Whamcloud - gitweb
LU-2740 utils: Add support for --version option
[fs/lustre-release.git] / lustre / utils / lfs.c
index 41f76d5..7ca4a7e 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2012, Whamcloud, Inc.
+ * Copyright (c) 2011, 2013, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 #include <string.h>
 #include <mntent.h>
 #include <errno.h>
+#include <err.h>
 #include <pwd.h>
 #include <grp.h>
+#include <sys/quota.h>
 #include <sys/types.h>
 #include <sys/stat.h>
 #include <fcntl.h>
 # include <sys/quota.h>
 #endif
 
-/* For dirname() */
-#include <libgen.h>
-
-#include <lnet/lnetctl.h>
-
-#include <liblustre.h>
-#include <lustre/lustreapi.h>
-
+#include <libcfs/libcfs.h>
 #include <libcfs/libcfsutil.h>
-#include "obdctl.h"
+#include <lustre/lustreapi.h>
+#include <lustre_ver.h>
 
 /* all functions */
 static int lfs_setstripe(int argc, char **argv);
 static int lfs_find(int argc, char **argv);
 static int lfs_getstripe(int argc, char **argv);
+static int lfs_getdirstripe(int argc, char **argv);
+static int lfs_setdirstripe(int argc, char **argv);
+static int lfs_rmentry(int argc, char **argv);
 static int lfs_osts(int argc, char **argv);
 static int lfs_mdts(int argc, char **argv);
 static int lfs_df(int argc, char **argv);
@@ -104,33 +103,88 @@ static int lfs_changelog_clear(int argc, char **argv);
 static int lfs_fid2path(int argc, char **argv);
 static int lfs_path2fid(int argc, char **argv);
 static int lfs_data_version(int argc, char **argv);
+static int lfs_hsm_state(int argc, char **argv);
+static int lfs_hsm_set(int argc, char **argv);
+static int lfs_hsm_clear(int argc, char **argv);
+static int lfs_hsm_action(int argc, char **argv);
+static int lfs_hsm_archive(int argc, char **argv);
+static int lfs_hsm_restore(int argc, char **argv);
+static int lfs_hsm_release(int argc, char **argv);
+static int lfs_hsm_remove(int argc, char **argv);
+static int lfs_hsm_cancel(int argc, char **argv);
+static int lfs_swap_layouts(int argc, char **argv);
+static int lfs_mv(int argc, char **argv);
+
+#define SETSTRIPE_USAGE(_cmd, _tgt) \
+       "usage: "_cmd" [--stripe-count|-c <stripe_count>]\n"\
+       "                 [--stripe-index|-i <start_ost_idx>]\n"\
+       "                 [--stripe-size|-S <stripe_size>]\n"\
+       "                 [--pool|-p <pool_name>]\n"\
+       "                 [--block|-b] "_tgt"\n"\
+       "\tstripe_size:  Number of bytes on each OST (0 filesystem default)\n"\
+       "\t              Can be specified with k, m or g (in KB, MB and GB\n"\
+       "\t              respectively)\n"\
+       "\tstart_ost_idx: OST index of first stripe (-1 default)\n"\
+       "\tstripe_count: Number of OSTs to stripe over (0 default, -1 all)\n"\
+       "\tpool_name:    Name of OST pool to use (default none)\n"\
+       "\tblock:        Block file access during data migration"
 
 /* all avaialable commands */
 command_t cmdlist[] = {
-        {"setstripe", lfs_setstripe, 0,
-         "Create a new file with a specific striping pattern or\n"
-         "set the default striping pattern on an existing directory or\n"
-         "delete the default striping pattern from an existing directory\n"
-         "usage: setstripe [--stripe-count|-c <stripe_count>]\n"
-         "                 [--stripe-index|-i <start_ost_idx>]\n"
-         "                 [--stripe-size|-S <stripe_size>]\n"
-         "                 [--pool|-p <pool_name>] <directory|filename>\n"
-         " or\n"
-         "       setstripe -d <directory>   (to delete default striping)\n"
-         "\tstripe_size:  Number of bytes on each OST (0 filesystem default)\n"
-         "\t              Can be specified with k, m or g (in KB, MB and GB\n"
-         "\t              respectively)\n"
-         "\tstart_ost_idx: OST index of first stripe (-1 default)\n"
-         "\tstripe_count: Number of OSTs to stripe over (0 default, -1 all)\n"
-         "\tpool_name:    Name of OST pool to use (default none)"},
-        {"getstripe", lfs_getstripe, 0,
-         "To list the striping info for a given file or files in a\n"
-         "directory or recursively for all files in a directory tree.\n"
-         "usage: getstripe [--ost|-O <uuid>] [--quiet | -q] [--verbose | -v]\n"
-         "                 [--stripe-count|-c] [--stripe-index|-i]\n"
-         "                 [--pool|-p] [--stripe-size|-S] [--directory|-d]\n"
-         "                 [--mdt-index|-M] [--recursive|-r] [--raw|-R]\n"
-         "                 <directory|filename> ..."},
+       {"setstripe", lfs_setstripe, 0,
+        "Create a new file with a specific striping pattern or\n"
+        "set the default striping pattern on an existing directory or\n"
+        "delete the default striping pattern from an existing directory\n"
+        "usage: setstripe -d <directory>   (to delete default striping)\n"\
+        " or\n"
+        SETSTRIPE_USAGE("setstripe", "<directory|filename>")},
+       {"getstripe", lfs_getstripe, 0,
+        "To list the striping info for a given file or files in a\n"
+        "directory or recursively for all files in a directory tree.\n"
+        "usage: getstripe [--ost|-O <uuid>] [--quiet | -q] [--verbose | -v]\n"
+        "                 [--stripe-count|-c] [--stripe-index|-i]\n"
+        "                 [--pool|-p] [--stripe-size|-S] [--directory|-d]\n"
+        "                 [--mdt-index|-M] [--recursive|-r] [--raw|-R]\n"
+        "                 [--layout|-L]\n"
+        "                 <directory|filename> ..."},
+       {"setdirstripe", lfs_setdirstripe, 0,
+        "To create a striped directory on a specified MDT. This can only\n"
+        "be done on MDT0 with the right of administrator.\n"
+        "usage: setdirstripe <--count|-c stripe_count>\n"
+        "              [--index|-i mdt_index] [--hash-type|-t hash_type]\n"
+        "              [--default_stripe|-D ] [--mode|-m mode] <dir>\n"
+        "\tstripe_count: stripe count of the striped directory\n"
+        "\tmdt_index:  MDT index of first stripe\n"
+        "\thash_type:  hash type of the striped directory. Hash types:\n"
+        "      fnv_1a_64 FNV-1a hash algorithm (default)\n"
+        "      all_char  sum of characters % MDT_COUNT (not recommended)\n"
+        "\tdefault_stripe: set default dirstripe of the directory\n"
+        "\tmode: the mode of the directory\n"},
+       {"getdirstripe", lfs_getdirstripe, 0,
+        "To list the striping info for a given directory\n"
+        "or recursively for all directories in a directory tree.\n"
+        "usage: getdirstripe [--obd|-O <uuid>] [--quiet|-q] [--verbose|-v]\n"
+        "               [--count|-c ] [--index|-i ] [--raw|-R]\n"
+        "               [--recursive | -r] [ --default_stripe | -D ] <dir> "},
+       {"mkdir", lfs_setdirstripe, 0,
+        "To create a striped directory on a specified MDT. This can only\n"
+        "be done on MDT0 with the right of administrator.\n"
+        "usage: mkdir <--count|-c stripe_count>\n"
+        "              [--index|-i mdt_index] [--hash-type|-t hash_type]\n"
+        "              [--default_stripe|-D ] [--mode|-m mode] <dir>\n"
+        "\tstripe_count: stripe count of the striped directory\n"
+        "\tmdt_index:  MDT index of first stripe\n"
+        "\thash_type:  hash type of the striped directory. Hash types:\n"
+        "      fnv_1a_64 FNV-1a hash algorithm (default)\n"
+        "      all_char  sum of characters % MDT_COUNT (not recommended)\n"
+        "\tdefault_stripe: set default dirstripe of the directory\n"
+        "\tmode: the mode of the directory\n"},
+       {"rm_entry", lfs_rmentry, 0,
+        "To remove the name entry of the remote directory. Note: This\n"
+        "command will only delete the name entry, i.e. the remote directory\n"
+        "will become inaccessable after this command. This can only be done\n"
+        "by the administrator\n"
+        "usage: rm_entry <dir>\n"},
         {"pool_list", lfs_poollist, 0,
          "List pools or pool OSTs\n"
          "usage: pool_list <fsname>[.<pool>] | <pathname>\n"},
@@ -147,6 +201,7 @@ command_t cmdlist[] = {
          "     [[!] --stripe-size|-S [+-]N[kMGT]] [[!] --type|-t <filetype>]\n"
          "     [[!] --gid|-g|--group|-G <gid>|<gname>]\n"
          "     [[!] --uid|-u|--user|-U <uid>|<uname>] [[!] --pool <pool>]\n"
+        "     [[!] --layout|-L released,raid0]\n"
          "\t !: used before an option indicates 'NOT' requested attribute\n"
          "\t -: used before a value indicates 'AT MOST' requested value\n"
          "\t +: used before a value indicates 'AT LEAST' requested value\n"},
@@ -194,9 +249,17 @@ command_t cmdlist[] = {
          "       -b can be used instead of --block-softlimit/--block-grace\n"
          "       -B can be used instead of --block-hardlimit\n"
          "       -i can be used instead of --inode-softlimit/--inode-grace\n"
-         "       -I can be used instead of --inode-hardlimit"},
+        "       -I can be used instead of --inode-hardlimit\n\n"
+        "Note: The total quota space will be split into many qunits and\n"
+        "      balanced over all server targets, the minimal qunit size is\n"
+        "      1M bytes for block space and 1K inodes for inode space.\n\n"
+        "      Quota space rebalancing process will stop when this mininum\n"
+        "      value is reached. As a result, quota exceeded can be returned\n"
+        "      while many targets still have 1MB or 1K inodes of spare\n"
+        "      quota space."},
         {"quota", lfs_quota, 0, "Display disk usage and limits.\n"
-         "usage: quota [-q] [-v] [-o <obd_uuid>|-i <mdt_idx>|-I <ost_idx>]\n"
+        "usage: quota [-q] [-v] [-h] [-o <obd_uuid>|-i <mdt_idx>|-I "
+                      "<ost_idx>]\n"
          "             [<-u|-g> <uname>|<uid>|<gname>|<gid>] <filesystem>\n"
          "       quota [-o <obd_uuid>|-i <mdt_idx>|-I <ost_idx>] -t <-u|-g> <filesystem>"},
 #endif
@@ -228,198 +291,460 @@ command_t cmdlist[] = {
          "interest to consumer <id>, allowing the system to free up space.\n"
          "An <endrec> of 0 means all records.\n"
          "usage: changelog_clear <mdtname> <id> <endrec>"},
-        {"fid2path", lfs_fid2path, 0,
-         "Resolve the full path to a given FID. For a specific hardlink "
-         "specify link number <linkno>.\n"
-         /* "For a historical name, specify changelog record <recno>.\n" */
-         "usage: fid2path <fsname|rootpath> <fid> [--link <linkno>]"
-                /*[--rec <recno>]*/},
-        {"path2fid", lfs_path2fid, 0, "Display the fid for a given path.\n"
-         "usage: path2fid <path>"},
-        {"data_version", lfs_data_version, 0, "Display file data version for "
-         "a given path.\n" "usage: data_version [-n] <path>"},
-        {"help", Parser_help, 0, "help"},
-        {"exit", Parser_quit, 0, "quit"},
-        {"quit", Parser_quit, 0, "quit"},
-        { 0, 0, 0, NULL }
+       {"fid2path", lfs_fid2path, 0,
+        "Resolve the full path(s) for given FID(s). For a specific hardlink "
+        "specify link number <linkno>.\n"
+       /* "For a historical link name, specify changelog record <recno>.\n" */
+        "usage: fid2path [--link <linkno>] <fsname|rootpath> <fid> ..."
+               /* [ --rec <recno> ] */ },
+       {"path2fid", lfs_path2fid, 0, "Display the fid(s) for a given path(s).\n"
+        "usage: path2fid <path> ..."},
+       {"data_version", lfs_data_version, 0, "Display file data version for "
+        "a given path.\n" "usage: data_version -[n|r|w] <path>"},
+       {"hsm_state", lfs_hsm_state, 0, "Display the HSM information (states, "
+        "undergoing actions) for given files.\n usage: hsm_state <file> ..."},
+       {"hsm_set", lfs_hsm_set, 0, "Set HSM user flag on specified files.\n"
+        "usage: hsm_set [--norelease] [--noarchive] [--dirty] [--exists] "
+        "[--archived] [--lost] <file> ..."},
+       {"hsm_clear", lfs_hsm_clear, 0, "Clear HSM user flag on specified "
+        "files.\n"
+        "usage: hsm_clear [--norelease] [--noarchive] [--dirty] [--exists] "
+        "[--archived] [--lost] <file> ..."},
+       {"hsm_action", lfs_hsm_action, 0, "Display current HSM request for "
+        "given files.\n" "usage: hsm_action <file> ..."},
+       {"hsm_archive", lfs_hsm_archive, 0,
+        "Archive file to external storage.\n"
+        "usage: hsm_archive [--filelist FILELIST] [--data DATA] [--archive NUM] "
+        "<file> ..."},
+       {"hsm_restore", lfs_hsm_restore, 0,
+        "Restore file from external storage.\n"
+        "usage: hsm_restore [--filelist FILELIST] [--data DATA] <file> ..."},
+       {"hsm_release", lfs_hsm_release, 0,
+        "Release files from Lustre.\n"
+        "usage: hsm_release [--filelist FILELIST] [--data DATA] <file> ..."},
+       {"hsm_remove", lfs_hsm_remove, 0,
+        "Remove file copy from external storage.\n"
+        "usage: hsm_remove [--filelist FILELIST] [--data DATA] <file> ..."},
+       {"hsm_cancel", lfs_hsm_cancel, 0,
+        "Cancel requests related to specified files.\n"
+        "usage: hsm_cancel [--filelist FILELIST] [--data DATA] <file> ..."},
+       {"swap_layouts", lfs_swap_layouts, 0, "Swap layouts between 2 files.\n"
+        "usage: swap_layouts <path1> <path2>"},
+       {"migrate", lfs_setstripe, 0, "migrate file from one OST layout to "
+        "another (may be not safe with concurent writes).\n"
+        SETSTRIPE_USAGE("migrate  ", "<filename>")},
+       {"mv", lfs_mv, 0,
+        "To move directories between MDTs.\n"
+        "usage: mv <directory|filename> [--mdt-index|-M] <mdt_index> "
+        "[--verbose|-v]\n"},
+       {"help", Parser_help, 0, "help"},
+       {"exit", Parser_quit, 0, "quit"},
+       {"quit", Parser_quit, 0, "quit"},
+       {"--version", Parser_version, 0,
+        "output build version of the utility and exit"},
+       { 0, 0, 0, NULL }
 };
 
-static int isnumber(const char *str)
-{
-        const char *ptr;
+#define MIGRATION_BLOCKS 1
 
-        if (str[0] != '-' && !isdigit(str[0]))
-                return 0;
-
-        for (ptr = str + 1; *ptr != '\0'; ptr++) {
-                if (!isdigit(*ptr))
-                        return 0;
-        }
+static int lfs_migrate(char *name, unsigned long long stripe_size,
+                      int stripe_offset, int stripe_count,
+                      int stripe_pattern, char *pool_name,
+                      __u64 migration_flags)
+{
+       int                      fd, fdv;
+       char                     volatile_file[PATH_MAX +
+                                               LUSTRE_VOLATILE_HDR_LEN + 4];
+       char                     parent[PATH_MAX];
+       char                    *ptr;
+       int                      rc;
+       __u64                    dv1;
+       struct lov_user_md      *lum = NULL;
+       int                      lumsz;
+       int                      bufsz;
+       void                    *buf = NULL;
+       int                      rsize, wsize;
+       __u64                    rpos, wpos, bufoff;
+       int                      gid = 0, sz;
+       int                      have_gl = 0;
+       struct stat              st, stv;
+
+       /* find the right size for the IO and allocate the buffer */
+       lumsz = lov_user_md_size(LOV_MAX_STRIPE_COUNT, LOV_USER_MAGIC_V3);
+       lum = malloc(lumsz);
+       if (lum == NULL) {
+               rc = -ENOMEM;
+               goto free;
+       }
+
+       rc = llapi_file_get_stripe(name, lum);
+       /* failure can come from may case and some may be not real error
+        * (eg: no stripe)
+        * in case of a real error, a later call will failed with a better
+        * error management */
+       if (rc < 0)
+               bufsz = 1024*1024;
+       else
+               bufsz = lum->lmm_stripe_size;
+       rc = posix_memalign(&buf, getpagesize(), bufsz);
+       if (rc != 0) {
+               rc = -rc;
+               goto free;
+       }
+
+       if (migration_flags & MIGRATION_BLOCKS) {
+               /* generate a random id for the grouplock */
+               fd = open("/dev/urandom", O_RDONLY);
+               if (fd == -1) {
+                       rc = -errno;
+                       fprintf(stderr, "cannot open /dev/urandom (%s)\n",
+                               strerror(-rc));
+                       goto free;
+               }
+               sz = sizeof(gid);
+               rc = read(fd, &gid, sz);
+               close(fd);
+               if (rc < sz) {
+                       rc = -errno;
+                       fprintf(stderr, "cannot read %d bytes from"
+                               " /dev/urandom (%s)\n", sz, strerror(-rc));
+                       goto free;
+               }
+       }
+
+       /* search for file directory pathname */
+       if (strlen(name) > sizeof(parent)-1) {
+               rc = -E2BIG;
+               goto free;
+       }
+       strncpy(parent, name, sizeof(parent));
+       ptr = strrchr(parent, '/');
+       if (ptr == NULL) {
+               if (getcwd(parent, sizeof(parent)) == NULL) {
+                       rc = -errno;
+                       goto free;
+               }
+       } else {
+               if (ptr == parent)
+                       strcpy(parent, "/");
+               else
+                       *ptr = '\0';
+       }
+       rc = snprintf(volatile_file, sizeof(volatile_file), "%s/%s::", parent,
+                     LUSTRE_VOLATILE_HDR);
+       if (rc >= sizeof(volatile_file)) {
+               rc = -E2BIG;
+               goto free;
+       }
+
+       /* create, open a volatile file, use caching (ie no directio) */
+       /* exclusive create is not needed because volatile files cannot
+        * conflict on name by construction */
+       fdv = llapi_file_open_pool(volatile_file, O_CREAT | O_WRONLY,
+                                  0644, stripe_size, stripe_offset,
+                                  stripe_count, stripe_pattern, pool_name);
+       if (fdv < 0) {
+               rc = fdv;
+               fprintf(stderr, "cannot create volatile file in %s (%s)\n",
+                       parent, strerror(-rc));
+               goto free;
+       }
+
+       /* open file, direct io */
+       /* even if the file is only read, WR mode is nedeed to allow
+        * layout swap on fd */
+       fd = open(name, O_RDWR | O_DIRECT);
+       if (fd == -1) {
+               rc = -errno;
+               fprintf(stderr, "cannot open %s (%s)\n", name, strerror(-rc));
+               close(fdv);
+               goto free;
+       }
+
+       /* Not-owner (root?) special case.
+        * Need to set owner/group of volatile file like original.
+        * This will allow to pass related check during layout_swap.
+        */
+       rc = fstat(fd, &st);
+       if (rc != 0) {
+               rc = -errno;
+               fprintf(stderr, "cannot stat %s (%s)\n", name,
+                       strerror(errno));
+               goto error;
+       }
+       rc = fstat(fdv, &stv);
+       if (rc != 0) {
+               rc = -errno;
+               fprintf(stderr, "cannot stat %s (%s)\n", volatile_file,
+                       strerror(errno));
+               goto error;
+       }
+       if (st.st_uid != stv.st_uid || st.st_gid != stv.st_gid) {
+               rc = fchown(fdv, st.st_uid, st.st_gid);
+               if (rc != 0) {
+                       rc = -errno;
+                       fprintf(stderr, "cannot chown %s (%s)\n", name,
+                               strerror(errno));
+                       goto error;
+               }
+       }
+
+       /* get file data version */
+       rc = llapi_get_data_version(fd, &dv1, LL_DV_RD_FLUSH);
+       if (rc != 0) {
+               fprintf(stderr, "cannot get dataversion on %s (%s)\n",
+                       name, strerror(-rc));
+               goto error;
+       }
+
+       if (migration_flags & MIGRATION_BLOCKS) {
+               /* take group lock to limit concurent access
+                * this will be no more needed when exclusive access will
+                * be implemented (see LU-2919) */
+               /* group lock is taken after data version read because it
+                * blocks data version call */
+               if (ioctl(fd, LL_IOC_GROUP_LOCK, gid) == -1) {
+                       rc = -errno;
+                       fprintf(stderr, "cannot get group lock on %s (%s)\n",
+                               name, strerror(-rc));
+                       goto error;
+               }
+               have_gl = 1;
+       }
+
+       /* copy data */
+       rpos = 0;
+       wpos = 0;
+       bufoff = 0;
+       rsize = -1;
+       do {
+               /* read new data only if we have written all
+                * previously read data */
+               if (wpos == rpos) {
+                       rsize = read(fd, buf, bufsz);
+                       if (rsize < 0) {
+                               rc = -errno;
+                               fprintf(stderr, "read failed on %s"
+                                       " (%s)\n", name,
+                                       strerror(-rc));
+                               goto error;
+                       }
+                       rpos += rsize;
+                       bufoff = 0;
+               }
+               /* eof ? */
+               if (rsize == 0)
+                       break;
+               wsize = write(fdv, buf + bufoff, rpos - wpos);
+               if (wsize < 0) {
+                       rc = -errno;
+                       fprintf(stderr, "write failed on volatile"
+                               " for %s (%s)\n", name, strerror(-rc));
+                       goto error;
+               }
+               wpos += wsize;
+               bufoff += wsize;
+       } while (1);
+
+       /* flush data */
+       fsync(fdv);
+
+       if (migration_flags & MIGRATION_BLOCKS) {
+               /* give back group lock */
+               if (ioctl(fd, LL_IOC_GROUP_UNLOCK, gid) == -1) {
+                       rc = -errno;
+                       fprintf(stderr, "cannot put group lock on %s (%s)\n",
+                               name, strerror(-rc));
+               }
+               have_gl = 0;
+       }
+
+       /* swap layouts
+        * for a migration we need to:
+        * - check data version on file did not change
+        * - keep file mtime
+        * - keep file atime
+        */
+       rc = llapi_fswap_layouts(fd, fdv, dv1, 0,
+                                SWAP_LAYOUTS_CHECK_DV1 |
+                                SWAP_LAYOUTS_KEEP_MTIME |
+                                SWAP_LAYOUTS_KEEP_ATIME);
+       if (rc == -EAGAIN) {
+               fprintf(stderr, "%s: dataversion changed during copy, "
+                       "migration aborted\n", name);
+               goto error;
+       }
+       if (rc != 0)
+               fprintf(stderr, "%s: swap layout to new file failed: %s\n",
+                       name, strerror(-rc));
 
-        return 1;
+error:
+       /* give back group lock */
+       if ((migration_flags & MIGRATION_BLOCKS) && have_gl &&
+           (ioctl(fd, LL_IOC_GROUP_UNLOCK, gid) == -1)) {
+               /* we keep in rc the original error */
+               fprintf(stderr, "cannot put group lock on %s (%s)\n",
+                       name, strerror(-errno));
+       }
+
+       close(fdv);
+       close(fd);
+free:
+       if (lum)
+               free(lum);
+       if (buf)
+               free(buf);
+       return rc;
 }
 
 /* functions */
 static int lfs_setstripe(int argc, char **argv)
 {
-        char *fname;
-        int result;
-        unsigned long long st_size;
-        int  st_offset, st_count;
-        char *end;
-        int c;
-        int delete = 0;
-        char *stripe_size_arg = NULL;
-        char *stripe_off_arg = NULL;
-        char *stripe_count_arg = NULL;
-        char *pool_name_arg = NULL;
-        unsigned long long size_units = 1;
-
-        struct option long_opts[] = {
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --count option"
-#else
-                /* This formerly implied "stripe-count", but was explicitly
-                 * made "stripe-count" for consistency with other options,
-                 * and to separate it from "mdt-count" when DNE arrives. */
-                {"count",        required_argument, 0, 'c'},
+       char                    *fname;
+       int                      result;
+       unsigned long long       st_size;
+       int                      st_offset, st_count;
+       char                    *end;
+       int                      c;
+       int                      delete = 0;
+       char                    *stripe_size_arg = NULL;
+       char                    *stripe_off_arg = NULL;
+       char                    *stripe_count_arg = NULL;
+       char                    *pool_name_arg = NULL;
+       unsigned long long       size_units = 1;
+       int                      migrate_mode = 0;
+       __u64                    migration_flags = 0;
+
+       struct option            long_opts[] = {
+               /* valid only in migrate mode */
+               {"block",        no_argument,       0, 'b'},
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               /* This formerly implied "stripe-count", but was explicitly
+                * made "stripe-count" for consistency with other options,
+                * and to separate it from "mdt-count" when DNE arrives. */
+               {"count",        required_argument, 0, 'c'},
 #endif
-                {"stripe-count", required_argument, 0, 'c'},
-                {"stripe_count", required_argument, 0, 'c'},
-                {"delete",       no_argument,       0, 'd'},
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --index option"
-#else
-                /* This formerly implied "stripe-index", but was explicitly
-                 * made "stripe-index" for consistency with other options,
-                 * and to separate it from "mdt-index" when DNE arrives. */
-                {"index",        required_argument, 0, 'i'},
+               {"stripe-count", required_argument, 0, 'c'},
+               {"stripe_count", required_argument, 0, 'c'},
+               {"delete",       no_argument,       0, 'd'},
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               /* This formerly implied "stripe-index", but was explicitly
+                * made "stripe-index" for consistency with other options,
+                * and to separate it from "mdt-index" when DNE arrives. */
+               {"index",        required_argument, 0, 'i'},
 #endif
-                {"stripe-index", required_argument, 0, 'i'},
-                {"stripe_index", required_argument, 0, 'i'},
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --offset option"
-#else
-                /* This formerly implied "stripe-index", but was confusing
-                 * with "file offset" (which will eventually be needed for
-                 * with different layouts by offset), so deprecate it. */
-                {"offset",       required_argument, 0, 'o'},
+               {"stripe-index", required_argument, 0, 'i'},
+               {"stripe_index", required_argument, 0, 'i'},
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               /* This formerly implied "stripe-index", but was confusing
+                * with "file offset" (which will eventually be needed for
+                * with different layouts by offset), so deprecate it. */
+               {"offset",       required_argument, 0, 'o'},
 #endif
-                {"pool",         required_argument, 0, 'p'},
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --size option"
-#else
-                /* This formerly implied "--stripe-size", but was confusing
-                 * with "lfs find --size|-s", which means "file size", so use
-                 * the consistent "--stripe-size|-S" for all commands. */
-                {"size",         required_argument, 0, 's'},
+               {"pool",         required_argument, 0, 'p'},
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               /* This formerly implied "--stripe-size", but was confusing
+                * with "lfs find --size|-s", which means "file size", so use
+                * the consistent "--stripe-size|-S" for all commands. */
+               {"size",         required_argument, 0, 's'},
 #endif
-                {"stripe-size",  required_argument, 0, 'S'},
-                {"stripe_size",  required_argument, 0, 'S'},
-                {0, 0, 0, 0}
-        };
+               {"stripe-size",  required_argument, 0, 'S'},
+               {"stripe_size",  required_argument, 0, 'S'},
+               {0, 0, 0, 0}
+       };
 
         st_size = 0;
         st_offset = -1;
         st_count = 0;
 
-#if LUSTRE_VERSION < OBD_OCD_VERSION(2,4,50,0)
-        if (argc == 5 && argv[1][0] != '-' &&
-            isnumber(argv[2]) && isnumber(argv[3]) && isnumber(argv[4])) {
-                fprintf(stderr, "error: obsolete usage of setstripe "
-                        "positional parameters.  Use -c, -i, -S instead.\n");
-                return CMD_HELP;
-        } else
-#else
-#warning "remove obsolete positional parameter error"
-#endif
-        {
-                optind = 0;
-                while ((c = getopt_long(argc, argv, "c:di:o:p:s:S:",
-                                        long_opts, NULL)) >= 0) {
-                switch (c) {
-                case 0:
-                        /* Long options. */
-                        break;
-                case 'c':
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --count option"
-#elif LUSTRE_VERSION >= OBD_OCD_VERSION(2,6,50,0)
-                        if (strcmp(argv[optind - 1], "--count") == 0)
-                                fprintf(stderr, "warning: '--count' deprecated"
-                                        ", use '--stripe-count' instead\n");
+       if (strcmp(argv[0], "migrate") == 0)
+               migrate_mode = 1;
+
+       optind = 0;
+       while ((c = getopt_long(argc, argv, "c:di:o:p:s:S:",
+                               long_opts, NULL)) >= 0) {
+               switch (c) {
+               case 0:
+                       /* Long options. */
+                       break;
+               case 'b':
+                       if (migrate_mode == 0) {
+                               fprintf(stderr, "--block is valid only for"
+                                               " migrate mode");
+                               return CMD_HELP;
+                       }
+                       migration_flags |= MIGRATION_BLOCKS;
+                       break;
+               case 'c':
+#if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
+                       if (strcmp(argv[optind - 1], "--count") == 0)
+                               fprintf(stderr, "warning: '--count' deprecated"
+                                       ", use '--stripe-count' instead\n");
 #endif
-                        stripe_count_arg = optarg;
-                        break;
-                case 'd':
-                        /* delete the default striping pattern */
-                        delete = 1;
-                        break;
-                case 'o':
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,4,50,0)
-                        fprintf(stderr, "warning: '--offset|-o' deprecated, "
-                                "use '--stripe-index|-i' instead\n");
-#else
-                        if (strcmp(argv[optind - 1], "--offset") == 0)
-                                /* need --stripe-index established first */
-                                fprintf(stderr, "warning: '--offset' deprecated"
-                                        ", use '--index' instead\n");
+                       stripe_count_arg = optarg;
+                       break;
+               case 'd':
+                       /* delete the default striping pattern */
+                       delete = 1;
+                       break;
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               case 'o':
+                       fprintf(stderr, "warning: '--offset|-o' deprecated, "
+                               "use '--stripe-index|-i' instead\n");
 #endif
-                case 'i':
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --offset and --index options"
-#elif LUSTRE_VERSION >= OBD_OCD_VERSION(2,6,50,0)
-                        if (strcmp(argv[optind - 1], "--index") == 0)
-                                fprintf(stderr, "warning: '--index' deprecated"
-                                        ", use '--stripe-index' instead\n");
+               case 'i':
+#if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
+                       if (strcmp(argv[optind - 1], "--index") == 0)
+                               fprintf(stderr, "warning: '--index' deprecated"
+                                       ", use '--stripe-index' instead\n");
 #endif
-                        stripe_off_arg = optarg;
-                        break;
-                case 's':
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --size option"
-#elif LUSTRE_VERSION >= OBD_OCD_VERSION(2,6,50,0)
-                        fprintf(stderr, "warning: '--size|-s' deprecated, "
-                                "use '--stripe-size|-S' instead\n");
+                       stripe_off_arg = optarg;
+                       break;
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               case 's':
+#if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
+                       fprintf(stderr, "warning: '--size|-s' deprecated, "
+                               "use '--stripe-size|-S' instead\n");
 #endif
-                case 'S':
-                        stripe_size_arg = optarg;
-                        break;
-                case 'p':
-                        pool_name_arg = optarg;
-                        break;
-                default:
-                        return CMD_HELP;
-                }
-                }
-
-                fname = argv[optind];
-
-                if (delete &&
-                    (stripe_size_arg != NULL || stripe_off_arg != NULL ||
-                     stripe_count_arg != NULL || pool_name_arg != NULL)) {
-                        fprintf(stderr, "error: %s: cannot specify -d with "
-                                        "-s, -c, -o, or -p options\n",
-                                        argv[0]);
-                        return CMD_HELP;
-                }
-        }
-
-        if (optind == argc) {
-                fprintf(stderr, "error: %s: missing filename|dirname\n",
-                        argv[0]);
-                return CMD_HELP;
-        }
-
-        /* get the stripe size */
-        if (stripe_size_arg != NULL) {
-                result = parse_size(stripe_size_arg, &st_size, &size_units, 0);
-                if (result) {
-                        fprintf(stderr, "error: %s: bad stripe size '%s'\n",
-                                argv[0], stripe_size_arg);
-                        return result;
-                }
-        }
+#endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0) */
+               case 'S':
+                       stripe_size_arg = optarg;
+                       break;
+               case 'p':
+                       pool_name_arg = optarg;
+                       break;
+               default:
+                       return CMD_HELP;
+               }
+       }
+
+       fname = argv[optind];
+
+       if (delete &&
+           (stripe_size_arg != NULL || stripe_off_arg != NULL ||
+            stripe_count_arg != NULL || pool_name_arg != NULL)) {
+               fprintf(stderr, "error: %s: cannot specify -d with "
+                       "-s, -c, -o, or -p options\n",
+                       argv[0]);
+               return CMD_HELP;
+       }
+
+       if (optind == argc) {
+               fprintf(stderr, "error: %s: missing filename|dirname\n",
+                       argv[0]);
+               return CMD_HELP;
+       }
+
+       /* get the stripe size */
+       if (stripe_size_arg != NULL) {
+               result = llapi_parse_size(stripe_size_arg, &st_size,
+                                         &size_units, 0);
+               if (result) {
+                       fprintf(stderr, "error: %s: bad stripe size '%s'\n",
+                               argv[0], stripe_size_arg);
+                       return result;
+               }
+       }
         /* get the stripe offset */
         if (stripe_off_arg != NULL) {
                 st_offset = strtol(stripe_off_arg, &end, 0);
@@ -439,18 +764,26 @@ static int lfs_setstripe(int argc, char **argv)
                 }
         }
 
-        do {
-                result = llapi_file_create_pool(fname, st_size, st_offset,
-                                                st_count, 0, pool_name_arg);
-                if (result) {
-                        fprintf(stderr,"error: %s: create stripe file '%s' "
-                                "failed\n", argv[0], fname);
-                        break;
-                }
-                fname = argv[++optind];
-        } while (fname != NULL);
-
-        return result;
+       do {
+               if (migrate_mode)
+                       result = lfs_migrate(fname, st_size, st_offset,
+                                            st_count, 0, pool_name_arg,
+                                            migration_flags);
+               else
+                       result = llapi_file_create_pool(fname, st_size,
+                                                       st_offset, st_count,
+                                                       0, pool_name_arg);
+               if (result) {
+                       fprintf(stderr,
+                               "error: %s: %s stripe file '%s' failed\n",
+                               argv[0], migrate_mode ? "migrate" : "create",
+                               fname);
+                       break;
+               }
+               fname = argv[++optind];
+       } while (fname != NULL);
+
+       return result;
 }
 
 static int lfs_poollist(int argc, char **argv)
@@ -543,12 +876,35 @@ static int id2name(char **name, unsigned int id, int type)
         return 0;
 }
 
+static int name2layout(__u32 *layout, char *name)
+{
+       char *ptr, *lyt;
+
+       *layout = 0;
+       for (ptr = name; ; ptr = NULL) {
+               lyt = strtok(ptr, ",");
+               if (lyt == NULL)
+                       break;
+               if (strcmp(lyt, "released") == 0)
+                       *layout |= LOV_PATTERN_F_RELEASED;
+               else if (strcmp(lyt, "raid0") == 0)
+                       *layout |= LOV_PATTERN_RAID0;
+               else
+                       return -1;
+       }
+       return 0;
+}
+
 #define FIND_POOL_OPT 3
 static int lfs_find(int argc, char **argv)
 {
-        int c, ret;
+       int c, rc;
+       int ret = 0;
         time_t t;
-        struct find_param param = { .maxdepth = -1, .quiet = 1 };
+       struct find_param param = {
+               .fp_max_depth = -1,
+               .quiet = 1,
+       };
         struct option long_opts[] = {
                 {"atime",        required_argument, 0, 'A'},
                 {"stripe-count", required_argument, 0, 'c'},
@@ -559,6 +915,7 @@ static int lfs_find(int argc, char **argv)
                 {"group",        required_argument, 0, 'G'},
                 {"stripe-index", required_argument, 0, 'i'},
                 {"stripe_index", required_argument, 0, 'i'},
+               {"layout",       required_argument, 0, 'L'},
                 {"mdt",          required_argument, 0, 'm'},
                 {"mtime",        required_argument, 0, 'M'},
                 {"name",         required_argument, 0, 'n'},
@@ -587,11 +944,11 @@ static int lfs_find(int argc, char **argv)
 
         time(&t);
 
-        optind = 0;
-        /* when getopt_long_only() hits '!' it returns 1, puts "!" in optarg */
-        while ((c = getopt_long_only(argc, argv,
-                                     "-A:c:C:D:g:G:i:m:M:n:O:Ppqrs:S:t:u:U:v",
-                                     long_opts, NULL)) >= 0) {
+       optind = 0;
+       /* when getopt_long_only() hits '!' it returns 1, puts "!" in optarg */
+       while ((c = getopt_long_only(argc, argv,
+                                    "-A:c:C:D:g:G:i:L:m:M:n:O:Ppqrs:S:t:u:U:v",
+                                    long_opts, NULL)) >= 0) {
                 xtime = NULL;
                 xsign = NULL;
                 if (neg_opt)
@@ -626,32 +983,32 @@ static int lfs_find(int argc, char **argv)
                         if (strcmp(optarg, "!") == 0)
                                 neg_opt = 2;
                         break;
-                case 'A':
-                        xtime = &param.atime;
-                        xsign = &param.asign;
-                        param.exclude_atime = !!neg_opt;
-                        /* no break, this falls through to 'C' for ctime */
-                case 'C':
-                        if (c == 'C') {
-                                xtime = &param.ctime;
-                                xsign = &param.csign;
-                                param.exclude_ctime = !!neg_opt;
-                        }
-                        /* no break, this falls through to 'M' for mtime */
-                case 'M':
-                        if (c == 'M') {
-                                xtime = &param.mtime;
-                                xsign = &param.msign;
-                                param.exclude_mtime = !!neg_opt;
-                        }
-                        ret = set_time(&t, xtime, optarg);
-                        if (ret == INT_MAX) {
-                                ret = -1;
-                                goto err;
-                        }
-                        if (ret)
-                                *xsign = ret;
-                        break;
+               case 'A':
+                       xtime = &param.fp_atime;
+                       xsign = &param.fp_asign;
+                       param.fp_exclude_atime = !!neg_opt;
+                       /* no break, this falls through to 'C' for ctime */
+               case 'C':
+                       if (c == 'C') {
+                               xtime = &param.fp_ctime;
+                               xsign = &param.fp_csign;
+                               param.fp_exclude_ctime = !!neg_opt;
+                       }
+                       /* no break, this falls through to 'M' for mtime */
+               case 'M':
+                       if (c == 'M') {
+                               xtime = &param.fp_mtime;
+                               xsign = &param.fp_msign;
+                               param.fp_exclude_mtime = !!neg_opt;
+                       }
+                       rc = set_time(&t, xtime, optarg);
+                       if (rc == INT_MAX) {
+                               ret = -1;
+                               goto err;
+                       }
+                       if (rc)
+                               *xsign = rc;
+                       break;
                 case 'c':
                         if (optarg[0] == '+') {
                                 param.stripecount_sign = -1;
@@ -671,14 +1028,14 @@ static int lfs_find(int argc, char **argv)
                         param.check_stripecount = 1;
                         param.exclude_stripecount = !!neg_opt;
                         break;
-                case 'D':
-                        param.maxdepth = strtol(optarg, 0, 0);
-                        break;
-                case 'g':
-                case 'G':
-                        ret = name2id(&param.gid, optarg, GROUP);
-                        if (ret) {
-                                param.gid = strtoul(optarg, &endptr, 10);
+               case 'D':
+                       param.fp_max_depth = strtol(optarg, 0, 0);
+                       break;
+               case 'g':
+               case 'G':
+                       rc = name2id(&param.fp_gid, optarg, GROUP);
+                       if (rc) {
+                               param.fp_gid = strtoul(optarg, &endptr, 10);
                                 if (*endptr != '\0') {
                                         fprintf(stderr, "Group/GID: %s cannot "
                                                 "be found.\n", optarg);
@@ -686,14 +1043,21 @@ static int lfs_find(int argc, char **argv)
                                         goto err;
                                 }
                         }
-                        param.exclude_gid = !!neg_opt;
-                        param.check_gid = 1;
-                        break;
+                       param.fp_exclude_gid = !!neg_opt;
+                       param.fp_check_gid = 1;
+                        break;
+               case 'L':
+                       ret = name2layout(&param.layout, optarg);
+                       if (ret)
+                               goto err;
+                       param.exclude_layout = !!neg_opt;
+                       param.check_layout = 1;
+                       break;
                 case 'u':
                 case 'U':
-                        ret = name2id(&param.uid, optarg, USER);
-                        if (ret) {
-                                param.uid = strtoul(optarg, &endptr, 10);
+                       rc = name2id(&param.fp_uid, optarg, USER);
+                       if (rc) {
+                               param.fp_uid = strtoul(optarg, &endptr, 10);
                                 if (*endptr != '\0') {
                                         fprintf(stderr, "User/UID: %s cannot "
                                                 "be found.\n", optarg);
@@ -701,8 +1065,8 @@ static int lfs_find(int argc, char **argv)
                                         goto err;
                                 }
                         }
-                        param.exclude_uid = !!neg_opt;
-                        param.check_uid = 1;
+                       param.fp_exclude_uid = !!neg_opt;
+                       param.fp_check_uid = 1;
                         break;
                 case FIND_POOL_OPT:
                         if (strlen(optarg) > LOV_MAXPOOLNAME) {
@@ -753,8 +1117,11 @@ static int lfs_find(int argc, char **argv)
                                 tmp = realloc(param.mdtuuid,
                                               param.num_alloc_mdts *
                                               sizeof(*param.mdtuuid));
-                                if (tmp == NULL)
-                                        GOTO(err_free, ret = -ENOMEM);
+                               if (tmp == NULL) {
+                                       ret = -ENOMEM;
+                                       goto err_free;
+                               }
+
                                 param.mdtuuid = tmp;
                         } else {
                                 param.exclude_obd = !!neg_opt;
@@ -762,25 +1129,36 @@ static int lfs_find(int argc, char **argv)
                                 tmp = realloc(param.obduuid,
                                               param.num_alloc_obds *
                                               sizeof(*param.obduuid));
-                                if (tmp == NULL)
-                                        GOTO(err_free, ret = -ENOMEM);
+                               if (tmp == NULL) {
+                                       ret = -ENOMEM;
+                                       goto err_free;
+                               }
+
                                 param.obduuid = tmp;
                         }
                         for (token = buf; token && *token; token = next) {
-                                char *uuid;
-                                if (c == 'm')
-                                        uuid =
-                                          param.mdtuuid[param.num_mdts++].uuid;
-                                else
-                                        uuid =
-                                          param.obduuid[param.num_obds++].uuid;
+                               struct obd_uuid *puuid;
+                               if (c == 'm') {
+                                       puuid =
+                                         &param.mdtuuid[param.num_mdts++];
+                               } else {
+                                       puuid =
+                                         &param.obduuid[param.num_obds++];
+                               }
                                 p = strchr(token, ',');
                                 next = 0;
                                 if (p) {
                                         *p = 0;
                                         next = p+1;
                                 }
-                                strcpy((char *)uuid, token);
+
+                               if (strlen(token) > sizeof(puuid->uuid) - 1) {
+                                       ret = -E2BIG;
+                                       goto err_free;
+                               }
+
+                               strncpy(puuid->uuid, token,
+                                       sizeof(puuid->uuid));
                         }
 err_free:
                         if (buf)
@@ -792,63 +1170,75 @@ err_free:
                         break;
                 case 'P':
                         break;
-                case 's':
-                        if (optarg[0] == '+') {
-                                param.size_sign = -1;
-                                optarg++;
-                        } else if (optarg[0] == '-') {
-                                param.size_sign =  1;
-                                optarg++;
-                        }
-
-                        ret = parse_size(optarg, &param.size,
-                                         &param.size_units, 0);
-                        if (ret) {
-                                fprintf(stderr, "error: bad file size '%s'\n",
-                                        optarg);
-                                goto err;
-                        }
-                        param.check_size = 1;
-                        param.exclude_size = !!neg_opt;
-                        break;
-                case 'S':
-                        if (optarg[0] == '+') {
-                                param.stripesize_sign = -1;
-                                optarg++;
-                        } else if (optarg[0] == '-') {
-                                param.stripesize_sign =  1;
-                                optarg++;
-                        }
-
-                        ret = parse_size(optarg, &param.stripesize,
-                                         &param.stripesize_units, 0);
-                        if (ret) {
-                                fprintf(stderr, "error: bad stripe_size '%s'\n",
-                                        optarg);
-                                goto err;
-                        }
-                        param.check_stripesize = 1;
-                        param.exclude_stripesize = !!neg_opt;
-                        break;
-                case 't':
-                        param.exclude_type = !!neg_opt;
-                        switch(optarg[0]) {
-                        case 'b': param.type = S_IFBLK; break;
-                        case 'c': param.type = S_IFCHR; break;
-                        case 'd': param.type = S_IFDIR; break;
-                        case 'f': param.type = S_IFREG; break;
-                        case 'l': param.type = S_IFLNK; break;
-                        case 'p': param.type = S_IFIFO; break;
-                        case 's': param.type = S_IFSOCK; break;
-#ifdef S_IFDOOR /* Solaris only */
-                        case 'D': param.type = S_IFDOOR; break;
-#endif
-                        default: fprintf(stderr, "error: %s: bad type '%s'\n",
-                                         argv[0], optarg);
-                                 ret = CMD_HELP;
-                                 goto err;
-                        };
-                        break;
+               case 's':
+                       if (optarg[0] == '+') {
+                               param.size_sign = -1;
+                               optarg++;
+                       } else if (optarg[0] == '-') {
+                               param.size_sign =  1;
+                               optarg++;
+                       }
+
+                       ret = llapi_parse_size(optarg, &param.size,
+                                              &param.size_units, 0);
+                       if (ret) {
+                               fprintf(stderr, "error: bad file size '%s'\n",
+                                       optarg);
+                               goto err;
+                       }
+                       param.check_size = 1;
+                       param.exclude_size = !!neg_opt;
+                       break;
+               case 'S':
+                       if (optarg[0] == '+') {
+                               param.stripesize_sign = -1;
+                               optarg++;
+                       } else if (optarg[0] == '-') {
+                               param.stripesize_sign =  1;
+                               optarg++;
+                       }
+
+                       ret = llapi_parse_size(optarg, &param.stripesize,
+                                              &param.stripesize_units, 0);
+                       if (ret) {
+                               fprintf(stderr, "error: bad stripe_size '%s'\n",
+                                       optarg);
+                               goto err;
+                       }
+                       param.check_stripesize = 1;
+                       param.exclude_stripesize = !!neg_opt;
+                       break;
+               case 't':
+                       param.fp_exclude_type = !!neg_opt;
+                       switch (optarg[0]) {
+                       case 'b':
+                               param.fp_type = S_IFBLK;
+                               break;
+                       case 'c':
+                               param.fp_type = S_IFCHR;
+                               break;
+                       case 'd':
+                               param.fp_type = S_IFDIR;
+                               break;
+                       case 'f':
+                               param.fp_type = S_IFREG;
+                               break;
+                       case 'l':
+                               param.fp_type = S_IFLNK;
+                               break;
+                       case 'p':
+                               param.fp_type = S_IFIFO;
+                               break;
+                       case 's':
+                               param.fp_type = S_IFSOCK;
+                               break;
+                       default:
+                               fprintf(stderr, "error: %s: bad type '%s'\n",
+                                       argv[0], optarg);
+                               ret = CMD_HELP;
+                               goto err;
+                       };
+                       break;
                 default:
                         ret = CMD_HELP;
                         goto err;
@@ -865,9 +1255,11 @@ err_free:
                 pathend = argc;
         }
 
-        do {
-                ret = llapi_find(argv[pathstart], &param);
-        } while (++pathstart < pathend && !ret);
+       do {
+               rc = llapi_find(argv[pathstart], &param);
+               if (rc != 0 && ret == 0)
+                       ret = rc;
+       } while (++pathstart < pathend);
 
         if (ret)
                 fprintf(stderr, "error: %s failed for %s.\n",
@@ -882,182 +1274,176 @@ err:
         return ret;
 }
 
-static int lfs_getstripe(int argc, char **argv)
+static int lfs_getstripe_internal(int argc, char **argv,
+                                 struct find_param *param)
 {
-        struct option long_opts[] = {
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --count option"
-#else
-                /* This formerly implied "stripe-count", but was explicitly
-                 * made "stripe-count" for consistency with other options,
-                 * and to separate it from "mdt-count" when DNE arrives. */
-                {"count",        no_argument,       0, 'c'},
+       struct option long_opts[] = {
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               /* This formerly implied "stripe-count", but was explicitly
+                * made "stripe-count" for consistency with other options,
+                * and to separate it from "mdt-count" when DNE arrives. */
+               {"count",               no_argument,            0, 'c'},
 #endif
-                {"stripe-count", no_argument,       0, 'c'},
-                {"stripe_count", no_argument,       0, 'c'},
-                {"directory",    no_argument,       0, 'd'},
-                {"generation",   no_argument,       0, 'g'},
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --index option"
-#else
-                /* This formerly implied "stripe-index", but was explicitly
-                 * made "stripe-index" for consistency with other options,
-                 * and to separate it from "mdt-index" when DNE arrives. */
-                {"index",        no_argument,       0, 'i'},
+               {"stripe-count",        no_argument,            0, 'c'},
+               {"stripe_count",        no_argument,            0, 'c'},
+               {"directory",           no_argument,            0, 'd'},
+               {"default",             no_argument,            0, 'D'},
+               {"generation",          no_argument,            0, 'g'},
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               /* This formerly implied "stripe-index", but was explicitly
+                * made "stripe-index" for consistency with other options,
+                * and to separate it from "mdt-index" when DNE arrives. */
+               {"index",               no_argument,            0, 'i'},
 #endif
-                {"stripe-index", no_argument,       0, 'i'},
-                {"stripe_index", no_argument,       0, 'i'},
-                {"mdt-index",    no_argument,       0, 'M'},
-                {"mdt_index",    no_argument,       0, 'M'},
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --offset option"
-#else
-                /* This formerly implied "stripe-index", but was confusing
-                 * with "file offset" (which will eventually be needed for
-                 * with different layouts by offset), so deprecate it. */
-                {"offset",       no_argument,       0, 'o'},
+               {"stripe-index",        no_argument,            0, 'i'},
+               {"stripe_index",        no_argument,            0, 'i'},
+               {"layout",              no_argument,            0, 'L'},
+               {"mdt-index",           no_argument,            0, 'M'},
+               {"mdt_index",           no_argument,            0, 'M'},
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               /* This formerly implied "stripe-index", but was confusing
+                * with "file offset" (which will eventually be needed for
+                * with different layouts by offset), so deprecate it. */
+               {"offset",              no_argument,            0, 'o'},
 #endif
-                {"obd",          required_argument, 0, 'O'},
-                {"ost",          required_argument, 0, 'O'},
-                {"pool",         no_argument,       0, 'p'},
-                {"quiet",        no_argument,       0, 'q'},
-                {"recursive",    no_argument,       0, 'r'},
-                {"raw",          no_argument,       0, 'R'},
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --size option"
-#else
-                /* This formerly implied "--stripe-size", but was confusing
-                 * with "lfs find --size|-s", which means "file size", so use
-                 * the consistent "--stripe-size|-S" for all commands. */
-                {"size",         no_argument,       0, 's'},
+               {"obd",                 required_argument,      0, 'O'},
+               {"ost",                 required_argument,      0, 'O'},
+               {"pool",                no_argument,            0, 'p'},
+               {"quiet",               no_argument,            0, 'q'},
+               {"recursive",           no_argument,            0, 'r'},
+               {"raw",                 no_argument,            0, 'R'},
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               /* This formerly implied "--stripe-size", but was confusing
+                * with "lfs find --size|-s", which means "file size", so use
+                * the consistent "--stripe-size|-S" for all commands. */
+               {"size",                no_argument,            0, 's'},
 #endif
-                {"stripe-size",  no_argument,       0, 'S'},
-                {"stripe_size",  no_argument,       0, 'S'},
-                {"verbose",      no_argument,       0, 'v'},
-                {0, 0, 0, 0}
-        };
-        int c, rc;
-        struct find_param param = { 0 };
-
-        param.maxdepth = 1;
-        optind = 0;
-        while ((c = getopt_long(argc, argv, "cdghiMoO:pqrRsSv",
-                                long_opts, NULL)) != -1) {
-                switch (c) {
-                case 'O':
-                        if (param.obduuid) {
-                                fprintf(stderr,
-                                        "error: %s: only one obduuid allowed",
-                                        argv[0]);
-                                return CMD_HELP;
-                        }
-                        param.obduuid = (struct obd_uuid *)optarg;
-                        break;
-                case 'q':
-                        param.quiet++;
-                        break;
-                case 'd':
-                        param.maxdepth = 0;
-                        break;
-                case 'r':
-                        param.recursive = 1;
-                        break;
-                case 'v':
-                        param.verbose = VERBOSE_ALL | VERBOSE_DETAIL;
-                        break;
-                case 'c':
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --count option"
-#elif LUSTRE_VERSION >= OBD_OCD_VERSION(2,6,50,0)
-                        if (strcmp(argv[optind - 1], "--count") == 0)
-                                fprintf(stderr, "warning: '--count' deprecated,"
-                                        " use '--stripe-count' instead\n");
+               {"stripe-size",         no_argument,            0, 'S'},
+               {"stripe_size",         no_argument,            0, 'S'},
+               {"verbose",             no_argument,            0, 'v'},
+               {0, 0, 0, 0}
+       };
+       int c, rc;
+
+       param->fp_max_depth = 1;
+       optind = 0;
+       while ((c = getopt_long(argc, argv, "cdDghiLMoO:pqrRsSv",
+                               long_opts, NULL)) != -1) {
+               switch (c) {
+               case 'O':
+                       if (param->obduuid) {
+                               fprintf(stderr,
+                                       "error: %s: only one obduuid allowed",
+                                       argv[0]);
+                               return CMD_HELP;
+                       }
+                       param->obduuid = (struct obd_uuid *)optarg;
+                       break;
+               case 'q':
+                       param->quiet++;
+                       break;
+               case 'd':
+                       param->fp_max_depth = 0;
+                       break;
+               case 'D':
+                       param->get_default_lmv = 1;
+                       break;
+               case 'r':
+                       param->recursive = 1;
+                       break;
+               case 'v':
+                       param->verbose = VERBOSE_ALL | VERBOSE_DETAIL;
+                       break;
+               case 'c':
+#if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
+                       if (strcmp(argv[optind - 1], "--count") == 0)
+                               fprintf(stderr, "warning: '--count' deprecated,"
+                                       " use '--stripe-count' instead\n");
 #endif
-                        if (!(param.verbose & VERBOSE_DETAIL)) {
-                                param.verbose |= VERBOSE_COUNT;
-                                param.maxdepth = 0;
-                        }
-                        break;
-                case 's':
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --size option"
-#elif LUSTRE_VERSION >= OBD_OCD_VERSION(2,6,50,0)
-                        fprintf(stderr, "warning: '--size|-s' deprecated, "
-                                "use '--stripe-size|-S' instead\n");
+                       if (!(param->verbose & VERBOSE_DETAIL)) {
+                               param->verbose |= VERBOSE_COUNT;
+                               param->fp_max_depth = 0;
+                       }
+                       break;
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               case 's':
+#if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
+                       fprintf(stderr, "warning: '--size|-s' deprecated, "
+                               "use '--stripe-size|-S' instead\n");
 #endif
-                case 'S':
-                        if (!(param.verbose & VERBOSE_DETAIL)) {
-                                param.verbose |= VERBOSE_SIZE;
-                                param.maxdepth = 0;
-                        }
-                        break;
-                case 'o':
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,4,50,0)
-                        fprintf(stderr, "warning: '--offset|-o' deprecated, "
-                                "use '--stripe-index|-i' instead\n");
-#else
-                        if (strcmp(argv[optind - 1], "--offset") == 0)
-                                /* need --stripe-index established first */
-                                fprintf(stderr, "warning: '--offset' deprecated"
-                                        ", use '--index' instead\n");
+#endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0) */
+               case 'S':
+                       if (!(param->verbose & VERBOSE_DETAIL)) {
+                               param->verbose |= VERBOSE_SIZE;
+                               param->fp_max_depth = 0;
+                       }
+                       break;
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 53, 0)
+               case 'o':
+                       fprintf(stderr, "warning: '--offset|-o' deprecated, "
+                               "use '--stripe-index|-i' instead\n");
 #endif
-                case 'i':
-#if LUSTRE_VERSION >= OBD_OCD_VERSION(2,9,50,0)
-#warning "remove deprecated --offset and --index options"
-#elif LUSTRE_VERSION >= OBD_OCD_VERSION(2,6,50,0)
-                        if (strcmp(argv[optind - 1], "--index") == 0)
-                                fprintf(stderr, "warning: '--index' deprecated"
-                                        ", use '--stripe-index' instead\n");
+               case 'i':
+#if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
+                       if (strcmp(argv[optind - 1], "--index") == 0)
+                               fprintf(stderr, "warning: '--index' deprecated"
+                                       ", use '--stripe-index' instead\n");
 #endif
-                        if (!(param.verbose & VERBOSE_DETAIL)) {
-                                param.verbose |= VERBOSE_OFFSET;
-                                param.maxdepth = 0;
-                        }
-                        break;
-                case 'p':
-                        if (!(param.verbose & VERBOSE_DETAIL)) {
-                                param.verbose |= VERBOSE_POOL;
-                                param.maxdepth = 0;
-                        }
-                        break;
-                case 'g':
-                        if (!(param.verbose & VERBOSE_DETAIL)) {
-                                param.verbose |= VERBOSE_GENERATION;
-                                param.maxdepth = 0;
-                        }
-                        break;
-                case 'M':
-                        if (!(param.verbose & VERBOSE_DETAIL))
-                                param.maxdepth = 0;
-                        param.verbose |= VERBOSE_MDTINDEX;
-                        break;
-                case 'R':
-                        param.raw = 1;
-                        break;
-                default:
-                        return CMD_HELP;
-                }
-        }
-
-        if (optind >= argc)
-                return CMD_HELP;
-
-        if (param.recursive)
-                param.maxdepth = -1;
-
-        if (!param.verbose)
-                param.verbose = VERBOSE_ALL;
-        if (param.quiet)
-                param.verbose = VERBOSE_OBJID;
-
-        do {
-                rc = llapi_getstripe(argv[optind], &param);
-        } while (++optind < argc && !rc);
-
-        if (rc)
-                fprintf(stderr, "error: %s failed for %s.\n",
-                        argv[0], argv[optind - 1]);
-        return rc;
+                       if (!(param->verbose & VERBOSE_DETAIL)) {
+                               param->verbose |= VERBOSE_OFFSET;
+                               param->fp_max_depth = 0;
+                       }
+                       break;
+               case 'p':
+                       if (!(param->verbose & VERBOSE_DETAIL)) {
+                               param->verbose |= VERBOSE_POOL;
+                               param->fp_max_depth = 0;
+                       }
+                       break;
+               case 'g':
+                       if (!(param->verbose & VERBOSE_DETAIL)) {
+                               param->verbose |= VERBOSE_GENERATION;
+                               param->fp_max_depth = 0;
+                       }
+                       break;
+               case 'L':
+                       if (!(param->verbose & VERBOSE_DETAIL)) {
+                               param->verbose |= VERBOSE_LAYOUT;
+                               param->fp_max_depth = 0;
+                       }
+                       break;
+               case 'M':
+                       if (!(param->verbose & VERBOSE_DETAIL))
+                               param->fp_max_depth = 0;
+                       param->verbose |= VERBOSE_MDTINDEX;
+                       break;
+               case 'R':
+                       param->raw = 1;
+                       break;
+               default:
+                       return CMD_HELP;
+               }
+       }
+
+       if (optind >= argc)
+               return CMD_HELP;
+
+       if (param->recursive)
+               param->fp_max_depth = -1;
+
+       if (!param->verbose)
+               param->verbose = VERBOSE_ALL;
+       if (param->quiet)
+               param->verbose = VERBOSE_OBJID;
+
+       do {
+               rc = llapi_getstripe(argv[optind], param);
+       } while (++optind < argc && !rc);
+
+       if (rc)
+               fprintf(stderr, "error: %s failed for %s.\n",
+                       argv[0], argv[optind - 1]);
+       return rc;
 }
 
 static int lfs_tgts(int argc, char **argv)
@@ -1098,6 +1484,241 @@ static int lfs_tgts(int argc, char **argv)
         return rc;
 }
 
+static int lfs_getstripe(int argc, char **argv)
+{
+       struct find_param param = { 0 };
+       return lfs_getstripe_internal(argc, argv, &param);
+}
+
+/* functions */
+static int lfs_getdirstripe(int argc, char **argv)
+{
+       struct find_param param = { 0 };
+
+       param.get_lmv = 1;
+       return lfs_getstripe_internal(argc, argv, &param);
+}
+
+/* functions */
+static int lfs_setdirstripe(int argc, char **argv)
+{
+       char                    *dname;
+       int                     result;
+       unsigned int            stripe_offset = -1;
+       unsigned int            stripe_count = 1;
+       enum lmv_hash_type      hash_type;
+       char                    *end;
+       int                     c;
+       char                    *stripe_offset_opt = NULL;
+       char                    *stripe_count_opt = NULL;
+       char                    *stripe_hash_opt = NULL;
+       char                    *mode_opt = NULL;
+       int                     default_stripe = 0;
+       mode_t                  mode = S_IRWXU | S_IRWXG | S_IRWXO;
+       mode_t                  previous_mode = 0;
+
+       struct option long_opts[] = {
+               {"count",       required_argument, 0, 'c'},
+               {"index",       required_argument, 0, 'i'},
+               {"mode",        required_argument, 0, 'm'},
+               {"hash-type",   required_argument, 0, 't'},
+               {"default_stripe", no_argument, 0, 'D'},
+               {0, 0, 0, 0}
+       };
+
+       optind = 0;
+
+       while ((c = getopt_long(argc, argv, "c:Di:m:t:", long_opts,
+                               NULL)) >= 0) {
+               switch (c) {
+               case 0:
+                       /* Long options. */
+                       break;
+               case 'c':
+                       stripe_count_opt = optarg;
+                       break;
+               case 'D':
+                       default_stripe = 1;
+                       break;
+               case 'i':
+                       stripe_offset_opt = optarg;
+                       break;
+               case 'm':
+                       mode_opt = optarg;
+                       break;
+               case 't':
+                       stripe_hash_opt = optarg;
+                       break;
+               default:
+                       fprintf(stderr, "error: %s: option '%s' "
+                                       "unrecognized\n",
+                                       argv[0], argv[optind - 1]);
+                       return CMD_HELP;
+               }
+       }
+
+       if (optind == argc) {
+               fprintf(stderr, "error: %s: missing dirname\n",
+                       argv[0]);
+               return CMD_HELP;
+       }
+
+       if (stripe_offset_opt == NULL && stripe_count_opt == NULL) {
+               fprintf(stderr, "error: %s: missing stripe offset and count.\n",
+                       argv[0]);
+               return CMD_HELP;
+       }
+
+       if (stripe_offset_opt != NULL) {
+               /* get the stripe offset */
+               stripe_offset = strtoul(stripe_offset_opt, &end, 0);
+               if (*end != '\0') {
+                       fprintf(stderr, "error: %s: bad stripe offset '%s'\n",
+                               argv[0], stripe_offset_opt);
+                       return CMD_HELP;
+               }
+       }
+
+       if (mode_opt != NULL) {
+               mode = strtoul(mode_opt, &end, 8);
+               if (*end != '\0') {
+                       fprintf(stderr, "error: %s: bad mode '%s'\n",
+                               argv[0], mode_opt);
+                       return CMD_HELP;
+               }
+               previous_mode = umask(0);
+       }
+
+       if (stripe_hash_opt == NULL ||
+           strcmp(stripe_hash_opt, LMV_HASH_NAME_FNV_1A_64) == 0) {
+               hash_type = LMV_HASH_TYPE_FNV_1A_64;
+       } else if (strcmp(stripe_hash_opt, LMV_HASH_NAME_ALL_CHARS) == 0) {
+               hash_type = LMV_HASH_TYPE_ALL_CHARS;
+       } else {
+               fprintf(stderr, "error: %s: bad stripe hash type '%s'\n",
+                       argv[0], stripe_hash_opt);
+               return CMD_HELP;
+       }
+
+       /* get the stripe count */
+       if (stripe_count_opt != NULL) {
+               stripe_count = strtoul(stripe_count_opt, &end, 0);
+               if (*end != '\0') {
+                       fprintf(stderr, "error: %s: bad stripe count '%s'\n",
+                               argv[0], stripe_count_opt);
+                       return CMD_HELP;
+               }
+       }
+
+       dname = argv[optind];
+       do {
+               if (default_stripe == 1) {
+                       result = llapi_dir_set_default_lmv_stripe(dname,
+                                                   stripe_offset, stripe_count,
+                                                   hash_type, NULL);
+               } else {
+                       result = llapi_dir_create_pool(dname, mode,
+                                                      stripe_offset,
+                                                      stripe_count, hash_type,
+                                                      NULL);
+               }
+
+               if (result) {
+                       fprintf(stderr, "error: %s: create stripe dir '%s' "
+                               "failed\n", argv[0], dname);
+                       break;
+               }
+               dname = argv[++optind];
+       } while (dname != NULL);
+
+       if (mode_opt != NULL)
+               umask(previous_mode);
+
+       return result;
+}
+
+/* functions */
+static int lfs_rmentry(int argc, char **argv)
+{
+       char *dname;
+       int   index;
+       int   result = 0;
+
+       if (argc <= 1) {
+               fprintf(stderr, "error: %s: missing dirname\n",
+                       argv[0]);
+               return CMD_HELP;
+       }
+
+       index = 1;
+       dname = argv[index];
+       while (dname != NULL) {
+               result = llapi_direntry_remove(dname);
+               if (result) {
+                       fprintf(stderr, "error: %s: remove dir entry '%s' "
+                               "failed\n", argv[0], dname);
+                       break;
+               }
+               dname = argv[++index];
+       }
+       return result;
+}
+
+static int lfs_mv(int argc, char **argv)
+{
+       struct  find_param param = {
+               .fp_max_depth = -1,
+               .mdtindex = -1,
+       };
+       char   *end;
+       int     c;
+       int     rc = 0;
+       struct option long_opts[] = {
+               {"--mdt-index", required_argument, 0, 'M'},
+               {"verbose",     no_argument,       0, 'v'},
+               {0, 0, 0, 0}
+       };
+
+       while ((c = getopt_long(argc, argv, "M:v", long_opts, NULL)) != -1) {
+               switch (c) {
+               case 'M': {
+                       param.mdtindex = strtoul(optarg, &end, 0);
+                       if (*end != '\0') {
+                               fprintf(stderr, "%s: invalid MDT index'%s'\n",
+                                       argv[0], optarg);
+                               return CMD_HELP;
+                       }
+                       break;
+               }
+               case 'v': {
+                       param.verbose = VERBOSE_DETAIL;
+                       break;
+               }
+               default:
+                       fprintf(stderr, "error: %s: unrecognized option '%s'\n",
+                               argv[0], argv[optind - 1]);
+                       return CMD_HELP;
+               }
+       }
+
+       if (param.mdtindex == -1) {
+               fprintf(stderr, "%s MDT index must be indicated\n", argv[0]);
+               return CMD_HELP;
+       }
+
+       if (optind >= argc) {
+               fprintf(stderr, "%s missing operand path\n", argv[0]);
+               return CMD_HELP;
+       }
+
+       param.migrate = 1;
+       rc = llapi_mv(argv[optind], &param);
+       if (rc != 0)
+               fprintf(stderr, "cannot migrate '%s' to MDT%04x: %s\n",
+                       argv[optind], param.mdtindex, strerror(-rc));
+       return rc;
+}
+
 static int lfs_osts(int argc, char **argv)
 {
         return lfs_tgts(argc, argv);
@@ -1132,7 +1753,10 @@ static int showdf(char *mntdir, struct obd_statfs *stat,
         double ratio = 0;
         char *suffix = "KMGTPEZY";
         /* Note if we have >2^64 bytes/fs these buffers will need to be grown */
-        char tbuf[20], ubuf[20], abuf[20], rbuf[20];
+       char tbuf[3 * sizeof(__u64)];
+       char ubuf[3 * sizeof(__u64)];
+       char abuf[3 * sizeof(__u64)];
+       char rbuf[3 * sizeof(__u64)];
 
         if (!uuid || !stat)
                 return -EINVAL;
@@ -1213,16 +1837,17 @@ struct ll_stat_type {
 static int mntdf(char *mntdir, char *fsname, char *pool, int ishow,
                int cooked, int lazy)
 {
-        struct obd_statfs stat_buf, sum = { .os_bsize = 1 };
-        struct obd_uuid uuid_buf;
-        char *poolname = NULL;
-        struct ll_stat_type types[] = { { LL_STATFS_LMV, "MDT" },
-                                        { LL_STATFS_LOV, "OST" },
-                                        { 0, NULL } };
-        struct ll_stat_type *tp;
-        __u32 index;
-        __u32 type;
-        int rc;
+       struct obd_statfs stat_buf, sum = { .os_bsize = 1 };
+       struct obd_uuid uuid_buf;
+       char *poolname = NULL;
+       struct ll_stat_type types[] = { { LL_STATFS_LMV, "MDT" },
+                                       { LL_STATFS_LOV, "OST" },
+                                       { 0, NULL } };
+       struct ll_stat_type *tp;
+       __u64 ost_ffree = 0;
+       __u32 index;
+       __u32 type;
+       int rc;
 
         if (pool) {
                 poolname = strchr(pool, '.');
@@ -1276,23 +1901,32 @@ static int mntdf(char *mntdir, char *fsname, char *pool, int ishow,
                                         sum.os_ffree += stat_buf.os_ffree;
                                         sum.os_files += stat_buf.os_files;
                                 } else /* if (tp->st_op == LL_STATFS_LOV) */ {
-                                        sum.os_blocks += stat_buf.os_blocks *
-                                                stat_buf.os_bsize;
-                                        sum.os_bfree  += stat_buf.os_bfree *
-                                                stat_buf.os_bsize;
-                                        sum.os_bavail += stat_buf.os_bavail *
-                                                stat_buf.os_bsize;
-                                }
-                        } else if (rc == -EINVAL || rc == -EFAULT) {
-                                break;
-                        }
-                }
-        }
-
-        printf("\n");
-        showdf(mntdir, &sum, "filesystem summary:", ishow, cooked, NULL, 0,0);
-        printf("\n");
-        return 0;
+                                       sum.os_blocks += stat_buf.os_blocks *
+                                               stat_buf.os_bsize;
+                                       sum.os_bfree  += stat_buf.os_bfree *
+                                               stat_buf.os_bsize;
+                                       sum.os_bavail += stat_buf.os_bavail *
+                                               stat_buf.os_bsize;
+                                       ost_ffree += stat_buf.os_ffree;
+                               }
+                       } else if (rc == -EINVAL || rc == -EFAULT) {
+                               break;
+                       }
+               }
+       }
+
+       /* If we don't have as many objects free on the OST as inodes
+        * on the MDS, we reduce the total number of inodes to
+        * compensate, so that the "inodes in use" number is correct.
+        * Matches ll_statfs_internal() so the results are consistent. */
+       if (ost_ffree < sum.os_ffree) {
+               sum.os_files = (sum.os_files - sum.os_ffree) + ost_ffree;
+               sum.os_ffree = ost_ffree;
+       }
+       printf("\n");
+       showdf(mntdir, &sum, "filesystem summary:", ishow, cooked, NULL, 0, 0);
+       printf("\n");
+       return 0;
 }
 
 static int lfs_df(int argc, char **argv)
@@ -1308,24 +1942,25 @@ static int lfs_df(int argc, char **argv)
                 {0, 0, 0, 0}
         };
 
-        optind = 0;
+       optind = 0;
        while ((c = getopt_long(argc, argv, "hilp:", long_opts, NULL)) != -1) {
-                switch (c) {
-                case 'i':
-                        ishow = 1;
-                        break;
-                case 'h':
-                        cooked = 1;
-                        break;
-                case 'l':
-                        lazy = 1;
-                case 'p':
-                        pool_name = optarg;
-                        break;
-                default:
-                        return CMD_HELP;
-                }
-        }
+               switch (c) {
+               case 'i':
+                       ishow = 1;
+                       break;
+               case 'h':
+                       cooked = 1;
+                       break;
+               case 'l':
+                       lazy = 1;
+                       break;
+               case 'p':
+                       pool_name = optarg;
+                       break;
+               default:
+                       return CMD_HELP;
+               }
+       }
         if (optind < argc && !realpath(argv[optind], path)) {
                 rc = -errno;
                 fprintf(stderr, "error: invalid path '%s': %s\n",
@@ -1423,9 +2058,7 @@ static int lfs_check(int argc, char **argv)
                 return rc;
         }
 
-        rc = llapi_target_iterate(num_types, obd_types,
-                                  mntdir, llapi_ping_target);
-
+       rc = llapi_target_check(num_types, obd_types, mntdir);
         if (rc)
                 fprintf(stderr, "error: %s: %s status failed\n",
                                 argv[0],argv[1]);
@@ -1714,17 +2347,17 @@ error:
         return ULONG_MAX;
 }
 
-#define ARG2ULL(nr, str, def_units)                                     \
-do {                                                                    \
-        unsigned long long limit, units = def_units;                    \
-        int rc;                                                         \
-                                                                        \
-        rc = parse_size(str, &limit, &units, 1);                        \
-        if (rc < 0) {                                                   \
-                fprintf(stderr, "error: bad limit value %s\n", str);    \
-                return CMD_HELP;                                        \
-        }                                                               \
-        nr = limit;                                                     \
+#define ARG2ULL(nr, str, def_units)                                    \
+do {                                                                   \
+       unsigned long long limit, units = def_units;                    \
+       int rc;                                                         \
+                                                                       \
+       rc = llapi_parse_size(str, &limit, &units, 1);                  \
+       if (rc < 0) {                                                   \
+               fprintf(stderr, "error: bad limit value %s\n", str);    \
+               return CMD_HELP;                                        \
+       }                                                               \
+       nr = limit;                                                     \
 } while (0)
 
 static inline int has_times_option(int argc, char **argv)
@@ -1874,19 +2507,43 @@ int lfs_setquota(int argc, char **argv)
                         ARG2ULL(dqb->dqb_bsoftlimit, optarg, 1024);
                         dqb->dqb_bsoftlimit >>= 10;
                         limit_mask |= BSLIMIT;
+                       if (dqb->dqb_bsoftlimit &&
+                           dqb->dqb_bsoftlimit <= 1024) /* <= 1M? */
+                               fprintf(stderr, "warning: block softlimit is "
+                                       "smaller than the miminal qunit size, "
+                                       "please see the help of setquota or "
+                                       "Lustre manual for details.\n");
                         break;
                 case 'B':
                         ARG2ULL(dqb->dqb_bhardlimit, optarg, 1024);
                         dqb->dqb_bhardlimit >>= 10;
                         limit_mask |= BHLIMIT;
+                       if (dqb->dqb_bhardlimit &&
+                           dqb->dqb_bhardlimit <= 1024) /* <= 1M? */
+                               fprintf(stderr, "warning: block hardlimit is "
+                                       "smaller than the miminal qunit size, "
+                                       "please see the help of setquota or "
+                                       "Lustre manual for details.\n");
                         break;
                 case 'i':
                         ARG2ULL(dqb->dqb_isoftlimit, optarg, 1);
                         limit_mask |= ISLIMIT;
+                       if (dqb->dqb_isoftlimit &&
+                           dqb->dqb_isoftlimit <= 1024) /* <= 1K inodes? */
+                               fprintf(stderr, "warning: inode softlimit is "
+                                       "smaller than the miminal qunit size, "
+                                       "please see the help of setquota or "
+                                       "Lustre manual for details.\n");
                         break;
                 case 'I':
                         ARG2ULL(dqb->dqb_ihardlimit, optarg, 1);
                         limit_mask |= IHLIMIT;
+                       if (dqb->dqb_ihardlimit &&
+                           dqb->dqb_ihardlimit <= 1024) /* <= 1K inodes? */
+                               fprintf(stderr, "warning: inode hardlimit is "
+                                       "smaller than the miminal qunit size, "
+                                       "please see the help of setquota or "
+                                       "Lustre manual for details.\n");
                         break;
                 default: /* getopt prints error message for us when opterr != 0 */
                         return CMD_HELP;
@@ -2026,29 +2683,50 @@ static void diff2str(time_t seconds, char *buf, time_t now)
         __sec2str(seconds - now, buf);
 }
 
-static void print_quota_title(char *name, struct if_quotactl *qctl)
+static void print_quota_title(char *name, struct if_quotactl *qctl,
+                             bool human_readable)
+{
+       printf("Disk quotas for %s %s (%cid %u):\n",
+              type2name(qctl->qc_type), name,
+              *type2name(qctl->qc_type), qctl->qc_id);
+       printf("%15s%8s %7s%8s%8s%8s %7s%8s%8s\n",
+              "Filesystem", human_readable ? "used" : "kbytes",
+              "quota", "limit", "grace",
+              "files", "quota", "limit", "grace");
+}
+
+static void kbytes2str(__u64 num, char *buf, bool h)
 {
-        printf("Disk quotas for %s %s (%cid %u):\n",
-               type2name(qctl->qc_type), name,
-               *type2name(qctl->qc_type), qctl->qc_id);
-        printf("%15s%8s %7s%8s%8s%8s %7s%8s%8s\n",
-               "Filesystem",
-               "kbytes", "quota", "limit", "grace",
-               "files", "quota", "limit", "grace");
+       if (!h) {
+               sprintf(buf, LPU64, num);
+       } else {
+               if (num >> 30)
+                       sprintf(buf, "%5.4gT", (double)num / (1 << 30));
+               else if (num >> 20)
+                       sprintf(buf, "%5.4gG", (double)num / (1 << 20));
+               else if (num >> 10)
+                       sprintf(buf, "%5.4gM", (double)num / (1 << 10));
+               else
+                       sprintf(buf, LPU64"%s", num, "k");
+       }
 }
 
-static void print_quota(char *mnt, struct if_quotactl *qctl, int type, int rc)
+static void print_quota(char *mnt, struct if_quotactl *qctl, int type,
+                       int rc, bool h)
 {
         time_t now;
 
         time(&now);
 
         if (qctl->qc_cmd == LUSTRE_Q_GETQUOTA || qctl->qc_cmd == Q_GETOQUOTA) {
-                int bover = 0, iover = 0;
-                struct obd_dqblk *dqb = &qctl->qc_dqblk;
+               int bover = 0, iover = 0;
+               struct obd_dqblk *dqb = &qctl->qc_dqblk;
+               char numbuf[3][32];
+               char timebuf[40];
+               char strbuf[32];
 
                 if (dqb->dqb_bhardlimit &&
-                    toqb(dqb->dqb_curspace) >= dqb->dqb_bhardlimit) {
+                   lustre_stoqb(dqb->dqb_curspace) >= dqb->dqb_bhardlimit) {
                         bover = 1;
                 } else if (dqb->dqb_bsoftlimit && dqb->dqb_btime) {
                         if (dqb->dqb_btime > now) {
@@ -2062,68 +2740,67 @@ static void print_quota(char *mnt, struct if_quotactl *qctl, int type, int rc)
                     dqb->dqb_curinodes >= dqb->dqb_ihardlimit) {
                         iover = 1;
                 } else if (dqb->dqb_isoftlimit && dqb->dqb_itime) {
-                        if (dqb->dqb_btime > now) {
-                                iover = 2;
-                        } else {
-                                iover = 3;
-                        }
+                       if (dqb->dqb_itime > now) {
+                               iover = 2;
+                       } else {
+                               iover = 3;
+                       }
                 }
 
-#if 0           /* XXX: always print quotas even when no usages */
-                if (dqb->dqb_curspace || dqb->dqb_curinodes)
-#endif
-                {
-                        char numbuf[3][32];
-                        char timebuf[40];
 
-                        if (strlen(mnt) > 15)
-                                printf("%s\n%15s", mnt, "");
-                        else
-                                printf("%15s", mnt);
+               if (strlen(mnt) > 15)
+                       printf("%s\n%15s", mnt, "");
+               else
+                       printf("%15s", mnt);
+
+               if (bover)
+                       diff2str(dqb->dqb_btime, timebuf, now);
+
+               kbytes2str(lustre_stoqb(dqb->dqb_curspace), strbuf, h);
+               if (rc == -EREMOTEIO)
+                       sprintf(numbuf[0], "%s*", strbuf);
+               else
+                       sprintf(numbuf[0], (dqb->dqb_valid & QIF_SPACE) ?
+                               "%s" : "[%s]", strbuf);
+
+               kbytes2str(dqb->dqb_bsoftlimit, strbuf, h);
+               if (type == QC_GENERAL)
+                       sprintf(numbuf[1], (dqb->dqb_valid & QIF_BLIMITS) ?
+                               "%s" : "[%s]", strbuf);
+               else
+                       sprintf(numbuf[1], "%s", "-");
+
+               kbytes2str(dqb->dqb_bhardlimit, strbuf, h);
+               sprintf(numbuf[2], (dqb->dqb_valid & QIF_BLIMITS) ?
+                       "%s" : "[%s]", strbuf);
+
+               printf(" %7s%c %6s %7s %7s",
+                      numbuf[0], bover ? '*' : ' ', numbuf[1],
+                      numbuf[2], bover > 1 ? timebuf : "-");
+
+               if (iover)
+                       diff2str(dqb->dqb_itime, timebuf, now);
+
+               sprintf(numbuf[0], (dqb->dqb_valid & QIF_INODES) ?
+                       LPU64 : "["LPU64"]", dqb->dqb_curinodes);
+
+               if (type == QC_GENERAL)
+                       sprintf(numbuf[1], (dqb->dqb_valid & QIF_ILIMITS) ?
+                               LPU64 : "["LPU64"]", dqb->dqb_isoftlimit);
+               else
+                       sprintf(numbuf[1], "%s", "-");
+
+               sprintf(numbuf[2], (dqb->dqb_valid & QIF_ILIMITS) ?
+                       LPU64 : "["LPU64"]", dqb->dqb_ihardlimit);
+
+               if (type != QC_OSTIDX)
+                       printf(" %7s%c %6s %7s %7s",
+                              numbuf[0], iover ? '*' : ' ', numbuf[1],
+                              numbuf[2], iover > 1 ? timebuf : "-");
+               else
+                       printf(" %7s %7s %7s %7s", "-", "-", "-", "-");
+               printf("\n");
 
-                        if (bover)
-                                diff2str(dqb->dqb_btime, timebuf, now);
-                        if (rc == -EREMOTEIO)
-                                sprintf(numbuf[0], LPU64"*",
-                                        toqb(dqb->dqb_curspace));
-                        else
-                                sprintf(numbuf[0],
-                                        (dqb->dqb_valid & QIF_SPACE) ?
-                                        LPU64 : "["LPU64"]",
-                                        toqb(dqb->dqb_curspace));
-                        if (type == QC_GENERAL)
-                                sprintf(numbuf[1], (dqb->dqb_valid & QIF_BLIMITS)
-                                        ? LPU64 : "["LPU64"]",
-                                        dqb->dqb_bsoftlimit);
-                        else
-                                sprintf(numbuf[1], "%s", "-");
-                        sprintf(numbuf[2], (dqb->dqb_valid & QIF_BLIMITS)
-                                ? LPU64 : "["LPU64"]", dqb->dqb_bhardlimit);
-                        printf(" %7s%c %6s %7s %7s",
-                               numbuf[0], bover ? '*' : ' ', numbuf[1],
-                               numbuf[2], bover > 1 ? timebuf : "-");
-
-                        if (iover)
-                                diff2str(dqb->dqb_itime, timebuf, now);
-
-                        sprintf(numbuf[0], (dqb->dqb_valid & QIF_INODES) ?
-                                LPU64 : "["LPU64"]", dqb->dqb_curinodes);
-                       if (type == QC_GENERAL)
-                                sprintf(numbuf[1], (dqb->dqb_valid & QIF_ILIMITS)
-                                        ? LPU64 : "["LPU64"]",
-                                        dqb->dqb_isoftlimit);
-                        else
-                                sprintf(numbuf[1], "%s", "-");
-                        sprintf(numbuf[2], (dqb->dqb_valid & QIF_ILIMITS) ?
-                                LPU64 : "["LPU64"]", dqb->dqb_ihardlimit);
-                        if (type != QC_OSTIDX)
-                                printf(" %7s%c %6s %7s %7s",
-                                       numbuf[0], iover ? '*' : ' ', numbuf[1],
-                                       numbuf[2], iover > 1 ? timebuf : "-");
-                        else
-                                printf(" %7s %7s %7s %7s", "-", "-", "-", "-");
-                        printf("\n");
-                }
         } else if (qctl->qc_cmd == LUSTRE_Q_GETINFO ||
                    qctl->qc_cmd == Q_GETOINFO) {
                 char bgtimebuf[40];
@@ -2136,7 +2813,8 @@ static void print_quota(char *mnt, struct if_quotactl *qctl, int type, int rc)
         }
 }
 
-static int print_obd_quota(char *mnt, struct if_quotactl *qctl, int is_mdt)
+static int print_obd_quota(char *mnt, struct if_quotactl *qctl, int is_mdt,
+                          bool h, __u64 *total)
 {
         int rc = 0, rc1 = 0, count = 0;
         __u32 valid = qctl->qc_valid;
@@ -2165,29 +2843,33 @@ static int print_obd_quota(char *mnt, struct if_quotactl *qctl, int is_mdt)
                         continue;
                 }
 
-                print_quota(obd_uuid2str(&qctl->obd_uuid), qctl, qctl->qc_valid, 0);
-        }
-
+               print_quota(obd_uuid2str(&qctl->obd_uuid), qctl,
+                           qctl->qc_valid, 0, h);
+               *total += is_mdt ? qctl->qc_dqblk.dqb_ihardlimit :
+                                  qctl->qc_dqblk.dqb_bhardlimit;
+       }
 out:
-        qctl->qc_valid = valid;
-        return rc ? : rc1;
+       qctl->qc_valid = valid;
+       return rc ? : rc1;
 }
 
 static int lfs_quota(int argc, char **argv)
 {
-        int c;
-        char *mnt, *name = NULL;
-        struct if_quotactl qctl = { .qc_cmd = LUSTRE_Q_GETQUOTA,
-                                    .qc_type = UGQUOTA };
-        char *obd_type = (char *)qctl.obd_type;
-        char *obd_uuid = (char *)qctl.obd_uuid.uuid;
-        int rc, rc1 = 0, rc2 = 0, rc3 = 0,
-            verbose = 0, pass = 0, quiet = 0, inacc;
-        char *endptr;
-        __u32 valid = QC_GENERAL, idx = 0;
-
-        optind = 0;
-        while ((c = getopt(argc, argv, "gi:I:o:qtuv")) != -1) {
+       int c;
+       char *mnt, *name = NULL;
+       struct if_quotactl qctl = { .qc_cmd = LUSTRE_Q_GETQUOTA,
+                                   .qc_type = UGQUOTA };
+       char *obd_type = (char *)qctl.obd_type;
+       char *obd_uuid = (char *)qctl.obd_uuid.uuid;
+       int rc, rc1 = 0, rc2 = 0, rc3 = 0,
+           verbose = 0, pass = 0, quiet = 0, inacc;
+       char *endptr;
+       __u32 valid = QC_GENERAL, idx = 0;
+       __u64 total_ialloc = 0, total_balloc = 0;
+       bool human_readable = false;
+
+       optind = 0;
+       while ((c = getopt(argc, argv, "gi:I:o:qtuvh")) != -1) {
                 switch (c) {
                 case 'u':
                         if (qctl.qc_type != UGQUOTA) {
@@ -2208,7 +2890,7 @@ static int lfs_quota(int argc, char **argv)
                         break;
                 case 'o':
                         valid = qctl.qc_valid = QC_UUID;
-                        strncpy(obd_uuid, optarg, sizeof(qctl.obd_uuid));
+                       strlcpy(obd_uuid, optarg, sizeof(qctl.obd_uuid));
                         break;
                 case 'i':
                         valid = qctl.qc_valid = QC_MDTIDX;
@@ -2224,6 +2906,9 @@ static int lfs_quota(int argc, char **argv)
                 case 'q':
                         quiet = 1;
                         break;
+               case 'h':
+                       human_readable = true;
+                       break;
                 default:
                         fprintf(stderr, "error: %s: option '-%c' "
                                         "unrecognized\n", argv[0], c);
@@ -2294,8 +2979,8 @@ ug_output:
                 }
         }
 
-        if (qctl.qc_cmd == LUSTRE_Q_GETQUOTA && !quiet)
-                print_quota_title(name, &qctl);
+       if (qctl.qc_cmd == LUSTRE_Q_GETQUOTA && !quiet)
+               print_quota_title(name, &qctl, human_readable);
 
         if (rc1 && *obd_type)
                 fprintf(stderr, "%s %s ", obd_type, obd_uuid);
@@ -2303,15 +2988,24 @@ ug_output:
         if (qctl.qc_valid != QC_GENERAL)
                 mnt = "";
 
-        inacc = (qctl.qc_cmd == LUSTRE_Q_GETQUOTA) &&
-                ((qctl.qc_dqblk.dqb_valid&(QIF_LIMITS|QIF_USAGE))!=(QIF_LIMITS|QIF_USAGE));
+       inacc = (qctl.qc_cmd == LUSTRE_Q_GETQUOTA) &&
+               ((qctl.qc_dqblk.dqb_valid & (QIF_LIMITS|QIF_USAGE)) !=
+                (QIF_LIMITS|QIF_USAGE));
 
-        print_quota(mnt, &qctl, QC_GENERAL, rc1);
+       print_quota(mnt, &qctl, QC_GENERAL, rc1, human_readable);
 
-        if (qctl.qc_valid == QC_GENERAL && qctl.qc_cmd != LUSTRE_Q_GETINFO && verbose) {
-                rc2 = print_obd_quota(mnt, &qctl, 1);
-                rc3 = print_obd_quota(mnt, &qctl, 0);
-        }
+       if (qctl.qc_valid == QC_GENERAL && qctl.qc_cmd != LUSTRE_Q_GETINFO &&
+           verbose) {
+               char strbuf[32];
+
+               rc2 = print_obd_quota(mnt, &qctl, 1, human_readable,
+                                     &total_ialloc);
+               rc3 = print_obd_quota(mnt, &qctl, 0, human_readable,
+                                     &total_balloc);
+               kbytes2str(total_balloc, strbuf, human_readable);
+               printf("Total allocated inode limit: "LPU64", total "
+                      "allocated block limit: %s\n", total_ialloc, strbuf);
+       }
 
         if (rc1 || rc2 || rc3 || inacc)
                 printf("Some errors happened when getting quota info. "
@@ -2348,10 +3042,10 @@ static int flushctx_ioctl(char *mp)
 
 static int lfs_flushctx(int argc, char **argv)
 {
-        int     kdestroy = 0, c;
-       FILE   *proc = NULL;
-        char    procline[PATH_MAX], *line;
-        int     rc = 0;
+       int     kdestroy = 0, c;
+       char    mntdir[PATH_MAX] = {'\0'};
+       int     index = 0;
+       int     rc = 0;
 
         optind = 0;
         while ((c = getopt(argc, argv, "k")) != -1) {
@@ -2367,46 +3061,24 @@ static int lfs_flushctx(int argc, char **argv)
         }
 
         if (kdestroy) {
-            int rc;
             if ((rc = system("kdestroy > /dev/null")) != 0) {
                 rc = WEXITSTATUS(rc);
                 fprintf(stderr, "error destroying tickets: %d, continuing\n", rc);
             }
         }
 
-        if (optind >= argc) {
-                /* flush for all mounted lustre fs. */
-                proc = fopen("/proc/mounts", "r");
-                if (!proc) {
-                        fprintf(stderr, "error: %s: can't open /proc/mounts\n",
-                                argv[0]);
-                        return -1;
-                }
+       if (optind >= argc) {
+               /* flush for all mounted lustre fs. */
+               while (!llapi_search_mounts(NULL, index++, mntdir, NULL)) {
+                       /* Check if we have a mount point */
+                       if (mntdir[0] == '\0')
+                               continue;
 
-                while ((line = fgets(procline, PATH_MAX, proc)) != NULL) {
-                        char dev[PATH_MAX];
-                        char mp[PATH_MAX];
-                        char fs[PATH_MAX];
-
-                        if (sscanf(line, "%s %s %s", dev, mp, fs) != 3) {
-                                fprintf(stderr, "%s: unexpected format in "
-                                                "/proc/mounts\n",
-                                        argv[0]);
+                       if (flushctx_ioctl(mntdir))
                                rc = -1;
-                               goto out;
-                        }
-
-                        if (strcmp(fs, "lustre") != 0)
-                                continue;
-                        /* we use '@' to determine it's a client. are there
-                         * any other better way?
-                         */
-                        if (strchr(dev, '@') == NULL)
-                                continue;
 
-                        if (flushctx_ioctl(mp))
-                                rc = -1;
-                }
+                       mntdir[0] = '\0'; /* avoid matching in next loop */
+               }
         } else {
                 /* flush fs as specified */
                 while (optind < argc) {
@@ -2414,10 +3086,6 @@ static int lfs_flushctx(int argc, char **argv)
                                 rc = -1;
                 }
         }
-
-out:
-       if (proc != NULL)
-               fclose(proc);
         return rc;
 }
 
@@ -2528,7 +3196,8 @@ static int lfs_changelog(int argc, char **argv)
                         /* namespace rec includes parent and filename */
                        printf(" p="DFID" %.*s", PFID(&rec->cr_pfid),
                                rec->cr_namelen, rec->cr_name);
-               if (fid_is_sane(&rec->cr_sfid))
+
+               if (!fid_is_zero(&rec->cr_sfid))
                        printf(" s="DFID" sp="DFID" %.*s",
                                PFID(&rec->cr_sfid), PFID(&rec->cr_spfid),
                                changelog_rec_snamelen(rec),
@@ -2577,7 +3246,7 @@ static int lfs_fid2path(int argc, char **argv)
         int linkno = -1;
         int lnktmp;
         int printcur = 0;
-        int rc;
+       int rc = 0;
 
         optind = 0;
 
@@ -2601,113 +3270,592 @@ static int lfs_fid2path(int argc, char **argv)
                         return CMD_HELP;
                 }
         }
-        device = argv[optind++];
-        fid = argv[optind++];
-        if (optind != argc)
-                return CMD_HELP;
 
-        path = calloc(1, PATH_MAX);
+       if (argc < 3)
+               return CMD_HELP;
+
+       device = argv[optind++];
+       path = calloc(1, PATH_MAX);
+       if (path == NULL) {
+               fprintf(stderr, "error: Not enough memory\n");
+               return -errno;
+       }
+
+       rc = 0;
+       while (optind < argc) {
+               fid = argv[optind++];
+
+               lnktmp = (linkno >= 0) ? linkno : 0;
+               while (1) {
+                       int oldtmp = lnktmp;
+                       long long rectmp = recno;
+                       int rc2;
+                       rc2 = llapi_fid2path(device, fid, path, PATH_MAX,
+                                            &rectmp, &lnktmp);
+                       if (rc2 < 0) {
+                               fprintf(stderr, "%s: error on FID %s: %s\n",
+                                       argv[0], fid, strerror(errno = -rc2));
+                               if (rc == 0)
+                                       rc = rc2;
+                               break;
+                       }
+
+                       if (printcur)
+                               fprintf(stdout, "%lld ", rectmp);
+                       if (device[0] == '/') {
+                               fprintf(stdout, "%s", device);
+                               if (device[strlen(device) - 1] != '/')
+                                       fprintf(stdout, "/");
+                       } else if (path[0] == '\0') {
+                               fprintf(stdout, "/");
+                       }
+                       fprintf(stdout, "%s\n", path);
+
+                       if (linkno >= 0)
+                               /* specified linkno */
+                               break;
+                       if (oldtmp == lnktmp)
+                               /* no more links */
+                               break;
+               }
+       }
+
+       free(path);
+       return rc;
+}
 
-        lnktmp = (linkno >= 0) ? linkno : 0;
-        while (1) {
-                int oldtmp = lnktmp;
-                long long rectmp = recno;
-                rc = llapi_fid2path(device, fid, path, PATH_MAX, &rectmp,
-                                    &lnktmp);
-                if (rc < 0) {
-                        fprintf(stderr, "%s error: %s\n", argv[0],
-                                strerror(errno = -rc));
-                        break;
-                }
+static int lfs_path2fid(int argc, char **argv)
+{
+       char **path;
+       const char *sep = "";
+       lustre_fid fid;
+       int rc = 0;
+
+       if (argc < 2)
+               return CMD_HELP;
+       else if (argc > 2)
+               sep = ": ";
+
+       path = argv + 1;
+       while (*path != NULL) {
+               int err = llapi_path2fid(*path, &fid);
+
+               if (err) {
+                       fprintf(stderr, "%s: can't get fid for %s: %s\n",
+                               argv[0], *path, strerror(-err));
+                       if (rc == 0) {
+                               rc = err;
+                               errno = -err;
+                       }
+                       goto out;
+               }
+               printf("%s%s"DFID"\n", *sep != '\0' ? *path : "", sep,
+                      PFID(&fid));
+out:
+               path++;
+       }
 
-                if (printcur)
-                        fprintf(stdout, "%lld ", rectmp);
-                if (device[0] == '/') {
-                        fprintf(stdout, "%s", device);
-                        if (device[strlen(device) - 1] != '/')
-                                fprintf(stdout, "/");
-                } else if (path[0] == '\0') {
-                        fprintf(stdout, "/");
-                }
-                fprintf(stdout, "%s\n", path);
+       return rc;
+}
 
-                if (linkno >= 0)
-                        /* specified linkno */
-                        break;
-                if (oldtmp == lnktmp)
-                        /* no more links */
-                        break;
-        }
+static int lfs_data_version(int argc, char **argv)
+{
+       char *path;
+       __u64 data_version;
+       int fd;
+       int rc;
+       int c;
+       int data_version_flags = LL_DV_RD_FLUSH; /* Read by default */
+
+       if (argc < 2)
+               return CMD_HELP;
+
+       optind = 0;
+       while ((c = getopt(argc, argv, "nrw")) != -1) {
+               switch (c) {
+               case 'n':
+                       data_version_flags = 0;
+                       break;
+               case 'r':
+                       data_version_flags |= LL_DV_RD_FLUSH;
+                       break;
+               case 'w':
+                       data_version_flags |= LL_DV_WR_FLUSH;
+                       break;
+               default:
+                       return CMD_HELP;
+               }
+       }
+       if (optind == argc)
+               return CMD_HELP;
+
+       path = argv[optind];
+       fd = open(path, O_RDONLY);
+       if (fd < 0)
+               err(errno, "cannot open file %s", path);
+
+       rc = llapi_get_data_version(fd, &data_version, data_version_flags);
+       if (rc < 0)
+               err(errno, "cannot get version for %s", path);
+       else
+               printf(LPU64 "\n", data_version);
 
-        free(path);
-        return rc;
+       close(fd);
+       return rc;
 }
 
-static int lfs_path2fid(int argc, char **argv)
+static int lfs_hsm_state(int argc, char **argv)
 {
-        char *path;
-        lustre_fid fid;
-        int rc;
+       int rc;
+       int i = 1;
+       char *path;
+       struct hsm_user_state hus;
+
+       if (argc < 2)
+               return CMD_HELP;
+
+       do {
+               path = argv[i];
+
+               rc = llapi_hsm_state_get(path, &hus);
+               if (rc) {
+                       fprintf(stderr, "can't get hsm state for %s: %s\n",
+                               path, strerror(errno = -rc));
+                       return rc;
+               }
+
+               /* Display path name and status flags */
+               printf("%s: (0x%08x)", path, hus.hus_states);
+
+               if (hus.hus_states & HS_RELEASED)
+                       printf(" released");
+               if (hus.hus_states & HS_EXISTS)
+                       printf(" exists");
+               if (hus.hus_states & HS_DIRTY)
+                       printf(" dirty");
+               if (hus.hus_states & HS_ARCHIVED)
+                       printf(" archived");
+               /* Display user-settable flags */
+               if (hus.hus_states & HS_NORELEASE)
+                       printf(" never_release");
+               if (hus.hus_states & HS_NOARCHIVE)
+                       printf(" never_archive");
+               if (hus.hus_states & HS_LOST)
+                       printf(" lost_from_hsm");
+
+               if (hus.hus_archive_id != 0)
+                       printf(", archive_id:%d", hus.hus_archive_id);
+               printf("\n");
 
-        if (argc != 2)
-                return CMD_HELP;
+       } while (++i < argc);
 
-        path = argv[1];
-        rc = llapi_path2fid(path, &fid);
-        if (rc) {
-                fprintf(stderr, "can't get fid for %s: %s\n", path,
-                        strerror(errno = -rc));
-                return rc;
-        }
+       return 0;
+}
 
-        printf(DFID"\n", PFID(&fid));
+#define LFS_HSM_SET   0
+#define LFS_HSM_CLEAR 1
 
-        return 0;
+/**
+ * Generic function to set or clear HSM flags.
+ * Used by hsm_set and hsm_clear.
+ *
+ * @mode  if LFS_HSM_SET, set the flags, if LFS_HSM_CLEAR, clear the flags.
+ */
+static int lfs_hsm_change_flags(int argc, char **argv, int mode)
+{
+       struct option long_opts[] = {
+               {"lost", 0, 0, 'l'},
+               {"norelease", 0, 0, 'r'},
+               {"noarchive", 0, 0, 'a'},
+               {"archived", 0, 0, 'A'},
+               {"dirty", 0, 0, 'd'},
+               {"exists", 0, 0, 'e'},
+               {0, 0, 0, 0}
+       };
+       char short_opts[] = "lraAde";
+       __u64 mask = 0;
+       int c, rc;
+       char *path;
+
+       if (argc < 3)
+               return CMD_HELP;
+
+       optind = 0;
+       while ((c = getopt_long(argc, argv, short_opts,
+                               long_opts, NULL)) != -1) {
+               switch (c) {
+               case 'l':
+                       mask |= HS_LOST;
+                       break;
+               case 'a':
+                       mask |= HS_NOARCHIVE;
+                       break;
+               case 'A':
+                       mask |= HS_ARCHIVED;
+                       break;
+               case 'r':
+                       mask |= HS_NORELEASE;
+                       break;
+               case 'd':
+                       mask |= HS_DIRTY;
+                       break;
+               case 'e':
+                       mask |= HS_EXISTS;
+                       break;
+               case '?':
+                       return CMD_HELP;
+               default:
+                       fprintf(stderr, "error: %s: option '%s' unrecognized\n",
+                               argv[0], argv[optind - 1]);
+                       return CMD_HELP;
+               }
+       }
+
+       /* User should have specified a flag */
+       if (mask == 0)
+               return CMD_HELP;
+
+       while (optind < argc) {
+
+               path = argv[optind];
+
+               /* If mode == 0, this means we apply the mask. */
+               if (mode == LFS_HSM_SET)
+                       rc = llapi_hsm_state_set(path, mask, 0, 0);
+               else
+                       rc = llapi_hsm_state_set(path, 0, mask, 0);
+
+               if (rc != 0) {
+                       fprintf(stderr, "Can't change hsm flags for %s: %s\n",
+                               path, strerror(errno = -rc));
+                       return rc;
+               }
+               optind++;
+       }
+
+       return 0;
 }
 
-static int lfs_data_version(int argc, char **argv)
+static int lfs_hsm_action(int argc, char **argv)
 {
-        char *path;
-        __u64 data_version;
-        int fd;
-        int rc;
-        int c;
-        int nolock = 0;
+       int                              rc;
+       int                              i = 1;
+       char                            *path;
+       struct hsm_current_action        hca;
+       struct hsm_extent                he;
+       enum hsm_user_action             hua;
+       enum hsm_progress_states         hps;
+
+       if (argc < 2)
+               return CMD_HELP;
+
+       do {
+               path = argv[i];
+
+               rc = llapi_hsm_current_action(path, &hca);
+               if (rc) {
+                       fprintf(stderr, "can't get hsm action for %s: %s\n",
+                               path, strerror(errno = -rc));
+                       return rc;
+               }
+               he = hca.hca_location;
+               hua = hca.hca_action;
+               hps = hca.hca_state;
+
+               printf("%s: %s", path, hsm_user_action2name(hua));
+
+               /* Skip file without action */
+               if (hca.hca_action == HUA_NONE) {
+                       printf("\n");
+                       continue;
+               }
+
+               printf(" %s ", hsm_progress_state2name(hps));
+
+               if ((hps == HPS_RUNNING) &&
+                   (hua == HUA_ARCHIVE || hua == HUA_RESTORE))
+                       printf("("LPX64 " bytes moved)\n", he.length);
+               else if ((he.offset + he.length) == LUSTRE_EOF)
+                       printf("(from "LPX64 " to EOF)\n", he.offset);
+               else
+                       printf("(from "LPX64 " to "LPX64")\n",
+                              he.offset, he.offset + he.length);
+
+       } while (++i < argc);
+
+       return 0;
+}
 
-        if (argc < 2)
-                return CMD_HELP;
+static int lfs_hsm_set(int argc, char **argv)
+{
+       return lfs_hsm_change_flags(argc, argv, LFS_HSM_SET);
+}
 
-        optind = 0;
-        while ((c = getopt(argc, argv, "n")) != -1) {
-                switch (c) {
-                case 'n':
-                        nolock = LL_DV_NOFLUSH;
-                        break;
-                default:
-                        return CMD_HELP;
-                }
-        }
-        if (optind == argc)
-                return CMD_HELP;
+static int lfs_hsm_clear(int argc, char **argv)
+{
+       return lfs_hsm_change_flags(argc, argv, LFS_HSM_CLEAR);
+}
 
-        path = argv[optind];
-        fd = open(path, O_RDONLY);
-        if (fd < 0) {
-                fprintf(stderr, "can't open %s: %s\n", path,
-                        strerror(errno));
-                return errno;
-        }
+/**
+ * Check file state and return its fid, to be used by lfs_hsm_request().
+ *
+ * \param[in]     file      Path to file to check
+ * \param[in,out] fid       Pointer to allocated lu_fid struct.
+ * \param[in,out] last_dev  Pointer to last device id used.
+ *
+ * \return 0 on success.
+ */
+static int lfs_hsm_prepare_file(char *file, struct lu_fid *fid,
+                               dev_t *last_dev)
+{
+       struct stat     st;
+       int             rc;
+
+       rc = lstat(file, &st);
+       if (rc) {
+               fprintf(stderr, "Cannot stat %s: %s\n", file, strerror(errno));
+               return -errno;
+       }
+       /* A request should be ... */
+       if (*last_dev != st.st_dev && *last_dev != 0) {
+               fprintf(stderr, "All files should be "
+                       "on the same filesystem: %s\n", file);
+               return -EINVAL;
+       }
+       *last_dev = st.st_dev;
+
+       rc = llapi_path2fid(file, fid);
+       if (rc) {
+               fprintf(stderr, "Cannot read FID of %s: %s\n",
+                       file, strerror(-rc));
+               return rc;
+       }
+       return 0;
+}
 
-        rc = llapi_get_data_version(fd, &data_version, nolock);
-        if (rc) {
-                fprintf(stderr, "can't get version for %s: %s\n", path,
-                        strerror(errno = -rc));
-       } else
-               printf(LPU64 "\n", data_version);
+static int lfs_hsm_request(int argc, char **argv, int action)
+{
+       struct option            long_opts[] = {
+               {"filelist", 1, 0, 'l'},
+               {"data", 1, 0, 'D'},
+               {"archive", 1, 0, 'a'},
+               {0, 0, 0, 0}
+       };
+       dev_t                    last_dev = 0;
+       char                     short_opts[] = "l:D:a:";
+       struct hsm_user_request *hur, *oldhur;
+       int                      c, i;
+       size_t                   len;
+       int                      nbfile;
+       char                    *line = NULL;
+       char                    *filelist = NULL;
+       char                     fullpath[PATH_MAX];
+       char                    *opaque = NULL;
+       int                      opaque_len = 0;
+       int                      archive_id = 0;
+       FILE                    *fp;
+       int                      nbfile_alloc = 0;
+       char                     some_file[PATH_MAX+1] = "";
+       int                      rc;
+
+       if (argc < 2)
+               return CMD_HELP;
+
+       optind = 0;
+       while ((c = getopt_long(argc, argv, short_opts,
+                               long_opts, NULL)) != -1) {
+               switch (c) {
+               case 'l':
+                       filelist = optarg;
+                       break;
+               case 'D':
+                       opaque = optarg;
+                       break;
+               case 'a':
+                       if (action != HUA_ARCHIVE) {
+                               fprintf(stderr,
+                                       "error: -a is supported only "
+                                       "when archiving\n");
+                               return CMD_HELP;
+                       }
+                       archive_id = atoi(optarg);
+                       break;
+               case '?':
+                       return CMD_HELP;
+               default:
+                       fprintf(stderr, "error: %s: option '%s' unrecognized\n",
+                               argv[0], argv[optind - 1]);
+                       return CMD_HELP;
+               }
+       }
+
+       /* All remaining args are files, so we have at least nbfile */
+       nbfile = argc - optind;
+
+       if ((nbfile == 0) && (filelist == NULL))
+               return CMD_HELP;
+
+       if (opaque != NULL)
+               opaque_len = strlen(opaque);
+
+       /* Alloc the request structure with enough place to store all files
+        * from command line. */
+       hur = llapi_hsm_user_request_alloc(nbfile, opaque_len);
+       if (hur == NULL) {
+               fprintf(stderr, "Cannot create the request: %s\n",
+                       strerror(errno));
+               return errno;
+       }
+       nbfile_alloc = nbfile;
+
+       hur->hur_request.hr_action = action;
+       hur->hur_request.hr_archive_id = archive_id;
+       hur->hur_request.hr_flags = 0;
+
+       /* All remaining args are files, add them */
+       if (nbfile != 0) {
+               if (strlen(argv[optind]) > sizeof(some_file)-1) {
+                       free(hur);
+                       return -E2BIG;
+               }
+               strncpy(some_file, argv[optind], sizeof(some_file));
+       }
+
+       for (i = 0; i < nbfile; i++) {
+               hur->hur_user_item[i].hui_extent.length = -1;
+               rc = lfs_hsm_prepare_file(argv[optind + i],
+                                         &hur->hur_user_item[i].hui_fid,
+                                         &last_dev);
+               hur->hur_request.hr_itemcount++;
+               if (rc)
+                       goto out_free;
+       }
+
+       /* from here stop using nb_file, use hur->hur_request.hr_itemcount */
+
+       /* If a filelist was specified, read the filelist from it. */
+       if (filelist != NULL) {
+               fp = fopen(filelist, "r");
+               if (fp == NULL) {
+                       fprintf(stderr, "Cannot read the file list %s: %s\n",
+                               filelist, strerror(errno));
+                       rc = -errno;
+                       goto out_free;
+               }
+
+               while ((rc = getline(&line, &len, fp)) != -1) {
+                       struct hsm_user_item *hui;
+
+                       /* If allocated buffer was too small, gets something
+                        * bigger */
+                       if (nbfile_alloc <= hur->hur_request.hr_itemcount) {
+                               ssize_t size;
+                               nbfile_alloc = nbfile_alloc * 2 + 1;
+                               oldhur = hur;
+                               hur = llapi_hsm_user_request_alloc(nbfile_alloc,
+                                                                  opaque_len);
+                               if (hur == NULL) {
+                                       fprintf(stderr, "Cannot allocate "
+                                               "the request: %s\n",
+                                               strerror(errno));
+                                       hur = oldhur;
+                                       rc = -errno;
+                                       fclose(fp);
+                                       goto out_free;
+                               }
+                               size = hur_len(oldhur);
+                               if (size < 0) {
+                                       fprintf(stderr, "Cannot allocate "
+                                               "the requested size\n");
+                                       hur = oldhur;
+                                       rc = -E2BIG;
+                                       fclose(fp);
+                                       goto out_free;
+                               }
+                               memcpy(hur, oldhur, size);
+                               free(oldhur);
+                       }
+
+                       /* Chop CR */
+                       if (line[strlen(line) - 1] == '\n')
+                               line[strlen(line) - 1] = '\0';
+
+                       hui =
+                            &hur->hur_user_item[hur->hur_request.hr_itemcount];
+                       hui->hui_extent.length = -1;
+                       rc = lfs_hsm_prepare_file(line, &hui->hui_fid,
+                                                 &last_dev);
+                       hur->hur_request.hr_itemcount++;
+                       if (rc) {
+                               fclose(fp);
+                               goto out_free;
+                       }
+
+                       if ((some_file[0] == '\0') &&
+                           (strlen(line) < sizeof(some_file)))
+                               strcpy(some_file, line);
+               }
+
+               rc = fclose(fp);
+               if (line)
+                       free(line);
+       }
+
+       /* If a --data was used, add it to the request */
+       hur->hur_request.hr_data_len = opaque_len;
+       if (opaque != NULL)
+               memcpy(hur_data(hur), opaque, opaque_len);
+
+       /* Send the HSM request */
+       if (realpath(some_file, fullpath) == NULL) {
+               fprintf(stderr, "Could not find path '%s': %s\n",
+                       some_file, strerror(errno));
+       }
+       rc = llapi_hsm_request(fullpath, hur);
+       if (rc) {
+               fprintf(stderr, "Cannot send HSM request (use of %s): %s\n",
+                       some_file, strerror(-rc));
+               goto out_free;
+       }
+
+out_free:
+       free(hur);
+       return rc;
+}
 
-       close(fd);
+static int lfs_hsm_archive(int argc, char **argv)
+{
+       return lfs_hsm_request(argc, argv, HUA_ARCHIVE);
+}
 
-       return rc;
+static int lfs_hsm_restore(int argc, char **argv)
+{
+       return lfs_hsm_request(argc, argv, HUA_RESTORE);
+}
+
+static int lfs_hsm_release(int argc, char **argv)
+{
+       return lfs_hsm_request(argc, argv, HUA_RELEASE);
+}
+
+static int lfs_hsm_remove(int argc, char **argv)
+{
+       return lfs_hsm_request(argc, argv, HUA_REMOVE);
+}
+
+static int lfs_hsm_cancel(int argc, char **argv)
+{
+       return lfs_hsm_request(argc, argv, HUA_CANCEL);
+}
+
+static int lfs_swap_layouts(int argc, char **argv)
+{
+       if (argc != 3)
+               return CMD_HELP;
+
+       return llapi_swap_layouts(argv[1], argv[2], 0, 0,
+                                 SWAP_LAYOUTS_KEEP_MTIME |
+                                 SWAP_LAYOUTS_KEEP_ATIME);
 }
 
 int main(int argc, char **argv)
@@ -2716,11 +3864,7 @@ int main(int argc, char **argv)
 
         setlinebuf(stdout);
 
-        ptl_initialize(argc, argv);
-        if (obd_initialize(argc, argv) < 0)
-                exit(2);
-
-        Parser_init("lfs > ", cmdlist);
+       Parser_init("lfs > ", cmdlist);
 
         if (argc > 1) {
                 rc = Parser_execarg(argc - 1, argv + 1, cmdlist);
@@ -2728,7 +3872,10 @@ int main(int argc, char **argv)
                 rc = Parser_commands();
         }
 
-        obd_finalize(argc, argv);
         return rc < 0 ? -rc : rc;
 }
 
+#ifdef _LUSTRE_IDL_H_
+/* Everything we need here should be included by lustreapi.h. */
+# error "lfs should not depend on lustre_idl.h"
+#endif /* _LUSTRE_IDL_H_ */