cvmfs.cc

Go to the documentation of this file.

// TODO(jblomer): the file system root should probably always return 1 for an
// inode.  See also integration test #23.

#define ENOATTR ENODATA  
#ifndef __STDC_FORMAT_MACROS
#define __STDC_FORMAT_MACROS
#endif

// sys/xattr.h conflicts with linux/xattr.h and needs to be loaded very early
#include <sys/xattr.h>  // NOLINT

#include "cvmfs_config.h"
#include "cvmfs.h"

#include <dirent.h>
#include <errno.h>
#include <fcntl.h>
#include <google/dense_hash_map>
#include <inttypes.h>
#include <pthread.h>
#include <stddef.h>
#include <stdint.h>
#include <sys/errno.h>
#include <sys/file.h>
#include <sys/mount.h>
#include <sys/resource.h>
#include <sys/stat.h>
#include <sys/time.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <unistd.h>

#include <algorithm>
#include <cassert>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <ctime>
#include <functional>
#include <map>
#include <new>
#include <string>
#include <vector>

#include "authz/authz_session.h"
#include "auto_umount.h"
#include "backoff.h"
#include "cache.h"
#include "cache_posix.h"
#include "cache_stream.h"
#include "catalog_mgr_client.h"
#include "clientctx.h"
#include "compat.h"
#include "compression.h"
#include "crypto/crypto_util.h"
#include "crypto/hash.h"
#include "crypto/signature.h"
#include "directory_entry.h"
#include "duplex_fuse.h"
#include "fence.h"
#include "fetch.h"
#include "file_chunk.h"
#include "fuse_inode_gen.h"
#include "fuse_remount.h"
#include "globals.h"
#include "glue_buffer.h"
#include "history_sqlite.h"
#include "interrupt.h"
#include "loader.h"
#include "lru_md.h"
#include "magic_xattr.h"
#include "manifest_fetch.h"
#include "monitor.h"
#include "mountpoint.h"
#include "network/download.h"
#include "nfs_maps.h"
#include "notification_client.h"
#include "options.h"
#include "quota_listener.h"
#include "quota_posix.h"
#include "shortstring.h"
#include "sqlitemem.h"
#include "sqlitevfs.h"
#include "statistics.h"
#include "talk.h"
#include "telemetry_aggregator.h"
#include "tracer.h"
#include "util/algorithm.h"
#include "util/atomic.h"
#include "util/concurrency.h"
#include "util/exception.h"
#include "util/logging.h"
#include "util/platform.h"
#include "util/smalloc.h"
#include "util/uuid.h"
#include "wpad.h"
#include "xattr.h"

using namespace std;  // NOLINT

namespace cvmfs {

FileSystem *file_system_ = NULL;
MountPoint *mount_point_ = NULL;
TalkManager *talk_mgr_ = NULL;
NotificationClient *notification_client_ = NULL;
Watchdog *watchdog_ = NULL;
FuseRemounter *fuse_remounter_ = NULL;
InodeGenerationInfo inode_generation_info_;


struct DirectoryListing {
  char *buffer;  
  // Not really used anymore.  But directory listing needs to be migrated during
  // hotpatch. If buffer is allocated by smmap, capacity is zero.
  size_t size;
  size_t capacity;

  DirectoryListing() : buffer(NULL), size(0), capacity(0) { }
};

const loader::LoaderExports *loader_exports_ = NULL;
OptionsManager *options_mgr_ = NULL;
pid_t pid_ = 0;  
quota::ListenerHandle *watchdog_listener_ = NULL;
quota::ListenerHandle *unpin_listener_ = NULL;


typedef google::dense_hash_map<uint64_t, DirectoryListing,
                               hash_murmur<uint64_t> >
        DirectoryHandles;
DirectoryHandles *directory_handles_ = NULL;
pthread_mutex_t lock_directory_handles_ = PTHREAD_MUTEX_INITIALIZER;
uint64_t next_directory_handle_ = 0;

unsigned max_open_files_; 
bool check_fd_overflow_ = true;
const int kNumReservedFd = 512;
const unsigned int kMinOpenFiles = 8192;


class FuseInterruptCue : public InterruptCue {
 public:
  explicit FuseInterruptCue(fuse_req_t *r) : req_ptr_(r) { }
  virtual ~FuseInterruptCue() { }
  virtual bool IsCanceled() { return fuse_req_interrupted(*req_ptr_); }
 private:
  fuse_req_t *req_ptr_;
};

struct FuseState {
  FuseState() : version(0), cache_symlinks(false), has_dentry_expire(false) {}
  unsigned version;
  bool cache_symlinks;
  bool has_dentry_expire;
};


static inline bool IncAndCheckNoOpenFiles() {
  const int64_t no_open_files = perf::Xadd(file_system_->no_open_files(), 1);
  if (!check_fd_overflow_)
    return true;
  return no_open_files < (static_cast<int>(max_open_files_) - kNumReservedFd);
}

static inline double GetKcacheTimeout() {
  if (!fuse_remounter_->IsCaching())
    return 0.0;
  return mount_point_->kcache_timeout_sec();
}


void GetReloadStatus(bool *drainout_mode, bool *maintenance_mode) {
  *drainout_mode = fuse_remounter_->IsInDrainoutMode();
  *maintenance_mode = fuse_remounter_->IsInMaintenanceMode();
}


static bool UseWatchdog() {
  if (loader_exports_ == NULL || loader_exports_->version < 2) {
    return true;  // spawn watchdog by default
                  // Note: with library versions before 2.1.8 it might not
                  //       create stack traces properly in all cases
  }

  return !loader_exports_->disable_watchdog;
}

std::string PrintInodeGeneration() {
  return "init-catalog-revision: " +
    StringifyInt(inode_generation_info_.initial_revision) + "  " +
    "current-catalog-revision: " +
    StringifyInt(mount_point_->catalog_mgr()->GetRevision()) + "  " +
    "incarnation: " + StringifyInt(inode_generation_info_.incarnation) + "  " +
    "inode generation: " + StringifyInt(inode_generation_info_.inode_generation)
    + "\n";
}


static bool CheckVoms(const fuse_ctx &fctx) {
  if (!mount_point_->has_membership_req())
    return true;
  string mreq = mount_point_->membership_req();
  LogCvmfs(kLogCvmfs, kLogDebug, "Got VOMS authz %s from filesystem "
           "properties", mreq.c_str());

  if (fctx.uid == 0)
    return true;

  return mount_point_->authz_session_mgr()->IsMemberOf(fctx.pid, mreq);
}

static bool MayBeInPageCacheTracker(const catalog::DirectoryEntry &dirent) {
  return dirent.IsRegular() &&
         (dirent.inode() < mount_point_->catalog_mgr()->GetRootInode());
}

static bool HasDifferentContent(
  const catalog::DirectoryEntry &dirent,
  const shash::Any &hash,
  const struct stat &info)
{
  if (hash == dirent.checksum())
    return false;
  // For chunked files, we don't want to load the full list of chunk hashes
  // so we only check the last modified timestamp
  if (dirent.IsChunkedFile() && (info.st_mtime == dirent.mtime()))
    return false;
  return true;
}

static bool FixupOpenInode(const PathString &path,
                           catalog::DirectoryEntry *dirent)
{
  if (!MayBeInPageCacheTracker(*dirent))
    return false;

  shash::Any hash_open;
  struct stat info;
  bool is_open = mount_point_->page_cache_tracker()->GetInfoIfOpen(
    dirent->inode(), &hash_open, &info);
  if (!is_open)
    return false;
  if (!HasDifferentContent(*dirent, hash_open, info))
    return false;

  // Overwrite dirent with inode from current generation
  bool found = mount_point_->catalog_mgr()->LookupPath(
      path, catalog::kLookupDefault, dirent);
  assert(found);

  return true;
}

static bool GetDirentForInode(const fuse_ino_t ino,
                              catalog::DirectoryEntry *dirent)
{
  // Lookup inode in cache
  if (mount_point_->inode_cache()->Lookup(ino, dirent))
    return true;

  // Look in the catalogs in 2 steps: lookup inode->path, lookup path
  static catalog::DirectoryEntry dirent_negative =
    catalog::DirectoryEntry(catalog::kDirentNegative);
  // Reset directory entry.  If the function returns false and dirent is no
  // the kDirentNegative, it was an I/O error
  *dirent = catalog::DirectoryEntry();

  catalog::ClientCatalogManager *catalog_mgr = mount_point_->catalog_mgr();

  if (file_system_->IsNfsSource()) {
    // NFS mode
    PathString path;
    bool retval = file_system_->nfs_maps()->GetPath(ino, &path);
    if (!retval) {
      *dirent = dirent_negative;
      return false;
    }
    if (catalog_mgr->LookupPath(path, catalog::kLookupDefault, dirent)) {
      // Fix inodes
      dirent->set_inode(ino);
      mount_point_->inode_cache()->Insert(ino, *dirent);
      return true;
    }
    return false;  // Not found in catalog or catalog load error
  }

  // Non-NFS mode
  PathString path;
  if (ino == catalog_mgr->GetRootInode()) {
    bool retval =
      catalog_mgr->LookupPath(PathString(), catalog::kLookupDefault, dirent);

    if (!AssertOrLog(retval, kLogCvmfs, kLogSyslogWarn | kLogDebug,
                     "GetDirentForInode: Race condition? Not found dirent %s",
                     dirent->name().c_str())) {
      return false;
    }

    dirent->set_inode(ino);
    mount_point_->inode_cache()->Insert(ino, *dirent);
    return true;
  }

  glue::InodeEx inode_ex(ino, glue::InodeEx::kUnknownType);
  bool retval = mount_point_->inode_tracker()->FindPath(&inode_ex, &path);
  if (!retval) {
    // This may be a retired inode whose stat information is only available
    // in the page cache tracker because there is still an open file
    LogCvmfs(kLogCvmfs, kLogDebug,
             "GetDirentForInode inode lookup failure %" PRId64, ino);
    *dirent = dirent_negative;
    // Indicate that the inode was not found in the tracker rather than not
    // found in the catalog
    dirent->set_inode(ino);
    return false;
  }
  if (catalog_mgr->LookupPath(path, catalog::kLookupDefault, dirent)) {
    if (!inode_ex.IsCompatibleFileType(dirent->mode())) {
      LogCvmfs(kLogCvmfs, kLogDebug,
               "Warning: inode %" PRId64 " (%s) changed file type",
               ino, path.c_str());
      // TODO(jblomer): we detect this issue but let it continue unhandled.
      // Fix me.
    }

    // Fix inodes
    dirent->set_inode(ino);
    mount_point_->inode_cache()->Insert(ino, *dirent);
    return true;
  }

  // Can happen after reload of catalogs or on catalog load failure
  LogCvmfs(kLogCvmfs, kLogDebug, "GetDirentForInode path lookup failure");
  return false;
}


static uint64_t GetDirentForPath(const PathString &path,
                                 catalog::DirectoryEntry *dirent)
{
  uint64_t live_inode = 0;
  if (!file_system_->IsNfsSource())
    live_inode = mount_point_->inode_tracker()->FindInode(path);

  shash::Md5 md5path(path.GetChars(), path.GetLength());
  if (mount_point_->md5path_cache()->Lookup(md5path, dirent)) {
    if (dirent->GetSpecial() == catalog::kDirentNegative)
      return false;
    // We may have initially stored the entry with an old inode in the
    // md5path cache and now should update it with the new one.
    if (!file_system_->IsNfsSource() && (live_inode != 0))
      dirent->set_inode(live_inode);
    return 1;
  }

  catalog::ClientCatalogManager *catalog_mgr = mount_point_->catalog_mgr();

  // Lookup inode in catalog TODO: not twice md5 calculation
  bool retval;
  retval = catalog_mgr->LookupPath(path, catalog::kLookupDefault, dirent);
  if (retval) {
    if (file_system_->IsNfsSource()) {
      dirent->set_inode(file_system_->nfs_maps()->GetInode(path));
    } else if (live_inode != 0) {
      dirent->set_inode(live_inode);
      if (FixupOpenInode(path, dirent)) {
        LogCvmfs(kLogCvmfs, kLogDebug,
          "content of %s change, replacing inode %" PRIu64 " --> %" PRIu64,
          path.c_str(), live_inode, dirent->inode());
        return live_inode;
        // Do not populate the md5path cache until the inode tracker is fixed
      }
    }
    mount_point_->md5path_cache()->Insert(md5path, *dirent);
    return 1;
  }

  LogCvmfs(kLogCvmfs, kLogDebug, "GetDirentForPath, no entry");
  // Only insert ENOENT results into negative cache.  Otherwise it was an
  // error loading nested catalogs
  if (dirent->GetSpecial() == catalog::kDirentNegative)
    mount_point_->md5path_cache()->InsertNegative(md5path);
  return 0;
}


static bool GetPathForInode(const fuse_ino_t ino, PathString *path) {
  // Check the path cache first
  if (mount_point_->path_cache()->Lookup(ino, path))
    return true;

  if (file_system_->IsNfsSource()) {
    // NFS mode, just a lookup
    LogCvmfs(kLogCvmfs, kLogDebug, "MISS %d - lookup in NFS maps", ino);
    if (file_system_->nfs_maps()->GetPath(ino, path)) {
      mount_point_->path_cache()->Insert(ino, *path);
      return true;
    }
    return false;
  }

  if (ino == mount_point_->catalog_mgr()->GetRootInode())
    return true;

  LogCvmfs(kLogCvmfs, kLogDebug, "MISS %d - looking in inode tracker", ino);
  glue::InodeEx inode_ex(ino, glue::InodeEx::kUnknownType);
  bool retval = mount_point_->inode_tracker()->FindPath(&inode_ex, path);

  if (!AssertOrLog(retval, kLogCvmfs, kLogSyslogWarn | kLogDebug,
                   "GetPathForInode: Race condition? "
                   "Inode not found in inode tracker at path %s",
                   path->c_str())) {
    return false;
  }


  mount_point_->path_cache()->Insert(ino, *path);
  return true;
}

static void DoTraceInode(const int event,
                          fuse_ino_t ino,
                          const std::string &msg)
{
  PathString path;
  bool found = GetPathForInode(ino, &path);
  if (!found) {
    LogCvmfs(kLogCvmfs, kLogDebug,
      "Tracing: Could not find path for inode %" PRIu64, uint64_t(ino));
    mount_point_->tracer()->Trace(event, PathString("@UNKNOWN"), msg);
  } else {
    mount_point_->tracer()->Trace(event, path, msg);
  }
}

static void inline TraceInode(const int event,
                                fuse_ino_t ino,
                                const std::string &msg)
{
  if (mount_point_->tracer()->IsActive()) DoTraceInode(event, ino, msg);
}

static void cvmfs_lookup(fuse_req_t req, fuse_ino_t parent, const char *name) {
  HighPrecisionTimer guard_timer(file_system_->hist_fs_lookup());

  perf::Inc(file_system_->n_fs_lookup());
  const struct fuse_ctx *fuse_ctx = fuse_req_ctx(req);
  FuseInterruptCue ic(&req);
  ClientCtxGuard ctx_guard(fuse_ctx->uid, fuse_ctx->gid, fuse_ctx->pid, &ic);
  fuse_remounter_->TryFinish();

  fuse_remounter_->fence()->Enter();
  catalog::ClientCatalogManager *catalog_mgr = mount_point_->catalog_mgr();

  fuse_ino_t parent_fuse = parent;
  parent = catalog_mgr->MangleInode(parent);
  LogCvmfs(kLogCvmfs, kLogDebug,
           "cvmfs_lookup in parent inode: %" PRIu64 " for name: %s",
           uint64_t(parent), name);

  PathString path;
  PathString parent_path;
  uint64_t live_inode = 0;
  catalog::DirectoryEntry dirent;
  struct fuse_entry_param result;

  memset(&result, 0, sizeof(result));
  double timeout = GetKcacheTimeout();
  result.attr_timeout = timeout;
  result.entry_timeout = timeout;

  // Special NFS lookups: . and ..
  if ((strcmp(name, ".") == 0) || (strcmp(name, "..") == 0)) {
    if (GetDirentForInode(parent, &dirent)) {
      if (strcmp(name, ".") == 0) {
        goto lookup_reply_positive;
      } else {
        // Lookup for ".."
        if (dirent.inode() == catalog_mgr->GetRootInode()) {
          dirent.set_inode(1);
          goto lookup_reply_positive;
        }
        if (!GetPathForInode(parent, &parent_path))
          goto lookup_reply_negative;
        if (GetDirentForPath(GetParentPath(parent_path), &dirent) > 0)
          goto lookup_reply_positive;
      }
    }
    // No entry for "." or no entry for ".."
    if (dirent.GetSpecial() == catalog::kDirentNegative)
      goto lookup_reply_negative;
    else
      goto lookup_reply_error;
    assert(false);
  }

  if (!GetPathForInode(parent, &parent_path)) {
    LogCvmfs(kLogCvmfs, kLogDebug, "no path for parent inode found");
    goto lookup_reply_negative;
  }

  path.Assign(parent_path);
  path.Append("/", 1);
  path.Append(name, strlen(name));
  mount_point_->tracer()->Trace(Tracer::kEventLookup, path, "lookup()");
  live_inode = GetDirentForPath(path, &dirent);
  if (live_inode == 0) {
    if (dirent.GetSpecial() == catalog::kDirentNegative)
      goto lookup_reply_negative;
    else
      goto lookup_reply_error;
  }

 lookup_reply_positive:
  if (!file_system_->IsNfsSource()) {
    if (live_inode > 1) {
      // live inode is stale (open file), we replace it
      assert(dirent.IsRegular());
      assert(dirent.inode() != live_inode);
      // The new inode is put in the tracker with refcounter == 0
      bool replaced = mount_point_->inode_tracker()->ReplaceInode(
        live_inode, glue::InodeEx(dirent.inode(), dirent.mode()));
      if (replaced)
        perf::Inc(file_system_->n_fs_inode_replace());
    }
    mount_point_->inode_tracker()->VfsGet(
      glue::InodeEx(dirent.inode(), dirent.mode()), path);
  }
  // We do _not_ track (and evict) positive replies; among other things, test
  // 076 fails with the following line uncommented
  //
  // WARNING! ENABLING THIS BREAKS ANY TYPE OF MOUNTPOINT POINTING TO THIS INODE
  //
  // only safe if fuse_expire_entry is available
  if (mount_point_->fuse_expire_entry()
      || (mount_point_->cache_symlinks() && dirent.IsLink())) {
    LogCvmfs(kLogCache, kLogDebug, "Dentry to evict: %s", name);
    mount_point_->dentry_tracker()->Add(parent_fuse, name,
                                        static_cast<uint64_t>(timeout));
  }

  fuse_remounter_->fence()->Leave();
  result.ino = dirent.inode();
  result.attr = dirent.GetStatStructure();
  fuse_reply_entry(req, &result);
  return;

 lookup_reply_negative:
  // Will be a no-op if there is no fuse cache eviction
  mount_point_->dentry_tracker()->Add(parent_fuse, name, uint64_t(timeout));
  fuse_remounter_->fence()->Leave();
  perf::Inc(file_system_->n_fs_lookup_negative());
  result.ino = 0;
  fuse_reply_entry(req, &result);
  return;

 lookup_reply_error:
  fuse_remounter_->fence()->Leave();

  LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr, "EIO (01) on %s", name);
  perf::Inc(file_system_->n_eio_total());
  perf::Inc(file_system_->n_eio_01());

  fuse_reply_err(req, EIO);
}


static void cvmfs_forget(
  fuse_req_t req,
  fuse_ino_t ino,
#if CVMFS_USE_LIBFUSE == 2
  unsigned long nlookup  // NOLINT
#else
  uint64_t nlookup
#endif
) {
  HighPrecisionTimer guard_timer(file_system_->hist_fs_forget());

  perf::Inc(file_system_->n_fs_forget());

  // The libfuse high-level library does the same
  if (ino == FUSE_ROOT_ID) {
    fuse_reply_none(req);
    return;
  }

  fuse_remounter_->fence()->Enter();
  ino = mount_point_->catalog_mgr()->MangleInode(ino);
  // This has been seen to deadlock on the debug log mutex on SL5.  Problem of
  // old kernel/fuse?
#if CVMFS_USE_LIBCVMFS == 2
  LogCvmfs(kLogCvmfs, kLogDebug, "forget on inode %" PRIu64 " by %u",
           uint64_t(ino), nlookup);
#else
  LogCvmfs(kLogCvmfs, kLogDebug, "forget on inode %" PRIu64 " by %" PRIu64,
           uint64_t(ino), nlookup);
#endif
  if (!file_system_->IsNfsSource()) {
    bool removed =
      mount_point_->inode_tracker()->GetVfsPutRaii().VfsPut(ino, nlookup);
    if (removed)
      mount_point_->page_cache_tracker()->GetEvictRaii().Evict(ino);
  }
  fuse_remounter_->fence()->Leave();
  fuse_reply_none(req);
}


#if (FUSE_VERSION >= 29)
static void cvmfs_forget_multi(
  fuse_req_t req,
  size_t count,
  struct fuse_forget_data *forgets
) {
  HighPrecisionTimer guard_timer(file_system_->hist_fs_forget_multi());

  perf::Xadd(file_system_->n_fs_forget(), count);
  if (file_system_->IsNfsSource()) {
    fuse_reply_none(req);
    return;
  }

  fuse_remounter_->fence()->Enter();
  {
    glue::InodeTracker::VfsPutRaii vfs_put_raii =
      mount_point_->inode_tracker()->GetVfsPutRaii();
    glue::PageCacheTracker::EvictRaii evict_raii =
      mount_point_->page_cache_tracker()->GetEvictRaii();
    for (size_t i = 0; i < count; ++i) {
      if (forgets[i].ino == FUSE_ROOT_ID) {
        continue;
      }

      uint64_t ino = mount_point_->catalog_mgr()->MangleInode(forgets[i].ino);
      LogCvmfs(kLogCvmfs, kLogDebug, "forget on inode %" PRIu64 " by %" PRIu64,
               ino, forgets[i].nlookup);

      bool removed = vfs_put_raii.VfsPut(ino, forgets[i].nlookup);
      if (removed)
        evict_raii.Evict(ino);
    }
  }
  fuse_remounter_->fence()->Leave();

  fuse_reply_none(req);
}
#endif  // FUSE_VERSION >= 29


static void ReplyNegative(const catalog::DirectoryEntry &dirent,
                          fuse_req_t req)
{
  if (dirent.GetSpecial() == catalog::kDirentNegative) {
    fuse_reply_err(req, ENOENT);
  } else {
    const char * name = dirent.name().c_str();
    const char * link = dirent.symlink().c_str();

    LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
       "EIO (02) name=%s symlink=%s",
       name ? name: "<unset>",
       link ? link: "<unset>");

    perf::Inc(file_system_->n_eio_total());
    perf::Inc(file_system_->n_eio_02());
    fuse_reply_err(req, EIO);
  }
}


static void cvmfs_getattr(fuse_req_t req, fuse_ino_t ino,
                          struct fuse_file_info *fi)
{
  HighPrecisionTimer guard_timer(file_system_->hist_fs_getattr());

  perf::Inc(file_system_->n_fs_stat());
  const struct fuse_ctx *fuse_ctx = fuse_req_ctx(req);
  FuseInterruptCue ic(&req);
  ClientCtxGuard ctx_guard(fuse_ctx->uid, fuse_ctx->gid, fuse_ctx->pid, &ic);
  fuse_remounter_->TryFinish();

  fuse_remounter_->fence()->Enter();
  ino = mount_point_->catalog_mgr()->MangleInode(ino);
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_getattr (stat) for inode: %" PRIu64,
           uint64_t(ino));

  if (!CheckVoms(*fuse_ctx)) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, EACCES);
    return;
  }
  catalog::DirectoryEntry dirent;
  bool found = GetDirentForInode(ino, &dirent);
  TraceInode(Tracer::kEventGetAttr, ino, "getattr()");
  if ((!found && (dirent.inode() == ino)) || MayBeInPageCacheTracker(dirent)) {
    // Serve retired inode from page cache tracker; even if we find it in the
    // catalog, we replace the dirent by the page cache tracker version to
    // not confuse open file handles
    LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_getattr %" PRIu64 " "
             "served from page cache tracker", ino);
    shash::Any hash;
    struct stat info;
    bool is_open =
      mount_point_->page_cache_tracker()->GetInfoIfOpen(ino, &hash, &info);
    if (is_open) {
      fuse_remounter_->fence()->Leave();
      if (found && HasDifferentContent(dirent, hash, info)) {
        // We should from now on provide the new inode information instead
        // of the stale one. To this end, we need to invalidate the dentry to
        // trigger a fresh LOOKUP call
        uint64_t parent_ino;
        NameString name;
        if (mount_point_->inode_tracker()->FindDentry(
              dirent.inode(), &parent_ino, &name))
        {
          fuse_remounter_->InvalidateDentry(parent_ino, name);
        }
        perf::Inc(file_system_->n_fs_stat_stale());
      }
      fuse_reply_attr(req, &info, GetKcacheTimeout());
      return;
    }
  }
  fuse_remounter_->fence()->Leave();

  if (!found) {
    ReplyNegative(dirent, req);
    return;
  }

  struct stat info = dirent.GetStatStructure();

  fuse_reply_attr(req, &info, GetKcacheTimeout());
}


static void cvmfs_readlink(fuse_req_t req, fuse_ino_t ino) {
  HighPrecisionTimer guard_timer(file_system_->hist_fs_readlink());

  perf::Inc(file_system_->n_fs_readlink());
  const struct fuse_ctx *fuse_ctx = fuse_req_ctx(req);
  FuseInterruptCue ic(&req);
  ClientCtxGuard ctx_guard(fuse_ctx->uid, fuse_ctx->gid, fuse_ctx->pid, &ic);

  fuse_remounter_->fence()->Enter();
  ino = mount_point_->catalog_mgr()->MangleInode(ino);
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_readlink on inode: %" PRIu64,
           uint64_t(ino));

  catalog::DirectoryEntry dirent;
  const bool found = GetDirentForInode(ino, &dirent);
  TraceInode(Tracer::kEventReadlink, ino, "readlink()");
  fuse_remounter_->fence()->Leave();

  if (!found) {
    ReplyNegative(dirent, req);
    return;
  }

  if (!dirent.IsLink()) {
    fuse_reply_err(req, EINVAL);
    return;
  }

  fuse_reply_readlink(req, dirent.symlink().c_str());
}


static void AddToDirListing(const fuse_req_t req,
                            const char *name, const struct stat *stat_info,
                            BigVector<char> *listing)
{
  LogCvmfs(kLogCvmfs, kLogDebug, "Add to listing: %s, inode %" PRIu64,
           name, uint64_t(stat_info->st_ino));
  size_t remaining_size = listing->capacity() - listing->size();
  const size_t entry_size = fuse_add_direntry(req, NULL, 0, name, stat_info, 0);

  while (entry_size > remaining_size) {
    listing->DoubleCapacity();
    remaining_size = listing->capacity() - listing->size();
  }

  char *buffer;
  bool large_alloc;
  listing->ShareBuffer(&buffer, &large_alloc);
  fuse_add_direntry(req, buffer + listing->size(),
                    remaining_size, name, stat_info,
                    listing->size() + entry_size);
  listing->SetSize(listing->size() + entry_size);
}


static void cvmfs_opendir(fuse_req_t req, fuse_ino_t ino,
                          struct fuse_file_info *fi)
{
  HighPrecisionTimer guard_timer(file_system_->hist_fs_opendir());

  const struct fuse_ctx *fuse_ctx = fuse_req_ctx(req);
  FuseInterruptCue ic(&req);
  ClientCtxGuard ctx_guard(fuse_ctx->uid, fuse_ctx->gid, fuse_ctx->pid, &ic);
  fuse_remounter_->TryFinish();

  fuse_remounter_->fence()->Enter();
  catalog::ClientCatalogManager *catalog_mgr = mount_point_->catalog_mgr();
  ino = catalog_mgr->MangleInode(ino);
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_opendir on inode: %" PRIu64,
           uint64_t(ino));
  if (!CheckVoms(*fuse_ctx)) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, EACCES);
    return;
  }

  TraceInode(Tracer::kEventOpenDir, ino, "opendir()");
  PathString path;
  catalog::DirectoryEntry d;
  bool found = GetPathForInode(ino, &path);
  if (!found) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, ENOENT);
    return;
  }
  found = GetDirentForInode(ino, &d);

  if (!found) {
    fuse_remounter_->fence()->Leave();
    ReplyNegative(d, req);
    return;
  }
  if (!d.IsDirectory()) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, ENOTDIR);
    return;
  }

  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_opendir on inode: %" PRIu64 ", path %s",
           uint64_t(ino), path.c_str());

  // Build listing
  BigVector<char> fuse_listing(512);

  // Add current directory link
  struct stat info;
  info = d.GetStatStructure();
  AddToDirListing(req, ".", &info, &fuse_listing);

  // Add parent directory link
  catalog::DirectoryEntry p;
  if (d.inode() != catalog_mgr->GetRootInode() &&
      (GetDirentForPath(GetParentPath(path), &p) > 0))
  {
    info = p.GetStatStructure();
    AddToDirListing(req, "..", &info, &fuse_listing);
  }

  // Add all names
  catalog::StatEntryList listing_from_catalog;
  bool retval = catalog_mgr->ListingStat(path, &listing_from_catalog);

  if (!retval) {
    fuse_remounter_->fence()->Leave();
    fuse_listing.Clear();  // Buffer is shared, empty manually

    LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
         "EIO (03) on %s", path.c_str());
    perf::Inc(file_system_->n_eio_total());
    perf::Inc(file_system_->n_eio_03());
    fuse_reply_err(req, EIO);
    return;
  }
  for (unsigned i = 0; i < listing_from_catalog.size(); ++i) {
    // Fix inodes
    PathString entry_path;
    entry_path.Assign(path);
    entry_path.Append("/", 1);
    entry_path.Append(listing_from_catalog.AtPtr(i)->name.GetChars(),
                      listing_from_catalog.AtPtr(i)->name.GetLength());

    catalog::DirectoryEntry entry_dirent;
    if (!GetDirentForPath(entry_path, &entry_dirent)) {
      LogCvmfs(kLogCvmfs, kLogDebug, "listing entry %s vanished, skipping",
               entry_path.c_str());
      continue;
    }

    struct stat fixed_info = listing_from_catalog.AtPtr(i)->info;
    fixed_info.st_ino = entry_dirent.inode();
    AddToDirListing(req, listing_from_catalog.AtPtr(i)->name.c_str(),
                    &fixed_info, &fuse_listing);
  }
  fuse_remounter_->fence()->Leave();

  DirectoryListing stream_listing;
  stream_listing.size = fuse_listing.size();
  stream_listing.capacity = fuse_listing.capacity();
  bool large_alloc;
  fuse_listing.ShareBuffer(&stream_listing.buffer, &large_alloc);
  if (large_alloc)
    stream_listing.capacity = 0;

  // Save the directory listing and return a handle to the listing
  {
    MutexLockGuard m(&lock_directory_handles_);
    LogCvmfs(kLogCvmfs, kLogDebug,
             "linking directory handle %d to dir inode: %" PRIu64,
             next_directory_handle_, uint64_t(ino));
    (*directory_handles_)[next_directory_handle_] = stream_listing;
    fi->fh = next_directory_handle_;
    ++next_directory_handle_;
  }
  perf::Inc(file_system_->n_fs_dir_open());
  perf::Inc(file_system_->no_open_dirs());

#if (FUSE_VERSION >= 30)
#ifdef CVMFS_ENABLE_FUSE3_CACHE_READDIR
  // This affects only reads on the same open directory handle (e.g. multiple
  // reads with rewinddir() between them).  A new opendir on the same directory
  // will trigger readdir calls independently of this setting.
  fi->cache_readdir = 1;
#endif
#endif
  fuse_reply_open(req, fi);
}


static void cvmfs_releasedir(fuse_req_t req, fuse_ino_t ino,
                             struct fuse_file_info *fi)
{
  HighPrecisionTimer guard_timer(file_system_->hist_fs_releasedir());

  ino = mount_point_->catalog_mgr()->MangleInode(ino);
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_releasedir on inode %" PRIu64
           ", handle %d", uint64_t(ino), fi->fh);

  int reply = 0;

  {
    MutexLockGuard m(&lock_directory_handles_);
    DirectoryHandles::iterator iter_handle = directory_handles_->find(fi->fh);
    if (iter_handle != directory_handles_->end()) {
      if (iter_handle->second.capacity == 0)
        smunmap(iter_handle->second.buffer);
      else
        free(iter_handle->second.buffer);
      directory_handles_->erase(iter_handle);
      perf::Dec(file_system_->no_open_dirs());
    } else {
      reply = EINVAL;
    }
  }

  fuse_reply_err(req, reply);
}


static void ReplyBufferSlice(const fuse_req_t req, const char *buffer,
                             const size_t buffer_size, const off_t offset,
                             const size_t max_size)
{
  if (offset < static_cast<int>(buffer_size)) {
    fuse_reply_buf(req, buffer + offset,
      std::min(static_cast<size_t>(buffer_size - offset), max_size));
  } else {
    fuse_reply_buf(req, NULL, 0);
  }
}


static void cvmfs_readdir(fuse_req_t req, fuse_ino_t ino, size_t size,
                          off_t off, struct fuse_file_info *fi)
{
  HighPrecisionTimer guard_timer(file_system_->hist_fs_readdir());

  LogCvmfs(kLogCvmfs, kLogDebug,
           "cvmfs_readdir on inode %" PRIu64 " reading %d bytes from offset %d",
           uint64_t(mount_point_->catalog_mgr()->MangleInode(ino)), size, off);

  DirectoryListing listing;

  MutexLockGuard m(&lock_directory_handles_);
  DirectoryHandles::const_iterator iter_handle =
    directory_handles_->find(fi->fh);
  if (iter_handle != directory_handles_->end()) {
    listing = iter_handle->second;

    ReplyBufferSlice(req, listing.buffer, listing.size, off, size);
    return;
  }

  fuse_reply_err(req, EINVAL);
}

static void FillOpenFlags(const glue::PageCacheTracker::OpenDirectives od,
                          struct fuse_file_info *fi)
{
  assert(!TestBit(glue::PageCacheTracker::kBitDirectIo, fi->fh));
  fi->keep_cache = od.keep_cache;
  fi->direct_io = od.direct_io;
  if (fi->direct_io)
    SetBit(glue::PageCacheTracker::kBitDirectIo, &fi->fh);
}


#ifdef __APPLE__
// On macOS, xattr on a symlink opens and closes the file (with O_SYMLINK)
// around the actual getxattr call. In order to not run into an I/O error
// we use a special file handle for symlinks, from which one cannot read.
static const uint64_t kFileHandleIgnore = static_cast<uint64_t>(2) << 60;
#endif

static void cvmfs_open(fuse_req_t req, fuse_ino_t ino,
                       struct fuse_file_info *fi)
{
  HighPrecisionTimer guard_timer(file_system_->hist_fs_open());

  const struct fuse_ctx *fuse_ctx = fuse_req_ctx(req);
  FuseInterruptCue ic(&req);
  ClientCtxGuard ctx_guard(fuse_ctx->uid, fuse_ctx->gid, fuse_ctx->pid, &ic);
  fuse_remounter_->fence()->Enter();
  catalog::ClientCatalogManager *catalog_mgr = mount_point_->catalog_mgr();
  ino = catalog_mgr->MangleInode(ino);
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_open on inode: %" PRIu64,
           uint64_t(ino));

  int fd = -1;
  catalog::DirectoryEntry dirent;
  PathString path;

  bool found = GetPathForInode(ino, &path);
  if (!found) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, ENOENT);
    return;
  }
  found = GetDirentForInode(ino, &dirent);
  if (!found) {
    fuse_remounter_->fence()->Leave();
    ReplyNegative(dirent, req);
    return;
  }

  if (!CheckVoms(*fuse_ctx)) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, EACCES);
    return;
  }

  mount_point_->tracer()->Trace(Tracer::kEventOpen, path, "open()");
  // Don't check.  Either done by the OS or one wants to purposefully work
  // around wrong open flags
  // if ((fi->flags & 3) != O_RDONLY) {
  //   fuse_reply_err(req, EROFS);
  //   return;
  // }
#ifdef __APPLE__
  if ((fi->flags & O_SHLOCK) || (fi->flags & O_EXLOCK)) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, EOPNOTSUPP);
    return;
  }
  if (fi->flags & O_SYMLINK) {
    fuse_remounter_->fence()->Leave();
    fi->fh = kFileHandleIgnore;
    fuse_reply_open(req, fi);
    return;
  }
#endif
  if (fi->flags & O_EXCL) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, EEXIST);
    return;
  }

  perf::Inc(file_system_->n_fs_open());  // Count actual open / fetch operations

  glue::PageCacheTracker::OpenDirectives open_directives;
  if (!dirent.IsChunkedFile()) {
    if (dirent.IsDirectIo()) {
      open_directives = mount_point_->page_cache_tracker()->OpenDirect();
    } else {
      open_directives =
        mount_point_->page_cache_tracker()->Open(
          ino, dirent.checksum(), dirent.GetStatStructure());
    }
    fuse_remounter_->fence()->Leave();
  } else {
    LogCvmfs(kLogCvmfs, kLogDebug,
             "chunked file %s opened (download delayed to read() call)",
             path.c_str());

    if (!IncAndCheckNoOpenFiles()) {
      perf::Dec(file_system_->no_open_files());
      fuse_remounter_->fence()->Leave();
      LogCvmfs(kLogCvmfs, kLogSyslogErr, "open file descriptor limit exceeded");
      fuse_reply_err(req, EMFILE);
      return;
    }

    // Figure out unique inode from annotated catalog
    // TODO(jblomer): we only need to lookup if the inode is not from the
    // current generation
    catalog::DirectoryEntry dirent_origin;
    if (!catalog_mgr->LookupPath(path, catalog::kLookupDefault,
                                 &dirent_origin)) {
      fuse_remounter_->fence()->Leave();
      LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
               "chunked file %s vanished unexpectedly", path.c_str());
      fuse_reply_err(req, ENOENT);
      return;
    }
    const uint64_t unique_inode = dirent_origin.inode();

    ChunkTables *chunk_tables = mount_point_->chunk_tables();
    chunk_tables->Lock();
    if (!chunk_tables->inode2chunks.Contains(unique_inode)) {
      chunk_tables->Unlock();

      // Retrieve File chunks from the catalog
      UniquePtr<FileChunkList> chunks(new FileChunkList());
      if (!catalog_mgr->ListFileChunks(path, dirent.hash_algorithm(),
                                       chunks.weak_ref()) ||
          chunks->IsEmpty())
      {
        fuse_remounter_->fence()->Leave();
        LogCvmfs(kLogCvmfs, kLogDebug| kLogSyslogErr,
           "EIO (04) file %s is marked as 'chunked', but no chunks found.",
           path.c_str());
        perf::Inc(file_system_->n_eio_total());
        perf::Inc(file_system_->n_eio_04());
        fuse_reply_err(req, EIO);
        return;
      }
      fuse_remounter_->fence()->Leave();

      chunk_tables->Lock();
      // Check again to avoid race
      if (!chunk_tables->inode2chunks.Contains(unique_inode)) {
        chunk_tables->inode2chunks.Insert(
          unique_inode, FileChunkReflist(chunks.Release(), path,
                                         dirent.compression_algorithm(),
                                         dirent.IsExternalFile()));
        chunk_tables->inode2references.Insert(unique_inode, 1);
      } else {
        uint32_t refctr;
        bool retval =
          chunk_tables->inode2references.Lookup(unique_inode, &refctr);
        assert(retval);
        chunk_tables->inode2references.Insert(unique_inode, refctr+1);
      }
    } else {
      fuse_remounter_->fence()->Leave();
      uint32_t refctr;
      bool retval =
        chunk_tables->inode2references.Lookup(unique_inode, &refctr);
      assert(retval);
      chunk_tables->inode2references.Insert(unique_inode, refctr+1);
    }

    // Update the chunk handle list
    LogCvmfs(kLogCvmfs, kLogDebug,
             "linking chunk handle %d to unique inode: %" PRIu64,
             chunk_tables->next_handle, uint64_t(unique_inode));
    chunk_tables->handle2fd.Insert(chunk_tables->next_handle, ChunkFd());
    chunk_tables->handle2uniqino.Insert(chunk_tables->next_handle,
                                        unique_inode);

    // Generate artificial content hash as hash over chunk hashes
    // TODO(jblomer): we may want to cache the result in the chunk tables
    FileChunkReflist chunk_reflist;
    bool retval =
        chunk_tables->inode2chunks.Lookup(unique_inode, &chunk_reflist);
    assert(retval);

    fi->fh = chunk_tables->next_handle;
    if (dirent.IsDirectIo()) {
      open_directives = mount_point_->page_cache_tracker()->OpenDirect();
    } else {
      open_directives = mount_point_->page_cache_tracker()->Open(
        ino, chunk_reflist.HashChunkList(), dirent.GetStatStructure());
    }
    FillOpenFlags(open_directives, fi);
    fi->fh = static_cast<uint64_t>(-static_cast<int64_t>(fi->fh));
    ++chunk_tables->next_handle;
    chunk_tables->Unlock();

    fuse_reply_open(req, fi);
    return;
  }

  Fetcher *this_fetcher = dirent.IsExternalFile()
    ? mount_point_->external_fetcher()
    : mount_point_->fetcher();
  CacheManager::Label label;
  label.path = path.ToString();
  label.size = dirent.size();
  label.zip_algorithm = dirent.compression_algorithm();
  if (mount_point_->catalog_mgr()->volatile_flag())
    label.flags |= CacheManager::kLabelVolatile;
  if (dirent.IsExternalFile())
    label.flags |= CacheManager::kLabelExternal;
  fd =
    this_fetcher->Fetch(CacheManager::LabeledObject(dirent.checksum(), label));

  if (fd >= 0) {
    if (IncAndCheckNoOpenFiles()) {
      LogCvmfs(kLogCvmfs, kLogDebug, "file %s opened (fd %d)",
               path.c_str(), fd);
      fi->fh = fd;
      FillOpenFlags(open_directives, fi);
      fuse_reply_open(req, fi);
      return;
    } else {
      if (file_system_->cache_mgr()->Close(fd) == 0)
        perf::Dec(file_system_->no_open_files());
      LogCvmfs(kLogCvmfs, kLogSyslogErr, "open file descriptor limit exceeded");
      fuse_reply_err(req, EMFILE);
      return;
    }
    assert(false);
  }

  // fd < 0
  LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
           "failed to open inode: %" PRIu64 ", CAS key %s, error code %d",
           uint64_t(ino), dirent.checksum().ToString().c_str(), errno);
  if (errno == EMFILE) {
    fuse_reply_err(req, EMFILE);
    return;
  }

  mount_point_->backoff_throttle()->Throttle();

  mount_point_->file_system()->io_error_info()->AddIoError();
  if (EIO == errno  || EIO == -fd) {
    LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
        "EIO (06) on %s", path.c_str() );
    perf::Inc(file_system_->n_eio_total());
    perf::Inc(file_system_->n_eio_06());
  }

  fuse_reply_err(req, -fd);
}


static void cvmfs_read(fuse_req_t req, fuse_ino_t ino, size_t size, off_t off,
                       struct fuse_file_info *fi)
{
  HighPrecisionTimer guard_timer(file_system_->hist_fs_read());

  LogCvmfs(kLogCvmfs, kLogDebug,
           "cvmfs_read inode: %" PRIu64 " reading %d bytes from offset %d "
           "fd %d", uint64_t(mount_point_->catalog_mgr()->MangleInode(ino)),
           size, off, fi->fh);
  perf::Inc(file_system_->n_fs_read());

#ifdef __APPLE__
  if (fi->fh == kFileHandleIgnore) {
    fuse_reply_err(req, EBADF);
    return;
  }
#endif

  // Get data chunk (<=128k guaranteed by Fuse)
  char *data = static_cast<char *>(alloca(size));
  unsigned int overall_bytes_fetched = 0;

  int64_t fd = static_cast<int64_t>(fi->fh);
  uint64_t abs_fd = (fd < 0) ? -fd : fd;
  ClearBit(glue::PageCacheTracker::kBitDirectIo, &abs_fd);

  // Do we have a a chunked file?
  if (fd < 0) {
    const struct fuse_ctx *fuse_ctx = fuse_req_ctx(req);
    FuseInterruptCue ic(&req);
    ClientCtxGuard ctx_guard(fuse_ctx->uid, fuse_ctx->gid, fuse_ctx->pid, &ic);

    const uint64_t chunk_handle = abs_fd;
    uint64_t unique_inode;
    ChunkFd chunk_fd;
    FileChunkReflist chunks;
    bool retval;

    // Fetch unique inode, chunk list and file descriptor
    ChunkTables *chunk_tables = mount_point_->chunk_tables();
    chunk_tables->Lock();
    retval = chunk_tables->handle2uniqino.Lookup(chunk_handle, &unique_inode);
    if (!retval) {
      LogCvmfs(kLogCvmfs, kLogDebug, "no unique inode, fall back to fuse ino");
      unique_inode = ino;
    }
    retval = chunk_tables->inode2chunks.Lookup(unique_inode, &chunks);
    assert(retval);
    chunk_tables->Unlock();

    unsigned chunk_idx = chunks.FindChunkIdx(off);

    // Lock chunk handle
    pthread_mutex_t *handle_lock = chunk_tables->Handle2Lock(chunk_handle);
    MutexLockGuard m(handle_lock);
    chunk_tables->Lock();
    retval = chunk_tables->handle2fd.Lookup(chunk_handle, &chunk_fd);
    assert(retval);
    chunk_tables->Unlock();

    // Fetch all needed chunks and read the requested data
    off_t offset_in_chunk = off - chunks.list->AtPtr(chunk_idx)->offset();
    do {
      // Open file descriptor to chunk
      if ((chunk_fd.fd == -1) || (chunk_fd.chunk_idx != chunk_idx)) {
        if (chunk_fd.fd != -1) file_system_->cache_mgr()->Close(chunk_fd.fd);
        Fetcher *this_fetcher = chunks.external_data
          ? mount_point_->external_fetcher()
          : mount_point_->fetcher();
        CacheManager::Label label;
        label.path = chunks.path.ToString();
        label.size = chunks.list->AtPtr(chunk_idx)->size();
        label.zip_algorithm = chunks.compression_alg;
        label.flags |= CacheManager::kLabelChunked;
        if (mount_point_->catalog_mgr()->volatile_flag())
          label.flags |= CacheManager::kLabelVolatile;
        if (chunks.external_data) {
          label.flags |= CacheManager::kLabelExternal;
          label.range_offset = chunks.list->AtPtr(chunk_idx)->offset();
        }
        chunk_fd.fd = this_fetcher->Fetch(CacheManager::LabeledObject(
          chunks.list->AtPtr(chunk_idx)->content_hash(), label));
        if (chunk_fd.fd < 0) {
          chunk_fd.fd = -1;
          chunk_tables->Lock();
          chunk_tables->handle2fd.Insert(chunk_handle, chunk_fd);
          chunk_tables->Unlock();

          LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
              "EIO (05) on %s", chunks.path.ToString().c_str() );
          perf::Inc(file_system_->n_eio_total());
          perf::Inc(file_system_->n_eio_05());
          fuse_reply_err(req, EIO);
          return;
        }
        chunk_fd.chunk_idx = chunk_idx;
      }

      LogCvmfs(kLogCvmfs, kLogDebug, "reading from chunk fd %d",
               chunk_fd.fd);
      // Read data from chunk
      const size_t bytes_to_read = size - overall_bytes_fetched;
      const size_t remaining_bytes_in_chunk =
        chunks.list->AtPtr(chunk_idx)->size() - offset_in_chunk;
      size_t bytes_to_read_in_chunk =
        std::min(bytes_to_read, remaining_bytes_in_chunk);
      const int64_t bytes_fetched = file_system_->cache_mgr()->Pread(
        chunk_fd.fd,
        data + overall_bytes_fetched,
        bytes_to_read_in_chunk,
        offset_in_chunk);

      if (bytes_fetched < 0) {
        LogCvmfs(kLogCvmfs, kLogSyslogErr, "read err no %" PRId64 " (%s)",
                 bytes_fetched, chunks.path.ToString().c_str());
        chunk_tables->Lock();
        chunk_tables->handle2fd.Insert(chunk_handle, chunk_fd);
        chunk_tables->Unlock();
        if ( EIO == errno || EIO == -bytes_fetched ) {
          LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
             "EIO (07) on %s", chunks.path.ToString().c_str() );
          perf::Inc(file_system_->n_eio_total());
          perf::Inc(file_system_->n_eio_07());
        }
        fuse_reply_err(req, -bytes_fetched);
        return;
      }
      overall_bytes_fetched += bytes_fetched;

      // Proceed to the next chunk to keep on reading data
      ++chunk_idx;
      offset_in_chunk = 0;
    } while ((overall_bytes_fetched < size) &&
             (chunk_idx < chunks.list->size()));

    // Update chunk file descriptor
    chunk_tables->Lock();
    chunk_tables->handle2fd.Insert(chunk_handle, chunk_fd);
    chunk_tables->Unlock();
    LogCvmfs(kLogCvmfs, kLogDebug, "released chunk file descriptor %d",
             chunk_fd.fd);
  } else {
    int64_t nbytes = file_system_->cache_mgr()->Pread(abs_fd, data, size, off);
    if (nbytes < 0) {
      if ( EIO == errno || EIO == -nbytes ) {
        PathString path;
        bool found = GetPathForInode(ino, &path);
        if ( found ) {
          LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
             "EIO (08) on %s", path.ToString().c_str() );
        } else {
          LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
             "EIO (08) on <unknown inode>");
        }
        perf::Inc(file_system_->n_eio_total());
        perf::Inc(file_system_->n_eio_08());
      }
      fuse_reply_err(req, -nbytes);
      return;
    }
    overall_bytes_fetched = nbytes;
  }

  // Push it to user
  fuse_reply_buf(req, data, overall_bytes_fetched);
  LogCvmfs(kLogCvmfs, kLogDebug, "pushed %d bytes to user",
           overall_bytes_fetched);
}


static void cvmfs_release(fuse_req_t req, fuse_ino_t ino,
                          struct fuse_file_info *fi)
{
  HighPrecisionTimer guard_timer(file_system_->hist_fs_release());

  ino = mount_point_->catalog_mgr()->MangleInode(ino);
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_release on inode: %" PRIu64,
           uint64_t(ino));

#ifdef __APPLE__
  if (fi->fh == kFileHandleIgnore) {
    fuse_reply_err(req, 0);
    return;
  }
#endif

  int64_t fd = static_cast<int64_t>(fi->fh);
  uint64_t abs_fd = (fd < 0) ? -fd : fd;
  if (!TestBit(glue::PageCacheTracker::kBitDirectIo, abs_fd)) {
    mount_point_->page_cache_tracker()->Close(ino);
  }
  ClearBit(glue::PageCacheTracker::kBitDirectIo, &abs_fd);

  // do we have a chunked file?
  if (fd < 0) {
    const uint64_t chunk_handle = abs_fd;
    LogCvmfs(kLogCvmfs, kLogDebug, "releasing chunk handle %" PRIu64,
             chunk_handle);
    uint64_t unique_inode;
    ChunkFd chunk_fd;
    FileChunkReflist chunks;
    uint32_t refctr;
    bool retval;

    ChunkTables *chunk_tables = mount_point_->chunk_tables();
    chunk_tables->Lock();
    retval = chunk_tables->handle2uniqino.Lookup(chunk_handle, &unique_inode);
    if (!retval) {
      LogCvmfs(kLogCvmfs, kLogDebug, "no unique inode, fall back to fuse ino");
      unique_inode = ino;
    } else {
      chunk_tables->handle2uniqino.Erase(chunk_handle);
    }
    retval = chunk_tables->handle2fd.Lookup(chunk_handle, &chunk_fd);
    assert(retval);
    chunk_tables->handle2fd.Erase(chunk_handle);

    retval = chunk_tables->inode2references.Lookup(unique_inode, &refctr);
    assert(retval);
    refctr--;
    if (refctr == 0) {
      LogCvmfs(kLogCvmfs, kLogDebug, "releasing chunk list for inode %" PRIu64,
               uint64_t(unique_inode));
      FileChunkReflist to_delete;
      retval = chunk_tables->inode2chunks.Lookup(unique_inode, &to_delete);
      assert(retval);
      chunk_tables->inode2references.Erase(unique_inode);
      chunk_tables->inode2chunks.Erase(unique_inode);
      delete to_delete.list;
    } else {
      chunk_tables->inode2references.Insert(unique_inode, refctr);
    }
    chunk_tables->Unlock();

    if (chunk_fd.fd != -1)
      file_system_->cache_mgr()->Close(chunk_fd.fd);
    perf::Dec(file_system_->no_open_files());
  } else {
    if (file_system_->cache_mgr()->Close(abs_fd) == 0) {
      perf::Dec(file_system_->no_open_files());
    }
  }
  fuse_reply_err(req, 0);
}

static void cvmfs_statfs(fuse_req_t req, fuse_ino_t ino) {
  ino = mount_point_->catalog_mgr()->MangleInode(ino);
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_statfs on inode: %" PRIu64,
           uint64_t(ino));

  TraceInode(Tracer::kEventStatFs, ino, "statfs()");

  perf::Inc(file_system_->n_fs_statfs());

  // Unmanaged cache (no lock needed - statfs is never modified)
  if (!file_system_->cache_mgr()->quota_mgr()->HasCapability(
       QuotaManager::kCapIntrospectSize))
  {
    LogCvmfs(kLogCvmfs, kLogDebug, "QuotaManager does not support statfs");
    fuse_reply_statfs(req, (mount_point_->statfs_cache()->info()));
    return;
  }

  MutexLockGuard m(mount_point_->statfs_cache()->lock());

  const uint64_t deadline = *mount_point_->statfs_cache()->expiry_deadline();
  struct statvfs *info = mount_point_->statfs_cache()->info();

  // cached version still valid
  if ( platform_monotonic_time() < deadline ) {
    perf::Inc(file_system_->n_fs_statfs_cached());
    fuse_reply_statfs(req, info);
    return;
  }

  uint64_t available = 0;
  uint64_t size = file_system_->cache_mgr()->quota_mgr()->GetSize();
  uint64_t capacity = file_system_->cache_mgr()->quota_mgr()->GetCapacity();
  // Fuse/OS X doesn't like values < 512
  info->f_bsize = info->f_frsize = 512;

  if (capacity == (uint64_t)(-1)) {
    // Unknown capacity, set capacity = size
    info->f_blocks = size / info->f_bsize;
  } else {
    // Take values from LRU module
    info->f_blocks = capacity / info->f_bsize;
    available = capacity - size;
  }

  info->f_bfree = info->f_bavail = available / info->f_bsize;

  // Inodes / entries
  fuse_remounter_->fence()->Enter();
  uint64_t all_inodes = mount_point_->catalog_mgr()->all_inodes();
  uint64_t loaded_inode = mount_point_->catalog_mgr()->loaded_inodes();
  info->f_files = all_inodes;
  info->f_ffree = info->f_favail = all_inodes - loaded_inode;
  fuse_remounter_->fence()->Leave();

  *mount_point_->statfs_cache()->expiry_deadline() =
      platform_monotonic_time()
    + mount_point_->statfs_cache()->cache_timeout();

  fuse_reply_statfs(req, info);
}

#ifdef __APPLE__
static void cvmfs_getxattr(fuse_req_t req, fuse_ino_t ino, const char *name,
                           size_t size, uint32_t position)
#else
static void cvmfs_getxattr(fuse_req_t req, fuse_ino_t ino, const char *name,
                           size_t size)
#endif
{
  const struct fuse_ctx *fuse_ctx = fuse_req_ctx(req);
  FuseInterruptCue ic(&req);
  ClientCtxGuard ctx_guard(fuse_ctx->uid, fuse_ctx->gid, fuse_ctx->pid, &ic);

  fuse_remounter_->fence()->Enter();
  catalog::ClientCatalogManager *catalog_mgr = mount_point_->catalog_mgr();
  ino = catalog_mgr->MangleInode(ino);
  LogCvmfs(kLogCvmfs, kLogDebug,
           "cvmfs_getxattr on inode: %" PRIu64 " for xattr: %s",
           uint64_t(ino), name);
  if (!CheckVoms(*fuse_ctx)) {
    fuse_remounter_->fence()->Leave();
    fuse_reply_err(req, EACCES);
    return;
  }
  TraceInode(Tracer::kEventGetXAttr, ino, "getxattr()");

  const string attr = name;
  catalog::DirectoryEntry d;
  const bool found = GetDirentForInode(ino, &d);
  bool retval;
  XattrList xattrs;

  PathString path;
  retval = GetPathForInode(ino, &path);
  assert(retval);
  if (d.IsLink()) {
    catalog::LookupOptions lookup_options = static_cast<catalog::LookupOptions>(
      catalog::kLookupDefault | catalog::kLookupRawSymlink);
    catalog::DirectoryEntry raw_symlink;
    retval = catalog_mgr->LookupPath(path, lookup_options, &raw_symlink);
    assert(retval);
    d.set_symlink(raw_symlink.symlink());
  }
  if (d.HasXattrs()) {
    retval = catalog_mgr->LookupXattrs(path, &xattrs);

    if (!AssertOrLog(retval, kLogCvmfs, kLogSyslogWarn | kLogDebug,
                     "cvmfs_statfs: Race condition? "
                     "LookupXattrs did not succeed for path %s",
                     path.c_str())) {
      fuse_remounter_->fence()->Leave();
      fuse_reply_err(req, ESTALE);
    }
  }

  bool magic_xattr_success = true;
  MagicXattrRAIIWrapper magic_xattr(mount_point_->magic_xattr_mgr()->GetLocked(
    attr, path, &d));
  if (!magic_xattr.IsNull()) {
    magic_xattr_success = magic_xattr->
                              PrepareValueFencedProtected(fuse_ctx->gid);
  }

  fuse_remounter_->fence()->Leave();

  if (!found) {
    ReplyNegative(d, req);
    return;
  }

  if (!magic_xattr_success) {
    fuse_reply_err(req, ENOATTR);
    return;
  }

  string attribute_value;

  if (!magic_xattr.IsNull()) {
    attribute_value = magic_xattr->GetValue();
  } else {
    if (!xattrs.Get(attr, &attribute_value)) {
      fuse_reply_err(req, ENOATTR);
      return;
    }
  }

  if (size == 0) {
    fuse_reply_xattr(req, attribute_value.length());
  } else if (size >= attribute_value.length()) {
    fuse_reply_buf(req, &attribute_value[0], attribute_value.length());
  } else {
    fuse_reply_err(req, ERANGE);
  }
}


static void cvmfs_listxattr(fuse_req_t req, fuse_ino_t ino, size_t size) {
  const struct fuse_ctx *fuse_ctx = fuse_req_ctx(req);
  FuseInterruptCue ic(&req);
  ClientCtxGuard ctx_guard(fuse_ctx->uid, fuse_ctx->gid, fuse_ctx->pid, &ic);

  fuse_remounter_->fence()->Enter();
  catalog::ClientCatalogManager *catalog_mgr = mount_point_->catalog_mgr();
  ino = catalog_mgr->MangleInode(ino);
  TraceInode(Tracer::kEventListAttr, ino, "listxattr()");
  LogCvmfs(kLogCvmfs, kLogDebug,
           "cvmfs_listxattr on inode: %" PRIu64 ", size %u [visibility %d]",
           uint64_t(ino), size,
           mount_point_->magic_xattr_mgr()->visibility());

  catalog::DirectoryEntry d;
  const bool found = GetDirentForInode(ino, &d);
  XattrList xattrs;
  if (d.HasXattrs()) {
    PathString path;
    bool retval = GetPathForInode(ino, &path);

    if (!AssertOrLog(retval, kLogCvmfs, kLogSyslogWarn | kLogDebug,
                     "cvmfs_listxattr: Race condition? "
                     "GetPathForInode did not succeed for ino %lu",
                     ino)) {
      fuse_remounter_->fence()->Leave();
      fuse_reply_err(req, ESTALE);
      return;
    }

    retval = catalog_mgr->LookupXattrs(path, &xattrs);
    if (!AssertOrLog(retval, kLogCvmfs, kLogSyslogWarn | kLogDebug,
                     "cvmfs_listxattr: Race condition? "
                     "LookupXattrs did not succeed for ino %lu",
                     ino)) {
      fuse_remounter_->fence()->Leave();
      fuse_reply_err(req, ESTALE);
      return;
    }
  }
  fuse_remounter_->fence()->Leave();

  if (!found) {
    ReplyNegative(d, req);
    return;
  }

  string attribute_list;
  attribute_list = mount_point_->magic_xattr_mgr()->GetListString(&d);
  attribute_list += xattrs.ListKeysPosix(attribute_list);

  if (size == 0) {
    fuse_reply_xattr(req, attribute_list.length());
  } else if (size >= attribute_list.length()) {
    if (attribute_list.empty())
      fuse_reply_buf(req, NULL, 0);
    else
      fuse_reply_buf(req, &attribute_list[0], attribute_list.length());
  } else {
    fuse_reply_err(req, ERANGE);
  }
}

bool Evict(const string &path) {
  catalog::DirectoryEntry dirent;
  fuse_remounter_->fence()->Enter();
  const bool found = (GetDirentForPath(PathString(path), &dirent) > 0);

  if (!found || !dirent.IsRegular()) {
    fuse_remounter_->fence()->Leave();
    return false;
  }

  if (!dirent.IsChunkedFile()) {
    fuse_remounter_->fence()->Leave();
  } else {
    FileChunkList chunks;
    mount_point_->catalog_mgr()->ListFileChunks(
      PathString(path), dirent.hash_algorithm(), &chunks);
    fuse_remounter_->fence()->Leave();
    for (unsigned i = 0; i < chunks.size(); ++i) {
        file_system_->cache_mgr()->quota_mgr()
           ->Remove(chunks.AtPtr(i)->content_hash());
    }
  }
  file_system_->cache_mgr()->quota_mgr()->Remove(dirent.checksum());
  return true;
}


bool Pin(const string &path) {
  catalog::DirectoryEntry dirent;
  fuse_remounter_->fence()->Enter();
  const bool found = (GetDirentForPath(PathString(path), &dirent) > 0);
  if (!found || !dirent.IsRegular()) {
    fuse_remounter_->fence()->Leave();
    return false;
  }

  Fetcher *this_fetcher = dirent.IsExternalFile()
    ? mount_point_->external_fetcher()
    : mount_point_->fetcher();

  if (!dirent.IsChunkedFile()) {
    fuse_remounter_->fence()->Leave();
  } else {
    FileChunkList chunks;
    mount_point_->catalog_mgr()->ListFileChunks(
      PathString(path), dirent.hash_algorithm(), &chunks);
    fuse_remounter_->fence()->Leave();
    for (unsigned i = 0; i < chunks.size(); ++i) {
      bool retval =
        file_system_->cache_mgr()->quota_mgr()->Pin(
          chunks.AtPtr(i)->content_hash(),
          chunks.AtPtr(i)->size(),
          "Part of " + path,
          false);
      if (!retval)
        return false;
      int fd = -1;
      CacheManager::Label label;
      label.path = path;
      label.size = chunks.AtPtr(i)->size();
      label.zip_algorithm = dirent.compression_algorithm();
      label.flags |= CacheManager::kLabelPinned;
      label.flags |= CacheManager::kLabelChunked;
      if (dirent.IsExternalFile()) {
        label.flags |= CacheManager::kLabelExternal;
        label.range_offset = chunks.AtPtr(i)->offset();
      }
      fd = this_fetcher->Fetch(CacheManager::LabeledObject(
        chunks.AtPtr(i)->content_hash(), label));
      if (fd < 0) {
        return false;
      }
      file_system_->cache_mgr()->Close(fd);
    }
    return true;
  }

  bool retval = file_system_->cache_mgr()->quota_mgr()->Pin(
    dirent.checksum(), dirent.size(), path, false);
  if (!retval)
    return false;
  CacheManager::Label label;
  label.flags = CacheManager::kLabelPinned;
  label.size = dirent.size();
  label.path = path;
  label.zip_algorithm = dirent.compression_algorithm();
  int fd = this_fetcher->Fetch(
    CacheManager::LabeledObject(dirent.checksum(), label));
  if (fd < 0) {
    return false;
  }
  file_system_->cache_mgr()->Close(fd);
  return true;
}


static void cvmfs_init(void *userdata, struct fuse_conn_info *conn) {
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_init");

  // NFS support
#ifdef CVMFS_NFS_SUPPORT
  conn->want |= FUSE_CAP_EXPORT_SUPPORT;
#endif

  if (mount_point_->enforce_acls()) {
#ifdef FUSE_CAP_POSIX_ACL
    if ((conn->capable & FUSE_CAP_POSIX_ACL) == 0) {
      PANIC(kLogDebug | kLogSyslogErr,
            "FUSE: ACL support requested but missing fuse kernel support, "
            "aborting");
    }
    conn->want |= FUSE_CAP_POSIX_ACL;
    LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslog, "enforcing ACLs");
#else
    PANIC(kLogDebug | kLogSyslogErr,
          "FUSE: ACL support requested but not available in this version of "
          "libfuse %d, aborting", FUSE_VERSION);
#endif
  }

  if ( mount_point_->cache_symlinks() ) {
#ifdef FUSE_CAP_CACHE_SYMLINKS
    if ((conn->capable & FUSE_CAP_CACHE_SYMLINKS) == FUSE_CAP_CACHE_SYMLINKS) {
      conn->want |= FUSE_CAP_CACHE_SYMLINKS;
      LogCvmfs(kLogCvmfs, kLogDebug, "FUSE: Enable symlink caching");
      #ifndef FUSE_CAP_EXPIRE_ONLY
        LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogWarn,
          "FUSE: Symlink caching enabled but no support for fuse_expire_entry. "
          "Symlinks will be cached but mountpoints on top of symlinks will "
          "break! "
          "Current libfuse %d is too old; required: libfuse >= 3.16, "
          "kernel >= 6.2-rc1",
          FUSE_VERSION);
      #endif
    } else {
      mount_point_->DisableCacheSymlinks();
      LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogWarn,
           "FUSE: Symlink caching requested but missing fuse kernel support, "
           "falling back to no caching");
    }
#else
    mount_point_->DisableCacheSymlinks();
    LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogWarn,
          "FUSE: Symlink caching requested but missing libfuse support, "
          "falling back to no caching. Current libfuse %d",
          FUSE_VERSION);
#endif
  }

#ifdef FUSE_CAP_EXPIRE_ONLY
  if ((conn->capable & FUSE_CAP_EXPIRE_ONLY) == FUSE_CAP_EXPIRE_ONLY &&
       FUSE_VERSION >= FUSE_MAKE_VERSION(3, 16)) {
    mount_point_->EnableFuseExpireEntry();
    LogCvmfs(kLogCvmfs, kLogDebug, "FUSE: Enable fuse_expire_entry ");
  } else if (mount_point_->cache_symlinks()) {
    LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogWarn,
      "FUSE: Symlink caching enabled but no support for fuse_expire_entry. "
      "Symlinks will be cached but mountpoints on top of symlinks will break! "
      "Current libfuse %d; required: libfuse >= 3.16, kernel >= 6.2-rc1",
      FUSE_VERSION);
  }
#endif
}

static void cvmfs_destroy(void *unused __attribute__((unused))) {
  // The debug log is already closed at this point
  LogCvmfs(kLogCvmfs, kLogDebug, "cvmfs_destroy");
}

static void SetCvmfsOperations(struct fuse_lowlevel_ops *cvmfs_operations) {
  memset(cvmfs_operations, 0, sizeof(*cvmfs_operations));

  // Init/Fini
  cvmfs_operations->init         = cvmfs_init;
  cvmfs_operations->destroy      = cvmfs_destroy;

  cvmfs_operations->lookup       = cvmfs_lookup;
  cvmfs_operations->getattr      = cvmfs_getattr;
  cvmfs_operations->readlink     = cvmfs_readlink;
  cvmfs_operations->open         = cvmfs_open;
  cvmfs_operations->read         = cvmfs_read;
  cvmfs_operations->release      = cvmfs_release;
  cvmfs_operations->opendir      = cvmfs_opendir;
  cvmfs_operations->readdir      = cvmfs_readdir;
  cvmfs_operations->releasedir   = cvmfs_releasedir;
  cvmfs_operations->statfs       = cvmfs_statfs;
  cvmfs_operations->getxattr     = cvmfs_getxattr;
  cvmfs_operations->listxattr    = cvmfs_listxattr;
  cvmfs_operations->forget       = cvmfs_forget;
#if (FUSE_VERSION >= 29)
  cvmfs_operations->forget_multi = cvmfs_forget_multi;
#endif
}

// Called by cvmfs_talk when switching into read-only cache mode
void UnregisterQuotaListener() {
  if (cvmfs::unpin_listener_) {
    quota::UnregisterListener(cvmfs::unpin_listener_);
    cvmfs::unpin_listener_ = NULL;
  }
  if (cvmfs::watchdog_listener_) {
    quota::UnregisterListener(cvmfs::watchdog_listener_);
    cvmfs::watchdog_listener_ = NULL;
  }
}

bool SendFuseFd(const std::string &socket_path) {
  int fuse_fd;
#if (FUSE_VERSION >= 30)
  fuse_fd = fuse_session_fd(*reinterpret_cast<struct fuse_session**>(
    loader_exports_->fuse_channel_or_session));
#else
  fuse_fd = fuse_chan_fd(*reinterpret_cast<struct fuse_chan**>(
    loader_exports_->fuse_channel_or_session));
#endif
  assert(fuse_fd >= 0);
  int sock_fd = ConnectSocket(socket_path);
  if (sock_fd < 0) {
    LogCvmfs(kLogCvmfs, kLogDebug, "cannot connect to socket %s: %d",
             socket_path.c_str(), errno);
    return false;
  }
  bool retval = SendFd2Socket(sock_fd, fuse_fd);
  close(sock_fd);
  return retval;
}

}  // namespace cvmfs


string *g_boot_error = NULL;

__attribute__((visibility("default")))
loader::CvmfsExports *g_cvmfs_exports = NULL;

class ExpiresMagicXattr : public BaseMagicXattr {
  time_t catalogs_valid_until_;

  virtual bool PrepareValueFenced() {
    catalogs_valid_until_ = cvmfs::fuse_remounter_->catalogs_valid_until();
    return true;
  }

  virtual std::string GetValue() {
    if (catalogs_valid_until_ == MountPoint::kIndefiniteDeadline) {
      return "never (fixed root catalog)";
    } else {
      time_t now = time(NULL);
      return StringifyInt( (catalogs_valid_until_ - now) / 60);
    }
  }
};

class InodeMaxMagicXattr : public BaseMagicXattr {
  virtual std::string GetValue() {
    return StringifyInt(
      cvmfs::inode_generation_info_.inode_generation +
      xattr_mgr_->mount_point()->catalog_mgr()->inode_gauge());
  }
};

class MaxFdMagicXattr : public BaseMagicXattr {
  virtual std::string GetValue() {
    return StringifyInt(cvmfs::max_open_files_ - cvmfs::kNumReservedFd);
  }
};

class PidMagicXattr : public BaseMagicXattr {
  virtual std::string GetValue() { return StringifyInt(cvmfs::pid_); }
};

class UptimeMagicXattr : public BaseMagicXattr {
  virtual std::string GetValue() {
    time_t now = time(NULL);
    uint64_t uptime = now - cvmfs::loader_exports_->boot_time;
    return StringifyInt(uptime / 60);
  }
};

static void RegisterMagicXattrs() {
  MagicXattrManager *mgr = cvmfs::mount_point_->magic_xattr_mgr();
  mgr->Register("user.expires", new ExpiresMagicXattr());
  mgr->Register("user.inode_max", new InodeMaxMagicXattr());
  mgr->Register("user.pid", new PidMagicXattr());
  mgr->Register("user.maxfd", new MaxFdMagicXattr());
  mgr->Register("user.uptime", new UptimeMagicXattr());

  mgr->Freeze();
}

static FileSystem *InitSystemFs(
  const string &mount_path,
  const string &fqrn,
  FileSystem::FileSystemInfo fs_info)
{
  fs_info.wait_workspace = false;
  FileSystem *file_system = FileSystem::Create(fs_info);

  if (file_system->boot_status() == loader::kFailLockWorkspace) {
    string fqrn_from_xattr;
    int retval = platform_getxattr(mount_path, "user.fqrn", &fqrn_from_xattr);
    if (!retval) {
      // Cvmfs not mounted anymore, but another cvmfs process is still in
      // shutdown procedure.  Try again and wait for lock
      delete file_system;
      fs_info.wait_workspace = true;
      file_system = FileSystem::Create(fs_info);
    } else {
      if (fqrn_from_xattr == fqrn) {
        LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogWarn,
                 "repository already mounted on %s", mount_path.c_str());
        file_system->set_boot_status(loader::kFailDoubleMount);
      } else {
        LogCvmfs(kLogCvmfs, kLogDebug | kLogSyslogErr,
                 "CernVM-FS repository %s already mounted on %s",
                 fqrn.c_str(), mount_path.c_str());
        file_system->set_boot_status(loader::kFailOtherMount);
      }
    }
  }

  return file_system;
}


static void InitOptionsMgr(const loader::LoaderExports *loader_exports) {
  if (loader_exports->version >= 3 && loader_exports->simple_options_parsing) {
    cvmfs::options_mgr_ = new SimpleOptionsParser(
      new DefaultOptionsTemplateManager(loader_exports->repository_name));
  } else {
    cvmfs::options_mgr_ = new BashOptionsManager(
      new DefaultOptionsTemplateManager(loader_exports->repository_name));
  }

  if (loader_exports->config_files != "") {
    vector<string> tokens = SplitString(loader_exports->config_files, ':');
    for (unsigned i = 0, s = tokens.size(); i < s; ++i) {
      cvmfs::options_mgr_->ParsePath(tokens[i], false);
    }
  } else {
    cvmfs::options_mgr_->ParseDefault(loader_exports->repository_name);
  }
}


static unsigned CheckMaxOpenFiles() {
  static unsigned max_open_files;
  static bool     already_done = false;

  // check number of open files (lazy evaluation)
  if (!already_done) {
    unsigned soft_limit = 0;
    unsigned hard_limit = 0;
    GetLimitNoFile(&soft_limit, &hard_limit);

    if (soft_limit < cvmfs::kMinOpenFiles) {
      LogCvmfs(kLogCvmfs, kLogSyslogWarn | kLogDebug,
               "Warning: current limits for number of open files are "
               "(%lu/%lu)\n"
               "CernVM-FS is likely to run out of file descriptors, "
               "set ulimit -n to at least %lu",
               soft_limit, hard_limit, cvmfs::kMinOpenFiles);
    }
    max_open_files = soft_limit;
    already_done   = true;
  }

  return max_open_files;
}


static int Init(const loader::LoaderExports *loader_exports) {
  g_boot_error = new string("unknown error");
  cvmfs::loader_exports_ = loader_exports;

  crypto::SetupLibcryptoMt();

  InitOptionsMgr(loader_exports);

  // We need logging set up before forking the watchdog
  FileSystem::SetupLoggingStandalone(
    *cvmfs::options_mgr_, loader_exports->repository_name);

  // Monitor, check for maximum number of open files
  if (cvmfs::UseWatchdog()) {
    auto_umount::SetMountpoint(loader_exports->mount_point);
    cvmfs::watchdog_ = Watchdog::Create(auto_umount::UmountOnCrash);
    if (cvmfs::watchdog_ == NULL) {
      *g_boot_error = "failed to initialize watchdog.";
      return loader::kFailMonitor;
    }
  }
  cvmfs::max_open_files_ = CheckMaxOpenFiles();

  FileSystem::FileSystemInfo fs_info;
  fs_info.type = FileSystem::kFsFuse;
  fs_info.name = loader_exports->repository_name;
  fs_info.exe_path = loader_exports->program_name;
  fs_info.options_mgr = cvmfs::options_mgr_;
  fs_info.foreground = loader_exports->foreground;
  cvmfs::file_system_ = InitSystemFs(
    loader_exports->mount_point,
    loader_exports->repository_name,
    fs_info);
  if (!cvmfs::file_system_->IsValid()) {
    *g_boot_error = cvmfs::file_system_->boot_error();
    return cvmfs::file_system_->boot_status();
  }
  if ((cvmfs::file_system_->cache_mgr()->id() == kPosixCacheManager) &&
      dynamic_cast<PosixCacheManager *>(
        cvmfs::file_system_->cache_mgr())->do_refcount())
  {
    cvmfs::check_fd_overflow_ = false;
  }

  cvmfs::mount_point_ = MountPoint::Create(loader_exports->repository_name,
                                           cvmfs::file_system_);
  if (!cvmfs::mount_point_->IsValid()) {
    *g_boot_error = cvmfs::mount_point_->boot_error();
    return cvmfs::mount_point_->boot_status();
  }

  RegisterMagicXattrs();

  cvmfs::directory_handles_ = new cvmfs::DirectoryHandles();
  cvmfs::directory_handles_->set_empty_key((uint64_t)(-1));
  cvmfs::directory_handles_->set_deleted_key((uint64_t)(-2));

  LogCvmfs(kLogCvmfs, kLogDebug, "fuse inode size is %d bits",
           sizeof(fuse_ino_t) * 8);

  cvmfs::inode_generation_info_.initial_revision =
    cvmfs::mount_point_->catalog_mgr()->GetRevision();
  cvmfs::inode_generation_info_.inode_generation =
    cvmfs::mount_point_->inode_annotation()->GetGeneration();
  LogCvmfs(kLogCvmfs, kLogDebug, "root inode is %" PRIu64,
           uint64_t(cvmfs::mount_point_->catalog_mgr()->GetRootInode()));

  void **channel_or_session = NULL;
  if (loader_exports->version >= 4) {
    channel_or_session = loader_exports->fuse_channel_or_session;
  }

  bool fuse_notify_invalidation = true;
  std::string buf;
  if (cvmfs::options_mgr_->GetValue("CVMFS_FUSE_NOTIFY_INVALIDATION",
                                    &buf)) {
    if (!cvmfs::options_mgr_->IsOn(buf)) {
      fuse_notify_invalidation = false;
      cvmfs::mount_point_->dentry_tracker()->Disable();
    }
  }
  cvmfs::fuse_remounter_ =
      new FuseRemounter(cvmfs::mount_point_, &cvmfs::inode_generation_info_,
                        channel_or_session, fuse_notify_invalidation);

  // Control & command interface
  cvmfs::talk_mgr_ = TalkManager::Create(
    cvmfs::mount_point_->talk_socket_path(),
    cvmfs::mount_point_,
    cvmfs::fuse_remounter_);
  if ((cvmfs::mount_point_->talk_socket_uid() != 0) ||
      (cvmfs::mount_point_->talk_socket_gid() != 0))
  {
    uid_t tgt_uid = cvmfs::mount_point_->talk_socket_uid();
    gid_t tgt_gid = cvmfs::mount_point_->talk_socket_gid();
    int rvi = chown(cvmfs::mount_point_->talk_socket_path().c_str(),
                    tgt_uid, tgt_gid);
    if (rvi != 0) {
      *g_boot_error = std::string("failed to set talk socket ownership - ")
        + "target " + StringifyInt(tgt_uid) + ":" + StringifyInt(tgt_uid)
        + ", user " + StringifyInt(geteuid()) + ":" + StringifyInt(getegid());
      return loader::kFailTalk;
    }
  }
  if (cvmfs::talk_mgr_ == NULL) {
    *g_boot_error = "failed to initialize talk socket (" +
                    StringifyInt(errno) + ")";
    return loader::kFailTalk;
  }

  // Notification system client
  {
    OptionsManager* options = cvmfs::file_system_->options_mgr();
    if (options->IsDefined("CVMFS_NOTIFICATION_SERVER")) {
      std::string config;
      options->GetValue("CVMFS_NOTIFICATION_SERVER", &config);
      const std::string repo_name = cvmfs::mount_point_->fqrn();
      cvmfs::notification_client_ =
          new NotificationClient(config, repo_name, cvmfs::fuse_remounter_,
                                 cvmfs::mount_point_->download_mgr(),
                                 cvmfs::mount_point_->signature_mgr());
    }
  }

  return loader::kFailOk;
}


static void Spawn() {
  // First thing: kick off the watchdog while we still have a single-threaded
  // well-defined state
  cvmfs::pid_ = getpid();
  if (cvmfs::watchdog_) {
    cvmfs::watchdog_->Spawn(GetCurrentWorkingDirectory() + "/stacktrace." +
                            cvmfs::mount_point_->fqrn());
  }

  cvmfs::fuse_remounter_->Spawn();
  if (cvmfs::mount_point_->dentry_tracker()->is_active()) {
    cvmfs::mount_point_->dentry_tracker()->SpawnCleaner(
        // Usually every minute
        static_cast<unsigned int>(cvmfs::mount_point_->kcache_timeout_sec()));
  }

  cvmfs::mount_point_->download_mgr()->Spawn();
  cvmfs::mount_point_->external_download_mgr()->Spawn();
  if (cvmfs::mount_point_->resolv_conf_watcher() != NULL) {
    cvmfs::mount_point_->resolv_conf_watcher()->Spawn();
  }
  QuotaManager *quota_mgr = cvmfs::file_system_->cache_mgr()->quota_mgr();
  quota_mgr->Spawn();
  if (quota_mgr->HasCapability(QuotaManager::kCapListeners)) {
    cvmfs::watchdog_listener_ = quota::RegisterWatchdogListener(
      quota_mgr,
      cvmfs::mount_point_->uuid()->uuid() + "-watchdog");
    cvmfs::unpin_listener_ = quota::RegisterUnpinListener(
      quota_mgr,
      cvmfs::mount_point_->catalog_mgr(),
      cvmfs::mount_point_->uuid()->uuid() + "-unpin");
  }
  cvmfs::mount_point_->tracer()->Spawn();
  cvmfs::talk_mgr_->Spawn();

  if (cvmfs::notification_client_ != NULL) {
    cvmfs::notification_client_->Spawn();
  }

  if (cvmfs::file_system_->nfs_maps() != NULL) {
    cvmfs::file_system_->nfs_maps()->Spawn();
  }

  cvmfs::file_system_->cache_mgr()->Spawn();

  if (cvmfs::mount_point_->telemetry_aggr() != NULL) {
    cvmfs::mount_point_->telemetry_aggr()->Spawn();
  }
}


static string GetErrorMsg() {
  if (g_boot_error)
    return *g_boot_error;
  return "";
}


static void ShutdownMountpoint() {
  delete cvmfs::talk_mgr_;
  cvmfs::talk_mgr_ = NULL;

  delete cvmfs::notification_client_;
  cvmfs::notification_client_ = NULL;

  // The remonter has a reference to the mount point and the inode generation
  delete cvmfs::fuse_remounter_;
  cvmfs::fuse_remounter_ = NULL;

  // The unpin listener requires the catalog, so this must be unregistered
  // before the catalog manager is removed
  if (cvmfs::unpin_listener_ != NULL) {
    quota::UnregisterListener(cvmfs::unpin_listener_);
    cvmfs::unpin_listener_ = NULL;
  }
  if (cvmfs::watchdog_listener_ != NULL) {
    quota::UnregisterListener(cvmfs::watchdog_listener_);
    cvmfs::watchdog_listener_ = NULL;
  }

  delete cvmfs::directory_handles_;
  delete cvmfs::mount_point_;
  cvmfs::directory_handles_ = NULL;
  cvmfs::mount_point_ = NULL;
}


static void Fini() {
  ShutdownMountpoint();

  delete cvmfs::file_system_;
  delete cvmfs::options_mgr_;
  cvmfs::file_system_ = NULL;
  cvmfs::options_mgr_ = NULL;

  delete cvmfs::watchdog_;
  cvmfs::watchdog_ = NULL;

  delete g_boot_error;
  g_boot_error = NULL;
  auto_umount::SetMountpoint("");

  crypto::CleanupLibcryptoMt();
}


static int AltProcessFlavor(int argc, char **argv) {
  if (strcmp(argv[1], "__cachemgr__") == 0) {
    return PosixQuotaManager::MainCacheManager(argc, argv);
  }
  if (strcmp(argv[1], "__wpad__") == 0) {
    return download::MainResolveProxyDescription(argc, argv);
  }
  return 1;
}


static bool MaintenanceMode(const int fd_progress) {
  SendMsg2Socket(fd_progress, "Entering maintenance mode\n");
  string msg_progress = "Draining out kernel caches (";
  if (FuseInvalidator::HasFuseNotifyInval())
    msg_progress += "up to ";
  msg_progress += StringifyInt(static_cast<int>(
                               cvmfs::mount_point_->kcache_timeout_sec())) +
                  "s)\n";
  SendMsg2Socket(fd_progress, msg_progress);
  cvmfs::fuse_remounter_->EnterMaintenanceMode();
  return true;
}


static bool SaveState(const int fd_progress, loader::StateList *saved_states) {
  string msg_progress;

  unsigned num_open_dirs = cvmfs::directory_handles_->size();
  if (num_open_dirs != 0) {
#ifdef DEBUGMSG
    for (cvmfs::DirectoryHandles::iterator i =
         cvmfs::directory_handles_->begin(),
         iEnd = cvmfs::directory_handles_->end(); i != iEnd; ++i)
    {
      LogCvmfs(kLogCvmfs, kLogDebug, "saving dirhandle %d", i->first);
    }
#endif

    msg_progress = "Saving open directory handles (" +
      StringifyInt(num_open_dirs) + " handles)\n";
    SendMsg2Socket(fd_progress, msg_progress);

    // TODO(jblomer): should rather be saved just in a malloc'd memory block
    cvmfs::DirectoryHandles *saved_handles =
      new cvmfs::DirectoryHandles(*cvmfs::directory_handles_);
    loader::SavedState *save_open_dirs = new loader::SavedState();
    save_open_dirs->state_id = loader::kStateOpenDirs;
    save_open_dirs->state = saved_handles;
    saved_states->push_back(save_open_dirs);
  }

  if (!cvmfs::file_system_->IsNfsSource()) {
    msg_progress = "Saving inode tracker\n";
    SendMsg2Socket(fd_progress, msg_progress);
    glue::InodeTracker *saved_inode_tracker =
      new glue::InodeTracker(*cvmfs::mount_point_->inode_tracker());
    loader::SavedState *state_glue_buffer = new loader::SavedState();
    state_glue_buffer->state_id = loader::kStateGlueBufferV4;
    state_glue_buffer->state = saved_inode_tracker;
    saved_states->push_back(state_glue_buffer);
  }

  msg_progress = "Saving negative entry cache\n";
  SendMsg2Socket(fd_progress, msg_progress);
  glue::DentryTracker *saved_dentry_tracker =
    new glue::DentryTracker(*cvmfs::mount_point_->dentry_tracker());
  loader::SavedState *state_dentry_tracker = new loader::SavedState();
  state_dentry_tracker->state_id = loader::kStateDentryTracker;
  state_dentry_tracker->state = saved_dentry_tracker;
  saved_states->push_back(state_dentry_tracker);

  msg_progress = "Saving page cache entry tracker\n";
  SendMsg2Socket(fd_progress, msg_progress);
  glue::PageCacheTracker *saved_page_cache_tracker =
    new glue::PageCacheTracker(*cvmfs::mount_point_->page_cache_tracker());
  loader::SavedState *state_page_cache_tracker = new loader::SavedState();
  state_page_cache_tracker->state_id = loader::kStatePageCacheTracker;
  state_page_cache_tracker->state = saved_page_cache_tracker;
  saved_states->push_back(state_page_cache_tracker);

  msg_progress = "Saving chunk tables\n";
  SendMsg2Socket(fd_progress, msg_progress);
  ChunkTables *saved_chunk_tables = new ChunkTables(
    *cvmfs::mount_point_->chunk_tables());
  loader::SavedState *state_chunk_tables = new loader::SavedState();
  state_chunk_tables->state_id = loader::kStateOpenChunksV4;
  state_chunk_tables->state = saved_chunk_tables;
  saved_states->push_back(state_chunk_tables);

  msg_progress = "Saving inode generation\n";
  SendMsg2Socket(fd_progress, msg_progress);
  cvmfs::inode_generation_info_.inode_generation +=
    cvmfs::mount_point_->catalog_mgr()->inode_gauge();
  cvmfs::InodeGenerationInfo *saved_inode_generation =
    new cvmfs::InodeGenerationInfo(cvmfs::inode_generation_info_);
  loader::SavedState *state_inode_generation = new loader::SavedState();
  state_inode_generation->state_id = loader::kStateInodeGeneration;
  state_inode_generation->state = saved_inode_generation;
  saved_states->push_back(state_inode_generation);

  msg_progress = "Saving fuse state\n";
  SendMsg2Socket(fd_progress, msg_progress);
  cvmfs::FuseState *saved_fuse_state = new cvmfs::FuseState();
  saved_fuse_state->cache_symlinks = cvmfs::mount_point_->cache_symlinks();
  saved_fuse_state->has_dentry_expire =
    cvmfs::mount_point_->fuse_expire_entry();
  loader::SavedState *state_fuse = new loader::SavedState();
  state_fuse->state_id = loader::kStateFuse;
  state_fuse->state = saved_fuse_state;
  saved_states->push_back(state_fuse);

  // Close open file catalogs
  ShutdownMountpoint();

  loader::SavedState *state_cache_mgr = new loader::SavedState();
  state_cache_mgr->state_id = loader::kStateOpenFiles;
  state_cache_mgr->state =
    cvmfs::file_system_->cache_mgr()->SaveState(fd_progress);
  saved_states->push_back(state_cache_mgr);

  msg_progress = "Saving open files counter\n";
  uint32_t *saved_num_fd =
    new uint32_t(cvmfs::file_system_->no_open_files()->Get());
  loader::SavedState *state_num_fd = new loader::SavedState();
  state_num_fd->state_id = loader::kStateOpenFilesCounter;
  state_num_fd->state = saved_num_fd;
  saved_states->push_back(state_num_fd);

  return true;
}


static bool RestoreState(const int fd_progress,
                         const loader::StateList &saved_states)
{
  // If we have no saved version of the page cache tracker, it is unsafe
  // to start using it.  The page cache tracker has to run for the entire
  // lifetime of the mountpoint or not at all.
  cvmfs::mount_point_->page_cache_tracker()->Disable();

  for (unsigned i = 0, l = saved_states.size(); i < l; ++i) {
    if (saved_states[i]->state_id == loader::kStateOpenDirs) {
      SendMsg2Socket(fd_progress, "Restoring open directory handles... ");
      delete cvmfs::directory_handles_;
      cvmfs::DirectoryHandles *saved_handles =
        (cvmfs::DirectoryHandles *)saved_states[i]->state;
      cvmfs::directory_handles_ = new cvmfs::DirectoryHandles(*saved_handles);
      cvmfs::file_system_->no_open_dirs()->Set(
        cvmfs::directory_handles_->size());
      cvmfs::DirectoryHandles::const_iterator i =
        cvmfs::directory_handles_->begin();
      for (; i != cvmfs::directory_handles_->end(); ++i) {
        if (i->first >= cvmfs::next_directory_handle_)
          cvmfs::next_directory_handle_ = i->first + 1;
      }

      SendMsg2Socket(fd_progress,
        StringifyInt(cvmfs::directory_handles_->size()) + " handles\n");
    }

    if (saved_states[i]->state_id == loader::kStateGlueBuffer) {
      SendMsg2Socket(fd_progress, "Migrating inode tracker (v1 to v4)... ");
      compat::inode_tracker::InodeTracker *saved_inode_tracker =
        (compat::inode_tracker::InodeTracker *)saved_states[i]->state;
      compat::inode_tracker::Migrate(
        saved_inode_tracker, cvmfs::mount_point_->inode_tracker());
      SendMsg2Socket(fd_progress, " done\n");
    }

    if (saved_states[i]->state_id == loader::kStateGlueBufferV2) {
      SendMsg2Socket(fd_progress, "Migrating inode tracker (v2 to v4)... ");
      compat::inode_tracker_v2::InodeTracker *saved_inode_tracker =
        (compat::inode_tracker_v2::InodeTracker *)saved_states[i]->state;
      compat::inode_tracker_v2::Migrate(saved_inode_tracker,
                                        cvmfs::mount_point_->inode_tracker());
      SendMsg2Socket(fd_progress, " done\n");
    }

    if (saved_states[i]->state_id == loader::kStateGlueBufferV3) {
      SendMsg2Socket(fd_progress, "Migrating inode tracker (v3 to v4)... ");
      compat::inode_tracker_v3::InodeTracker *saved_inode_tracker =
        (compat::inode_tracker_v3::InodeTracker *)saved_states[i]->state;
      compat::inode_tracker_v3::Migrate(saved_inode_tracker,
                                        cvmfs::mount_point_->inode_tracker());
      SendMsg2Socket(fd_progress, " done\n");
    }

    if (saved_states[i]->state_id == loader::kStateGlueBufferV4) {
      SendMsg2Socket(fd_progress, "Restoring inode tracker... ");
      cvmfs::mount_point_->inode_tracker()->~InodeTracker();
      glue::InodeTracker *saved_inode_tracker =
        (glue::InodeTracker *)saved_states[i]->state;
      new (cvmfs::mount_point_->inode_tracker())
        glue::InodeTracker(*saved_inode_tracker);
      SendMsg2Socket(fd_progress, " done\n");
    }

    if (saved_states[i]->state_id == loader::kStateDentryTracker) {
      SendMsg2Socket(fd_progress, "Restoring dentry tracker... ");
      cvmfs::mount_point_->dentry_tracker()->~DentryTracker();
      glue::DentryTracker *saved_dentry_tracker =
        static_cast<glue::DentryTracker *>(saved_states[i]->state);
      new (cvmfs::mount_point_->dentry_tracker())
        glue::DentryTracker(*saved_dentry_tracker);
      SendMsg2Socket(fd_progress, " done\n");
    }

    if (saved_states[i]->state_id == loader::kStatePageCacheTracker) {
      SendMsg2Socket(fd_progress, "Restoring page cache entry tracker... ");
      cvmfs::mount_point_->page_cache_tracker()->~PageCacheTracker();
      glue::PageCacheTracker *saved_page_cache_tracker =
        (glue::PageCacheTracker *)saved_states[i]->state;
      new (cvmfs::mount_point_->page_cache_tracker())
        glue::PageCacheTracker(*saved_page_cache_tracker);
      SendMsg2Socket(fd_progress, " done\n");
    }

    ChunkTables *chunk_tables = cvmfs::mount_point_->chunk_tables();

    if (saved_states[i]->state_id == loader::kStateOpenChunks) {
      SendMsg2Socket(fd_progress, "Migrating chunk tables (v1 to v4)... ");
      compat::chunk_tables::ChunkTables *saved_chunk_tables =
        (compat::chunk_tables::ChunkTables *)saved_states[i]->state;
      compat::chunk_tables::Migrate(saved_chunk_tables, chunk_tables);
      SendMsg2Socket(fd_progress,
        StringifyInt(chunk_tables->handle2fd.size()) + " handles\n");
    }

    if (saved_states[i]->state_id == loader::kStateOpenChunksV2) {
      SendMsg2Socket(fd_progress, "Migrating chunk tables (v2 to v4)... ");
      compat::chunk_tables_v2::ChunkTables *saved_chunk_tables =
        (compat::chunk_tables_v2::ChunkTables *)saved_states[i]->state;
      compat::chunk_tables_v2::Migrate(saved_chunk_tables, chunk_tables);
      SendMsg2Socket(fd_progress,
        StringifyInt(chunk_tables->handle2fd.size()) + " handles\n");
    }

    if (saved_states[i]->state_id == loader::kStateOpenChunksV3) {
      SendMsg2Socket(fd_progress, "Migrating chunk tables (v3 to v4)... ");
      compat::chunk_tables_v3::ChunkTables *saved_chunk_tables =
        (compat::chunk_tables_v3::ChunkTables *)saved_states[i]->state;
      compat::chunk_tables_v3::Migrate(saved_chunk_tables, chunk_tables);
      SendMsg2Socket(fd_progress,
        StringifyInt(chunk_tables->handle2fd.size()) + " handles\n");
    }

    if (saved_states[i]->state_id == loader::kStateOpenChunksV4) {
      SendMsg2Socket(fd_progress, "Restoring chunk tables... ");
      chunk_tables->~ChunkTables();
      ChunkTables *saved_chunk_tables = reinterpret_cast<ChunkTables *>(
        saved_states[i]->state);
      new (chunk_tables) ChunkTables(*saved_chunk_tables);
      SendMsg2Socket(fd_progress, " done\n");
    }

    if (saved_states[i]->state_id == loader::kStateInodeGeneration) {
      SendMsg2Socket(fd_progress, "Restoring inode generation... ");
      cvmfs::InodeGenerationInfo *old_info =
        (cvmfs::InodeGenerationInfo *)saved_states[i]->state;
      if (old_info->version == 1) {
        // Migration
        cvmfs::inode_generation_info_.initial_revision =
          old_info->initial_revision;
        cvmfs::inode_generation_info_.incarnation = old_info->incarnation;
        // Note: in the rare case of inode generation being 0 before, inode
        // can clash after reload before remount
      } else {
        cvmfs::inode_generation_info_ = *old_info;
      }
      ++cvmfs::inode_generation_info_.incarnation;
      SendMsg2Socket(fd_progress, " done\n");
    }

    if (saved_states[i]->state_id == loader::kStateOpenFilesCounter) {
      SendMsg2Socket(fd_progress, "Restoring open files counter... ");
      cvmfs::file_system_->no_open_files()->Set(*(reinterpret_cast<uint32_t *>(
        saved_states[i]->state)));
      SendMsg2Socket(fd_progress, " done\n");
    }

    if (saved_states[i]->state_id == loader::kStateOpenFiles) {
      int old_root_fd = cvmfs::mount_point_->catalog_mgr()->root_fd();

      // TODO(jblomer): make this less hacky

      CacheManagerIds saved_type =
        cvmfs::file_system_->cache_mgr()->PeekState(saved_states[i]->state);
      int fixup_root_fd = -1;

      if ((saved_type == kStreamingCacheManager) &&
          (cvmfs::file_system_->cache_mgr()->id() != kStreamingCacheManager))
      {
        // stick to the streaming cache manager
        StreamingCacheManager *new_cache_mgr = new
          StreamingCacheManager(cvmfs::max_open_files_,
                                cvmfs::file_system_->cache_mgr(),
                                cvmfs::mount_point_->download_mgr(),
                                cvmfs::mount_point_->external_download_mgr());
        fixup_root_fd = new_cache_mgr->PlantFd(old_root_fd);
        cvmfs::file_system_->ReplaceCacheManager(new_cache_mgr);
        cvmfs::mount_point_->fetcher()->ReplaceCacheManager(new_cache_mgr);
        cvmfs::mount_point_->external_fetcher()->ReplaceCacheManager(
          new_cache_mgr);
      }

      if ((cvmfs::file_system_->cache_mgr()->id() == kStreamingCacheManager) &&
          (saved_type != kStreamingCacheManager))
      {
        // stick to the cache manager wrapped into the streaming cache
        CacheManager *wrapped_cache_mgr = dynamic_cast<StreamingCacheManager *>(
            cvmfs::file_system_->cache_mgr())->MoveOutBackingCacheMgr(
              &fixup_root_fd);
        delete cvmfs::file_system_->cache_mgr();
        cvmfs::file_system_->ReplaceCacheManager(wrapped_cache_mgr);
        cvmfs::mount_point_->fetcher()->ReplaceCacheManager(wrapped_cache_mgr);
        cvmfs::mount_point_->external_fetcher()->ReplaceCacheManager(
          wrapped_cache_mgr);
      }

      int new_root_fd = cvmfs::file_system_->cache_mgr()->RestoreState(
        fd_progress, saved_states[i]->state);
      LogCvmfs(kLogCvmfs, kLogDebug, "new root file catalog descriptor @%d",
               new_root_fd);
      if (new_root_fd >= 0) {
        cvmfs::file_system_->RemapCatalogFd(old_root_fd, new_root_fd);
      } else if (fixup_root_fd >= 0) {
        LogCvmfs(kLogCvmfs, kLogDebug,
                 "new root file catalog descriptor (fixup) @%d", fixup_root_fd);
        cvmfs::file_system_->RemapCatalogFd(old_root_fd, fixup_root_fd);
      }
    }

    if (saved_states[i]->state_id == loader::kStateFuse) {
      SendMsg2Socket(fd_progress, "Restoring fuse state... ");
      cvmfs::FuseState *fuse_state =
        static_cast<cvmfs::FuseState *>(saved_states[i]->state);
      if (!fuse_state->cache_symlinks)
        cvmfs::mount_point_->DisableCacheSymlinks();
      if (fuse_state->has_dentry_expire)
        cvmfs::mount_point_->EnableFuseExpireEntry();
      SendMsg2Socket(fd_progress, " done\n");
    }
  }
  if (cvmfs::mount_point_->inode_annotation()) {
    uint64_t saved_generation = cvmfs::inode_generation_info_.inode_generation;
    cvmfs::mount_point_->inode_annotation()->IncGeneration(saved_generation);
  }

  return true;
}


static void FreeSavedState(const int fd_progress,
                           const loader::StateList &saved_states)
{
  for (unsigned i = 0, l = saved_states.size(); i < l; ++i) {
    switch (saved_states[i]->state_id) {
      case loader::kStateOpenDirs:
        SendMsg2Socket(fd_progress, "Releasing saved open directory handles\n");
        delete static_cast<cvmfs::DirectoryHandles *>(saved_states[i]->state);
        break;
      case loader::kStateGlueBuffer:
        SendMsg2Socket(
          fd_progress, "Releasing saved glue buffer (version 1)\n");
        delete static_cast<compat::inode_tracker::InodeTracker *>(
          saved_states[i]->state);
        break;
      case loader::kStateGlueBufferV2:
        SendMsg2Socket(
          fd_progress, "Releasing saved glue buffer (version 2)\n");
        delete static_cast<compat::inode_tracker_v2::InodeTracker *>(
          saved_states[i]->state);
        break;
      case loader::kStateGlueBufferV3:
        SendMsg2Socket(
          fd_progress, "Releasing saved glue buffer (version 3)\n");
        delete static_cast<compat::inode_tracker_v3::InodeTracker *>(
          saved_states[i]->state);
        break;
      case loader::kStateGlueBufferV4:
        SendMsg2Socket(fd_progress, "Releasing saved glue buffer\n");
        delete static_cast<glue::InodeTracker *>(saved_states[i]->state);
        break;
      case loader::kStateDentryTracker:
        SendMsg2Socket(fd_progress, "Releasing saved dentry tracker\n");
        delete static_cast<glue::DentryTracker *>(saved_states[i]->state);
        break;
      case loader::kStatePageCacheTracker:
        SendMsg2Socket(fd_progress, "Releasing saved page cache entry cache\n");
        delete static_cast<glue::PageCacheTracker *>(saved_states[i]->state);
        break;
      case loader::kStateOpenChunks:
        SendMsg2Socket(fd_progress, "Releasing chunk tables (version 1)\n");
        delete static_cast<compat::chunk_tables::ChunkTables *>(
          saved_states[i]->state);
        break;
      case loader::kStateOpenChunksV2:
        SendMsg2Socket(fd_progress, "Releasing chunk tables (version 2)\n");
        delete static_cast<compat::chunk_tables_v2::ChunkTables *>(
          saved_states[i]->state);
        break;
      case loader::kStateOpenChunksV3:
        SendMsg2Socket(fd_progress, "Releasing chunk tables (version 3)\n");
        delete static_cast<compat::chunk_tables_v3::ChunkTables *>(
          saved_states[i]->state);
        break;
      case loader::kStateOpenChunksV4:
        SendMsg2Socket(fd_progress, "Releasing chunk tables\n");
        delete static_cast<ChunkTables *>(saved_states[i]->state);
        break;
      case loader::kStateInodeGeneration:
        SendMsg2Socket(fd_progress, "Releasing saved inode generation info\n");
        delete static_cast<cvmfs::InodeGenerationInfo *>(
          saved_states[i]->state);
        break;
      case loader::kStateOpenFiles:
        cvmfs::file_system_->cache_mgr()->FreeState(
          fd_progress, saved_states[i]->state);
        break;
      case loader::kStateOpenFilesCounter:
        SendMsg2Socket(fd_progress, "Releasing open files counter\n");
        delete static_cast<uint32_t *>(saved_states[i]->state);
        break;
      case loader::kStateFuse:
        SendMsg2Socket(fd_progress, "Releasing fuse state\n");
        delete static_cast<cvmfs::FuseState *>(saved_states[i]->state);
        break;
      default:
        break;
    }
  }
}


static void __attribute__((constructor)) LibraryMain() {
  g_cvmfs_exports = new loader::CvmfsExports();
  g_cvmfs_exports->so_version = PACKAGE_VERSION;
  g_cvmfs_exports->fnAltProcessFlavor = AltProcessFlavor;
  g_cvmfs_exports->fnInit = Init;
  g_cvmfs_exports->fnSpawn = Spawn;
  g_cvmfs_exports->fnFini = Fini;
  g_cvmfs_exports->fnGetErrorMsg = GetErrorMsg;
  g_cvmfs_exports->fnMaintenanceMode = MaintenanceMode;
  g_cvmfs_exports->fnSaveState = SaveState;
  g_cvmfs_exports->fnRestoreState = RestoreState;
  g_cvmfs_exports->fnFreeSavedState = FreeSavedState;
  cvmfs::SetCvmfsOperations(&g_cvmfs_exports->cvmfs_operations);
}


static void __attribute__((destructor)) LibraryExit() {
  delete g_cvmfs_exports;
  g_cvmfs_exports = NULL;
}