Whamcloud - gitweb
7fa9c5cf9eda37b59d55350bfe6eea445e8cfaa7
[fs/lustre-release.git] / lustre / utils / lfs.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2016, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/utils/lfs.c
33  *
34  * Author: Peter J. Braam <braam@clusterfs.com>
35  * Author: Phil Schwan <phil@clusterfs.com>
36  * Author: Robert Read <rread@clusterfs.com>
37  */
38
39 /* for O_DIRECTORY */
40 #ifndef _GNU_SOURCE
41 #define _GNU_SOURCE
42 #endif
43
44 #include <stdlib.h>
45 #include <stdio.h>
46 #include <getopt.h>
47 #include <string.h>
48 #include <mntent.h>
49 #include <unistd.h>
50 #include <errno.h>
51 #include <err.h>
52 #include <pwd.h>
53 #include <grp.h>
54 #include <sys/ioctl.h>
55 #include <sys/quota.h>
56 #include <sys/time.h>
57 #include <sys/types.h>
58 #include <sys/stat.h>
59 #include <fcntl.h>
60 #include <dirent.h>
61 #include <time.h>
62 #include <ctype.h>
63
64 #include <libcfs/util/string.h>
65 #include <libcfs/util/ioctl.h>
66 #include <libcfs/util/parser.h>
67 #include <lustre/lustreapi.h>
68 #include <lustre_ver.h>
69 #include <linux/lustre_param.h>
70
71 #ifndef ARRAY_SIZE
72 # define ARRAY_SIZE(a) ((sizeof(a)) / (sizeof((a)[0])))
73 #endif /* !ARRAY_SIZE */
74
75 /* all functions */
76 static int lfs_setstripe(int argc, char **argv);
77 static int lfs_find(int argc, char **argv);
78 static int lfs_getstripe(int argc, char **argv);
79 static int lfs_getdirstripe(int argc, char **argv);
80 static int lfs_setdirstripe(int argc, char **argv);
81 static int lfs_rmentry(int argc, char **argv);
82 static int lfs_osts(int argc, char **argv);
83 static int lfs_mdts(int argc, char **argv);
84 static int lfs_df(int argc, char **argv);
85 static int lfs_getname(int argc, char **argv);
86 static int lfs_check(int argc, char **argv);
87 #ifdef HAVE_SYS_QUOTA_H
88 static int lfs_setquota(int argc, char **argv);
89 static int lfs_quota(int argc, char **argv);
90 #endif
91 static int lfs_flushctx(int argc, char **argv);
92 static int lfs_cp(int argc, char **argv);
93 static int lfs_ls(int argc, char **argv);
94 static int lfs_poollist(int argc, char **argv);
95 static int lfs_changelog(int argc, char **argv);
96 static int lfs_changelog_clear(int argc, char **argv);
97 static int lfs_fid2path(int argc, char **argv);
98 static int lfs_path2fid(int argc, char **argv);
99 static int lfs_data_version(int argc, char **argv);
100 static int lfs_hsm_state(int argc, char **argv);
101 static int lfs_hsm_set(int argc, char **argv);
102 static int lfs_hsm_clear(int argc, char **argv);
103 static int lfs_hsm_action(int argc, char **argv);
104 static int lfs_hsm_archive(int argc, char **argv);
105 static int lfs_hsm_restore(int argc, char **argv);
106 static int lfs_hsm_release(int argc, char **argv);
107 static int lfs_hsm_remove(int argc, char **argv);
108 static int lfs_hsm_cancel(int argc, char **argv);
109 static int lfs_swap_layouts(int argc, char **argv);
110 static int lfs_mv(int argc, char **argv);
111 static int lfs_ladvise(int argc, char **argv);
112 static int lfs_list_commands(int argc, char **argv);
113
114 /* Setstripe and migrate share mostly the same parameters */
115 #define SSM_CMD_COMMON(cmd) \
116         "usage: "cmd" [--stripe-count|-c <stripe_count>]\n"             \
117         "                 [--stripe-index|-i <start_ost_idx>]\n"        \
118         "                 [--stripe-size|-S <stripe_size>]\n"           \
119         "                 [--pool|-p <pool_name>]\n"                    \
120         "                 [--ost|-o <ost_indices>]\n"                   \
121         "                 [--component-end|-E <comp_end>]\n"
122
123 #define SSM_HELP_COMMON \
124         "\tstripe_size:  Number of bytes on each OST (0 filesystem default)\n" \
125         "\t              Can be specified with k, m or g (in KB, MB and GB\n" \
126         "\t              respectively)\n"                               \
127         "\tstart_ost_idx: OST index of first stripe (-1 default)\n"     \
128         "\tstripe_count: Number of OSTs to stripe over (0 default, -1 all)\n" \
129         "\tpool_name:    Name of OST pool to use (default none)\n"      \
130         "\tost_indices:  List of OST indices, can be repeated multiple times\n"\
131         "\t              Indices be specified in a format of:\n"        \
132         "\t                -o <ost_1>,<ost_i>-<ost_j>,<ost_n>\n"        \
133         "\t              Or:\n"                                         \
134         "\t                -o <ost_1> -o <ost_i>-<ost_j> -o <ost_n>\n"  \
135         "\t              If --pool is set with --ost, then the OSTs\n"  \
136         "\t              must be the members of the pool."              \
137         "\tcomp_end:     Extent end of the component\n"                 \
138         "\t              Can be specified with k, m or g (in KB, MB and GB\n" \
139         "\t              respectively, -1 for EOF), it must be aligned with\n"\
140         "\t              the stripe_size\n"
141
142 #define SETSTRIPE_USAGE                                         \
143         SSM_CMD_COMMON("setstripe")                             \
144         "                 <directory|filename>\n"               \
145         SSM_HELP_COMMON                                         \
146
147 #define MIGRATE_USAGE                                                   \
148         SSM_CMD_COMMON("migrate  ")                                     \
149         "                 [--block|-b]\n"                               \
150         "                 [--non-block|-n]\n"                           \
151         "                 <filename>\n"                                 \
152         SSM_HELP_COMMON                                                 \
153         "\n"                                                            \
154         "\tblock:        Block file access during data migration (default)\n" \
155         "\tnon-block:    Abort migrations if concurrent access is detected\n" \
156
157 #define SETDIRSTRIPE_USAGE                                      \
158         "               [--mdt-count|-c stripe_count>\n"        \
159         "               [--mdt-index|-i mdt_index]\n"           \
160         "               [--mdt-hash|-H mdt_hash]\n"             \
161         "               [--default|-D] [--mode|-m mode] <dir>\n"        \
162         "\tstripe_count: stripe count of the striped directory\n"       \
163         "\tmdt_index: MDT index of first stripe\n"                      \
164         "\tmdt_hash:  hash type of the striped directory. mdt types:\n" \
165         "       fnv_1a_64 FNV-1a hash algorithm (default)\n"            \
166         "       all_char  sum of characters % MDT_COUNT (not recommended)\n" \
167         "\tdefault_stripe: set default dirstripe of the directory\n"    \
168         "\tmode: the mode of the directory\n"
169
170 static const char       *progname;
171 static bool              file_lease_supported = true;
172
173 /* all available commands */
174 command_t cmdlist[] = {
175         {"setstripe", lfs_setstripe, 0,
176          "Create a new file with a specific striping pattern or\n"
177          "set the default striping pattern on an existing directory or\n"
178          "delete the default striping pattern from an existing directory or\n"
179          "add layout component(s) to an existing composite file or\n"
180          "delete specified component(s) from an existing composite file\n\n"
181          "To delete default striping from an existing directory:\n"
182          "usage: setstripe -d <directory>\n"
183          " or\n"
184          "To delete component(s) from an existing composite file:\n"
185          "usage: setstripe --component-del [--component-id|-I <comp_id>]\n"
186          "                               [--component-flags|-F <comp_flags>]\n"
187          "                               <filename>\n"
188          "\tcomp_id:     Unique component ID\n"
189          "\tcomp_flags:  'init' indicating all instantiated components\n"
190          "\t             '^init' indicating all uninstantiated components\n"
191          "\t-I and -F can't be specified at the same time\n"
192          " or\n"
193          "To add component(s) to an existing composite file:\n"
194          SSM_CMD_COMMON("setstripe --component-add")
195          " or\n"
196          "To create a file with specified striping/composite layout:\n"
197          SETSTRIPE_USAGE},
198         {"getstripe", lfs_getstripe, 0,
199          "To list the striping info for a given file or files in a\n"
200          "directory or recursively for all files in a directory tree.\n"
201          "usage: getstripe [--ost|-O <uuid>] [--quiet|-q] [--verbose|-v]\n"
202          "                 [--stripe-count|-c] [--stripe-index|-i]\n"
203          "                 [--pool|-p] [--stripe-size|-S] [--directory|-d]\n"
204          "                 [--mdt|-m] [--recursive|-r] [--raw|-R] [--yaml|-y]\n"
205          "                 [--layout|-L] [--fid|-F] [--generation|-g]\n"
206          "                 [--component-id[=comp_id]|-I[comp_id]]\n"
207          "                 [--component-flags[=comp_flags]]\n"
208          "                 [--component-count]\n"
209          "                 [--component-start[=[+-]comp_start]]\n"
210          "                 [--component-end[=[+-]comp_end]|-E[[+-]comp_end]]\n"
211          "                 <directory|filename> ..."},
212         {"setdirstripe", lfs_setdirstripe, 0,
213          "To create a striped directory on a specified MDT. This can only\n"
214          "be done on MDT0 with the right of administrator.\n"
215          "usage: setdirstripe [OPTION] <directory>\n"
216          SETDIRSTRIPE_USAGE},
217         {"getdirstripe", lfs_getdirstripe, 0,
218          "To list the striping info for a given directory\n"
219          "or recursively for all directories in a directory tree.\n"
220          "usage: getdirstripe [--obd|-O <uuid>] [--mdt-count|-c]\n"
221          "                    [--mdt-index|-i] [--mdt-hash|-t]\n"
222          "                    [--recursive|-r] [--yaml|-y]\n"
223          "                    [--default|-D] <dir> ..."},
224         {"mkdir", lfs_setdirstripe, 0,
225          "To create a striped directory on a specified MDT. This can only\n"
226          "be done on MDT0 with the right of administrator.\n"
227          "usage: mkdir [OPTION] <directory>\n"
228          SETDIRSTRIPE_USAGE},
229         {"rm_entry", lfs_rmentry, 0,
230          "To remove the name entry of the remote directory. Note: This\n"
231          "command will only delete the name entry, i.e. the remote directory\n"
232          "will become inaccessable after this command. This can only be done\n"
233          "by the administrator\n"
234          "usage: rm_entry <dir>\n"},
235         {"pool_list", lfs_poollist, 0,
236          "List pools or pool OSTs\n"
237          "usage: pool_list <fsname>[.<pool>] | <pathname>\n"},
238         {"find", lfs_find, 0,
239          "find files matching given attributes recursively in directory tree.\n"
240          "usage: find <directory|filename> ...\n"
241          "     [[!] --atime|-A [+-]N] [[!] --ctime|-C [+-]N]\n"
242          "     [[!] --mtime|-M [+-]N] [[!] --mdt|-m <uuid|index,...>]\n"
243          "     [--maxdepth|-D N] [[!] --name|-n <pattern>]\n"
244          "     [[!] --ost|-O <uuid|index,...>] [--print|-p] [--print0|-P]\n"
245          "     [[!] --size|-s [+-]N[bkMGTPE]]\n"
246          "     [[!] --stripe-count|-c [+-]<stripes>]\n"
247          "     [[!] --stripe-index|-i <index,...>]\n"
248          "     [[!] --stripe-size|-S [+-]N[kMGT]] [[!] --type|-t <filetype>]\n"
249          "     [[!] --gid|-g|--group|-G <gid>|<gname>]\n"
250          "     [[!] --uid|-u|--user|-U <uid>|<uname>] [[!] --pool <pool>]\n"
251          "     [[!] --projid <projid>]\n"
252          "     [[!] --layout|-L released,raid0]\n"
253          "     [[!] --component-count [+-]<comp_cnt>]\n"
254          "     [[!] --component-start [+-]N[kMGTPE]]\n"
255          "     [[!] --component-end|-E [+-]N[kMGTPE]]\n"
256          "     [[!] --component-flags <comp_flags>]\n"
257          "     [[!] --mdt-count|-T [+-]<stripes>]\n"
258          "     [[!] --mdt-hash|-H <hashtype>\n"
259          "\t !: used before an option indicates 'NOT' requested attribute\n"
260          "\t -: used before a value indicates less than requested value\n"
261          "\t +: used before a value indicates more than requested value\n"
262          "\tmdt-hash:   hash type of the striped directory.\n"
263          "\t            fnv_1a_64 FNV-1a hash algorithm\n"
264          "\t            all_char  sum of characters % MDT_COUNT\n"},
265         {"check", lfs_check, 0,
266          "Display the status of MDS or OSTs (as specified in the command)\n"
267          "or all the servers (MDS and OSTs).\n"
268          "usage: check <osts|mds|servers>"},
269         {"osts", lfs_osts, 0, "list OSTs connected to client "
270          "[for specified path only]\n" "usage: osts [path]"},
271         {"mdts", lfs_mdts, 0, "list MDTs connected to client "
272          "[for specified path only]\n" "usage: mdts [path]"},
273         {"df", lfs_df, 0,
274          "report filesystem disk space usage or inodes usage"
275          "of each MDS and all OSDs or a batch belonging to a specific pool .\n"
276          "Usage: df [-i] [-h] [--lazy|-l] [--pool|-p <fsname>[.<pool>] [path]"},
277         {"getname", lfs_getname, 0, "list instances and specified mount points "
278          "[for specified path only]\n"
279          "Usage: getname [-h]|[path ...] "},
280 #ifdef HAVE_SYS_QUOTA_H
281         {"setquota", lfs_setquota, 0, "Set filesystem quotas.\n"
282          "usage: setquota <-u|-g|-p> <uname>|<uid>|<gname>|<gid>|<projid>\n"
283          "                -b <block-softlimit> -B <block-hardlimit>\n"
284          "                -i <inode-softlimit> -I <inode-hardlimit> <filesystem>\n"
285          "       setquota <-u|--user|-g|--group|-p|--projid> <uname>|<uid>|<gname>|<gid>|<projid>\n"
286          "                [--block-softlimit <block-softlimit>]\n"
287          "                [--block-hardlimit <block-hardlimit>]\n"
288          "                [--inode-softlimit <inode-softlimit>]\n"
289          "                [--inode-hardlimit <inode-hardlimit>] <filesystem>\n"
290          "       setquota [-t] <-u|--user|-g|--group|-p|--projid>\n"
291          "                [--block-grace <block-grace>]\n"
292          "                [--inode-grace <inode-grace>] <filesystem>\n"
293          "       -b can be used instead of --block-softlimit/--block-grace\n"
294          "       -B can be used instead of --block-hardlimit\n"
295          "       -i can be used instead of --inode-softlimit/--inode-grace\n"
296          "       -I can be used instead of --inode-hardlimit\n\n"
297          "Note: The total quota space will be split into many qunits and\n"
298          "      balanced over all server targets, the minimal qunit size is\n"
299          "      1M bytes for block space and 1K inodes for inode space.\n\n"
300          "      Quota space rebalancing process will stop when this mininum\n"
301          "      value is reached. As a result, quota exceeded can be returned\n"
302          "      while many targets still have 1MB or 1K inodes of spare\n"
303          "      quota space."},
304         {"quota", lfs_quota, 0, "Display disk usage and limits.\n"
305          "usage: quota [-q] [-v] [-h] [-o <obd_uuid>|-i <mdt_idx>|-I "
306                        "<ost_idx>]\n"
307          "             [<-u|-g|-p> <uname>|<uid>|<gname>|<gid>|<projid>] <filesystem>\n"
308          "       quota [-o <obd_uuid>|-i <mdt_idx>|-I <ost_idx>] -t <-u|-g|-p> <filesystem>"},
309 #endif
310         {"flushctx", lfs_flushctx, 0, "Flush security context for current user.\n"
311          "usage: flushctx [-k] [mountpoint...]"},
312         {"cp", lfs_cp, 0,
313          "Remote user copy files and directories.\n"
314          "usage: cp [OPTION]... [-T] SOURCE DEST\n\tcp [OPTION]... SOURCE... DIRECTORY\n\tcp [OPTION]... -t DIRECTORY SOURCE..."},
315         {"ls", lfs_ls, 0,
316          "Remote user list directory contents.\n"
317          "usage: ls [OPTION]... [FILE]..."},
318         {"changelog", lfs_changelog, 0,
319          "Show the metadata changes on an MDT."
320          "\nusage: changelog <mdtname> [startrec [endrec]]"},
321         {"changelog_clear", lfs_changelog_clear, 0,
322          "Indicate that old changelog records up to <endrec> are no longer of "
323          "interest to consumer <id>, allowing the system to free up space.\n"
324          "An <endrec> of 0 means all records.\n"
325          "usage: changelog_clear <mdtname> <id> <endrec>"},
326         {"fid2path", lfs_fid2path, 0,
327          "Resolve the full path(s) for given FID(s). For a specific hardlink "
328          "specify link number <linkno>.\n"
329         /* "For a historical link name, specify changelog record <recno>.\n" */
330          "usage: fid2path [--link <linkno>] <fsname|rootpath> <fid> ..."
331                 /* [ --rec <recno> ] */ },
332         {"path2fid", lfs_path2fid, 0, "Display the fid(s) for a given path(s).\n"
333          "usage: path2fid [--parents] <path> ..."},
334         {"data_version", lfs_data_version, 0, "Display file data version for "
335          "a given path.\n" "usage: data_version -[n|r|w] <path>"},
336         {"hsm_state", lfs_hsm_state, 0, "Display the HSM information (states, "
337          "undergoing actions) for given files.\n usage: hsm_state <file> ..."},
338         {"hsm_set", lfs_hsm_set, 0, "Set HSM user flag on specified files.\n"
339          "usage: hsm_set [--norelease] [--noarchive] [--dirty] [--exists] "
340          "[--archived] [--lost] <file> ..."},
341         {"hsm_clear", lfs_hsm_clear, 0, "Clear HSM user flag on specified "
342          "files.\n"
343          "usage: hsm_clear [--norelease] [--noarchive] [--dirty] [--exists] "
344          "[--archived] [--lost] <file> ..."},
345         {"hsm_action", lfs_hsm_action, 0, "Display current HSM request for "
346          "given files.\n" "usage: hsm_action <file> ..."},
347         {"hsm_archive", lfs_hsm_archive, 0,
348          "Archive file to external storage.\n"
349          "usage: hsm_archive [--filelist FILELIST] [--data DATA] [--archive NUM] "
350          "<file> ..."},
351         {"hsm_restore", lfs_hsm_restore, 0,
352          "Restore file from external storage.\n"
353          "usage: hsm_restore [--filelist FILELIST] [--data DATA] <file> ..."},
354         {"hsm_release", lfs_hsm_release, 0,
355          "Release files from Lustre.\n"
356          "usage: hsm_release [--filelist FILELIST] [--data DATA] <file> ..."},
357         {"hsm_remove", lfs_hsm_remove, 0,
358          "Remove file copy from external storage.\n"
359          "usage: hsm_remove [--filelist FILELIST] [--data DATA]\n"
360          "                  [--mntpath MOUNTPATH] [--archive NUM] <file|FID> ...\n"
361          "\n"
362          "Note: To remove files from the archive that have been deleted on\n"
363          "Lustre, set mntpath and optionally archive. In that case, all the\n"
364          "positional arguments and entries in the file list must be FIDs."
365         },
366         {"hsm_cancel", lfs_hsm_cancel, 0,
367          "Cancel requests related to specified files.\n"
368          "usage: hsm_cancel [--filelist FILELIST] [--data DATA] <file> ..."},
369         {"swap_layouts", lfs_swap_layouts, 0, "Swap layouts between 2 files.\n"
370          "usage: swap_layouts <path1> <path2>"},
371         {"migrate", lfs_setstripe, 0,
372          "migrate a directory between MDTs.\n"
373          "usage: migrate --mdt-index <mdt_idx> [--verbose|-v] "
374          "<directory>\n"
375          "\tmdt_idx:      index of the destination MDT\n"
376          "\n"
377          "migrate file objects from one OST "
378          "layout\nto another (may be not safe with concurent writes).\n"
379          "usage: migrate  "
380          "[--stripe-count|-c] <stripe_count>\n"
381          "              [--stripe-index|-i] <start_ost_index>\n"
382          "              [--stripe-size|-S] <stripe_size>\n"
383          "              [--pool|-p] <pool_name>\n"
384          "              [--ost-list|-o] <ost_indices>\n"
385          "              [--block|-b]\n"
386          "              [--non-block|-n]\n"
387          "              <file|directory>\n"
388          "\tstripe_count:     number of OSTs to stripe a file over\n"
389          "\tstripe_ost_index: index of the first OST to stripe a file over\n"
390          "\tstripe_size:      number of bytes to store before moving to the next OST\n"
391          "\tpool_name:        name of the predefined pool of OSTs\n"
392          "\tost_indices:      OSTs to stripe over, in order\n"
393          "\tblock:            wait for the operation to return before continuing\n"
394          "\tnon-block:        do not wait for the operation to return.\n"},
395         {"mv", lfs_mv, 0,
396          "To move directories between MDTs. This command is deprecated, "
397          "use \"migrate\" instead.\n"
398          "usage: mv <directory|filename> [--mdt-index|-M] <mdt_index> "
399          "[--verbose|-v]\n"},
400         {"ladvise", lfs_ladvise, 0,
401          "Provide servers with advice about access patterns for a file.\n"
402          "usage: ladvise [--advice|-a ADVICE] [--start|-s START[kMGT]]\n"
403          "               [--background|-b]\n"
404          "               {[--end|-e END[kMGT]] | [--length|-l LENGTH[kMGT]]}\n"
405          "               <file> ...\n"},
406         {"help", Parser_help, 0, "help"},
407         {"exit", Parser_quit, 0, "quit"},
408         {"quit", Parser_quit, 0, "quit"},
409         {"--version", Parser_version, 0,
410          "output build version of the utility and exit"},
411         {"--list-commands", lfs_list_commands, 0,
412          "list commands supported by the utility and exit"},
413         { 0, 0, 0, NULL }
414 };
415
416
417 #define MIGRATION_NONBLOCK      1
418
419 static int check_hashtype(const char *hashtype)
420 {
421         int i;
422
423         for (i = LMV_HASH_TYPE_ALL_CHARS; i < LMV_HASH_TYPE_MAX; i++)
424                 if (strcmp(hashtype, mdt_hash_name[i]) == 0)
425                         return i;
426
427         return 0;
428 }
429
430 /**
431  * Internal helper for migrate_copy_data(). Check lease and report error if
432  * need be.
433  *
434  * \param[in]  fd           File descriptor on which to check the lease.
435  * \param[out] lease_broken Set to true if the lease was broken.
436  * \param[in]  group_locked Whether a group lock was taken or not.
437  * \param[in]  path         Name of the file being processed, for error
438  *                          reporting
439  *
440  * \retval 0       Migration can keep on going.
441  * \retval -errno  Error occurred, abort migration.
442  */
443 static int check_lease(int fd, bool *lease_broken, bool group_locked,
444                        const char *path)
445 {
446         int rc;
447
448         if (!file_lease_supported)
449                 return 0;
450
451         rc = llapi_lease_check(fd);
452         if (rc > 0)
453                 return 0; /* llapi_check_lease returns > 0 on success. */
454
455         if (!group_locked) {
456                 fprintf(stderr, "%s: cannot migrate '%s': file busy\n",
457                         progname, path);
458                 rc = rc ? rc : -EAGAIN;
459         } else {
460                 fprintf(stderr, "%s: external attempt to access file '%s' "
461                         "blocked until migration ends.\n", progname, path);
462                 rc = 0;
463         }
464         *lease_broken = true;
465         return rc;
466 }
467
468 static int migrate_copy_data(int fd_src, int fd_dst, size_t buf_size,
469                              bool group_locked, const char *fname)
470 {
471         void    *buf = NULL;
472         ssize_t  rsize = -1;
473         ssize_t  wsize = 0;
474         size_t   rpos = 0;
475         size_t   wpos = 0;
476         off_t    bufoff = 0;
477         int      rc;
478         bool     lease_broken = false;
479
480         /* Use a page-aligned buffer for direct I/O */
481         rc = posix_memalign(&buf, getpagesize(), buf_size);
482         if (rc != 0)
483                 return -rc;
484
485         while (1) {
486                 /* read new data only if we have written all
487                  * previously read data */
488                 if (wpos == rpos) {
489                         if (!lease_broken) {
490                                 rc = check_lease(fd_src, &lease_broken,
491                                                  group_locked, fname);
492                                 if (rc < 0)
493                                         goto out;
494                         }
495                         rsize = read(fd_src, buf, buf_size);
496                         if (rsize < 0) {
497                                 rc = -errno;
498                                 fprintf(stderr, "%s: %s: read failed: %s\n",
499                                         progname, fname, strerror(-rc));
500                                 goto out;
501                         }
502                         rpos += rsize;
503                         bufoff = 0;
504                 }
505                 /* eof ? */
506                 if (rsize == 0)
507                         break;
508
509                 wsize = write(fd_dst, buf + bufoff, rpos - wpos);
510                 if (wsize < 0) {
511                         rc = -errno;
512                         fprintf(stderr,
513                                 "%s: %s: write failed on volatile: %s\n",
514                                 progname, fname, strerror(-rc));
515                         goto out;
516                 }
517                 wpos += wsize;
518                 bufoff += wsize;
519         }
520
521         rc = fsync(fd_dst);
522         if (rc < 0) {
523                 rc = -errno;
524                 fprintf(stderr, "%s: %s: fsync failed: %s\n",
525                         progname, fname, strerror(-rc));
526         }
527
528 out:
529         free(buf);
530         return rc;
531 }
532
533 static int migrate_copy_timestamps(int fdv, const struct stat *st)
534 {
535         struct timeval  tv[2] = {
536                 {.tv_sec = st->st_atime},
537                 {.tv_sec = st->st_mtime}
538         };
539
540         return futimes(fdv, tv);
541 }
542
543 static int migrate_block(int fd, int fdv, const struct stat *st,
544                          size_t buf_size, const char *name)
545 {
546         __u64   dv1;
547         int     gid;
548         int     rc;
549         int     rc2;
550
551         rc = llapi_get_data_version(fd, &dv1, LL_DV_RD_FLUSH);
552         if (rc < 0) {
553                 fprintf(stderr, "%s: %s: cannot get dataversion: %s\n",
554                         progname, name, strerror(-rc));
555                 return rc;
556         }
557
558         do
559                 gid = random();
560         while (gid == 0);
561
562         /* The grouplock blocks all concurrent accesses to the file.
563          * It has to be taken after llapi_get_data_version as it would
564          * block it too. */
565         rc = llapi_group_lock(fd, gid);
566         if (rc < 0) {
567                 fprintf(stderr, "%s: %s: cannot get group lock: %s\n",
568                         progname, name, strerror(-rc));
569                 return rc;
570         }
571
572         rc = migrate_copy_data(fd, fdv, buf_size, true, name);
573         if (rc < 0) {
574                 fprintf(stderr, "%s: %s: data copy failed\n", progname, name);
575                 goto out_unlock;
576         }
577
578         /* Make sure we keep original atime/mtime values */
579         rc = migrate_copy_timestamps(fdv, st);
580         if (rc < 0) {
581                 fprintf(stderr, "%s: %s: timestamp copy failed\n",
582                         progname, name);
583                 goto out_unlock;
584         }
585
586         /* swap layouts
587          * for a migration we need to check data version on file did
588          * not change.
589          *
590          * Pass in gid=0 since we already own grouplock. */
591         rc = llapi_fswap_layouts_grouplock(fd, fdv, dv1, 0, 0,
592                                            SWAP_LAYOUTS_CHECK_DV1);
593         if (rc == -EAGAIN) {
594                 fprintf(stderr, "%s: %s: dataversion changed during copy, "
595                         "migration aborted\n", progname, name);
596                 goto out_unlock;
597         } else if (rc < 0) {
598                 fprintf(stderr, "%s: %s: cannot swap layouts: %s\n", progname,
599                         name, strerror(-rc));
600                 goto out_unlock;
601         }
602
603 out_unlock:
604         rc2 = llapi_group_unlock(fd, gid);
605         if (rc2 < 0 && rc == 0) {
606                 fprintf(stderr, "%s: %s: putting group lock failed: %s\n",
607                         progname, name, strerror(-rc2));
608                 rc = rc2;
609         }
610
611         return rc;
612 }
613
614 static int migrate_nonblock(int fd, int fdv, const struct stat *st,
615                             size_t buf_size, const char *name)
616 {
617         __u64   dv1;
618         __u64   dv2;
619         int     rc;
620
621         rc = llapi_get_data_version(fd, &dv1, LL_DV_RD_FLUSH);
622         if (rc < 0) {
623                 fprintf(stderr, "%s: %s: cannot get data version: %s\n",
624                         progname, name, strerror(-rc));
625                 return rc;
626         }
627
628         rc = migrate_copy_data(fd, fdv, buf_size, false, name);
629         if (rc < 0) {
630                 fprintf(stderr, "%s: %s: data copy failed\n", progname, name);
631                 return rc;
632         }
633
634         rc = llapi_get_data_version(fd, &dv2, LL_DV_RD_FLUSH);
635         if (rc != 0) {
636                 fprintf(stderr, "%s: %s: cannot get data version: %s\n",
637                         progname, name, strerror(-rc));
638                 return rc;
639         }
640
641         if (dv1 != dv2) {
642                 rc = -EAGAIN;
643                 fprintf(stderr, "%s: %s: data version changed during "
644                                 "migration\n",
645                         progname, name);
646                 return rc;
647         }
648
649         /* Make sure we keep original atime/mtime values */
650         rc = migrate_copy_timestamps(fdv, st);
651         if (rc < 0) {
652                 fprintf(stderr, "%s: %s: timestamp copy failed\n",
653                         progname, name);
654                 return rc;
655         }
656
657         /* Atomically put lease, swap layouts and close.
658          * for a migration we need to check data version on file did
659          * not change. */
660         rc = llapi_fswap_layouts(fd, fdv, 0, 0, SWAP_LAYOUTS_CLOSE);
661         if (rc < 0) {
662                 fprintf(stderr, "%s: %s: cannot swap layouts: %s\n",
663                         progname, name, strerror(-rc));
664                 return rc;
665         }
666
667         return 0;
668 }
669
670 static int lfs_component_set(char *fname, int comp_id, __u32 flags)
671 {
672         return -ENOTSUP;
673 }
674
675 static int lfs_component_del(char *fname, __u32 comp_id, __u32 flags)
676 {
677         int     rc = 0;
678
679         if (flags != 0 && comp_id != 0)
680                 return -EINVAL;
681
682         /* LCME_FL_INIT is the only supported flag in PFL */
683         if (flags != 0) {
684                 if (flags & ~LCME_KNOWN_FLAGS) {
685                         fprintf(stderr, "Invalid component flags %#x\n", flags);
686                         return -EINVAL;
687                 }
688         } else if (comp_id > LCME_ID_MAX) {
689                 fprintf(stderr, "Invalid component id %u\n", comp_id);
690                 return -EINVAL;
691         }
692
693         rc = llapi_layout_file_comp_del(fname, comp_id, flags);
694         if (rc)
695                 fprintf(stderr, "Delete component %#x from %s failed. %s\n",
696                         comp_id, fname, strerror(errno));
697         return rc;
698 }
699
700 static int lfs_component_add(char *fname, struct llapi_layout *layout)
701 {
702         int     rc;
703
704         if (layout == NULL)
705                 return -EINVAL;
706
707         rc = llapi_layout_file_comp_add(fname, layout);
708         if (rc)
709                 fprintf(stderr, "Add layout component(s) to %s failed. %s\n",
710                         fname, strerror(errno));
711         return rc;
712 }
713
714 static int lfs_component_create(char *fname, int open_flags, mode_t open_mode,
715                                 struct llapi_layout *layout)
716 {
717         struct stat     st;
718         int     fd;
719
720         if (layout == NULL)
721                 return -EINVAL;
722
723         fd = lstat(fname, &st);
724         if (fd == 0 && S_ISDIR(st.st_mode))
725                 open_flags = O_DIRECTORY | O_RDONLY;
726
727         fd = llapi_layout_file_open(fname, open_flags, open_mode, layout);
728         if (fd < 0)
729                 fprintf(stderr, "%s %s failed. %s\n",
730                         S_ISDIR(st.st_mode) ?
731                                 "Set default composite layout to " :
732                                 "Create composite file",
733                         fname, strerror(errno));
734         return fd;
735 }
736
737 static int lfs_migrate(char *name, __u64 migration_flags,
738                        struct llapi_stripe_param *param,
739                        struct llapi_layout *layout)
740 {
741         int                      fd = -1;
742         int                      fdv = -1;
743         char                     parent[PATH_MAX];
744         int                      mdt_index;
745         int                      random_value;
746         char                     volatile_file[sizeof(parent) +
747                                                LUSTRE_VOLATILE_HDR_LEN +
748                                                2 * sizeof(mdt_index) +
749                                                2 * sizeof(random_value) + 4];
750         char                    *ptr;
751         int                      rc;
752         struct lov_user_md      *lum = NULL;
753         int                      lum_size;
754         int                      buf_size = 1024 * 1024 * 4;
755         bool                     have_lease_rdlck = false;
756         struct stat              st;
757         struct stat              stv;
758
759         /* find the right size for the IO and allocate the buffer */
760         lum_size = lov_user_md_size(LOV_MAX_STRIPE_COUNT, LOV_USER_MAGIC_V3);
761         lum = malloc(lum_size);
762         if (lum == NULL) {
763                 rc = -ENOMEM;
764                 goto free;
765         }
766
767         rc = llapi_file_get_stripe(name, lum);
768         /* failure can happen for many reasons and some may be not real errors
769          * (eg: no stripe)
770          * in case of a real error, a later call will fail with better
771          * error management */
772         if (rc == 0) {
773                 if ((lum->lmm_magic == LOV_USER_MAGIC_V1 ||
774                      lum->lmm_magic == LOV_USER_MAGIC_V3) &&
775                     lum->lmm_stripe_size != 0)
776                         buf_size = lum->lmm_stripe_size;
777         }
778
779         /* open file, direct io */
780         /* even if the file is only read, WR mode is nedeed to allow
781          * layout swap on fd */
782         fd = open(name, O_RDWR | O_DIRECT);
783         if (fd == -1) {
784                 rc = -errno;
785                 fprintf(stderr, "%s: %s: cannot open: %s\n", progname, name,
786                         strerror(-rc));
787                 goto free;
788         }
789
790         if (file_lease_supported) {
791                 rc = llapi_lease_get(fd, LL_LEASE_RDLCK);
792                 if (rc == -EOPNOTSUPP) {
793                         /* Older servers do not support file lease.
794                          * Disable related checks. This opens race conditions
795                          * as explained in LU-4840 */
796                         file_lease_supported = false;
797                 } else if (rc < 0) {
798                         fprintf(stderr, "%s: %s: cannot get open lease: %s\n",
799                                 progname, name, strerror(-rc));
800                         goto error;
801                 } else {
802                         have_lease_rdlck = true;
803                 }
804         }
805
806         /* search for file directory pathname */
807         if (strlen(name) > sizeof(parent)-1) {
808                 rc = -E2BIG;
809                 goto error;
810         }
811         strncpy(parent, name, sizeof(parent));
812         ptr = strrchr(parent, '/');
813         if (ptr == NULL) {
814                 if (getcwd(parent, sizeof(parent)) == NULL) {
815                         rc = -errno;
816                         goto error;
817                 }
818         } else {
819                 if (ptr == parent)
820                         strcpy(parent, "/");
821                 else
822                         *ptr = '\0';
823         }
824
825         rc = llapi_file_fget_mdtidx(fd, &mdt_index);
826         if (rc < 0) {
827                 fprintf(stderr, "%s: %s: cannot get MDT index: %s\n",
828                         progname, name, strerror(-rc));
829                 goto error;
830         }
831
832         do {
833                 int open_flags = O_WRONLY | O_CREAT | O_EXCL | O_NOFOLLOW;
834                 mode_t open_mode = S_IRUSR | S_IWUSR;
835
836                 random_value = random();
837                 rc = snprintf(volatile_file, sizeof(volatile_file),
838                               "%s/%s:%.4X:%.4X", parent, LUSTRE_VOLATILE_HDR,
839                               mdt_index, random_value);
840                 if (rc >= sizeof(volatile_file)) {
841                         rc = -E2BIG;
842                         goto error;
843                 }
844
845                 /* create, open a volatile file, use caching (ie no directio) */
846                 if (param != NULL)
847                         fdv = llapi_file_open_param(volatile_file, open_flags,
848                                                     open_mode, param);
849                 else if (layout != NULL)
850                         fdv = lfs_component_create(volatile_file, open_flags,
851                                                    open_mode, layout);
852                 else
853                         fdv = -EINVAL;
854         } while (fdv == -EEXIST);
855
856         if (fdv < 0) {
857                 rc = fdv;
858                 fprintf(stderr, "%s: %s: cannot create volatile file in"
859                                 " directory: %s\n",
860                         progname, parent, strerror(-rc));
861                 goto error;
862         }
863
864         /* In case the MDT does not support creation of volatile files
865          * we should try to unlink it. */
866         (void)unlink(volatile_file);
867
868         /* Not-owner (root?) special case.
869          * Need to set owner/group of volatile file like original.
870          * This will allow to pass related check during layout_swap.
871          */
872         rc = fstat(fd, &st);
873         if (rc != 0) {
874                 rc = -errno;
875                 fprintf(stderr, "%s: %s: cannot stat: %s\n", progname, name,
876                         strerror(errno));
877                 goto error;
878         }
879         rc = fstat(fdv, &stv);
880         if (rc != 0) {
881                 rc = -errno;
882                 fprintf(stderr, "%s: %s: cannot stat: %s\n", progname,
883                         volatile_file, strerror(errno));
884                 goto error;
885         }
886         if (st.st_uid != stv.st_uid || st.st_gid != stv.st_gid) {
887                 rc = fchown(fdv, st.st_uid, st.st_gid);
888                 if (rc != 0) {
889                         rc = -errno;
890                         fprintf(stderr, "%s: %s: cannot chown: %s\n", progname,
891                                 name, strerror(errno));
892                         goto error;
893                 }
894         }
895
896         if (migration_flags & MIGRATION_NONBLOCK && file_lease_supported) {
897                 rc = migrate_nonblock(fd, fdv, &st, buf_size, name);
898                 if (rc == 0) {
899                         have_lease_rdlck = false;
900                         fdv = -1; /* The volatile file is closed as we put the
901                                    * lease in non-blocking mode. */
902                 }
903         } else {
904                 /* Blocking mode (forced if servers do not support file lease).
905                  * It is also the default mode, since we cannot distinguish
906                  * between a broken lease and a server that does not support
907                  * atomic swap/close (LU-6785) */
908                 rc = migrate_block(fd, fdv, &st, buf_size, name);
909         }
910
911 error:
912         if (have_lease_rdlck)
913                 llapi_lease_put(fd);
914
915         if (fd >= 0)
916                 close(fd);
917
918         if (fdv >= 0)
919                 close(fdv);
920
921 free:
922         if (lum)
923                 free(lum);
924
925         return rc;
926 }
927
928 /**
929  * Parse a string containing an OST index list into an array of integers.
930  *
931  * The input string contains a comma delimited list of individual
932  * indices and ranges, for example "1,2-4,7". Add the indices into the
933  * \a osts array and remove duplicates.
934  *
935  * \param[out] osts    array to store indices in
936  * \param[in] size     size of \a osts array
937  * \param[in] offset   starting index in \a osts
938  * \param[in] arg      string containing OST index list
939  *
940  * \retval positive    number of indices in \a osts
941  * \retval -EINVAL     unable to parse \a arg
942  */
943 static int parse_targets(__u32 *osts, int size, int offset, char *arg)
944 {
945         int rc;
946         int nr = offset;
947         int slots = size - offset;
948         char *ptr = NULL;
949         bool end_of_loop;
950
951         if (arg == NULL)
952                 return -EINVAL;
953
954         end_of_loop = false;
955         while (!end_of_loop) {
956                 int start_index;
957                 int end_index;
958                 int i;
959                 char *endptr = NULL;
960
961                 rc = -EINVAL;
962
963                 ptr = strchrnul(arg, ',');
964
965                 end_of_loop = *ptr == '\0';
966                 *ptr = '\0';
967
968                 start_index = strtol(arg, &endptr, 0);
969                 if (endptr == arg) /* no data at all */
970                         break;
971                 if (*endptr != '-' && *endptr != '\0') /* has invalid data */
972                         break;
973                 if (start_index < 0)
974                         break;
975
976                 end_index = start_index;
977                 if (*endptr == '-') {
978                         end_index = strtol(endptr + 1, &endptr, 0);
979                         if (*endptr != '\0')
980                                 break;
981                         if (end_index < start_index)
982                                 break;
983                 }
984
985                 for (i = start_index; i <= end_index && slots > 0; i++) {
986                         int j;
987
988                         /* remove duplicate */
989                         for (j = 0; j < offset; j++) {
990                                 if (osts[j] == i)
991                                         break;
992                         }
993                         if (j == offset) { /* no duplicate */
994                                 osts[nr++] = i;
995                                 --slots;
996                         }
997                 }
998                 if (slots == 0 && i < end_index)
999                         break;
1000
1001                 *ptr = ',';
1002                 arg = ++ptr;
1003                 offset = nr;
1004                 rc = 0;
1005         }
1006         if (!end_of_loop && ptr != NULL)
1007                 *ptr = ',';
1008
1009         return rc < 0 ? rc : nr;
1010 }
1011
1012 static int verify_pool_name(char *prog_name, char *pool_name)
1013 {
1014         char *ptr;
1015
1016         if (pool_name == NULL)
1017                 return 0;
1018
1019         ptr = strchr(pool_name, '.');
1020         if (ptr != NULL && ptr == pool_name) {
1021                 fprintf(stderr, "error: %s: fsname is empty in pool name '%s'\n",
1022                         prog_name, pool_name);
1023                 return -EINVAL;
1024         }
1025
1026         return 0;
1027 }
1028
1029 struct lfs_setstripe_args {
1030         unsigned long long       lsa_comp_end;
1031         unsigned long long       lsa_stripe_size;
1032         int                      lsa_stripe_count;
1033         int                      lsa_stripe_off;
1034         __u32                    lsa_comp_flags;
1035         int                      lsa_nr_osts;
1036         __u32                   *lsa_osts;
1037         char                    *lsa_pool_name;
1038 };
1039
1040 static inline void setstripe_args_init(struct lfs_setstripe_args *lsa)
1041 {
1042         memset(lsa, 0, sizeof(*lsa));
1043         lsa->lsa_stripe_off = -1;
1044 }
1045
1046 static inline bool setstripe_args_specified(struct lfs_setstripe_args *lsa)
1047 {
1048         return (lsa->lsa_stripe_size != 0 || lsa->lsa_stripe_count != 0 ||
1049                 lsa->lsa_stripe_off != -1 || lsa->lsa_pool_name != NULL ||
1050                 lsa->lsa_comp_end != 0);
1051 }
1052
1053 static int comp_args_to_layout(struct llapi_layout **composite,
1054                                struct lfs_setstripe_args *lsa)
1055 {
1056         struct llapi_layout *layout = *composite;
1057         uint64_t prev_end = 0;
1058         int i = 0, rc;
1059
1060         if (layout == NULL) {
1061                 layout = llapi_layout_alloc();
1062                 if (layout == NULL) {
1063                         fprintf(stderr, "Alloc llapi_layout failed. %s\n",
1064                                 strerror(errno));
1065                         return -ENOMEM;
1066                 }
1067                 *composite = layout;
1068         } else {
1069                 uint64_t start;
1070
1071                 /* Get current component extent, current component
1072                  * must be the tail component. */
1073                 rc = llapi_layout_comp_extent_get(layout, &start, &prev_end);
1074                 if (rc) {
1075                         fprintf(stderr, "Get comp extent failed. %s\n",
1076                                 strerror(errno));
1077                         return rc;
1078                 }
1079
1080                 rc = llapi_layout_comp_add(layout);
1081                 if (rc) {
1082                         fprintf(stderr, "Add component failed. %s\n",
1083                                 strerror(errno));
1084                         return rc;
1085                 }
1086         }
1087
1088         rc = llapi_layout_comp_extent_set(layout, prev_end, lsa->lsa_comp_end);
1089         if (rc) {
1090                 fprintf(stderr, "Set extent [%lu, %llu) failed. %s\n",
1091                         prev_end, lsa->lsa_comp_end, strerror(errno));
1092                 return rc;
1093         }
1094
1095         if (lsa->lsa_stripe_size != 0) {
1096                 rc = llapi_layout_stripe_size_set(layout,
1097                                                   lsa->lsa_stripe_size);
1098                 if (rc) {
1099                         fprintf(stderr, "Set stripe size %llu failed. %s\n",
1100                                 lsa->lsa_stripe_size, strerror(errno));
1101                         return rc;
1102                 }
1103         }
1104
1105         if (lsa->lsa_stripe_count != 0) {
1106                 rc = llapi_layout_stripe_count_set(layout,
1107                                                    lsa->lsa_stripe_count == -1 ?
1108                                                    LLAPI_LAYOUT_WIDE :
1109                                                    lsa->lsa_stripe_count);
1110                 if (rc) {
1111                         fprintf(stderr, "Set stripe count %d failed. %s\n",
1112                                 lsa->lsa_stripe_count, strerror(errno));
1113                         return rc;
1114                 }
1115         }
1116
1117         if (lsa->lsa_pool_name != NULL) {
1118                 rc = llapi_layout_pool_name_set(layout, lsa->lsa_pool_name);
1119                 if (rc) {
1120                         fprintf(stderr, "Set pool name: %s failed. %s\n",
1121                                 lsa->lsa_pool_name, strerror(errno));
1122                         return rc;
1123                 }
1124         }
1125
1126         if (lsa->lsa_nr_osts > 0) {
1127                 if (lsa->lsa_stripe_count > 0 &&
1128                     lsa->lsa_nr_osts != lsa->lsa_stripe_count) {
1129                         fprintf(stderr, "stripe_count(%d) != nr_osts(%d)\n",
1130                                 lsa->lsa_stripe_count, lsa->lsa_nr_osts);
1131                         return -EINVAL;
1132                 }
1133                 for (i = 0; i < lsa->lsa_nr_osts; i++) {
1134                         rc = llapi_layout_ost_index_set(layout, i,
1135                                                         lsa->lsa_osts[i]);
1136                         if (rc)
1137                                 break;
1138                 }
1139         } else if (lsa->lsa_stripe_off != -1) {
1140                 rc = llapi_layout_ost_index_set(layout, 0, lsa->lsa_stripe_off);
1141         }
1142         if (rc) {
1143                 fprintf(stderr, "Set ost index %d failed. %s\n",
1144                         i, strerror(errno));
1145                 return rc;
1146         }
1147
1148         return 0;
1149 }
1150
1151 /* In 'lfs setstripe --component-add' mode, we need to fetch the extent
1152  * end of the last component in the existing file, and adjust the
1153  * first extent start of the components to be added accordingly. */
1154 static int adjust_first_extent(char *fname, struct llapi_layout *layout)
1155 {
1156         struct llapi_layout *head;
1157         uint64_t start, end, stripe_size, prev_end = 0;
1158         int rc;
1159
1160         if (layout == NULL)
1161                 return -EINVAL;
1162
1163         errno = 0;
1164         head = llapi_layout_get_by_path(fname, 0);
1165         if (head == NULL) {
1166                 fprintf(stderr, "Read layout from %s failed. %s\n",
1167                         fname, strerror(errno));
1168                 return -EINVAL;
1169         } else if (errno == ENODATA) {
1170                 /* file without LOVEA, this component-add will be turned
1171                  * into a component-create. */
1172                 llapi_layout_free(head);
1173                 return -ENODATA;
1174         } else {
1175                 /* Current component of 'head' should be tail of component
1176                  * list by default, but we do an extra move cursor operation
1177                  * here to test if the layout is non-composite. */
1178                 rc = llapi_layout_comp_use(head, LLAPI_LAYOUT_COMP_USE_LAST);
1179                 if (rc < 0) {
1180                         fprintf(stderr, "'%s' isn't a composite file?\n",
1181                                 fname);
1182                         llapi_layout_free(head);
1183                         return rc;
1184                 }
1185         }
1186
1187         rc = llapi_layout_comp_extent_get(head, &start, &prev_end);
1188         if (rc) {
1189                 fprintf(stderr, "Get prev extent failed. %s\n",
1190                         strerror(errno));
1191                 llapi_layout_free(head);
1192                 return rc;
1193         }
1194
1195         llapi_layout_free(head);
1196
1197         /* Make sure we use the first component of the layout to be added. */
1198         rc = llapi_layout_comp_use(layout, LLAPI_LAYOUT_COMP_USE_FIRST);
1199         if (rc < 0) {
1200                 fprintf(stderr, "Move component cursor failed. %s\n",
1201                         strerror(errno));
1202                 return rc;
1203         }
1204
1205         rc = llapi_layout_comp_extent_get(layout, &start, &end);
1206         if (rc) {
1207                 fprintf(stderr, "Get extent failed. %s\n", strerror(errno));
1208                 return rc;
1209         }
1210
1211         if (start > prev_end || end <= prev_end) {
1212                 fprintf(stderr, "First extent to be set [%lu, %lu) isn't "
1213                         "adjacent with the existing file extent end: %lu\n",
1214                         start, end, prev_end);
1215                 return -EINVAL;
1216         }
1217
1218         rc = llapi_layout_stripe_size_get(layout, &stripe_size);
1219         if (rc) {
1220                 fprintf(stderr, "Get stripe size failed. %s\n",
1221                         strerror(errno));
1222                 return rc;
1223         }
1224
1225         if (stripe_size != LLAPI_LAYOUT_DEFAULT &&
1226             (prev_end & (stripe_size - 1))) {
1227                 fprintf(stderr, "Stripe size %lu not aligned with %lu\n",
1228                         stripe_size, prev_end);
1229                 return -EINVAL;
1230         }
1231
1232         rc = llapi_layout_comp_extent_set(layout, prev_end, end);
1233         if (rc) {
1234                 fprintf(stderr, "Set component extent [%lu, %lu) failed. %s\n",
1235                         prev_end, end, strerror(errno));
1236                 return rc;
1237         }
1238
1239         return 0;
1240 }
1241
1242 static inline bool comp_flags_is_neg(__u32 flags)
1243 {
1244         return flags & LCME_FL_NEG;
1245 }
1246
1247 static inline void comp_flags_set_neg(__u32 *flags)
1248 {
1249         *flags |= LCME_FL_NEG;
1250 }
1251
1252 static inline void comp_flags_clear_neg(__u32 *flags)
1253 {
1254         *flags &= ~LCME_FL_NEG;
1255 }
1256
1257 static int comp_str2flags(__u32 *flags, char *string)
1258 {
1259         char *name;
1260         __u32 neg_flags = 0;
1261
1262         if (string == NULL)
1263                 return -EINVAL;
1264
1265         *flags = 0;
1266         for (name = strtok(string, ","); name; name = strtok(NULL, ",")) {
1267                 bool found = false;
1268                 int i;
1269
1270                 for (i = 0; i < ARRAY_SIZE(comp_flags_table); i++) {
1271                         __u32 comp_flag = comp_flags_table[i].cfn_flag;
1272                         const char *comp_name = comp_flags_table[i].cfn_name;
1273
1274                         if (strcmp(name, comp_name) == 0) {
1275                                 *flags |= comp_flag;
1276                                 found = true;
1277                         } else if (strncmp(name, "^", 1) == 0 &&
1278                                    strcmp(name + 1, comp_name) == 0) {
1279                                 neg_flags |= comp_flag;
1280                                 found = true;
1281                         }
1282                 }
1283                 if (!found) {
1284                         llapi_printf(LLAPI_MSG_ERROR, "Component flag "
1285                                      "'%s' is not supported.\n", name);
1286                         return -EINVAL;
1287                 }
1288         }
1289
1290         if (*flags == 0 && neg_flags == 0)
1291                 return -EINVAL;
1292         /* don't support mixed flags for now */
1293         if (*flags && neg_flags)
1294                 return -EINVAL;
1295
1296         if (neg_flags) {
1297                 *flags = neg_flags;
1298                 comp_flags_set_neg(flags);
1299         }
1300
1301         return 0;
1302 }
1303
1304 static inline bool arg_is_eof(char *arg)
1305 {
1306         return !strncmp(arg, "-1", strlen("-1")) ||
1307                !strncmp(arg, "EOF", strlen("EOF")) ||
1308                !strncmp(arg, "eof", strlen("eof"));
1309 }
1310
1311 enum {
1312         LFS_POOL_OPT = 3,
1313         LFS_COMP_COUNT_OPT,
1314         LFS_COMP_START_OPT,
1315         LFS_COMP_FLAGS_OPT,
1316         LFS_COMP_DEL_OPT,
1317         LFS_COMP_SET_OPT,
1318         LFS_COMP_ADD_OPT,
1319         LFS_PROJID_OPT,
1320 };
1321
1322 /* functions */
1323 static int lfs_setstripe(int argc, char **argv)
1324 {
1325         struct lfs_setstripe_args        lsa;
1326         struct llapi_stripe_param       *param = NULL;
1327         struct find_param                migrate_mdt_param = {
1328                 .fp_max_depth = -1,
1329                 .fp_mdt_index = -1,
1330         };
1331         char                            *fname;
1332         int                              result;
1333         int                              result2 = 0;
1334         char                            *end;
1335         int                              c;
1336         int                              delete = 0;
1337         char                            *mdt_idx_arg = NULL;
1338         unsigned long long               size_units = 1;
1339         bool                             migrate_mode = false;
1340         bool                             migration_block = false;
1341         __u64                            migration_flags = 0;
1342         __u32                            osts[LOV_MAX_STRIPE_COUNT] = { 0 };
1343         int                              comp_del = 0, comp_set = 0;
1344         int                              comp_add = 0;
1345         __u32                            comp_id = 0;
1346         struct llapi_layout             *layout = NULL;
1347
1348         struct option            long_opts[] = {
1349                 /* --block is only valid in migrate mode */
1350                 {"block",        no_argument,       0, 'b'},
1351                 {"comp-add",     no_argument,       0, LFS_COMP_ADD_OPT},
1352                 {"component-add", no_argument,      0, LFS_COMP_ADD_OPT},
1353                 {"comp-del",     no_argument,       0, LFS_COMP_DEL_OPT},
1354                 {"component-del", no_argument,      0, LFS_COMP_DEL_OPT},
1355                 {"comp-flags",   required_argument, 0, LFS_COMP_FLAGS_OPT},
1356                 {"component-flags", required_argument, 0, LFS_COMP_FLAGS_OPT},
1357                 {"comp-set",     no_argument,       0, LFS_COMP_SET_OPT},
1358                 {"component-set", no_argument,      0, LFS_COMP_SET_OPT},
1359 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
1360                 /* This formerly implied "stripe-count", but was explicitly
1361                  * made "stripe-count" for consistency with other options,
1362                  * and to separate it from "mdt-count" when DNE arrives. */
1363                 {"count",        required_argument, 0, 'c'},
1364 #endif
1365                 {"stripe-count", required_argument, 0, 'c'},
1366                 {"stripe_count", required_argument, 0, 'c'},
1367                 {"delete",       no_argument,       0, 'd'},
1368                 {"comp-end",     required_argument, 0, 'E'},
1369                 {"component-end", required_argument, 0, 'E'},
1370                 /* dirstripe {"mdt-hash",     required_argument, 0, 'H'}, */
1371 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
1372                 /* This formerly implied "stripe-index", but was explicitly
1373                  * made "stripe-index" for consistency with other options,
1374                  * and to separate it from "mdt-index" when DNE arrives. */
1375                 {"index",        required_argument, 0, 'i'},
1376 #endif
1377                 {"stripe-index", required_argument, 0, 'i'},
1378                 {"stripe_index", required_argument, 0, 'i'},
1379                 {"comp-id",      required_argument, 0, 'I'},
1380                 {"component-id", required_argument, 0, 'I'},
1381                 {"mdt",          required_argument, 0, 'm'},
1382                 {"mdt-index",    required_argument, 0, 'm'},
1383                 {"mdt_index",    required_argument, 0, 'm'},
1384                 /* --non-block is only valid in migrate mode */
1385                 {"non-block",    no_argument,       0, 'n'},
1386                 {"ost",          required_argument, 0, 'o'},
1387 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1388                 {"ost-list",     required_argument, 0, 'o'},
1389                 {"ost_list",     required_argument, 0, 'o'},
1390 #endif
1391                 {"pool",         required_argument, 0, 'p'},
1392 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
1393                 /* This formerly implied "--stripe-size", but was confusing
1394                  * with "lfs find --size|-s", which means "file size", so use
1395                  * the consistent "--stripe-size|-S" for all commands. */
1396                 {"size",         required_argument, 0, 's'},
1397 #endif
1398                 {"stripe-size",  required_argument, 0, 'S'},
1399                 {"stripe_size",  required_argument, 0, 'S'},
1400                 /* dirstripe {"mdt-count",    required_argument, 0, 'T'}, */
1401                 /* --verbose is only valid in migrate mode */
1402                 {"verbose",      no_argument,       0, 'v'},
1403                 {0, 0, 0, 0}
1404         };
1405
1406         setstripe_args_init(&lsa);
1407
1408         if (strcmp(argv[0], "migrate") == 0)
1409                 migrate_mode = true;
1410
1411         while ((c = getopt_long(argc, argv, "bc:dE:i:I:m:no:p:s:S:v",
1412                                 long_opts, NULL)) >= 0) {
1413                 switch (c) {
1414                 case 0:
1415                         /* Long options. */
1416                         break;
1417                 case LFS_COMP_ADD_OPT:
1418                         comp_add = 1;
1419                         break;
1420                 case LFS_COMP_DEL_OPT:
1421                         comp_del = 1;
1422                         break;
1423                 case LFS_COMP_FLAGS_OPT:
1424                         result = comp_str2flags(&lsa.lsa_comp_flags, optarg);
1425                         if (result != 0) {
1426                                 fprintf(stderr, "error: %s: bad comp flags "
1427                                         "'%s'\n", argv[0], optarg);
1428                                 goto error;
1429                         }
1430                         break;
1431                 case LFS_COMP_SET_OPT:
1432                         comp_set = 1;
1433                         break;
1434                 case 'b':
1435                         if (!migrate_mode) {
1436                                 fprintf(stderr, "--block is valid only for"
1437                                                 " migrate mode\n");
1438                                 goto error;
1439                         }
1440                         migration_block = true;
1441                         break;
1442                 case 'c':
1443 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
1444                         if (strcmp(argv[optind - 1], "--count") == 0)
1445                                 fprintf(stderr, "warning: '--count' deprecated"
1446                                         ", use '--stripe-count' instead\n");
1447 #endif
1448                         lsa.lsa_stripe_count = strtoul(optarg, &end, 0);
1449                         if (*end != '\0') {
1450                                 fprintf(stderr, "error: %s: bad stripe count "
1451                                         "'%s'\n", argv[0], optarg);
1452                                 goto error;
1453                         }
1454                         break;
1455                 case 'd':
1456                         /* delete the default striping pattern */
1457                         delete = 1;
1458                         break;
1459                 case 'E':
1460                         if (lsa.lsa_comp_end != 0) {
1461                                 result = comp_args_to_layout(&layout, &lsa);
1462                                 if (result)
1463                                         goto error;
1464
1465                                 setstripe_args_init(&lsa);
1466                         }
1467
1468                         if (arg_is_eof(optarg)) {
1469                                 lsa.lsa_comp_end = LUSTRE_EOF;
1470                         } else {
1471                                 result = llapi_parse_size(optarg,
1472                                                         &lsa.lsa_comp_end,
1473                                                         &size_units, 0);
1474                                 if (result) {
1475                                         fprintf(stderr, "error: %s: "
1476                                                 "bad component end '%s'\n",
1477                                                 argv[0], optarg);
1478                                         goto error;
1479                                 }
1480                         }
1481                         break;
1482                 case 'i':
1483                         if (strcmp(argv[optind - 1], "--index") == 0)
1484                                 fprintf(stderr, "warning: '--index' deprecated"
1485                                         ", use '--stripe-index' instead\n");
1486                         lsa.lsa_stripe_off = strtol(optarg, &end, 0);
1487                         if (*end != '\0') {
1488                                 fprintf(stderr, "error: %s: bad stripe offset "
1489                                         "'%s'\n", argv[0], optarg);
1490                                 goto error;
1491                         }
1492                         break;
1493                 case 'I':
1494                         comp_id = strtoul(optarg, &end, 0);
1495                         if (*end != '\0' || comp_id == 0 ||
1496                             comp_id > LCME_ID_MAX) {
1497                                 fprintf(stderr, "error: %s: bad comp ID "
1498                                         "'%s'\n", argv[0], optarg);
1499                                 goto error;
1500                         }
1501                         break;
1502                 case 'm':
1503                         if (!migrate_mode) {
1504                                 fprintf(stderr, "--mdt-index is valid only for"
1505                                                 " migrate mode\n");
1506                                 goto error;
1507                         }
1508                         mdt_idx_arg = optarg;
1509                         break;
1510                 case 'n':
1511                         if (!migrate_mode) {
1512                                 fprintf(stderr, "--non-block is valid only for"
1513                                                 " migrate mode\n");
1514                                 goto error;
1515                         }
1516                         migration_flags |= MIGRATION_NONBLOCK;
1517                         break;
1518                 case 'o':
1519                         lsa.lsa_nr_osts = parse_targets(osts,
1520                                                 sizeof(osts) / sizeof(__u32),
1521                                                 lsa.lsa_nr_osts, optarg);
1522                         if (lsa.lsa_nr_osts < 0) {
1523                                 fprintf(stderr,
1524                                         "error: %s: bad OST indices '%s'\n",
1525                                         argv[0], optarg);
1526                                 goto error;
1527                         }
1528
1529                         lsa.lsa_osts = osts;
1530                         if (lsa.lsa_stripe_off == -1)
1531                                 lsa.lsa_stripe_off = osts[0];
1532                         break;
1533                 case 'p':
1534                         result = verify_pool_name(argv[0], optarg);
1535                         if (result)
1536                                 goto error;
1537                         lsa.lsa_pool_name = optarg;
1538                         break;
1539 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
1540                 case 's':
1541 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
1542                         fprintf(stderr, "warning: '--size|-s' deprecated, "
1543                                 "use '--stripe-size|-S' instead\n");
1544 #endif
1545 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0) */
1546                 case 'S':
1547                         result = llapi_parse_size(optarg, &lsa.lsa_stripe_size,
1548                                                   &size_units, 0);
1549                         if (result) {
1550                                 fprintf(stderr, "error: %s: bad stripe size "
1551                                         "'%s'\n", argv[0], optarg);
1552                                 goto error;
1553                         }
1554                         break;
1555                 case 'v':
1556                         if (!migrate_mode) {
1557                                 fprintf(stderr, "--verbose is valid only for"
1558                                                 " migrate mode\n");
1559                                 goto error;
1560                         }
1561                         migrate_mdt_param.fp_verbose = VERBOSE_DETAIL;
1562                         break;
1563                 default:
1564                         goto error;
1565                 }
1566         }
1567
1568         fname = argv[optind];
1569
1570         if (lsa.lsa_comp_end != 0) {
1571                 result = comp_args_to_layout(&layout, &lsa);
1572                 if (result)
1573                         goto error;
1574         }
1575
1576         if (optind == argc) {
1577                 fprintf(stderr, "error: %s: missing filename|dirname\n",
1578                         argv[0]);
1579                 goto error;
1580         }
1581
1582         /* Only LCME_FL_INIT flags is used in PFL, and it shouldn't be
1583          * altered by user space tool, so we don't need to support the
1584          * --component-set for this moment. */
1585         if (comp_set != 0) {
1586                 fprintf(stderr, "error: %s: --component-set isn't supported.\n",
1587                         argv[0]);
1588                 goto error;
1589         }
1590
1591         if ((delete + comp_set + comp_del + comp_add) > 1) {
1592                 fprintf(stderr, "error: %s: can't specify --component-set, "
1593                         "--component-del, --component-add or -d together\n",
1594                         argv[0]);
1595                 goto error;
1596         }
1597
1598         if (delete && (setstripe_args_specified(&lsa) || comp_id != 0 ||
1599                        lsa.lsa_comp_flags != 0 || layout != NULL)) {
1600                 fprintf(stderr, "error: %s: can't specify -d with "
1601                         "-s, -c, -o, -p, -I, -F or -E options\n",
1602                         argv[0]);
1603                 goto error;
1604         }
1605
1606         if ((comp_set || comp_del) &&
1607             (setstripe_args_specified(&lsa) || layout != NULL)) {
1608                 fprintf(stderr, "error: %s: can't specify --component-del or "
1609                         "--component-set with -s, -c, -o, -p or -E options.\n",
1610                         argv[0]);
1611                 goto error;
1612         }
1613
1614         if (comp_del && comp_id != 0 && lsa.lsa_comp_flags != 0) {
1615                 fprintf(stderr, "error: %s: can't specify both -I and -F for "
1616                         "--component-del option.\n", argv[0]);
1617                 goto error;
1618         }
1619
1620         if (comp_add || comp_del) {
1621                 struct stat st;
1622
1623                 result = lstat(fname, &st);
1624                 if (result == 0 && S_ISDIR(st.st_mode)) {
1625                         fprintf(stderr, "error: %s: can't use --component-add "
1626                                 "or --component-del for directory.\n",
1627                                 argv[0]);
1628                         goto error;
1629                 }
1630         }
1631
1632         if (comp_add) {
1633                 if (layout == NULL) {
1634                         fprintf(stderr, "error: %s: -E option must be present"
1635                                 "in --component-add mode.\n", argv[0]);
1636                         goto error;
1637                 }
1638                 result = adjust_first_extent(fname, layout);
1639                 if (result == -ENODATA)
1640                         comp_add = 0;
1641                 else if (result != 0)
1642                         goto error;
1643         }
1644
1645         if (mdt_idx_arg != NULL && optind > 3) {
1646                 fprintf(stderr, "error: %s: cannot specify -m with other "
1647                         "options\n", argv[0]);
1648                 goto error;
1649         }
1650
1651         if ((migration_flags & MIGRATION_NONBLOCK) && migration_block) {
1652                 fprintf(stderr,
1653                         "error: %s: cannot specify --non-block and --block\n",
1654                         argv[0]);
1655                 goto error;
1656         }
1657
1658         if (!comp_del && !comp_set && comp_id != 0) {
1659                 fprintf(stderr, "error: %s: -I can only be used with "
1660                         "--component-del.\n", argv[0]);
1661                 goto error;
1662         }
1663
1664         if (mdt_idx_arg != NULL) {
1665                 /* initialize migrate mdt parameters */
1666                 migrate_mdt_param.fp_mdt_index = strtoul(mdt_idx_arg, &end, 0);
1667                 if (*end != '\0') {
1668                         fprintf(stderr, "error: %s: bad MDT index '%s'\n",
1669                                 argv[0], mdt_idx_arg);
1670                         goto error;
1671                 }
1672                 migrate_mdt_param.fp_migrate = 1;
1673         } else if (layout == NULL) {
1674                 /* initialize stripe parameters */
1675                 param = calloc(1, offsetof(typeof(*param),
1676                                lsp_osts[lsa.lsa_nr_osts]));
1677                 if (param == NULL) {
1678                         fprintf(stderr, "error: %s: %s\n", argv[0],
1679                                 strerror(ENOMEM));
1680                         goto error;
1681                 }
1682
1683                 param->lsp_stripe_size = lsa.lsa_stripe_size;
1684                 param->lsp_stripe_offset = lsa.lsa_stripe_off;
1685                 param->lsp_stripe_count = lsa.lsa_stripe_count;
1686                 param->lsp_stripe_pattern = 0;
1687                 param->lsp_pool = lsa.lsa_pool_name;
1688                 param->lsp_is_specific = false;
1689                 if (lsa.lsa_nr_osts > 0) {
1690                         if (lsa.lsa_stripe_count > 0 &&
1691                             lsa.lsa_nr_osts != lsa.lsa_stripe_count) {
1692                                 fprintf(stderr, "error: %s: stripe count '%d' "
1693                                         "doesn't match the number of OSTs: %d\n"
1694                                         , argv[0], lsa.lsa_stripe_count,
1695                                         lsa.lsa_nr_osts);
1696                                 free(param);
1697                                 goto error;
1698                         }
1699
1700                         param->lsp_is_specific = true;
1701                         param->lsp_stripe_count = lsa.lsa_nr_osts;
1702                         memcpy(param->lsp_osts, osts,
1703                                sizeof(*osts) * lsa.lsa_nr_osts);
1704                 }
1705         }
1706
1707         for (fname = argv[optind]; fname != NULL; fname = argv[++optind]) {
1708                 char *op;
1709                 if (mdt_idx_arg != NULL) {
1710                         result = llapi_migrate_mdt(fname, &migrate_mdt_param);
1711                         op = "migrate mdt objects of";
1712                 } else if (migrate_mode) {
1713                         result = lfs_migrate(fname, migration_flags, param,
1714                                              layout);
1715                         op = "migrate ost objects of";
1716                 } else if (comp_set != 0) {
1717                         result = lfs_component_set(fname, comp_id,
1718                                                    lsa.lsa_comp_flags);
1719                         op = "modify component flags of";
1720                 } else if (comp_del != 0) {
1721                         result = lfs_component_del(fname, comp_id,
1722                                                    lsa.lsa_comp_flags);
1723                         op = "delete component of";
1724                 } else if (comp_add != 0) {
1725                         result = lfs_component_add(fname, layout);
1726                         op = "add component to";
1727                 } else if (layout != NULL) {
1728                         result = lfs_component_create(fname, O_CREAT | O_WRONLY,
1729                                                       0644, layout);
1730                         if (result >= 0) {
1731                                 close(result);
1732                                 result = 0;
1733                         }
1734                         op = "create composite";
1735                 } else {
1736                         result = llapi_file_open_param(fname,
1737                                                        O_CREAT | O_WRONLY,
1738                                                        0644, param);
1739                         if (result >= 0) {
1740                                 close(result);
1741                                 result = 0;
1742                         }
1743                         op = "create striped";
1744                 }
1745                 if (result) {
1746                         /* Save the first error encountered. */
1747                         if (result2 == 0)
1748                                 result2 = result;
1749                         fprintf(stderr, "error: %s: %s file '%s' failed: %s\n",
1750                                 argv[0], op, fname,
1751                                 lsa.lsa_pool_name != NULL && result == EINVAL ?
1752                                 "OST not in pool?" : strerror(errno));
1753                         continue;
1754                 }
1755         }
1756
1757         free(param);
1758         llapi_layout_free(layout);
1759         return result2;
1760 error:
1761         llapi_layout_free(layout);
1762         return CMD_HELP;
1763 }
1764
1765 static int lfs_poollist(int argc, char **argv)
1766 {
1767         if (argc != 2)
1768                 return CMD_HELP;
1769
1770         return llapi_poollist(argv[1]);
1771 }
1772
1773 static int set_time(time_t *time, time_t *set, char *str)
1774 {
1775         time_t t;
1776         int res = 0;
1777
1778         if (str[0] == '+')
1779                 res = 1;
1780         else if (str[0] == '-')
1781                 res = -1;
1782
1783         if (res)
1784                 str++;
1785
1786         t = strtol(str, NULL, 0);
1787         if (*time < t * 24 * 60 * 60) {
1788                 if (res)
1789                         str--;
1790                 fprintf(stderr, "Wrong time '%s' is specified.\n", str);
1791                 return INT_MAX;
1792         }
1793
1794         *set = *time - t * 24 * 60 * 60;
1795         return res;
1796 }
1797 static int name2uid(unsigned int *id, const char *name)
1798 {
1799         struct passwd *passwd;
1800
1801         passwd = getpwnam(name);
1802         if (passwd == NULL)
1803                 return -ENOENT;
1804         *id = passwd->pw_uid;
1805
1806         return 0;
1807 }
1808
1809 static int name2gid(unsigned int *id, const char *name)
1810 {
1811         struct group *group;
1812
1813         group = getgrnam(name);
1814         if (group == NULL)
1815                 return -ENOENT;
1816         *id = group->gr_gid;
1817
1818         return 0;
1819 }
1820
1821 static inline int name2projid(unsigned int *id, const char *name)
1822 {
1823         return -ENOTSUP;
1824 }
1825
1826 static int uid2name(char **name, unsigned int id)
1827 {
1828         struct passwd *passwd;
1829
1830         passwd = getpwuid(id);
1831         if (passwd == NULL)
1832                 return -ENOENT;
1833         *name = passwd->pw_name;
1834
1835         return 0;
1836 }
1837
1838 static inline int gid2name(char **name, unsigned int id)
1839 {
1840         struct group *group;
1841
1842         group = getgrgid(id);
1843         if (group == NULL)
1844                 return -ENOENT;
1845         *name = group->gr_name;
1846
1847         return 0;
1848 }
1849
1850 static int name2layout(__u32 *layout, char *name)
1851 {
1852         char *ptr, *lyt;
1853
1854         *layout = 0;
1855         for (ptr = name; ; ptr = NULL) {
1856                 lyt = strtok(ptr, ",");
1857                 if (lyt == NULL)
1858                         break;
1859                 if (strcmp(lyt, "released") == 0)
1860                         *layout |= LOV_PATTERN_F_RELEASED;
1861                 else if (strcmp(lyt, "raid0") == 0)
1862                         *layout |= LOV_PATTERN_RAID0;
1863                 else
1864                         return -1;
1865         }
1866         return 0;
1867 }
1868
1869 static int lfs_find(int argc, char **argv)
1870 {
1871         int c, rc;
1872         int ret = 0;
1873         time_t t;
1874         struct find_param param = {
1875                 .fp_max_depth = -1,
1876                 .fp_quiet = 1,
1877         };
1878         struct option long_opts[] = {
1879                 {"atime",        required_argument, 0, 'A'},
1880                 {"comp-count",   required_argument, 0, LFS_COMP_COUNT_OPT},
1881                 {"component-count", required_argument, 0, LFS_COMP_COUNT_OPT},
1882                 {"comp-flags",   required_argument, 0, LFS_COMP_FLAGS_OPT},
1883                 {"component-flags", required_argument, 0, LFS_COMP_FLAGS_OPT},
1884                 {"comp-start",   required_argument, 0, LFS_COMP_START_OPT},
1885                 {"component-start", required_argument, 0, LFS_COMP_START_OPT},
1886                 {"stripe-count", required_argument, 0, 'c'},
1887                 {"stripe_count", required_argument, 0, 'c'},
1888                 {"ctime",        required_argument, 0, 'C'},
1889                 {"maxdepth",     required_argument, 0, 'D'},
1890                 {"comp-end",     required_argument, 0, 'E'},
1891                 {"component-end", required_argument, 0, 'E'},
1892                 {"gid",          required_argument, 0, 'g'},
1893                 {"group",        required_argument, 0, 'G'},
1894                 {"mdt-hash",     required_argument, 0, 'H'},
1895                 {"stripe-index", required_argument, 0, 'i'},
1896                 {"stripe_index", required_argument, 0, 'i'},
1897                 /*{"component-id", required_argument, 0, 'I'},*/
1898                 {"layout",       required_argument, 0, 'L'},
1899                 {"mdt",          required_argument, 0, 'm'},
1900                 {"mdt-index",    required_argument, 0, 'm'},
1901                 {"mdt_index",    required_argument, 0, 'm'},
1902                 {"mtime",        required_argument, 0, 'M'},
1903                 {"name",         required_argument, 0, 'n'},
1904      /* reserve {"or",           no_argument,     , 0, 'o'}, to match find(1) */
1905                 {"obd",          required_argument, 0, 'O'},
1906                 {"ost",          required_argument, 0, 'O'},
1907                 /* no short option for pool, p/P already used */
1908                 {"pool",         required_argument, 0, LFS_POOL_OPT},
1909                 {"print0",       no_argument,       0, 'p'},
1910                 {"print",        no_argument,       0, 'P'},
1911                 {"projid",       required_argument, 0, LFS_PROJID_OPT},
1912                 {"size",         required_argument, 0, 's'},
1913                 {"stripe-size",  required_argument, 0, 'S'},
1914                 {"stripe_size",  required_argument, 0, 'S'},
1915                 {"type",         required_argument, 0, 't'},
1916                 {"mdt-count",    required_argument, 0, 'T'},
1917                 {"uid",          required_argument, 0, 'u'},
1918                 {"user",         required_argument, 0, 'U'},
1919                 {0, 0, 0, 0}
1920         };
1921         int pathstart = -1;
1922         int pathend = -1;
1923         int neg_opt = 0;
1924         time_t *xtime;
1925         int *xsign;
1926         int isoption;
1927         char *endptr;
1928
1929         time(&t);
1930
1931         /* when getopt_long_only() hits '!' it returns 1, puts "!" in optarg */
1932         while ((c = getopt_long_only(argc, argv,
1933                         "-A:c:C:D:E:g:G:H:i:L:m:M:n:O:Ppqrs:S:t:T:u:U:v",
1934                         long_opts, NULL)) >= 0) {
1935                 xtime = NULL;
1936                 xsign = NULL;
1937                 if (neg_opt)
1938                         --neg_opt;
1939                 /* '!' is part of option */
1940                 /* when getopt_long_only() finds a string which is not
1941                  * an option nor a known option argument it returns 1
1942                  * in that case if we already have found pathstart and pathend
1943                  * (i.e. we have the list of pathnames),
1944                  * the only supported value is "!"
1945                  */
1946                 isoption = (c != 1) || (strcmp(optarg, "!") == 0);
1947                 if (!isoption && pathend != -1) {
1948                         fprintf(stderr, "err: %s: filename|dirname must either "
1949                                         "precede options or follow options\n",
1950                                         argv[0]);
1951                         ret = CMD_HELP;
1952                         goto err;
1953                 }
1954                 if (!isoption && pathstart == -1)
1955                         pathstart = optind - 1;
1956                 if (isoption && pathstart != -1 && pathend == -1)
1957                         pathend = optind - 2;
1958                 switch (c) {
1959                 case 0:
1960                         /* Long options. */
1961                         break;
1962                 case 1:
1963                         /* unknown; opt is "!" or path component,
1964                          * checking done above.
1965                          */
1966                         if (strcmp(optarg, "!") == 0)
1967                                 neg_opt = 2;
1968                         break;
1969                 case 'A':
1970                         xtime = &param.fp_atime;
1971                         xsign = &param.fp_asign;
1972                         param.fp_exclude_atime = !!neg_opt;
1973                         /* no break, this falls through to 'C' for ctime */
1974                 case 'C':
1975                         if (c == 'C') {
1976                                 xtime = &param.fp_ctime;
1977                                 xsign = &param.fp_csign;
1978                                 param.fp_exclude_ctime = !!neg_opt;
1979                         }
1980                         /* no break, this falls through to 'M' for mtime */
1981                 case 'M':
1982                         if (c == 'M') {
1983                                 xtime = &param.fp_mtime;
1984                                 xsign = &param.fp_msign;
1985                                 param.fp_exclude_mtime = !!neg_opt;
1986                         }
1987                         rc = set_time(&t, xtime, optarg);
1988                         if (rc == INT_MAX) {
1989                                 ret = -1;
1990                                 goto err;
1991                         }
1992                         if (rc)
1993                                 *xsign = rc;
1994                         break;
1995                 case LFS_COMP_COUNT_OPT:
1996                         if (optarg[0] == '+') {
1997                                 param.fp_comp_count_sign = -1;
1998                                 optarg++;
1999                         } else if (optarg[0] == '-') {
2000                                 param.fp_comp_count_sign =  1;
2001                                 optarg++;
2002                         }
2003
2004                         param.fp_comp_count = strtoul(optarg, &endptr, 0);
2005                         if (*endptr != '\0') {
2006                                 fprintf(stderr, "error: bad component count "
2007                                         "'%s'\n", optarg);
2008                                 goto err;
2009                         }
2010                         param.fp_check_comp_count = 1;
2011                         param.fp_exclude_comp_count = !!neg_opt;
2012                         break;
2013                 case LFS_COMP_FLAGS_OPT:
2014                         rc = comp_str2flags(&param.fp_comp_flags, optarg);
2015                         if (rc || comp_flags_is_neg(param.fp_comp_flags)) {
2016                                 fprintf(stderr, "error: bad component flags "
2017                                         "'%s'\n", optarg);
2018                                 goto err;
2019                         }
2020                         param.fp_check_comp_flags = 1;
2021                         param.fp_exclude_comp_flags = !!neg_opt;
2022                         break;
2023                 case LFS_COMP_START_OPT:
2024                         if (optarg[0] == '+') {
2025                                 param.fp_comp_start_sign = -1;
2026                                 optarg++;
2027                         } else if (optarg[0] == '-') {
2028                                 param.fp_comp_start_sign =  1;
2029                                 optarg++;
2030                         }
2031
2032                         rc = llapi_parse_size(optarg, &param.fp_comp_start,
2033                                               &param.fp_comp_start_units, 0);
2034                         if (rc) {
2035                                 fprintf(stderr, "error: bad component start "
2036                                         "'%s'\n", optarg);
2037                                 goto err;
2038                         }
2039                         param.fp_check_comp_start = 1;
2040                         param.fp_exclude_comp_start = !!neg_opt;
2041                         break;
2042                 case 'c':
2043                         if (optarg[0] == '+') {
2044                                 param.fp_stripe_count_sign = -1;
2045                                 optarg++;
2046                         } else if (optarg[0] == '-') {
2047                                 param.fp_stripe_count_sign =  1;
2048                                 optarg++;
2049                         }
2050
2051                         param.fp_stripe_count = strtoul(optarg, &endptr, 0);
2052                         if (*endptr != '\0') {
2053                                 fprintf(stderr,"error: bad stripe_count '%s'\n",
2054                                         optarg);
2055                                 ret = -1;
2056                                 goto err;
2057                         }
2058                         param.fp_check_stripe_count = 1;
2059                         param.fp_exclude_stripe_count = !!neg_opt;
2060                         break;
2061                 case 'D':
2062                         param.fp_max_depth = strtol(optarg, 0, 0);
2063                         break;
2064                 case 'E':
2065                         if (optarg[0] == '+') {
2066                                 param.fp_comp_end_sign = -1;
2067                                 optarg++;
2068                         } else if (optarg[0] == '-') {
2069                                 param.fp_comp_end_sign =  1;
2070                                 optarg++;
2071                         }
2072
2073                         if (arg_is_eof(optarg)) {
2074                                 param.fp_comp_end = LUSTRE_EOF;
2075                                 param.fp_comp_end_units = 1;
2076                                 rc = 0;
2077                         } else {
2078                                 rc = llapi_parse_size(optarg,
2079                                                 &param.fp_comp_end,
2080                                                 &param.fp_comp_end_units, 0);
2081                         }
2082                         if (rc) {
2083                                 fprintf(stderr, "error: bad component end "
2084                                         "'%s'\n", optarg);
2085                                 goto err;
2086                         }
2087                         param.fp_check_comp_end = 1;
2088                         param.fp_exclude_comp_end = !!neg_opt;
2089                         break;
2090                 case 'g':
2091                 case 'G':
2092                         rc = name2gid(&param.fp_gid, optarg);
2093                         if (rc) {
2094                                 param.fp_gid = strtoul(optarg, &endptr, 10);
2095                                 if (*endptr != '\0') {
2096                                         fprintf(stderr, "Group/GID: %s cannot "
2097                                                 "be found.\n", optarg);
2098                                         ret = -1;
2099                                         goto err;
2100                                 }
2101                         }
2102                         param.fp_exclude_gid = !!neg_opt;
2103                         param.fp_check_gid = 1;
2104                         break;
2105                 case 'H':
2106                         param.fp_hash_type = check_hashtype(optarg);
2107                         if (param.fp_hash_type == 0) {
2108                                 fprintf(stderr, "error: bad hash_type '%s'\n",
2109                                         optarg);
2110                                 ret = -1;
2111                                 goto err;
2112                         }
2113                         param.fp_check_hash_type = 1;
2114                         param.fp_exclude_hash_type = !!neg_opt;
2115                         break;
2116                 case 'L':
2117                         ret = name2layout(&param.fp_layout, optarg);
2118                         if (ret)
2119                                 goto err;
2120                         param.fp_exclude_layout = !!neg_opt;
2121                         param.fp_check_layout = 1;
2122                         break;
2123                 case 'u':
2124                 case 'U':
2125                         rc = name2uid(&param.fp_uid, optarg);
2126                         if (rc) {
2127                                 param.fp_uid = strtoul(optarg, &endptr, 10);
2128                                 if (*endptr != '\0') {
2129                                         fprintf(stderr, "User/UID: %s cannot "
2130                                                 "be found.\n", optarg);
2131                                         ret = -1;
2132                                         goto err;
2133                                 }
2134                         }
2135                         param.fp_exclude_uid = !!neg_opt;
2136                         param.fp_check_uid = 1;
2137                         break;
2138                 case LFS_POOL_OPT:
2139                         if (strlen(optarg) > LOV_MAXPOOLNAME) {
2140                                 fprintf(stderr,
2141                                         "Pool name %s is too long"
2142                                         " (max is %d)\n", optarg,
2143                                         LOV_MAXPOOLNAME);
2144                                 ret = -1;
2145                                 goto err;
2146                         }
2147                         /* we do check for empty pool because empty pool
2148                          * is used to find V1 lov attributes */
2149                         strncpy(param.fp_poolname, optarg, LOV_MAXPOOLNAME);
2150                         param.fp_poolname[LOV_MAXPOOLNAME] = '\0';
2151                         param.fp_exclude_pool = !!neg_opt;
2152                         param.fp_check_pool = 1;
2153                         break;
2154                 case 'n':
2155                         param.fp_pattern = (char *)optarg;
2156                         param.fp_exclude_pattern = !!neg_opt;
2157                         break;
2158                 case 'm':
2159                 case 'i':
2160                 case 'O': {
2161                         char *buf, *token, *next, *p;
2162                         int len = 1;
2163                         void *tmp;
2164
2165                         buf = strdup(optarg);
2166                         if (buf == NULL) {
2167                                 ret = -ENOMEM;
2168                                 goto err;
2169                         }
2170
2171                         param.fp_exclude_obd = !!neg_opt;
2172
2173                         token = buf;
2174                         while (token && *token) {
2175                                 token = strchr(token, ',');
2176                                 if (token) {
2177                                         len++;
2178                                         token++;
2179                                 }
2180                         }
2181                         if (c == 'm') {
2182                                 param.fp_exclude_mdt = !!neg_opt;
2183                                 param.fp_num_alloc_mdts += len;
2184                                 tmp = realloc(param.fp_mdt_uuid,
2185                                               param.fp_num_alloc_mdts *
2186                                               sizeof(*param.fp_mdt_uuid));
2187                                 if (tmp == NULL) {
2188                                         ret = -ENOMEM;
2189                                         goto err_free;
2190                                 }
2191
2192                                 param.fp_mdt_uuid = tmp;
2193                         } else {
2194                                 param.fp_exclude_obd = !!neg_opt;
2195                                 param.fp_num_alloc_obds += len;
2196                                 tmp = realloc(param.fp_obd_uuid,
2197                                               param.fp_num_alloc_obds *
2198                                               sizeof(*param.fp_obd_uuid));
2199                                 if (tmp == NULL) {
2200                                         ret = -ENOMEM;
2201                                         goto err_free;
2202                                 }
2203
2204                                 param.fp_obd_uuid = tmp;
2205                         }
2206                         for (token = buf; token && *token; token = next) {
2207                                 struct obd_uuid *puuid;
2208                                 if (c == 'm') {
2209                                         puuid =
2210                                         &param.fp_mdt_uuid[param.fp_num_mdts++];
2211                                 } else {
2212                                         puuid =
2213                                         &param.fp_obd_uuid[param.fp_num_obds++];
2214                                 }
2215                                 p = strchr(token, ',');
2216                                 next = 0;
2217                                 if (p) {
2218                                         *p = 0;
2219                                         next = p+1;
2220                                 }
2221
2222                                 if (strlen(token) > sizeof(puuid->uuid) - 1) {
2223                                         ret = -E2BIG;
2224                                         goto err_free;
2225                                 }
2226
2227                                 strncpy(puuid->uuid, token,
2228                                         sizeof(puuid->uuid));
2229                         }
2230 err_free:
2231                         if (buf)
2232                                 free(buf);
2233                         break;
2234                 }
2235                 case 'p':
2236                         param.fp_zero_end = 1;
2237                         break;
2238                 case 'P':
2239                         break;
2240                 case LFS_PROJID_OPT:
2241                         rc = name2projid(&param.fp_projid, optarg);
2242                         if (rc) {
2243                                 param.fp_projid = strtoul(optarg, &endptr, 10);
2244                                 if (*endptr != '\0') {
2245                                         fprintf(stderr,
2246                                                 "Invalid project ID: %s",
2247                                                 optarg);
2248                                         ret = -1;
2249                                         goto err;
2250                                 }
2251                         }
2252                         param.fp_exclude_projid = !!neg_opt;
2253                         param.fp_check_projid = 1;
2254                         break;
2255                 case 's':
2256                         if (optarg[0] == '+') {
2257                                 param.fp_size_sign = -1;
2258                                 optarg++;
2259                         } else if (optarg[0] == '-') {
2260                                 param.fp_size_sign =  1;
2261                                 optarg++;
2262                         }
2263
2264                         ret = llapi_parse_size(optarg, &param.fp_size,
2265                                                &param.fp_size_units, 0);
2266                         if (ret) {
2267                                 fprintf(stderr, "error: bad file size '%s'\n",
2268                                         optarg);
2269                                 goto err;
2270                         }
2271                         param.fp_check_size = 1;
2272                         param.fp_exclude_size = !!neg_opt;
2273                         break;
2274                 case 'S':
2275                         if (optarg[0] == '+') {
2276                                 param.fp_stripe_size_sign = -1;
2277                                 optarg++;
2278                         } else if (optarg[0] == '-') {
2279                                 param.fp_stripe_size_sign =  1;
2280                                 optarg++;
2281                         }
2282
2283                         ret = llapi_parse_size(optarg, &param.fp_stripe_size,
2284                                                &param.fp_stripe_size_units, 0);
2285                         if (ret) {
2286                                 fprintf(stderr, "error: bad stripe_size '%s'\n",
2287                                         optarg);
2288                                 goto err;
2289                         }
2290                         param.fp_check_stripe_size = 1;
2291                         param.fp_exclude_stripe_size = !!neg_opt;
2292                         break;
2293                 case 't':
2294                         param.fp_exclude_type = !!neg_opt;
2295                         switch (optarg[0]) {
2296                         case 'b':
2297                                 param.fp_type = S_IFBLK;
2298                                 break;
2299                         case 'c':
2300                                 param.fp_type = S_IFCHR;
2301                                 break;
2302                         case 'd':
2303                                 param.fp_type = S_IFDIR;
2304                                 break;
2305                         case 'f':
2306                                 param.fp_type = S_IFREG;
2307                                 break;
2308                         case 'l':
2309                                 param.fp_type = S_IFLNK;
2310                                 break;
2311                         case 'p':
2312                                 param.fp_type = S_IFIFO;
2313                                 break;
2314                         case 's':
2315                                 param.fp_type = S_IFSOCK;
2316                                 break;
2317                         default:
2318                                 fprintf(stderr, "error: %s: bad type '%s'\n",
2319                                         argv[0], optarg);
2320                                 ret = CMD_HELP;
2321                                 goto err;
2322                         };
2323                         break;
2324                 case 'T':
2325                         if (optarg[0] == '+') {
2326                                 param.fp_mdt_count_sign = -1;
2327                                 optarg++;
2328                         } else if (optarg[0] == '-') {
2329                                 param.fp_mdt_count_sign =  1;
2330                                 optarg++;
2331                         }
2332
2333                         param.fp_mdt_count = strtoul(optarg, &endptr, 0);
2334                         if (*endptr != '\0') {
2335                                 fprintf(stderr, "error: bad mdt_count '%s'\n",
2336                                         optarg);
2337                                 ret = -1;
2338                                 goto err;
2339                         }
2340                         param.fp_check_mdt_count = 1;
2341                         param.fp_exclude_mdt_count = !!neg_opt;
2342                         break;
2343                 default:
2344                         ret = CMD_HELP;
2345                         goto err;
2346                 };
2347         }
2348
2349         if (pathstart == -1) {
2350                 fprintf(stderr, "error: %s: no filename|pathname\n",
2351                         argv[0]);
2352                 ret = CMD_HELP;
2353                 goto err;
2354         } else if (pathend == -1) {
2355                 /* no options */
2356                 pathend = argc;
2357         }
2358
2359         do {
2360                 rc = llapi_find(argv[pathstart], &param);
2361                 if (rc != 0 && ret == 0)
2362                         ret = rc;
2363         } while (++pathstart < pathend);
2364
2365         if (ret)
2366                 fprintf(stderr, "error: %s failed for %s.\n",
2367                         argv[0], argv[optind - 1]);
2368 err:
2369         if (param.fp_obd_uuid && param.fp_num_alloc_obds)
2370                 free(param.fp_obd_uuid);
2371
2372         if (param.fp_mdt_uuid && param.fp_num_alloc_mdts)
2373                 free(param.fp_mdt_uuid);
2374
2375         return ret;
2376 }
2377
2378 static int lfs_getstripe_internal(int argc, char **argv,
2379                                   struct find_param *param)
2380 {
2381         struct option long_opts[] = {
2382                 {"comp-count",          no_argument, 0, LFS_COMP_COUNT_OPT},
2383                 {"component-count",     no_argument, 0, LFS_COMP_COUNT_OPT},
2384                 {"comp-flags",      optional_argument, 0, LFS_COMP_FLAGS_OPT},
2385                 {"component-flags", optional_argument, 0, LFS_COMP_FLAGS_OPT},
2386                 {"comp-start",      optional_argument, 0, LFS_COMP_START_OPT},
2387                 {"component-start", optional_argument, 0, LFS_COMP_START_OPT},
2388 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
2389                 /* This formerly implied "stripe-count", but was explicitly
2390                  * made "stripe-count" for consistency with other options,
2391                  * and to separate it from "mdt-count" when DNE arrives. */
2392                 {"count",               no_argument,            0, 'c'},
2393 #endif
2394                 {"stripe-count",        no_argument,            0, 'c'},
2395                 {"stripe_count",        no_argument,            0, 'c'},
2396                 {"directory",           no_argument,            0, 'd'},
2397                 {"default",             no_argument,            0, 'D'},
2398                 {"comp-end",            optional_argument,      0, 'E'},
2399                 {"component-end",       optional_argument,      0, 'E'},
2400                 {"fid",                 no_argument,            0, 'F'},
2401                 {"generation",          no_argument,            0, 'g'},
2402                 /* dirstripe {"mdt-hash",     required_argument, 0, 'H'}, */
2403 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
2404                 /* This formerly implied "stripe-index", but was explicitly
2405                  * made "stripe-index" for consistency with other options,
2406                  * and to separate it from "mdt-index" when DNE arrives. */
2407                 {"index",               no_argument,            0, 'i'},
2408 #endif
2409                 {"stripe-index",        no_argument,            0, 'i'},
2410                 {"stripe_index",        no_argument,            0, 'i'},
2411                 {"comp-id",             optional_argument,      0, 'I'},
2412                 {"component-id",        optional_argument,      0, 'I'},
2413                 {"layout",              no_argument,            0, 'L'},
2414                 {"mdt",                 no_argument,            0, 'm'},
2415                 {"mdt-index",           no_argument,            0, 'm'},
2416                 {"mdt_index",           no_argument,            0, 'm'},
2417 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2418                 {"mdt-index",           no_argument,            0, 'M'},
2419                 {"mdt_index",           no_argument,            0, 'M'},
2420 #endif
2421 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
2422                 /* This formerly implied "stripe-index", but was confusing
2423                  * with "file offset" (which will eventually be needed for
2424                  * with different layouts by offset), so deprecate it. */
2425                 {"offset",              no_argument,            0, 'o'},
2426 #endif
2427                 {"obd",                 required_argument,      0, 'O'},
2428                 {"ost",                 required_argument,      0, 'O'},
2429                 {"pool",                no_argument,            0, 'p'},
2430                 {"quiet",               no_argument,            0, 'q'},
2431                 {"recursive",           no_argument,            0, 'r'},
2432                 {"raw",                 no_argument,            0, 'R'},
2433 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
2434                 /* This formerly implied "--stripe-size", but was confusing
2435                  * with "lfs find --size|-s", which means "file size", so use
2436                  * the consistent "--stripe-size|-S" for all commands. */
2437                 {"size",                no_argument,            0, 's'},
2438 #endif
2439                 {"stripe-size",         no_argument,            0, 'S'},
2440                 {"stripe_size",         no_argument,            0, 'S'},
2441                 /* dirstripe {"mdt-count",    required_argument, 0, 'T'}, */
2442                 {"verbose",             no_argument,            0, 'v'},
2443                 {"yaml",                no_argument,            0, 'y'},
2444                 {0, 0, 0, 0}
2445         };
2446         int c, rc;
2447         char *end, *tmp;
2448
2449         while ((c = getopt_long(argc, argv, "cdDE::FghiI::LmMoO:pqrRsSvy",
2450                                 long_opts, NULL)) != -1) {
2451                 switch (c) {
2452                 case 'c':
2453                         if (strcmp(argv[optind - 1], "--count") == 0)
2454                                 fprintf(stderr, "warning: '--count' deprecated,"
2455                                         " use '--stripe-count' instead\n");
2456                         if (!(param->fp_verbose & VERBOSE_DETAIL)) {
2457                                 param->fp_verbose |= VERBOSE_COUNT;
2458                                 param->fp_max_depth = 0;
2459                         }
2460                         break;
2461                 case LFS_COMP_COUNT_OPT:
2462                         param->fp_verbose |= VERBOSE_COMP_COUNT;
2463                         param->fp_max_depth = 0;
2464                         break;
2465                 case LFS_COMP_FLAGS_OPT:
2466                         if (optarg != NULL) {
2467                                 __u32 *flags = &param->fp_comp_flags;
2468                                 rc = comp_str2flags(flags, optarg);
2469                                 if (rc != 0) {
2470                                         fprintf(stderr, "error: %s bad "
2471                                                 "component flags '%s'.\n",
2472                                                 argv[0], optarg);
2473                                         return CMD_HELP;
2474                                 } else {
2475                                         param->fp_check_comp_flags = 1;
2476                                         param->fp_exclude_comp_flags =
2477                                                 comp_flags_is_neg(*flags);
2478                                         comp_flags_clear_neg(flags);
2479                                 }
2480                         } else {
2481                                 param->fp_verbose |= VERBOSE_COMP_FLAGS;
2482                                 param->fp_max_depth = 0;
2483                         }
2484                         break;
2485                 case LFS_COMP_START_OPT:
2486                         if (optarg != NULL) {
2487                                 tmp = optarg;
2488                                 if (tmp[0] == '+') {
2489                                         param->fp_comp_start_sign = -1;
2490                                         tmp++;
2491                                 } else if (tmp[0] == '-') {
2492                                         param->fp_comp_start_sign = 1;
2493                                         tmp++;
2494                                 }
2495                                 rc = llapi_parse_size(tmp,
2496                                                 &param->fp_comp_start,
2497                                                 &param->fp_comp_start_units, 0);
2498                                 if (rc != 0) {
2499                                         fprintf(stderr, "error: %s bad "
2500                                                 "component start '%s'.\n",
2501                                                 argv[0], tmp);
2502                                         return CMD_HELP;
2503                                 } else {
2504                                         param->fp_check_comp_start = 1;
2505                                 }
2506                         } else {
2507                                 param->fp_verbose |= VERBOSE_COMP_START;
2508                                 param->fp_max_depth = 0;
2509                         }
2510                         break;
2511                 case 'd':
2512                         param->fp_max_depth = 0;
2513                         break;
2514                 case 'D':
2515                         param->fp_get_default_lmv = 1;
2516                         break;
2517                 case 'E':
2518                         if (optarg != NULL) {
2519                                 tmp = optarg;
2520                                 if (tmp[0] == '+') {
2521                                         param->fp_comp_end_sign = -1;
2522                                         tmp++;
2523                                 } else if (tmp[0] == '-') {
2524                                         param->fp_comp_end_sign = 1;
2525                                         tmp++;
2526                                 }
2527
2528                                 if (arg_is_eof(tmp)) {
2529                                         param->fp_comp_end = LUSTRE_EOF;
2530                                         param->fp_comp_end_units = 1;
2531                                         rc = 0;
2532                                 } else {
2533                                         rc = llapi_parse_size(tmp,
2534                                                 &param->fp_comp_end,
2535                                                 &param->fp_comp_end_units, 0);
2536                                 }
2537                                 if (rc != 0) {
2538                                         fprintf(stderr, "error: %s bad "
2539                                                 "component end '%s'.\n",
2540                                                 argv[0], tmp);
2541                                         return CMD_HELP;
2542                                 }
2543                                 param->fp_check_comp_end = 1;
2544                         } else {
2545                                 param->fp_verbose |= VERBOSE_COMP_END;
2546                                 param->fp_max_depth = 0;
2547                         }
2548                         break;
2549                 case 'F':
2550                         if (!(param->fp_verbose & VERBOSE_DETAIL)) {
2551                                 param->fp_verbose |= VERBOSE_DFID;
2552                                 param->fp_max_depth = 0;
2553                         }
2554                         break;
2555                 case 'g':
2556                         if (!(param->fp_verbose & VERBOSE_DETAIL)) {
2557                                 param->fp_verbose |= VERBOSE_GENERATION;
2558                                 param->fp_max_depth = 0;
2559                         }
2560                         break;
2561 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
2562                 case 'o':
2563                         fprintf(stderr, "warning: '--offset|-o' deprecated, "
2564                                 "use '--stripe-index|-i' instead\n");
2565 #endif
2566                 case 'i':
2567 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 6, 53, 0)
2568                         if (strcmp(argv[optind - 1], "--index") == 0)
2569                                 fprintf(stderr, "warning: '--index' deprecated"
2570                                         ", use '--stripe-index' instead\n");
2571 #endif
2572                         if (!(param->fp_verbose & VERBOSE_DETAIL)) {
2573                                 param->fp_verbose |= VERBOSE_OFFSET;
2574                                 param->fp_max_depth = 0;
2575                         }
2576                         break;
2577                 case 'I':
2578                         if (optarg != NULL) {
2579                                 param->fp_comp_id = strtoul(optarg, &end, 0);
2580                                 if (*end != '\0' || param->fp_comp_id == 0 ||
2581                                     param->fp_comp_id > LCME_ID_MAX) {
2582                                         fprintf(stderr, "error: %s bad "
2583                                                 "component id '%s'\n",
2584                                                 argv[0], optarg);
2585                                         return CMD_HELP;
2586                                 } else {
2587                                         param->fp_check_comp_id = 1;
2588                                 }
2589                         } else {
2590                                 param->fp_max_depth = 0;
2591                                 param->fp_verbose |= VERBOSE_COMP_ID;
2592                         }
2593                         break;
2594                 case 'L':
2595                         if (!(param->fp_verbose & VERBOSE_DETAIL)) {
2596                                 param->fp_verbose |= VERBOSE_LAYOUT;
2597                                 param->fp_max_depth = 0;
2598                         }
2599                         break;
2600 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2601                 case 'M':
2602 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 11, 53, 0)
2603                         fprintf(stderr, "warning: '-M' deprecated"
2604                                 ", use '-m' instead\n");
2605 #endif
2606 #endif
2607                 case 'm':
2608                         if (!(param->fp_verbose & VERBOSE_DETAIL))
2609                                 param->fp_max_depth = 0;
2610                         param->fp_verbose |= VERBOSE_MDTINDEX;
2611                         break;
2612                 case 'O':
2613                         if (param->fp_obd_uuid) {
2614                                 fprintf(stderr,
2615                                         "error: %s: only one obduuid allowed",
2616                                         argv[0]);
2617                                 return CMD_HELP;
2618                         }
2619                         param->fp_obd_uuid = (struct obd_uuid *)optarg;
2620                         break;
2621                 case 'p':
2622                         if (!(param->fp_verbose & VERBOSE_DETAIL)) {
2623                                 param->fp_verbose |= VERBOSE_POOL;
2624                                 param->fp_max_depth = 0;
2625                         }
2626                         break;
2627                 case 'q':
2628                         param->fp_quiet++;
2629                         break;
2630                 case 'r':
2631                         param->fp_recursive = 1;
2632                         break;
2633                 case 'R':
2634                         param->fp_raw = 1;
2635                         break;
2636 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0)
2637                 case 's':
2638                         fprintf(stderr, "warning: '--size|-s' deprecated, "
2639                                 "use '--stripe-size|-S' instead\n");
2640 #endif /* LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 9, 59, 0) */
2641                 case 'S':
2642                         if (!(param->fp_verbose & VERBOSE_DETAIL)) {
2643                                 param->fp_verbose |= VERBOSE_SIZE;
2644                                 param->fp_max_depth = 0;
2645                         }
2646                         break;
2647                 case 'v':
2648                         param->fp_verbose = VERBOSE_DEFAULT | VERBOSE_DETAIL;
2649                         break;
2650                 case 'y':
2651                         param->fp_yaml = 1;
2652                         break;
2653                 default:
2654                         return CMD_HELP;
2655                 }
2656         }
2657
2658         if (optind >= argc)
2659                 return CMD_HELP;
2660
2661         if (param->fp_recursive)
2662                 param->fp_max_depth = -1;
2663         else if (param->fp_verbose & VERBOSE_DETAIL)
2664                 param->fp_max_depth = 1;
2665
2666         if (!param->fp_verbose)
2667                 param->fp_verbose = VERBOSE_DEFAULT;
2668         if (param->fp_quiet)
2669                 param->fp_verbose = VERBOSE_OBJID;
2670
2671         do {
2672                 rc = llapi_getstripe(argv[optind], param);
2673         } while (++optind < argc && !rc);
2674
2675         if (rc)
2676                 fprintf(stderr, "error: %s failed for %s.\n",
2677                         argv[0], argv[optind - 1]);
2678         return rc;
2679 }
2680
2681 static int lfs_tgts(int argc, char **argv)
2682 {
2683         char mntdir[PATH_MAX] = {'\0'}, path[PATH_MAX] = {'\0'};
2684         struct find_param param;
2685         int index = 0, rc=0;
2686
2687         if (argc > 2)
2688                 return CMD_HELP;
2689
2690         if (argc == 2 && !realpath(argv[1], path)) {
2691                 rc = -errno;
2692                 fprintf(stderr, "error: invalid path '%s': %s\n",
2693                         argv[1], strerror(-rc));
2694                 return rc;
2695         }
2696
2697         while (!llapi_search_mounts(path, index++, mntdir, NULL)) {
2698                 /* Check if we have a mount point */
2699                 if (mntdir[0] == '\0')
2700                         continue;
2701
2702                 memset(&param, 0, sizeof(param));
2703                 if (!strcmp(argv[0], "mdts"))
2704                         param.fp_get_lmv = 1;
2705
2706                 rc = llapi_ostlist(mntdir, &param);
2707                 if (rc) {
2708                         fprintf(stderr, "error: %s: failed on %s\n",
2709                                 argv[0], mntdir);
2710                 }
2711                 if (path[0] != '\0')
2712                         break;
2713                 memset(mntdir, 0, PATH_MAX);
2714         }
2715
2716         return rc;
2717 }
2718
2719 static int lfs_getstripe(int argc, char **argv)
2720 {
2721         struct find_param param = { 0 };
2722
2723         param.fp_max_depth = 1;
2724         return lfs_getstripe_internal(argc, argv, &param);
2725 }
2726
2727 /* functions */
2728 static int lfs_getdirstripe(int argc, char **argv)
2729 {
2730         struct find_param param = { 0 };
2731         struct option long_opts[] = {
2732 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2733                 {"mdt-count",   no_argument,            0, 'c'},
2734 #endif
2735                 {"mdt-hash",    no_argument,            0, 'H'},
2736                 {"mdt-index",   no_argument,            0, 'i'},
2737                 {"recursive",   no_argument,            0, 'r'},
2738 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2739                 {"mdt-hash",    no_argument,            0, 't'},
2740 #endif
2741                 {"default",     no_argument,            0, 'D'},
2742                 {"obd",         required_argument,      0, 'O'},
2743                 {"mdt-count",   no_argument,            0, 'T'},
2744                 {"yaml",        no_argument,            0, 'y'},
2745                 {0, 0, 0, 0}
2746         };
2747         int c, rc;
2748
2749         param.fp_get_lmv = 1;
2750
2751         while ((c = getopt_long(argc, argv,
2752                                 "cDHiO:rtTy", long_opts, NULL)) != -1)
2753         {
2754                 switch (c) {
2755                 case 'O':
2756                         if (param.fp_obd_uuid) {
2757                                 fprintf(stderr,
2758                                         "error: %s: only one obduuid allowed",
2759                                         argv[0]);
2760                                 return CMD_HELP;
2761                         }
2762                         param.fp_obd_uuid = (struct obd_uuid *)optarg;
2763                         break;
2764 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2765                 case 'c':
2766 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 10, 50, 0)
2767                         fprintf(stderr, "warning: '-c' deprecated"
2768                                 ", use '-T' instead\n");
2769 #endif
2770 #endif
2771                 case 'T':
2772                         param.fp_verbose |= VERBOSE_COUNT;
2773                         break;
2774                 case 'i':
2775                         param.fp_verbose |= VERBOSE_OFFSET;
2776                         break;
2777 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2778                 case 't':
2779 #endif
2780                 case 'H':
2781                         param.fp_verbose |= VERBOSE_HASH_TYPE;
2782                         break;
2783                 case 'D':
2784                         param.fp_get_default_lmv = 1;
2785                         break;
2786                 case 'r':
2787                         param.fp_recursive = 1;
2788                         break;
2789                 case 'y':
2790                         param.fp_yaml = 1;
2791                         break;
2792                 default:
2793                         return CMD_HELP;
2794                 }
2795         }
2796
2797         if (optind >= argc)
2798                 return CMD_HELP;
2799
2800         if (param.fp_recursive)
2801                 param.fp_max_depth = -1;
2802
2803         if (!param.fp_verbose)
2804                 param.fp_verbose = VERBOSE_DEFAULT;
2805
2806         do {
2807                 rc = llapi_getstripe(argv[optind], &param);
2808         } while (++optind < argc && !rc);
2809
2810         if (rc)
2811                 fprintf(stderr, "error: %s failed for %s.\n",
2812                         argv[0], argv[optind - 1]);
2813         return rc;
2814 }
2815
2816 /* functions */
2817 static int lfs_setdirstripe(int argc, char **argv)
2818 {
2819         char                    *dname;
2820         int                     result;
2821         unsigned int            stripe_offset = -1;
2822         unsigned int            stripe_count = 1;
2823         enum lmv_hash_type      hash_type;
2824         char                    *end;
2825         int                     c;
2826         char                    *stripe_offset_opt = NULL;
2827         char                    *stripe_count_opt = NULL;
2828         char                    *stripe_hash_opt = NULL;
2829         char                    *mode_opt = NULL;
2830         bool                    default_stripe = false;
2831         mode_t                  mode = S_IRWXU | S_IRWXG | S_IRWXO;
2832         mode_t                  previous_mode = 0;
2833         bool                    delete = false;
2834
2835         struct option long_opts[] = {
2836 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2837                 {"count",       required_argument, 0, 'c'},
2838 #endif
2839                 {"mdt-count",   required_argument, 0, 'c'},
2840                 {"delete",      no_argument, 0, 'd'},
2841 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2842                 {"index",       required_argument, 0, 'i'},
2843 #endif
2844                 {"mdt-index",   required_argument, 0, 'i'},
2845                 {"mode",        required_argument, 0, 'm'},
2846 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2847                 {"hash-type",   required_argument, 0, 't'},
2848                 {"mdt-hash",    required_argument, 0, 't'},
2849 #endif
2850                 {"mdt-hash",    required_argument, 0, 'H'},
2851 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2852                 {"default_stripe", no_argument, 0, 'D'},
2853 #endif
2854                 {"default",     no_argument, 0, 'D'},
2855                 {0, 0, 0, 0}
2856         };
2857
2858         while ((c = getopt_long(argc, argv, "c:dDi:H:m:t:", long_opts,
2859                                 NULL)) >= 0) {
2860                 switch (c) {
2861                 case 0:
2862                         /* Long options. */
2863                         break;
2864                 case 'c':
2865 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 11, 53, 0)
2866                         if (strcmp(argv[optind - 1], "--count") == 0)
2867                                 fprintf(stderr, "warning: '--count' deprecated"
2868                                         ", use '--mdt-count' instead\n");
2869 #endif
2870                         stripe_count_opt = optarg;
2871                         break;
2872                 case 'd':
2873                         delete = true;
2874                         default_stripe = true;
2875                         break;
2876                 case 'D':
2877                         default_stripe = true;
2878                         break;
2879                 case 'i':
2880 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 11, 53, 0)
2881                         if (strcmp(argv[optind - 1], "--index") == 0)
2882                                 fprintf(stderr, "warning: '--index' deprecated"
2883                                         ", use '--mdt-index' instead\n");
2884 #endif
2885                         stripe_offset_opt = optarg;
2886                         break;
2887                 case 'm':
2888                         mode_opt = optarg;
2889                         break;
2890 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
2891                 case 't':
2892 #endif
2893                 case 'H':
2894 #if LUSTRE_VERSION_CODE >= OBD_OCD_VERSION(2, 11, 53, 0)
2895                         if (strcmp(argv[optind - 1], "--hash-type") == 0)
2896                                 fprintf(stderr, "warning: '--hash-type' "
2897                                         "deprecated, use '--mdt-hash' "
2898                                         "instead\n");
2899 #endif
2900                         stripe_hash_opt = optarg;
2901                         break;
2902                 default:
2903                         fprintf(stderr, "error: %s: option '%s' "
2904                                         "unrecognized\n",
2905                                         argv[0], argv[optind - 1]);
2906                         return CMD_HELP;
2907                 }
2908         }
2909
2910         if (optind == argc) {
2911                 fprintf(stderr, "error: %s: missing dirname\n",
2912                         argv[0]);
2913                 return CMD_HELP;
2914         }
2915
2916         if (!delete && stripe_offset_opt == NULL && stripe_count_opt == NULL) {
2917                 fprintf(stderr, "error: %s: missing stripe offset and count.\n",
2918                         argv[0]);
2919                 return CMD_HELP;
2920         }
2921
2922         if (stripe_offset_opt != NULL) {
2923                 /* get the stripe offset */
2924                 stripe_offset = strtoul(stripe_offset_opt, &end, 0);
2925                 if (*end != '\0') {
2926                         fprintf(stderr, "error: %s: bad stripe offset '%s'\n",
2927                                 argv[0], stripe_offset_opt);
2928                         return CMD_HELP;
2929                 }
2930         }
2931
2932         if (delete) {
2933                 if (stripe_offset_opt != NULL || stripe_count_opt != NULL) {
2934                         fprintf(stderr, "error: %s: cannot specify -d with -s,"
2935                                 " or -i options.\n", argv[0]);
2936                         return CMD_HELP;
2937                 } else {
2938                         stripe_count = 0;
2939                 }
2940         }
2941
2942
2943         if (mode_opt != NULL) {
2944                 mode = strtoul(mode_opt, &end, 8);
2945                 if (*end != '\0') {
2946                         fprintf(stderr, "error: %s: bad mode '%s'\n",
2947                                 argv[0], mode_opt);
2948                         return CMD_HELP;
2949                 }
2950                 previous_mode = umask(0);
2951         }
2952
2953         if (stripe_hash_opt == NULL) {
2954                 hash_type = LMV_HASH_TYPE_FNV_1A_64;
2955         } else {
2956                 hash_type = check_hashtype(stripe_hash_opt);
2957                 if (hash_type == 0) {
2958                         fprintf(stderr,
2959                                 "error: %s: bad stripe hash type '%s'\n",
2960                                 argv[0], stripe_hash_opt);
2961                         return CMD_HELP;
2962                 }
2963         }
2964
2965         /* get the stripe count */
2966         if (stripe_count_opt != NULL) {
2967                 stripe_count = strtoul(stripe_count_opt, &end, 0);
2968                 if (*end != '\0') {
2969                         fprintf(stderr, "error: %s: bad stripe count '%s'\n",
2970                                 argv[0], stripe_count_opt);
2971                         return CMD_HELP;
2972                 }
2973         }
2974
2975         dname = argv[optind];
2976         do {
2977                 if (default_stripe) {
2978                         result = llapi_dir_set_default_lmv_stripe(dname,
2979                                                     stripe_offset, stripe_count,
2980                                                     hash_type, NULL);
2981                 } else {
2982                         result = llapi_dir_create_pool(dname, mode,
2983                                                        stripe_offset,
2984                                                        stripe_count, hash_type,
2985                                                        NULL);
2986                 }
2987
2988                 if (result) {
2989                         fprintf(stderr, "error: %s: create stripe dir '%s' "
2990                                 "failed\n", argv[0], dname);
2991                         break;
2992                 }
2993                 dname = argv[++optind];
2994         } while (dname != NULL);
2995
2996         if (mode_opt != NULL)
2997                 umask(previous_mode);
2998
2999         return result;
3000 }
3001
3002 /* functions */
3003 static int lfs_rmentry(int argc, char **argv)
3004 {
3005         char *dname;
3006         int   index;
3007         int   result = 0;
3008
3009         if (argc <= 1) {
3010                 fprintf(stderr, "error: %s: missing dirname\n",
3011                         argv[0]);
3012                 return CMD_HELP;
3013         }
3014
3015         index = 1;
3016         dname = argv[index];
3017         while (dname != NULL) {
3018                 result = llapi_direntry_remove(dname);
3019                 if (result) {
3020                         fprintf(stderr, "error: %s: remove dir entry '%s' "
3021                                 "failed\n", argv[0], dname);
3022                         break;
3023                 }
3024                 dname = argv[++index];
3025         }
3026         return result;
3027 }
3028
3029 static int lfs_mv(int argc, char **argv)
3030 {
3031         struct  find_param param = {
3032                 .fp_max_depth = -1,
3033                 .fp_mdt_index = -1,
3034         };
3035         char   *end;
3036         int     c;
3037         int     rc = 0;
3038         struct option long_opts[] = {
3039                 {"mdt-index", required_argument, 0, 'M'},
3040                 {"verbose",     no_argument,       0, 'v'},
3041                 {0, 0, 0, 0}
3042         };
3043
3044         while ((c = getopt_long(argc, argv, "M:v", long_opts, NULL)) != -1) {
3045                 switch (c) {
3046                 case 'M': {
3047                         param.fp_mdt_index = strtoul(optarg, &end, 0);
3048                         if (*end != '\0') {
3049                                 fprintf(stderr, "%s: invalid MDT index'%s'\n",
3050                                         argv[0], optarg);
3051                                 return CMD_HELP;
3052                         }
3053                         break;
3054                 }
3055                 case 'v': {
3056                         param.fp_verbose = VERBOSE_DETAIL;
3057                         break;
3058                 }
3059                 default:
3060                         fprintf(stderr, "error: %s: unrecognized option '%s'\n",
3061                                 argv[0], argv[optind - 1]);
3062                         return CMD_HELP;
3063                 }
3064         }
3065
3066         if (param.fp_mdt_index == -1) {
3067                 fprintf(stderr, "%s: MDT index must be specified\n", argv[0]);
3068                 return CMD_HELP;
3069         }
3070
3071         if (optind >= argc) {
3072                 fprintf(stderr, "%s: missing operand path\n", argv[0]);
3073                 return CMD_HELP;
3074         }
3075
3076         param.fp_migrate = 1;
3077         rc = llapi_migrate_mdt(argv[optind], &param);
3078         if (rc != 0)
3079                 fprintf(stderr, "%s: cannot migrate '%s' to MDT%04x: %s\n",
3080                         argv[0], argv[optind], param.fp_mdt_index,
3081                         strerror(-rc));
3082         return rc;
3083 }
3084
3085 static int lfs_osts(int argc, char **argv)
3086 {
3087         return lfs_tgts(argc, argv);
3088 }
3089
3090 static int lfs_mdts(int argc, char **argv)
3091 {
3092         return lfs_tgts(argc, argv);
3093 }
3094
3095 #define COOK(value)                                                     \
3096 ({                                                                      \
3097         int radix = 0;                                                  \
3098         while (value > 1024) {                                          \
3099                 value /= 1024;                                          \
3100                 radix++;                                                \
3101         }                                                               \
3102         radix;                                                          \
3103 })
3104 #define UUF     "%-20s"
3105 #define CSF     "%11s"
3106 #define CDF     "%11llu"
3107 #define HDF     "%8.1f%c"
3108 #define RSF     "%4s"
3109 #define RDF     "%3d%%"
3110
3111 enum mntdf_flags {
3112         MNTDF_INODES    = 0x0001,
3113         MNTDF_COOKED    = 0x0002,
3114         MNTDF_LAZY      = 0x0004,
3115         MNTDF_VERBOSE   = 0x0008,
3116 };
3117
3118 static int showdf(char *mntdir, struct obd_statfs *stat,
3119                   char *uuid, enum mntdf_flags flags,
3120                   char *type, int index, int rc)
3121 {
3122         long long avail, used, total;
3123         double ratio = 0;
3124         char *suffix = "KMGTPEZY";
3125         /* Note if we have >2^64 bytes/fs these buffers will need to be grown */
3126         char tbuf[3 * sizeof(__u64)];
3127         char ubuf[3 * sizeof(__u64)];
3128         char abuf[3 * sizeof(__u64)];
3129         char rbuf[3 * sizeof(__u64)];
3130
3131         if (!uuid || !stat)
3132                 return -EINVAL;
3133
3134         switch (rc) {
3135         case 0:
3136                 if (flags & MNTDF_INODES) {
3137                         avail = stat->os_ffree;
3138                         used = stat->os_files - stat->os_ffree;
3139                         total = stat->os_files;
3140                 } else {
3141                         int shift = flags & MNTDF_COOKED ? 0 : 10;
3142
3143                         avail = (stat->os_bavail * stat->os_bsize) >> shift;
3144                         used  = ((stat->os_blocks - stat->os_bfree) *
3145                                  stat->os_bsize) >> shift;
3146                         total = (stat->os_blocks * stat->os_bsize) >> shift;
3147                 }
3148
3149                 if ((used + avail) > 0)
3150                         ratio = (double)used / (double)(used + avail);
3151
3152                 if (flags & MNTDF_COOKED) {
3153                         int i;
3154                         double cook_val;
3155
3156                         cook_val = (double)total;
3157                         i = COOK(cook_val);
3158                         if (i > 0)
3159                                 snprintf(tbuf, sizeof(tbuf), HDF, cook_val,
3160                                          suffix[i - 1]);
3161                         else
3162                                 snprintf(tbuf, sizeof(tbuf), CDF, total);
3163
3164                         cook_val = (double)used;
3165                         i = COOK(cook_val);
3166                         if (i > 0)
3167                                 snprintf(ubuf, sizeof(ubuf), HDF, cook_val,
3168                                          suffix[i - 1]);
3169                         else
3170                                 snprintf(ubuf, sizeof(ubuf), CDF, used);
3171
3172                         cook_val = (double)avail;
3173                         i = COOK(cook_val);
3174                         if (i > 0)
3175                                 snprintf(abuf, sizeof(abuf), HDF, cook_val,
3176                                          suffix[i - 1]);
3177                         else
3178                                 snprintf(abuf, sizeof(abuf), CDF, avail);
3179                 } else {
3180                         snprintf(tbuf, sizeof(tbuf), CDF, total);
3181                         snprintf(ubuf, sizeof(tbuf), CDF, used);
3182                         snprintf(abuf, sizeof(tbuf), CDF, avail);
3183                 }
3184
3185                 sprintf(rbuf, RDF, (int)(ratio * 100 + 0.5));
3186                 printf(UUF" "CSF" "CSF" "CSF" "RSF" %-s",
3187                        uuid, tbuf, ubuf, abuf, rbuf, mntdir);
3188                 if (type)
3189                         printf("[%s:%d]", type, index);
3190
3191                 if (stat->os_state) {
3192                         /*
3193                          * Each character represents the matching
3194                          * OS_STATE_* bit.
3195                          */
3196                         const char state_names[] = "DRSI";
3197                         __u32      state;
3198                         __u32      i;
3199
3200                         printf(" ");
3201                         for (i = 0, state = stat->os_state;
3202                              state && i < sizeof(state_names); i++) {
3203                                 if (!(state & (1 << i)))
3204                                         continue;
3205                                 printf("%c", state_names[i]);
3206                                 state ^= 1 << i;
3207                         }
3208                 }
3209
3210                 printf("\n");
3211                 break;
3212         case -ENODATA:
3213                 printf(UUF": inactive device\n", uuid);
3214                 break;
3215         default:
3216                 printf(UUF": %s\n", uuid, strerror(-rc));
3217                 break;
3218         }
3219
3220         return 0;
3221 }
3222
3223 struct ll_stat_type {
3224         int   st_op;
3225         char *st_name;
3226 };
3227
3228 static int mntdf(char *mntdir, char *fsname, char *pool, enum mntdf_flags flags)
3229 {
3230         struct obd_statfs stat_buf, sum = { .os_bsize = 1 };
3231         struct obd_uuid uuid_buf;
3232         char *poolname = NULL;
3233         struct ll_stat_type types[] = { { LL_STATFS_LMV, "MDT" },
3234                                         { LL_STATFS_LOV, "OST" },
3235                                         { 0, NULL } };
3236         struct ll_stat_type *tp;
3237         __u64 ost_ffree = 0;
3238         __u32 index;
3239         __u32 type;
3240         int fd;
3241         int rc = 0;
3242         int rc2;
3243
3244         if (pool) {
3245                 poolname = strchr(pool, '.');
3246                 if (poolname != NULL) {
3247                         if (strncmp(fsname, pool, strlen(fsname))) {
3248                                 fprintf(stderr, "filesystem name incorrect\n");
3249                                 return -ENODEV;
3250                         }
3251                         poolname++;
3252                 } else
3253                         poolname = pool;
3254         }
3255
3256         fd = open(mntdir, O_RDONLY);
3257         if (fd < 0) {
3258                 rc = -errno;
3259                 fprintf(stderr, "%s: cannot open '%s': %s\n", progname, mntdir,
3260                         strerror(errno));
3261                 return rc;
3262         }
3263
3264         if (flags & MNTDF_INODES)
3265                 printf(UUF" "CSF" "CSF" "CSF" "RSF" %-s\n",
3266                        "UUID", "Inodes", "IUsed", "IFree",
3267                        "IUse%", "Mounted on");
3268         else
3269                 printf(UUF" "CSF" "CSF" "CSF" "RSF" %-s\n",
3270                        "UUID", flags & MNTDF_COOKED ? "bytes" : "1K-blocks",
3271                        "Used", "Available", "Use%", "Mounted on");
3272
3273         for (tp = types; tp->st_name != NULL; tp++) {
3274                 for (index = 0; ; index++) {
3275                         memset(&stat_buf, 0, sizeof(struct obd_statfs));
3276                         memset(&uuid_buf, 0, sizeof(struct obd_uuid));
3277                         type = flags & MNTDF_LAZY ?
3278                                 tp->st_op | LL_STATFS_NODELAY : tp->st_op;
3279                         rc2 = llapi_obd_fstatfs(fd, type, index,
3280                                                &stat_buf, &uuid_buf);
3281                         if (rc2 == -ENODEV)
3282                                 break;
3283                         if (rc2 == -EAGAIN)
3284                                 continue;
3285                         if (rc2 == -ENODATA) { /* Inactive device, OK. */
3286                                 if (!(flags & MNTDF_VERBOSE))
3287                                         continue;
3288                         } else if (rc2 < 0 && rc == 0) {
3289                                 rc = rc2;
3290                         }
3291
3292                         if (poolname && tp->st_op == LL_STATFS_LOV &&
3293                             llapi_search_ost(fsname, poolname,
3294                                              obd_uuid2str(&uuid_buf)) != 1)
3295                                 continue;
3296
3297                         /* the llapi_obd_statfs() call may have returned with
3298                          * an error, but if it filled in uuid_buf we will at
3299                          * lease use that to print out a message for that OBD.
3300                          * If we didn't get anything in the uuid_buf, then fill
3301                          * it in so that we can print an error message. */
3302                         if (uuid_buf.uuid[0] == '\0')
3303                                 snprintf(uuid_buf.uuid, sizeof(uuid_buf.uuid),
3304                                          "%s%04x", tp->st_name, index);
3305                         showdf(mntdir, &stat_buf, obd_uuid2str(&uuid_buf),
3306                                flags, tp->st_name, index, rc2);
3307
3308                         if (rc2 == 0) {
3309                                 if (tp->st_op == LL_STATFS_LMV) {
3310                                         sum.os_ffree += stat_buf.os_ffree;
3311                                         sum.os_files += stat_buf.os_files;
3312                                 } else /* if (tp->st_op == LL_STATFS_LOV) */ {
3313                                         sum.os_blocks += stat_buf.os_blocks *
3314                                                 stat_buf.os_bsize;
3315                                         sum.os_bfree  += stat_buf.os_bfree *
3316                                                 stat_buf.os_bsize;
3317                                         sum.os_bavail += stat_buf.os_bavail *
3318                                                 stat_buf.os_bsize;
3319                                         ost_ffree += stat_buf.os_ffree;
3320                                 }
3321                         }
3322                 }
3323         }
3324
3325         close(fd);
3326
3327         /* If we don't have as many objects free on the OST as inodes
3328          * on the MDS, we reduce the total number of inodes to
3329          * compensate, so that the "inodes in use" number is correct.
3330          * Matches ll_statfs_internal() so the results are consistent. */
3331         if (ost_ffree < sum.os_ffree) {
3332                 sum.os_files = (sum.os_files - sum.os_ffree) + ost_ffree;
3333                 sum.os_ffree = ost_ffree;
3334         }
3335         printf("\n");
3336         showdf(mntdir, &sum, "filesystem_summary:", flags, NULL, 0, 0);
3337         printf("\n");
3338
3339         return rc;
3340 }
3341
3342 static int lfs_df(int argc, char **argv)
3343 {
3344         char mntdir[PATH_MAX] = {'\0'}, path[PATH_MAX] = {'\0'};
3345         enum mntdf_flags flags = 0;
3346         int c, rc = 0, index = 0;
3347         char fsname[PATH_MAX] = "", *pool_name = NULL;
3348         struct option long_opts[] = {
3349                 {"human-readable", 0, 0, 'h'},
3350                 {"inodes", 0, 0, 'i'},
3351                 {"lazy", 0, 0, 'l'},
3352                 {"pool", required_argument, 0, 'p'},
3353                 {"verbose", 0, 0, 'v'},
3354