[openafs.git] / src / vol / volume.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
 *
 * Portions Copyright (c) 2005-2008 Sine Nomine Associates
 */

/* 1/1/89: NB:  this stuff is all going to be replaced.  Don't take it too seriously */
/*

        System:         VICE-TWO
        Module:         volume.c
        Institution:    The Information Technology Center, Carnegie-Mellon University

 */

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

#include <roken.h>
#include <afs/opr.h>

#include <ctype.h>
#include <stddef.h>

#ifdef HAVE_SYS_FILE_H
#include <sys/file.h>
#endif

#ifdef AFS_PTHREAD_ENV
# include <opr/lock.h>
#else
# include <opr/lockstub.h>
#endif
#include <opr/ffs.h>
#include <opr/jhash.h>

#include <afs/afsint.h>

#include <rx/rx_queue.h>

#ifndef AFS_NT40_ENV
#if !defined(AFS_SGI_ENV)
#ifdef AFS_VFSINCL_ENV
#define VFS
#ifdef  AFS_SUN5_ENV
#include <sys/fs/ufs_fs.h>
#else
#if defined(AFS_DARWIN_ENV) || defined(AFS_XBSD_ENV)
#include <ufs/ufs/dinode.h>
#include <ufs/ffs/fs.h>
#else
#include <ufs/fs.h>
#endif
#endif
#else /* AFS_VFSINCL_ENV */
#if !defined(AFS_AIX_ENV) && !defined(AFS_LINUX20_ENV) && !defined(AFS_XBSD_ENV) && !defined(AFS_DARWIN_ENV)
#include <sys/fs.h>
#endif
#endif /* AFS_VFSINCL_ENV */
#endif /* AFS_SGI_ENV */
#endif /* !AFS_NT40_ENV */

#ifdef  AFS_AIX_ENV
#include <sys/vfs.h>
#else
#ifdef  AFS_HPUX_ENV
#include <mntent.h>
#else
#if     defined(AFS_SUN_ENV) || defined(AFS_SUN5_ENV)
#ifdef  AFS_SUN5_ENV
#include <sys/mnttab.h>
#include <sys/mntent.h>
#else
#include <mntent.h>
#endif
#else
#ifndef AFS_NT40_ENV
#if defined(AFS_SGI_ENV)
#include <mntent.h>
#else
#ifndef AFS_LINUX20_ENV
#include <fstab.h>              /* Need to find in libc 5, present in libc 6 */
#endif
#endif
#endif /* AFS_SGI_ENV */
#endif
#endif /* AFS_HPUX_ENV */
#endif

#include "nfs.h"
#include <afs/errors.h>
#include "lock.h"
#include "lwp.h"
#include <afs/afssyscalls.h>
#include "ihandle.h"
#include <afs/afsutil.h>
#include "daemon_com.h"
#include "fssync.h"
#include "salvsync.h"
#include "vnode.h"
#include "volume.h"
#include "partition.h"
#include "volume_inline.h"
#include "common.h"
#include "vutils.h"
#include <afs/dir.h>

#ifdef AFS_PTHREAD_ENV
pthread_mutex_t vol_glock_mutex;
pthread_mutex_t vol_trans_mutex;
pthread_cond_t vol_put_volume_cond;
pthread_cond_t vol_sleep_cond;
pthread_cond_t vol_init_attach_cond;
pthread_cond_t vol_vinit_cond;
int vol_attach_threads = 1;
#endif /* AFS_PTHREAD_ENV */

#ifdef AFS_DEMAND_ATTACH_FS
pthread_mutex_t vol_salvsync_mutex;

/*
 * Set this to 1 to disallow SALVSYNC communication in all threads; used
 * during shutdown, since the salvageserver may have gone away.
 */
static volatile sig_atomic_t vol_disallow_salvsync = 0;
#endif /* AFS_DEMAND_ATTACH_FS */

/**
 * has VShutdown_r been called / is VShutdown_r running?
 */
static int vol_shutting_down = 0;

/* Forward declarations */
static Volume *attach2(Error * ec, VolumeId volumeId, char *path,
                       struct DiskPartition64 *partp, Volume * vp,
                       int isbusy, int mode, int *acheckedOut);
static void ReallyFreeVolume(Volume * vp);
#ifdef AFS_DEMAND_ATTACH_FS
static void FreeVolume(Volume * vp);
#else /* !AFS_DEMAND_ATTACH_FS */
#define FreeVolume(vp) ReallyFreeVolume(vp)
static void VScanUpdateList(void);
#endif /* !AFS_DEMAND_ATTACH_FS */
static void VInitVolumeHeaderCache(afs_uint32 howMany);
static int GetVolumeHeader(Volume * vp);
static void ReleaseVolumeHeader(struct volHeader *hd);
static void FreeVolumeHeader(Volume * vp);
static void AddVolumeToHashTable(Volume * vp, VolumeId hashid);
static void DeleteVolumeFromHashTable(Volume * vp);
static int VHold_r(Volume * vp);
static void VGetBitmap_r(Error * ec, Volume * vp, VnodeClass class);
static void VReleaseVolumeHandles_r(Volume * vp);
static void VCloseVolumeHandles_r(Volume * vp);
static void LoadVolumeHeader(Error * ec, Volume * vp);
static int VCheckOffline(Volume * vp);
static int VCheckDetach(Volume * vp);
static Volume * GetVolume(Error * ec, Error * client_ec, VolumeId volumeId,
                          Volume * hint, const struct timespec *ts);

ProgramType programType;        /* The type of program using the package */
static VolumePackageOptions vol_opts;

/* extended volume package statistics */
VolPkgStats VStats;

#ifdef VOL_LOCK_DEBUG
pthread_t vol_glock_holder = 0;
#endif


/* this parameter needs to be tunable at runtime.
 * 128 was really inadequate for largish servers -- at 16384 volumes this
 * puts average chain length at 128, thus an average 65 deref's to find a volptr.
 * talk about bad spatial locality...
 *
 * an AVL or splay tree might work a lot better, but we'll just increase
 * the default hash table size for now
 */
#define DEFAULT_VOLUME_HASH_BITS 10
#define DEFAULT_VOLUME_HASH_SIZE opr_jhash_size(DEFAULT_VOLUME_HASH_BITS)
#define DEFAULT_VOLUME_HASH_MASK opr_jhash_mask(DEFAULT_VOLUME_HASH_BITS)
#define VOLUME_HASH(volumeId) \
    (opr_jhash_int(volumeId, 0) & VolumeHashTable.Mask)

/*
 * turn volume hash chains into partially ordered lists.
 * when the threshold is exceeded between two adjacent elements,
 * perform a chain rebalancing operation.
 *
 * keep the threshold high in order to keep cache line invalidates
 * low "enough" on SMPs
 */
#define VOLUME_HASH_REORDER_THRESHOLD 200

/*
 * when possible, don't just reorder single elements, but reorder
 * entire chains of elements at once.  a chain of elements that
 * exceed the element previous to the pivot by at least CHAIN_THRESH
 * accesses are moved in front of the chain whose elements have at
 * least CHAIN_THRESH less accesses than the pivot element
 */
#define VOLUME_HASH_REORDER_CHAIN_THRESH (VOLUME_HASH_REORDER_THRESHOLD / 2)

/*
 * The per volume uniquifier is bumped by 200 and and written to disk
 * every 200 file creates.
 */
#define VOLUME_UPDATE_UNIQUIFIER_BUMP 200

#include "rx/rx_queue.h"


VolumeHashTable_t VolumeHashTable = {
    DEFAULT_VOLUME_HASH_SIZE,
    DEFAULT_VOLUME_HASH_MASK,
    NULL
};


static void VInitVolumeHash(void);


#ifdef AFS_PTHREAD_ENV
/**
 * disk partition queue element
 */
typedef struct diskpartition_queue_t {
    struct rx_queue queue;             /**< queue header */
    struct DiskPartition64 *diskP;     /**< disk partition table entry */
} diskpartition_queue_t;

#ifndef AFS_DEMAND_ATTACH_FS

typedef struct vinitvolumepackage_thread_t {
    struct rx_queue queue;
    pthread_cond_t thread_done_cv;
    int n_threads_complete;
} vinitvolumepackage_thread_t;
static void * VInitVolumePackageThread(void * args);

#else  /* !AFS_DEMAND_ATTTACH_FS */
#define VINIT_BATCH_MAX_SIZE 512

/**
 * disk partition work queue
 */
struct partition_queue {
    struct rx_queue head;              /**< diskpartition_queue_t queue */
    pthread_mutex_t mutex;
    pthread_cond_t cv;
};

/**
 * volumes parameters for preattach
 */
struct volume_init_batch {
    struct rx_queue queue;               /**< queue header */
    int thread;                          /**< posting worker thread */
    int last;                            /**< indicates thread is done */
    int size;                            /**< number of volume ids in batch */
    Volume *batch[VINIT_BATCH_MAX_SIZE]; /**< volumes ids to preattach */
};

/**
 * volume parameters work queue
 */
struct volume_init_queue {
    struct rx_queue head;                /**< volume_init_batch queue */
    pthread_mutex_t mutex;
    pthread_cond_t cv;
};

/**
 * volume init worker thread parameters
 */
struct vinitvolumepackage_thread_param {
    int nthreads;                        /**< total number of worker threads */
    int thread;                          /**< thread number for this worker thread */
    struct partition_queue *pq;          /**< queue partitions to scan */
    struct volume_init_queue *vq;        /**< queue of volume to preattach */
};

static void *VInitVolumePackageThread(void *args);
static struct DiskPartition64 *VInitNextPartition(struct partition_queue *pq);
static VolumeId VInitNextVolumeId(DIR *dirp);
static int VInitPreAttachVolumes(int nthreads, struct volume_init_queue *vq);

#endif /* !AFS_DEMAND_ATTACH_FS */
#endif /* AFS_PTHREAD_ENV */

#ifndef AFS_DEMAND_ATTACH_FS
static int VAttachVolumesByPartition(struct DiskPartition64 *diskP,
                                     int * nAttached, int * nUnattached);
#endif /* AFS_DEMAND_ATTACH_FS */


#ifdef AFS_DEMAND_ATTACH_FS
/* demand attach fileserver extensions */

/* XXX
 * in the future we will support serialization of VLRU state into the fs_state
 * disk dumps
 *
 * these structures are the beginning of that effort
 */
struct VLRU_DiskHeader {
    struct versionStamp stamp;            /* magic and structure version number */
    afs_uint32 mtime;                     /* time of dump to disk */
    afs_uint32 num_records;               /* number of VLRU_DiskEntry records */
};

struct VLRU_DiskEntry {
    VolumeId vid;                       /* volume ID */
    afs_uint32 idx;                       /* generation */
    afs_uint32 last_get;                  /* timestamp of last get */
};

struct VLRU_StartupQueue {
    struct VLRU_DiskEntry * entry;
    int num_entries;
    int next_idx;
};

typedef struct vshutdown_thread_t {
    struct rx_queue q;
    pthread_mutex_t lock;
    pthread_cond_t cv;
    pthread_cond_t master_cv;
    int n_threads;
    int n_threads_complete;
    int vol_remaining;
    int schedule_version;
    int pass;
    byte n_parts;
    byte n_parts_done_pass;
    byte part_thread_target[VOLMAXPARTS+1];
    byte part_done_pass[VOLMAXPARTS+1];
    struct rx_queue * part_pass_head[VOLMAXPARTS+1];
    int stats[4][VOLMAXPARTS+1];
} vshutdown_thread_t;
static void * VShutdownThread(void * args);


static Volume * VAttachVolumeByVp_r(Error * ec, Volume * vp, int mode);
static int VCheckFree(Volume * vp);

/* VByP List */
static void AddVolumeToVByPList_r(Volume * vp);
static void DeleteVolumeFromVByPList_r(Volume * vp);
static void VVByPListBeginExclusive_r(struct DiskPartition64 * dp);
static void VVByPListEndExclusive_r(struct DiskPartition64 * dp);
static void VVByPListWait_r(struct DiskPartition64 * dp);

/* online salvager */
typedef enum {
    VCHECK_SALVAGE_OK = 0,         /**< no pending salvage */
    VCHECK_SALVAGE_SCHEDULED = 1,  /**< salvage has been scheduled */
    VCHECK_SALVAGE_ASYNC = 2,      /**< salvage being scheduled */
    VCHECK_SALVAGE_DENIED = 3,     /**< salvage not scheduled; denied */
    VCHECK_SALVAGE_FAIL = 4        /**< salvage not scheduled; failed */
} vsalvage_check;
static int VCheckSalvage(Volume * vp);
#if defined(SALVSYNC_BUILD_CLIENT) || defined(FSSYNC_BUILD_CLIENT)
static int VScheduleSalvage_r(Volume * vp);
#endif

/* Volume hash table */
static void VReorderHash_r(VolumeHashChainHead * head, Volume * pp, Volume * vp);
static void VHashBeginExclusive_r(VolumeHashChainHead * head);
static void VHashEndExclusive_r(VolumeHashChainHead * head);
static void VHashWait_r(VolumeHashChainHead * head);

/* shutdown */
static int ShutdownVByPForPass_r(struct DiskPartition64 * dp, int pass);
static int ShutdownVolumeWalk_r(struct DiskPartition64 * dp, int pass,
                                struct rx_queue ** idx);
static void ShutdownController(vshutdown_thread_t * params);
static void ShutdownCreateSchedule(vshutdown_thread_t * params);

/* VLRU */
static void VLRU_ComputeConstants(void);
static void VInitVLRU(void);
static void VLRU_Init_Node_r(Volume * vp);
static void VLRU_Add_r(Volume * vp);
static void VLRU_Delete_r(Volume * vp);
static void VLRU_UpdateAccess_r(Volume * vp);
static void * VLRU_ScannerThread(void * args);
static void VLRU_Scan_r(int idx);
static void VLRU_Promote_r(int idx);
static void VLRU_Demote_r(int idx);
static void VLRU_SwitchQueues(Volume * vp, int new_idx, int append);

/* soft detach */
static int VCheckSoftDetach(Volume * vp, afs_uint32 thresh);
static int VCheckSoftDetachCandidate(Volume * vp, afs_uint32 thresh);
static int VSoftDetachVolume_r(Volume * vp, afs_uint32 thresh);


pthread_key_t VThread_key;
VThreadOptions_t VThread_defaults = {
    0                           /**< allow salvsync */
};
#endif /* AFS_DEMAND_ATTACH_FS */


struct Lock vol_listLock;       /* Lock obtained when listing volumes:
                                 * prevents a volume from being missed
                                 * if the volume is attached during a
                                 * list volumes */


/* Common message used when the volume goes off line */
char *VSalvageMessage =
    "Files in this volume are currently unavailable; call operations";

int VInit;                      /* 0 - uninitialized,
                                 * 1 - initialized but not all volumes have been attached,
                                 * 2 - initialized and all volumes have been attached,
                                 * 3 - initialized, all volumes have been attached, and
                                 * VConnectFS() has completed. */

static int vinit_attach_abort = 0;

bit32 VolumeCacheCheck;         /* Incremented everytime a volume goes on line--
                                 * used to stamp volume headers and in-core
                                 * vnodes.  When the volume goes on-line the
                                 * vnode will be invalidated
                                 * access only with VOL_LOCK held */




/***************************************************/
/* Startup routines                                */
/***************************************************/

#if defined(FAST_RESTART) && defined(AFS_DEMAND_ATTACH_FS)
# error FAST_RESTART and DAFS are incompatible. For the DAFS equivalent \
        of FAST_RESTART, use the -unsafe-nosalvage fileserver argument
#endif

/**
 * assign default values to a VolumePackageOptions struct.
 *
 * Always call this on a VolumePackageOptions struct first, then set any
 * specific options you want, then call VInitVolumePackage2.
 *
 * @param[in]  pt   caller's program type
 * @param[out] opts volume package options
 */
void
VOptDefaults(ProgramType pt, VolumePackageOptions *opts)
{
    opts->nLargeVnodes = opts->nSmallVnodes = 5;
    opts->volcache = 0;

    opts->canScheduleSalvage = 0;
    opts->canUseFSSYNC = 0;
    opts->canUseSALVSYNC = 0;

    opts->interrupt_rxcall = NULL;
    opts->offline_timeout = -1;
    opts->offline_shutdown_timeout = -1;
    opts->usage_threshold = 128;
    opts->usage_rate_limit = 5;

#ifdef FAST_RESTART
    opts->unsafe_attach = 1;
#else /* !FAST_RESTART */
    opts->unsafe_attach = 0;
#endif /* !FAST_RESTART */

    switch (pt) {
    case fileServer:
        opts->canScheduleSalvage = 1;
        opts->canUseSALVSYNC = 1;
        break;

    case salvageServer:
        opts->canUseFSSYNC = 1;
        break;

    case volumeServer:
        opts->nLargeVnodes = 0;
        opts->nSmallVnodes = 0;

        opts->canScheduleSalvage = 1;
        opts->canUseFSSYNC = 1;
        break;

    default:
        /* noop */
        break;
    }
}

/**
 * Set VInit to a certain value, and signal waiters.
 *
 * @param[in] value  the value to set VInit to
 *
 * @pre VOL_LOCK held
 */
static void
VSetVInit_r(int value)
{
    VInit = value;
    opr_cv_broadcast(&vol_vinit_cond);
}

static_inline void
VLogOfflineTimeout(const char *type, afs_int32 timeout)
{
    if (timeout < 0) {
        return;
    }
    if (timeout == 0) {
        Log("VInitVolumePackage: Interrupting clients accessing %s "
            "immediately\n", type);
    } else {
        Log("VInitVolumePackage: Interrupting clients accessing %s "
            "after %ld second%s\n", type, (long)timeout, timeout==1?"":"s");
    }
}

int
VInitVolumePackage2(ProgramType pt, VolumePackageOptions * opts)
{
    int errors = 0;             /* Number of errors while finding vice partitions. */

    programType = pt;
    vol_opts = *opts;

#ifndef AFS_PTHREAD_ENV
    if (opts->offline_timeout != -1 || opts->offline_shutdown_timeout != -1) {
        Log("VInitVolumePackage: offline_timeout and/or "
            "offline_shutdown_timeout was specified, but the volume package "
            "does not support these for LWP builds\n");
        return -1;
    }
#endif
    VLogOfflineTimeout("volumes going offline", opts->offline_timeout);
    VLogOfflineTimeout("volumes going offline during shutdown",
                       opts->offline_shutdown_timeout);

    memset(&VStats, 0, sizeof(VStats));
    VStats.hdr_cache_size = 200;

    VInitPartitionPackage();
    VInitVolumeHash();
#ifdef AFS_DEMAND_ATTACH_FS
    if (programType == fileServer) {
        VInitVLRU();
    } else {
        VLRU_SetOptions(VLRU_SET_ENABLED, 0);
    }
    opr_Verify(pthread_key_create(&VThread_key, NULL) == 0);
#endif

    opr_mutex_init(&vol_glock_mutex);
    opr_mutex_init(&vol_trans_mutex);
    opr_cv_init(&vol_put_volume_cond);
    opr_cv_init(&vol_sleep_cond);
    opr_cv_init(&vol_init_attach_cond);
    opr_cv_init(&vol_vinit_cond);
#ifndef AFS_PTHREAD_ENV
    IOMGR_Initialize();
#endif /* AFS_PTHREAD_ENV */
    Lock_Init(&vol_listLock);

    srandom(time(0));           /* For VGetVolumeInfo */

#ifdef AFS_DEMAND_ATTACH_FS
    opr_mutex_init(&vol_salvsync_mutex);
#endif /* AFS_DEMAND_ATTACH_FS */

    /* Ok, we have done enough initialization that fileserver can
     * start accepting calls, even though the volumes may not be
     * available just yet.
     */
    VInit = 1;

#if defined(AFS_DEMAND_ATTACH_FS) && defined(SALVSYNC_BUILD_SERVER)
    if (programType == salvageServer) {
        SALVSYNC_salvInit();
    }
#endif /* AFS_DEMAND_ATTACH_FS */
#ifdef FSSYNC_BUILD_SERVER
    if (programType == fileServer) {
        FSYNC_fsInit();
    }
#endif
#if defined(AFS_DEMAND_ATTACH_FS) && defined(SALVSYNC_BUILD_CLIENT)
    if (VCanUseSALVSYNC()) {
        /* establish a connection to the salvager at this point */
        opr_Verify(VConnectSALV() != 0);
    }
#endif /* AFS_DEMAND_ATTACH_FS */

    if (opts->volcache > VStats.hdr_cache_size)
        VStats.hdr_cache_size = opts->volcache;
    VInitVolumeHeaderCache(VStats.hdr_cache_size);

    VInitVnodes(vLarge, opts->nLargeVnodes);
    VInitVnodes(vSmall, opts->nSmallVnodes);


    errors = VAttachPartitions();
    if (errors)
        return -1;

    if (programType != fileServer) {
        errors = VInitAttachVolumes(programType);
        if (errors) {
            return -1;
        }
    }

#ifdef FSSYNC_BUILD_CLIENT
    if (VCanUseFSSYNC()) {
        if (!VConnectFS()) {
#ifdef AFS_DEMAND_ATTACH_FS
            if (programType == salvageServer) {
                Log("Unable to connect to file server; aborted\n");
                exit(1);
            }
#endif /* AFS_DEMAND_ATTACH_FS */
            Log("Unable to connect to file server; will retry at need\n");
        }
    }
#endif /* FSSYNC_BUILD_CLIENT */
    return 0;
}


#if !defined(AFS_PTHREAD_ENV)
/**
 * Attach volumes in vice partitions
 *
 * @param[in]  pt         calling program type
 *
 * @return 0
 * @note This is the original, non-threaded version of attach parititions.
 *
 * @post VInit state is 2
 */
int
VInitAttachVolumes(ProgramType pt)
{
    opr_Assert(VInit==1);
    if (pt == fileServer) {
        struct DiskPartition64 *diskP;
        /* Attach all the volumes in this partition */
        for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
            int nAttached = 0, nUnattached = 0;
            opr_Verify(VAttachVolumesByPartition(diskP,
                                                 &nAttached, &nUnattached)
                            == 0);
        }
    }
    VOL_LOCK;
    VSetVInit_r(2);                     /* Initialized, and all volumes have been attached */
    LWP_NoYieldSignal(VInitAttachVolumes);
    VOL_UNLOCK;
    return 0;
}
#endif /* !AFS_PTHREAD_ENV */

#if defined(AFS_PTHREAD_ENV) && !defined(AFS_DEMAND_ATTACH_FS)
/**
 * Attach volumes in vice partitions
 *
 * @param[in]  pt         calling program type
 *
 * @return 0
 * @note Threaded version of attach parititions.
 *
 * @post VInit state is 2
 */
int
VInitAttachVolumes(ProgramType pt)
{
    opr_Assert(VInit==1);
    if (pt == fileServer) {
        struct DiskPartition64 *diskP;
        struct vinitvolumepackage_thread_t params;
        struct diskpartition_queue_t * dpq;
        int i, threads, parts;
        pthread_t tid;
        pthread_attr_t attrs;

        opr_cv_init(&params.thread_done_cv);
        queue_Init(&params);
        params.n_threads_complete = 0;

        /* create partition work queue */
        for (parts=0, diskP = DiskPartitionList; diskP; diskP = diskP->next, parts++) {
            dpq = malloc(sizeof(struct diskpartition_queue_t));
            opr_Assert(dpq != NULL);
            dpq->diskP = diskP;
            queue_Append(&params,dpq);
        }

        threads = min(parts, vol_attach_threads);

        if (threads > 1) {
            /* spawn off a bunch of initialization threads */
            opr_Verify(pthread_attr_init(&attrs) == 0);
            opr_Verify(pthread_attr_setdetachstate(&attrs,
                                                   PTHREAD_CREATE_DETACHED)
                            == 0);

            Log("VInitVolumePackage: beginning parallel fileserver startup\n");
            Log("VInitVolumePackage: using %d threads to attach volumes on %d partitions\n",
                threads, parts);

            VOL_LOCK;
            for (i=0; i < threads; i++) {
                AFS_SIGSET_DECL;
                AFS_SIGSET_CLEAR();
                opr_Verify(pthread_create(&tid, &attrs,
                                          &VInitVolumePackageThread,
                                          &params) == 0);
                AFS_SIGSET_RESTORE();
            }

            while(params.n_threads_complete < threads) {
                VOL_CV_WAIT(&params.thread_done_cv);
            }
            VOL_UNLOCK;

            opr_Verify(pthread_attr_destroy(&attrs) == 0);
        } else {
            /* if we're only going to run one init thread, don't bother creating
             * another LWP */
            Log("VInitVolumePackage: beginning single-threaded fileserver startup\n");
            Log("VInitVolumePackage: using 1 thread to attach volumes on %d partition(s)\n",
                parts);

            VInitVolumePackageThread(&params);
        }

        opr_cv_destroy(&params.thread_done_cv);
    }
    VOL_LOCK;
    VSetVInit_r(2);                     /* Initialized, and all volumes have been attached */
    opr_cv_broadcast(&vol_init_attach_cond);
    VOL_UNLOCK;
    return 0;
}

static void *
VInitVolumePackageThread(void * args) {

    struct DiskPartition64 *diskP;
    struct vinitvolumepackage_thread_t * params;
    struct diskpartition_queue_t * dpq;

    params = (vinitvolumepackage_thread_t *) args;


    VOL_LOCK;
    /* Attach all the volumes in this partition */
    while (queue_IsNotEmpty(params)) {
        int nAttached = 0, nUnattached = 0;

        if (vinit_attach_abort) {
            Log("Aborting initialization\n");
            goto done;
        }

        dpq = queue_First(params,diskpartition_queue_t);
        queue_Remove(dpq);
        VOL_UNLOCK;
        diskP = dpq->diskP;
        free(dpq);

        opr_Verify(VAttachVolumesByPartition(diskP, &nAttached,
                                             &nUnattached) == 0);

        VOL_LOCK;
    }

done:
    params->n_threads_complete++;
    opr_cv_signal(&params->thread_done_cv);
    VOL_UNLOCK;
    return NULL;
}
#endif /* AFS_PTHREAD_ENV && !AFS_DEMAND_ATTACH_FS */

#if defined(AFS_DEMAND_ATTACH_FS)
/**
 * Attach volumes in vice partitions
 *
 * @param[in]  pt         calling program type
 *
 * @return 0
 * @note Threaded version of attach partitions.
 *
 * @post VInit state is 2
 */
int
VInitAttachVolumes(ProgramType pt)
{
    opr_Assert(VInit==1);
    if (pt == fileServer) {

        struct DiskPartition64 *diskP;
        struct partition_queue pq;
        struct volume_init_queue vq;

        int i, threads, parts;
        pthread_t tid;
        pthread_attr_t attrs;

        /* create partition work queue */
        queue_Init(&pq);
        opr_cv_init(&pq.cv);
        opr_mutex_init(&pq.mutex);
        for (parts = 0, diskP = DiskPartitionList; diskP; diskP = diskP->next, parts++) {
            struct diskpartition_queue_t *dp;
            dp = malloc(sizeof(struct diskpartition_queue_t));
            opr_Assert(dp != NULL);
            dp->diskP = diskP;
            queue_Append(&pq, dp);
        }

        /* number of worker threads; at least one, not to exceed the number of partitions */
        threads = min(parts, vol_attach_threads);

        /* create volume work queue */
        queue_Init(&vq);
        opr_cv_init(&vq.cv);
        opr_mutex_init(&vq.mutex);

        opr_Verify(pthread_attr_init(&attrs) == 0);
        opr_Verify(pthread_attr_setdetachstate(&attrs,
                                               PTHREAD_CREATE_DETACHED) == 0);

        Log("VInitVolumePackage: beginning parallel fileserver startup\n");
        Log("VInitVolumePackage: using %d threads to pre-attach volumes on %d partitions\n",
                threads, parts);

        /* create threads to scan disk partitions. */
        for (i=0; i < threads; i++) {
            struct vinitvolumepackage_thread_param *params;
            AFS_SIGSET_DECL;

            params = malloc(sizeof(struct vinitvolumepackage_thread_param));
            opr_Assert(params);
            params->pq = &pq;
            params->vq = &vq;
            params->nthreads = threads;
            params->thread = i+1;

            AFS_SIGSET_CLEAR();
            opr_Verify(pthread_create(&tid, &attrs,
                                      &VInitVolumePackageThread,
                                      (void*)params) == 0);
            AFS_SIGSET_RESTORE();
        }

        VInitPreAttachVolumes(threads, &vq);

        opr_Verify(pthread_attr_destroy(&attrs) == 0);
        opr_cv_destroy(&pq.cv);
        opr_mutex_destroy(&pq.mutex);
        opr_cv_destroy(&vq.cv);
        opr_mutex_destroy(&vq.mutex);
    }

    VOL_LOCK;
    VSetVInit_r(2);                     /* Initialized, and all volumes have been attached */
    opr_cv_broadcast(&vol_init_attach_cond);
    VOL_UNLOCK;

    return 0;
}

/**
 * Volume package initialization worker thread. Scan partitions for volume
 * header files. Gather batches of volume ids and dispatch them to
 * the main thread to be preattached.  The volume preattachement is done
 * in the main thread to avoid global volume lock contention.
 */
static void *
VInitVolumePackageThread(void *args)
{
    struct vinitvolumepackage_thread_param *params;
    struct DiskPartition64 *partition;
    struct partition_queue *pq;
    struct volume_init_queue *vq;
    struct volume_init_batch *vb;

    opr_Assert(args);
    params = (struct vinitvolumepackage_thread_param *)args;
    pq = params->pq;
    vq = params->vq;
    opr_Assert(pq);
    opr_Assert(vq);

    vb = malloc(sizeof(struct volume_init_batch));
    opr_Assert(vb);
    vb->thread = params->thread;
    vb->last = 0;
    vb->size = 0;

    Log("Scanning partitions on thread %d of %d\n", params->thread, params->nthreads);
    while((partition = VInitNextPartition(pq))) {
        DIR *dirp;
        VolumeId vid;

        Log("Partition %s: pre-attaching volumes\n", partition->name);
        dirp = opendir(VPartitionPath(partition));
        if (!dirp) {
            Log("opendir on Partition %s failed, errno=%d!\n", partition->name, errno);
            continue;
        }
        while ((vid = VInitNextVolumeId(dirp))) {
            Volume *vp = calloc(1, sizeof(Volume));
            opr_Assert(vp);
            vp->device = partition->device;
            vp->partition = partition;
            vp->hashid = vid;
            queue_Init(&vp->vnode_list);
            queue_Init(&vp->rx_call_list);
            opr_cv_init(&V_attachCV(vp));

            vb->batch[vb->size++] = vp;
            if (vb->size == VINIT_BATCH_MAX_SIZE) {
                opr_mutex_enter(&vq->mutex);
                queue_Append(vq, vb);
                opr_cv_broadcast(&vq->cv);
                opr_mutex_exit(&vq->mutex);

                vb = malloc(sizeof(struct volume_init_batch));
                opr_Assert(vb);
                vb->thread = params->thread;
                vb->size = 0;
                vb->last = 0;
            }
        }
        closedir(dirp);
    }

    vb->last = 1;
    opr_mutex_enter(&vq->mutex);
    queue_Append(vq, vb);
    opr_cv_broadcast(&vq->cv);
    opr_mutex_exit(&vq->mutex);

    Log("Partition scan thread %d of %d ended\n", params->thread, params->nthreads);
    free(params);
    return NULL;
}

/**
 * Read next element from the pre-populated partition list.
 */
static struct DiskPartition64*
VInitNextPartition(struct partition_queue *pq)
{
    struct DiskPartition64 *partition;
    struct diskpartition_queue_t *dp; /* queue element */

    if (vinit_attach_abort) {
        Log("Aborting volume preattach thread.\n");
        return NULL;
    }

    /* get next partition to scan */
    opr_mutex_enter(&pq->mutex);
    if (queue_IsEmpty(pq)) {
        opr_mutex_exit(&pq->mutex);
        return NULL;
    }
    dp = queue_First(pq, diskpartition_queue_t);
    queue_Remove(dp);
    opr_mutex_exit(&pq->mutex);

    opr_Assert(dp);
    opr_Assert(dp->diskP);

    partition = dp->diskP;
    free(dp);
    return partition;
}

/**
 * Find next volume id on the partition.
 */
static VolumeId
VInitNextVolumeId(DIR *dirp)
{
    struct dirent *d;
    VolumeId vid = 0;
    char *ext;

    while((d = readdir(dirp))) {
        if (vinit_attach_abort) {
            Log("Aborting volume preattach thread.\n");
            break;
        }
        ext = strrchr(d->d_name, '.');
        if (d->d_name[0] == 'V' && ext && strcmp(ext, VHDREXT) == 0) {
            vid = VolumeNumber(d->d_name);
            if (vid) {
               break;
            }
            Log("Warning: bogus volume header file: %s\n", d->d_name);
        }
    }
    return vid;
}

/**
 * Preattach volumes in batches to avoid lock contention.
 */
static int
VInitPreAttachVolumes(int nthreads, struct volume_init_queue *vq)
{
    struct volume_init_batch *vb;
    int i;

    while (nthreads) {
        /* dequeue next volume */
        opr_mutex_enter(&vq->mutex);
        if (queue_IsEmpty(vq)) {
            opr_cv_wait(&vq->cv, &vq->mutex);
        }
        vb = queue_First(vq, volume_init_batch);
        queue_Remove(vb);
        opr_mutex_exit(&vq->mutex);

        if (vb->size) {
            VOL_LOCK;
            for (i = 0; i<vb->size; i++) {
                Volume *vp;
                Volume *dup;
                Error ec = 0;

                vp = vb->batch[i];
                dup = VLookupVolume_r(&ec, vp->hashid, NULL);
                if (ec) {
                    Log("Error looking up volume, code=%d\n", ec);
                }
                else if (dup) {
                    Log("Warning: Duplicate volume id %" AFS_VOLID_FMT " detected.\n", afs_printable_VolumeId_lu(vp->hashid));
                }
                else {
                    /* put pre-attached volume onto the hash table
                     * and bring it up to the pre-attached state */
                    AddVolumeToHashTable(vp, vp->hashid);
                    AddVolumeToVByPList_r(vp);
                    VLRU_Init_Node_r(vp);
                    VChangeState_r(vp, VOL_STATE_PREATTACHED);
                }
            }
            VOL_UNLOCK;
        }

        if (vb->last) {
            nthreads--;
        }
        free(vb);
    }
    return 0;
}
#endif /* AFS_DEMAND_ATTACH_FS */

#if !defined(AFS_DEMAND_ATTACH_FS)
/*
 * attach all volumes on a given disk partition
 */
static int
VAttachVolumesByPartition(struct DiskPartition64 *diskP, int * nAttached, int * nUnattached)
{
  DIR * dirp;
  struct dirent * dp;
  int ret = 0;

  Log("Partition %s: attaching volumes\n", diskP->name);
  dirp = opendir(VPartitionPath(diskP));
  if (!dirp) {
    Log("opendir on Partition %s failed!\n", diskP->name);
    return 1;
  }

  while ((dp = readdir(dirp))) {
    char *p;
    p = strrchr(dp->d_name, '.');

    if (vinit_attach_abort) {
      Log("Partition %s: abort attach volumes\n", diskP->name);
      goto done;
    }

    if (p != NULL && strcmp(p, VHDREXT) == 0) {
      Error error;
      Volume *vp;
      vp = VAttachVolumeByName(&error, diskP->name, dp->d_name,
                               V_VOLUPD);
      (*(vp ? nAttached : nUnattached))++;
      if (error == VOFFLINE)
        Log("Volume %d stays offline (/vice/offline/%s exists)\n", VolumeNumber(dp->d_name), dp->d_name);
      else if (GetLogLevel() >= 5) {
        Log("Partition %s: attached volume %d (%s)\n",
            diskP->name, VolumeNumber(dp->d_name),
            dp->d_name);
      }
      if (vp) {
        VPutVolume(vp);
      }
    }
  }

  Log("Partition %s: attached %d volumes; %d volumes not attached\n", diskP->name, *nAttached, *nUnattached);
done:
  closedir(dirp);
  return ret;
}
#endif /* !AFS_DEMAND_ATTACH_FS */

/***************************************************/
/* Shutdown routines                               */
/***************************************************/

/*
 * demand attach fs
 * highly multithreaded volume package shutdown
 *
 * with the demand attach fileserver extensions,
 * VShutdown has been modified to be multithreaded.
 * In order to achieve optimal use of many threads,
 * the shutdown code involves one control thread and
 * n shutdown worker threads.  The control thread
 * periodically examines the number of volumes available
 * for shutdown on each partition, and produces a worker
 * thread allocation schedule.  The idea is to eliminate
 * redundant scheduling computation on the workers by
 * having a single master scheduler.
 *
 * The scheduler's objectives are:
 * (1) fairness
 *   each partition with volumes remaining gets allocated
 *   at least 1 thread (assuming sufficient threads)
 * (2) performance
 *   threads are allocated proportional to the number of
 *   volumes remaining to be offlined.  This ensures that
 *   the OS I/O scheduler has many requests to elevator
 *   seek on partitions that will (presumably) take the
 *   longest amount of time (from now) to finish shutdown
 * (3) keep threads busy
 *   when there are extra threads, they are assigned to
 *   partitions using a simple round-robin algorithm
 *
 * In the future, we may wish to add the ability to adapt
 * to the relative performance patterns of each disk
 * partition.
 *
 *
 * demand attach fs
 * multi-step shutdown process
 *
 * demand attach shutdown is a four-step process. Each
 * shutdown "pass" shuts down increasingly more difficult
 * volumes.  The main purpose is to achieve better cache
 * utilization during shutdown.
 *
 * pass 0
 *   shutdown volumes in the unattached, pre-attached
 *   and error states
 * pass 1
 *   shutdown attached volumes with cached volume headers
 * pass 2
 *   shutdown all volumes in non-exclusive states
 * pass 3
 *   shutdown all remaining volumes
 */

#ifdef AFS_DEMAND_ATTACH_FS

void
VShutdown_r(void)
{
    int i;
    struct DiskPartition64 * diskP;
    struct diskpartition_queue_t * dpq;
    vshutdown_thread_t params;
    pthread_t tid;
    pthread_attr_t attrs;

    memset(&params, 0, sizeof(vshutdown_thread_t));

    if (VInit < 2) {
        Log("VShutdown:  aborting attach volumes\n");
        vinit_attach_abort = 1;
        VOL_CV_WAIT(&vol_init_attach_cond);
    }

    for (params.n_parts=0, diskP = DiskPartitionList;
         diskP; diskP = diskP->next, params.n_parts++);

    Log("VShutdown:  shutting down on-line volumes on %d partition%s...\n",
        params.n_parts, params.n_parts > 1 ? "s" : "");

    vol_shutting_down = 1;

    if (vol_attach_threads > 1) {
        /* prepare for parallel shutdown */
        params.n_threads = vol_attach_threads;
        opr_mutex_init(&params.lock);
        opr_cv_init(&params.cv);
        opr_cv_init(&params.master_cv);
        opr_Verify(pthread_attr_init(&attrs) == 0);
        opr_Verify(pthread_attr_setdetachstate(&attrs,
                                               PTHREAD_CREATE_DETACHED) == 0);
        queue_Init(&params);

        /* setup the basic partition information structures for
         * parallel shutdown */
        for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
            /* XXX debug */
            struct rx_queue * qp, * nqp;
            Volume * vp;
            int count = 0;

            VVByPListWait_r(diskP);
            VVByPListBeginExclusive_r(diskP);

            /* XXX debug */
            for (queue_Scan(&diskP->vol_list, qp, nqp, rx_queue)) {
                vp = (Volume *)((char *)qp - offsetof(Volume, vol_list));
                if (vp->header)
                    count++;
            }
            Log("VShutdown: partition %s has %d volumes with attached headers\n",
                VPartitionPath(diskP), count);


            /* build up the pass 0 shutdown work queue */
            dpq = malloc(sizeof(struct diskpartition_queue_t));
            opr_Assert(dpq != NULL);
            dpq->diskP = diskP;
            queue_Prepend(&params, dpq);

            params.part_pass_head[diskP->index] = queue_First(&diskP->vol_list, rx_queue);
        }

        Log("VShutdown:  beginning parallel fileserver shutdown\n");
        Log("VShutdown:  using %d threads to offline volumes on %d partition%s\n",
            vol_attach_threads, params.n_parts, params.n_parts > 1 ? "s" : "" );

        /* do pass 0 shutdown */
        opr_mutex_enter(&params.lock);
        for (i=0; i < params.n_threads; i++) {
            opr_Verify(pthread_create(&tid, &attrs, &VShutdownThread,
                                      &params) == 0);
        }

        /* wait for all the pass 0 shutdowns to complete */
        while (params.n_threads_complete < params.n_threads) {
            CV_WAIT(&params.master_cv, &params.lock);
        }
        params.n_threads_complete = 0;
        params.pass = 1;
        opr_cv_broadcast(&params.cv);
        opr_mutex_exit(&params.lock);

        Log("VShutdown:  pass 0 completed using the 1 thread per partition algorithm\n");
        Log("VShutdown:  starting passes 1 through 3 using finely-granular mp-fast algorithm\n");

        /* run the parallel shutdown scheduler. it will drop the glock internally */
        ShutdownController(&params);

        /* wait for all the workers to finish pass 3 and terminate */
        while (params.pass < 4) {
            VOL_CV_WAIT(&params.cv);
        }

        opr_Verify(pthread_attr_destroy(&attrs) == 0);
        opr_cv_destroy(&params.cv);
        opr_cv_destroy(&params.master_cv);
        opr_mutex_destroy(&params.lock);

        /* drop the VByPList exclusive reservations */
        for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
            VVByPListEndExclusive_r(diskP);
            Log("VShutdown:  %s stats : (pass[0]=%d, pass[1]=%d, pass[2]=%d, pass[3]=%d)\n",
                VPartitionPath(diskP),
                params.stats[0][diskP->index],
                params.stats[1][diskP->index],
                params.stats[2][diskP->index],
                params.stats[3][diskP->index]);
        }

        Log("VShutdown:  shutdown finished using %d threads\n", params.n_threads);
    } else {
        /* if we're only going to run one shutdown thread, don't bother creating
         * another LWP */
        Log("VShutdown:  beginning single-threaded fileserver shutdown\n");

        for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
            VShutdownByPartition_r(diskP);
        }
    }

    Log("VShutdown:  complete.\n");
}

#else /* AFS_DEMAND_ATTACH_FS */

void
VShutdown_r(void)
{
    int i;
    Volume *vp, *np;
    afs_int32 code;

    if (VInit < 2) {
        Log("VShutdown:  aborting attach volumes\n");
        vinit_attach_abort = 1;
#ifdef AFS_PTHREAD_ENV
        VOL_CV_WAIT(&vol_init_attach_cond);
#else
        LWP_WaitProcess(VInitAttachVolumes);
#endif /* AFS_PTHREAD_ENV */
    }

    Log("VShutdown:  shutting down on-line volumes...\n");
    vol_shutting_down = 1;
    for (i = 0; i < VolumeHashTable.Size; i++) {
        /* try to hold first volume in the hash table */
        for (queue_Scan(&VolumeHashTable.Table[i],vp,np,Volume)) {
            code = VHold_r(vp);
            if (code == 0) {
                if (GetLogLevel() >= 5)
                    Log("VShutdown:  Attempting to take volume %" AFS_VOLID_FMT " offline.\n",
                        afs_printable_VolumeId_lu(vp->hashid));

                /* next, take the volume offline (drops reference count) */
                VOffline_r(vp, "File server was shut down");
            }
        }
    }
    Log("VShutdown:  complete.\n");
}
#endif /* AFS_DEMAND_ATTACH_FS */


void
VShutdown(void)
{
    opr_Assert(VInit>0);
    VOL_LOCK;
    VShutdown_r();
    VOL_UNLOCK;
}

/**
 * stop new activity (e.g. SALVSYNC) from occurring
 *
 * Use this to make the volume package less busy; for example, during
 * shutdown. This doesn't actually shutdown/detach anything in the
 * volume package, but prevents certain processes from ocurring. For
 * example, preventing new SALVSYNC communication in DAFS. In theory, we
 * could also use this to prevent new volume attachment, or prevent
 * other programs from checking out volumes, etc.
 */
void
VSetTranquil(void)
{
#ifdef AFS_DEMAND_ATTACH_FS
    /* make sure we don't try to contact the salvageserver, since it may
     * not be around anymore */
    vol_disallow_salvsync = 1;
#endif
}

#ifdef AFS_DEMAND_ATTACH_FS
/*
 * demand attach fs
 * shutdown control thread
 */
static void
ShutdownController(vshutdown_thread_t * params)
{
    /* XXX debug */
    struct DiskPartition64 * diskP;
    Device id;
    vshutdown_thread_t shadow;

    ShutdownCreateSchedule(params);

    while ((params->pass < 4) &&
           (params->n_threads_complete < params->n_threads)) {
        /* recompute schedule once per second */

        memcpy(&shadow, params, sizeof(vshutdown_thread_t));

        VOL_UNLOCK;
        /* XXX debug */
        Log("ShutdownController:  schedule version=%d, vol_remaining=%d, pass=%d\n",
            shadow.schedule_version, shadow.vol_remaining, shadow.pass);
        Log("ShutdownController:  n_threads_complete=%d, n_parts_done_pass=%d\n",
            shadow.n_threads_complete, shadow.n_parts_done_pass);
        for (diskP = DiskPartitionList; diskP; diskP=diskP->next) {
            id = diskP->index;
            Log("ShutdownController:  part[%d] : (len=%d, thread_target=%d, done_pass=%d, pass_head=%p)\n",
                id,
                diskP->vol_list.len,
                shadow.part_thread_target[id],
                shadow.part_done_pass[id],
                shadow.part_pass_head[id]);
        }

        sleep(1);
        VOL_LOCK;

        ShutdownCreateSchedule(params);
    }
}

/* create the shutdown thread work schedule.
 * this scheduler tries to implement fairness
 * by allocating at least 1 thread to each
 * partition with volumes to be shutdown,
 * and then it attempts to allocate remaining
 * threads based upon the amount of work left
 */
static void
ShutdownCreateSchedule(vshutdown_thread_t * params)
{
    struct DiskPartition64 * diskP;
    int sum, thr_workload, thr_left;
    int part_residue[VOLMAXPARTS+1];
    Device id;

    /* compute the total number of outstanding volumes */
    sum = 0;
    for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
        sum += diskP->vol_list.len;
    }

    params->schedule_version++;
    params->vol_remaining = sum;

    if (!sum)
        return;

    /* compute average per-thread workload */
    thr_workload = sum / params->n_threads;
    if (sum % params->n_threads)
        thr_workload++;

    thr_left = params->n_threads;
    memset(&part_residue, 0, sizeof(part_residue));

    /* for fairness, give every partition with volumes remaining
     * at least one thread */
    for (diskP = DiskPartitionList; diskP && thr_left; diskP = diskP->next) {
        id = diskP->index;
        if (diskP->vol_list.len) {
            params->part_thread_target[id] = 1;
            thr_left--;
        } else {
            params->part_thread_target[id] = 0;
        }
    }

    if (thr_left && thr_workload) {
        /* compute length-weighted workloads */
        int delta;

        for (diskP = DiskPartitionList; diskP && thr_left; diskP = diskP->next) {
            id = diskP->index;
            delta = (diskP->vol_list.len / thr_workload) -
                params->part_thread_target[id];
            if (delta < 0) {
                continue;
            }
            if (delta < thr_left) {
                params->part_thread_target[id] += delta;
                thr_left -= delta;
            } else {
                params->part_thread_target[id] += thr_left;
                thr_left = 0;
                break;
            }
        }
    }

    if (thr_left) {
        /* try to assign any leftover threads to partitions that
         * had volume lengths closer to needing thread_target+1 */
        int max_residue, max_id = 0;

        /* compute the residues */
        for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
            id = diskP->index;
            part_residue[id] = diskP->vol_list.len -
                (params->part_thread_target[id] * thr_workload);
        }

        /* now try to allocate remaining threads to partitions with the
         * highest residues */
        while (thr_left) {
            max_residue = 0;
            for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
                id = diskP->index;
                if (part_residue[id] > max_residue) {
                    max_residue = part_residue[id];
                    max_id = id;
                }
            }

            if (!max_residue) {
                break;
            }

            params->part_thread_target[max_id]++;
            thr_left--;
            part_residue[max_id] = 0;
        }
    }

    if (thr_left) {
        /* punt and give any remaining threads equally to each partition */
        int alloc;
        if (thr_left >= params->n_parts) {
            alloc = thr_left / params->n_parts;
            for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
                id = diskP->index;
                params->part_thread_target[id] += alloc;
                thr_left -= alloc;
            }
        }

        /* finish off the last of the threads */
        for (diskP = DiskPartitionList; thr_left && diskP; diskP = diskP->next) {
            id = diskP->index;
            params->part_thread_target[id]++;
            thr_left--;
        }
    }
}

/* worker thread for parallel shutdown */
static void *
VShutdownThread(void * args)
{
    vshutdown_thread_t * params;
    int found, pass, schedule_version_save, count;
    struct DiskPartition64 *diskP;
    struct diskpartition_queue_t * dpq;
    Device id;

    params = (vshutdown_thread_t *) args;

    /* acquire the shutdown pass 0 lock */
    opr_mutex_enter(&params->lock);

    /* if there's still pass 0 work to be done,
     * get a work entry, and do a pass 0 shutdown */
    if (queue_IsNotEmpty(params)) {
        dpq = queue_First(params, diskpartition_queue_t);
        queue_Remove(dpq);
        opr_mutex_exit(&params->lock);
        diskP = dpq->diskP;
        free(dpq);
        id = diskP->index;

        count = 0;
        while (ShutdownVolumeWalk_r(diskP, 0, &params->part_pass_head[id]))
            count++;
        params->stats[0][diskP->index] = count;
        opr_mutex_enter(&params->lock);
    }

    params->n_threads_complete++;
    if (params->n_threads_complete == params->n_threads) {
        /* notify control thread that all workers have completed pass 0 */
        opr_cv_signal(&params->master_cv);
    }
    while (params->pass == 0) {
        opr_cv_wait(&params->cv, &params->lock);
    }

    /* switch locks */
    opr_mutex_exit(&params->lock);
    VOL_LOCK;

    pass = params->pass;
    opr_Assert(pass > 0);

    /* now escalate through the more complicated shutdowns */
    while (pass <= 3) {
        schedule_version_save = params->schedule_version;
        found = 0;
        /* find a disk partition to work on */
        for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
            id = diskP->index;
            if (params->part_thread_target[id] && !params->part_done_pass[id]) {
                params->part_thread_target[id]--;
                found = 1;
                break;
            }
        }

        if (!found) {
            /* hmm. for some reason the controller thread couldn't find anything for
             * us to do. let's see if there's anything we can do */
            for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
                id = diskP->index;
                if (diskP->vol_list.len && !params->part_done_pass[id]) {
                    found = 1;
                    break;
                } else if (!params->part_done_pass[id]) {
                    params->part_done_pass[id] = 1;
                    params->n_parts_done_pass++;
                    if (pass == 3) {
                        Log("VShutdown:  done shutting down volumes on partition %s.\n",
                            VPartitionPath(diskP));
                    }
                }
            }
        }

        /* do work on this partition until either the controller
         * creates a new schedule, or we run out of things to do
         * on this partition */
        if (found) {
            count = 0;
            while (!params->part_done_pass[id] &&
                   (schedule_version_save == params->schedule_version)) {
                /* ShutdownVolumeWalk_r will drop the glock internally */
                if (!ShutdownVolumeWalk_r(diskP, pass, &params->part_pass_head[id])) {
                    if (!params->part_done_pass[id]) {
                        params->part_done_pass[id] = 1;
                        params->n_parts_done_pass++;
                        if (pass == 3) {
                            Log("VShutdown:  done shutting down volumes on partition %s.\n",
                                VPartitionPath(diskP));
                        }
                    }
                    break;
                }
                count++;
            }

            params->stats[pass][id] += count;
        } else {
            /* ok, everyone is done this pass, proceed */

            /* barrier lock */
            params->n_threads_complete++;
            while (params->pass == pass) {
                if (params->n_threads_complete == params->n_threads) {
                    /* we are the last thread to complete, so we will
                     * reinitialize worker pool state for the next pass */
                    params->n_threads_complete = 0;
                    params->n_parts_done_pass = 0;
                    params->pass++;
                    for (diskP = DiskPartitionList; diskP; diskP = diskP->next) {
                        id = diskP->index;
                        params->part_done_pass[id] = 0;
                        params->part_pass_head[id] = queue_First(&diskP->vol_list, rx_queue);
                    }

                    /* compute a new thread schedule before releasing all the workers */
                    ShutdownCreateSchedule(params);

                    /* wake up all the workers */
                    opr_cv_broadcast(&params->cv);

                    VOL_UNLOCK;
                    Log("VShutdown:  pass %d completed using %d threads on %d partitions\n",
                        pass, params->n_threads, params->n_parts);
                    VOL_LOCK;
                } else {
                    VOL_CV_WAIT(&params->cv);
                }
            }
            pass = params->pass;
        }

        /* for fairness */
        VOL_UNLOCK;
        pthread_yield();
        VOL_LOCK;
    }

    VOL_UNLOCK;

    return NULL;
}

/* shut down all volumes on a given disk partition
 *
 * note that this function will not allow mp-fast
 * shutdown of a partition */
int
VShutdownByPartition_r(struct DiskPartition64 * dp)
{
    int pass;
    int pass_stats[4];
    int total;

    /* wait for other exclusive ops to finish */
    VVByPListWait_r(dp);

    /* begin exclusive access */
    VVByPListBeginExclusive_r(dp);

    /* pick the low-hanging fruit first,
     * then do the complicated ones last
     * (has the advantage of keeping
     *  in-use volumes up until the bitter end) */
    for (pass = 0, total=0; pass < 4; pass++) {
        pass_stats[pass] = ShutdownVByPForPass_r(dp, pass);
        total += pass_stats[pass];
    }

    /* end exclusive access */
    VVByPListEndExclusive_r(dp);

    Log("VShutdownByPartition:  shut down %d volumes on %s (pass[0]=%d, pass[1]=%d, pass[2]=%d, pass[3]=%d)\n",
        total, VPartitionPath(dp), pass_stats[0], pass_stats[1], pass_stats[2], pass_stats[3]);

    return 0;
}

/* internal shutdown functionality
 *
 * for multi-pass shutdown:
 * 0 to only "shutdown" {pre,un}attached and error state volumes
 * 1 to also shutdown attached volumes w/ volume header loaded
 * 2 to also shutdown attached volumes w/o volume header loaded
 * 3 to also shutdown exclusive state volumes
 *
 * caller MUST hold exclusive access on the hash chain
 * because we drop vol_glock_mutex internally
 *
 * this function is reentrant for passes 1--3
 * (e.g. multiple threads can cooperate to
 *  shutdown a partition mp-fast)
 *
 * pass 0 is not scaleable because the volume state data is
 * synchronized by vol_glock mutex, and the locking overhead
 * is too high to drop the lock long enough to do linked list
 * traversal
 */
static int
ShutdownVByPForPass_r(struct DiskPartition64 * dp, int pass)
{
    struct rx_queue * q = queue_First(&dp->vol_list, rx_queue);
    int i = 0;
    const char *pass_strs[4] = {"{un/pre}attached vols", "vols w/ vol header loaded", "vols w/o vol header loaded", "vols with exclusive state"};

    while (ShutdownVolumeWalk_r(dp, pass, &q)) {
        i++;
        if (0 == i%100) {
            Log("VShutdownByPartition:  ... shut down %d volumes on %s in pass %d (%s)\n", i, VPartitionPath(dp), pass, pass_strs[pass]);
        }
    }

    return i;
}

/* conditionally shutdown one volume on partition dp
 * returns 1 if a volume was shutdown in this pass,
 * 0 otherwise */
static int
ShutdownVolumeWalk_r(struct DiskPartition64 * dp, int pass,
                     struct rx_queue ** idx)
{
    struct rx_queue *qp, *nqp;
    Volume * vp;

    qp = *idx;

    for (queue_ScanFrom(&dp->vol_list, qp, qp, nqp, rx_queue)) {
        vp = (Volume *) (((char *)qp) - offsetof(Volume, vol_list));

        switch (pass) {
        case 0:
            if ((V_attachState(vp) != VOL_STATE_UNATTACHED) &&
                (V_attachState(vp) != VOL_STATE_ERROR) &&
                (V_attachState(vp) != VOL_STATE_DELETED) &&
                (V_attachState(vp) != VOL_STATE_PREATTACHED)) {
                break;
            }
        case 1:
            if ((V_attachState(vp) == VOL_STATE_ATTACHED) &&
                (vp->header == NULL)) {
                break;
            }
        case 2:
            if (VIsExclusiveState(V_attachState(vp))) {
                break;
            }
        case 3:
            *idx = nqp;
            DeleteVolumeFromVByPList_r(vp);
            VShutdownVolume_r(vp);
            vp = NULL;
            return 1;
        }
    }

    return 0;
}

/*
 * shutdown a specific volume
 */
/* caller MUST NOT hold a heavyweight ref on vp */
int
VShutdownVolume_r(Volume * vp)
{
    int code;

    VCreateReservation_r(vp);

    if (GetLogLevel() >= 5) {
        Log("VShutdownVolume_r:  vid=%" AFS_VOLID_FMT ", device=%d, state=%u\n",
            afs_printable_VolumeId_lu(vp->hashid), vp->partition->device,
            (unsigned int) V_attachState(vp));
    }

    /* wait for other blocking ops to finish */
    VWaitExclusiveState_r(vp);

    opr_Assert(VIsValidState(V_attachState(vp)));

    switch(V_attachState(vp)) {
    case VOL_STATE_SALVAGING:
        /* Leave salvaging volumes alone. Any in-progress salvages will
         * continue working after viced shuts down. This is intentional.
         */

    case VOL_STATE_PREATTACHED:
    case VOL_STATE_ERROR:
        VChangeState_r(vp, VOL_STATE_UNATTACHED);
    case VOL_STATE_UNATTACHED:
    case VOL_STATE_DELETED:
        break;
    case VOL_STATE_GOING_OFFLINE:
    case VOL_STATE_SHUTTING_DOWN:
    case VOL_STATE_ATTACHED:
        code = VHold_r(vp);
        if (!code) {
            if (GetLogLevel() >= 5)
                Log("VShutdown:  Attempting to take volume %" AFS_VOLID_FMT " offline.\n",
                    afs_printable_VolumeId_lu(vp->hashid));

            /* take the volume offline (drops reference count) */
            VOffline_r(vp, "File server was shut down");
        }
        break;
    default:
        break;
    }

    VCancelReservation_r(vp);
    vp = NULL;
    return 0;
}
#endif /* AFS_DEMAND_ATTACH_FS */


/***************************************************/
/* Header I/O routines                             */
/***************************************************/

static const char *
HeaderName(bit32 magic)
{
    switch (magic) {
    case VOLUMEINFOMAGIC:
        return "volume info";
    case SMALLINDEXMAGIC:
        return "small index";
    case LARGEINDEXMAGIC:
        return "large index";
    case LINKTABLEMAGIC:
        return "link table";
    }
    return "unknown";
}

/* open a descriptor for the inode (h),
 * read in an on-disk structure into buffer (to) of size (size),
 * verify versionstamp in structure has magic (magic) and
 * optionally verify version (version) if (version) is nonzero
 */
static void
ReadHeader(Error * ec, IHandle_t * h, char *to, int size, bit32 magic,
           bit32 version)
{
    struct versionStamp *vsn;
    FdHandle_t *fdP;
    afs_sfsize_t nbytes;
    afs_ino_str_t stmp;

    *ec = 0;
    if (h == NULL) {
        Log("ReadHeader: Null inode handle argument for %s header file.\n",
            HeaderName(magic));
        *ec = VSALVAGE;
        return;
    }

    fdP = IH_OPEN(h);
    if (fdP == NULL) {
        Log("ReadHeader: Failed to open %s header file "
            "(volume=%" AFS_VOLID_FMT ", inode=%s); errno=%d\n", HeaderName(magic), afs_printable_VolumeId_lu(h->ih_vid),
            PrintInode(stmp, h->ih_ino), errno);
        *ec = VSALVAGE;
        return;
    }

    vsn = (struct versionStamp *)to;
    nbytes = FDH_PREAD(fdP, to, size, 0);
    if (nbytes < 0) {
        Log("ReadHeader: Failed to read %s header file "
            "(volume=%" AFS_VOLID_FMT ", inode=%s); errno=%d\n", HeaderName(magic), afs_printable_VolumeId_lu(h->ih_vid),
            PrintInode(stmp, h->ih_ino), errno);
        *ec = VSALVAGE;
        FDH_REALLYCLOSE(fdP);
        return;
    }
    if (nbytes != size) {
        Log("ReadHeader: Incorrect number of bytes read from %s header file "
            "(volume=%" AFS_VOLID_FMT ", inode=%s); expected=%d, read=%d\n",
            HeaderName(magic), afs_printable_VolumeId_lu(h->ih_vid), 
            PrintInode(stmp, h->ih_ino), size, (int)nbytes);
        *ec = VSALVAGE;
        FDH_REALLYCLOSE(fdP);
        return;
    }
    if (vsn->magic != magic) {
        Log("ReadHeader: Incorrect magic for %s header file "
            "(volume=%" AFS_VOLID_FMT ", inode=%s); expected=0x%x, read=0x%x\n",
            HeaderName(magic), afs_printable_VolumeId_lu(h->ih_vid),
            PrintInode(stmp, h->ih_ino), magic, vsn->magic);
        *ec = VSALVAGE;
        FDH_REALLYCLOSE(fdP);
        return;
    }

    FDH_CLOSE(fdP);

    /* Check is conditional, in case caller wants to inspect version himself */
    if (version && vsn->version != version) {
        Log("ReadHeader: Incorrect version for %s header file "
            "(volume=%" AFS_VOLID_FMT ", inode=%s); expected=%x, read=%x\n",
            HeaderName(magic), afs_printable_VolumeId_lu(h->ih_vid), PrintInode(stmp, h->ih_ino),
            version, vsn->version);
        *ec = VSALVAGE;
    }
}

void
WriteVolumeHeader_r(Error * ec, Volume * vp)
{
    IHandle_t *h = V_diskDataHandle(vp);
    FdHandle_t *fdP;

    *ec = 0;

    fdP = IH_OPEN(h);
    if (fdP == NULL) {
        *ec = VSALVAGE;
        return;
    }
    if (FDH_PWRITE(fdP, (char *)&V_disk(vp), sizeof(V_disk(vp)), 0)
        != sizeof(V_disk(vp))) {
        *ec = VSALVAGE;
        FDH_REALLYCLOSE(fdP);
        return;
    }
    FDH_CLOSE(fdP);
}

/* VolumeHeaderToDisk
 * Allows for storing 64 bit inode numbers in on-disk volume header
 * file.
 */
/* convert in-memory representation of a volume header to the
 * on-disk representation of a volume header */
void
VolumeHeaderToDisk(VolumeDiskHeader_t * dh, VolumeHeader_t * h)
{

    memset(dh, 0, sizeof(VolumeDiskHeader_t));
    dh->stamp = h->stamp;
    dh->id = h->id;
    dh->parent = h->parent;

#ifdef AFS_64BIT_IOPS_ENV
    dh->volumeInfo_lo = (afs_int32) h->volumeInfo & 0xffffffff;
    dh->volumeInfo_hi = (afs_int32) (h->volumeInfo >> 32) & 0xffffffff;
    dh->smallVnodeIndex_lo = (afs_int32) h->smallVnodeIndex & 0xffffffff;
    dh->smallVnodeIndex_hi =
        (afs_int32) (h->smallVnodeIndex >> 32) & 0xffffffff;
    dh->largeVnodeIndex_lo = (afs_int32) h->largeVnodeIndex & 0xffffffff;
    dh->largeVnodeIndex_hi =
        (afs_int32) (h->largeVnodeIndex >> 32) & 0xffffffff;
    dh->linkTable_lo = (afs_int32) h->linkTable & 0xffffffff;
    dh->linkTable_hi = (afs_int32) (h->linkTable >> 32) & 0xffffffff;
#else
    dh->volumeInfo_lo = h->volumeInfo;
    dh->smallVnodeIndex_lo = h->smallVnodeIndex;
    dh->largeVnodeIndex_lo = h->largeVnodeIndex;
    dh->linkTable_lo = h->linkTable;
#endif
}

/* DiskToVolumeHeader
 * Converts an on-disk representation of a volume header to
 * the in-memory representation of a volume header.
 *
 * Makes the assumption that AFS has *always*
 * zero'd the volume header file so that high parts of inode
 * numbers are 0 in older (SGI EFS) volume header files.
 */
void
DiskToVolumeHeader(VolumeHeader_t * h, VolumeDiskHeader_t * dh)
{
    memset(h, 0, sizeof(VolumeHeader_t));
    h->stamp = dh->stamp;
    h->id = dh->id;
    h->parent = dh->parent;

#ifdef AFS_64BIT_IOPS_ENV
    h->volumeInfo =
        (Inode) dh->volumeInfo_lo | ((Inode) dh->volumeInfo_hi << 32);

    h->smallVnodeIndex =
        (Inode) dh->smallVnodeIndex_lo | ((Inode) dh->
                                          smallVnodeIndex_hi << 32);

    h->largeVnodeIndex =
        (Inode) dh->largeVnodeIndex_lo | ((Inode) dh->
                                          largeVnodeIndex_hi << 32);
    h->linkTable =
        (Inode) dh->linkTable_lo | ((Inode) dh->linkTable_hi << 32);
#else
    h->volumeInfo = dh->volumeInfo_lo;
    h->smallVnodeIndex = dh->smallVnodeIndex_lo;
    h->largeVnodeIndex = dh->largeVnodeIndex_lo;
    h->linkTable = dh->linkTable_lo;
#endif
}


/***************************************************/
/* Volume Attachment routines                      */
/***************************************************/

#ifdef AFS_DEMAND_ATTACH_FS
/**
 * pre-attach a volume given its path.
 *
 * @param[out] ec         outbound error code
 * @param[in]  partition  partition path string
 * @param[in]  name       volume id string
 *
 * @return volume object pointer
 *
 * @note A pre-attached volume will only have its partition
 *       and hashid fields initialized.  At first call to
 *       VGetVolume, the volume will be fully attached.
 *
 */
Volume *
VPreAttachVolumeByName(Error * ec, char *partition, char *name)
{
    Volume * vp;
    VOL_LOCK;
    vp = VPreAttachVolumeByName_r(ec, partition, name);
    VOL_UNLOCK;
    return vp;
}

/**
 * pre-attach a volume given its path.
 *
 * @param[out] ec         outbound error code
 * @param[in]  partition  path to vice partition
 * @param[in]  name       volume id string
 *
 * @return volume object pointer
 *
 * @pre VOL_LOCK held
 *
 * @internal volume package internal use only.
 */
Volume *
VPreAttachVolumeByName_r(Error * ec, char *partition, char *name)
{
    return VPreAttachVolumeById_r(ec,
                                  partition,
                                  VolumeNumber(name));
}

/**
 * pre-attach a volume given its path and numeric volume id.
 *
 * @param[out] ec          error code return
 * @param[in]  partition   path to vice partition
 * @param[in]  volumeId    numeric volume id
 *
 * @return volume object pointer
 *
 * @pre VOL_LOCK held
 *
 * @internal volume package internal use only.
 */
Volume *
VPreAttachVolumeById_r(Error * ec,
                       char * partition,
                       VolumeId volumeId)
{
    Volume *vp;
    struct DiskPartition64 *partp;

    *ec = 0;

    opr_Assert(programType == fileServer);

    if (!(partp = VGetPartition_r(partition, 0))) {
        *ec = VNOVOL;
        Log("VPreAttachVolumeById_r:  Error getting partition (%s)\n", partition);
        return NULL;
    }

    /* ensure that any vp we pass to VPreAttachVolumeByVp_r
     * is NOT in exclusive state.
     */
 retry:
    vp = VLookupVolume_r(ec, volumeId, NULL);

    if (*ec) {
        return NULL;
    }

    if (vp && VIsExclusiveState(V_attachState(vp))) {
        VCreateReservation_r(vp);
        VWaitExclusiveState_r(vp);
        VCancelReservation_r(vp);
        vp = NULL;
        goto retry;    /* look up volume again */
    }

    /* vp == NULL or vp not exclusive both OK */

    return VPreAttachVolumeByVp_r(ec, partp, vp, volumeId);
}

/**
 * preattach a volume.
 *
 * @param[out] ec     outbound error code
 * @param[in]  partp  pointer to partition object
 * @param[in]  vp     pointer to volume object
 * @param[in]  vid    volume id
 *
 * @return volume object pointer
 *
 * @pre VOL_LOCK is held.
 *
 * @pre vp (if specified) must not be in exclusive state.
 *
 * @warning Returned volume object pointer does not have to
 *          equal the pointer passed in as argument vp.  There
 *          are potential race conditions which can result in
 *          the pointers having different values.  It is up to
 *          the caller to make sure that references are handled
 *          properly in this case.
 *
 * @note If there is already a volume object registered with
 *       the same volume id, its pointer MUST be passed as
 *       argument vp.  Failure to do so will result in a silent
 *       failure to preattach.
 *
 * @internal volume package internal use only.
 */
Volume *
VPreAttachVolumeByVp_r(Error * ec,
                       struct DiskPartition64 * partp,
                       Volume * vp,
                       VolumeId vid)
{
    Volume *nvp = NULL;

    *ec = 0;

    /* don't proceed unless it's safe */
    if (vp) {
        opr_Assert(!VIsExclusiveState(V_attachState(vp)));
    }

    /* check to see if pre-attach already happened */
    if (vp &&
        (V_attachState(vp) != VOL_STATE_UNATTACHED) &&
        (V_attachState(vp) != VOL_STATE_DELETED) &&
        (V_attachState(vp) != VOL_STATE_PREATTACHED) &&
        !VIsErrorState(V_attachState(vp))) {
        /*
         * pre-attach is a no-op in all but the following cases:
         *
         *   - volume is unattached
         *   - volume is in an error state
         *   - volume is pre-attached
         */
        Log("VPreattachVolumeByVp_r: volume %" AFS_VOLID_FMT " not in quiescent state (state %u flags 0x%x)\n",
            afs_printable_VolumeId_lu(vid), V_attachState(vp),
            V_attachFlags(vp));
        goto done;
    } else if (vp) {
        /* we're re-attaching a volume; clear out some old state */
        memset(&vp->salvage, 0, sizeof(struct VolumeOnlineSalvage));

        if (V_partition(vp) != partp) {
            /* XXX potential race */
            DeleteVolumeFromVByPList_r(vp);
        }
    } else {
        /* if we need to allocate a new Volume struct,
         * go ahead and drop the vol glock, otherwise
         * do the basic setup synchronised, as it's
         * probably not worth dropping the lock */
        VOL_UNLOCK;

        /* allocate the volume structure */
        vp = nvp = calloc(1, sizeof(Volume));
        opr_Assert(vp != NULL);
        queue_Init(&vp->vnode_list);
        queue_Init(&vp->rx_call_list);
        opr_cv_init(&V_attachCV(vp));
    }

    /* link the volume with its associated vice partition */
    vp->device = partp->device;
    vp->partition = partp;

    vp->hashid = vid;
    vp->specialStatus = 0;

    /* if we dropped the lock, reacquire the lock,
     * check for pre-attach races, and then add
     * the volume to the hash table */
    if (nvp) {
        VOL_LOCK;
        nvp = VLookupVolume_r(ec, vid, NULL);
        if (*ec) {
            free(vp);
            vp = NULL;
            goto done;
        } else if (nvp) { /* race detected */
            free(vp);
            vp = nvp;
            goto done;
        } else {
          /* hack to make up for VChangeState_r() decrementing
           * the old state counter */
          VStats.state_levels[0]++;
        }
    }

    /* put pre-attached volume onto the hash table
     * and bring it up to the pre-attached state */
    AddVolumeToHashTable(vp, vp->hashid);
    AddVolumeToVByPList_r(vp);
    VLRU_Init_Node_r(vp);
    VChangeState_r(vp, VOL_STATE_PREATTACHED);

    if (GetLogLevel() >= 5)
        Log("VPreAttachVolumeByVp_r:  volume %" AFS_VOLID_FMT " pre-attached\n", afs_printable_VolumeId_lu(vp->hashid));

  done:
    if (*ec)
        return NULL;
    else
        return vp;
}
#endif /* AFS_DEMAND_ATTACH_FS */

/* Attach an existing volume, given its pathname, and return a
   pointer to the volume header information.  The volume also
   normally goes online at this time.  An offline volume
   must be reattached to make it go online */
Volume *
VAttachVolumeByName(Error * ec, char *partition, char *name, int mode)
{
    Volume *retVal;
    VOL_LOCK;
    retVal = VAttachVolumeByName_r(ec, partition, name, mode);
    VOL_UNLOCK;
    return retVal;
}

Volume *
VAttachVolumeByName_r(Error * ec, char *partition, char *name, int mode)
{
    Volume *vp = NULL;
    struct DiskPartition64 *partp;
    char path[64];
    int isbusy = 0;
    VolumeId volumeId;
    int checkedOut;
#ifdef AFS_DEMAND_ATTACH_FS
    VolumeStats stats_save;
    Volume *svp = NULL;
#endif /* AFS_DEMAND_ATTACH_FS */

    *ec = 0;

    volumeId = VolumeNumber(name);

    if (!(partp = VGetPartition_r(partition, 0))) {
        *ec = VNOVOL;
        Log("VAttachVolume: Error getting partition (%s)\n", partition);
        goto done;
    }

    if (VRequiresPartLock()) {
        opr_Assert(VInit == 3);
        VLockPartition_r(partition);
    } else if (programType == fileServer) {
#ifdef AFS_DEMAND_ATTACH_FS
        /* lookup the volume in the hash table */
        vp = VLookupVolume_r(ec, volumeId, NULL);
        if (*ec) {
            return NULL;
        }

        if (vp) {
            /* save any counters that are supposed to
             * be monotonically increasing over the
             * lifetime of the fileserver */
            memcpy(&stats_save, &vp->stats, sizeof(VolumeStats));
        } else {
            memset(&stats_save, 0, sizeof(VolumeStats));
        }

        /* if there's something in the hash table, and it's not
         * in the pre-attach state, then we may need to detach
         * it before proceeding */
        if (vp && (V_attachState(vp) != VOL_STATE_PREATTACHED)) {
            VCreateReservation_r(vp);
            VWaitExclusiveState_r(vp);

            /* at this point state must be one of:
             *   - UNATTACHED
             *   - ATTACHED
             *   - SHUTTING_DOWN
             *   - GOING_OFFLINE
             *   - SALVAGING
             *   - ERROR
             *   - DELETED
             */

            if (vp->specialStatus == VBUSY)
                isbusy = 1;

            /* if it's already attached, see if we can return it */
            if (V_attachState(vp) == VOL_STATE_ATTACHED) {
                VGetVolumeByVp_r(ec, vp);
                if (V_inUse(vp) == fileServer) {
                    VCancelReservation_r(vp);
                    return vp;
                }

                /* otherwise, we need to detach, and attempt to re-attach */
                VDetachVolume_r(ec, vp);
                if (*ec) {
                    Log("VAttachVolume: Error detaching old volume instance (%s)\n", name);
                }
            } else {
                /* if it isn't fully attached, delete from the hash tables,
                   and let the refcounter handle the rest */
                DeleteVolumeFromHashTable(vp);
                DeleteVolumeFromVByPList_r(vp);
            }

            VCancelReservation_r(vp);
            vp = NULL;
        }

        /* pre-attach volume if it hasn't been done yet */
        if (!vp ||
            (V_attachState(vp) == VOL_STATE_UNATTACHED) ||
            (V_attachState(vp) == VOL_STATE_DELETED) ||
            (V_attachState(vp) == VOL_STATE_ERROR)) {
            svp = vp;
            vp = VPreAttachVolumeByVp_r(ec, partp, vp, volumeId);
            if (*ec) {
                return NULL;
            }
        }

        opr_Assert(vp != NULL);

        /* handle pre-attach races
         *
         * multiple threads can race to pre-attach a volume,
         * but we can't let them race beyond that
         *
         * our solution is to let the first thread to bring
         * the volume into an exclusive state win; the other
         * threads just wait until it finishes bringing the
         * volume online, and then they do a vgetvolumebyvp
         */
        if (svp && (svp != vp)) {
            /* wait for other exclusive ops to finish */
            VCreateReservation_r(vp);
            VWaitExclusiveState_r(vp);

            /* get a heavyweight ref, kill the lightweight ref, and return */
            VGetVolumeByVp_r(ec, vp);
            VCancelReservation_r(vp);
            return vp;
        }

        /* at this point, we are chosen as the thread to do
         * demand attachment for this volume. all other threads
         * doing a getvolume on vp->hashid will block until we finish */

        /* make sure any old header cache entries are invalidated
         * before proceeding */
        FreeVolumeHeader(vp);

        VChangeState_r(vp, VOL_STATE_ATTACHING);

        /* restore any saved counters */
        memcpy(&vp->stats, &stats_save, sizeof(VolumeStats));
#else /* AFS_DEMAND_ATTACH_FS */
        vp = VGetVolume_r(ec, volumeId);
        if (vp) {
            if (V_inUse(vp) == fileServer)
                return vp;
            if (vp->specialStatus == VBUSY)
                isbusy = 1;
            VDetachVolume_r(ec, vp);
            if (*ec) {
                Log("VAttachVolume: Error detaching volume (%s)\n", name);
            }
            vp = NULL;
        }
#endif /* AFS_DEMAND_ATTACH_FS */
    }

    *ec = 0;
    strcpy(path, VPartitionPath(partp));

    VOL_UNLOCK;

    strcat(path, OS_DIRSEP);
    strcat(path, name);

    if (!vp) {
      vp = (Volume *) calloc(1, sizeof(Volume));
      opr_Assert(vp != NULL);
      vp->hashid = volumeId;
      vp->device = partp->device;
      vp->partition = partp;
      queue_Init(&vp->vnode_list);
      queue_Init(&vp->rx_call_list);
#ifdef AFS_DEMAND_ATTACH_FS
      opr_cv_init(&V_attachCV(vp));
#endif /* AFS_DEMAND_ATTACH_FS */
    }

    /* attach2 is entered without any locks, and returns
     * with vol_glock_mutex held */
    vp = attach2(ec, volumeId, path, partp, vp, isbusy, mode, &checkedOut);

    if (VCanUseFSSYNC() && vp) {
#ifdef AFS_DEMAND_ATTACH_FS
        if ((mode == V_VOLUPD) || (VolumeWriteable(vp) && (mode == V_CLONE))) {
            /* mark volume header as in use so that volser crashes lead to a
             * salvage attempt */
            VUpdateVolume_r(ec, vp, 0);
        }
        /* for dafs, we should tell the fileserver, except for V_PEEK
         * where we know it is not necessary */
        if (mode == V_PEEK) {
            vp->needsPutBack = 0;
        } else {
            vp->needsPutBack = VOL_PUTBACK;
        }
#else /* !AFS_DEMAND_ATTACH_FS */
        /* duplicate computation in fssync.c about whether the server
         * takes the volume offline or not.  If the volume isn't
         * offline, we must not return it when we detach the volume,
         * or the server will abort */
        if (mode == V_READONLY || mode == V_PEEK
            || (!VolumeWriteable(vp) && (mode == V_CLONE || mode == V_DUMP)))
            vp->needsPutBack = 0;
        else
            vp->needsPutBack = VOL_PUTBACK;
#endif /* !AFS_DEMAND_ATTACH_FS */
    }
#ifdef FSSYNC_BUILD_CLIENT
    /* Only give back the vol to the fileserver if we checked it out; attach2
     * will set checkedOut only if we successfully checked it out from the
     * fileserver. */
    if (VCanUseFSSYNC() && vp == NULL && checkedOut) {

#ifdef AFS_DEMAND_ATTACH_FS
        /* If we couldn't attach but we scheduled a salvage, we already
         * notified the fileserver; don't online it now */
        if (*ec != VSALVAGING)
#endif /* AFS_DEMAND_ATTACH_FS */
        FSYNC_VolOp(volumeId, partition, FSYNC_VOL_ON, 0, NULL);
    } else
#endif
    if (programType == fileServer && vp) {
#ifdef AFS_DEMAND_ATTACH_FS
        /*
         * we can get here in cases where we don't "own"
         * the volume (e.g. volume owned by a utility).
         * short circuit around potential disk header races.
         */
        if (V_attachState(vp) != VOL_STATE_ATTACHED) {
            goto done;
        }
#endif
        VUpdateVolume_r(ec, vp, 0);
        if (*ec) {
            Log("VAttachVolume: Error updating volume\n");
            if (vp)
                VPutVolume_r(vp);
            goto done;
        }
        if (VolumeWriteable(vp) && V_dontSalvage(vp) == 0) {
#ifndef AFS_DEMAND_ATTACH_FS
            /* This is a hack: by temporarily setting the incore
             * dontSalvage flag ON, the volume will be put back on the
             * Update list (with dontSalvage OFF again).  It will then
             * come back in N minutes with DONT_SALVAGE eventually
             * set.  This is the way that volumes that have never had
             * it set get it set; or that volumes that have been
             * offline without DONT SALVAGE having been set also
             * eventually get it set */
            V_dontSalvage(vp) = DONT_SALVAGE;
#endif /* !AFS_DEMAND_ATTACH_FS */
            VAddToVolumeUpdateList_r(ec, vp);
            if (*ec) {
                Log("VAttachVolume: Error adding volume to update list\n");
                if (vp)
                    VPutVolume_r(vp);
                goto done;
            }
        }
        if (GetLogLevel() != 0)
          Log("VOnline:  volume %" AFS_VOLID_FMT " (%s) attached and online\n", afs_printable_VolumeId_lu(V_id(vp)),
                V_name(vp));
    }

  done:
    if (VRequiresPartLock()) {
        VUnlockPartition_r(partition);
    }
    if (*ec) {
#ifdef AFS_DEMAND_ATTACH_FS
        /* attach failed; make sure we're in error state */
        if (vp && !VIsErrorState(V_attachState(vp))) {
            VChangeState_r(vp, VOL_STATE_ERROR);
        }
#endif /* AFS_DEMAND_ATTACH_FS */
        return NULL;
    } else {
        return vp;
    }
}

#ifdef AFS_DEMAND_ATTACH_FS
/* VAttachVolumeByVp_r
 *
 * finish attaching a volume that is
 * in a less than fully attached state
 */
/* caller MUST hold a ref count on vp */
static Volume *
VAttachVolumeByVp_r(Error * ec, Volume * vp, int mode)
{
    char name[VMAXPATHLEN];
    int reserve = 0;
    struct DiskPartition64 *partp;
    char path[64];
    int isbusy = 0;
    VolumeId volumeId;
    Volume * nvp = NULL;
    VolumeStats stats_save;
    int checkedOut;
    *ec = 0;

    /* volume utility should never call AttachByVp */
    opr_Assert(programType == fileServer);

    volumeId = vp->hashid;
    partp = vp->partition;
    VolumeExternalName_r(volumeId, name, sizeof(name));


    /* if another thread is performing a blocking op, wait */
    VWaitExclusiveState_r(vp);

    memcpy(&stats_save, &vp->stats, sizeof(VolumeStats));

    /* if it's already attached, see if we can return it */
    if (V_attachState(vp) == VOL_STATE_ATTACHED) {
        VGetVolumeByVp_r(ec, vp);
        if (V_inUse(vp) == fileServer) {
            return vp;
        } else {
            if (vp->specialStatus == VBUSY)
                isbusy = 1;
            VDetachVolume_r(ec, vp);
            if (*ec) {
                Log("VAttachVolume: Error detaching volume (%s)\n", name);
            }
            vp = NULL;
        }
    }

    /* pre-attach volume if it hasn't been done yet */
    if (!vp ||
        (V_attachState(vp) == VOL_STATE_UNATTACHED) ||
        (V_attachState(vp) == VOL_STATE_DELETED) ||
        (V_attachState(vp) == VOL_STATE_ERROR)) {
        nvp = VPreAttachVolumeByVp_r(ec, partp, vp, volumeId);
        if (*ec) {
            return NULL;
        }
        if (nvp != vp) {
            reserve = 1;
            VCreateReservation_r(nvp);
            vp = nvp;
        }
    }

    opr_Assert(vp != NULL);
    VChangeState_r(vp, VOL_STATE_ATTACHING);

    /* restore monotonically increasing stats */
    memcpy(&vp->stats, &stats_save, sizeof(VolumeStats));

    *ec = 0;

    /* compute path to disk header */
    strcpy(path, VPartitionPath(partp));

    VOL_UNLOCK;

    strcat(path, OS_DIRSEP);
    strcat(path, name);

    /* do volume attach
     *
     * NOTE: attach2 is entered without any locks, and returns
     * with vol_glock_mutex held */
    vp = attach2(ec, volumeId, path, partp, vp, isbusy, mode, &checkedOut);

    /*
     * the event that an error was encountered, or
     * the volume was not brought to an attached state
     * for any reason, skip to the end.  We cannot
     * safely call VUpdateVolume unless we "own" it.
     */
    if (*ec ||
        (vp == NULL) ||
        (V_attachState(vp) != VOL_STATE_ATTACHED)) {
        goto done;
    }

    VUpdateVolume_r(ec, vp, 0);
    if (*ec) {
        Log("VAttachVolume: Error updating volume %" AFS_VOLID_FMT "\n",
            afs_printable_VolumeId_lu(vp->hashid));
        VPutVolume_r(vp);
        goto done;
    }
    if (VolumeWriteable(vp) && V_dontSalvage(vp) == 0) {
#ifndef AFS_DEMAND_ATTACH_FS
        /* This is a hack: by temporarily setting the incore
         * dontSalvage flag ON, the volume will be put back on the
         * Update list (with dontSalvage OFF again).  It will then
         * come back in N minutes with DONT_SALVAGE eventually
         * set.  This is the way that volumes that have never had
         * it set get it set; or that volumes that have been
         * offline without DONT SALVAGE having been set also
         * eventually get it set */
        V_dontSalvage(vp) = DONT_SALVAGE;
#endif /* !AFS_DEMAND_ATTACH_FS */
        VAddToVolumeUpdateList_r(ec, vp);
        if (*ec) {
            Log("VAttachVolume: Error adding volume %" AFS_VOLID_FMT " to update list\n",
                afs_printable_VolumeId_lu(vp->hashid));
            if (vp)
                VPutVolume_r(vp);
            goto done;
        }
    }
    if (GetLogLevel() != 0)
        Log("VOnline:  volume %" AFS_VOLID_FMT " (%s) attached and online\n",
            afs_printable_VolumeId_lu(V_id(vp)), V_name(vp));
  done:
    if (reserve) {
        VCancelReservation_r(nvp);
        reserve = 0;
    }
    if (*ec && (*ec != VOFFLINE) && (*ec != VSALVAGE)) {
        if (vp && !VIsErrorState(V_attachState(vp))) {
            VChangeState_r(vp, VOL_STATE_ERROR);
        }
        return NULL;
    } else {
        return vp;
    }
}

/**
 * lock a volume on disk (non-blocking).
 *
 * @param[in] vp  The volume to lock
 * @param[in] locktype READ_LOCK or WRITE_LOCK
 *
 * @return operation status
 *  @retval 0 success, lock was obtained
 *  @retval EBUSY a conflicting lock was held by another process
 *  @retval EIO   error acquiring lock
 *
 * @pre If we're in the fileserver, vp is in an exclusive state
 *
 * @pre vp is not already locked
 */
static int
VLockVolumeNB(Volume *vp, int locktype)
{
    int code;

    opr_Assert(programType != fileServer
               || VIsExclusiveState(V_attachState(vp)));
    opr_Assert(!(V_attachFlags(vp) & VOL_LOCKED));

    code = VLockVolumeByIdNB(vp->hashid, vp->partition, locktype);
    if (code == 0) {
        V_attachFlags(vp) |= VOL_LOCKED;
    }

    return code;
}

/**
 * unlock a volume on disk that was locked with VLockVolumeNB.
 *
 * @param[in] vp  volume to unlock
 *
 * @pre If we're in the fileserver, vp is in an exclusive state
 *
 * @pre vp has already been locked
 */
static void
VUnlockVolume(Volume *vp)
{
    opr_Assert(programType != fileServer
               || VIsExclusiveState(V_attachState(vp)));
    opr_Assert((V_attachFlags(vp) & VOL_LOCKED));

    VUnlockVolumeById(vp->hashid, vp->partition);

    V_attachFlags(vp) &= ~VOL_LOCKED;
}
#endif /* AFS_DEMAND_ATTACH_FS */

/**
 * read in a vol header, possibly lock the vol header, and possibly check out
 * the vol header from the fileserver, as part of volume attachment.
 *
 * @param[out] ec     error code
 * @param[in] vp      volume pointer object
 * @param[in] partp   disk partition object of the attaching partition
 * @param[in] mode    attachment mode such as V_VOLUPD, V_DUMP, etc (see
 *                    volume.h)
 * @param[in] peek    1 to just try to read in the volume header and make sure
 *                    we don't try to lock the vol, or check it out from
 *                    FSSYNC or anything like that; 0 otherwise, for 'normal'
 *                    operation
 * @param[out] acheckedOut   If we successfully checked-out the volume from
 *                           the fileserver (if we needed to), this is set
 *                           to 1, otherwise it is untouched.
 *
 * @note As part of DAFS volume attachment, the volume header may be either
 *       read- or write-locked to ensure mutual exclusion of certain volume
 *       operations. In some cases in order to determine whether we need to
 *       read- or write-lock the header, we need to read in the header to see
 *       if the volume is RW or not. So, if we read in the header under a
 *       read-lock and determine that we actually need a write-lock on the
 *       volume header, this function will drop the read lock, acquire a write
 *       lock, and read the header in again.
 */
static void
attach_volume_header(Error *ec, Volume *vp, struct DiskPartition64 *partp,
                     int mode, int peek, int *acheckedOut)
{
    struct VolumeDiskHeader diskHeader;
    struct VolumeHeader header;
    int code;
    int first_try = 1;
    int lock_tries = 0, checkout_tries = 0;
    int retry;
    VolumeId volid = vp->hashid;
#ifdef FSSYNC_BUILD_CLIENT
    int checkout, done_checkout = 0;
#endif /* FSSYNC_BUILD_CLIENT */
#ifdef AFS_DEMAND_ATTACH_FS
    int locktype = 0, use_locktype = -1;
#endif /* AFS_DEMAND_ATTACH_FS */

 retry:
    retry = 0;
    *ec = 0;

    if (lock_tries > VOL_MAX_CHECKOUT_RETRIES) {
        Log("VAttachVolume: retried too many times trying to lock header for "
            "vol %lu part %s; giving up\n", afs_printable_uint32_lu(volid),
            VPartitionPath(partp));
        *ec = VNOVOL;
        goto done;
    }
    if (checkout_tries > VOL_MAX_CHECKOUT_RETRIES) {
        Log("VAttachVolume: retried too many times trying to checkout "
            "vol %lu part %s; giving up\n", afs_printable_uint32_lu(volid),
            VPartitionPath(partp));
        *ec = VNOVOL;
        goto done;
    }

    if (VReadVolumeDiskHeader(volid, partp, NULL)) {
        /* short-circuit the 'volume does not exist' case */
        *ec = VNOVOL;
        goto done;
    }

#ifdef FSSYNC_BUILD_CLIENT
    checkout = !done_checkout;
    done_checkout = 1;
    if (!peek && checkout && VMustCheckoutVolume(mode)) {
        SYNC_response res;
        memset(&res, 0, sizeof(res));

        if (FSYNC_VolOp(volid, partp->name, FSYNC_VOL_NEEDVOLUME, mode, &res)
            != SYNC_OK) {

            if (res.hdr.reason == FSYNC_SALVAGE) {
                Log("VAttachVolume: file server says volume %lu is salvaging\n",
                     afs_printable_uint32_lu(volid));
                *ec = VSALVAGING;
            } else {
                Log("VAttachVolume: attach of volume %lu apparently denied by file server\n",
                     afs_printable_uint32_lu(volid));
                *ec = VNOVOL;   /* XXXX */
            }
            goto done;
        }
        *acheckedOut = 1;
    }
#endif

#ifdef AFS_DEMAND_ATTACH_FS
    if (use_locktype < 0) {
        /* don't know whether vol is RO or RW; assume it's RO and we can retry
         * if it turns out to be RW */
        locktype = VVolLockType(mode, 0);

    } else {
        /* a previous try says we should use use_locktype to lock the volume,
         * so use that */
        locktype = use_locktype;
    }

    if (!peek && locktype) {
        code = VLockVolumeNB(vp, locktype);
        if (code) {
            if (code == EBUSY) {
                Log("VAttachVolume: another program has vol %lu locked\n",
                    afs_printable_uint32_lu(volid));
            } else {
                Log("VAttachVolume: error %d trying to lock vol %lu\n",
                    code, afs_printable_uint32_lu(volid));
            }

            *ec = VNOVOL;
            goto done;
        }
    }
#endif /* AFS_DEMAND_ATTACH_FS */

    code = VReadVolumeDiskHeader(volid, partp, &diskHeader);
    if (code) {
        if (code == EIO) {
            *ec = VSALVAGE;
        } else {
            *ec = VNOVOL;
        }
        goto done;
    }

    DiskToVolumeHeader(&header, &diskHeader);

    IH_INIT(vp->vnodeIndex[vLarge].handle, partp->device, header.parent,
            header.largeVnodeIndex);
    IH_INIT(vp->vnodeIndex[vSmall].handle, partp->device, header.parent,
            header.smallVnodeIndex);
    IH_INIT(vp->diskDataHandle, partp->device, header.parent,
            header.volumeInfo);
    IH_INIT(vp->linkHandle, partp->device, header.parent, header.linkTable);

    if (first_try) {
        /* only need to do this once */
        VOL_LOCK;
        GetVolumeHeader(vp);
        VOL_UNLOCK;
    }

#if defined(AFS_DEMAND_ATTACH_FS) && defined(FSSYNC_BUILD_CLIENT)
    /* demand attach changes the V_PEEK mechanism
     *
     * we can now suck the current disk data structure over
     * the fssync interface without going to disk
     *
     * (technically, we don't need to restrict this feature
     *  to demand attach fileservers.  However, I'm trying
     *  to limit the number of common code changes)
     */
    if (VCanUseFSSYNC() && (mode == V_PEEK || peek)) {
        SYNC_response res;
        res.payload.len = sizeof(VolumeDiskData);
        res.payload.buf = &(V_disk(vp));

        if (FSYNC_VolOp(vp->hashid,
                        partp->name,
                        FSYNC_VOL_QUERY_HDR,
                        FSYNC_WHATEVER,
                        &res) == SYNC_OK) {
            goto disk_header_loaded;
        }
    }
#endif /* AFS_DEMAND_ATTACH_FS && FSSYNC_BUILD_CLIENT */
    (void)ReadHeader(ec, V_diskDataHandle(vp), (char *)&V_disk(vp),
                     sizeof(V_disk(vp)), VOLUMEINFOMAGIC, VOLUMEINFOVERSION);

#ifdef AFS_DEMAND_ATTACH_FS
    /* update stats */
    VOL_LOCK;
    IncUInt64(&VStats.hdr_loads);
    IncUInt64(&vp->stats.hdr_loads);
    VOL_UNLOCK;
#endif /* AFS_DEMAND_ATTACH_FS */

    if (*ec) {
        Log("VAttachVolume: Error reading diskDataHandle header for vol %lu; "
            "error=%u\n", afs_printable_uint32_lu(volid), *ec);
        goto done;
    }

#ifdef AFS_DEMAND_ATTACH_FS
# ifdef FSSYNC_BUILD_CLIENT
 disk_header_loaded:
# endif /* FSSYNC_BUILD_CLIENT */

    /* if the lock type we actually used to lock the volume is different than
     * the lock type we should have used, retry with the lock type we should
     * use */
    use_locktype = VVolLockType(mode, VolumeWriteable(vp));
    if (locktype != use_locktype) {
        retry = 1;
        lock_tries++;
    }
#endif /* AFS_DEMAND_ATTACH_FS */

    *ec = 0;

 done:
#if defined(AFS_DEMAND_ATTACH_FS) && defined(FSSYNC_BUILD_CLIENT)
    if (!peek && *ec == 0 && retry == 0 && VMustCheckoutVolume(mode)) {

        code = FSYNC_VerifyCheckout(volid, partp->name, FSYNC_VOL_NEEDVOLUME, mode);

        if (code == SYNC_DENIED) {
            /* must retry checkout; fileserver no longer thinks we have
             * the volume */
            retry = 1;
            checkout_tries++;
            done_checkout = 0;

        } else if (code != SYNC_OK) {
            *ec = VNOVOL;
        }
    }
#endif /* AFS_DEMAND_ATTACH_FS && FSSYNC_BUILD_CLIENT */

    if (*ec || retry) {
        /* either we are going to be called again for a second pass, or we
         * encountered an error; clean up in either case */

#ifdef AFS_DEMAND_ATTACH_FS
        if ((V_attachFlags(vp) & VOL_LOCKED)) {
            VUnlockVolume(vp);
        }
#endif /* AFS_DEMAND_ATTACH_FS */
        if (vp->linkHandle) {
            IH_RELEASE(vp->vnodeIndex[vLarge].handle);
            IH_RELEASE(vp->vnodeIndex[vSmall].handle);
            IH_RELEASE(vp->diskDataHandle);
            IH_RELEASE(vp->linkHandle);
        }
    }

    if (*ec) {
        VOL_LOCK;
        FreeVolumeHeader(vp);
        VOL_UNLOCK;
        return;
    }
    if (retry) {
        first_try = 0;
        goto retry;
    }
}

#ifdef AFS_DEMAND_ATTACH_FS
static void
attach_check_vop(Error *ec, VolumeId volid, struct DiskPartition64 *partp,
                 Volume *vp, int *acheckedOut)
{
    *ec = 0;

    if (vp->pending_vol_op) {

        VOL_LOCK;

        if (vp->pending_vol_op->vol_op_state == FSSYNC_VolOpRunningUnknown) {
            int code;
            code = VVolOpLeaveOnlineNoHeader_r(vp, vp->pending_vol_op);
            if (code == 1) {
                vp->pending_vol_op->vol_op_state = FSSYNC_VolOpRunningOnline;
            } else if (code == 0) {
                vp->pending_vol_op->vol_op_state = FSSYNC_VolOpRunningOffline;

            } else {
                /* we need the vol header to determine if the volume can be
                 * left online for the vop, so... get the header */

                VOL_UNLOCK;

                /* attach header with peek=1 to avoid checking out the volume
                 * or locking it; we just want the header info, we're not
                 * messing with the volume itself at all */
                attach_volume_header(ec, vp, partp, V_PEEK, 1, acheckedOut);
                if (*ec) {
                    return;
                }

                VOL_LOCK;

                if (VVolOpLeaveOnline_r(vp, vp->pending_vol_op)) {
                    vp->pending_vol_op->vol_op_state = FSSYNC_VolOpRunningOnline;
                } else {
                    vp->pending_vol_op->vol_op_state = FSSYNC_VolOpRunningOffline;
                }

                /* make sure we grab a new vol header and re-open stuff on
                 * actual attachment; we can't keep the data we grabbed, since
                 * it was not done under a lock and thus not safe */
                FreeVolumeHeader(vp);
                VReleaseVolumeHandles_r(vp);
            }
        }
        /* see if the pending volume op requires exclusive access */
        switch (vp->pending_vol_op->vol_op_state) {
        case FSSYNC_VolOpPending:
            /* this should never happen */
            opr_Assert(vp->pending_vol_op->vol_op_state
                            != FSSYNC_VolOpPending);
            break;

        case FSSYNC_VolOpRunningUnknown:
            /* this should never happen; we resolved 'unknown' above */
            opr_Assert(vp->pending_vol_op->vol_op_state
                            != FSSYNC_VolOpRunningUnknown);
            break;

        case FSSYNC_VolOpRunningOffline:
            /* mark the volume down */
            *ec = VOFFLINE;
            VChangeState_r(vp, VOL_STATE_UNATTACHED);

            /* do not set V_offlineMessage here; we don't have ownership of
             * the volume (and probably do not have the header loaded), so we
             * can't alter the disk header */

            /* check to see if we should set the specialStatus flag */
            if (VVolOpSetVBusy_r(vp, vp->pending_vol_op)) {
                /* don't overwrite specialStatus if it was already set to
                 * something else (e.g. VMOVED) */
                if (!vp->specialStatus) {
                    vp->specialStatus = VBUSY;
                }
            }
            break;

        default:
            break;
        }

        VOL_UNLOCK;
    }
}
#endif /* AFS_DEMAND_ATTACH_FS */

/**
 * volume attachment helper function.
 *
 * @param[out] ec      error code
 * @param[in] volumeId volume ID of the attaching volume
 * @param[in] path     full path to the volume header .vol file
 * @param[in] partp    disk partition object for the attaching partition
 * @param[in] vp       volume object; vp->hashid, vp->device, vp->partition,
 *                     vp->vnode_list, vp->rx_call_list, and V_attachCV (for
 *                     DAFS) should already be initialized
 * @param[in] isbusy   1 if vp->specialStatus should be set to VBUSY; that is,
 *                     if there is a volume operation running for this volume
 *                     that should set the volume to VBUSY during its run. 0
 *                     otherwise. (see VVolOpSetVBusy_r)
 * @param[in] mode     attachment mode such as V_VOLUPD, V_DUMP, etc (see
 *                     volume.h)
 * @param[out] acheckedOut   If we successfully checked-out the volume from
 *                           the fileserver (if we needed to), this is set
 *                           to 1, otherwise it is 0.
 *
 * @return pointer to the semi-attached volume pointer
 *  @retval NULL an error occurred (check value of *ec)
 *  @retval vp volume successfully attaching
 *
 * @pre no locks held
 *
 * @post VOL_LOCK held
 */
static Volume *
attach2(Error * ec, VolumeId volumeId, char *path, struct DiskPartition64 *partp,
        Volume * vp, int isbusy, int mode, int *acheckedOut)
{
    /* have we read in the header successfully? */
    int read_header = 0;

#ifdef AFS_DEMAND_ATTACH_FS
    /* should we FreeVolume(vp) instead of VCheckFree(vp) in the error
     * cleanup? */
    int forcefree = 0;

    /* in the case of an error, to what state should the volume be
     * transitioned? */
    VolState error_state = VOL_STATE_ERROR;
#endif /* AFS_DEMAND_ATTACH_FS */

    *ec = 0;

    vp->vnodeIndex[vLarge].handle = NULL;
    vp->vnodeIndex[vSmall].handle = NULL;
    vp->diskDataHandle = NULL;
    vp->linkHandle = NULL;

    *acheckedOut = 0;

#ifdef AFS_DEMAND_ATTACH_FS
    attach_check_vop(ec, volumeId, partp, vp, acheckedOut);
    if (!*ec) {
        attach_volume_header(ec, vp, partp, mode, 0, acheckedOut);
    }
#else
    attach_volume_header(ec, vp, partp, mode, 0, acheckedOut);
#endif /* !AFS_DEMAND_ATTACH_FS */

    if (*ec == VNOVOL) {
        /* if the volume doesn't exist, skip straight to 'error' so we don't
         * request a salvage */
        goto unlocked_error;
    }

    if (!*ec) {
        read_header = 1;

        /* ensure that we don't override specialStatus if it was set to
         * something else (e.g. VMOVED) */
        if (isbusy && !vp->specialStatus) {
            vp->specialStatus = VBUSY;
        }
        vp->shuttingDown = 0;
        vp->goingOffline = 0;
        vp->nUsers = 1;
#ifdef AFS_DEMAND_ATTACH_FS
        vp->stats.last_attach = FT_ApproxTime();
        vp->stats.attaches++;
#endif

        VOL_LOCK;
        IncUInt64(&VStats.attaches);
        vp->cacheCheck = ++VolumeCacheCheck;
        /* just in case this ever rolls over */
        if (!vp->cacheCheck)
            vp->cacheCheck = ++VolumeCacheCheck;
        VOL_UNLOCK;

#ifdef AFS_DEMAND_ATTACH_FS
        V_attachFlags(vp) |= VOL_HDR_LOADED;
        vp->stats.last_hdr_load = vp->stats.last_attach;
#endif /* AFS_DEMAND_ATTACH_FS */
    }

    if (!*ec) {
        struct IndexFileHeader iHead;

        /*
         * We just read in the diskstuff part of the header.  If the detailed
         * volume stats area has not yet been initialized, we should bzero the
         * area and mark it as initialized.
         */
        if (!(V_stat_initialized(vp))) {
            memset((V_stat_area(vp)), 0, VOL_STATS_BYTES);
            V_stat_initialized(vp) = 1;
        }

        (void)ReadHeader(ec, vp->vnodeIndex[vSmall].handle,
                         (char *)&iHead, sizeof(iHead),
                         SMALLINDEXMAGIC, SMALLINDEXVERSION);

        if (*ec) {
            Log("VAttachVolume: Error reading smallVnode vol header %s; error=%u\n", path, *ec);
        }
    }

    if (!*ec) {
        struct IndexFileHeader iHead;

        (void)ReadHeader(ec, vp->vnodeIndex[vLarge].handle,
                         (char *)&iHead, sizeof(iHead),
                         LARGEINDEXMAGIC, LARGEINDEXVERSION);

        if (*ec) {
            Log("VAttachVolume: Error reading largeVnode vol header %s; error=%u\n", path, *ec);
        }
    }

#ifdef AFS_NAMEI_ENV
    if (!*ec) {
        struct versionStamp stamp;

        (void)ReadHeader(ec, V_linkHandle(vp), (char *)&stamp,
                         sizeof(stamp), LINKTABLEMAGIC, LINKTABLEVERSION);

        if (*ec) {
            Log("VAttachVolume: Error reading namei vol header %s; error=%u\n", path, *ec);
        }
    }
#endif /* AFS_NAMEI_ENV */

#if defined(AFS_DEMAND_ATTACH_FS)
    if (*ec && ((*ec != VOFFLINE) || (V_attachState(vp) != VOL_STATE_UNATTACHED))) {
        VOL_LOCK;
        if (!VCanScheduleSalvage()) {
            Log("VAttachVolume: Error attaching volume %s; volume needs salvage; error=%u\n", path, *ec);
        }
        VRequestSalvage_r(ec, vp, SALVSYNC_ERROR, VOL_SALVAGE_NO_OFFLINE);
        vp->nUsers = 0;

        goto locked_error;
    } else if (*ec) {
        /* volume operation in progress */
        VOL_LOCK;
        /* we have already transitioned the vp away from ATTACHING state, so we
         * can go right to the end of attach2, and we do not need to transition
         * to ERROR. */
        goto error_notbroken;
    }
#else /* AFS_DEMAND_ATTACH_FS */
    if (*ec) {
        Log("VAttachVolume: Error attaching volume %s; volume needs salvage; error=%u\n", path, *ec);
        goto unlocked_error;
    }
#endif /* AFS_DEMAND_ATTACH_FS */

    if (V_needsSalvaged(vp)) {
        if (vp->specialStatus)
            vp->specialStatus = 0;
        VOL_LOCK;
#if defined(AFS_DEMAND_ATTACH_FS)
        if (!VCanScheduleSalvage()) {
            Log("VAttachVolume: volume salvage flag is ON for %s; volume needs salvage\n", path);
        }
        VRequestSalvage_r(ec, vp, SALVSYNC_NEEDED, VOL_SALVAGE_NO_OFFLINE);
        vp->nUsers = 0;

#else /* AFS_DEMAND_ATTACH_FS */
        *ec = VSALVAGE;
#endif /* AFS_DEMAND_ATTACH_FS */

        goto locked_error;
    }

    VOL_LOCK;
    vp->nextVnodeUnique = V_uniquifier(vp);

    if (VShouldCheckInUse(mode) && V_inUse(vp) && VolumeWriteable(vp)) {
        if (!V_needsSalvaged(vp)) {
            V_needsSalvaged(vp) = 1;
            VUpdateVolume_r(ec, vp, 0);
        }
#if defined(AFS_DEMAND_ATTACH_FS)
        if (!VCanScheduleSalvage()) {
            Log("VAttachVolume: volume %s needs to be salvaged; not attached.\n", path);
        }
        VRequestSalvage_r(ec, vp, SALVSYNC_NEEDED, VOL_SALVAGE_NO_OFFLINE);
        vp->nUsers = 0;

#else /* AFS_DEMAND_ATTACH_FS */
        Log("VAttachVolume: volume %s needs to be salvaged; not attached.\n", path);
        *ec = VSALVAGE;
#endif /* AFS_DEMAND_ATTACH_FS */

        goto locked_error;
    }

    if (programType == fileServer && V_destroyMe(vp) == DESTROY_ME) {
        /* Only check destroyMe if we are the fileserver, since the
         * volserver et al sometimes need to work with volumes with
         * destroyMe set. Examples are 'temporary' volumes the
         * volserver creates, and when we create a volume (destroyMe
         * is set on creation; sometimes a separate volserver
         * transaction is created to clear destroyMe).
         */

#if defined(AFS_DEMAND_ATTACH_FS)
        /* schedule a salvage so the volume goes away on disk */
        VRequestSalvage_r(ec, vp, SALVSYNC_ERROR, VOL_SALVAGE_NO_OFFLINE);
        VChangeState_r(vp, VOL_STATE_ERROR);
        vp->nUsers = 0;
        forcefree = 1;
#endif /* AFS_DEMAND_ATTACH_FS */
        Log("VAttachVolume: volume %s is junk; it should be destroyed at next salvage\n", path);
        *ec = VNOVOL;
        goto locked_error;
    }

    vp->vnodeIndex[vSmall].bitmap = vp->vnodeIndex[vLarge].bitmap = NULL;
#ifndef BITMAP_LATER
    if (programType == fileServer && VolumeWriteable(vp)) {
        int i;
        for (i = 0; i < nVNODECLASSES; i++) {
            VGetBitmap_r(ec, vp, i);
            if (*ec) {
#ifdef AFS_DEMAND_ATTACH_FS
                VRequestSalvage_r(ec, vp, SALVSYNC_ERROR, VOL_SALVAGE_NO_OFFLINE);
                vp->nUsers = 0;
#endif /* AFS_DEMAND_ATTACH_FS */
                Log("VAttachVolume: error getting bitmap for volume (%s)\n",
                    path);
                goto locked_error;
            }
        }
    }
#endif /* BITMAP_LATER */

    if (VInit >= 2 && V_needsCallback(vp)) {
        if (V_BreakVolumeCallbacks) {
            Log("VAttachVolume: Volume %lu was changed externally; breaking callbacks\n",
                afs_printable_uint32_lu(V_id(vp)));
            V_needsCallback(vp) = 0;
            VOL_UNLOCK;
            (*V_BreakVolumeCallbacks) (V_id(vp));
            VOL_LOCK;

            VUpdateVolume_r(ec, vp, 0);
        }
#ifdef FSSYNC_BUILD_CLIENT
        else if (VCanUseFSSYNC()) {
            afs_int32 fsync_code;

            V_needsCallback(vp) = 0;
            VOL_UNLOCK;
            fsync_code = FSYNC_VolOp(V_id(vp), NULL, FSYNC_VOL_BREAKCBKS, FSYNC_WHATEVER, NULL);
            VOL_LOCK;

            if (fsync_code) {
                V_needsCallback(vp) = 1;
                Log("Error trying to tell the fileserver to break callbacks for "
                    "changed volume %lu; error code %ld\n",
                    afs_printable_uint32_lu(V_id(vp)),
                    afs_printable_int32_ld(fsync_code));
            } else {
                VUpdateVolume_r(ec, vp, 0);
            }
        }
#endif /* FSSYNC_BUILD_CLIENT */

        if (*ec) {
            Log("VAttachVolume: error %d clearing needsCallback on volume "
                "%lu; needs salvage\n", (int)*ec,
                afs_printable_uint32_lu(V_id(vp)));
#ifdef AFS_DEMAND_ATTACH_FS
            VRequestSalvage_r(ec, vp, SALVSYNC_ERROR, VOL_SALVAGE_NO_OFFLINE);
            vp->nUsers = 0;
#else /* !AFS_DEMAND_ATTACH_FS */
            *ec = VSALVAGE;
#endif /* !AFS_DEMAND_ATTACh_FS */
            goto locked_error;
        }
    }

    if (programType == fileServer) {
        if (vp->specialStatus)
            vp->specialStatus = 0;
        if (V_blessed(vp) && V_inService(vp) && !V_needsSalvaged(vp)) {
            V_inUse(vp) = fileServer;
            V_offlineMessage(vp)[0] = '\0';
        }
#ifdef AFS_DEMAND_ATTACH_FS
        /* check if the volume is actually usable. only do this for DAFS; for
         * non-DAFS, volumes that are not inService/blessed can still be
         * attached, even if clients cannot access them. this is relevant
         * because for non-DAFS, we try to attach the volume when e.g.
         * volserver gives us back then vol when its done with it, but
         * volserver may give us back a volume that is not inService/blessed. */

        if (!V_inUse(vp)) {
            *ec = VNOVOL;
            /* Put the vol into PREATTACHED state, so if someone tries to
             * access it again, we try to attach, see that we're not blessed,
             * and give a VNOVOL error again. Putting it into UNATTACHED state
             * would result in a VOFFLINE error instead. */
            error_state = VOL_STATE_PREATTACHED;

            /* mimic e.g. GetVolume errors */
            if (!V_blessed(vp)) {
                Log("Volume %lu offline: not blessed\n", afs_printable_uint32_lu(V_id(vp)));
                FreeVolumeHeader(vp);
            } else if (!V_inService(vp)) {
                Log("Volume %lu offline: not in service\n", afs_printable_uint32_lu(V_id(vp)));
                /* the volume is offline and should be unattached */
                *ec = VOFFLINE;
                error_state = VOL_STATE_UNATTACHED;
                FreeVolumeHeader(vp);
            } else {
                Log("Volume %lu offline: needs salvage\n", afs_printable_uint32_lu(V_id(vp)));
                *ec = VSALVAGE;
                error_state = VOL_STATE_ERROR;
                /* see if we can recover */
                VRequestSalvage_r(ec, vp, SALVSYNC_NEEDED, VOL_SALVAGE_NO_OFFLINE);
            }
            vp->nUsers = 0;
            goto locked_error;
        }
#endif /* AFS_DEMAND_ATTACH_FS */
    } else {
#ifdef AFS_DEMAND_ATTACH_FS
        if ((mode != V_PEEK) && (mode != V_SECRETLY) && (mode != V_READONLY))
            V_inUse(vp) = programType;
#endif /* AFS_DEMAND_ATTACH_FS */
        V_checkoutMode(vp) = mode;
    }

    AddVolumeToHashTable(vp, vp->hashid);
#ifdef AFS_DEMAND_ATTACH_FS
    if (VCanUnlockAttached() && (V_attachFlags(vp) & VOL_LOCKED)) {
        VUnlockVolume(vp);
    }
    if ((programType != fileServer) ||
        (V_inUse(vp) == fileServer)) {
        AddVolumeToVByPList_r(vp);
        VLRU_Add_r(vp);
        VChangeState_r(vp, VOL_STATE_ATTACHED);
    } else {
        VChangeState_r(vp, VOL_STATE_UNATTACHED);
    }
#endif

    return vp;

unlocked_error:
    VOL_LOCK;
locked_error:
#ifdef AFS_DEMAND_ATTACH_FS
    if (!VIsErrorState(V_attachState(vp))) {
        if (programType != fileServer && *ec == VNOVOL) {
            /* do not log anything in this case; it is common for
             * non-fileserver programs to fail here with VNOVOL, since that
             * is what happens when they simply try to use a volume, but that
             * volume doesn't exist. */

        } else if (VIsErrorState(error_state)) {
            Log("attach2: forcing vol %" AFS_VOLID_FMT " to error state (state %u flags 0x%x ec %d)\n",
                afs_printable_VolumeId_lu(vp->hashid), V_attachState(vp),
                V_attachFlags(vp), *ec);
        }
        VChangeState_r(vp, error_state);
    }
#endif /* AFS_DEMAND_ATTACH_FS */

    if (read_header) {
        VReleaseVolumeHandles_r(vp);
    }

#ifdef AFS_DEMAND_ATTACH_FS
 error_notbroken:
    if (VCheckSalvage(vp) == VCHECK_SALVAGE_FAIL) {
        /* The salvage could not be scheduled with the salvage server
         * due to a hard error. Reset the error code to prevent retry loops by
         * callers. */
        if (*ec == VSALVAGING) {
            *ec = VSALVAGE;
        }
    }
    if (forcefree) {
        FreeVolume(vp);
    } else {
        VCheckFree(vp);
    }
#else /* !AFS_DEMAND_ATTACH_FS */
    FreeVolume(vp);
#endif /* !AFS_DEMAND_ATTACH_FS */
    return NULL;
}

/* Attach an existing volume.
   The volume also normally goes online at this time.
   An offline volume must be reattached to make it go online.
 */

Volume *
VAttachVolume(Error * ec, VolumeId volumeId, int mode)
{
    Volume *retVal;
    VOL_LOCK;
    retVal = VAttachVolume_r(ec, volumeId, mode);
    VOL_UNLOCK;
    return retVal;
}

Volume *
VAttachVolume_r(Error * ec, VolumeId volumeId, int mode)
{
    char *part, *name;
    VGetVolumePath(ec, volumeId, &part, &name);
    if (*ec) {
        Volume *vp;
        Error error;
        vp = VGetVolume_r(&error, volumeId);
        if (vp) {
            opr_Assert(V_inUse(vp) == 0);
            VDetachVolume_r(ec, vp);
        }
        return NULL;
    }
    return VAttachVolumeByName_r(ec, part, name, mode);
}

/* Increment a reference count to a volume, sans context swaps.  Requires
 * possibly reading the volume header in from the disk, since there's
 * an invariant in the volume package that nUsers>0 ==> vp->header is valid.
 *
 * N.B. This call can fail if we can't read in the header!!  In this case
 * we still guarantee we won't context swap, but the ref count won't be
 * incremented (otherwise we'd violate the invariant).
 */
/* NOTE: with the demand attach fileserver extensions, the global lock
 * is dropped within VHold */
#ifdef AFS_DEMAND_ATTACH_FS
static int
VHold_r(Volume * vp)
{
    Error error;

    VCreateReservation_r(vp);
    VWaitExclusiveState_r(vp);

    LoadVolumeHeader(&error, vp);
    if (error) {
        VCancelReservation_r(vp);
        return error;
    }
    vp->nUsers++;
    VCancelReservation_r(vp);
    return 0;
}
#else /* AFS_DEMAND_ATTACH_FS */
static int
VHold_r(Volume * vp)
{
    Error error;

    LoadVolumeHeader(&error, vp);
    if (error)
        return error;
    vp->nUsers++;
    return 0;
}
#endif /* AFS_DEMAND_ATTACH_FS */

/**** volume timeout-related stuff ****/

#ifdef AFS_PTHREAD_ENV

static struct timespec *shutdown_timeout;
static pthread_once_t shutdown_timeout_once = PTHREAD_ONCE_INIT;

static_inline int
VTimedOut(const struct timespec *ts)
{
    struct timeval tv;
    int code;

    if (ts->tv_sec == 0) {
        /* short-circuit; this will have always timed out */
        return 1;
    }

    code = gettimeofday(&tv, NULL);
    if (code) {
        Log("Error %d from gettimeofday, assuming we have not timed out\n", errno);
        /* assume no timeout; failure mode is we just wait longer than normal
         * instead of returning errors when we shouldn't */
        return 0;
    }

    if (tv.tv_sec < ts->tv_sec ||
        (tv.tv_sec == ts->tv_sec && tv.tv_usec*1000 < ts->tv_nsec)) {

        return 0;
    }

    return 1;
}

/**
 * Calculate an absolute timeout.
 *
 * @param[out] ts  A timeout that is "timeout" seconds from now, if we return
 *                 NULL, the memory is not touched
 * @param[in]  timeout  How long the timeout should be from now
 *
 * @return timeout to use
 *  @retval NULL      no timeout; wait forever
 *  @retval non-NULL  the given value for "ts"
 *
 * @internal
 */
static struct timespec *
VCalcTimeout(struct timespec *ts, afs_int32 timeout)
{
    struct timeval now;
    int code;

    if (timeout < 0) {
        return NULL;
    }

    if (timeout == 0) {
        ts->tv_sec = ts->tv_nsec = 0;
        return ts;
    }

    code = gettimeofday(&now, NULL);
    if (code) {
        Log("Error %d from gettimeofday, falling back to 'forever' timeout\n", errno);
        return NULL;
    }

    ts->tv_sec = now.tv_sec + timeout;
    ts->tv_nsec = now.tv_usec * 1000;

    return ts;
}

/**
 * Initialize the shutdown_timeout global.
 */
static void
VShutdownTimeoutInit(void)
{
    struct timespec *ts;

    ts = malloc(sizeof(*ts));

    shutdown_timeout = VCalcTimeout(ts, vol_opts.offline_shutdown_timeout);

    if (!shutdown_timeout) {
        free(ts);
    }
}

/**
 * Figure out the timeout that should be used for waiting for offline volumes.
 *
 * @param[out] ats  Storage space for a local timeout value if needed
 *
 * @return The timeout value that should be used
 *   @retval NULL      No timeout; wait forever for offlining volumes
 *   @retval non-NULL  A pointer to the absolute time that should be used as
 *                     the deadline for waiting for offlining volumes.
 *
 * @note If we return non-NULL, the pointer we return may or may not be the
 *       same as "ats"
 */
static const struct timespec *
VOfflineTimeout(struct timespec *ats)
{
    if (vol_shutting_down) {
        opr_Verify(pthread_once(&shutdown_timeout_once,
                                VShutdownTimeoutInit) == 0);
        return shutdown_timeout;
    } else {
        return VCalcTimeout(ats, vol_opts.offline_timeout);
    }
}

#else /* AFS_PTHREAD_ENV */

/* Waiting a certain amount of time for offlining volumes is not supported
 * for LWP due to a lack of primitives. So, we never time out */
# define VTimedOut(x) (0)
# define VOfflineTimeout(x) (NULL)

#endif /* !AFS_PTHREAD_ENV */

static afs_int32
VIsGoingOffline_r(struct Volume *vp)
{
    afs_int32 code = 0;

    if (vp->goingOffline) {
        if (vp->specialStatus) {
            code = vp->specialStatus;
        } else if (V_inService(vp) == 0 || V_blessed(vp) == 0) {
            code = VNOVOL;
        } else {
            code = VOFFLINE;
        }
    }

    return code;
}

/**
 * Tell the caller if a volume is waiting to go offline.
 *
 * @param[in] vp  The volume we want to know about
 *
 * @return volume status
 *   @retval 0 volume is not waiting to go offline, go ahead and use it
 *   @retval nonzero volume is waiting to offline, and give the returned code
 *           as an error to anyone accessing the volume
 *
 * @pre VOL_LOCK is NOT held
 * @pre caller holds a heavyweight reference on vp
 */
afs_int32
VIsGoingOffline(struct Volume *vp)
{
    afs_int32 code;

    VOL_LOCK;
    code = VIsGoingOffline_r(vp);
    VOL_UNLOCK;

    return code;
}

/**
 * Register an RX call with a volume.
 *
 * @param[inout] ec        Error code; if unset when passed in, may be set if
 *                         the volume starts going offline
 * @param[out]   client_ec @see GetVolume
 * @param[in] vp   Volume struct
 * @param[in] cbv  VCallByVol struct containing the RX call to register
 *
 * @pre VOL_LOCK held
 * @pre caller holds heavy ref on vp
 *
 * @internal
 */
static void
VRegisterCall_r(Error *ec, Error *client_ec, Volume *vp, struct VCallByVol *cbv)
{
    if (vp && cbv) {
#ifdef AFS_DEMAND_ATTACH_FS
        if (!*ec) {
            /* just in case the volume started going offline after we got the
             * reference to it... otherwise, if the volume started going
             * offline right at the end of GetVolume(), we might race with the
             * RX call scanner, and return success and add our cbv to the
             * rx_call_list _after_ the scanner has scanned the list. */
            *ec = VIsGoingOffline_r(vp);
            if (client_ec) {
                *client_ec = *ec;
            }
        }

        while (V_attachState(vp) == VOL_STATE_SCANNING_RXCALLS) {
            VWaitStateChange_r(vp);
        }
#endif /* AFS_DEMAND_ATTACH_FS */

        queue_Prepend(&vp->rx_call_list, cbv);
    }
}

/**
 * Deregister an RX call with a volume.
 *
 * @param[in] vp   Volume struct
 * @param[in] cbv  VCallByVol struct containing the RX call to deregister
 *
 * @pre VOL_LOCK held
 * @pre caller holds heavy ref on vp
 *
 * @internal
 */
static void
VDeregisterCall_r(Volume *vp, struct VCallByVol *cbv)
{
    if (cbv && queue_IsOnQueue(cbv)) {
#ifdef AFS_DEMAND_ATTACH_FS
        while (V_attachState(vp) == VOL_STATE_SCANNING_RXCALLS) {
            VWaitStateChange_r(vp);
        }
#endif /* AFS_DEMAND_ATTACH_FS */

        queue_Remove(cbv);
    }
}

/***************************************************/
/* get and put volume routines                     */
/***************************************************/

/**
 * put back a heavyweight reference to a volume object.
 *
 * @param[in] vp  volume object pointer
 *
 * @pre VOL_LOCK held
 *
 * @post heavyweight volume reference put back.
 *       depending on state, volume may have been taken offline,
 *       detached, salvaged, freed, etc.
 *
 * @internal volume package internal use only
 */
void
VPutVolume_r(Volume * vp)
{
    opr_Verify(--vp->nUsers >= 0);
    if (vp->nUsers == 0) {
        VCheckOffline(vp);
        ReleaseVolumeHeader(vp->header);
#ifdef AFS_DEMAND_ATTACH_FS
        if (!VCheckDetach(vp)) {
            VCheckSalvage(vp);
            VCheckFree(vp);
        }
#else /* AFS_DEMAND_ATTACH_FS */
        VCheckDetach(vp);
#endif /* AFS_DEMAND_ATTACH_FS */
    }
}

void
VPutVolume(Volume * vp)
{
    VOL_LOCK;
    VPutVolume_r(vp);
    VOL_UNLOCK;
}

/**
 * Puts a volume reference obtained with VGetVolumeWithCall.
 *
 * @param[in] vp  Volume struct
 * @param[in] cbv VCallByVol struct given to VGetVolumeWithCall, or NULL if none
 *
 * @pre VOL_LOCK is NOT held
 */
void
VPutVolumeWithCall(Volume *vp, struct VCallByVol *cbv)
{
    VOL_LOCK;
    VDeregisterCall_r(vp, cbv);
    VPutVolume_r(vp);
    VOL_UNLOCK;
}

/* Get a pointer to an attached volume.  The pointer is returned regardless
   of whether or not the volume is in service or on/off line.  An error
   code, however, is returned with an indication of the volume's status */
Volume *
VGetVolume(Error * ec, Error * client_ec, VolumeId volumeId)
{
    Volume *retVal;
    VOL_LOCK;
    retVal = GetVolume(ec, client_ec, volumeId, NULL, 0);
    VOL_UNLOCK;
    return retVal;
}

/**
 * Get a volume reference associated with an RX call.
 *
 * @param[out] ec @see GetVolume
 * @param[out] client_ec @see GetVolume
 * @param[in] volumeId @see GetVolume
 * @param[in] ts  How long to wait for going-offline volumes (absolute time).
 *                If NULL, wait forever. If ts->tv_sec == 0, return immediately
 *                with an error if the volume is going offline.
 * @param[in] cbv Contains an RX call to be associated with this volume
 *                reference. This call may be interrupted if the volume is
 *                requested to go offline while we hold a ref on it. Give NULL
 *                to not associate an RX call with this reference.
 *
 * @return @see GetVolume
 *
 * @note for LWP builds, ts must be NULL
 *
 * @note A reference obtained with this function MUST be put back with
 *       VPutVolumeWithCall
 */
Volume *
VGetVolumeWithCall(Error * ec, Error * client_ec, VolumeId volumeId,
                   const struct timespec *ts, struct VCallByVol *cbv)
{
    Volume *retVal;
    VOL_LOCK;
    retVal = GetVolume(ec, client_ec, volumeId, NULL, ts);
    VRegisterCall_r(ec, client_ec, retVal, cbv);
    VOL_UNLOCK;
    return retVal;
}

Volume *
VGetVolume_r(Error * ec, VolumeId volumeId)
{
    return GetVolume(ec, NULL, volumeId, NULL, NULL);
}

/* try to get a volume we've previously looked up */
/* for demand attach fs, caller MUST NOT hold a ref count on vp */
Volume *
VGetVolumeByVp_r(Error * ec, Volume * vp)
{
    return GetVolume(ec, NULL, vp->hashid, vp, NULL);
}

/**
 * private interface for getting a volume handle
 *
 * @param[out] ec         error code (0 if no error)
 * @param[out] client_ec  wire error code to be given to clients
 * @param[in]  volumeId   ID of the volume we want
 * @param[in]  hint       optional hint for hash lookups, or NULL
 * @param[in]  timeout    absolute deadline for waiting for the volume to go
 *                        offline, if it is going offline. NULL to wait forever.
 *
 * @return a volume handle for the specified volume
 *  @retval NULL an error occurred, or the volume is in such a state that
 *               we cannot load a header or return any volume struct
 *
 * @note for DAFS, caller must NOT hold a ref count on 'hint'
 *
 * @note 'timeout' is only checked if the volume is actually going offline; so
 *       if you pass timeout->tv_sec = 0, this will exhibit typical
 *       nonblocking behavior.
 *
 * @note for LWP builds, 'timeout' must be NULL
 */
static Volume *
GetVolume(Error * ec, Error * client_ec, VolumeId volumeId, Volume * hint,
          const struct timespec *timeout)
{
    Volume *vp = hint;
#ifdef AFS_DEMAND_ATTACH_FS
    Volume *avp, * rvp = hint;
#endif

    /*
     * if VInit is zero, the volume package dynamic
     * data structures have not been initialized yet,
     * and we must immediately return an error
     */
    if (VInit == 0) {
        vp = NULL;
        *ec = VOFFLINE;
        if (client_ec) {
            *client_ec = VOFFLINE;
        }
        goto not_inited;
    }

#ifdef AFS_DEMAND_ATTACH_FS
    if (rvp) {
        VCreateReservation_r(rvp);
    }
#endif /* AFS_DEMAND_ATTACH_FS */

    for (;;) {
        *ec = 0;
        if (client_ec)
            *client_ec = 0;

        vp = VLookupVolume_r(ec, volumeId, vp);
        if (*ec) {
            vp = NULL;
            break;
        }

#ifdef AFS_DEMAND_ATTACH_FS
        if (rvp && (rvp != vp)) {
            /* break reservation on old vp */
            VCancelReservation_r(rvp);
            rvp = NULL;
        }
#endif /* AFS_DEMAND_ATTACH_FS */

        if (!vp) {
            if (VInit < 2) {
                /* Until we have reached an initialization level of 2
                 * we don't know whether this volume exists or not.
                 * We can't sleep and retry later because before a volume
                 * is attached, the caller tries to get it first.  Just
                 * return VOFFLINE and the caller can choose whether to
                 * retry the command or not. */
                *ec = VOFFLINE;
                break;
            }

            *ec = VNOVOL;
            break;
        }

        IncUInt64(&VStats.hdr_gets);

#ifdef AFS_DEMAND_ATTACH_FS
        /* block if someone else is performing an exclusive op on this volume */
        if (rvp != vp) {
            rvp = vp;
            VCreateReservation_r(rvp);
        }
        VWaitExclusiveState_r(vp);

        /* short circuit with VNOVOL in the following circumstances:
         *
         *   - VOL_STATE_ERROR
         *   - VOL_STATE_SHUTTING_DOWN
         */
        if ((V_attachState(vp) == VOL_STATE_ERROR) ||
            (V_attachState(vp) == VOL_STATE_SHUTTING_DOWN)) {
            *ec = VNOVOL;
            vp = NULL;
            break;
        }

        /*
         * short circuit with VOFFLINE for VOL_STATE_UNATTACHED/GOING_OFFLINE and
         *                    VNOVOL   for VOL_STATE_DELETED
         */
       if ((V_attachState(vp) == VOL_STATE_UNATTACHED) ||
           (V_attachState(vp) == VOL_STATE_GOING_OFFLINE) ||
           (V_attachState(vp) == VOL_STATE_DELETED)) {
           if (vp->specialStatus) {
               *ec = vp->specialStatus;
           } else if (V_attachState(vp) == VOL_STATE_DELETED) {
               *ec = VNOVOL;
           } else {
               *ec = VOFFLINE;
           }
           vp = NULL;
           break;
       }

        /* allowable states:
         *   - PREATTACHED
         *   - ATTACHED
         *   - SALVAGING
         *   - SALVAGE_REQ
         */

        if (vp->salvage.requested) {
            VUpdateSalvagePriority_r(vp);
        }

        if (V_attachState(vp) == VOL_STATE_PREATTACHED) {
            if (vp->specialStatus) {
                *ec = vp->specialStatus;
                vp = NULL;
                break;
            }
            avp = VAttachVolumeByVp_r(ec, vp, 0);
            if (avp) {
                if (vp != avp) {
                    /* VAttachVolumeByVp_r can return a pointer
                     * != the vp passed to it under certain
                     * conditions; make sure we don't leak
                     * reservations if that happens */
                    vp = avp;
                    VCancelReservation_r(rvp);
                    rvp = avp;
                    VCreateReservation_r(rvp);
                }
                VPutVolume_r(avp);
            }
            if (*ec) {
                int endloop = 0;
                switch (*ec) {
                case VSALVAGING:
                    break;
                case VOFFLINE:
                    endloop = 1;
                    if (vp->specialStatus) {
                        *ec = vp->specialStatus;
                    }
                    break;

                default:
                    if (vp->specialStatus) {
                        *ec = vp->specialStatus;
                    } else {
                        *ec = VNOVOL;
                    }
                    endloop = 1;
                }
                if (endloop) {
                    vp = NULL;
                    break;
                }
            }
        }

        if (VIsSalvaging(vp) || (*ec == VSALVAGING)) {
            if (client_ec) {
                /* see CheckVnode() in afsfileprocs.c for an explanation
                 * of this error code logic */
                afs_uint32 now = FT_ApproxTime();
                if ((vp->stats.last_salvage + (10 * 60)) >= now) {
                    *client_ec = VBUSY;
                } else {
                    *client_ec = VRESTARTING;
                }
            }
            *ec = VSALVAGING;
            vp = NULL;
            break;
        }

        if (VIsErrorState(V_attachState(vp))) {
            /* make sure we don't take a vp in VOL_STATE_ERROR state and use
             * it, or transition it out of that state */
            if (!*ec) {
                *ec = VNOVOL;
            }
            vp = NULL;
            break;
        }

        /*
         * this test MUST happen after VAttachVolymeByVp, so we have no
         * conflicting vol op. (attach2 would have errored out if we had one;
         * specifically attach_check_vop must have detected a conflicting vop)
         */
         opr_Assert(!vp->pending_vol_op || vp->pending_vol_op->vol_op_state == FSSYNC_VolOpRunningOnline);

#endif /* AFS_DEMAND_ATTACH_FS */

        LoadVolumeHeader(ec, vp);
        if (*ec) {
            /* Only log the error if it was a totally unexpected error.  Simply
             * a missing inode is likely to be caused by the volume being deleted */
            if (errno != ENXIO || GetLogLevel() != 0)
                Log("Volume %" AFS_VOLID_FMT ": couldn't reread volume header\n",
                    afs_printable_VolumeId_lu(vp->hashid));
#ifdef AFS_DEMAND_ATTACH_FS
            if (VCanScheduleSalvage()) {
                VRequestSalvage_r(ec, vp, SALVSYNC_ERROR, 0 /*flags*/);
            } else {
                FreeVolume(vp);
                vp = NULL;
            }
#else /* AFS_DEMAND_ATTACH_FS */
            FreeVolume(vp);
            vp = NULL;
#endif /* AFS_DEMAND_ATTACH_FS */
            break;
        }

        if (vp->shuttingDown) {
            *ec = VNOVOL;
            vp = NULL;
            break;
        }

        if (programType == fileServer) {
            if (vp->goingOffline) {
                if (timeout && VTimedOut(timeout)) {
                    /* we've timed out; don't wait for the vol */
                } else {
#ifdef AFS_DEMAND_ATTACH_FS
                    /* wait for the volume to go offline */
                    if (V_attachState(vp) == VOL_STATE_GOING_OFFLINE) {
                        VTimedWaitStateChange_r(vp, timeout, NULL);
                    }
#elif defined(AFS_PTHREAD_ENV)
                    VOL_CV_TIMEDWAIT(&vol_put_volume_cond, timeout, NULL);
#else /* AFS_PTHREAD_ENV */
                    /* LWP has no timed wait, so the caller better not be
                     * expecting one */
                    opr_Assert(!timeout);
                    LWP_WaitProcess(VPutVolume);
#endif /* AFS_PTHREAD_ENV */
                    continue;
                }
            }
            if (vp->specialStatus) {
                *ec = vp->specialStatus;
            } else if (V_inService(vp) == 0 || V_blessed(vp) == 0) {
                *ec = VNOVOL;
            } else if (V_inUse(vp) == 0 || vp->goingOffline) {
                *ec = VOFFLINE;
            }
        }
        break;
    }

#ifdef AFS_DEMAND_ATTACH_FS
    /* if no error, bump nUsers */
    if (vp) {
        vp->nUsers++;
        VLRU_UpdateAccess_r(vp);
    }
    if (rvp) {
        VCancelReservation_r(rvp);
        rvp = NULL;
    }
    if (client_ec && !*client_ec) {
        *client_ec = *ec;
    }
#else /* AFS_DEMAND_ATTACH_FS */
    /* if no error, bump nUsers */
    if (vp) {
        vp->nUsers++;
    }
    if (client_ec) {
        *client_ec = *ec;
    }
#endif /* AFS_DEMAND_ATTACH_FS */

 not_inited:
    opr_Assert(vp || *ec);
    return vp;
}


/***************************************************/
/* Volume offline/detach routines                  */
/***************************************************/

/* caller MUST hold a heavyweight ref on vp */
#ifdef AFS_DEMAND_ATTACH_FS
void
VTakeOffline_r(Volume * vp)
{
    Error error;

    opr_Assert(vp->nUsers > 0);
    opr_Assert(programType == fileServer);

    VCreateReservation_r(vp);
    VWaitExclusiveState_r(vp);

    vp->goingOffline = 1;
    V_needsSalvaged(vp) = 1;

    VRequestSalvage_r(&error, vp, SALVSYNC_ERROR, 0);
    VCancelReservation_r(vp);
}
#else /* AFS_DEMAND_ATTACH_FS */
void
VTakeOffline_r(Volume * vp)
{
    opr_Assert(vp->nUsers > 0);
    opr_Assert(programType == fileServer);

    vp->goingOffline = 1;
    V_needsSalvaged(vp) = 1;
}
#endif /* AFS_DEMAND_ATTACH_FS */

void
VTakeOffline(Volume * vp)
{
    VOL_LOCK;
    VTakeOffline_r(vp);
    VOL_UNLOCK;
}

/**
 * force a volume offline.
 *
 * @param[in] vp     volume object pointer
 * @param[in] flags  flags (see note below)
 *
 * @note the flag VOL_FORCEOFF_NOUPDATE is a recursion control flag
 *       used when VUpdateVolume_r needs to call VForceOffline_r
 *       (which in turn would normally call VUpdateVolume_r)
 *
 * @see VUpdateVolume_r
 *
 * @pre VOL_LOCK must be held.
 *      for DAFS, caller must hold ref.
 *
 * @note for DAFS, it _is safe_ to call this function from an
 *       exclusive state
 *
 * @post needsSalvaged flag is set.
 *       for DAFS, salvage is requested.
 *       no further references to the volume through the volume
 *       package will be honored.
 *       all file descriptor and vnode caches are invalidated.
 *
 * @warning this is a heavy-handed interface.  it results in
 *          a volume going offline regardless of the current
 *          reference count state.
 *
 * @internal  volume package internal use only
 */
void
VForceOffline_r(Volume * vp, int flags)
{
    Error error;
    if (!V_inUse(vp)) {
#ifdef AFS_DEMAND_ATTACH_FS
        VChangeState_r(vp, VOL_STATE_ERROR);
#endif
        return;
    }

    strcpy(V_offlineMessage(vp),
           "Forced offline due to internal error: volume needs to be salvaged");
    Log("Volume %" AFS_VOLID_FMT " forced offline:  it needs salvaging!\n", afs_printable_VolumeId_lu(V_id(vp)));

    V_inUse(vp) = 0;
    vp->goingOffline = 0;
    V_needsSalvaged(vp) = 1;
    if (!(flags & VOL_FORCEOFF_NOUPDATE)) {
        VUpdateVolume_r(&error, vp, VOL_UPDATE_NOFORCEOFF);
    }

#ifdef AFS_DEMAND_ATTACH_FS
    VRequestSalvage_r(&error, vp, SALVSYNC_ERROR, 0 /*flags*/);
#endif /* AFS_DEMAND_ATTACH_FS */

#ifdef AFS_PTHREAD_ENV
    opr_cv_broadcast(&vol_put_volume_cond);
#else /* AFS_PTHREAD_ENV */
    LWP_NoYieldSignal(VPutVolume);
#endif /* AFS_PTHREAD_ENV */

    VReleaseVolumeHandles_r(vp);
}

/**
 * force a volume offline.
 *
 * @param[in] vp  volume object pointer
 *
 * @see VForceOffline_r
 */
void
VForceOffline(Volume * vp)
{
    VOL_LOCK;
    VForceOffline_r(vp, 0);
    VOL_UNLOCK;
}

/**
 * Iterate over the RX calls associated with a volume, and interrupt them.
 *
 * @param[in] vp The volume whose RX calls we want to scan
 *
 * @pre VOL_LOCK held
 */
static void
VScanCalls_r(struct Volume *vp)
{
    struct VCallByVol *cbv, *ncbv;
    afs_int32 err;
#ifdef AFS_DEMAND_ATTACH_FS
    VolState state_save;
#endif

    if (queue_IsEmpty(&vp->rx_call_list))
        return; /* no calls to interrupt */
    if (!vol_opts.interrupt_rxcall)
        return; /* we have no function with which to interrupt calls */
    err = VIsGoingOffline_r(vp);
    if (!err)
        return; /* we're not going offline anymore */

#ifdef AFS_DEMAND_ATTACH_FS
    VWaitExclusiveState_r(vp);
    state_save = VChangeState_r(vp, VOL_STATE_SCANNING_RXCALLS);
    VOL_UNLOCK;
#endif /* AFS_DEMAND_ATTACH_FS */

    for(queue_Scan(&vp->rx_call_list, cbv, ncbv, VCallByVol)) {
        if (GetLogLevel() != 0) {
            struct rx_peer *peer;
            char hoststr[16];
            peer = rx_PeerOf(rx_ConnectionOf(cbv->call));

            Log("Offlining volume %" AFS_VOLID_FMT " while client %s:%u is trying to read "
                "from it; kicking client off with error %ld\n",
                afs_printable_VolumeId_lu(vp->hashid),
                afs_inet_ntoa_r(rx_HostOf(peer), hoststr),
                (unsigned) ntohs(rx_PortOf(peer)),
                (long) err);
        }
        (*vol_opts.interrupt_rxcall) (cbv->call, err);
    }

#ifdef AFS_DEMAND_ATTACH_FS
    VOL_LOCK;
    VChangeState_r(vp, state_save);
#endif /* AFS_DEMAND_ATTACH_FS */
}

#ifdef AFS_DEMAND_ATTACH_FS
/**
 * Wait for a vp to go offline.
 *
 * @param[out] ec 1 if a salvage on the volume has been requested and
 *                salvok == 0, 0 otherwise
 * @param[in] vp  The volume to wait for
 * @param[in] salvok  If 0, we return immediately with *ec = 1 if the volume
 *                    has been requested to salvage. Otherwise we keep waiting
 *                    until the volume has gone offline.
 *
 * @pre VOL_LOCK held
 * @pre caller holds a lightweight ref on vp
 *
 * @note DAFS only
 */
static void
VWaitForOfflineByVp_r(Error *ec, struct Volume *vp, int salvok)
{
    struct timespec timeout_ts;
    const struct timespec *ts;
    int timedout = 0;

    ts = VOfflineTimeout(&timeout_ts);

    *ec = 0;

    while (!VIsOfflineState(V_attachState(vp)) && !timedout) {
        if (!salvok && vp->salvage.requested) {
            *ec = 1;
            return;
        }
        VTimedWaitStateChange_r(vp, ts, &timedout);
    }
    if (!timedout) {
        /* we didn't time out, so the volume must be offline, so we're done */
        return;
    }

    /* If we got here, we timed out waiting for the volume to go offline.
     * Kick off the accessing RX calls and wait again */

    VScanCalls_r(vp);

    while (!VIsOfflineState(V_attachState(vp))) {
        if (!salvok && vp->salvage.requested) {
            *ec = 1;
            return;
        }

        VWaitStateChange_r(vp);
    }
}

#else /* AFS_DEMAND_ATTACH_FS */

/**
 * Wait for a volume to go offline.
 *
 * @pre VOL_LOCK held
 *
 * @note non-DAFS only (for DAFS, use @see WaitForOfflineByVp_r)
 */
static void