lustre/utils/lhsmtool_posix.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.htm
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * (C) Copyright 2012 Commissariat a l'energie atomique et aux energies
  24  *     alternatives
  25  *
  26  * Copyright (c) 2013, 2016, Intel Corporation.
  27  */
  28 /* HSM copytool program for POSIX filesystem-based HSM's.
  29  *
  30  * An HSM copytool daemon acts on action requests from Lustre to copy files
  31  * to and from an HSM archive system. This one in particular makes regular
  32  * POSIX filesystem calls to a given path, where an HSM is presumably mounted.
  33  *
  34  * This particular tool can also import an existing HSM archive.
  35  */
  36
  37 #ifndef _GNU_SOURCE
  38 #define _GNU_SOURCE
  39 #endif
  40 #include <ctype.h>
  41 #include <fcntl.h>
  42 #include <stdio.h>
  43 #include <stdint.h>
  44 #include <stdlib.h>
  45 #include <string.h>
  46 #include <dirent.h>
  47 #include <errno.h>
  48 #include <getopt.h>
  49 #include <pthread.h>
  50 #include <time.h>
  51 #include <unistd.h>
  52 #include <utime.h>
  53 #include <signal.h>
  54 #include <sys/time.h>
  55 #include <sys/xattr.h>
  56 #include <sys/syscall.h>
  57 #include <sys/types.h>
  58
  59 #include <libcfs/util/string.h>
  60 #include <linux/lustre/lustre_fid.h>
  61 #include <lustre/lustreapi.h>
  62
  63 /* Progress reporting period */
  64 #define REPORT_INTERVAL_DEFAULT 30
  65 /* HSM hash subdir permissions */
  66 #define DIR_PERM S_IRWXU
  67 /* HSM hash file permissions */
  68 #define FILE_PERM (S_IRUSR | S_IWUSR)
  69
  70 #define ONE_MB 0x100000
  71
  72 #ifndef NSEC_PER_SEC
  73 # define NSEC_PER_SEC 1000000000UL
  74 #endif
  75
  76 enum ct_action {
  77         CA_IMPORT = 1,
  78         CA_REBIND,
  79         CA_MAXSEQ,
  80 };
  81
  82 struct options {
  83         int                      o_copy_attrs;
  84         int                      o_daemonize;
  85         int                      o_dry_run;
  86         int                      o_abort_on_error;
  87         int                      o_shadow_tree;
  88         int                      o_verbose;
  89         int                      o_copy_xattrs;
  90         int                      o_archive_id_used;
  91         int                      o_archive_id_cnt;
  92         int                     *o_archive_id;
  93         int                      o_report_int;
  94         unsigned long long       o_bandwidth;
  95         size_t                   o_chunk_size;
  96         enum ct_action           o_action;
  97         char                    *o_event_fifo;
  98         char                    *o_mnt;
  99         int                      o_mnt_fd;
 100         char                    *o_hsm_root;
 101         char                    *o_src; /* for import, or rebind */
 102         char                    *o_dst; /* for import, or rebind */
 103 };
 104
 105 /* everything else is zeroed */
 106 struct options opt = {
 107         .o_copy_attrs = 1,
 108         .o_shadow_tree = 1,
 109         .o_verbose = LLAPI_MSG_INFO,
 110         .o_copy_xattrs = 1,
 111         .o_report_int = REPORT_INTERVAL_DEFAULT,
 112         .o_chunk_size = ONE_MB,
 113 };
 114
 115 /* hsm_copytool_private will hold an open FD on the lustre mount point
 116  * for us. Additionally open one on the archive FS root to make sure
 117  * it doesn't drop out from under us (and remind the admin to shutdown
 118  * the copytool before unmounting). */
 119
 120 static int arc_fd = -1;
 121
 122 static int err_major;
 123 static int err_minor;
 124
 125 static char cmd_name[PATH_MAX];
 126 static char fs_name[MAX_OBD_NAME + 1];
 127
 128 static struct hsm_copytool_private *ctdata;
 129
 130 static inline double ct_now(void)
 131 {
 132         struct timeval tv;
 133
 134         gettimeofday(&tv, NULL);
 135
 136         return tv.tv_sec + 0.000001 * tv.tv_usec;
 137 }
 138
 139 #define CT_ERROR(_rc, _format, ...)                                     \
 140         llapi_error(LLAPI_MSG_ERROR, _rc,                               \
 141                     "%f %s[%ld]: "_format,                              \
 142                     ct_now(), cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 143
 144 #define CT_DEBUG(_format, ...)                                          \
 145         llapi_error(LLAPI_MSG_DEBUG | LLAPI_MSG_NO_ERRNO, 0,            \
 146                     "%f %s[%ld]: "_format,                              \
 147                     ct_now(), cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 148
 149 #define CT_WARN(_format, ...) \
 150         llapi_error(LLAPI_MSG_WARN | LLAPI_MSG_NO_ERRNO, 0,             \
 151                     "%f %s[%ld]: "_format,                              \
 152                     ct_now(), cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 153
 154 #define CT_TRACE(_format, ...)                                          \
 155         llapi_error(LLAPI_MSG_INFO | LLAPI_MSG_NO_ERRNO, 0,             \
 156                     "%f %s[%ld]: "_format,                              \
 157                     ct_now(), cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 158
 159 static void usage(const char *name, int rc)
 160 {
 161         fprintf(stdout,
 162         " Usage: %s [options]... <mode> <lustre_mount_point>\n"
 163         "The Lustre HSM Posix copy tool can be used as a daemon or "
 164         "as a command line tool\n"
 165         "The Lustre HSM daemon acts on action requests from Lustre\n"
 166         "to copy files to and from an HSM archive system.\n"
 167         "This POSIX-flavored daemon makes regular POSIX filesystem calls\n"
 168         "to an HSM mounted at a given hsm_root.\n"
 169         "   --daemon            Daemon mode, run in background\n"
 170         " Options:\n"
 171         "   --no-attr           Don't copy file attributes\n"
 172         "   --no-shadow         Don't create shadow namespace in archive\n"
 173         "   --no-xattr          Don't copy file extended attributes\n"
 174         "The Lustre HSM tool performs administrator-type actions\n"
 175         "on a Lustre HSM archive.\n"
 176         "This POSIX-flavored tool can link an existing HSM namespace\n"
 177         "into a Lustre filesystem.\n"
 178         " Usage:\n"
 179         "   %s [options] --import <src> <dst> <lustre_mount_point>\n"
 180         "      import an archived subtree from\n"
 181         "       <src> (FID or relative path to hsm_root) into the Lustre\n"
 182         "             filesystem at\n"
 183         "       <dst> (absolute path)\n"
 184         "   %s [options] --rebind <old_FID> <new_FID> <lustre_mount_point>\n"
 185         "      rebind an entry in the HSM to a new FID\n"
 186         "       <old_FID> old FID the HSM entry is bound to\n"
 187         "       <new_FID> new FID to bind the HSM entry to\n"
 188         "   %s [options] --rebind <list_file> <lustre_mount_point>\n"
 189         "      perform the rebind operation for all FID in the list file\n"
 190         "       each line of <list_file> consists of <old_FID> <new_FID>\n"
 191         "   %s [options] --max-sequence <fsname>\n"
 192         "       return the max fid sequence of archived files\n"
 193         "   --abort-on-error          Abort operation on major error\n"
 194         "   -A, --archive <#>         Archive number (repeatable)\n"
 195         "   -b, --bandwidth <bw>      Limit I/O bandwidth (unit can be used\n,"
 196         "                             default is MB)\n"
 197         "   --dry-run                 Don't run, just show what would be done\n"
 198         "   -c, --chunk-size <sz>     I/O size used during data copy\n"
 199         "                             (unit can be used, default is MB)\n"
 200         "   -f, --event-fifo <path>   Write events stream to fifo\n"
 201         "   -p, --hsm-root <path>     Target HSM mount point\n"
 202         "   -q, --quiet               Produce less verbose output\n"
 203         "   -u, --update-interval <s> Interval between progress reports sent\n"
 204         "                             to Coordinator\n"
 205         "   -v, --verbose             Produce more verbose output\n",
 206         cmd_name, cmd_name, cmd_name, cmd_name, cmd_name);
 207
 208         exit(rc);
 209 }
 210
 211 static int ct_parseopts(int argc, char * const *argv)
 212 {
 213         struct option long_opts[] = {
 214         { .val = 1,     .name = "abort-on-error",
 215           .flag = &opt.o_abort_on_error,        .has_arg = no_argument },
 216         { .val = 1,     .name = "abort_on_error",
 217           .flag = &opt.o_abort_on_error,        .has_arg = no_argument },
 218         { .val = 'A',   .name = "archive",      .has_arg = required_argument },
 219         { .val = 'b',   .name = "bandwidth",    .has_arg = required_argument },
 220         { .val = 'c',   .name = "chunk-size",   .has_arg = required_argument },
 221         { .val = 'c',   .name = "chunk_size",   .has_arg = required_argument },
 222         { .val = 1,     .name = "daemon",       .has_arg = no_argument,
 223           .flag = &opt.o_daemonize },
 224         { .val = 'f',   .name = "event-fifo",   .has_arg = required_argument },
 225         { .val = 'f',   .name = "event_fifo",   .has_arg = required_argument },
 226         { .val = 1,     .name = "dry-run",      .has_arg = no_argument,
 227           .flag = &opt.o_dry_run },
 228         { .val = 'h',   .name = "help",         .has_arg = no_argument },
 229         { .val = 'i',   .name = "import",       .has_arg = no_argument },
 230         { .val = 'M',   .name = "max-sequence", .has_arg = no_argument },
 231         { .val = 'M',   .name = "max_sequence", .has_arg = no_argument },
 232         { .val = 0,     .name = "no-attr",      .has_arg = no_argument,
 233           .flag = &opt.o_copy_attrs },
 234         { .val = 0,     .name = "no_attr",      .has_arg = no_argument,
 235           .flag = &opt.o_copy_attrs },
 236         { .val = 0,     .name = "no-shadow",    .has_arg = no_argument,
 237           .flag = &opt.o_shadow_tree },
 238         { .val = 0,     .name = "no_shadow",    .has_arg = no_argument,
 239           .flag = &opt.o_shadow_tree },
 240         { .val = 0,     .name = "no-xattr",     .has_arg = no_argument,
 241           .flag = &opt.o_copy_xattrs },
 242         { .val = 0,     .name = "no_xattr",     .has_arg = no_argument,
 243           .flag = &opt.o_copy_xattrs },
 244         { .val = 'p',   .name = "hsm-root",     .has_arg = required_argument },
 245         { .val = 'p',   .name = "hsm_root",     .has_arg = required_argument },
 246         { .val = 'q',   .name = "quiet",        .has_arg = no_argument },
 247         { .val = 'r',   .name = "rebind",       .has_arg = no_argument },
 248         { .val = 'u',   .name = "update-interval",
 249                                                 .has_arg = required_argument },
 250         { .val = 'u',   .name = "update_interval",
 251                                                 .has_arg = required_argument },
 252         { .val = 'v',   .name = "verbose",      .has_arg = no_argument },
 253         { .name = NULL } };
 254         unsigned long long value;
 255         unsigned long long unit;
 256         bool all_id = false;
 257         int c, rc;
 258         int i;
 259
 260         optind = 0;
 261
 262         opt.o_archive_id_cnt = LL_HSM_ORIGIN_MAX_ARCHIVE;
 263         opt.o_archive_id = malloc(opt.o_archive_id_cnt *
 264                                   sizeof(*opt.o_archive_id));
 265         if (opt.o_archive_id == NULL)
 266                 return -ENOMEM;
 267 repeat:
 268         while ((c = getopt_long(argc, argv, "A:b:c:f:hiMp:qru:v",
 269                                 long_opts, NULL)) != -1) {
 270                 switch (c) {
 271                 case 'A': {
 272                         char *end = NULL;
 273                         int val = strtol(optarg, &end, 10);
 274
 275                         if (*end != '\0') {
 276                                 rc = -EINVAL;
 277                                 CT_ERROR(rc, "invalid archive-id: '%s'",
 278                                          optarg);
 279                                 return rc;
 280                         }
 281                         /* if archiveID is zero, any archiveID is accepted */
 282                         if (all_id == true)
 283                                 goto repeat;
 284
 285                         if (val == 0) {
 286                                 free(opt.o_archive_id);
 287                                 opt.o_archive_id = NULL;
 288                                 opt.o_archive_id_cnt = 0;
 289                                 opt.o_archive_id_used = 0;
 290                                 all_id = true;
 291                                 CT_WARN("archive-id = 0 is found, any backend"
 292                                         "will be served\n");
 293                                 goto repeat;
 294                         }
 295
 296                         /* skip the duplicated id */
 297                         for (i = 0; i < opt.o_archive_id_used; i++) {
 298                                 if (opt.o_archive_id[i] == val)
 299                                         goto repeat;
 300                         }
 301                         /* extend the space */
 302                         if (opt.o_archive_id_used >= opt.o_archive_id_cnt) {
 303                                 int *tmp;
 304
 305                                 opt.o_archive_id_cnt *= 2;
 306                                 tmp = realloc(opt.o_archive_id,
 307                                               sizeof(*opt.o_archive_id) *
 308                                               opt.o_archive_id_cnt);
 309                                 if (tmp == NULL)
 310                                         return -ENOMEM;
 311
 312                                 opt.o_archive_id = tmp;
 313                         }
 314
 315                         opt.o_archive_id[opt.o_archive_id_used++] = val;
 316                         break;
 317                 }
 318                 case 'b': /* -b and -c have both a number with unit as arg */
 319                 case 'c':
 320                         unit = ONE_MB;
 321                         if (llapi_parse_size(optarg, &value, &unit, 0) < 0) {
 322                                 rc = -EINVAL;
 323                                 CT_ERROR(rc, "bad value for -%c '%s'", c,
 324                                          optarg);
 325                                 return rc;
 326                         }
 327                         if (c == 'c')
 328                                 opt.o_chunk_size = value;
 329                         else
 330                                 opt.o_bandwidth = value;
 331                         break;
 332                 case 'f':
 333                         opt.o_event_fifo = optarg;
 334                         break;
 335                 case 'h':
 336                         usage(argv[0], 0);
 337                 case 'i':
 338                         opt.o_action = CA_IMPORT;
 339                         break;
 340                 case 'M':
 341                         opt.o_action = CA_MAXSEQ;
 342                         break;
 343                 case 'p':
 344                         opt.o_hsm_root = optarg;
 345                         break;
 346                 case 'q':
 347                         opt.o_verbose--;
 348                         break;
 349                 case 'r':
 350                         opt.o_action = CA_REBIND;
 351                         break;
 352                 case 'u':
 353                         opt.o_report_int = atoi(optarg);
 354                         if (opt.o_report_int < 0) {
 355                                 rc = -EINVAL;
 356                                 CT_ERROR(rc, "bad value for -%c '%s'", c,
 357                                          optarg);
 358                                 return rc;
 359                         }
 360                         break;
 361                 case 'v':
 362                         opt.o_verbose++;
 363                         break;
 364                 case 0:
 365                         break;
 366                 default:
 367                         return -EINVAL;
 368                 }
 369         }
 370
 371         switch (opt.o_action) {
 372         case CA_IMPORT:
 373                 /* src dst mount_point */
 374                 if (argc != optind + 3) {
 375                         rc = -EINVAL;
 376                         CT_ERROR(rc, "--import requires 2 arguments");
 377                         return rc;
 378                 }
 379                 opt.o_src = argv[optind++];
 380                 opt.o_dst = argv[optind++];
 381                 break;
 382         case CA_REBIND:
 383                 /* FID1 FID2 mount_point or FILE mount_point */
 384                 if (argc == optind + 2) {
 385                         opt.o_src = argv[optind++];
 386                         opt.o_dst = NULL;
 387                 } else if (argc == optind + 3) {
 388                         opt.o_src = argv[optind++];
 389                         opt.o_dst = argv[optind++];
 390                 } else {
 391                         rc = -EINVAL;
 392                         CT_ERROR(rc, "--rebind requires 1 or 2 arguments");
 393                         return rc;
 394                 }
 395                 break;
 396         case CA_MAXSEQ:
 397         default:
 398                 /* just mount point */
 399                 break;
 400         }
 401
 402         if (argc != optind + 1) {
 403                 rc = -EINVAL;
 404                 CT_ERROR(rc, "no mount point specified");
 405                 return rc;
 406         }
 407
 408         opt.o_mnt = argv[optind];
 409         opt.o_mnt_fd = -1;
 410
 411         CT_TRACE("action=%d src=%s dst=%s mount_point=%s",
 412                  opt.o_action, opt.o_src, opt.o_dst, opt.o_mnt);
 413
 414         if (opt.o_hsm_root == NULL) {
 415                 rc = -EINVAL;
 416                 CT_ERROR(rc, "must specify a root directory for the backend");
 417                 return rc;
 418         }
 419
 420         if (opt.o_action == CA_IMPORT) {
 421                 if (opt.o_src && opt.o_src[0] == '/') {
 422                         rc = -EINVAL;
 423                         CT_ERROR(rc,
 424                                  "source path must be relative to HSM root");
 425                         return rc;
 426                 }
 427
 428                 if (opt.o_dst && opt.o_dst[0] != '/') {
 429                         rc = -EINVAL;
 430                         CT_ERROR(rc, "destination path must be absolute");
 431                         return rc;
 432                 }
 433         }
 434
 435         return 0;
 436 }
 437
 438 /* mkdir -p path */
 439 static int ct_mkdir_p(const char *path)
 440 {
 441         char    *saved, *ptr;
 442         int      rc;
 443
 444         ptr = strdup(path);
 445         if (ptr == NULL)
 446                 return -errno;
 447
 448         saved = ptr;
 449         while (*ptr == '/')
 450                 ptr++;
 451
 452         while ((ptr = strchr(ptr, '/')) != NULL) {
 453                 *ptr = '\0';
 454                 rc = mkdir(saved, DIR_PERM);
 455                 *ptr = '/';
 456                 if (rc < 0 && errno != EEXIST) {
 457                         rc = -errno;
 458                         CT_ERROR(rc, "cannot mkdir '%s'", path);
 459                         free(saved);
 460                         return rc;
 461                 }
 462                 ptr++;
 463         }
 464
 465         free(saved);
 466
 467         return 0;
 468 }
 469
 470 static int ct_save_stripe(int src_fd, const char *src, const char *dst)
 471 {
 472         char                     lov_file[PATH_MAX + 8];
 473         char                     lov_buf[XATTR_SIZE_MAX];
 474         struct lov_user_md      *lum;
 475         int                      rc;
 476         int                      fd;
 477         ssize_t                  xattr_size;
 478
 479         snprintf(lov_file, sizeof(lov_file), "%s.lov", dst);
 480         CT_TRACE("saving stripe info of '%s' in %s", src, lov_file);
 481
 482         xattr_size = fgetxattr(src_fd, XATTR_LUSTRE_LOV, lov_buf,
 483                                sizeof(lov_buf));
 484         if (xattr_size < 0) {
 485                 rc = -errno;
 486                 CT_ERROR(rc, "cannot get stripe info on '%s'", src);
 487                 return rc;
 488         }
 489
 490         lum = (struct lov_user_md *)lov_buf;
 491
 492         if (lum->lmm_magic == LOV_USER_MAGIC_V1 ||
 493             lum->lmm_magic == LOV_USER_MAGIC_V3) {
 494                 /* Set stripe_offset to -1 so that it is not interpreted as a
 495                  * hint on restore. */
 496                 lum->lmm_stripe_offset = -1;
 497         }
 498
 499         fd = open(lov_file, O_TRUNC | O_CREAT | O_WRONLY, FILE_PERM);
 500         if (fd < 0) {
 501                 rc = -errno;
 502                 CT_ERROR(rc, "cannot open '%s'", lov_file);
 503                 goto err_cleanup;
 504         }
 505
 506         rc = write(fd, lum, xattr_size);
 507         if (rc < 0) {
 508                 rc = -errno;
 509                 CT_ERROR(rc, "cannot write %zd bytes to '%s'",
 510                          xattr_size, lov_file);
 511                 close(fd);
 512                 goto err_cleanup;
 513         }
 514
 515         rc = close(fd);
 516         if (rc < 0) {
 517                 rc = -errno;
 518                 CT_ERROR(rc, "cannot close '%s'", lov_file);
 519                 goto err_cleanup;
 520         }
 521
 522         return 0;
 523
 524 err_cleanup:
 525         unlink(lov_file);
 526
 527         return rc;
 528 }
 529
 530 static int ct_load_stripe(const char *src, void *lovea, size_t *lovea_size)
 531 {
 532         char     lov_file[PATH_MAX + 4];
 533         int      rc;
 534         int      fd;
 535
 536         snprintf(lov_file, sizeof(lov_file), "%s.lov", src);
 537         CT_TRACE("reading stripe rules from '%s' for '%s'", lov_file, src);
 538
 539         fd = open(lov_file, O_RDONLY);
 540         if (fd < 0) {
 541                 CT_ERROR(errno, "cannot open '%s'", lov_file);
 542                 return -ENODATA;
 543         }
 544
 545         rc = read(fd, lovea, *lovea_size);
 546         if (rc < 0) {
 547                 CT_ERROR(errno, "cannot read %zu bytes from '%s'",
 548                          *lovea_size, lov_file);
 549                 close(fd);
 550                 return -ENODATA;
 551         }
 552
 553         *lovea_size = rc;
 554         close(fd);
 555
 556         return 0;
 557 }
 558
 559 static int ct_restore_stripe(const char *src, const char *dst, int dst_fd,
 560                              const void *lovea, size_t lovea_size)
 561 {
 562         int     rc;
 563
 564         rc = fsetxattr(dst_fd, XATTR_LUSTRE_LOV, lovea, lovea_size,
 565                        XATTR_CREATE);
 566         if (rc < 0) {
 567                 rc = -errno;
 568                 CT_ERROR(rc, "cannot set lov EA on '%s'", dst);
 569         }
 570
 571         return rc;
 572 }
 573
 574 static int ct_copy_data(struct hsm_copyaction_private *hcp, const char *src,
 575                         const char *dst, int src_fd, int dst_fd,
 576                         const struct hsm_action_item *hai, long hal_flags)
 577 {
 578         struct hsm_extent        he;
 579         __u64                    offset = hai->hai_extent.offset;
 580         struct stat              src_st;
 581         struct stat              dst_st;
 582         char                    *buf = NULL;
 583         __u64                    write_total = 0;
 584         __u64                    length = hai->hai_extent.length;
 585         time_t                   last_report_time;
 586         int                      rc = 0;
 587         double                   start_ct_now = ct_now();
 588         /* Bandwidth Control */
 589         time_t                  start_time;
 590         time_t                  now;
 591         time_t                  last_bw_print;
 592
 593         if (fstat(src_fd, &src_st) < 0) {
 594                 rc = -errno;
 595                 CT_ERROR(rc, "cannot stat '%s'", src);
 596                 return rc;
 597         }
 598
 599         if (!S_ISREG(src_st.st_mode)) {
 600                 rc = -EINVAL;
 601                 CT_ERROR(rc, "'%s' is not a regular file", src);
 602                 return rc;
 603         }
 604
 605         if (hai->hai_extent.offset > (__u64)src_st.st_size) {
 606                 rc = -EINVAL;
 607                 CT_ERROR(rc, "Trying to start reading past end (%ju > "
 608                          "%jd) of '%s' source file",
 609                          (uintmax_t)hai->hai_extent.offset,
 610                          (intmax_t)src_st.st_size, src);
 611                 return rc;
 612         }
 613
 614         if (fstat(dst_fd, &dst_st) < 0) {
 615                 rc = -errno;
 616                 CT_ERROR(rc, "cannot stat '%s'", dst);
 617                 return rc;
 618         }
 619
 620         if (!S_ISREG(dst_st.st_mode)) {
 621                 rc = -EINVAL;
 622                 CT_ERROR(rc, "'%s' is not a regular file", dst);
 623                 return rc;
 624         }
 625
 626         /* Don't read beyond a given extent */
 627         if (length > src_st.st_size - hai->hai_extent.offset)
 628                 length = src_st.st_size - hai->hai_extent.offset;
 629
 630         start_time = last_bw_print = last_report_time = time(NULL);
 631
 632         he.offset = offset;
 633         he.length = 0;
 634         rc = llapi_hsm_action_progress(hcp, &he, length, 0);
 635         if (rc < 0) {
 636                 /* Action has been canceled or something wrong
 637                  * is happening. Stop copying data. */
 638                 CT_ERROR(rc, "progress ioctl for copy '%s'->'%s' failed",
 639                          src, dst);
 640                 goto out;
 641         }
 642
 643         errno = 0;
 644
 645         buf = malloc(opt.o_chunk_size);
 646         if (buf == NULL) {
 647                 rc = -ENOMEM;
 648                 goto out;
 649         }
 650
 651         CT_TRACE("start copy of %ju bytes from '%s' to '%s'",
 652                  (uintmax_t)length, src, dst);
 653
 654         while (write_total < length) {
 655                 ssize_t rsize;
 656                 ssize_t wsize;
 657                 int     chunk = (length - write_total > opt.o_chunk_size) ?
 658                                  opt.o_chunk_size : length - write_total;
 659
 660                 rsize = pread(src_fd, buf, chunk, offset);
 661                 if (rsize == 0)
 662                         /* EOF */
 663                         break;
 664
 665                 if (rsize < 0) {
 666                         rc = -errno;
 667                         CT_ERROR(rc, "cannot read from '%s'", src);
 668                         break;
 669                 }
 670
 671                 wsize = pwrite(dst_fd, buf, rsize, offset);
 672                 if (wsize < 0) {
 673                         rc = -errno;
 674                         CT_ERROR(rc, "cannot write to '%s'", dst);
 675                         break;
 676                 }
 677
 678                 write_total += wsize;
 679                 offset += wsize;
 680
 681                 now = time(NULL);
 682                 /* sleep if needed, to honor bandwidth limits */
 683                 if (opt.o_bandwidth != 0) {
 684                         unsigned long long write_theory;
 685
 686                         write_theory = (now - start_time) * opt.o_bandwidth;
 687
 688                         if (write_theory < write_total) {
 689                                 unsigned long long      excess;
 690                                 struct timespec         delay;
 691
 692                                 excess = write_total - write_theory;
 693
 694                                 delay.tv_sec = excess / opt.o_bandwidth;
 695                                 delay.tv_nsec = (excess % opt.o_bandwidth) *
 696                                         NSEC_PER_SEC / opt.o_bandwidth;
 697
 698                                 if (now >= last_bw_print + opt.o_report_int) {
 699                                         CT_TRACE("bandwith control: %lluB/s "
 700                                                  "excess=%llu sleep for "
 701                                                  "%lld.%09lds",
 702                                                  opt.o_bandwidth, excess,
 703                                                  (long long)delay.tv_sec,
 704                                                  delay.tv_nsec);
 705                                         last_bw_print = now;
 706                                 }
 707
 708                                 do {
 709                                         rc = nanosleep(&delay, &delay);
 710                                 } while (rc < 0 && errno == EINTR);
 711                                 if (rc < 0) {
 712                                         CT_ERROR(errno, "delay for bandwidth "
 713                                                  "control failed to sleep: "
 714                                                  "residual=%lld.%09lds",
 715                                                  (long long)delay.tv_sec,
 716                                                  delay.tv_nsec);
 717                                         rc = 0;
 718                                 }
 719                         }
 720                 }
 721
 722                 now = time(NULL);
 723                 if (now >= last_report_time + opt.o_report_int) {
 724                         last_report_time = now;
 725                         CT_TRACE("%%%ju ", (uintmax_t)(100 * write_total / length));
 726                         /* only give the length of the write since the last
 727                          * progress report */
 728                         he.length = offset - he.offset;
 729                         rc = llapi_hsm_action_progress(hcp, &he, length, 0);
 730                         if (rc < 0) {
 731                                 /* Action has been canceled or something wrong
 732                                  * is happening. Stop copying data. */
 733                                 CT_ERROR(rc, "progress ioctl for copy"
 734                                          " '%s'->'%s' failed", src, dst);
 735                                 goto out;
 736                         }
 737                         he.offset = offset;
 738                 }
 739                 rc = 0;
 740         }
 741
 742 out:
 743         /*
 744          * truncate restored file
 745          * size is taken from the archive this is done to support
 746          * restore after a force release which leaves the file with the
 747          * wrong size (can big bigger than the new size)
 748          */
 749         if ((hai->hai_action == HSMA_RESTORE) &&
 750             (src_st.st_size < dst_st.st_size)) {
 751                 /*
 752                  * make sure the file is on disk before reporting success.
 753                  */
 754                 rc = ftruncate(dst_fd, src_st.st_size);
 755                 if (rc < 0) {
 756                         rc = -errno;
 757                         CT_ERROR(rc, "cannot truncate '%s' to size %jd",
 758                                  dst, (intmax_t)src_st.st_size);
 759                         err_major++;
 760                 }
 761         }
 762
 763         if (buf != NULL)
 764                 free(buf);
 765
 766         CT_TRACE("copied %ju bytes in %f seconds",
 767                  (uintmax_t)length, ct_now() - start_ct_now);
 768
 769         return rc;
 770 }
 771
 772 /* Copy file attributes from file src to file dest */
 773 static int ct_copy_attr(const char *src, const char *dst, int src_fd,
 774                         int dst_fd)
 775 {
 776         struct stat     st;
 777         struct timeval  times[2];
 778         int             rc;
 779
 780         if (fstat(src_fd, &st) < 0) {
 781                 rc = -errno;
 782                 CT_ERROR(rc, "cannot stat '%s'", src);
 783                 return rc;
 784         }
 785
 786         times[0].tv_sec = st.st_atime;
 787         times[0].tv_usec = 0;
 788         times[1].tv_sec = st.st_mtime;
 789         times[1].tv_usec = 0;
 790         if (fchmod(dst_fd, st.st_mode) < 0 ||
 791             fchown(dst_fd, st.st_uid, st.st_gid) < 0 ||
 792             futimes(dst_fd, times) < 0) {
 793                 rc = -errno;
 794                 CT_ERROR(rc, "cannot set attributes of '%s'", src);
 795                 return rc;
 796         }
 797
 798         return 0;
 799 }
 800
 801 static int ct_copy_xattr(const char *src, const char *dst, int src_fd,
 802                          int dst_fd, bool is_restore)
 803 {
 804         char     list[XATTR_LIST_MAX];
 805         char     value[XATTR_SIZE_MAX];
 806         char    *name;
 807         ssize_t  list_len;
 808         int      rc;
 809
 810         list_len = flistxattr(src_fd, list, sizeof(list));
 811         if (list_len < 0)
 812                 return -errno;
 813
 814         name = list;
 815         while (name < list + list_len) {
 816                 rc = fgetxattr(src_fd, name, value, sizeof(value));
 817                 if (rc < 0)
 818                         return -errno;
 819
 820                 /* when we restore, we do not restore lustre xattr */
 821                 if (!is_restore ||
 822                     (strncmp(XATTR_TRUSTED_PREFIX, name,
 823                              sizeof(XATTR_TRUSTED_PREFIX) - 1) != 0)) {
 824                         rc = fsetxattr(dst_fd, name, value, rc, 0);
 825                         CT_TRACE("fsetxattr of '%s' on '%s' rc=%d (%s)",
 826                                  name, dst, rc, strerror(errno));
 827                         /* lustre.* attrs aren't supported on other FS's */
 828                         if (rc < 0 && errno != EOPNOTSUPP) {
 829                                 rc = -errno;
 830                                 CT_ERROR(rc, "cannot set extended attribute"
 831                                          " '%s' of '%s'",
 832                                          name, dst);
 833                                 return rc;
 834                         }
 835                 }
 836                 name += strlen(name) + 1;
 837         }
 838
 839         return 0;
 840 }
 841
 842 static int ct_path_lustre(char *buf, int sz, const char *mnt,
 843                           const struct lu_fid *fid)
 844 {
 845         return snprintf(buf, sz, "%s/%s/fid/"DFID_NOBRACE, mnt,
 846                         dot_lustre_name, PFID(fid));
 847 }
 848
 849 static int ct_path_archive(char *buf, int sz, const char *archive_dir,
 850                            const struct lu_fid *fid)
 851 {
 852         return snprintf(buf, sz, "%s/%04x/%04x/%04x/%04x/%04x/%04x/"
 853                         DFID_NOBRACE, archive_dir,
 854                         (fid)->f_oid       & 0xFFFF,
 855                         (fid)->f_oid >> 16 & 0xFFFF,
 856                         (unsigned int)((fid)->f_seq       & 0xFFFF),
 857                         (unsigned int)((fid)->f_seq >> 16 & 0xFFFF),
 858                         (unsigned int)((fid)->f_seq >> 32 & 0xFFFF),
 859                         (unsigned int)((fid)->f_seq >> 48 & 0xFFFF),
 860                         PFID(fid));
 861 }
 862
 863 static bool ct_is_retryable(int err)
 864 {
 865         return err == -ETIMEDOUT;
 866 }
 867
 868 static int ct_begin_restore(struct hsm_copyaction_private **phcp,
 869                             const struct hsm_action_item *hai,
 870                             int mdt_index, int open_flags)
 871 {
 872         char     src[PATH_MAX];
 873         int      rc;
 874
 875         rc = llapi_hsm_action_begin(phcp, ctdata, hai, mdt_index, open_flags,
 876                                     false);
 877         if (rc < 0) {
 878                 ct_path_lustre(src, sizeof(src), opt.o_mnt, &hai->hai_fid);
 879                 CT_ERROR(rc, "llapi_hsm_action_begin() on '%s' failed", src);
 880         }
 881
 882         return rc;
 883 }
 884
 885 static int ct_begin(struct hsm_copyaction_private **phcp,
 886                     const struct hsm_action_item *hai)
 887 {
 888         /* Restore takes specific parameters. Call the same function w/ default
 889          * values for all other operations. */
 890         return ct_begin_restore(phcp, hai, -1, 0);
 891 }
 892
 893 static int ct_fini(struct hsm_copyaction_private **phcp,
 894                    const struct hsm_action_item *hai, int hp_flags, int ct_rc)
 895 {
 896         struct hsm_copyaction_private   *hcp;
 897         char                             lstr[PATH_MAX];
 898         int                              rc;
 899
 900         CT_TRACE("Action completed, notifying coordinator "
 901                  "cookie=%#jx, FID="DFID", hp_flags=%d err=%d",
 902                  (uintmax_t)hai->hai_cookie, PFID(&hai->hai_fid),
 903                  hp_flags, -ct_rc);
 904
 905         ct_path_lustre(lstr, sizeof(lstr), opt.o_mnt, &hai->hai_fid);
 906
 907         if (phcp == NULL || *phcp == NULL) {
 908                 rc = llapi_hsm_action_begin(&hcp, ctdata, hai, -1, 0, true);
 909                 if (rc < 0) {
 910                         CT_ERROR(rc, "llapi_hsm_action_begin() on '%s' failed",
 911                                  lstr);
 912                         return rc;
 913                 }
 914                 phcp = &hcp;
 915         }
 916
 917         rc = llapi_hsm_action_end(phcp, &hai->hai_extent, hp_flags, abs(ct_rc));
 918         if (rc == -ECANCELED)
 919                 CT_ERROR(rc, "completed action on '%s' has been canceled: "
 920                          "cookie=%#jx, FID="DFID, lstr,
 921                          (uintmax_t)hai->hai_cookie, PFID(&hai->hai_fid));
 922         else if (rc < 0)
 923                 CT_ERROR(rc, "llapi_hsm_action_end() on '%s' failed", lstr);
 924         else
 925                 CT_TRACE("llapi_hsm_action_end() on '%s' ok (rc=%d)",
 926                          lstr, rc);
 927
 928         return rc;
 929 }
 930
 931 static int ct_archive(const struct hsm_action_item *hai, const long hal_flags)
 932 {
 933         struct hsm_copyaction_private   *hcp = NULL;
 934         char                             src[PATH_MAX];
 935         char                             dst[PATH_MAX + 4] = "";
 936         char                             root[PATH_MAX] = "";
 937         int                              rc;
 938         int                              rcf = 0;
 939         bool                             rename_needed = false;
 940         int                              hp_flags = 0;
 941         int                              open_flags;
 942         int                              src_fd = -1;
 943         int                              dst_fd = -1;
 944
 945         rc = ct_begin(&hcp, hai);
 946         if (rc < 0)
 947                 goto fini_major;
 948
 949         /* we fill archive so:
 950          * source = data FID
 951          * destination = lustre FID
 952          */
 953         ct_path_lustre(src, sizeof(src), opt.o_mnt, &hai->hai_dfid);
 954         ct_path_archive(root, sizeof(root), opt.o_hsm_root, &hai->hai_fid);
 955         if (hai->hai_extent.length == -1) {
 956                 /* whole file, write it to tmp location and atomically
 957                  * replace old archived file */
 958                 snprintf(dst, sizeof(dst), "%s_tmp", root);
 959                 /* we cannot rely on the same test because ct_copy_data()
 960                  * updates hai_extent.length */
 961                 rename_needed = true;
 962         } else {
 963                 snprintf(dst, sizeof(dst), "%s", root);
 964         }
 965
 966         CT_TRACE("archiving '%s' to '%s'", src, dst);
 967
 968         if (opt.o_dry_run) {
 969                 rc = 0;
 970                 goto fini_major;
 971         }
 972
 973         rc = ct_mkdir_p(dst);
 974         if (rc < 0) {
 975                 CT_ERROR(rc, "mkdir_p '%s' failed", dst);
 976                 goto fini_major;
 977         }
 978
 979         src_fd = llapi_hsm_action_get_fd(hcp);
 980         if (src_fd < 0) {
 981                 rc = src_fd;
 982                 CT_ERROR(rc, "cannot open '%s' for read", src);
 983                 goto fini_major;
 984         }
 985
 986         open_flags = O_WRONLY | O_NOFOLLOW;
 987         /* If extent is specified, don't truncate an old archived copy */
 988         open_flags |= ((hai->hai_extent.length == -1) ? O_TRUNC : 0) | O_CREAT;
 989
 990         dst_fd = open(dst, open_flags, FILE_PERM);
 991         if (dst_fd == -1) {
 992                 rc = -errno;
 993                 CT_ERROR(rc, "cannot open '%s' for write", dst);
 994                 goto fini_major;
 995         }
 996
 997         /* saving stripe is not critical */
 998         rc = ct_save_stripe(src_fd, src, dst);
 999         if (rc < 0)
1000                 CT_ERROR(rc, "cannot save file striping info of '%s' in '%s'",
1001                          src, dst);
1002
1003         rc = ct_copy_data(hcp, src, dst, src_fd, dst_fd, hai, hal_flags);
1004         if (rc < 0) {
1005                 CT_ERROR(rc, "data copy failed from '%s' to '%s'", src, dst);
1006                 goto fini_major;
1007         }
1008
1009         rc = fsync(dst_fd);
1010         if (rc < 0) {
1011                 rc = -errno;
1012                 CT_ERROR(rc, "cannot flush '%s' archive file '%s'", src, dst);
1013                 goto fini_major;
1014         }
1015
1016         CT_TRACE("data archiving for '%s' to '%s' done", src, dst);
1017
1018         /* attrs will remain on the MDS; no need to copy them, except possibly
1019           for disaster recovery */
1020         if (opt.o_copy_attrs) {
1021                 rc = ct_copy_attr(src, dst, src_fd, dst_fd);
1022                 if (rc < 0) {
1023                         CT_ERROR(rc, "cannot copy attr of '%s' to '%s'",
1024                                  src, dst);
1025                         rcf = rc;
1026                 }
1027                 CT_TRACE("attr file for '%s' saved to archive '%s'",
1028                          src, dst);
1029         }
1030
1031         /* xattrs will remain on the MDS; no need to copy them, except possibly
1032          for disaster recovery */
1033         if (opt.o_copy_xattrs) {
1034                 rc = ct_copy_xattr(src, dst, src_fd, dst_fd, false);
1035                 if (rc < 0) {
1036                         CT_ERROR(rc, "cannot copy xattr of '%s' to '%s'",
1037                                  src, dst);
1038                         rcf = rcf ? rcf : rc;
1039                 }
1040                 CT_TRACE("xattr file for '%s' saved to archive '%s'",
1041                          src, dst);
1042         }
1043
1044         if (rename_needed == true) {
1045                 char     tmp_src[PATH_MAX + 8];
1046                 char     tmp_dst[PATH_MAX + 8];
1047
1048                 /* atomically replace old archived file */
1049                 ct_path_archive(src, sizeof(src), opt.o_hsm_root,
1050                                 &hai->hai_fid);
1051                 rc = rename(dst, src);
1052                 if (rc < 0) {
1053                         rc = -errno;
1054                         CT_ERROR(rc, "cannot rename '%s' to '%s'", dst, src);
1055                         goto fini_major;
1056                 }
1057                 /* rename lov file */
1058                 snprintf(tmp_src, sizeof(tmp_src), "%s.lov", src);
1059                 snprintf(tmp_dst, sizeof(tmp_dst), "%s.lov", dst);
1060                 rc = rename(tmp_dst, tmp_src);
1061                 if (rc < 0)
1062                         CT_ERROR(errno, "cannot rename '%s' to '%s'",
1063                                  tmp_dst, tmp_src);
1064         }
1065
1066         if (opt.o_shadow_tree) {
1067                 /* Create a namespace of softlinks that shadows the original
1068                  * Lustre namespace.  This will only be current at
1069                  * time-of-archive (won't follow renames).
1070                  * WARNING: release won't kill these links; a manual
1071                  * cleanup of dead links would be required.
1072                  */
1073                 char             buf[PATH_MAX];
1074                 long long        recno = -1;
1075                 int              linkno = 0;
1076                 char            *ptr;
1077                 int              depth = 0;
1078                 ssize_t          sz;
1079
1080                 sprintf(buf, DFID, PFID(&hai->hai_fid));
1081                 sprintf(src, "%s/shadow/", opt.o_hsm_root);
1082
1083                 ptr = opt.o_hsm_root;
1084                 while (*ptr)
1085                         (*ptr++ == '/') ? depth-- : 0;
1086
1087                 rc = llapi_fid2path(opt.o_mnt, buf, src + strlen(src),
1088                                     sizeof(src) - strlen(src), &recno, &linkno);
1089                 if (rc < 0) {
1090                         CT_ERROR(rc, "cannot get FID of '%s'", buf);
1091                         rcf = rcf ? rcf : rc;
1092                         goto fini_minor;
1093                 }
1094
1095                 /* Figure out how many parent dirs to symlink back */
1096                 ptr = src;
1097                 while (*ptr)
1098                         (*ptr++ == '/') ? depth++ : 0;
1099                 sprintf(buf, "..");
1100                 while (--depth > 1)
1101                         strcat(buf, "/..");
1102
1103                 ct_path_archive(dst, sizeof(dst), buf, &hai->hai_fid);
1104
1105                 if (ct_mkdir_p(src)) {
1106                         CT_ERROR(errno, "mkdir_p '%s' failed", src);
1107                         rcf = rcf ? rcf : -errno;
1108                         goto fini_minor;
1109                 }
1110                 /* symlink already exists ? */
1111                 sz = readlink(src, buf, sizeof(buf));
1112                 /* detect truncation */
1113                 if (sz == sizeof(buf)) {
1114                         rcf = rcf ? rcf : -E2BIG;
1115                         CT_ERROR(rcf, "readlink '%s' truncated", src);
1116                         goto fini_minor;
1117                 }
1118                 if (sz >= 0) {
1119                         buf[sz] = '\0';
1120                         if (sz == 0 || strncmp(buf, dst, sz) != 0) {
1121                                 if (unlink(src) && errno != ENOENT) {
1122                                         CT_ERROR(errno,
1123                                                  "cannot unlink symlink '%s'",
1124                                                  src);
1125                                         rcf = rcf ? rcf : -errno;
1126                                         goto fini_minor;
1127                                 }
1128                                 /* unlink old symlink done */
1129                                 CT_TRACE("remove old symlink '%s' pointing"
1130                                          " to '%s'", src, buf);
1131                         } else {
1132                                 /* symlink already ok */
1133                                 CT_TRACE("symlink '%s' already pointing"
1134                                          " to '%s'", src, dst);
1135                                 rcf = 0;
1136                                 goto fini_minor;
1137                         }
1138                 }
1139                 if (symlink(dst, src)) {
1140                         CT_ERROR(errno, "cannot symlink '%s' to '%s'",
1141                                  src, dst);
1142                         rcf = rcf ? rcf : -errno;
1143                         goto fini_minor;
1144                 }
1145                 CT_TRACE("symlink '%s' to '%s' done", src, dst);
1146         }
1147 fini_minor:
1148         if (rcf)
1149                 err_minor++;
1150         goto out;
1151
1152
1153 fini_major:
1154         err_major++;
1155
1156         unlink(dst);
1157         if (ct_is_retryable(rc))
1158                 hp_flags |= HP_FLAG_RETRY;
1159
1160         rcf = rc;
1161
1162 out:
1163         if (!(src_fd < 0))
1164                 close(src_fd);
1165
1166         if (!(dst_fd < 0))
1167                 close(dst_fd);
1168
1169         rc = ct_fini(&hcp, hai, hp_flags, rcf);
1170
1171         return rc;
1172 }
1173
1174 static int ct_restore(const struct hsm_action_item *hai, const long hal_flags)
1175 {
1176         struct hsm_copyaction_private   *hcp = NULL;
1177         char                             src[PATH_MAX];
1178         char                             dst[PATH_MAX];
1179         char                             lov_buf[XATTR_SIZE_MAX];
1180         size_t                           lov_size = sizeof(lov_buf);
1181         int                              rc;
1182         int                              hp_flags = 0;
1183         int                              src_fd = -1;
1184         int                              dst_fd = -1;
1185         int                              mdt_index = -1;
1186         int                              open_flags = 0;
1187         bool                             set_lovea;
1188         struct lu_fid                    dfid;
1189         /* we fill lustre so:
1190          * source = lustre FID in the backend
1191          * destination = data FID = volatile file
1192          */
1193
1194         /* build backend file name from released file FID */
1195         ct_path_archive(src, sizeof(src), opt.o_hsm_root, &hai->hai_fid);
1196
1197         rc = llapi_get_mdt_index_by_fid(opt.o_mnt_fd, &hai->hai_fid,
1198                                         &mdt_index);
1199         if (rc < 0) {
1200                 CT_ERROR(rc, "cannot get mdt index "DFID"",
1201                          PFID(&hai->hai_fid));
1202                 return rc;
1203         }
1204         /* restore loads and sets the LOVEA w/o interpreting it to avoid
1205          * dependency on the structure format. */
1206         rc = ct_load_stripe(src, lov_buf, &lov_size);
1207         if (rc < 0) {
1208                 CT_WARN("cannot get stripe rules for '%s' (%s), use default",
1209                         src, strerror(-rc));
1210                 set_lovea = false;
1211         } else {
1212                 open_flags |= O_LOV_DELAY_CREATE;
1213                 set_lovea = true;
1214         }
1215
1216         rc = ct_begin_restore(&hcp, hai, mdt_index, open_flags);
1217         if (rc < 0)
1218                 goto fini;
1219
1220         /* get the FID of the volatile file */
1221         rc = llapi_hsm_action_get_dfid(hcp, &dfid);
1222         if (rc < 0) {
1223                 CT_ERROR(rc, "restoring "DFID
1224                          ", cannot get FID of created volatile file",
1225                          PFID(&hai->hai_fid));
1226                 goto fini;
1227         }
1228
1229         /* build volatile "file name", for messages */
1230         snprintf(dst, sizeof(dst), "{VOLATILE}="DFID, PFID(&dfid));
1231
1232         CT_TRACE("restoring data from '%s' to '%s'", src, dst);
1233
1234         if (opt.o_dry_run) {
1235                 rc = 0;
1236                 goto fini;
1237         }
1238
1239         src_fd = open(src, O_RDONLY | O_NOATIME | O_NOFOLLOW);
1240         if (src_fd < 0) {
1241                 rc = -errno;
1242                 CT_ERROR(rc, "cannot open '%s' for read", src);
1243                 goto fini;
1244         }
1245
1246         dst_fd = llapi_hsm_action_get_fd(hcp);
1247         if (dst_fd < 0) {
1248                 rc = dst_fd;
1249                 CT_ERROR(rc, "cannot open '%s' for write", dst);
1250                 goto fini;
1251         }
1252
1253         if (set_lovea) {
1254                 /* the layout cannot be allocated through .fid so we have to
1255                  * restore a layout */
1256                 rc = ct_restore_stripe(src, dst, dst_fd, lov_buf, lov_size);
1257                 if (rc < 0) {
1258                         CT_ERROR(rc, "cannot restore file striping info"
1259                                  " for '%s' from '%s'", dst, src);
1260                         err_major++;
1261                         goto fini;
1262                 }
1263         }
1264
1265         rc = ct_copy_data(hcp, src, dst, src_fd, dst_fd, hai, hal_flags);
1266         if (rc < 0) {
1267                 CT_ERROR(rc, "cannot copy data from '%s' to '%s'",
1268                          src, dst);
1269                 err_major++;
1270                 if (ct_is_retryable(rc))
1271                         hp_flags |= HP_FLAG_RETRY;
1272                 goto fini;
1273         }
1274
1275         CT_TRACE("data restore from '%s' to '%s' done", src, dst);
1276
1277 fini:
1278         rc = ct_fini(&hcp, hai, hp_flags, rc);
1279
1280         /* object swaping is done by cdt at copy end, so close of volatile file
1281          * cannot be done before */
1282         if (!(src_fd < 0))
1283                 close(src_fd);
1284
1285         if (!(dst_fd < 0))
1286                 close(dst_fd);
1287
1288         return rc;
1289 }
1290
1291 static int ct_remove(const struct hsm_action_item *hai, const long hal_flags)
1292 {
1293         struct hsm_copyaction_private   *hcp = NULL;
1294         char                             dst[PATH_MAX], attr[PATH_MAX + 4];
1295         int                              rc;
1296
1297         rc = ct_begin(&hcp, hai);
1298         if (rc < 0)
1299                 goto fini;
1300
1301         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, &hai->hai_fid);
1302
1303         CT_TRACE("removing file '%s'", dst);
1304
1305         if (opt.o_dry_run) {
1306                 rc = 0;
1307                 goto fini;
1308         }
1309
1310         rc = unlink(dst);
1311         if (rc < 0) {
1312                 rc = -errno;
1313                 CT_ERROR(rc, "cannot unlink '%s'", dst);
1314                 err_minor++;
1315                 goto fini;
1316         }
1317
1318         snprintf(attr, sizeof(attr), "%s.lov", dst);
1319         rc = unlink(attr);
1320         if (rc < 0) {
1321                 rc = -errno;
1322                 CT_ERROR(rc, "cannot unlink '%s'", attr);
1323                 err_minor++;
1324
1325                 /* ignore the error when lov file does not exist. */
1326                 if (rc == -ENOENT)
1327                         rc = 0;
1328                 else
1329                         goto fini;
1330         }
1331
1332 fini:
1333         rc = ct_fini(&hcp, hai, 0, rc);
1334
1335         return rc;
1336 }
1337
1338 static int ct_process_item(struct hsm_action_item *hai, const long hal_flags)
1339 {
1340         int     rc = 0;
1341
1342         if (opt.o_verbose >= LLAPI_MSG_INFO || opt.o_dry_run) {
1343                 /* Print the original path */
1344                 char            fid[128];
1345                 char            path[PATH_MAX];
1346                 long long       recno = -1;
1347                 int             linkno = 0;
1348
1349                 sprintf(fid, DFID, PFID(&hai->hai_fid));
1350                 CT_TRACE("'%s' action %s reclen %d, cookie=%#jx",
1351                          fid, hsm_copytool_action2name(hai->hai_action),
1352                          hai->hai_len, (uintmax_t)hai->hai_cookie);
1353                 rc = llapi_fid2path(opt.o_mnt, fid, path,
1354                                     sizeof(path), &recno, &linkno);
1355                 if (rc < 0)
1356                         CT_ERROR(rc, "cannot get path of FID %s", fid);
1357                 else
1358                         CT_TRACE("processing file '%s'", path);
1359         }
1360
1361         switch (hai->hai_action) {
1362         /* set err_major, minor inside these functions */
1363         case HSMA_ARCHIVE:
1364                 rc = ct_archive(hai, hal_flags);
1365                 break;
1366         case HSMA_RESTORE:
1367                 rc = ct_restore(hai, hal_flags);
1368                 break;
1369         case HSMA_REMOVE:
1370                 rc = ct_remove(hai, hal_flags);
1371                 break;
1372         case HSMA_CANCEL:
1373                 CT_TRACE("cancel not implemented for file system '%s'",
1374                          opt.o_mnt);
1375                 /* Don't report progress to coordinator for this cookie:
1376                  * the copy function will get ECANCELED when reporting
1377                  * progress. */
1378                 err_minor++;
1379                 return 0;
1380                 break;
1381         default:
1382                 rc = -EINVAL;
1383                 CT_ERROR(rc, "unknown action %d, on '%s'", hai->hai_action,
1384                          opt.o_mnt);
1385                 err_minor++;
1386                 ct_fini(NULL, hai, 0, rc);
1387         }
1388
1389         return 0;
1390 }
1391
1392 struct ct_th_data {
1393         long                     hal_flags;
1394         struct hsm_action_item  *hai;
1395 };
1396
1397 static void *ct_thread(void *data)
1398 {
1399         struct ct_th_data *cttd = data;
1400         int rc;
1401
1402         rc = ct_process_item(cttd->hai, cttd->hal_flags);
1403
1404         free(cttd->hai);
1405         free(cttd);
1406         pthread_exit((void *)(intptr_t)rc);
1407 }
1408
1409 static int ct_process_item_async(const struct hsm_action_item *hai,
1410                                  long hal_flags)
1411 {
1412         pthread_attr_t           attr;
1413         pthread_t                thread;
1414         struct ct_th_data       *data;
1415         int                      rc;
1416
1417         data = malloc(sizeof(*data));
1418         if (data == NULL)
1419                 return -ENOMEM;
1420
1421         data->hai = malloc(hai->hai_len);
1422         if (data->hai == NULL) {
1423                 free(data);
1424                 return -ENOMEM;
1425         }
1426
1427         memcpy(data->hai, hai, hai->hai_len);
1428         data->hal_flags = hal_flags;
1429
1430         rc = pthread_attr_init(&attr);
1431         if (rc != 0) {
1432                 CT_ERROR(rc, "pthread_attr_init failed for '%s' service",
1433                          opt.o_mnt);
1434                 free(data->hai);
1435                 free(data);
1436                 return -rc;
1437         }
1438
1439         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
1440
1441         rc = pthread_create(&thread, &attr, ct_thread, data);
1442         if (rc != 0)
1443                 CT_ERROR(rc, "cannot create thread for '%s' service",
1444                          opt.o_mnt);
1445
1446         pthread_attr_destroy(&attr);
1447         return 0;
1448 }
1449
1450 static int ct_import_one(const char *src, const char *dst)
1451 {
1452         char            newarc[PATH_MAX];
1453         struct lu_fid   fid;
1454         struct stat     st;
1455         int             rc;
1456
1457         CT_TRACE("importing '%s' from '%s'", dst, src);
1458
1459         if (stat(src, &st) < 0) {
1460                 rc = -errno;
1461                 CT_ERROR(rc, "cannot stat '%s'", src);
1462                 return rc;
1463         }
1464
1465         if (opt.o_dry_run)
1466                 return 0;
1467
1468         rc = llapi_hsm_import(dst,
1469                               opt.o_archive_id_used ?
1470                               opt.o_archive_id[0] : 0,
1471                               &st, 0, 0, 0, 0, NULL, &fid);
1472         if (rc < 0) {
1473                 CT_ERROR(rc, "cannot import '%s' from '%s'", dst, src);
1474                 return rc;
1475         }
1476
1477         ct_path_archive(newarc, sizeof(newarc), opt.o_hsm_root, &fid);
1478
1479         rc = ct_mkdir_p(newarc);
1480         if (rc < 0) {
1481                 CT_ERROR(rc, "mkdir_p '%s' failed", newarc);
1482                 err_major++;
1483                 return rc;
1484
1485         }
1486
1487         /* Lots of choices now: mv, ln, ln -s ? */
1488         rc = link(src, newarc); /* hardlink */
1489         if (rc < 0) {
1490                 rc = -errno;
1491                 CT_ERROR(rc, "cannot link '%s' to '%s'", newarc, src);
1492                 err_major++;
1493                 return rc;
1494         }
1495         CT_TRACE("imported '%s' from '%s'=='%s'", dst, newarc, src);
1496
1497         return 0;
1498 }
1499
1500 static char *path_concat(const char *dirname, const char *basename)
1501 {
1502         char    *result;
1503         int      rc;
1504
1505         rc = asprintf(&result, "%s/%s", dirname, basename);
1506         if (rc < 0)
1507                 return NULL;
1508
1509         return result;
1510 }
1511
1512 static int ct_import_fid(const struct lu_fid *import_fid)
1513 {
1514         char    fid_path[PATH_MAX];
1515         int     rc;
1516
1517         ct_path_lustre(fid_path, sizeof(fid_path), opt.o_mnt, import_fid);
1518         rc = access(fid_path, F_OK);
1519         if (rc == 0 || errno != ENOENT) {
1520                 rc = (errno == 0) ? -EEXIST : -errno;
1521                 CT_ERROR(rc, "cannot import '"DFID"'", PFID(import_fid));
1522                 return rc;
1523         }
1524
1525         ct_path_archive(fid_path, sizeof(fid_path), opt.o_hsm_root,
1526                         import_fid);
1527
1528         CT_TRACE("Resolving "DFID" to %s", PFID(import_fid), fid_path);
1529
1530         return ct_import_one(fid_path, opt.o_dst);
1531 }
1532
1533 static int ct_import_recurse(const char *relpath)
1534 {
1535         DIR             *dir;
1536         struct dirent   *ent;
1537         char            *srcpath, *newpath;
1538         struct lu_fid    import_fid;
1539         int              rc;
1540
1541         if (relpath == NULL)
1542                 return -EINVAL;
1543
1544         /* Is relpath a FID? In which case SFID should expand to three
1545          * elements. */
1546         rc = sscanf(relpath, SFID, RFID(&import_fid));
1547         if (rc == 3)
1548                 return ct_import_fid(&import_fid);
1549
1550         srcpath = path_concat(opt.o_hsm_root, relpath);
1551         if (srcpath == NULL) {
1552                 err_major++;
1553                 return -ENOMEM;
1554         }
1555
1556         dir = opendir(srcpath);
1557         if (dir == NULL) {
1558                 /* Not a dir, or error */
1559                 if (errno == ENOTDIR) {
1560                         /* Single regular file case, treat o_dst as absolute
1561                            final location. */
1562                         rc = ct_import_one(srcpath, opt.o_dst);
1563                 } else {
1564                         rc = -errno;
1565                         CT_ERROR(rc, "cannot opendir '%s'", srcpath);
1566                         err_major++;
1567                 }
1568                 free(srcpath);
1569                 return rc;
1570         }
1571         free(srcpath);
1572
1573         while ((ent = readdir(dir)) != NULL) {
1574                 if (!strcmp(ent->d_name, ".") ||
1575                     !strcmp(ent->d_name, ".."))
1576                         continue;
1577
1578                 /* New relative path */
1579                 newpath = path_concat(relpath, ent->d_name);
1580                 if (newpath == NULL) {
1581                         err_major++;
1582                         rc = -ENOMEM;
1583                         goto out;
1584                 }
1585
1586                 if (ent->d_type == DT_DIR) {
1587                         rc = ct_import_recurse(newpath);
1588                 } else {
1589                         char src[PATH_MAX];
1590                         char dst[PATH_MAX];
1591
1592                         sprintf(src, "%s/%s", opt.o_hsm_root, newpath);
1593                         sprintf(dst, "%s/%s", opt.o_dst, newpath);
1594                         /* Make the target dir in the Lustre fs */
1595                         rc = ct_mkdir_p(dst);
1596                         if (rc == 0) {
1597                                 /* Import the file */
1598                                 rc = ct_import_one(src, dst);
1599                         } else {
1600                                 CT_ERROR(rc, "ct_mkdir_p '%s' failed", dst);
1601                                 err_major++;
1602                         }
1603                 }
1604
1605                 if (rc != 0) {
1606                         CT_ERROR(rc, "cannot import '%s'", newpath);
1607                         if (err_major && opt.o_abort_on_error) {
1608                                 free(newpath);
1609                                 goto out;
1610                         }
1611                 }
1612                 free(newpath);
1613         }
1614
1615         rc = 0;
1616 out:
1617         closedir(dir);
1618         return rc;
1619 }
1620
1621 static int ct_rebind_one(const struct lu_fid *old_fid,
1622                          const struct lu_fid *new_fid)
1623 {
1624         char    src[PATH_MAX];
1625         char    dst[PATH_MAX];
1626         int     rc;
1627
1628         CT_TRACE("rebind "DFID" to "DFID, PFID(old_fid), PFID(new_fid));
1629
1630         ct_path_archive(src, sizeof(src), opt.o_hsm_root, old_fid);
1631         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, new_fid);
1632
1633         if (!opt.o_dry_run) {
1634                 char src_attr[PATH_MAX + 4];
1635                 char dst_attr[PATH_MAX + 4];
1636
1637                 ct_mkdir_p(dst);
1638                 if (rename(src, dst)) {
1639                         rc = -errno;
1640                         CT_ERROR(rc, "cannot rename '%s' to '%s'", src, dst);
1641                         return -errno;
1642                 }
1643                 /* rename lov file */
1644                 snprintf(src_attr, sizeof(src_attr), "%s.lov", src);
1645                 snprintf(dst_attr, sizeof(dst_attr), "%s.lov", dst);
1646                 if (rename(src_attr, dst_attr))
1647                         CT_ERROR(errno, "cannot rename '%s' to '%s'",
1648                                  src_attr, dst_attr);
1649
1650         }
1651         return 0;
1652 }
1653
1654 static bool fid_is_file(struct lu_fid *fid)
1655 {
1656         return fid_is_norm(fid) || fid_is_igif(fid);
1657 }
1658
1659 static bool should_ignore_line(const char *line)
1660 {
1661         int     i;
1662
1663         for (i = 0; line[i] != '\0'; i++) {
1664                 if (isspace(line[i]))
1665                         continue;
1666                 else if (line[i] == '#')
1667                         return true;
1668                 else
1669                         return false;
1670         }
1671
1672         return true;
1673 }
1674
1675 static int ct_rebind_list(const char *list)
1676 {
1677         int              rc;
1678         FILE            *filp;
1679         ssize_t          r;
1680         char            *line = NULL;
1681         size_t           line_size = 0;
1682         unsigned int     nl = 0;
1683         unsigned int     ok = 0;
1684
1685         filp = fopen(list, "r");
1686         if (filp == NULL) {
1687                 rc = -errno;
1688                 CT_ERROR(rc, "cannot open '%s'", list);
1689                 return rc;
1690         }
1691
1692         /* each line consists of 2 FID */
1693         while ((r = getline(&line, &line_size, filp)) != -1) {
1694                 struct lu_fid   old_fid;
1695                 struct lu_fid   new_fid;
1696
1697                 /* Ignore empty and commented out ('#...') lines. */
1698                 if (should_ignore_line(line))
1699                         continue;
1700
1701                 nl++;
1702
1703                 rc = sscanf(line, SFID" "SFID, RFID(&old_fid), RFID(&new_fid));
1704                 if (rc != 6 || !fid_is_file(&old_fid) ||
1705                     !fid_is_file(&new_fid)) {
1706                         CT_ERROR(EINVAL,
1707                                  "'%s' FID expected near '%s', line %u",
1708                                  list, line, nl);
1709                         err_major++;
1710                         continue;
1711                 }
1712
1713                 if (ct_rebind_one(&old_fid, &new_fid))
1714                         err_major++;
1715                 else
1716                         ok++;
1717         }
1718
1719         fclose(filp);
1720
1721         if (line)
1722                 free(line);
1723
1724         /* return 0 if all rebinds were successful */
1725         CT_TRACE("%u lines read from '%s', %u rebind successful", nl, list, ok);
1726
1727         return ok == nl ? 0 : -1;
1728 }
1729
1730 static int ct_rebind(void)
1731 {
1732         int     rc;
1733
1734         if (opt.o_dst) {
1735                 struct lu_fid old_fid;
1736                 struct lu_fid new_fid;
1737
1738                 if (sscanf(opt.o_src, SFID, RFID(&old_fid)) != 3 ||
1739                     !fid_is_file(&old_fid)) {
1740                         rc = -EINVAL;
1741                         CT_ERROR(rc, "'%s' invalid FID format", opt.o_src);
1742                         return rc;
1743                 }
1744
1745                 if (sscanf(opt.o_dst, SFID, RFID(&new_fid)) != 3 ||
1746                     !fid_is_file(&new_fid)) {
1747                         rc = -EINVAL;
1748                         CT_ERROR(rc, "'%s' invalid FID format", opt.o_dst);
1749                         return rc;
1750                 }
1751
1752                 rc = ct_rebind_one(&old_fid, &new_fid);
1753
1754                 return rc;
1755         }
1756
1757         /* o_src is a list file */
1758         rc = ct_rebind_list(opt.o_src);
1759
1760         return rc;
1761 }
1762
1763 static int ct_dir_level_max(const char *dirpath, __u16 *sub_seqmax)
1764 {
1765         DIR             *dir;
1766         int              rc;
1767         __u16            sub_seq;
1768         struct dirent *ent;
1769
1770         *sub_seqmax = 0;
1771
1772         dir = opendir(dirpath);
1773         if (dir == NULL) {
1774                 rc = -errno;
1775                 CT_ERROR(rc, "cannot open directory '%s'", opt.o_hsm_root);
1776                 return rc;
1777         }
1778
1779         do {
1780                 errno = 0;
1781                 ent = readdir(dir);
1782                 if (ent == NULL) {
1783                         /* end of directory.
1784                          * rc is 0 and seqmax contains the max value. */
1785                         rc = -errno;
1786                         if (rc)
1787                                 CT_ERROR(rc, "cannot readdir '%s'", dirpath);
1788                         goto out;
1789                 }
1790
1791                 if (!strcmp(ent->d_name, ".") || !strcmp(ent->d_name, ".."))
1792                         continue;
1793
1794                 if (sscanf(ent->d_name, "%hx", &sub_seq) != 1) {
1795                         CT_TRACE("'%s' has an unexpected dirname format, "
1796                                  "skip entry", ent->d_name);
1797                         continue;
1798                 }
1799                 if (sub_seq > *sub_seqmax)
1800                         *sub_seqmax = sub_seq;
1801         } while (1);
1802 out:
1803         closedir(dir);
1804         return rc;
1805 }
1806
1807 static int ct_max_sequence(void)
1808 {
1809         int     rc, i;
1810         char    path[PATH_MAX];
1811         __u64   seq = 0;
1812         __u16   subseq;
1813
1814         snprintf(path, sizeof(path), "%s", opt.o_hsm_root);
1815         /* FID sequence is stored in top-level directory names:
1816          * hsm_root/16bits (high weight)/16 bits/16 bits/16 bits (low weight).
1817          */
1818         for (i = 0; i < 4; i++) {
1819                 size_t path_len;
1820
1821                 rc = ct_dir_level_max(path, &subseq);
1822                 if (rc != 0)
1823                         return rc;
1824                 seq |= ((__u64)subseq << ((3 - i) * 16));
1825                 path_len = strlen(path);
1826                 rc = snprintf(path + path_len, sizeof(path) - path_len,
1827                               "/%04x", subseq);
1828                 if (rc >= (sizeof(path) - path_len))
1829                         return -E2BIG;
1830                 path[sizeof(path) - 1] = '\0';
1831         }
1832
1833         printf("max_sequence: %#jx\n", (uintmax_t)seq);
1834
1835         return 0;
1836 }
1837
1838 static void handler(int signal)
1839 {
1840         psignal(signal, "exiting");
1841         /* If we don't clean up upon interrupt, umount thinks there's a ref
1842          * and doesn't remove us from mtab (EINPROGRESS). The lustre client
1843          * does successfully unmount and the mount is actually gone, but the
1844          * mtab entry remains. So this just makes mtab happier. */
1845         llapi_hsm_copytool_unregister(&ctdata);
1846
1847         /* Also remove fifo upon signal as during normal/error exit */
1848         if (opt.o_event_fifo != NULL)
1849                 llapi_hsm_unregister_event_fifo(opt.o_event_fifo);
1850         _exit(1);
1851 }
1852
1853 /* Daemon waits for messages from the kernel; run it in the background. */
1854 static int ct_run(void)
1855 {
1856         struct sigaction cleanup_sigaction;
1857         int rc;
1858
1859         if (opt.o_daemonize) {
1860                 rc = daemon(1, 1);
1861                 if (rc < 0) {
1862                         rc = -errno;
1863                         CT_ERROR(rc, "cannot daemonize");
1864                         return rc;
1865                 }
1866         }
1867
1868         setbuf(stdout, NULL);
1869
1870         if (opt.o_event_fifo != NULL) {
1871                 rc = llapi_hsm_register_event_fifo(opt.o_event_fifo);
1872                 if (rc < 0) {
1873                         CT_ERROR(rc, "failed to register event fifo");
1874                         return rc;
1875                 }
1876                 llapi_error_callback_set(llapi_hsm_log_error);
1877         }
1878
1879         rc = llapi_hsm_copytool_register(&ctdata, opt.o_mnt,
1880                                          opt.o_archive_id_used,
1881                                          opt.o_archive_id, 0);
1882         if (rc < 0) {
1883                 CT_ERROR(rc, "cannot start copytool interface");
1884                 return rc;
1885         }
1886
1887         memset(&cleanup_sigaction, 0, sizeof(cleanup_sigaction));
1888         cleanup_sigaction.sa_handler = handler;
1889         sigemptyset(&cleanup_sigaction.sa_mask);
1890         sigaction(SIGINT, &cleanup_sigaction, NULL);
1891         sigaction(SIGTERM, &cleanup_sigaction, NULL);
1892
1893         while (1) {
1894                 struct hsm_action_list *hal;
1895                 struct hsm_action_item *hai;
1896                 int msgsize;
1897                 int i = 0;
1898
1899                 CT_TRACE("waiting for message from kernel");
1900
1901                 rc = llapi_hsm_copytool_recv(ctdata, &hal, &msgsize);
1902                 if (rc == -ESHUTDOWN) {
1903                         CT_TRACE("shutting down");
1904                         break;
1905                 } else if (rc < 0) {
1906                         CT_WARN("cannot receive action list: %s",
1907                                 strerror(-rc));
1908                         err_major++;
1909                         if (opt.o_abort_on_error)
1910                                 break;
1911                         else
1912                                 continue;
1913                 }
1914
1915                 CT_TRACE("copytool fs=%s archive#=%d item_count=%d",
1916                          hal->hal_fsname, hal->hal_archive_id, hal->hal_count);
1917
1918                 if (strcmp(hal->hal_fsname, fs_name) != 0) {
1919                         rc = -EINVAL;
1920                         CT_ERROR(rc, "'%s' invalid fs name, expecting: %s",
1921                                  hal->hal_fsname, fs_name);
1922                         err_major++;
1923                         if (opt.o_abort_on_error)
1924                                 break;
1925                         else
1926                                 continue;
1927                 }
1928
1929                 hai = hai_first(hal);
1930                 while (++i <= hal->hal_count) {
1931                         if ((char *)hai - (char *)hal > msgsize) {
1932                                 rc = -EPROTO;
1933                                 CT_ERROR(rc,
1934                                          "'%s' item %d past end of message!",
1935                                          opt.o_mnt, i);
1936                                 err_major++;
1937                                 break;
1938                         }
1939                         rc = ct_process_item_async(hai, hal->hal_flags);
1940                         if (rc < 0)
1941                                 CT_ERROR(rc, "'%s' item %d process",
1942                                          opt.o_mnt, i);
1943                         if (opt.o_abort_on_error && err_major)
1944                                 break;
1945                         hai = hai_next(hai);
1946                 }
1947
1948                 if (opt.o_abort_on_error && err_major)
1949                         break;
1950         }
1951
1952         llapi_hsm_copytool_unregister(&ctdata);
1953         if (opt.o_event_fifo != NULL)
1954                 llapi_hsm_unregister_event_fifo(opt.o_event_fifo);
1955
1956         return rc;
1957 }
1958
1959 static int ct_setup(void)
1960 {
1961         int     rc;
1962
1963         /* set llapi message level */
1964         llapi_msg_set_level(opt.o_verbose);
1965
1966         arc_fd = open(opt.o_hsm_root, O_RDONLY);
1967         if (arc_fd < 0) {
1968                 rc = -errno;
1969                 CT_ERROR(rc, "cannot open archive at '%s'", opt.o_hsm_root);
1970                 return rc;
1971         }
1972
1973         rc = llapi_search_fsname(opt.o_mnt, fs_name);
1974         if (rc < 0) {
1975                 CT_ERROR(rc, "cannot find a Lustre filesystem mounted at '%s'",
1976                          opt.o_mnt);
1977                 return rc;
1978         }
1979
1980         opt.o_mnt_fd = open(opt.o_mnt, O_RDONLY);
1981         if (opt.o_mnt_fd < 0) {
1982                 rc = -errno;
1983                 CT_ERROR(rc, "cannot open mount point at '%s'",
1984                          opt.o_mnt);
1985                 return rc;
1986         }
1987
1988         return rc;
1989 }
1990
1991 static int ct_cleanup(void)
1992 {
1993         int     rc;
1994
1995         if (opt.o_mnt_fd >= 0) {
1996                 rc = close(opt.o_mnt_fd);
1997                 if (rc < 0) {
1998                         rc = -errno;
1999                         CT_ERROR(rc, "cannot close mount point");
2000                         return rc;
2001                 }
2002         }
2003
2004         if (arc_fd >= 0) {
2005                 rc = close(arc_fd);
2006                 if (rc < 0) {
2007                         rc = -errno;
2008                         CT_ERROR(rc, "cannot close archive root directory");
2009                         return rc;
2010                 }
2011         }
2012
2013         if (opt.o_archive_id_cnt > 0) {
2014                 free(opt.o_archive_id);
2015                 opt.o_archive_id = NULL;
2016                 opt.o_archive_id_cnt = 0;
2017         }
2018
2019         return 0;
2020 }
2021
2022 int main(int argc, char **argv)
2023 {
2024         int     rc;
2025
2026         snprintf(cmd_name, sizeof(cmd_name), "%s", basename(argv[0]));
2027         rc = ct_parseopts(argc, argv);
2028         if (rc < 0) {
2029                 CT_WARN("try '%s --help' for more information", cmd_name);
2030                 return -rc;
2031         }
2032
2033         rc = ct_setup();
2034         if (rc < 0)
2035                 goto error_cleanup;
2036
2037         switch (opt.o_action) {
2038         case CA_IMPORT:
2039                 rc = ct_import_recurse(opt.o_src);
2040                 break;
2041         case CA_REBIND:
2042                 rc = ct_rebind();
2043                 break;
2044         case CA_MAXSEQ:
2045                 rc = ct_max_sequence();
2046                 break;
2047         default:
2048                 rc = ct_run();
2049                 break;
2050         }
2051
2052         if (opt.o_action != CA_MAXSEQ)
2053                 CT_TRACE("process finished, errs: %d major, %d minor,"
2054                          " rc=%d (%s)", err_major, err_minor, rc,
2055                          strerror(-rc));
2056
2057 error_cleanup:
2058         ct_cleanup();
2059
2060         return -rc;
2061 }