLen -= Position;
            if (size > dynrootLen)
                size = dynrootLen;
            if (size < 0)
                size = 0;
            code = afs_CFileWrite(file, 0, dynrootDir, size);
            afs_PutDynroot();

            if (code == size)
                code = 0;
            else
                code = -1;

            tdc->validPos = Position + size;
            afs_CFileTruncate(file, size);      /* prune it */
        } else if (afs_IsDynrootMount(avc)) {
            char *dynrootDir;
            int dynrootLen;

            afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);

            dynrootDir += Position;
            dynrootLen -= Position;
            if (size > dynrootLen)
                size = dynrootLen;
            if (size < 0)
                size = 0;
            code = afs_CFileWrite(file, 0, dynrootDir, size);
            afs_PutDynroot();

            if (code == size)
                code = 0;
            else
                code = -1;

            tdc->validPos = Position + size;
            afs_CFileTruncate(file, size);      /* prune it */
        } else
            /*
             * Not a dynamic vnode:  do the real fetch.
             */
            do {
                /*
                 * Locks held:
                 * avc->lock(R) if setLocks && !slowPass
                 * avc->lock(W) if !setLocks || slowPass
                 * tdc->lock(W)
                 */

                tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
                if (tc) {
#ifndef AFS_NOSTATS
                    numFetchLoops++;
                    if (fromReplica)
                        (accP->numReplicasAccessed)++;

#endif /* AFS_NOSTATS */
                    if (!setLocks || slowPass) {
                        avc->callback = tc->parent->srvr->server;
                    } else {
                        newCallback = tc->parent->srvr->server;
                        setNewCallback = 1;
                    }
                    i = osi_Time();
                    code = afs_CacheFetchProc(tc, rxconn, file, Position, tdc,
                                               avc, size, tsmall);
                } else
                   code = -1;

                if (code == 0) {
                    /* callback could have been broken (or expired) in a race here,
                     * but we return the data anyway.  It's as good as we knew about
                     * when we started. */
                    /*
                     * validPos is updated by CacheFetchProc, and can only be
                     * modifed under a dcache write lock, which we've blocked out
                     */
                    afs_size_t length;

                    size = tdc->validPos - Position;    /* actual segment size */
                    if (size < 0)
                        size = 0;
                    afs_CFileTruncate(file, size);      /* prune it */

                    /* Check that the amount of data that we fetched for the
                     * dcache makes sense. */
                    FillInt64(length, tsmall->OutStatus.Length_hi, tsmall->OutStatus.Length);
                    if (!IsDCacheSizeOK(tdc, avc, size,
                                        length,
                                        tsmall->OutStatus.DataVersion, 1)) {
                        code = EIO;
                    }
                }
                if (code) {
                    if (!setLocks || slowPass) {
                        afs_StaleVCacheFlags(avc, AFS_STALEVC_CLEARCB, CUnique);
                    } else {
                        /* Something lost.  Forget about performance, and go
                         * back with a vcache write lock.
                         */
                        afs_CFileTruncate(file, 0);
                        afs_AdjustSize(tdc, 0);
                        afs_CFileClose(file);
                        osi_FreeLargeSpace(tsmall);
                        tsmall = 0;
                        ReleaseWriteLock(&tdc->lock);
                        afs_PutDCache(tdc);
                        tdc = 0;

                        /*
                         * Call afs_Analyze to manage the connection references
                         * and handle the error code (possibly mark servers
                         * down, etc). We are going to retry getting the
                         * dcache regardless, so we just ignore the retry hint
                         * returned by afs_Analyze on this call.
                         */
                        (void)afs_Analyze(tc, rxconn, code, &avc->f.fid, areq,
                                          AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL);

                        ReleaseReadLock(&avc->lock);

                        slowPass = 1;
                        goto RetryGetDCache;
                    }
                }

            } while (afs_Analyze
                     (tc, rxconn, code, &avc->f.fid, areq,
                      AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));

        /*
         * Locks held:
         * avc->lock(R) if setLocks && !slowPass
         * avc->lock(W) if !setLocks || slowPass
         * tdc->lock(W)
         */

#ifndef AFS_NOSTATS
        /*
         * In the case of replicated access, jot down info on the number of
         * attempts it took before we got through or gave up.
         */
        if (fromReplica) {
            if (numFetchLoops <= 1)
                (accP->refFirstReplicaOK)++;
            if (numFetchLoops > accP->maxReplicasPerRef)
                accP->maxReplicasPerRef = numFetchLoops;
        }
#endif /* AFS_NOSTATS */

        tdc->dflags &= ~DFFetching;
        if (afs_osi_Wakeup(&tdc->validPos) == 0)
            afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
                       __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
                       tdc, ICL_TYPE_INT32, tdc->dflags);
        if (avc->execsOrWriters == 0)
            tdc->f.states &= ~DWriting;

        /* now, if code != 0, we have an error and should punt.
         * note that we have the vcache write lock, either because
         * !setLocks or slowPass.
         */
        if (code) {
            afs_CFileTruncate(file, 0);
            afs_AdjustSize(tdc, 0);
            afs_CFileClose(file);
            ZapDCE(tdc);        /* sets DFEntryMod */
            if (vType(avc) == VDIR) {
                DZap(tdc);
            }
            tdc->f.states &= ~(DRO|DBackup|DRW);
            afs_DCMoveBucket(tdc, 0, 0);
            ReleaseWriteLock(&tdc->lock);
            afs_PutDCache(tdc);
            if (!afs_IsDynroot(avc)) {
                afs_StaleVCacheFlags(avc, 0, CUnique);
                /*
                 * Locks held:
                 * avc->lock(W); assert(!setLocks || slowPass)
                 */
                osi_Assert(!setLocks || slowPass);
            }
            tdc = NULL;
            goto done;
        }

        /* otherwise we copy in the just-fetched info */
        afs_CFileClose(file);
        afs_AdjustSize(tdc, size);      /* new size */
        /*
         * Copy appropriate fields into vcache.  Status is
         * copied later where we selectively acquire the
         * vcache write lock.
         */
        if (slowPass)
            afs_ProcessFS(avc, &tsmall->OutStatus, areq);
        else
            setVcacheStatus = 1;
        hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
               tsmall->OutStatus.DataVersion);
        tdc->dflags |= DFEntryMod;
        afs_indexFlags[tdc->index] |= IFEverUsed;
        ConvertWToSLock(&tdc->lock);
    } /*Data version numbers don't match */
    else {
        /*
         * Data version numbers match.
         */
        afs_stats_cmperf.dcacheHits++;
    }                           /*Data version numbers match */

    updateV2DC(setLocks, avc, tdc, 335);        /* set hint */
  done:
    /*
     * Locks held:
     * avc->lock(R) if setLocks && !slowPass
     * avc->lock(W) if !setLocks || slowPass
     * tdc->lock(S) if tdc
     */

    /*
     * See if this was a reference to a file in the local cell.
     */
    if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
        afs_stats_cmperf.dlocalAccesses++;
    else
        afs_stats_cmperf.dremoteAccesses++;

    /* Fix up LRU info */

    if (tdc) {
        ObtainWriteLock(&afs_xdcache, 602);
        hset(afs_indexTimes[tdc->index], afs_indexCounter);
        hadd32(afs_indexCounter, 1);
        ReleaseWriteLock(&afs_xdcache);

        /* return the data */
        if (vType(avc) == VDIR)
            *aoffset = abyte;
        else
            *aoffset = AFS_CHUNKOFFSET(abyte);
        *alen = (tdc->f.chunkBytes - *aoffset);
        ReleaseSharedLock(&tdc->lock);
    }

    /*
     * Locks held:
     * avc->lock(R) if setLocks && !slowPass
     * avc->lock(W) if !setLocks || slowPass
     */

    /* Fix up the callback and status values in the vcache */
    doVcacheUpdate = 0;
    if (setLocks && !slowPass) {
        /* DCLOCKXXX
         *
         * This is our dirty little secret to parallel fetches.
         * We don't write-lock the vcache while doing the fetch,
         * but potentially we'll need to update the vcache after
         * the fetch is done.
         *
         * Drop the read lock and try to re-obtain the write
         * lock.  If the vcache still has the same DV, it's
         * ok to go ahead and install the new data.
         */
        afs_hyper_t currentDV, statusDV;

        hset(currentDV, avc->f.m.DataVersion);

        if (setNewCallback && avc->callback != newCallback)
            doVcacheUpdate = 1;

        if (tsmall) {
            hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
                   tsmall->OutStatus.DataVersion);

            if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
                doVcacheUpdate = 1;
            if (setVcacheStatus && !hsame(currentDV, statusDV))
                doVcacheUpdate = 1;
        }

        ReleaseReadLock(&avc->lock);

        if (doVcacheUpdate) {
            ObtainWriteLock(&avc->lock, 615);
            if (!hsame(avc->f.m.DataVersion, currentDV)) {
                /* We lose.  Someone will beat us to it. */
                doVcacheUpdate = 0;
                ReleaseWriteLock(&avc->lock);
            }
        }
    }

    /* With slow pass, we've already done all the updates */
    if (slowPass) {
        ReleaseWriteLock(&avc->lock);
    }

    /* Check if we need to perform any last-minute fixes with a write-lock */
    if (!setLocks || doVcacheUpdate) {
        if (setNewCallback)
            avc->callback = newCallback;
        if (tsmall && setVcacheStatus)
            afs_ProcessFS(avc, &tsmall->OutStatus, areq);
        if (setLocks)
            ReleaseWriteLock(&avc->lock);
    }

    if (tsmall)
        osi_FreeLargeSpace(tsmall);

    return tdc;
}                               /*afs_GetDCache */


/*
 * afs_WriteThroughDSlots
 *
 * Description:
 *      Sweep through the dcache slots and write out any modified
 *      in-memory data back on to our caching store.
 *
 * Parameters:
 *      None.
 *
 * Environment:
 *      The afs_xdcache is write-locked through this whole affair.
 */
int
afs_WriteThroughDSlots(void)
{
    struct dcache *tdc;
    afs_int32 i, touchedit = 0;
    int code = 0;

    struct afs_q DirtyQ, *tq;

    AFS_STATCNT(afs_WriteThroughDSlots);

    /*
     * Because of lock ordering, we can't grab dcache locks while
     * holding afs_xdcache.  So we enter xdcache, get a reference
     * for every dcache entry, and exit xdcache.
     */
    ObtainWriteLock(&afs_xdcache, 283);
    QInit(&DirtyQ);
    for (i = 0; i < afs_cacheFiles; i++) {
        tdc = afs_indexTable[i];

        /* Grab tlock in case the existing refcount isn't zero */
        if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
            ObtainWriteLock(&tdc->tlock, 623);
            tdc->refCount++;
            ReleaseWriteLock(&tdc->tlock);

            QAdd(&DirtyQ, &tdc->dirty);
        }
    }
    ReleaseWriteLock(&afs_xdcache);

    /*
     * Now, for each dcache entry we found, check if it's dirty.
     * If so, get write-lock, get afs_xdcache, which protects
     * afs_cacheInodep, and flush it.  Don't forget to put back
     * the refcounts.
     */

#define DQTODC(q)       ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))

    for (tq = DirtyQ.prev; tq != &DirtyQ && code == 0; tq = QPrev(tq)) {
        tdc = DQTODC(tq);
        if (tdc->dflags & DFEntryMod) {
            int wrLock;

            wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));

            /* Now that we have the write lock, double-check */
            if (wrLock && (tdc->dflags & DFEntryMod)) {
                tdc->dflags &= ~DFEntryMod;
                ObtainWriteLock(&afs_xdcache, 620);
                code = afs_WriteDCache(tdc, 1);
                ReleaseWriteLock(&afs_xdcache);
                if (code) {
                    /* We didn't successfully write out the dslot; make sure we
                     * try again later */
                    tdc->dflags |= DFEntryMod;
                } else {
                    touchedit = 1;
                }
            }
            if (wrLock)
                ReleaseWriteLock(&tdc->lock);
        }

        afs_PutDCache(tdc);
    }

    if (code) {
        return code;
    }

    ObtainWriteLock(&afs_xdcache, 617);
    if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
        /* Touch the file to make sure that the mtime on the file is kept
         * up-to-date to avoid losing cached files on cold starts because
         * their mtime seems old...
         */
        struct afs_fheader theader;

        afs_InitFHeader(&theader);
        afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
    }
    ReleaseWriteLock(&afs_xdcache);
    return 0;
}

/*
 * afs_MemGetDSlot
 *
 * Description:
 *      Return a pointer to an freshly initialized dcache entry using
 *      a memory-based cache.  The tlock will be read-locked.
 *
 * Parameters:
 *      aslot : Dcache slot to look at.
 *      type : What 'type' of dslot to get; see the dslot_state enum
 *
 * Environment:
 *      Must be called with afs_xdcache write-locked.
 */

struct dcache *
afs_MemGetDSlot(afs_int32 aslot, dslot_state type)
{
    struct dcache *tdc;
    int existing = 0;

    AFS_STATCNT(afs_MemGetDSlot);
    if (CheckLock(&afs_xdcache) != -1)
        osi_Panic("getdslot nolock");
    if (aslot < 0 || aslot >= afs_cacheFiles)
        osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
    tdc = afs_indexTable[aslot];
    if (tdc) {
        QRemove(&tdc->lruq);    /* move to queue head */
        QAdd(&afs_DLRU, &tdc->lruq);
        /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
        ObtainWriteLock(&tdc->tlock, 624);
        tdc->refCount++;
        ConvertWToRLock(&tdc->tlock);
        return tdc;
    }

    /* if we got here, the given slot is not in memory in our list of known
     * slots. for memcache, the only place a dslot can exist is in memory, so
     * if the caller is expecting to get back a known dslot, and we've reached
     * here, something is very wrong. DSLOT_NEW is the only type of dslot that
     * may not exist; for all others, the caller assumes the given dslot
     * already exists. so, 'type' had better be DSLOT_NEW here, or something is
     * very wrong. */
    osi_Assert(type == DSLOT_NEW);

    if (!afs_freeDSList)
        afs_GetDownDSlot(4);
    if (!afs_freeDSList) {
        /* none free, making one is better than a panic */
        afs_stats_cmperf.dcacheXAllocs++;       /* count in case we have a leak */
        tdc = afs_osi_Alloc(sizeof(struct dcache));
        osi_Assert(tdc != NULL);
#ifdef  KERNEL_HAVE_PIN
        pin((char *)tdc, sizeof(struct dcache));        /* XXX */
#endif
    } else {
        tdc = afs_freeDSList;
        afs_freeDSList = (struct dcache *)tdc->lruq.next;
        existing = 1;
    }
    tdc->dflags = 0;    /* up-to-date, not in free q */
    tdc->mflags = 0;
    QAdd(&afs_DLRU, &tdc->lruq);
    if (tdc->lruq.prev == &tdc->lruq)
        osi_Panic("lruq 3");

    /* initialize entry */
    tdc->f.fid.Cell = 0;
    tdc->f.fid.Fid.Volume = 0;
    tdc->f.chunk = -1;
    hones(tdc->f.versionNo);
    tdc->f.inode.mem = aslot;
    tdc->dflags |= DFEntryMod;
    tdc->refCount = 1;
    tdc->index = aslot;
    afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;

    if (existing) {
        osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
        osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
        osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
    }

    AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
    AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
    AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
    ObtainReadLock(&tdc->tlock);

    afs_indexTable[aslot] = tdc;
    return tdc;

}                               /*afs_MemGetDSlot */

static void
LogCacheError(int aslot, int off, int code, int target_size)
{
    struct osi_stat tstat;
    char *procname;

    if (afs_osi_Stat(afs_cacheInodep, &tstat)) {
        tstat.size = -1;
    }

    procname = osi_AllocSmallSpace(AFS_SMALLOCSIZ);
    if (procname != NULL) {
        osi_procname(procname, AFS_SMALLOCSIZ);
        procname[AFS_SMALLOCSIZ-1] = '\0';
    }

    afs_warn("afs: disk cache read error in CacheItems slot %d "
             "off %d/%d code %d/%d pid %d (%s)\n",
             aslot, off, (int)tstat.size, code, target_size,
             (int)MyPidxx2Pid(MyPidxx),
             procname ? procname : "");

    if (procname != NULL) {
        osi_FreeSmallSpace(procname);
        procname = NULL;
    }
}

unsigned int last_error = 0, lasterrtime = 0;

/*
 * afs_UFSGetDSlot
 *
 * Description:
 *      Return a pointer to an freshly initialized dcache entry using
 *      a UFS-based disk cache.  The dcache tlock will be read-locked.
 *
 * Parameters:
 *      aslot : Dcache slot to look at.
 *      type : What 'type' of dslot to get; see the dslot_state enum
 *
 * Environment:
 *      afs_xdcache lock write-locked.
 */
struct dcache *
afs_UFSGetDSlot(afs_int32 aslot, dslot_state type)
{
    afs_int32 code;
    struct dcache *tdc;
    int existing = 0;
    int entryok;
    int off;

    AFS_STATCNT(afs_UFSGetDSlot);
    if (CheckLock(&afs_xdcache) != -1)
        osi_Panic("getdslot nolock");
    if (aslot < 0 || aslot >= afs_cacheFiles)
        osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
    tdc = afs_indexTable[aslot];
    if (tdc) {
        QRemove(&tdc->lruq);    /* move to queue head */
        QAdd(&afs_DLRU, &tdc->lruq);
        /* Grab tlock in case refCount != 0 */
        ObtainWriteLock(&tdc->tlock, 625);
        tdc->refCount++;
        ConvertWToRLock(&tdc->tlock);
        return tdc;
    }

    /* otherwise we should read it in from the cache file */
    if (!afs_freeDSList)
        afs_GetDownDSlot(4);
    if (!afs_freeDSList) {
        /* none free, making one is better than a panic */
        afs_stats_cmperf.dcacheXAllocs++;       /* count in case we have a leak */
        tdc = afs_osi_Alloc(sizeof(struct dcache));
        osi_Assert(tdc != NULL);
#ifdef  KERNEL_HAVE_PIN
        pin((char *)tdc, sizeof(struct dcache));        /* XXX */
#endif
    } else {
        tdc = afs_freeDSList;
        afs_freeDSList = (struct dcache *)tdc->lruq.next;
        existing = 1;
    }
    tdc->dflags = 0;    /* up-to-date, not in free q */
    tdc->mflags = 0;
    QAdd(&afs_DLRU, &tdc->lruq);
    if (tdc->lruq.prev == &tdc->lruq)
        osi_Panic("lruq 3");

    /*
     * Seek to the aslot'th entry and read it in.
     */
    off = sizeof(struct fcache)*aslot + sizeof(struct afs_fheader);
    code =
        afs_osi_Read(afs_cacheInodep,
                     off, (char *)(&tdc->f),
                     sizeof(struct fcache));
    entryok = 1;
    if (code != sizeof(struct fcache)) {
        entryok = 0;
#if defined(KERNEL_HAVE_UERROR)
        last_error = getuerror();
#else
        last_error = code;
#endif
        lasterrtime = osi_Time();
        if (type != DSLOT_NEW) {
            /* If we are requesting a non-DSLOT_NEW slot, this is an error.
             * non-DSLOT_NEW slots are supposed to already exist, so if we
             * failed to read in the slot, something is wrong. */
            LogCacheError(aslot, off, code, sizeof(struct fcache));

            /* put tdc back on the free dslot list */
            QRemove(&tdc->lruq);
            tdc->index = NULLIDX;
            tdc->lruq.next = (struct afs_q *)afs_freeDSList;
            afs_freeDSList = tdc;
            return NULL;
        }
    }
    if (!afs_CellNumValid(tdc->f.fid.Cell)) {
        entryok = 0;
        if (type == DSLOT_VALID) {
            osi_Panic("afs: needed valid dcache but index %d off %d has "
                      "invalid cell num %d\n",
                      (int)aslot, off, (int)tdc->f.fid.Cell);
        }
    }

    if (type == DSLOT_VALID && tdc->f.fid.Fid.Volume == 0) {
        osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
                  (int)aslot, off);
    }

    if (type == DSLOT_UNUSED) {
        /* the requested dslot is known to exist, but contain invalid data
         * (this happens when we're using a dslot from the free or discard
         * list). be sure not to re-use the data in it, so force invalidation.
         */
        entryok = 0;
    }

    if (!entryok) {
        tdc->f.fid.Cell = 0;
        tdc->f.fid.Fid.Volume = 0;
        tdc->f.chunk = -1;
        hones(tdc->f.versionNo);
        tdc->dflags |= DFEntryMod;
        afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
        tdc->f.states &= ~(DRO|DBackup|DRW);
        afs_DCMoveBucket(tdc, 0, 0);
    } else {
        if (tdc->f.states & DRO) {
            afs_DCMoveBucket(tdc, 0, 2);
        } else if (tdc->f.states & DBackup) {
            afs_DCMoveBucket(tdc, 0, 1);
        } else {
            afs_DCMoveBucket(tdc, 0, 1);
        }
    }
    tdc->refCount = 1;
    tdc->index = aslot;
    if (tdc->f.chunk >= 0)
        tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
    else
        tdc->validPos = 0;

    if (existing) {
        osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
        osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
        osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
    }

    AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
    AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
    AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
    ObtainReadLock(&tdc->tlock);

    /*
     * If we didn't read into a temporary dcache region, update the
     * slot pointer table.
     */
    afs_indexTable[aslot] = tdc;
    return tdc;

}                               /*afs_UFSGetDSlot */



/*!
 * Write a particular dcache entry back to its home in the
 * CacheInfo file.
 *
 * \param adc Pointer to the dcache entry to write.
 * \param atime If true, set the modtime on the file to the current time.
 *
 * \note Environment:
 *      Must be called with the afs_xdcache lock at least read-locked,
 *      and dcache entry at least read-locked.
 *      The reference count is not changed.
 */

int
afs_WriteDCache(struct dcache *adc, int atime)
{
    afs_int32 code;

    if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
        return 0;
    AFS_STATCNT(afs_WriteDCache);
    osi_Assert(WriteLocked(&afs_xdcache));
    if (atime)
        adc->f.modTime = osi_Time();

    if ((afs_indexFlags[adc->index] & (IFFree | IFDiscarded)) == 0 &&
        adc->f.fid.Fid.Volume == 0) {
        /* If a dcache slot is not on the free or discard list, it must be
         * in the hash table. Thus, the volume must be non-zero, since that
         * is how we determine whether or not to unhash the entry when kicking
         * it out of the cache. Do this check now, since otherwise this can
         * cause hash table corruption and a panic later on after we read the
         * entry back in. */
        osi_Panic("afs_WriteDCache zero volume index %d flags 0x%x\n",
                  adc->index, (unsigned)afs_indexFlags[adc->index]);
    }

    /*
     * Seek to the right dcache slot and write the in-memory image out to disk.
     */
    afs_cellname_write();
    code =
        afs_osi_Write(afs_cacheInodep,
                      sizeof(struct fcache) * adc->index +
                      sizeof(struct afs_fheader), (char *)(&adc->f),
                      sizeof(struct fcache));
    if (code != sizeof(struct fcache)) {
        afs_warn("afs: failed to write to CacheItems off %ld code %d/%d\n",
                 (long)(sizeof(struct fcache) * adc->index + sizeof(struct afs_fheader)),
                 (int)code, (int)sizeof(struct fcache));
        return EIO;
    }
    return 0;
}



/*!
 * Wake up users of a particular file waiting for stores to take
 * place.
 *
 * \param avc Ptr to related vcache entry.
 *
 * \note Environment:
 *      Nothing interesting.
 */
int
afs_wakeup(struct vcache *avc)
{
    int i;
    struct brequest *tb;
    tb = afs_brs;
    AFS_STATCNT(afs_wakeup);
    for (i = 0; i < NBRS; i++, tb++) {
        /* if request is valid and for this file, we've found it */
        if (tb->refCount > 0 && avc == tb->vc) {

            /*
             * If CSafeStore is on, then we don't awaken the guy
             * waiting for the store until the whole store has finished.
             * Otherwise, we do it now.  Note that if CSafeStore is on,
             * the BStore routine actually wakes up the user, instead
             * of us.
             * I think this is redundant now because this sort of thing
             * is already being handled by the higher-level code.
             */
            if ((avc->f.states & CSafeStore) == 0) {
                tb->code_raw = tb->code_checkcode = 0;
                tb->flags |= BUVALID;
                if (tb->flags & BUWAIT) {
                    tb->flags &= ~BUWAIT;
                    afs_osi_Wakeup(tb);
                }
            }
            break;
        }
    }
    return 0;
}

/*!
 * Given a file name and inode, set up that file to be an
 * active member in the AFS cache.  This also involves checking
 * the usability of its data.
 *
 * \param afile Name of the cache file to initialize.
 * \param ainode Inode of the file.
 *
 * \note Environment:
 *      This function is called only during initialization.
 */
int
afs_InitCacheFile(char *afile, ino_t ainode)
{
    afs_int32 code;
    afs_int32 index;
    int fileIsBad;
    struct osi_file *tfile;
    struct osi_stat tstat;
    struct dcache *tdc;

    AFS_STATCNT(afs_InitCacheFile);
    index = afs_stats_cmperf.cacheNumEntries;
    if (index >= afs_cacheFiles)
        return EINVAL;

    ObtainWriteLock(&afs_xdcache, 282);
    tdc = afs_GetNewDSlot(index);
    ReleaseReadLock(&tdc->tlock);
    ReleaseWriteLock(&afs_xdcache);

    ObtainWriteLock(&tdc->lock, 621);
    ObtainWriteLock(&afs_xdcache, 622);
    if (!afile && !ainode) {
        tfile = NULL;
        fileIsBad = 1;
    } else {
        if (afile) {
            code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
            if (code) {
                ReleaseWriteLock(&afs_xdcache);
                ReleaseWriteLock(&tdc->lock);
                afs_PutDCache(tdc);
                return code;
            }
        } else {
            /* Add any other 'complex' inode types here ... */
#if !defined(AFS_LINUX26_ENV) && !defined(AFS_CACHE_VNODE_PATH)
            tdc->f.inode.ufs = ainode;
#else
            osi_Panic("Can't init cache with inode numbers when complex inodes are "
                      "in use\n");
#endif
        }
        fileIsBad = 0;
        if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
            fileIsBad = 1;
        tfile = osi_UFSOpen(&tdc->f.inode);
        if (!tfile) {
            ReleaseWriteLock(&afs_xdcache);
            ReleaseWriteLock(&tdc->lock);
            afs_PutDCache(tdc);
            return ENOENT;
        }

        code = afs_osi_Stat(tfile, &tstat);
        if (code)
            osi_Panic("initcachefile stat");

        /*
         * If file size doesn't match the cache info file, it's probably bad.
         */
        if (tdc->f.chunkBytes != tstat.size)
            fileIsBad = 1;
        /*
         * If file changed within T (120?) seconds of cache info file, it's
         * probably bad.  In addition, if slot changed within last T seconds,
         * the cache info file may be incorrectly identified, and so slot
         * may be bad.
         */
        if (cacheInfoModTime < tstat.mtime + 120)
            fileIsBad = 1;
        if (cacheInfoModTime < tdc->f.modTime + 120)
            fileIsBad = 1;
        /* In case write through is behind, make sure cache items entry is
         * at least as new as the chunk.
         */
        if (tdc->f.modTime < tstat.mtime)
            fileIsBad = 1;
    }
    tdc->f.chunkBytes = 0;

    if (fileIsBad) {
        tdc->f.fid.Fid.Volume = 0;      /* not in the hash table */
        if (tfile && tstat.size != 0)
            osi_UFSTruncate(tfile, 0);
        tdc->f.states &= ~(DRO|DBackup|DRW);
        afs_DCMoveBucket(tdc, 0, 0);
        /* put entry in free cache slot list */
        afs_dvnextTbl[tdc->index] = afs_freeDCList;
        afs_freeDCList = index;
        afs_freeDCCount++;
        afs_indexFlags[index] |= IFFree;
        afs_indexUnique[index] = 0;
    } else {
        /*
         * We must put this entry in the appropriate hash tables.
         * Note that i is still set from the above DCHash call
         */
        code = DCHash(&tdc->f.fid, tdc->f.chunk);
        afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
        afs_dchashTbl[code] = tdc->index;
        code = DVHash(&tdc->f.fid);
        afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
        afs_dvhashTbl[code] = tdc->index;
        afs_AdjustSize(tdc, tstat.size);        /* adjust to new size */
        if (tstat.size > 0)
            /* has nontrivial amt of data */
            afs_indexFlags[index] |= IFEverUsed;
        afs_stats_cmperf.cacheFilesReused++;
        /*
         * Initialize index times to file's mod times; init indexCounter
         * to max thereof
         */
        hset32(afs_indexTimes[index], tstat.atime);
        if (hgetlo(afs_indexCounter) < tstat.atime) {
            hset32(afs_indexCounter, tstat.atime);
        }
        afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
    }                           /*File is not bad */

    if (tfile)
        osi_UFSClose(tfile);
    tdc->f.states &= ~DWriting;
    tdc->dflags &= ~DFEntryMod;
    /* don't set f.modTime; we're just cleaning up */
    osi_Assert(afs_WriteDCache(tdc, 0) == 0);
    ReleaseWriteLock(&afs_xdcache);
    ReleaseWriteLock(&tdc->lock);
    afs_PutDCache(tdc);
    afs_stats_cmperf.cacheNumEntries++;
    return 0;
}


/*Max # of struct dcache's resident at any time*/
/*
 * If 'dchint' is enabled then in-memory dcache min is increased because of
 * crashes...
 */
#define DDSIZE 200

/*!
 * Initialize dcache related variables.
 *
 * \param afiles
 * \param ablocks
 * \param aDentries
 * \param achunk
 * \param aflags
 *
 */
int
afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
{
    struct dcache *tdp;
    int i;
    int code;
    int afs_dhashbits;

    afs_freeDCList = NULLIDX;
    afs_discardDCList = NULLIDX;
    afs_freeDCCount = 0;
    afs_freeDSList = NULL;
    hzero(afs_indexCounter);

    LOCK_INIT(&afs_xdcache, "afs_xdcache");

    /*
     * Set chunk size
     */
    if (achunk) {
        if (achunk < 0 || achunk > 30)
            achunk = 13;        /* Use default */
        AFS_SETCHUNKSIZE(achunk);
    }

    if (!aDentries)
        aDentries = DDSIZE;

    /* afs_dhashsize defaults to 1024 */
    if (aDentries > 512)
        afs_dhashsize = 2048;
    /* Try to keep the average chain length around two unless the table
     * would be ridiculously big. */
    if (aDentries > 4096) {
        afs_dhashbits = opr_fls(aDentries) - 3;
        /* Cap the hash tables to 32k entries. */
        if (afs_dhashbits > 15)
            afs_dhashbits = 15;
        afs_dhashsize = opr_jhash_size(afs_dhashbits);
    }
    /* initialize hash tables */
    afs_dvhashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
    osi_Assert(afs_dvhashTbl != NULL);
    afs_dchashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
    osi_Assert(afs_dchashTbl != NULL);
    for (i = 0; i < afs_dhashsize; i++) {
        afs_dvhashTbl[i] = NULLIDX;
        afs_dchashTbl[i] = NULLIDX;
    }
    afs_dvnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
    osi_Assert(afs_dvnextTbl != NULL);
    afs_dcnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
    osi_Assert(afs_dcnextTbl != NULL);
    for (i = 0; i < afiles; i++) {
        afs_dvnextTbl[i] = NULLIDX;
        afs_dcnextTbl[i] = NULLIDX;
    }

    /* Allocate and zero the pointer array to the dcache entries */
    afs_indexTable = afs_osi_Alloc(sizeof(struct dcache *) * afiles);
    osi_Assert(afs_indexTable != NULL);
    memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
    afs_indexTimes = afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
    osi_Assert(afs_indexTimes != NULL);
    memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
    afs_indexUnique = afs_osi_Alloc(afiles * sizeof(afs_uint32));
    osi_Assert(afs_indexUnique != NULL);
    memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
    afs_indexFlags = afs_osi_Alloc(afiles * sizeof(u_char));
    osi_Assert(afs_indexFlags != NULL);
    memset(afs_indexFlags, 0, afiles * sizeof(char));

    /* Allocate and thread the struct dcache entries themselves */
    tdp = afs_Initial_freeDSList =
        afs_osi_Alloc(aDentries * sizeof(struct dcache));
    osi_Assert(tdp != NULL);
    memset(tdp, 0, aDentries * sizeof(struct dcache));
#ifdef  KERNEL_HAVE_PIN
    pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles);      /* XXX */
    pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles);  /* XXX */
    pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
    pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles);   /* XXX */
    pin((char *)tdp, aDentries * sizeof(struct dcache));        /* XXX */
    pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize);      /* XXX */
    pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize);      /* XXX */
    pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles);     /* XXX */
    pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles);     /* XXX */
#endif

    afs_freeDSList = &tdp[0];
    for (i = 0; i < aDentries - 1; i++) {
        tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
        AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
        AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
        AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
    }
    tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
    AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
    AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
    AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");

    afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
        afs_cacheBlocks = ablocks;
    afs_ComputeCacheParms();    /* compute parms based on cache size */

    afs_dcentries = aDentries;
    afs_blocksUsed = 0;
    afs_stats_cmperf.cacheBucket0_Discarded =
        afs_stats_cmperf.cacheBucket1_Discarded =
        afs_stats_cmperf.cacheBucket2_Discarded = 0;
    afs_DCSizeInit();
    QInit(&afs_DLRU);

    if (aflags & AFSCALL_INIT_MEMCACHE) {
        /*
         * Use a memory cache instead of a disk cache
         */
        cacheDiskType = AFS_FCACHE_TYPE_MEM;
        afs_cacheType = &afs_MemCacheOps;
        afiles = (afiles < aDentries) ? afiles : aDentries;     /* min */
        ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
        /* ablocks is reported in 1K blocks */
        code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
        if (code != 0) {
            afs_warn("afsd: memory cache too large for available memory.\n");
            afs_warn("afsd: AFS files cannot be accessed.\n\n");
            dcacheDisabled = 1;
            return code;
        } else
            afs_warn("Memory cache: Allocating %d dcache entries...",
                   aDentries);
    } else {
        cacheDiskType = AFS_FCACHE_TYPE_UFS;
        afs_cacheType = &afs_UfsCacheOps;
    }
    return 0;
}

/*!
 * Shuts down the cache.
 *
 */
void
shutdown_dcache(void)
{
    int i;

#ifdef AFS_CACHE_VNODE_PATH
    if (cacheDiskType != AFS_FCACHE_TYPE_MEM) {
        struct dcache *tdc;
        for (i = 0; i < afs_cacheFiles; i++) {
            tdc = afs_indexTable[i];
            if (tdc) {
                afs_osi_FreeStr(tdc->f.inode.ufs);
            }
        }
    }
#endif

    afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
    afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
    afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
    afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
    afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
    afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
    afs_osi_Free(afs_Initial_freeDSList,
                 afs_dcentries * sizeof(struct dcache));
#ifdef  KERNEL_HAVE_PIN
    unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
    unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
    unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
    unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
    unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
    unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
    unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
#endif


    for (i = 0; i < afs_dhashsize; i++) {
        afs_dvhashTbl[i] = NULLIDX;
        afs_dchashTbl[i] = NULLIDX;
    }

    afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
    afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));

    afs_blocksUsed = afs_dcentries = 0;
    afs_stats_cmperf.cacheBucket0_Discarded =
        afs_stats_cmperf.cacheBucket1_Discarded =
        afs_stats_cmperf.cacheBucket2_Discarded = 0;
    hzero(afs_indexCounter);

    afs_freeDCCount = 0;
    afs_freeDCList = NULLIDX;
    afs_discardDCList = NULLIDX;
    afs_freeDSList = afs_Initial_freeDSList = 0;

    LOCK_INIT(&afs_xdcache, "afs_xdcache");
    QInit(&afs_DLRU);

}

/*!
 * Get a dcache ready for writing, respecting the current cache size limits
 *
 * len is required because afs_GetDCache with flag == 4 expects the length
 * field to be filled. It decides from this whether it's necessary to fetch
 * data into the chunk before writing or not (when the whole chunk is
 * overwritten!).
 *
 * \param avc           The vcache to fetch a dcache for
 * \param filePos       The start of the section to be written
 * \param len           The length of the section to be written
 * \param areq
 * \param noLock
 *
 * \return If successful, a reference counted dcache with tdc->lock held. Lock
 *         must be released and afs_PutDCache() called to free dcache.
 *         NULL on  failure
 *
 * \note avc->lock must be held on entry. Function may release and reobtain
 *       avc->lock and GLOCK.
 */

struct dcache *
afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
                           afs_size_t len, struct vrequest *areq,
                           int noLock)
{
    struct dcache *tdc = NULL;
    afs_size_t offset;

    /* read the cached info */
    if (noLock) {
        tdc = afs_FindDCache(avc, filePos);
        if (tdc)
            ObtainWriteLock(&tdc->lock, 657);
    } else if (afs_blocksUsed >
               PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
        tdc = afs_FindDCache(avc, filePos);
        if (tdc) {
            ObtainWriteLock(&tdc->lock, 658);
            if (!afs_IsDCacheFresh(tdc, avc)
                || (tdc->dflags & DFFetching)) {
                ReleaseWriteLock(&tdc->lock);
                afs_PutDCache(tdc);
                tdc = NULL;
            }
        }
        if (!tdc) {
            afs_MaybeWakeupTruncateDaemon();
            while (afs_blocksUsed >
                   PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
                ReleaseWriteLock(&avc->lock);
                if (afs_blocksUsed - afs_blocksDiscarded >
                    PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
                    if (afs_WaitForCacheDrain == 0)
                        afs_WaitForCacheDrainCount++;
                    afs_WaitForCacheDrain = 1;
                    afs_osi_Sleep(&afs_WaitForCacheDrain);
                }
                afs_MaybeFreeDiscardedDCache();
                afs_MaybeWakeupTruncateDaemon();
                ObtainWriteLock(&avc->lock, 509);
            }
            avc->f.states |= CDirty;
            tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
            if (tdc)
                ObtainWriteLock(&tdc->lock, 659);
        }
    } else {
        tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
        if (tdc)
            ObtainWriteLock(&tdc->lock, 660);
    }
    if (tdc) {
        if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
            afs_stats_cmperf.cacheCurrDirtyChunks++;
            afs_indexFlags[tdc->index] |= IFDataMod;    /* so it doesn't disappear */
        }
        if (!(tdc->f.states & DWriting)) {
            /* don't mark entry as mod if we don't have to */
            tdc->f.states |= DWriting;
            tdc->dflags |= DFEntryMod;
        }
    }
    return tdc;
}

/*!
 * Make a shadow copy of a dir's dcache. It's used for disconnected
 * operations like remove/create/rename to keep the original directory data.
 * On reconnection, we can diff the original data with the server and get the
 * server changes and with the local data to get the local changes.
 *
 * \param avc The dir vnode.
 * \param adc The dir dcache.
 *
 * \return 0 for success.
 *
 * \note The vcache entry must be write locked.
 * \note The dcache entry must be read locked.
 */
int
afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
{
    int i, code, ret_code = 0, written, trans_size;
    struct dcache *new_dc = NULL;
    struct osi_file *tfile_src = NULL, *tfile_dst = NULL;
    struct VenusFid shadow_fid;
    char *data = NULL;

    /* Is this a dir? */
    if (vType(avc) != VDIR)
        return ENOTDIR;

    if (avc->f.shadow.vnode || avc->f.shadow.unique)
        return EEXIST;

    /* Generate a fid for the shadow dir. */
    shadow_fid.Cell = avc->f.fid.Cell;
    shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
    afs_GenShadowFid(&shadow_fid);

    ObtainWriteLock(&afs_xdcache, 716);

    /* Get a fresh dcache. */
    (void)afs_AllocDCache(&new_dc, avc, 0, 0, &shadow_fid);
    osi_Assert(new_dc);

    ObtainReadLock(&adc->mflock);

    /* Set up the new fid. */
    /* Copy interesting data from original dir dcache. */
    new_dc->mflags = adc->mflags;
    new_dc->dflags = adc->dflags;
    new_dc->f.modTime = adc->f.modTime;
    new_dc->f.versionNo = adc->f.versionNo;
    new_dc->f.states = adc->f.states;
    new_dc->f.chunk= adc->f.chunk;
    new_dc->f.chunkBytes = adc->f.chunkBytes;

    ReleaseReadLock(&adc->mflock);

    /* Now add to the two hash chains */
    i = DCHash(&shadow_fid, 0);
    afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
    afs_dchashTbl[i] = new_dc->index;

    i = DVHash(&shadow_fid);
    afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
    afs_dvhashTbl[i] = new_dc->index;

    ReleaseWriteLock(&afs_xdcache);

    /* Alloc a 4k block. */
    data = afs_osi_Alloc(4096);
    if (!data) {
        afs_warn("afs_MakeShadowDir: could not alloc data\n");
        ret_code = ENOMEM;
        goto done;
    }

    /* Open the files. */
    tfile_src = afs_CFileOpen(&adc->f.inode);
    if (!tfile_src) {
        ret_code = EIO;
        goto done;
    }

    tfile_dst = afs_CFileOpen(&new_dc->f.inode);
    if (!tfile_dst) {
        ret_code = EIO;
        goto done;
    }

    /* And now copy dir dcache data into this dcache,
     * 4k at a time.
     */
    written = 0;
    while (written < adc->f.chunkBytes) {
        trans_size = adc->f.chunkBytes - written;
        if (trans_size > 4096)
            trans_size = 4096;

        /* Read a chunk from the dcache. */
        code = afs_CFileRead(tfile_src, written, data, trans_size);
        if (code < trans_size) {
            ret_code = EIO;
            break;
        }

        /* Write it to the new dcache. */
        code = afs_CFileWrite(tfile_dst, written, data, trans_size);
        if (code < trans_size) {
            ret_code = EIO;
            break;
        }

        written+=trans_size;
    }

 done:
    if (tfile_dst)
        afs_CFileClose(tfile_dst);
    if (tfile_src)
        afs_CFileClose(tfile_src);

    if (data)
        afs_osi_Free(data, 4096);

    ReleaseWriteLock(&new_dc->lock);
    afs_PutDCache(new_dc);

    if (!ret_code) {
        ObtainWriteLock(&afs_xvcache, 763);
        ObtainWriteLock(&afs_disconDirtyLock, 765);
        QAdd(&afs_disconShadow, &avc->shadowq);
        osi_Assert((afs_RefVCache(avc) == 0));
        ReleaseWriteLock(&afs_disconDirtyLock);
        ReleaseWriteLock(&afs_xvcache);

        avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
        avc->f.shadow.unique = shadow_fid.Fid.Unique;
    }

    return ret_code;
}

/*!
 * Delete the dcaches of a shadow dir.
 *
 * \param avc The vcache containing the shadow fid.
 *
 * \note avc must be write locked.
 */
void
afs_DeleteShadowDir(struct vcache *avc)
{
    struct dcache *tdc;
    struct VenusFid shadow_fid;

    shadow_fid.Cell = avc->f.fid.Cell;
    shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
    shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
    shadow_fid.Fid.Unique = avc->f.shadow.unique;

    tdc = afs_FindDCacheByFid(&shadow_fid);
    if (tdc) {
        afs_HashOutDCache(tdc, 1);
        afs_DiscardDCache(tdc);
        afs_PutDCache(tdc);
    }
    avc->f.shadow.vnode = avc->f.shadow.unique = 0;
    ObtainWriteLock(&afs_disconDirtyLock, 708);
    QRemove(&avc->shadowq);
    ReleaseWriteLock(&afs_disconDirtyLock);
    afs_PutVCache(avc); /* Because we held it when we added to the queue */
}

/*!
 * Populate a dcache with empty chunks up to a given file size,
 * used before extending a file in order to avoid 'holes' which
 * we can't access in disconnected mode.
 *
 * \param avc   The vcache which is being extended (locked)
 * \param alen  The new length of the file
 *
 */
void
afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
{
    struct dcache *tdc;
    afs_size_t len, offset;
    afs_int32 start, end;

    /* We're doing this to deal with the situation where we extend
     * by writing after lseek()ing past the end of the file . If that
     * extension skips chunks, then those chunks won't be created, and
     * GetDCache will assume that they have to be fetched from the server.
     * So, for each chunk between the current file position, and the new
     * length we GetDCache for that chunk.
     */

    if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
        return;

    if (avc->f.m.Length == 0)
        start = 0;
    else
        start = AFS_CHUNK(avc->f.m.Length)+1;

    end = AFS_CHUNK(apos);

    while (start<end) {
        len = AFS_CHUNKTOSIZE(start);
        tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);
        if (tdc)
            afs_PutDCache(tdc);
        start++;
    }
}