afs: Skip bulkstat if stat cache looks full

[openafs.git] / src / afs / VNOPS / afs_vnop_lookup.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:
 * afs_lookup
 * EvalMountPoint
 * afs_DoBulkStat
 */

#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/nfsclient.h"
#include "afs/exporter.h"
#include "afs/afs_osidnlc.h"
#include "afs/afs_dynroot.h"

extern struct vcache *afs_globalVp;

afs_int32 afs_bkvolpref = 0;
afs_int32 afs_bulkStatsDone;
static int bulkStatCounter = 0; /* counter for bulk stat seq. numbers */
int afs_fakestat_enable = 0;    /* 1: fakestat-all, 2: fakestat-crosscell */


/* this would be faster if it did comparison as int32word, but would be 
 * dependant on byte-order and alignment, and I haven't figured out
 * what "@sys" is in binary... */
#define AFS_EQ_ATSYS(name) (((name)[0]=='@')&&((name)[1]=='s')&&((name)[2]=='y')&&((name)[3]=='s')&&(!(name)[4]))

/* call under write lock, evaluate mvid.target_root field from a mt pt.
 * avc is the vnode of the mount point object; must be write-locked.
 * advc is the vnode of the containing directory (optional; if NULL and
 *   EvalMountPoint succeeds, caller must initialize *avolpp->dotdot)
 * avolpp is where we return a pointer to the volume named by the mount pt, if success
 * areq is the identity of the caller.
 *
 * NOTE: this function returns a held volume structure in *volpp if it returns 0!
 */
static int
EvalMountData(char type, char *data, afs_uint32 states, afs_uint32 cellnum,
              struct volume **avolpp, struct vrequest *areq,
              afs_uint32 *acellidxp, afs_uint32 *avolnump,
              afs_uint32 *avnoidp, afs_uint32 *auniqp)
{
    struct volume *tvp = 0;
    struct VenusFid tfid;
    struct cell *tcell;
    char *cpos, *volnamep = NULL;
    char *endptr;
    afs_int32 prefetch;         /* 1=>None  2=>RO  3=>BK */
    afs_int32 mtptCell, assocCell = 0, hac = 0;
    afs_int32 samecell, roname, len;
    afs_uint32 volid = 0, cellidx, vnoid = 0, uniq = 0;

    /* Start by figuring out and finding the cell */
    cpos = afs_strchr(data, ':');       /* if cell name present */
    if (cpos) {
        afs_uint32 mtptCellnum;
        volnamep = cpos + 1;
        *cpos = 0;
        if ((afs_strtoi_r(data, &endptr, &mtptCellnum) == 0) &&
            (endptr == cpos)) {
            tcell = afs_GetCell(mtptCellnum, READ_LOCK);
        } else {
            tcell = afs_GetCellByName(data, READ_LOCK);
        }
        *cpos = ':';
    } else if (cellnum) {
        volnamep = data;
        tcell = afs_GetCell(cellnum, READ_LOCK);
    } else {
        /* No cellname or cellnum; return ENODEV */
        return ENODEV;
    }
    if (!tcell) {
        /* no cell found; return ENODEV */
        return ENODEV;
    }

    cellidx = tcell->cellIndex;
    mtptCell = tcell->cellNum;  /* The cell for the mountpoint */
    if (tcell->lcellp) {
        hac = 1;                /* has associated cell */
        assocCell = tcell->lcellp->cellNum;     /* The associated cell */
    }
    afs_PutCell(tcell, READ_LOCK);

    /* If there's nothing to look up, we can't proceed */
    if (!*volnamep)
        return ENODEV;

    /* cell found. figure out volume */
    cpos = afs_strchr(volnamep, ':');
    if (cpos)
        *cpos = 0;

    /* Look for an all-numeric volume ID */
    if ((afs_strtoi_r(volnamep, &endptr, &volid) == 0) &&
        ((endptr == cpos) || (!*endptr)))
    {
        /* Ok. Is there a vnode and uniq? */
        if (cpos) {
            char *vnodep = (char *)(cpos + 1);
            char *uniqp = NULL;
            if ((!*vnodep) /* no vnode after colon */
                || !(uniqp = afs_strchr(vnodep, ':')) /* no colon for uniq */
                || (!*(++uniqp)) /* no uniq after colon */
                || (afs_strtoi_r(vnodep, &endptr, &vnoid) != 0) /* bad vno */
                || (*endptr != ':') /* bad vnode field */
                || (afs_strtoi_r(uniqp, &endptr, &uniq) != 0) /* bad uniq */
                || (*endptr)) /* anything after uniq */
            {
                *cpos = ':';
                /* sorry. vnode and uniq, or nothing */
                return ENODEV;
            }
        }
    } else
            volid = 0;

    /*
     * If the volume ID was all-numeric, and they didn't ask for a
     * pointer to the volume structure, then just return the number
     * as-is.  This is currently only used for handling name lookups
     * in the dynamic mount directory.
     */
    if (volid && !avolpp) {
        if (cpos)
            *cpos = ':';
        goto done;
    }

    /*
     * If the volume ID was all-numeric, and the type was '%', then
     * assume whoever made the mount point knew what they were doing,
     * and don't second-guess them by forcing use of a RW volume when
     * they gave the ID of something else.
     */
    if (volid && type == '%') {
        tfid.Fid.Volume = volid;        /* remember BK volume */
        tfid.Cell = mtptCell;
        tvp = afs_GetVolume(&tfid, areq, WRITE_LOCK);   /* get the new one */
        if (cpos) /* one way or another we're done */
            *cpos = ':';
        if (!tvp)
            return ENODEV; /* afs_GetVolume failed; return ENODEV */
        goto done;
    }

    /* Is volume name a "<n>.backup" or "<n>.readonly" name */
    len = strlen(volnamep);
    roname = ((len > 9) && (strcmp(&volnamep[len - 9], ".readonly") == 0))
        || ((len > 7) && (strcmp(&volnamep[len - 7], ".backup") == 0));

    /* When we cross mountpoint, do we stay in the same cell */
    samecell = (cellnum == mtptCell) || (hac && (cellnum == assocCell));

    /* Decide whether to prefetch the BK, or RO.  Also means we want the BK or
     * RO.
     * If this is a regular mountpoint with a RW volume name
     * - If BK preference is enabled AND we remain within the same cell AND
     *   start from a BK volume, then we will want to prefetch the BK volume.
     * - If we cross a cell boundary OR start from a RO volume, then we will
     *   want to prefetch the RO volume.
     */
    if ((type == '#') && !roname) {
        if (afs_bkvolpref && samecell && (states & CBackup))
            prefetch = 3;       /* Prefetch the BK */
        else if (!samecell || (states & CRO))
            prefetch = 2;       /* Prefetch the RO */
        else
            prefetch = 1;       /* Do not prefetch */
    } else {
        prefetch = 1;           /* Do not prefetch */
    }

    /* Get the volume struct. Unless this volume name has ".readonly" or
     * ".backup" in it, this will get the volume struct for the RW volume.
     * The RO volume will be prefetched if requested (but not returned).
     * Set up to use volname first.
     */
    tvp = afs_GetVolumeByName(volnamep, mtptCell, prefetch, areq, WRITE_LOCK);

    /* If no volume was found in this cell, try the associated linked cell */
    if (!tvp && hac && areq->volumeError) {
        tvp =
            afs_GetVolumeByName(volnamep, assocCell, prefetch, areq,
                                WRITE_LOCK);
    }

    /* done with volname */
    if (cpos)
        *cpos = ':';
    if (!tvp)
        return ENODEV;          /* Couldn't find the volume */
    else
        volid = tvp->volume;

    /* Don't cross mountpoint from a BK to a BK volume */
    if ((states & CBackup) && (tvp->states & VBackup)) {
        afs_PutVolume(tvp, WRITE_LOCK);
        return ENODEV;
    }

    /* If we want (prefetched) the BK and it exists, then drop the RW volume
     * and get the BK.
     * Otherwise, if we want (prefetched0 the RO and it exists, then drop the
     * RW volume and get the RO.
     * Otherwise, go with the RW.
     */
    if ((prefetch == 3) && tvp->backVol) {
        tfid.Fid.Volume = tvp->backVol; /* remember BK volume */
        tfid.Cell = tvp->cell;
        afs_PutVolume(tvp, WRITE_LOCK); /* release old volume */
        tvp = afs_GetVolume(&tfid, areq, WRITE_LOCK);   /* get the new one */
        if (!tvp)
            return ENODEV;      /* oops, can't do it */
    } else if ((prefetch >= 2) && tvp->roVol) {
        tfid.Fid.Volume = tvp->roVol;   /* remember RO volume */
        tfid.Cell = tvp->cell;
        afs_PutVolume(tvp, WRITE_LOCK); /* release old volume */
        tvp = afs_GetVolume(&tfid, areq, WRITE_LOCK);   /* get the new one */
        if (!tvp)
            return ENODEV;      /* oops, can't do it */
    }

done:
    if (acellidxp)
        *acellidxp = cellidx;
    if (avolnump)
        *avolnump = volid;
    if (avnoidp)
        *avnoidp = vnoid;
    if (auniqp)
        *auniqp = uniq;
    if (avolpp)
        *avolpp = tvp;
    else if (tvp)
        afs_PutVolume(tvp, WRITE_LOCK);
    return 0;
}

int
EvalMountPoint(struct vcache *avc, struct vcache *advc,
               struct volume **avolpp, struct vrequest *areq)
{
    afs_int32 code;
    afs_uint32 avnoid, auniq;

    AFS_STATCNT(EvalMountPoint);
    *avolpp = NULL;
    code = afs_HandleLink(avc, areq);
    if (code)
        return code;

    /* Determine which cell and volume the mointpoint goes to */
    code = EvalMountData(avc->linkData[0], avc->linkData + 1,
                         avc->f.states, avc->f.fid.Cell, avolpp, areq, 0, 0,
                         &avnoid, &auniq);
    if (code) return code;

    if (!avnoid)
        avnoid = 1;

    if (!auniq)
        auniq = 1;

    if (avc->mvid.target_root == NULL)
        avc->mvid.target_root = osi_AllocSmallSpace(sizeof(struct VenusFid));
    avc->mvid.target_root->Cell = (*avolpp)->cell;
    avc->mvid.target_root->Fid.Volume = (*avolpp)->volume;
    avc->mvid.target_root->Fid.Vnode = avnoid;
    avc->mvid.target_root->Fid.Unique = auniq;
    avc->f.states |= CMValid;

    /* Used to: if the mount point is stored within a backup volume,
     * then we should only update the parent pointer information if
     * there's none already set, so as to avoid updating a volume's ..
     * info with something in an OldFiles directory.
     *
     * Next two lines used to be under this if:
     *
     * if (!(avc->f.states & CBackup) || tvp->dotdot.Fid.Volume == 0)
     *
     * Now: update mount point back pointer on every call, so that we handle
     * multiple mount points better.  This way, when du tries to go back
     * via chddir(".."), it will end up exactly where it started, yet
     * cd'ing via a new path to a volume will reset the ".." pointer
     * to the new path.
     */
    (*avolpp)->mtpoint = avc->f.fid;    /* setup back pointer to mtpoint */
    
    if (advc)
        (*avolpp)->dotdot = advc->f.fid;

    return 0;
}

/*
 * afs_InitFakeStat
 *
 * Must be called on an afs_fakestat_state object before calling
 * afs_EvalFakeStat or afs_PutFakeStat.  Calling afs_PutFakeStat
 * without calling afs_EvalFakeStat is legal, as long as this
 * function is called.
 */
void
afs_InitFakeStat(struct afs_fakestat_state *state)
{
    if (!afs_fakestat_enable)
        return;

    state->valid = 1;
    state->did_eval = 0;
    state->need_release = 0;
}

/*
 * afs_EvalFakeStat_int
 *
 * The actual implementation of afs_EvalFakeStat and afs_TryEvalFakeStat,
 * which is called by those wrapper functions.
 *
 * Only issues RPCs if canblock is non-zero.
 */
static int
afs_EvalFakeStat_int(struct vcache **avcp, struct afs_fakestat_state *state,
                     struct vrequest *areq, int canblock)
{
    struct vcache *tvc, *root_vp;
    struct volume *tvolp = NULL;
    int code = 0;

    if (!afs_fakestat_enable)
        return 0;

    osi_Assert(state->valid == 1);
    osi_Assert(state->did_eval == 0);
    state->did_eval = 1;

    tvc = *avcp;
    if (tvc->mvstat != AFS_MVSTAT_MTPT)
        return 0;

    if (canblock) {
        /* Is the call to VerifyVCache really necessary? */
        code = afs_VerifyVCache(tvc, areq);
        if (code)
            goto done;

        ObtainWriteLock(&tvc->lock, 599);
        code = EvalMountPoint(tvc, NULL, &tvolp, areq);
        ReleaseWriteLock(&tvc->lock);
        if (code)
            goto done;
        if (tvolp) {
            tvolp->dotdot = tvc->f.fid;
            tvolp->dotdot.Fid.Vnode = tvc->f.parent.vnode;
            tvolp->dotdot.Fid.Unique = tvc->f.parent.unique;
        }
    }
    if (tvc->mvid.target_root && (tvc->f.states & CMValid)) {
        if (!canblock) {
            afs_int32 retry;

            do {
                retry = 0;
                ObtainReadLock(&afs_xvcache);
                root_vp = afs_FindVCache(tvc->mvid.target_root, &retry, 0);
                if (root_vp && retry) {
                    ReleaseReadLock(&afs_xvcache);
                    afs_PutVCache(root_vp);
                }
            } while (root_vp && retry);
            ReleaseReadLock(&afs_xvcache);
        } else {
            root_vp = afs_GetVCache(tvc->mvid.target_root, areq);
        }
        if (!root_vp) {
            code = canblock ? EIO : 0;
            goto done;
        }
#ifdef AFS_DARWIN80_ENV
        root_vp->f.m.Type = VDIR;
        AFS_GUNLOCK();
        code = afs_darwin_finalizevnode(root_vp, NULL, NULL, 0, 0);
        AFS_GLOCK();
        if (code) goto done;
        vnode_ref(AFSTOV(root_vp));
#endif
        if (tvolp && !afs_InReadDir(root_vp)) {
            /* Is this always kosher?  Perhaps we should instead use
             * NBObtainWriteLock to avoid potential deadlock.
             */
            ObtainWriteLock(&root_vp->lock, 598);
            if (!root_vp->mvid.parent)
                root_vp->mvid.parent = osi_AllocSmallSpace(sizeof(struct VenusFid));
            *root_vp->mvid.parent = tvolp->dotdot;
            ReleaseWriteLock(&root_vp->lock);
        }
        state->need_release = 1;
        state->root_vp = root_vp;
        *avcp = root_vp;
        code = 0;
    } else {
        code = canblock ? EIO : 0;
    }

  done:
    if (tvolp)
        afs_PutVolume(tvolp, WRITE_LOCK);
    return code;
}

/*
 * afs_EvalFakeStat
 *
 * Automatically does the equivalent of EvalMountPoint for vcache entries
 * which are mount points.  Remembers enough state to properly release
 * the volume root vcache when afs_PutFakeStat() is called.
 *
 * State variable must be initialized by afs_InitFakeState() beforehand.
 *
 * Returns 0 when everything succeeds and *avcp points to the vcache entry
 * that should be used for the real vnode operation.  Returns non-zero if
 * something goes wrong and the error code should be returned to the user.
 */
int
afs_EvalFakeStat(struct vcache **avcp, struct afs_fakestat_state *state,
                 struct vrequest *areq)
{
    return afs_EvalFakeStat_int(avcp, state, areq, 1);
}

/*
 * afs_TryEvalFakeStat
 *
 * Same as afs_EvalFakeStat, but tries not to talk to remote servers
 * and only evaluate the mount point if all the data is already in
 * local caches.
 *
 * Returns 0 if everything succeeds and *avcp points to a valid
 * vcache entry (possibly evaluated).
 */
int
afs_TryEvalFakeStat(struct vcache **avcp, struct afs_fakestat_state *state,
                    struct vrequest *areq)
{
    return afs_EvalFakeStat_int(avcp, state, areq, 0);
}

/*
 * afs_PutFakeStat
 *
 * Perform any necessary cleanup at the end of a vnode op, given that
 * afs_InitFakeStat was previously called with this state.
 */
void
afs_PutFakeStat(struct afs_fakestat_state *state)
{
    if (!afs_fakestat_enable)
        return;

    osi_Assert(state->valid == 1);
    if (state->need_release)
        afs_PutVCache(state->root_vp);
    state->valid = 0;
}

int
afs_ENameOK(char *aname)
{
    int tlen;

    AFS_STATCNT(ENameOK);
    tlen = strlen(aname);
    if (tlen >= 4 && strcmp(aname + tlen - 4, "@sys") == 0)
        return 0;
    return 1;
}

static int
afs_getsysname(struct vrequest *areq, struct vcache *adp,
               char *bufp, int *num, char **sysnamelist[])
{
    struct unixuser *au;
    afs_int32 error;

    AFS_STATCNT(getsysname);

    *sysnamelist = afs_sysnamelist;

    if (!afs_nfsexporter)
        strcpy(bufp, (*sysnamelist)[0]);
    else {
        au = afs_GetUser(areq->uid, adp->f.fid.Cell, READ_LOCK);
        if (au->exporter) {
            error = EXP_SYSNAME(au->exporter, (char *)0, sysnamelist, num, 0);
            if (error) {
                strcpy(bufp, "@sys");
                afs_PutUser(au, READ_LOCK);
                return -1;
            } else {
                strcpy(bufp, (*sysnamelist)[0]);
            }
        } else
            strcpy(bufp, afs_sysname);
        afs_PutUser(au, READ_LOCK);
    }
    return 0;
}

void
Check_AtSys(struct vcache *avc, const char *aname,
            struct sysname_info *state, struct vrequest *areq)
{
    int num = 0;
    char **sysnamelist[MAXNUMSYSNAMES];

    if (AFS_EQ_ATSYS(aname)) {
        state->offset = 0;
        state->name = osi_AllocLargeSpace(MAXSYSNAME);
        state->allocked = 1;
        state->index =
            afs_getsysname(areq, avc, state->name, &num, sysnamelist);
    } else {
        state->offset = -1;
        state->allocked = 0;
        state->index = 0;
        state->name = (char *)aname;
    }
}

int
Next_AtSys(struct vcache *avc, struct vrequest *areq,
           struct sysname_info *state)
{
    int num = afs_sysnamecount;
    char **sysnamelist[MAXNUMSYSNAMES];

    if (state->index == -1)
        return 0;               /* No list */

    /* Check for the initial state of aname != "@sys" in Check_AtSys */
    if (state->offset == -1 && state->allocked == 0) {
        char *tname;

        /* Check for .*@sys */
        for (tname = state->name; *tname; tname++)
            /*Move to the end of the string */ ;

        if ((tname > state->name + 4) && (AFS_EQ_ATSYS(tname - 4))) {
            state->offset = (tname - 4) - state->name;
            tname = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
            strncpy(tname, state->name, state->offset);
            state->name = tname;
            state->allocked = 1;
            num = 0;
            state->index =
                afs_getsysname(areq, avc, state->name + state->offset, &num,
                               sysnamelist);
            return 1;
        } else
            return 0;           /* .*@sys doesn't match either */
    } else {
        struct unixuser *au;
        afs_int32 error;

        *sysnamelist = afs_sysnamelist;

        if (afs_nfsexporter) {
            au = afs_GetUser(areq->uid, avc->f.fid.Cell, READ_LOCK);
            if (au->exporter) {
                error =
                    EXP_SYSNAME(au->exporter, (char *)0, sysnamelist, &num, 0);
                if (error) {
                    afs_PutUser(au, READ_LOCK);
                    return 0;
                }
            }
            afs_PutUser(au, READ_LOCK);
        }
        if (++(state->index) >= num || !(*sysnamelist)[(unsigned int)state->index])
            return 0;           /* end of list */
    }
    strcpy(state->name + state->offset, (*sysnamelist)[(unsigned int)state->index]);
    return 1;
}

static int
afs_CheckBulkStatus(struct afs_conn *tc, int nFids, AFSBulkStats *statParm,
                    AFSCBs *cbParm)
{
    int i;
    int code;

    if (statParm->AFSBulkStats_len != nFids || cbParm->AFSCBs_len != nFids) {
        afs_warn("afs: BulkFetchStatus length %u/%u, expected %u\n",
                 (unsigned)statParm->AFSBulkStats_len,
                 (unsigned)cbParm->AFSCBs_len, nFids);
        afs_BadFetchStatus(tc);
        return VBUSY;
    }
    for (i = 0; i < nFids; i++) {
        if (statParm->AFSBulkStats_val[i].errorCode) {
            continue;
        }
        code = afs_CheckFetchStatus(tc, &statParm->AFSBulkStats_val[i]);
        if (code) {
            return code;
        }
    }

    return 0;
}

extern int BlobScan(struct dcache * afile, afs_int32 ablob, afs_int32 *ablobOut);

/* called with an unlocked directory and directory cookie.  Areqp
 * describes who is making the call.
 * Scans the next N (about 30, typically) directory entries, and does
 * a bulk stat call to stat them all.
 *
 * Must be very careful when merging in RPC responses, since we dont
 * want to overwrite newer info that was added by a file system mutating
 * call that ran concurrently with our bulk stat call.
 *
 * We do that, as described below, by not merging in our info (always
 * safe to skip the merge) if the status info is valid in the vcache entry.
 *
 * If adapt ever implements the bulk stat RPC, then this code will need to
 * ensure that vcaches created for failed RPC's to older servers have the
 * CForeign bit set.
 */
static struct vcache *BStvc = NULL;

int
afs_DoBulkStat(struct vcache *adp, long dirCookie, struct vrequest *areqp)
{
    int nentries;               /* # of entries to prefetch */
    int nskip;                  /* # of slots in the LRU queue to skip */
#ifdef AFS_DARWIN80_ENV
    int npasses = 0;
    struct vnode *lruvp;
#endif
    struct vcache *lruvcp;      /* vcache ptr of our goal pos in LRU queue */
    struct dcache *dcp;         /* chunk containing the dir block */
    afs_size_t temp;            /* temp for holding chunk length, &c. */
    struct AFSFid *fidsp;       /* file IDs were collecting */
    struct AFSCallBack *cbsp;   /* call back pointers */
    struct AFSCallBack *tcbp;   /* temp callback ptr */
    struct AFSFetchStatus *statsp;      /* file status info */
    struct AFSVolSync volSync;  /* vol sync return info */
    struct vcache *tvcp;        /* temp vcp */
    struct afs_q *tq;           /* temp queue variable */
    AFSCBFids fidParm;          /* file ID parm for bulk stat */
    AFSBulkStats statParm;      /* stat info parm for bulk stat */
    int fidIndex = 0;           /* which file were stating */
    struct afs_conn *tcp = 0;   /* conn for call */
    AFSCBs cbParm;              /* callback parm for bulk stat */
    struct server *hostp = 0;   /* host we got callback from */
    long startTime;             /* time we started the call,
                                 * for callback expiration base
                                 */
#if defined(AFS_DARWIN_ENV)
    int ftype[4] = {VNON, VREG, VDIR, VLNK}; /* verify type is as expected */
#endif
    afs_size_t statSeqNo = 0;   /* Valued of file size to detect races */
    int code;                   /* error code */
    afs_int32 newIndex;         /* new index in the dir */
    struct DirBuffer entry;     /* Buffer for dir manipulation */
    struct DirEntry *dirEntryp; /* dir entry we are examining */
    int i;
    struct VenusFid afid;       /* file ID we are using now */
    struct VenusFid tfid;       /* another temp. file ID */
    afs_int32 retry;            /* handle low-level SGI MP race conditions */
    long volStates;             /* flags from vol structure */
    struct volume *volp = 0;    /* volume ptr */
    struct VenusFid dotdot = {0, {0, 0, 0}};
    int flagIndex = 0;          /* First file with bulk fetch flag set */
    struct rx_connection *rxconn;
    int attempt_i;
    XSTATS_DECLS;
    dotdot.Cell = 0;
    dotdot.Fid.Unique = 0;
    dotdot.Fid.Vnode = 0;

    /* first compute some basic parameters.  We dont want to prefetch more
     * than a fraction of the cache in any given call, and we want to preserve
     * a portion of the LRU queue in any event, so as to avoid thrashing
     * the entire stat cache (we will at least leave some of it alone).
     * presently dont stat more than 1/8 the cache in any one call.      */
    nentries = afs_cacheStats / 8;

    /* dont bother prefetching more than one calls worth of info */
    if (nentries > AFSCBMAX)
        nentries = AFSCBMAX;

    /* heuristic to make sure that things fit in 4K.  This means that
     * we shouldnt make it any bigger than 47 entries.  I am typically
     * going to keep it a little lower, since we don't want to load
     * too much of the stat cache.
     */
    if (nentries > 30)
        nentries = 30;

    /* now, to reduce the stack size, well allocate two 4K blocks,
     * one for fids and callbacks, and one for stat info.  Well set
     * up our pointers to the memory from there, too.
     */
    statsp = osi_Alloc(AFSCBMAX * sizeof(AFSFetchStatus));
    fidsp = osi_AllocLargeSpace(nentries * sizeof(AFSFid));
    cbsp = osi_Alloc(AFSCBMAX * sizeof(AFSCallBack));

    /* next, we must iterate over the directory, starting from the specified
     * cookie offset (dirCookie), and counting out nentries file entries.
     * We skip files that already have stat cache entries, since we
     * dont want to bulk stat files that are already in the cache.
     */
  tagain:
    code = afs_VerifyVCache(adp, areqp);
    if (code)
        goto done2;

    dcp = afs_GetDCache(adp, (afs_size_t) 0, areqp, &temp, &temp, 1);
    if (!dcp) {
        code = EIO;
        goto done2;
    }

    /* lock the directory cache entry */
    ObtainReadLock(&adp->lock);
    ObtainReadLock(&dcp->lock);

    /*
     * Make sure that the data in the cache is current. There are two
     * cases we need to worry about:
     * 1. The cache data is being fetched by another process.
     * 2. The cache data is no longer valid
     */
    while ((adp->f.states & CStatd)
           && (dcp->dflags & DFFetching)
           && afs_IsDCacheFresh(dcp, adp)) {
        afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING,
                   __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, dcp,
                   ICL_TYPE_INT32, dcp->dflags);
        ReleaseReadLock(&dcp->lock);
        ReleaseReadLock(&adp->lock);
        afs_osi_Sleep(&dcp->validPos);
        ObtainReadLock(&adp->lock);
        ObtainReadLock(&dcp->lock);
    }
    if (!(adp->f.states & CStatd)
        || !afs_IsDCacheFresh(dcp, adp)) {
        ReleaseReadLock(&dcp->lock);
        ReleaseReadLock(&adp->lock);
        afs_PutDCache(dcp);
        goto tagain;
    }

    /* Generate a sequence number so we can tell whether we should
     * store the attributes when processing the response. This number is
     * stored in the file size when we set the CBulkFetching bit. If the
     * CBulkFetching is still set and this value hasn't changed, then
     * we know we were the last to set CBulkFetching bit for this file,
     * and it is safe to set the status information for this file.
     */
    statSeqNo = bulkStatCounter++;
    /* ensure against wrapping */
    if (statSeqNo == 0)
        statSeqNo = bulkStatCounter++;

    /* now we have dir data in the cache, so scan the dir page */
    fidIndex = 0;
    flagIndex = 0;

    /*
     * Only examine at most the next 'nentries*4' entries to find dir entries
     * to stat. This is an arbitrary limit that we set so we don't waste time
     * scanning an entire dir that contains stat'd entries. For example, if a
     * dir contains 10k entries, and all or almost all of them are stat'd, then
     * we'll examine 10k entries for no benefit. For each entry, we run
     * afs_FindVCache, and grab and release afs_xvcache; doing this e.g. 10k
     * times can have significant impact if the client is under a lot of load.
     */
    for (attempt_i = 0; attempt_i < nentries * 4; attempt_i++) {

        /* look for first safe entry to examine in the directory.  BlobScan
         * looks for a the 1st allocated dir after the dirCookie slot.
         */
        code = BlobScan(dcp, (dirCookie >> 5), &newIndex);
        if (code || newIndex == 0)
            break;

        /* remember the updated directory cookie */
        dirCookie = newIndex << 5;

        /* get a ptr to the dir entry */
        code = afs_dir_GetBlob(dcp, newIndex, &entry);
        if (code)
            break;
        dirEntryp = (struct DirEntry *)entry.data;

        /* dont copy more than we have room for */
        if (fidIndex >= nentries) {
            DRelease(&entry, 0);
            break;
        }

        /* now, if the dir entry looks good, copy it out to our list.  Vnode
         * 0 means deleted, although it should also be free were it deleted.
         */
        if (dirEntryp->fid.vnode != 0) {
            /* dont copy entries we have in our cache.  This check will
             * also make us skip "." and probably "..", unless it has
             * disappeared from the cache since we did our namei call.
             */
            tfid.Cell = adp->f.fid.Cell;
            tfid.Fid.Volume = adp->f.fid.Fid.Volume;
            tfid.Fid.Vnode = ntohl(dirEntryp->fid.vnode);
            tfid.Fid.Unique = ntohl(dirEntryp->fid.vunique);
            do {
                retry = 0;
                ObtainSharedLock(&afs_xvcache, 130);
                tvcp = afs_FindVCache(&tfid, &retry, IS_SLOCK /* no stats | LRU */ );
                if (tvcp && retry) {
                    ReleaseSharedLock(&afs_xvcache);
                    afs_PutVCache(tvcp);
                }
            } while (tvcp && retry);
            if (!tvcp) {        /* otherwise, create manually */
                UpgradeSToWLock(&afs_xvcache, 129);
                tvcp = afs_NewBulkVCache(&tfid, hostp, statSeqNo);
                if (tvcp)
                {
                    ObtainWriteLock(&tvcp->lock, 505);
#ifdef AFS_DARWIN80_ENV
                    /* use even/odd hack to guess file versus dir.
                       let links be reaped. oh well. */
                    if (dirEntryp->fid.vnode & 1)
                        tvcp->f.m.Type = VDIR;
                    else
                        tvcp->f.m.Type = VREG;
                    /* finalize to a best guess */
                    afs_darwin_finalizevnode(tvcp, AFSTOV(adp), NULL, 0, 1);
                    /* re-acquire usecount that finalizevnode disposed of */
                    vnode_ref(AFSTOV(tvcp));
#endif
                    ReleaseWriteLock(&afs_xvcache);
                    afs_RemoveVCB(&tfid);
                    ReleaseWriteLock(&tvcp->lock);
                } else {
                    ReleaseWriteLock(&afs_xvcache);
                }
            } else {
                ReleaseSharedLock(&afs_xvcache);
            }
            if (!tvcp)
            {
                DRelease(&entry, 0);
                ReleaseReadLock(&dcp->lock);
                ReleaseReadLock(&adp->lock);
                afs_PutDCache(dcp);
                goto done;      /* can happen if afs_NewVCache fails */
            }

            /* WARNING: afs_DoBulkStat uses the Length field to store a
             * sequence number for each bulk status request. Under no
             * circumstances should afs_DoBulkStat store a sequence number
             * if the new length will be ignored when afs_ProcessFS is
             * called with new stats. */
#ifdef AFS_SGI_ENV
            if (!(tvcp->f.states & CStatd)
                && (!((tvcp->f.states & CBulkFetching) &&
                      (tvcp->f.m.Length != statSeqNo)))
                && (tvcp->execsOrWriters <= 0)
                && !afs_DirtyPages(tvcp)
                && !AFS_VN_MAPPED((vnode_t *) tvcp))
#else
            if (!(tvcp->f.states & CStatd)
                && (!((tvcp->f.states & CBulkFetching) &&
                      (tvcp->f.m.Length != statSeqNo)))
                && (tvcp->execsOrWriters <= 0)
                && !afs_DirtyPages(tvcp))
#endif

            {
                /* this entry doesnt exist in the cache, and is not
                 * already being fetched by someone else, so add it to the
                 * list of file IDs to obtain.
                 *
                 * We detect a callback breaking race condition by checking the
                 * CBulkFetching state bit and the value in the file size.
                 * It is safe to set the status only if the CBulkFetching
                 * flag is still set and the value in the file size does
                 * not change. NewBulkVCache sets us up for the new ones.
                 * Set up the rest here.
                 *
                 * Don't fetch status for dirty files. We need to
                 * preserve the value of the file size. We could
                 * flush the pages, but it wouldn't be worthwhile.
                 */
                if (!(tvcp->f.states & CBulkFetching)) {
                    tvcp->f.states |= CBulkFetching;
                    tvcp->f.m.Length = statSeqNo;
                }
                memcpy((char *)(fidsp + fidIndex), (char *)&tfid.Fid,
                       sizeof(*fidsp));
                fidIndex++;
            }
            afs_PutVCache(tvcp);
        }

        /* if dir vnode has non-zero entry */
        /* move to the next dir entry by adding in the # of entries
         * used by this dir entry.
         */
        temp = afs_dir_NameBlobs(dirEntryp->name) << 5;
        DRelease(&entry, 0);
        if (temp <= 0)
            break;
        dirCookie += temp;
    }                           /* for loop over dir entries */

    /* now release the dir lock and prepare to make the bulk RPC */
    ReleaseReadLock(&dcp->lock);
    ReleaseReadLock(&adp->lock);

    /* release the chunk */
    afs_PutDCache(dcp);

    /* dont make a null call */
    if (fidIndex == 0)
        goto done;

    do {
        /* setup the RPC parm structures */
        fidParm.AFSCBFids_len = fidIndex;
        fidParm.AFSCBFids_val = fidsp;
        statParm.AFSBulkStats_len = fidIndex;
        statParm.AFSBulkStats_val = statsp;
        cbParm.AFSCBs_len = fidIndex;
        cbParm.AFSCBs_val = cbsp;

        /* start the timer; callback expirations are relative to this */
        startTime = osi_Time();

        tcp = afs_Conn(&adp->f.fid, areqp, SHARED_LOCK, &rxconn);
        if (tcp) {
            hostp = tcp->parent->srvr->server;

            for (i = 0; i < fidIndex; i++) {
                /* we must set tvcp->callback before the BulkStatus call, so
                 * we can detect concurrent InitCallBackState's */

                afid.Cell = adp->f.fid.Cell;
                afid.Fid.Volume = adp->f.fid.Fid.Volume;
                afid.Fid.Vnode = fidsp[i].Vnode;
                afid.Fid.Unique = fidsp[i].Unique;

                do {
                    retry = 0;
                    ObtainReadLock(&afs_xvcache);
                    tvcp = afs_FindVCache(&afid, &retry, 0 /* !stats&!lru */);
                    ReleaseReadLock(&afs_xvcache);
                } while (tvcp && retry);

                if (!tvcp) {
                    continue;
                }

                if ((tvcp->f.states & CBulkFetching) &&
                     (tvcp->f.m.Length == statSeqNo)) {
                    tvcp->callback = hostp;
                }

                afs_PutVCache(tvcp);
                tvcp = NULL;
            }

            XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_BULKSTATUS);

            if (!(tcp->parent->srvr->server->flags & SNO_INLINEBULK)) {
                RX_AFS_GUNLOCK();
                code =
                    RXAFS_InlineBulkStatus(rxconn, &fidParm, &statParm,
                                           &cbParm, &volSync);
                RX_AFS_GLOCK();
                if (code == RXGEN_OPCODE) {
                    tcp->parent->srvr->server->flags |= SNO_INLINEBULK;
                    RX_AFS_GUNLOCK();
                    code =
                        RXAFS_BulkStatus(rxconn, &fidParm, &statParm,
                                         &cbParm, &volSync);
                    RX_AFS_GLOCK();
                }
            } else {
                RX_AFS_GUNLOCK();
                code =
                    RXAFS_BulkStatus(rxconn, &fidParm, &statParm, &cbParm,
                                     &volSync);
                RX_AFS_GLOCK();
            }
            XSTATS_END_TIME;

            if (code == 0) {
                code = afs_CheckBulkStatus(tcp, fidIndex, &statParm, &cbParm);
            }
        } else
            code = -1;
        /* make sure we give afs_Analyze a chance to retry,
         * but if the RPC succeeded we may have entries to merge.
         * if we wipe code with one entry's status we get bogus failures.
         */
    } while (afs_Analyze
             (tcp, rxconn, code ? code : (&statsp[0])->errorCode,
              &adp->f.fid, areqp, AFS_STATS_FS_RPCIDX_BULKSTATUS,
              SHARED_LOCK, NULL));

    /* now, if we didnt get the info, bail out. */
    if (code)
        goto done;

    /* we need vol flags to create the entries properly */
    dotdot.Fid.Volume = 0;
    volp = afs_GetVolume(&adp->f.fid, areqp, READ_LOCK);
    if (volp) {
        volStates = volp->states;
        if (volp->dotdot.Fid.Volume != 0)
            dotdot = volp->dotdot;
    } else
        volStates = 0;

    /* find the place to merge the info into  We do this by skipping
     * nskip entries in the LRU queue.  The more we skip, the more
     * we preserve, since the head of the VLRU queue is the most recently
     * referenced file.
     */
  reskip:
    nskip = afs_cacheStats / 2; /* preserved fraction of the cache */
    ObtainReadLock(&afs_xvcache);
#ifdef AFS_DARWIN80_ENV
 reskip2:
#endif
    if (QEmpty(&VLRU)) {
        /* actually a serious error, probably should panic. Probably will 
         * panic soon, oh well. */
        ReleaseReadLock(&afs_xvcache);
        afs_warnuser("afs_DoBulkStat: VLRU empty!");
        goto done;
    }
    if ((VLRU.next->prev != &VLRU) || (VLRU.prev->next != &VLRU)) {
        refpanic("Bulkstat VLRU inconsistent");
    }
    for (tq = VLRU.next; tq != &VLRU; tq = QNext(tq)) {
        if (--nskip <= 0) {
#ifdef AFS_DARWIN80_ENV
            if ((!(QTOV(tq)->f.states & CDeadVnode)&&!(QTOV(tq)->f.states & CVInit)))
#endif
                break;
        }
        if (QNext(QPrev(tq)) != tq) {
            BStvc = QTOV(tq);
            refpanic("BulkStat VLRU inconsistent");
        }
    }
    if (tq != &VLRU)
        lruvcp = QTOV(tq);
    else
        lruvcp = QTOV(VLRU.next);

    /* now we have to hold this entry, so that it does not get moved
     * into the free list while we're running.  It could still get
     * moved within the lru queue, but hopefully that will be rare; it
     * doesn't hurt nearly as much.
     */
    retry = 0;
#ifdef AFS_DARWIN80_ENV
    if (((lruvcp->f.states & CDeadVnode)||(lruvcp->f.states & CVInit))) {
        if (npasses == 0) {
            nskip = 1;
            npasses++;
            goto reskip2;
        } else
            panic("Can't find non-dead vnode in VLRU\n");
    }
    lruvp = AFSTOV(lruvcp);
    if (vnode_get(lruvp))       /* this bumps ref count */
        retry = 1;
    else if (vnode_ref(lruvp)) {
        AFS_GUNLOCK();
        /* AFSTOV(lruvcp) may be NULL */
        vnode_put(lruvp);
        AFS_GLOCK();
        retry = 1;
    }
#else
    if (osi_vnhold(lruvcp) != 0) {
        retry = 1;
    }
#endif
    ReleaseReadLock(&afs_xvcache);      /* could be read lock */
    if (retry)
        goto reskip;

    /* otherwise, merge in the info.  We have to be quite careful here,
     * since we need to ensure that we don't merge old info over newer
     * stuff in a stat cache entry.  We're very conservative here: we don't
     * do the merge at all unless we ourselves create the stat cache
     * entry.  That's pretty safe, and should work pretty well, since we
     * typically expect to do the stat cache creation ourselves.
     *
     * We also have to take into account racing token revocations.
     */
    for (i = 0; i < fidIndex; i++) {
        if ((&statsp[i])->errorCode)
            continue;
        afid.Cell = adp->f.fid.Cell;
        afid.Fid.Volume = adp->f.fid.Fid.Volume;
        afid.Fid.Vnode = fidsp[i].Vnode;
        afid.Fid.Unique = fidsp[i].Unique;
        do {
            retry = 0;
            ObtainReadLock(&afs_xvcache);
            tvcp = afs_FindVCache(&afid, &retry, 0/* !stats&!lru */);
            ReleaseReadLock(&afs_xvcache);
        } while (tvcp && retry);

        /* The entry may no longer exist */
        if (tvcp == NULL) {
            continue;
        }

        /* now we have the entry held, but we need to fill it in */
        ObtainWriteLock(&tvcp->lock, 131);

        /* if CBulkFetching is not set, or if the file size no longer
         * matches the value we placed there when we set the CBulkFetching
         * flag, then someone else has done something with this node,
         * and we may not have the latest status information for this
         * file.  Leave the entry alone. There's also a file type
         * change here, for OSX bulkstat support.
         */
        if (!(tvcp->f.states & CBulkFetching)
            || (tvcp->f.m.Length != statSeqNo)
#if defined(AFS_DARWIN_ENV)
            || (ftype[(&statsp[i])->FileType] != vType(tvcp))
#endif
           ) {
            flagIndex++;
            ReleaseWriteLock(&tvcp->lock);
            afs_PutVCache(tvcp);
            continue;
        }

        /* now copy ".." entry back out of volume structure, if necessary */
        if (tvcp->mvstat == AFS_MVSTAT_ROOT && (dotdot.Fid.Volume != 0)) {
            if (!tvcp->mvid.parent)
                tvcp->mvid.parent = osi_AllocSmallSpace(sizeof(struct VenusFid));
            *tvcp->mvid.parent = dotdot;
        }

#ifdef AFS_DARWIN80_ENV
        if (((lruvcp->f.states & CDeadVnode)||(lruvcp->f.states & CVInit)))
            panic("vlru control point went dead\n");
#endif

        ObtainWriteLock(&afs_xvcache, 132);
        if ((VLRU.next->prev != &VLRU) || (VLRU.prev->next != &VLRU)) {
            refpanic("Bulkstat VLRU inconsistent2");
        }
        if ((QNext(QPrev(&tvcp->vlruq)) != &tvcp->vlruq)
            || (QPrev(QNext(&tvcp->vlruq)) != &tvcp->vlruq)) {
            refpanic("Bulkstat VLRU inconsistent4");
        }
        if ((QNext(QPrev(&lruvcp->vlruq)) != &lruvcp->vlruq)
            || (QPrev(QNext(&lruvcp->vlruq)) != &lruvcp->vlruq)) {
            refpanic("Bulkstat VLRU inconsistent5");
        }

        if (tvcp != lruvcp) {   /* if they are == don't move it, don't corrupt vlru */
            QRemove(&tvcp->vlruq);
            QAdd(&lruvcp->vlruq, &tvcp->vlruq);
        }

        if ((VLRU.next->prev != &VLRU) || (VLRU.prev->next != &VLRU)) {
            refpanic("Bulkstat VLRU inconsistent3");
        }
        if ((QNext(QPrev(&tvcp->vlruq)) != &tvcp->vlruq)
            || (QPrev(QNext(&tvcp->vlruq)) != &tvcp->vlruq)) {
            refpanic("Bulkstat VLRU inconsistent5");
        }
        if ((QNext(QPrev(&lruvcp->vlruq)) != &lruvcp->vlruq)
            || (QPrev(QNext(&lruvcp->vlruq)) != &lruvcp->vlruq)) {
            refpanic("Bulkstat VLRU inconsistent6");
        }
        ReleaseWriteLock(&afs_xvcache);

        ObtainWriteLock(&afs_xcbhash, 494);

        /* We need to check the flags again. We may have missed
         * something while we were waiting for a lock.
         */
        if (!(tvcp->f.states & CBulkFetching) || (tvcp->f.m.Length != statSeqNo)) {
            flagIndex++;
            ReleaseWriteLock(&tvcp->lock);
            ReleaseWriteLock(&afs_xcbhash);
            afs_PutVCache(tvcp);
            continue;
        }

        /* now merge in the resulting status back into the vnode.
         * We only do this if the entry looks clear.
         */
        afs_ProcessFS(tvcp, &statsp[i], areqp);
#if defined(AFS_LINUX22_ENV)
        afs_fill_inode(AFSTOV(tvcp), NULL);     /* reset inode operations */
#endif

        /* do some accounting for bulk stats: mark this entry as
         * loaded, so we can tell if we use it before it gets
         * recycled.
         */
        tvcp->f.states |= CBulkStat;
        tvcp->f.states &= ~CBulkFetching;
        flagIndex++;
        afs_bulkStatsDone++;

        /* merge in vol info */
        if (volStates & VRO)
            tvcp->f.states |= CRO;
        if (volStates & VBackup)
            tvcp->f.states |= CBackup;
        if (volStates & VForeign)
            tvcp->f.states |= CForeign;

        /* merge in the callback info */
        tvcp->f.states |= CTruth;

        /* get ptr to the callback we are interested in */
        tcbp = cbsp + i;

        if (tcbp->ExpirationTime != 0) {
            tvcp->cbExpires = tcbp->ExpirationTime + startTime;
            tvcp->callback = hostp;
            tvcp->f.states |= CStatd;
            afs_QueueCallback(tvcp, CBHash(tcbp->ExpirationTime), volp);
        } else if (tvcp->f.states & CRO) {
            /* ordinary callback on a read-only volume -- AFS 3.2 style */
            tvcp->cbExpires = 3600 + startTime;
            tvcp->callback = hostp;
            tvcp->f.states |= CStatd;
            afs_QueueCallback(tvcp, CBHash(3600), volp);
        } else {
            afs_StaleVCacheFlags(tvcp,
                                 AFS_STALEVC_CBLOCKED | AFS_STALEVC_CLEARCB,
                                 CUnique);
        }
#ifdef AFS_DARWIN80_ENV
        /* reclaim->FlushVCache will need xcbhash */
        if (((tvcp->f.states & CDeadVnode)||(tvcp->f.states & CVInit))) {
            ReleaseWriteLock(&afs_xcbhash);
            /* passing in a parent hangs getting the vnode lock */
            code = afs_darwin_finalizevnode(tvcp, NULL, NULL, 0, 1);
            if (code) {
                /* It's gonna get recycled - shouldn't happen */
                afs_StaleVCacheFlags(tvcp,
                                     AFS_STALEVC_CBLOCKED | AFS_STALEVC_CLEARCB,
                                     CUnique);
            } else
                /* re-acquire the usecount that finalizevnode disposed of */
                vnode_ref(AFSTOV(tvcp));
        } else
#endif
        ReleaseWriteLock(&afs_xcbhash);

        ReleaseWriteLock(&tvcp->lock);
        /* finally, we're done with the entry */
        afs_PutVCache(tvcp);
    }                           /* for all files we got back */

    /* finally return the pointer into the LRU queue */
#ifdef AFS_DARWIN80_ENV
    if (((lruvcp->f.states & CDeadVnode)||(lruvcp->f.states & CVInit)))
        panic("vlru control point went dead before put\n");
    AFS_GUNLOCK();
    vnode_put(lruvp);
    vnode_rele(lruvp);
    AFS_GLOCK();
#else
    afs_PutVCache(lruvcp);
#endif

  done:
    /* Be sure to turn off the CBulkFetching flags */
    for (i = flagIndex; i < fidIndex; i++) {
        afid.Cell = adp->f.fid.Cell;
        afid.Fid.Volume = adp->f.fid.Fid.Volume;
        afid.Fid.Vnode = fidsp[i].Vnode;
        afid.Fid.Unique = fidsp[i].Unique;
        do {
            retry = 0;
            ObtainReadLock(&afs_xvcache);
            tvcp = afs_FindVCache(&afid, &retry, 0 /* !stats&!lru */);
            ReleaseReadLock(&afs_xvcache);
        } while (tvcp && retry);
        if (tvcp != NULL) {
            if ((tvcp->f.states & CBulkFetching)
                && (tvcp->f.m.Length == statSeqNo)) {
                tvcp->f.states &= ~CBulkFetching;
            }
            afs_PutVCache(tvcp);
        }
    }
    if (volp)
        afs_PutVolume(volp, READ_LOCK);

  done2:
    osi_FreeLargeSpace((char *)fidsp);
    osi_Free((char *)statsp, AFSCBMAX * sizeof(AFSFetchStatus));
    osi_Free((char *)cbsp, AFSCBMAX * sizeof(AFSCallBack));
    return code;
}

#ifdef AFS_DARWIN80_ENV
int AFSDOBULK = 0;
#endif

static int
afs_ShouldTryBulkStat(struct vcache *adp)
{
#ifdef AFS_DARWIN80_ENV
    if (!AFSDOBULK) {
        return 0;
    }
#endif
    if (AFS_IS_DISCONNECTED) {
        /* We can't prefetch entries if we're offline. */
        return 0;
    }
    if (adp->opens < 1) {
        /* Don't bother prefetching entries if nobody is holding the dir open
         * while we're doing a lookup. */
        return 0;
    }
    if (afs_VCacheStressed()) {
        /* If we already have too many vcaches, don't create more vcaches we
         * may not even use. */
        return 0;
    }
    if ((adp->f.states & CForeign)) {
        /* Don't bulkstat for dfs xlator dirs. */
        return 0;
    }
    if (afs_IsDynroot(adp)) {
        /* Don't prefetch dynroot entries; that's pointless, since we generate
         * those locally. */
        return 0;
    }
    if (afs_InReadDir(adp)) {
        /* Don't bulkstat if we're in the middle of servicing a readdir() in
         * the same process. */
        return 0;
    }
    return 1;
}

static_inline int
osi_lookup_isdot(const char *aname)
{
#ifdef AFS_SUN5_ENV
    if (!aname[0]) {
        /* in Solaris, we can get passed "" as a path component if we are the
         * root directory, e.g. after a call to chroot. It is equivalent to
         * looking up "." */
        return 1;
    }
#endif /* AFS_SUN5_ENV */
    if (aname[0] == '.' && !aname[1]) {
        return 1;
    }
    return 0;
}

int
#if defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
afs_lookup(OSI_VC_DECL(adp), char *aname, struct vcache **avcp, struct pathname *pnp, int flags, struct vnode *rdir, afs_ucred_t *acred)
#elif defined(UKERNEL)
afs_lookup(OSI_VC_DECL(adp), char *aname, struct vcache **avcp, afs_ucred_t *acred, int flags)
#else
afs_lookup(OSI_VC_DECL(adp), char *aname, struct vcache **avcp, afs_ucred_t *acred)
#endif
{
    struct vrequest *treq = NULL;
    char *tname = NULL;
    struct vcache *tvc = 0;
    afs_int32 code;
    afs_int32 bulkcode = 0;
    int pass = 0, hit = 0;
    int force_eval = afs_fakestat_enable ? 0 : 1;
    long dirCookie;
    afs_hyper_t versionNo;
    int no_read_access = 0;
    struct sysname_info sysState;       /* used only for @sys checking */
    int dynrootRetry = 1;
    struct afs_fakestat_state fakestate;
    int tryEvalOnly = 0;

    /* Don't allow ENOENT errors, except for a specific code path where
     * 'enoent_prohibited' is cleared below. */
    int enoent_prohibited = 1;

    OSI_VC_CONVERT(adp);

    AFS_STATCNT(afs_lookup);
    afs_InitFakeStat(&fakestate);

    AFS_DISCON_LOCK();

    if ((code = afs_CreateReq(&treq, acred)))
        goto done;

    if (afs_fakestat_enable && adp->mvstat == AFS_MVSTAT_MTPT) {
       if (strcmp(aname, ".directory") == 0)
           tryEvalOnly = 1;
    }

#if defined(AFS_DARWIN_ENV)
    /* Workaround for MacOSX Finder, which tries to look for
     * .DS_Store and Contents under every directory.
     */
    if (afs_fakestat_enable && adp->mvstat == AFS_MVSTAT_MTPT) {
        if (strcmp(aname, ".DS_Store") == 0)
            tryEvalOnly = 1;
        if (strcmp(aname, "Contents") == 0)
            tryEvalOnly = 1;
    }
    if (afs_fakestat_enable && adp->mvstat == AFS_MVSTAT_ROOT) {
        if (strncmp(aname, "._", 2) == 0)
            tryEvalOnly = 1;
    }
#endif

    if (tryEvalOnly)
        code = afs_TryEvalFakeStat(&adp, &fakestate, treq);
    else
        code = afs_EvalFakeStat(&adp, &fakestate, treq);

    /*printf("Code is %d\n", code);*/
    
    if (tryEvalOnly && adp->mvstat == AFS_MVSTAT_MTPT)
        code = ENODEV;
    if (code)
        goto done;

    /* come back to here if we encounter a non-existent object in a read-only
     * volume's directory */
  redo:
    *avcp = NULL;               /* Since some callers don't initialize it */
    bulkcode = 0;

    if (!(adp->f.states & CStatd) && !afs_InReadDir(adp)) {
        if ((code = afs_VerifyVCache2(adp, treq))) {
            goto done;
        }
    } else
        code = 0;

    /* watch for ".." in a volume root */
    if (adp->mvstat == AFS_MVSTAT_ROOT && aname[0] == '.' && aname[1] == '.' && !aname[2]) {
        /* looking up ".." in root via special hacks */
        if (adp->mvid.parent == (struct VenusFid *)0 || adp->mvid.parent->Fid.Volume == 0) {
            code = ENODEV;
            goto done;
        }
        /* otherwise we have the fid here, so we use it */
        /*printf("Getting vcache\n");*/
        tvc = afs_GetVCache(adp->mvid.parent, treq);
        afs_Trace3(afs_iclSetp, CM_TRACE_GETVCDOTDOT, ICL_TYPE_FID, adp->mvid.parent,
                   ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, code);
        *avcp = tvc;
        code = (tvc ? 0 : EIO);
        hit = 1;
        if (tvc && !VREFCOUNT_GT(tvc, 0)) {
            osi_Panic("TT1");
        }
        if (code) {
            /*printf("LOOKUP GETVCDOTDOT -> %d\n", code); */
        }
        goto done;
    }

    /* now check the access */
    if (treq->uid != adp->last_looker) {
        if (!afs_AccessOK(adp, PRSFS_LOOKUP, treq, CHECK_MODE_BITS)) {
            *avcp = NULL;
            code = EACCES;
            goto done;
        } else
            adp->last_looker = treq->uid;
    }

    /* Check for read access as well.  We need read access in order to
     * stat files, but not to stat subdirectories. */
    if (!afs_AccessOK(adp, PRSFS_READ, treq, CHECK_MODE_BITS))
        no_read_access = 1;

    /* special case lookup of ".".  Can we check for it sooner in this code,
     * for instance, way up before "redo:" ??
     * I'm not fiddling with the LRUQ here, either, perhaps I should, or else 
     * invent a lightweight version of GetVCache.
     */
    if (osi_lookup_isdot(aname)) {      /* special case */
        ObtainReadLock(&afs_xvcache);
        if (osi_vnhold(adp) != 0) {
            ReleaseReadLock(&afs_xvcache);
            code = EIO;
            goto done;
        }
        ReleaseReadLock(&afs_xvcache);
#ifdef AFS_DARWIN80_ENV
        vnode_get(AFSTOV(adp));
#endif
        code = 0;
        *avcp = tvc = adp;
        hit = 1;
        if (adp && !VREFCOUNT_GT(adp, 0)) {
            osi_Panic("TT2");
        }
        goto done;
    }

    /*
     * Special case lookup of ".." in the dynamic mount directory.
     * The parent of this directory is _always_ the AFS root volume.
     */
    if (afs_IsDynrootMount(adp) &&
        aname[0] == '.' && aname[1] == '.' && !aname[2]) {

        ObtainReadLock(&afs_xvcache);
        if (osi_vnhold(afs_globalVp) != 0) {
            ReleaseReadLock(&afs_xvcache);
            code = EIO;
            goto done;
        }
        ReleaseReadLock(&afs_xvcache);
#ifdef AFS_DARWIN80_ENV
        vnode_get(AFSTOV(afs_globalVp));
#endif
        code = 0;
        *avcp = tvc = afs_globalVp;
        hit = 1;
        goto done;
    }

    /*
     * Special case lookups in the dynamic mount directory.
     * The names here take the form cell:volume, similar to a mount point.
     * EvalMountData parses that and returns a cell and volume ID, which
     * we use to construct the appropriate dynroot Fid.
     */
    if (afs_IsDynrootMount(adp)) {
        struct VenusFid tfid;
        afs_uint32 cellidx, volid, vnoid, uniq;

        code = EvalMountData('%', aname, 0, 0, NULL, treq, &cellidx, &volid, &vnoid, &uniq);
        if (code)
            goto done;
        /* If a vnode was returned, it's not a real mount point */
        if (vnoid > 1) {
            struct cell *tcell = afs_GetCellByIndex(cellidx, READ_LOCK);
            tfid.Cell = tcell->cellNum;
            afs_PutCell(tcell, READ_LOCK);
            tfid.Fid.Vnode = vnoid;
            tfid.Fid.Volume = volid;
            tfid.Fid.Unique = uniq;
        } else {
            afs_GetDynrootMountFid(&tfid);
            tfid.Fid.Vnode = VNUM_FROM_TYPEID(VN_TYPE_MOUNT, cellidx << 2);
            tfid.Fid.Unique = volid;
        }
        *avcp = tvc = afs_GetVCache(&tfid, treq);
        code = (tvc ? 0 : EIO);
        hit = 1;
        goto done;
    }

#ifdef AFS_LINUX26_ENV
    /*
     * Special case of the dynamic mount volume in a static root.
     * This is really unfortunate, but we need this for the translator.
     */
    if (adp == afs_globalVp && !afs_GetDynrootEnable() &&
        !strcmp(aname, AFS_DYNROOT_MOUNTNAME)) {
        struct VenusFid tfid;

        afs_GetDynrootMountFid(&tfid);
        *avcp = tvc = afs_GetVCache(&tfid, treq);
        code = 0;
        hit = 1;
        goto done;
    }
#endif

    Check_AtSys(adp, aname, &sysState, treq);
    tname = sysState.name;

    /* 1st Check_AtSys and lookup by tname is required here, for now,
     * because the dnlc is *not* told to remove entries for the parent
     * dir of file/dir op that afs_LocalHero likes, but dnlc is informed
     * if the cached entry for the parent dir is invalidated for a
     * non-local change.
     * Otherwise, we'd be able to do a dnlc lookup on an entry ending
     * w/@sys and know the dnlc was consistent with reality. */
    tvc = osi_dnlc_lookup(adp, tname, WRITE_LOCK);
    *avcp = tvc;                /* maybe wasn't initialized, but it is now */
    if (tvc) {
        if (no_read_access && vType(tvc) != VDIR && vType(tvc) != VLNK) {
            /* need read access on dir to stat non-directory / non-link */
            afs_PutVCache(tvc);
            *avcp = NULL;
            code = EACCES;
            goto done;
        }
#ifdef AFS_LINUX22_ENV
        if (tvc->mvstat == AFS_MVSTAT_ROOT) {   /* we don't trust the dnlc for root vcaches */
            AFS_RELE(AFSTOV(tvc));
            *avcp = 0;
        } else {
            code = 0;
            hit = 1;
            goto done;
        }
#else /* non - LINUX */
        code = 0;
        hit = 1;
        goto done;
#endif /* linux22 */
    }

    {                           /* sub-block just to reduce stack usage */
        struct dcache *tdc;
        afs_size_t dirOffset, dirLen;
        struct VenusFid tfid;

        /* now we have to lookup the next fid */
        if (afs_InReadDir(adp))
            tdc = adp->dcreaddir;
        else
            tdc = afs_GetDCache(adp, (afs_size_t) 0, treq,
                                &dirOffset, &dirLen, 1);
        if (!tdc) {
            *avcp = NULL;       /* redundant, but harmless */
            code = EIO;
            goto done;
        }

        /* now we will just call dir package with appropriate inode.
         * Dirs are always fetched in their entirety for now */
        ObtainReadLock(&adp->lock);
        ObtainReadLock(&tdc->lock);

        /*
         * Make sure that the data in the cache is current. There are two
         * cases we need to worry about:
         * 1. The cache data is being fetched by another process.
         * 2. The cache data is no longer valid
         *
         * If a readdir is in progress _in this thread_, it has a shared
         * lock on the vcache and has obtained current data, so we just
         * use that.  This eliminates several possible deadlocks.  
         */
        if (!afs_InReadDir(adp)) {
            while ((adp->f.states & CStatd)
                   && (tdc->dflags & DFFetching)
                   && afs_IsDCacheFresh(tdc, adp)) {
                ReleaseReadLock(&tdc->lock);
                ReleaseReadLock(&adp->lock);
                afs_osi_Sleep(&tdc->validPos);
                ObtainReadLock(&adp->lock);
                ObtainReadLock(&tdc->lock);
            }
            if (!(adp->f.states & CStatd)
                || !afs_IsDCacheFresh(tdc, adp)) {
                ReleaseReadLock(&tdc->lock);
                ReleaseReadLock(&adp->lock);
                afs_PutDCache(tdc);
                if (tname && tname != aname)
                    osi_FreeLargeSpace(tname);
                goto redo;
            }
        }

        /* Save the version number for when we call osi_dnlc_enter */
        hset(versionNo, tdc->f.versionNo);

        /*
         * check for, and handle "@sys" if it's there.  We should be able
         * to avoid the alloc and the strcpy with a little work, but it's
         * not pressing.  If there aren't any remote users (ie, via the 
         * NFS translator), we have a slightly easier job.
         * the faster way to do this is to check for *aname == '@' and if 
         * it's there, check for @sys, otherwise, assume there's no @sys 
         * then, if the lookup fails, check for .*@sys...
         */
        /* above now implemented by Check_AtSys and Next_AtSys */

        /* lookup the name in the appropriate dir, and return a cache entry
         * on the resulting fid */
        code =
            afs_dir_LookupOffset(tdc, sysState.name, &tfid.Fid,
                                 &dirCookie);

        /* If the first lookup doesn't succeed, maybe it's got @sys in the name */
        while (code == ENOENT && Next_AtSys(adp, treq, &sysState))
            code =
                afs_dir_LookupOffset(tdc, sysState.name, &tfid.Fid,
                                     &dirCookie);
        tname = sysState.name;

        ReleaseReadLock(&tdc->lock);
        if (!afs_InReadDir(adp))
            afs_PutDCache(tdc);
        if (code == ENOENT && afs_IsDynroot(adp) && dynrootRetry && !tryEvalOnly) {
            struct cell *tc;
            char *cn = (tname[0] == '.') ? tname + 1 : tname;
            ReleaseReadLock(&adp->lock);
            /* confirm it's not just hushed */
            tc = afs_GetCellByName(cn, WRITE_LOCK);
            if (tc) {
                if (tc->states & CHush) {
                    tc->states &= ~CHush;
                    ReleaseWriteLock(&tc->lock);
                    afs_DynrootInvalidate();
                    goto redo;
                }
                ReleaseWriteLock(&tc->lock);
            }
            /* Allow a second dynroot retry if the cell was hushed before */
            dynrootRetry = 0;
            if (tname[0] == '.')
                afs_LookupAFSDB(tname + 1);
            else
                afs_LookupAFSDB(tname);
            if (tname && tname != aname)
                osi_FreeLargeSpace(tname);
            goto redo;
        } else {
            ReleaseReadLock(&adp->lock);
        }

        /* new fid has same cell and volume */
        tfid.Cell = adp->f.fid.Cell;
        tfid.Fid.Volume = adp->f.fid.Fid.Volume;
        afs_Trace4(afs_iclSetp, CM_TRACE_LOOKUP, ICL_TYPE_POINTER, adp,
                   ICL_TYPE_STRING, tname, ICL_TYPE_FID, &tfid,
                   ICL_TYPE_INT32, code);

        if (code) {
            if (code == ENOENT) {
                /* The target name really doesn't exist (according to
                 * afs_dir_LookupOffset, anyway). */
                enoent_prohibited = 0;
            }
            goto done;
        }

        /* prefetch some entries, if the dir is currently open.  The variable
         * dirCookie tells us where to start prefetching from.
         */
        if (afs_ShouldTryBulkStat(adp)) {
            afs_int32 retry;
            /* if the entry is not in the cache, or is in the cache,
             * but hasn't been statd, then do a bulk stat operation.
             */
            do {
                retry = 0;
                ObtainReadLock(&afs_xvcache);
                tvc = afs_FindVCache(&tfid, &retry, 0 /* !stats,!lru */ );
                ReleaseReadLock(&afs_xvcache);
            } while (tvc && retry);

            if (!tvc || !(tvc->f.states & CStatd))
                bulkcode = afs_DoBulkStat(adp, dirCookie, treq);
            else
                bulkcode = 0;

            /* if the vcache isn't usable, release it */
            if (tvc && !(tvc->f.states & CStatd)) {
                afs_PutVCache(tvc);
                tvc = NULL;
            }
        } else {
            tvc = NULL;
            bulkcode = 0;
        }

        /* now get the status info, if we don't already have it */
        /* This is kind of weird, but we might wind up accidentally calling
         * RXAFS_Lookup because we happened upon a file which legitimately
         * has a 0 uniquifier. That is the result of allowing unique to wrap
         * to 0. This was fixed in AFS 3.4. For CForeign, Unique == 0 means that
         * the file has not yet been looked up.
         */
        if (!tvc) {
            if (!tfid.Fid.Unique && (adp->f.states & CForeign)) {
                tvc = afs_LookupVCache(&tfid, treq, adp, tname);
            }
            if (!tvc && !bulkcode) {    /* lookup failed or wasn't called */
                tvc = afs_GetVCache(&tfid, treq);
            }
        }                       /* if !tvc */
    }                           /* sub-block just to reduce stack usage */

    if (tvc) {
        if (adp->f.states & CForeign)
            tvc->f.states |= CForeign;
        tvc->f.parent.vnode = adp->f.fid.Fid.Vnode;
        tvc->f.parent.unique = adp->f.fid.Fid.Unique;
        tvc->f.states &= ~CBulkStat;

        if (afs_fakestat_enable == 2 && tvc->mvstat == AFS_MVSTAT_MTPT) {
            ObtainSharedLock(&tvc->lock, 680);
            if (!tvc->linkData) {
                UpgradeSToWLock(&tvc->lock, 681);
                code = afs_HandleLink(tvc, treq);
                ConvertWToRLock(&tvc->lock);
            } else {
                ConvertSToRLock(&tvc->lock);
                code = 0;
            }
            if (!code && !afs_strchr(tvc->linkData, ':'))
                force_eval = 1;
            ReleaseReadLock(&tvc->lock);
        }
        if (tvc->mvstat == AFS_MVSTAT_MTPT && (tvc->f.states & CMValid) && tvc->mvid.target_root != NULL)
          force_eval = 1; /* This is now almost for free, get it correct */

#if defined(UKERNEL)
        if (!(flags & AFS_LOOKUP_NOEVAL))
            /* don't eval mount points */
#endif /* UKERNEL */
            if (tvc->mvstat == AFS_MVSTAT_MTPT && force_eval) {
                /* a mt point, possibly unevaluated */
                struct volume *tvolp;

                ObtainWriteLock(&tvc->lock, 133);
                code = EvalMountPoint(tvc, adp, &tvolp, treq);
                ReleaseWriteLock(&tvc->lock);

                if (code) {
                    afs_PutVCache(tvc);
                    if (tvolp)
                        afs_PutVolume(tvolp, WRITE_LOCK);
                    goto done;
                }

                /* next, we want to continue using the target of the mt point */
                if (tvc->mvid.target_root && (tvc->f.states & CMValid)) {
                    struct vcache *uvc;
                    /* now lookup target, to set .. pointer */
                    afs_Trace2(afs_iclSetp, CM_TRACE_LOOKUP1,
                               ICL_TYPE_POINTER, tvc, ICL_TYPE_FID,
                               &tvc->f.fid);
                    uvc = tvc;  /* remember for later */

                    if (tvolp && (tvolp->states & VForeign)) {
                        /* XXXX tvolp has ref cnt on but not locked! XXX */
                        tvc =
                            afs_GetRootVCache(tvc->mvid.target_root, treq, tvolp);
                    } else {
                        tvc = afs_GetVCache(tvc->mvid.target_root, treq);
                    }
                    afs_PutVCache(uvc); /* we're done with it */

                    if (!tvc) {
                        code = EIO;
                        if (tvolp) {
                            afs_PutVolume(tvolp, WRITE_LOCK);
                        }
                        goto done;
                    }

                    /* now, if we came via a new mt pt (say because of a new
                     * release of a R/O volume), we must reevaluate the ..
                     * ptr to point back to the appropriate place */
                    if (tvolp) {
                        ObtainWriteLock(&tvc->lock, 134);
                        if (tvc->mvid.parent == NULL) {
                            tvc->mvid.parent =
                                osi_AllocSmallSpace(sizeof(struct VenusFid));
                        }
                        /* setup backpointer */
                        *tvc->mvid.parent = tvolp->dotdot;
                        ReleaseWriteLock(&tvc->lock);
                        afs_PutVolume(tvolp, WRITE_LOCK);
                    }
                } else {
                    afs_PutVCache(tvc);
                    code = ENODEV;
                    if (tvolp)
                        afs_PutVolume(tvolp, WRITE_LOCK);
                    goto done;
                }
            }
        *avcp = tvc;
        if (tvc && !VREFCOUNT_GT(tvc, 0)) {
            osi_Panic("TT3");
        }
        code = 0;
    } else {
        /* if we get here, we found something in a directory that couldn't
         * be located (a Multics "connection failure").  If the volume is
         * read-only, we try flushing this entry from the cache and trying
         * again. */
        if (!AFS_IS_DISCONNECTED) {
            if (pass == 0) {
                struct volume *tv;
                tv = afs_GetVolume(&adp->f.fid, treq, READ_LOCK);
                if (tv) {
                    if (tv->states & VRO) {
                        pass = 1;       /* try this *once* */
                        /* re-stat to get later version */
                        afs_StaleVCache(adp);
                        afs_PutVolume(tv, READ_LOCK);
                        goto redo;
                    }
                    afs_PutVolume(tv, READ_LOCK);
                }
            }
            code = EIO;
        } else {
            code = ENETDOWN;
        }
    }

  done:
    /* put the network buffer back, if need be */
    if (tname != aname && tname)
        osi_FreeLargeSpace(tname);
    if (code == 0) {

        if (afs_mariner)
            afs_AddMarinerName(aname, tvc);

#if defined(UKERNEL)
        if (!(flags & AFS_LOOKUP_NOEVAL)) {
            /* Here we don't enter the name into the DNLC because we want the
             * evaluated mount dir to be there (the vcache for the mounted
             * volume) rather than the vc of the mount point itself.  We can
             * still find the mount point's vc in the vcache by its fid. */
#endif /* UKERNEL */
            if (!hit && (force_eval || tvc->mvstat != AFS_MVSTAT_MTPT)) {
                osi_dnlc_enter(adp, aname, tvc, &versionNo);
            } else {
#ifdef AFS_LINUX20_ENV
                /* So Linux inode cache is up to date. */
                code = afs_VerifyVCache(tvc, treq);
#else
                afs_PutFakeStat(&fakestate);
                afs_DestroyReq(treq);
                AFS_DISCON_UNLOCK();
                return 0;       /* can't have been any errors if hit and !code */
#endif
            }
#if defined(UKERNEL)
        }
#endif
    }
    if (bulkcode)
        code = bulkcode;

    code = afs_CheckCode(code, treq, 19);
    if (code) {
        /* If there is an error, make sure *avcp is null.
         * Alphas panic otherwise - defect 10719.
         */
        *avcp = NULL;
    }
    if (code == ENOENT && enoent_prohibited) {
        /*
         * We got an ENOENT error, but we didn't get it while looking up the
         * dir entry in the relevant dir blob. That means we likely hit some
         * other internal error; don't allow us to return ENOENT in this case,
         * since some platforms cache ENOENT errors, and the target path name
         * may actually exist.
         */
        code = EIO;
    }

    afs_PutFakeStat(&fakestate);
    afs_DestroyReq(treq);
    AFS_DISCON_UNLOCK();
    return code;
}