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  */
  27 /* HSM copytool program for POSIX filesystem-based HSM's.
  28  *
  29  * An HSM copytool daemon acts on action requests from Lustre to copy files
  30  * to and from an HSM archive system. This one in particular makes regular
  31  * POSIX filesystem calls to a given path, where an HSM is presumably mounted.
  32  *
  33  * This particular tool can also import an existing HSM archive.
  34  */
  35 #include <stdint.h>
  36 #include <stdlib.h>
  37 #include <dirent.h>
  38 #include <errno.h>
  39 #include <utime.h>
  40 #include <sys/xattr.h>
  41 #include <sys/syscall.h>
  42 #include <sys/types.h>
  43 #include <lustre/lustre_idl.h>
  44 #include <lustre/lustreapi.h>
  45
  46 /* Progress reporting period */
  47 #define REPORT_INTERVAL_DEFAULT 30
  48 /* HSM hash subdir permissions */
  49 #define DIR_PERM S_IRWXU
  50 /* HSM hash file permissions */
  51 #define FILE_PERM (S_IRUSR | S_IWUSR)
  52
  53 #define ONE_MB 0x100000
  54
  55 /* copytool uses a 32b bitmask field to register with kuc
  56  * archive num = 0 => all
  57  * archive num from 1 to 32
  58  */
  59 #define MAX_ARCHIVE_CNT (sizeof(__u32) * 8)
  60
  61 enum ct_action {
  62         CA_IMPORT = 1,
  63         CA_REBIND,
  64         CA_MAXSEQ,
  65 };
  66
  67 struct options {
  68         int                      o_copy_attrs;
  69         int                      o_daemonize;
  70         int                      o_dry_run;
  71         int                      o_abort_on_error;
  72         int                      o_shadow_tree;
  73         int                      o_verbose;
  74         int                      o_copy_xattrs;
  75         int                      o_archive_cnt;
  76         int                      o_archive_id[MAX_ARCHIVE_CNT];
  77         int                      o_report_int;
  78         unsigned long long       o_bandwidth;
  79         size_t                   o_chunk_size;
  80         enum ct_action           o_action;
  81         char                    *o_event_fifo;
  82         char                    *o_mnt;
  83         char                    *o_hsm_root;
  84         char                    *o_src; /* for import, or rebind */
  85         char                    *o_dst; /* for import, or rebind */
  86 };
  87
  88 /* everything else is zeroed */
  89 struct options opt = {
  90         .o_copy_attrs = 1,
  91         .o_shadow_tree = 1,
  92         .o_verbose = LLAPI_MSG_INFO,
  93         .o_copy_xattrs = 1,
  94         .o_report_int = REPORT_INTERVAL_DEFAULT,
  95         .o_chunk_size = ONE_MB,
  96 };
  97
  98 /* hsm_copytool_private will hold an open FD on the lustre mount point
  99  * for us. Additionally open one on the archive FS root to make sure
 100  * it doesn't drop out from under us (and remind the admin to shutdown
 101  * the copytool before unmounting). */
 102
 103 static int arc_fd = -1;
 104
 105 static int err_major;
 106 static int err_minor;
 107
 108 static char cmd_name[PATH_MAX];
 109 static char fs_name[MAX_OBD_NAME + 1];
 110
 111 static struct hsm_copytool_private *ctdata;
 112
 113
 114 #define CT_ERROR(_rc, _format, ...) \
 115         llapi_error(LLAPI_MSG_ERROR, _rc,                               \
 116                     "%s[%ld]: "_format,                                 \
 117                     cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 118 #define CT_DEBUG(_format, ...) \
 119         llapi_error(LLAPI_MSG_DEBUG | LLAPI_MSG_NO_ERRNO, 0,            \
 120                     "%s[%ld]: "_format,                                 \
 121                     cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 122 #define CT_WARN(_format, ...) \
 123         llapi_error(LLAPI_MSG_WARN | LLAPI_MSG_NO_ERRNO, 0,             \
 124                     "%s[%ld]: "_format,                                 \
 125                     cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 126 #define CT_TRACE(_format, ...) \
 127         llapi_error(LLAPI_MSG_INFO | LLAPI_MSG_NO_ERRNO, 0,             \
 128                     "%s[%ld]: "_format,                                 \
 129                     cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 130 #define CT_PRINTF(_format, ...) \
 131         llapi_printf(LLAPI_MSG_NORMAL,                                  \
 132                      "%s[%ld]: "_format,                                \
 133                      cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 134
 135 static void usage(const char *name, int rc)
 136 {
 137         fprintf(stdout,
 138         " Usage: %s [options]... <mode> <lustre_mount_point>\n"
 139         "The Lustre HSM Posix copy tool can be used as a daemon or "
 140         "as a command line tool\n"
 141         "The Lustre HSM daemon acts on action requests from Lustre\n"
 142         "to copy files to and from an HSM archive system.\n"
 143         "This POSIX-flavored daemon makes regular POSIX filesystem calls\n"
 144         "to an HSM mounted at a given hsm_root.\n"
 145         "   --daemon            Daemon mode, run in background\n"
 146         " Options:\n"
 147         "   --no-attr           Don't copy file attributes\n"
 148         "   --no-shadow         Don't create shadow namespace in archive\n"
 149         "   --no-xattr          Don't copy file extended attributes\n"
 150         "The Lustre HSM tool performs administrator-type actions\n"
 151         "on a Lustre HSM archive.\n"
 152         "This POSIX-flavored tool can link an existing HSM namespace\n"
 153         "into a Lustre filesystem.\n"
 154         " Usage:\n"
 155         "   %s [options] --import <src> <dst> <lustre_mount_point>\n"
 156         "      import an archived subtree at\n"
 157         "       <src> (relative to hsm_root) into the Lustre filesystem at\n"
 158         "       <dst> (absolute)\n"
 159         "   %s [options] --rebind <old_FID> <new_FID> <lustre_mount_point>\n"
 160         "      rebind an entry in the HSM to a new FID\n"
 161         "       <old_FID> old FID the HSM entry is bound to\n"
 162         "       <new_FID> new FID to bind the HSM entry to\n"
 163         "   %s [options] --rebind <list_file> <lustre_mount_point>\n"
 164         "      perform the rebind operation for all FID in the list file\n"
 165         "       each line of <list_file> consists of <old_FID> <new_FID>\n"
 166         "   %s [options] --max-sequence <fsname>\n"
 167         "       return the max fid sequence of archived files\n"
 168         "   --abort-on-error          Abort operation on major error\n"
 169         "   -A, --archive <#>         Archive number (repeatable)\n"
 170         "   -b, --bandwidth <bw>      Limit I/O bandwidth (unit can be used\n,"
 171         "                             default is MB)\n"
 172         "   --dry-run                 Don't run, just show what would be done\n"
 173         "   -c, --chunk-size <sz>     I/O size used during data copy\n"
 174         "                             (unit can be used, default is MB)\n"
 175         "   -f, --event-fifo <path>   Write events stream to fifo\n"
 176         "   -p, --hsm-root <path>     Target HSM mount point\n"
 177         "   -q, --quiet               Produce less verbose output\n"
 178         "   -u, --update-interval <s> Interval between progress reports sent\n"
 179         "                             to Coordinator\n"
 180         "   -v, --verbose             Produce more verbose output\n",
 181         cmd_name, cmd_name, cmd_name, cmd_name, cmd_name);
 182
 183         exit(rc);
 184 }
 185
 186 static int ct_parseopts(int argc, char * const *argv)
 187 {
 188         struct option long_opts[] = {
 189                 {"abort-on-error", no_argument,       &opt.o_abort_on_error, 1},
 190                 {"abort_on_error", no_argument,       &opt.o_abort_on_error, 1},
 191                 {"archive",        required_argument, NULL,                'A'},
 192                 {"bandwidth",      required_argument, NULL,                'b'},
 193                 {"chunk-size",     required_argument, NULL,                'c'},
 194                 {"chunk_size",     required_argument, NULL,                'c'},
 195                 {"daemon",         no_argument,       &opt.o_daemonize,     1},
 196                 {"event-fifo",     required_argument, NULL,                'f'},
 197                 {"event_fifo",     required_argument, NULL,                'f'},
 198                 {"dry-run",        no_argument,       &opt.o_dry_run,       1},
 199                 {"help",           no_argument,       NULL,                'h'},
 200                 {"hsm-root",       required_argument, NULL,                'p'},
 201                 {"hsm_root",       required_argument, NULL,                'p'},
 202                 {"import",         no_argument,       NULL,                'i'},
 203                 {"max-sequence",   no_argument,       NULL,                'M'},
 204                 {"max_sequence",   no_argument,       NULL,                'M'},
 205                 {"no-attr",        no_argument,       &opt.o_copy_attrs,    0},
 206                 {"no_attr",        no_argument,       &opt.o_copy_attrs,    0},
 207                 {"no-shadow",      no_argument,       &opt.o_shadow_tree,   0},
 208                 {"no_shadow",      no_argument,       &opt.o_shadow_tree,   0},
 209                 {"no-xattr",       no_argument,       &opt.o_copy_xattrs,   0},
 210                 {"no_xattr",       no_argument,       &opt.o_copy_xattrs,   0},
 211                 {"quiet",          no_argument,       NULL,                'q'},
 212                 {"rebind",         no_argument,       NULL,                'r'},
 213                 {"update-interval", required_argument,  NULL,              'u'},
 214                 {"update_interval", required_argument,  NULL,              'u'},
 215                 {"verbose",        no_argument,       NULL,                'v'},
 216                 {0, 0, 0, 0}
 217         };
 218         int                      c, rc;
 219         unsigned long long       value;
 220         unsigned long long       unit;
 221
 222         optind = 0;
 223         while ((c = getopt_long(argc, argv, "A:b:c:f:hiMp:qru:v",
 224                                 long_opts, NULL)) != -1) {
 225                 switch (c) {
 226                 case 'A':
 227                         if ((opt.o_archive_cnt >= MAX_ARCHIVE_CNT) ||
 228                             (atoi(optarg) >= MAX_ARCHIVE_CNT)) {
 229                                 rc = -E2BIG;
 230                                 CT_ERROR(rc, "archive number must be less"
 231                                          "than %zu", MAX_ARCHIVE_CNT);
 232                                 return rc;
 233                         }
 234                         opt.o_archive_id[opt.o_archive_cnt] = atoi(optarg);
 235                         opt.o_archive_cnt++;
 236                         break;
 237                 case 'b': /* -b and -c have both a number with unit as arg */
 238                 case 'c':
 239                         unit = ONE_MB;
 240                         if (llapi_parse_size(optarg, &value, &unit, 0) < 0) {
 241                                 rc = -EINVAL;
 242                                 CT_ERROR(rc, "bad value for -%c '%s'", c,
 243                                          optarg);
 244                                 return rc;
 245                         }
 246                         if (c == 'c')
 247                                 opt.o_chunk_size = value;
 248                         else
 249                                 opt.o_bandwidth = value;
 250                         break;
 251                 case 'f':
 252                         opt.o_event_fifo = optarg;
 253                         break;
 254                 case 'h':
 255                         usage(argv[0], 0);
 256                 case 'i':
 257                         opt.o_action = CA_IMPORT;
 258                         break;
 259                 case 'M':
 260                         opt.o_action = CA_MAXSEQ;
 261                         break;
 262                 case 'p':
 263                         opt.o_hsm_root = optarg;
 264                         break;
 265                 case 'q':
 266                         opt.o_verbose--;
 267                         break;
 268                 case 'r':
 269                         opt.o_action = CA_REBIND;
 270                         break;
 271                 case 'u':
 272                         opt.o_report_int = atoi(optarg);
 273                         if (opt.o_report_int < 0) {
 274                                 rc = -EINVAL;
 275                                 CT_ERROR(rc, "bad value for -%c '%s'", c,
 276                                          optarg);
 277                                 return rc;
 278                         }
 279                         break;
 280                 case 'v':
 281                         opt.o_verbose++;
 282                         break;
 283                 case 0:
 284                         break;
 285                 default:
 286                         return -EINVAL;
 287                 }
 288         }
 289
 290         switch (opt.o_action) {
 291         case CA_IMPORT:
 292                 /* src dst mount_point */
 293                 if (argc != optind + 3) {
 294                         rc = -EINVAL;
 295                         CT_ERROR(rc, "--import requires 2 arguments");
 296                         return rc;
 297                 }
 298                 opt.o_src = argv[optind++];
 299                 opt.o_dst = argv[optind++];
 300                 break;
 301         case CA_REBIND:
 302                 /* FID1 FID2 mount_point or FILE mount_point */
 303                 if (argc == optind + 2) {
 304                         opt.o_src = argv[optind++];
 305                         opt.o_dst = NULL;
 306                 } else if (argc == optind + 3) {
 307                         opt.o_src = argv[optind++];
 308                         opt.o_dst = argv[optind++];
 309                 } else {
 310                         rc = -EINVAL;
 311                         CT_ERROR(rc, "--rebind requires 1 or 2 arguments");
 312                         return rc;
 313                 }
 314                 break;
 315         case CA_MAXSEQ:
 316         default:
 317                 /* just mount point */
 318                 break;
 319         }
 320
 321         if (argc != optind + 1) {
 322                 rc = -EINVAL;
 323                 CT_ERROR(rc, "no mount point specified");
 324                 return rc;
 325         }
 326
 327         opt.o_mnt = argv[optind];
 328
 329         CT_TRACE("action=%d src=%s dst=%s mount_point=%s",
 330                  opt.o_action, opt.o_src, opt.o_dst, opt.o_mnt);
 331
 332         if (!opt.o_dry_run && opt.o_hsm_root == NULL) {
 333                 rc = -EINVAL;
 334                 CT_ERROR(rc, "must specify a root directory for the backend");
 335                 return rc;
 336         }
 337
 338         if (opt.o_action == CA_IMPORT) {
 339                 if (opt.o_src && opt.o_src[0] == '/') {
 340                         rc = -EINVAL;
 341                         CT_ERROR(rc,
 342                                  "source path must be relative to HSM root");
 343                         return rc;
 344                 }
 345
 346                 if (opt.o_dst && opt.o_dst[0] != '/') {
 347                         rc = -EINVAL;
 348                         CT_ERROR(rc, "destination path must be absolute");
 349                         return rc;
 350                 }
 351         }
 352
 353         return 0;
 354 }
 355
 356 /* mkdir -p path */
 357 static int ct_mkdir_p(const char *path)
 358 {
 359         char    *saved, *ptr;
 360         int      rc;
 361
 362         ptr = strdup(path);
 363         saved = ptr;
 364         while (*ptr == '/')
 365                 ptr++;
 366
 367         while ((ptr = strchr(ptr, '/')) != NULL) {
 368                 *ptr = '\0';
 369                 rc = mkdir(saved, DIR_PERM);
 370                 *ptr = '/';
 371                 if (rc < 0 && errno != EEXIST) {
 372                         rc = -errno;
 373                         CT_ERROR(rc, "cannot mkdir '%s'", path);
 374                         free(saved);
 375                         return rc;
 376                 }
 377                 ptr++;
 378         }
 379
 380         free(saved);
 381
 382         return 0;
 383 }
 384
 385 static int ct_save_stripe(int src_fd, const char *src, const char *dst)
 386 {
 387         char                     lov_file[PATH_MAX];
 388         char                     lov_buf[XATTR_SIZE_MAX];
 389         struct lov_user_md      *lum;
 390         int                      rc;
 391         int                      fd;
 392         ssize_t                  xattr_size;
 393
 394         snprintf(lov_file, sizeof(lov_file), "%s.lov", dst);
 395         CT_TRACE("saving stripe info of '%s' in %s", src, lov_file);
 396
 397         xattr_size = fgetxattr(src_fd, XATTR_LUSTRE_LOV, lov_buf,
 398                                sizeof(lov_buf));
 399         if (xattr_size < 0) {
 400                 rc = -errno;
 401                 CT_ERROR(rc, "cannot get stripe info on '%s'", src);
 402                 return rc;
 403         }
 404
 405         lum = (struct lov_user_md *)lov_buf;
 406
 407         if (lum->lmm_magic == LOV_USER_MAGIC_V1 ||
 408             lum->lmm_magic == LOV_USER_MAGIC_V3) {
 409                 /* Set stripe_offset to -1 so that it is not interpreted as a
 410                  * hint on restore. */
 411                 lum->lmm_stripe_offset = -1;
 412         }
 413
 414         fd = open(lov_file, O_TRUNC | O_CREAT | O_WRONLY, FILE_PERM);
 415         if (fd < 0) {
 416                 rc = -errno;
 417                 CT_ERROR(rc, "cannot open '%s'", lov_file);
 418                 goto err_cleanup;
 419         }
 420
 421         rc = write(fd, lum, xattr_size);
 422         if (rc < 0) {
 423                 rc = -errno;
 424                 CT_ERROR(rc, "cannot write %zd bytes to '%s'",
 425                          xattr_size, lov_file);
 426                 close(fd);
 427                 goto err_cleanup;
 428         }
 429
 430         rc = close(fd);
 431         if (rc < 0) {
 432                 rc = -errno;
 433                 CT_ERROR(rc, "cannot close '%s'", lov_file);
 434                 goto err_cleanup;
 435         }
 436
 437         return 0;
 438
 439 err_cleanup:
 440         unlink(lov_file);
 441
 442         return rc;
 443 }
 444
 445 static int ct_load_stripe(const char *src, void *lovea, size_t *lovea_size)
 446 {
 447         char     lov_file[PATH_MAX];
 448         int      rc;
 449         int      fd;
 450
 451         snprintf(lov_file, sizeof(lov_file), "%s.lov", src);
 452         CT_TRACE("reading stripe rules from '%s' for '%s'", lov_file, src);
 453
 454         fd = open(lov_file, O_RDONLY);
 455         if (fd < 0) {
 456                 CT_ERROR(errno, "cannot open '%s'", lov_file);
 457                 return -ENODATA;
 458         }
 459
 460         rc = read(fd, lovea, *lovea_size);
 461         if (rc < 0) {
 462                 CT_ERROR(errno, "cannot read %zu bytes from '%s'",
 463                          *lovea_size, lov_file);
 464                 close(fd);
 465                 return -ENODATA;
 466         }
 467
 468         *lovea_size = rc;
 469         close(fd);
 470
 471         return 0;
 472 }
 473
 474 static int ct_restore_stripe(const char *src, const char *dst, int dst_fd,
 475                              const void *lovea, size_t lovea_size)
 476 {
 477         int     rc;
 478
 479         rc = fsetxattr(dst_fd, XATTR_LUSTRE_LOV, lovea, lovea_size,
 480                        XATTR_CREATE);
 481         if (rc < 0) {
 482                 CT_ERROR(errno, "cannot set lov EA on '%s'", dst);
 483                 rc = -errno;
 484         }
 485
 486         return rc;
 487 }
 488
 489 static void bandwidth_ctl_delay(int wsize)
 490 {
 491         static unsigned long long       tot_bytes;
 492         static time_t                   start_time;
 493         static time_t                   last_time;
 494         time_t                          now = time(0);
 495         double                          tot_time;
 496         double                          excess;
 497         unsigned int                    sleep_time;
 498
 499         if (now > last_time + 5) {
 500                 tot_bytes = 0;
 501                 start_time = last_time = now;
 502         }
 503
 504         tot_bytes += wsize;
 505         tot_time = now - start_time;
 506         if (tot_time < 1)
 507                 tot_time = 1;
 508
 509         excess = tot_bytes - tot_time * opt.o_bandwidth;
 510         sleep_time = excess * 1000000 / opt.o_bandwidth;
 511         if ((now - start_time) % 10 == 1)
 512                 CT_TRACE("bandwith control: excess=%E sleep for %dus", excess,
 513                          sleep_time);
 514
 515         if (excess > 0)
 516                 usleep(sleep_time);
 517
 518         last_time = now;
 519 }
 520
 521 static int ct_copy_data(struct hsm_copyaction_private *hcp, const char *src,
 522                         const char *dst, int src_fd, int dst_fd,
 523                         const struct hsm_action_item *hai, long hal_flags)
 524 {
 525         struct hsm_extent        he;
 526         __u64                    offset = hai->hai_extent.offset;
 527         struct stat              src_st;
 528         struct stat              dst_st;
 529         char                    *buf = NULL;
 530         __u64                    write_total = 0;
 531         __u64                    length;
 532         time_t                   last_print_time = time(NULL);
 533         int                      rc = 0;
 534
 535         CT_TRACE("going to copy data from '%s' to '%s'", src, dst);
 536
 537         if (fstat(src_fd, &src_st) < 0) {
 538                 rc = -errno;
 539                 CT_ERROR(rc, "cannot stat '%s'", src);
 540                 return rc;
 541         }
 542
 543         if (!S_ISREG(src_st.st_mode)) {
 544                 rc = -EINVAL;
 545                 CT_ERROR(rc, "'%s' is not a regular file", src);
 546                 return rc;
 547         }
 548
 549         if (fstat(dst_fd, &dst_st) < 0) {
 550                 rc = -errno;
 551                 CT_ERROR(rc, "cannot stat '%s'", dst);
 552                 return rc;
 553         }
 554
 555         if (!S_ISREG(dst_st.st_mode)) {
 556                 rc = -EINVAL;
 557                 CT_ERROR(rc, "'%s' is not a regular file", dst);
 558                 return rc;
 559         }
 560
 561         rc = lseek(src_fd, hai->hai_extent.offset, SEEK_SET);
 562         if (rc < 0) {
 563                 rc = -errno;
 564                 CT_ERROR(rc,
 565                          "cannot seek for read to "LPU64" (len %jd) in '%s'",
 566                          hai->hai_extent.offset, (intmax_t)src_st.st_size, src);
 567                 return rc;
 568         }
 569
 570         /* Don't read beyond a given extent */
 571         length = min(hai->hai_extent.length, src_st.st_size);
 572
 573         he.offset = offset;
 574         he.length = 0;
 575         rc = llapi_hsm_action_progress(hcp, &he, length, 0);
 576         if (rc < 0) {
 577                 /* Action has been canceled or something wrong
 578                  * is happening. Stop copying data. */
 579                 CT_ERROR(rc, "progress ioctl for copy '%s'->'%s' failed",
 580                          src, dst);
 581                 goto out;
 582         }
 583
 584         errno = 0;
 585
 586         buf = malloc(opt.o_chunk_size);
 587         if (buf == NULL) {
 588                 rc = -ENOMEM;
 589                 goto out;
 590         }
 591
 592         CT_DEBUG("Going to copy "LPU64" bytes %s -> %s\n", length, src, dst);
 593
 594         while (write_total < length) {
 595                 ssize_t rsize;
 596                 ssize_t wsize;
 597                 int     chunk = (length - write_total > opt.o_chunk_size) ?
 598                                  opt.o_chunk_size : length - write_total;
 599
 600                 rsize = pread(src_fd, buf, chunk, offset);
 601                 if (rsize == 0)
 602                         /* EOF */
 603                         break;
 604
 605                 if (rsize < 0) {
 606                         rc = -errno;
 607                         CT_ERROR(rc, "cannot read from '%s'", src);
 608                         break;
 609                 }
 610
 611                 wsize = pwrite(dst_fd, buf, rsize, offset);
 612                 if (wsize < 0) {
 613                         rc = -errno;
 614                         CT_ERROR(rc, "cannot write to '%s'", dst);
 615                         break;
 616                 }
 617
 618                 write_total += wsize;
 619                 offset += wsize;
 620
 621                 if (opt.o_bandwidth != 0)
 622                         /* sleep if needed, to honor bandwidth limits */
 623                         bandwidth_ctl_delay(wsize);
 624
 625                 if (time(0) >= last_print_time + opt.o_report_int) {
 626                         last_print_time = time(0);
 627                         CT_TRACE("%%"LPU64" ", 100 * write_total / length);
 628                         he.length = write_total;
 629                         rc = llapi_hsm_action_progress(hcp, &he, length, 0);
 630                         if (rc < 0) {
 631                                 /* Action has been canceled or something wrong
 632                                  * is happening. Stop copying data. */
 633                                 CT_ERROR(rc, "progress ioctl for copy"
 634                                          " '%s'->'%s' failed", src, dst);
 635                                 goto out;
 636                         }
 637                 }
 638                 rc = 0;
 639         }
 640
 641 out:
 642         /*
 643          * truncate restored file
 644          * size is taken from the archive this is done to support
 645          * restore after a force release which leaves the file with the
 646          * wrong size (can big bigger than the new size)
 647          */
 648         if ((hai->hai_action == HSMA_RESTORE) &&
 649             (src_st.st_size < dst_st.st_size)) {
 650                 /*
 651                  * make sure the file is on disk before reporting success.
 652                  */
 653                 rc = ftruncate(dst_fd, src_st.st_size);
 654                 if (rc < 0) {
 655                         rc = -errno;
 656                         CT_ERROR(rc, "cannot truncate '%s' to size %jd",
 657                                  dst, (intmax_t)src_st.st_size);
 658                         err_major++;
 659                 }
 660         }
 661
 662         if (buf != NULL)
 663                 free(buf);
 664
 665         return rc;
 666 }
 667
 668 /* Copy file attributes from file src to file dest */
 669 static int ct_copy_attr(const char *src, const char *dst, int src_fd,
 670                         int dst_fd)
 671 {
 672         struct stat     st;
 673         struct timeval  times[2];
 674         int             rc;
 675
 676         if (fstat(src_fd, &st) < 0) {
 677                 rc = -errno;
 678                 CT_ERROR(rc, "cannot stat '%s'", src);
 679                 return rc;
 680         }
 681
 682         times[0].tv_sec = st.st_atime;
 683         times[0].tv_usec = 0;
 684         times[1].tv_sec = st.st_mtime;
 685         times[1].tv_usec = 0;
 686         if (fchmod(dst_fd, st.st_mode) < 0 ||
 687             fchown(dst_fd, st.st_uid, st.st_gid) < 0 ||
 688             futimes(dst_fd, times) < 0) {
 689                 rc = -errno;
 690                 CT_ERROR(rc, "cannot set attributes of '%s'", src);
 691                 return rc;
 692         }
 693
 694         return 0;
 695 }
 696
 697 static int ct_copy_xattr(const char *src, const char *dst, int src_fd,
 698                          int dst_fd, bool is_restore)
 699 {
 700         char     list[XATTR_LIST_MAX];
 701         char     value[XATTR_SIZE_MAX];
 702         char    *name;
 703         ssize_t  list_len;
 704         int      rc;
 705
 706         list_len = flistxattr(src_fd, list, sizeof(list));
 707         if (list_len < 0)
 708                 return -errno;
 709
 710         name = list;
 711         while (name < list + list_len) {
 712                 rc = fgetxattr(src_fd, name, value, sizeof(value));
 713                 if (rc < 0)
 714                         return -errno;
 715
 716                 /* when we restore, we do not restore lustre xattr */
 717                 if (!is_restore ||
 718                     (strncmp(XATTR_TRUSTED_PREFIX, name,
 719                              sizeof(XATTR_TRUSTED_PREFIX) - 1) != 0)) {
 720                         rc = fsetxattr(dst_fd, name, value, rc, 0);
 721                         CT_TRACE("fsetxattr of '%s' on '%s' rc=%d (%s)",
 722                                  name, dst, rc, strerror(errno));
 723                         /* lustre.* attrs aren't supported on other FS's */
 724                         if (rc < 0 && errno != EOPNOTSUPP) {
 725                                 rc = -errno;
 726                                 CT_ERROR(rc, "cannot set extended attribute"
 727                                          " '%s' of '%s'",
 728                                          name, dst);
 729                                 return rc;
 730                         }
 731                 }
 732                 name += strlen(name) + 1;
 733         }
 734
 735         return 0;
 736 }
 737
 738 static int ct_path_lustre(char *buf, int sz, const char *mnt,
 739                           const lustre_fid *fid)
 740 {
 741         return snprintf(buf, sz, "%s/%s/fid/"DFID_NOBRACE, mnt,
 742                         dot_lustre_name, PFID(fid));
 743 }
 744
 745 static int ct_path_archive(char *buf, int sz, const char *archive_dir,
 746                            const lustre_fid *fid)
 747 {
 748         return snprintf(buf, sz, "%s/%04x/%04x/%04x/%04x/%04x/%04x/"
 749                         DFID_NOBRACE, archive_dir,
 750                         (fid)->f_oid       & 0xFFFF,
 751                         (fid)->f_oid >> 16 & 0xFFFF,
 752                         (unsigned int)((fid)->f_seq       & 0xFFFF),
 753                         (unsigned int)((fid)->f_seq >> 16 & 0xFFFF),
 754                         (unsigned int)((fid)->f_seq >> 32 & 0xFFFF),
 755                         (unsigned int)((fid)->f_seq >> 48 & 0xFFFF),
 756                         PFID(fid));
 757 }
 758
 759 static bool ct_is_retryable(int err)
 760 {
 761         return err == -ETIMEDOUT;
 762 }
 763
 764 static int ct_begin_restore(struct hsm_copyaction_private **phcp,
 765                             const struct hsm_action_item *hai,
 766                             int mdt_index, int open_flags)
 767 {
 768         char     src[PATH_MAX];
 769         int      rc;
 770
 771         rc = llapi_hsm_action_begin(phcp, ctdata, hai, mdt_index, open_flags,
 772                                     false);
 773         if (rc < 0) {
 774                 ct_path_lustre(src, sizeof(src), opt.o_mnt, &hai->hai_fid);
 775                 CT_ERROR(rc, "llapi_hsm_action_begin() on '%s' failed", src);
 776         }
 777
 778         return rc;
 779 }
 780
 781 static int ct_begin(struct hsm_copyaction_private **phcp,
 782                     const struct hsm_action_item *hai)
 783 {
 784         /* Restore takes specific parameters. Call the same function w/ default
 785          * values for all other operations. */
 786         return ct_begin_restore(phcp, hai, -1, 0);
 787 }
 788
 789 static int ct_fini(struct hsm_copyaction_private **phcp,
 790                    const struct hsm_action_item *hai, int hp_flags, int ct_rc)
 791 {
 792         struct hsm_copyaction_private   *hcp;
 793         char                             lstr[PATH_MAX];
 794         int                              rc;
 795
 796         CT_TRACE("Action completed, notifying coordinator "
 797                  "cookie="LPX64", FID="DFID", hp_flags=%d err=%d",
 798                  hai->hai_cookie, PFID(&hai->hai_fid),
 799                  hp_flags, -ct_rc);
 800
 801         ct_path_lustre(lstr, sizeof(lstr), opt.o_mnt, &hai->hai_fid);
 802
 803         if (phcp == NULL || *phcp == NULL) {
 804                 rc = llapi_hsm_action_begin(&hcp, ctdata, hai, -1, 0, true);
 805                 if (rc < 0) {
 806                         CT_ERROR(rc, "llapi_hsm_action_begin() on '%s' failed",
 807                                  lstr);
 808                         return rc;
 809                 }
 810                 phcp = &hcp;
 811         }
 812
 813         rc = llapi_hsm_action_end(phcp, &hai->hai_extent, hp_flags, abs(ct_rc));
 814         if (rc == -ECANCELED)
 815                 CT_ERROR(rc, "completed action on '%s' has been canceled: "
 816                          "cookie="LPX64", FID="DFID, lstr, hai->hai_cookie,
 817                          PFID(&hai->hai_fid));
 818         else if (rc < 0)
 819                 CT_ERROR(rc, "llapi_hsm_action_end() on '%s' failed", lstr);
 820         else
 821                 CT_TRACE("llapi_hsm_action_end() on '%s' ok (rc=%d)",
 822                          lstr, rc);
 823
 824         return rc;
 825 }
 826
 827 static int ct_archive(const struct hsm_action_item *hai, const long hal_flags)
 828 {
 829         struct hsm_copyaction_private   *hcp = NULL;
 830         char                             src[PATH_MAX];
 831         char                             dst[PATH_MAX] = "";
 832         int                              rc;
 833         int                              rcf = 0;
 834         bool                             rename_needed = false;
 835         int                              hp_flags = 0;
 836         int                              open_flags;
 837         int                              src_fd = -1;
 838         int                              dst_fd = -1;
 839
 840         rc = ct_begin(&hcp, hai);
 841         if (rc < 0)
 842                 goto fini_major;
 843
 844         /* we fill archive so:
 845          * source = data FID
 846          * destination = lustre FID
 847          */
 848         ct_path_lustre(src, sizeof(src), opt.o_mnt, &hai->hai_dfid);
 849         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, &hai->hai_fid);
 850         if (hai->hai_extent.length == -1) {
 851                 /* whole file, write it to tmp location and atomically
 852                  * replace old archived file */
 853                 strncat(dst, "_tmp", sizeof(dst) - strlen(dst) - 1);
 854                 /* we cannot rely on the same test because ct_copy_data()
 855                  * updates hai_extent.length */
 856                 rename_needed = true;
 857         }
 858
 859         CT_TRACE("archiving '%s' to '%s'", src, dst);
 860
 861         if (opt.o_dry_run) {
 862                 rc = 0;
 863                 goto fini_major;
 864         }
 865
 866         rc = ct_mkdir_p(dst);
 867         if (rc < 0) {
 868                 CT_ERROR(rc, "mkdir_p '%s' failed", dst);
 869                 goto fini_major;
 870         }
 871
 872         src_fd = llapi_hsm_action_get_fd(hcp);
 873         if (src_fd < 0) {
 874                 rc = -errno;
 875                 CT_ERROR(rc, "cannot open '%s' for read", src);
 876                 goto fini_major;
 877         }
 878
 879         open_flags = O_WRONLY | O_NOFOLLOW | O_NONBLOCK;
 880         /* If extent is specified, don't truncate an old archived copy */
 881         open_flags |= ((hai->hai_extent.length == -1) ? O_TRUNC : 0) | O_CREAT;
 882
 883         dst_fd = open(dst, open_flags, FILE_PERM);
 884         if (dst_fd == -1) {
 885                 rc = -errno;
 886                 CT_ERROR(rc, "cannot open '%s' for write", dst);
 887                 goto fini_major;
 888         }
 889
 890         /* saving stripe is not critical */
 891         rc = ct_save_stripe(src_fd, src, dst);
 892         if (rc < 0)
 893                 CT_ERROR(rc, "cannot save file striping info of '%s' in '%s'",
 894                          src, dst);
 895
 896         rc = ct_copy_data(hcp, src, dst, src_fd, dst_fd, hai, hal_flags);
 897         if (rc < 0) {
 898                 CT_ERROR(rc, "data copy failed from '%s' to '%s'", src, dst);
 899                 goto fini_major;
 900         }
 901
 902         rc = fsync(dst_fd);
 903         if (rc < 0) {
 904                 rc = -errno;
 905                 CT_ERROR(rc, "cannot flush '%s' archive file '%s'", src, dst);
 906                 goto fini_major;
 907         }
 908
 909         CT_TRACE("data archiving for '%s' to '%s' done", src, dst);
 910
 911         /* attrs will remain on the MDS; no need to copy them, except possibly
 912           for disaster recovery */
 913         if (opt.o_copy_attrs) {
 914                 rc = ct_copy_attr(src, dst, src_fd, dst_fd);
 915                 if (rc < 0) {
 916                         CT_ERROR(rc, "cannot copy attr of '%s' to '%s'",
 917                                  src, dst);
 918                         rcf = rc;
 919                 }
 920                 CT_TRACE("attr file for '%s' saved to archive '%s'",
 921                          src, dst);
 922         }
 923
 924         /* xattrs will remain on the MDS; no need to copy them, except possibly
 925          for disaster recovery */
 926         if (opt.o_copy_xattrs) {
 927                 rc = ct_copy_xattr(src, dst, src_fd, dst_fd, false);
 928                 if (rc < 0) {
 929                         CT_ERROR(rc, "cannot copy xattr of '%s' to '%s'",
 930                                  src, dst);
 931                         rcf = rcf ? rcf : rc;
 932                 }
 933                 CT_TRACE("xattr file for '%s' saved to archive '%s'",
 934                          src, dst);
 935         }
 936
 937         if (rename_needed == true) {
 938                 char     tmp_src[PATH_MAX];
 939                 char     tmp_dst[PATH_MAX];
 940
 941                 /* atomically replace old archived file */
 942                 ct_path_archive(src, sizeof(src), opt.o_hsm_root,
 943                                 &hai->hai_fid);
 944                 rc = rename(dst, src);
 945                 if (rc < 0) {
 946                         rc = -errno;
 947                         CT_ERROR(rc, "cannot rename '%s' to '%s'", dst, src);
 948                         goto fini_major;
 949                 }
 950                 /* rename lov file */
 951                 snprintf(tmp_src, sizeof(tmp_src), "%s.lov", src);
 952                 snprintf(tmp_dst, sizeof(tmp_dst), "%s.lov", dst);
 953                 rc = rename(tmp_dst, tmp_src);
 954                 if (rc < 0)
 955                         CT_ERROR(errno, "cannot rename '%s' to '%s'",
 956                                  tmp_dst, tmp_src);
 957         }
 958
 959         if (opt.o_shadow_tree) {
 960                 /* Create a namespace of softlinks that shadows the original
 961                  * Lustre namespace.  This will only be current at
 962                  * time-of-archive (won't follow renames).
 963                  * WARNING: release won't kill these links; a manual
 964                  * cleanup of dead links would be required.
 965                  */
 966                 char             buf[PATH_MAX];
 967                 long long        recno = -1;
 968                 int              linkno = 0;
 969                 char            *ptr;
 970                 int              depth = 0;
 971                 ssize_t          sz;
 972
 973                 sprintf(buf, DFID, PFID(&hai->hai_fid));
 974                 sprintf(src, "%s/shadow/", opt.o_hsm_root);
 975
 976                 ptr = opt.o_hsm_root;
 977                 while (*ptr)
 978                         (*ptr++ == '/') ? depth-- : 0;
 979
 980                 rc = llapi_fid2path(opt.o_mnt, buf, src + strlen(src),
 981                                     sizeof(src) - strlen(src), &recno, &linkno);
 982                 if (rc < 0) {
 983                         CT_ERROR(rc, "cannot get FID of '%s'", buf);
 984                         rcf = rcf ? rcf : rc;
 985                         goto fini_minor;
 986                 }
 987
 988                 /* Figure out how many parent dirs to symlink back */
 989                 ptr = src;
 990                 while (*ptr)
 991                         (*ptr++ == '/') ? depth++ : 0;
 992                 sprintf(buf, "..");
 993                 while (--depth > 1)
 994                         strcat(buf, "/..");
 995
 996                 ct_path_archive(dst, sizeof(dst), buf, &hai->hai_fid);
 997
 998                 if (ct_mkdir_p(src)) {
 999                         CT_ERROR(errno, "mkdir_p '%s' failed", src);
1000                         rcf = rcf ? rcf : -errno;
1001                         goto fini_minor;
1002                 }
1003                 /* symlink already exists ? */
1004                 sz = readlink(src, buf, sizeof(buf));
1005                 /* detect truncation */
1006                 if (sz == sizeof(buf)) {
1007                         rcf = rcf ? rcf : -E2BIG;
1008                         CT_ERROR(rcf, "readlink '%s' truncated", src);
1009                         goto fini_minor;
1010                 }
1011                 if (sz >= 0) {
1012                         buf[sz] = '\0';
1013                         if (sz == 0 || strncmp(buf, dst, sz) != 0) {
1014                                 if (unlink(src) && errno != ENOENT) {
1015                                         CT_ERROR(errno,
1016                                                  "cannot unlink symlink '%s'",
1017                                                  src);
1018                                         rcf = rcf ? rcf : -errno;
1019                                         goto fini_minor;
1020                                 /* unlink old symlink done */
1021                                 CT_TRACE("remove old symlink '%s' pointing"
1022                                          " to '%s'", src, buf);
1023                                 }
1024                         } else {
1025                                 /* symlink already ok */
1026                                 CT_TRACE("symlink '%s' already pointing"
1027                                          " to '%s'", src, dst);
1028                                 rcf = 0;
1029                                 goto fini_minor;
1030                         }
1031                 }
1032                 if (symlink(dst, src)) {
1033                         CT_ERROR(errno, "cannot symlink '%s' to '%s'",
1034                                  src, dst);
1035                         rcf = rcf ? rcf : -errno;
1036                         goto fini_minor;
1037                 }
1038                 CT_TRACE("symlink '%s' to '%s' done", src, dst);
1039         }
1040 fini_minor:
1041         if (rcf)
1042                 err_minor++;
1043         goto out;
1044
1045
1046 fini_major:
1047         err_major++;
1048
1049         unlink(dst);
1050         if (ct_is_retryable(rc))
1051                 hp_flags |= HP_FLAG_RETRY;
1052
1053         rcf = rc;
1054
1055 out:
1056         if (!(src_fd < 0))
1057                 close(src_fd);
1058
1059         if (!(dst_fd < 0))
1060                 close(dst_fd);
1061
1062         rc = ct_fini(&hcp, hai, hp_flags, rcf);
1063
1064         return rc;
1065 }
1066
1067 static int ct_restore(const struct hsm_action_item *hai, const long hal_flags)
1068 {
1069         struct hsm_copyaction_private   *hcp = NULL;
1070         char                             src[PATH_MAX];
1071         char                             dst[PATH_MAX];
1072         char                             lov_buf[XATTR_SIZE_MAX];
1073         size_t                           lov_size = sizeof(lov_buf);
1074         int                              rc;
1075         int                              hp_flags = 0;
1076         int                              src_fd = -1;
1077         int                              dst_fd = -1;
1078         int                              mdt_index = -1; /* Not implemented */
1079         int                              open_flags = 0;
1080         bool                             set_lovea;
1081         struct lu_fid                    dfid;
1082         /* we fill lustre so:
1083          * source = lustre FID in the backend
1084          * destination = data FID = volatile file
1085          */
1086
1087         /* build backend file name from released file FID */
1088         ct_path_archive(src, sizeof(src), opt.o_hsm_root, &hai->hai_fid);
1089
1090         /* restore loads and sets the LOVEA w/o interpreting it to avoid
1091          * dependency on the structure format. */
1092         rc = ct_load_stripe(src, lov_buf, &lov_size);
1093         if (rc < 0) {
1094                 CT_WARN("cannot get stripe rules for '%s' (%s), use default",
1095                         src, strerror(-rc));
1096                 set_lovea = false;
1097         } else {
1098                 open_flags |= O_LOV_DELAY_CREATE;
1099                 set_lovea = true;
1100         }
1101
1102         rc = ct_begin_restore(&hcp, hai, mdt_index, open_flags);
1103         if (rc < 0)
1104                 goto fini;
1105
1106         /* get the FID of the volatile file */
1107         rc = llapi_hsm_action_get_dfid(hcp, &dfid);
1108         if (rc < 0) {
1109                 CT_ERROR(rc, "restoring "DFID
1110                          ", cannot get FID of created volatile file",
1111                          PFID(&hai->hai_fid));
1112                 goto fini;
1113         }
1114
1115         /* build volatile "file name", for messages */
1116         snprintf(dst, sizeof(dst), "{VOLATILE}="DFID, PFID(&dfid));
1117
1118         CT_TRACE("restoring data from '%s' to '%s'", src, dst);
1119
1120         if (opt.o_dry_run) {
1121                 rc = 0;
1122                 goto fini;
1123         }
1124
1125         src_fd = open(src, O_RDONLY | O_NOATIME | O_NONBLOCK | O_NOFOLLOW);
1126         if (src_fd < 0) {
1127                 rc = -errno;
1128                 CT_ERROR(rc, "cannot open '%s' for read", src);
1129                 goto fini;
1130         }
1131
1132         dst_fd = llapi_hsm_action_get_fd(hcp);
1133
1134         if (set_lovea) {
1135                 /* the layout cannot be allocated through .fid so we have to
1136                  * restore a layout */
1137                 rc = ct_restore_stripe(src, dst, dst_fd, lov_buf, lov_size);
1138                 if (rc < 0) {
1139                         CT_ERROR(rc, "cannot restore file striping info"
1140                                  " for '%s' from '%s'", dst, src);
1141                         err_major++;
1142                         goto fini;
1143                 }
1144         }
1145
1146         rc = ct_copy_data(hcp, src, dst, src_fd, dst_fd, hai, hal_flags);
1147         if (rc < 0) {
1148                 CT_ERROR(rc, "cannot copy data from '%s' to '%s'",
1149                          src, dst);
1150                 err_major++;
1151                 if (ct_is_retryable(rc))
1152                         hp_flags |= HP_FLAG_RETRY;
1153                 goto fini;
1154         }
1155
1156         CT_TRACE("data restore from '%s' to '%s' done", src, dst);
1157
1158 fini:
1159         rc = ct_fini(&hcp, hai, hp_flags, rc);
1160
1161         /* object swaping is done by cdt at copy end, so close of volatile file
1162          * cannot be done before */
1163         if (!(src_fd < 0))
1164                 close(src_fd);
1165
1166         if (!(dst_fd < 0))
1167                 close(dst_fd);
1168
1169         return rc;
1170 }
1171
1172 static int ct_remove(const struct hsm_action_item *hai, const long hal_flags)
1173 {
1174         struct hsm_copyaction_private   *hcp = NULL;
1175         char                             dst[PATH_MAX];
1176         int                              rc;
1177
1178         rc = ct_begin(&hcp, hai);
1179         if (rc < 0)
1180                 goto fini;
1181
1182         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, &hai->hai_fid);
1183
1184         CT_TRACE("removing file '%s'", dst);
1185
1186         if (opt.o_dry_run) {
1187                 rc = 0;
1188                 goto fini;
1189         }
1190
1191         rc = unlink(dst);
1192         if (rc < 0) {
1193                 rc = -errno;
1194                 CT_ERROR(rc, "cannot unlink '%s'", dst);
1195                 err_minor++;
1196                 goto fini;
1197         }
1198
1199         strncat(dst, ".lov", sizeof(dst) - strlen(dst) - 1);
1200         rc = unlink(dst);
1201         if (rc < 0) {
1202                 rc = -errno;
1203                 CT_ERROR(rc, "cannot unlink '%s'", dst);
1204                 err_minor++;
1205                 goto fini;
1206         }
1207
1208 fini:
1209         rc = ct_fini(&hcp, hai, 0, rc);
1210
1211         return rc;
1212 }
1213
1214 static int ct_process_item(struct hsm_action_item *hai, const long hal_flags)
1215 {
1216         int     rc = 0;
1217
1218         if (opt.o_verbose >= LLAPI_MSG_INFO || opt.o_dry_run) {
1219                 /* Print the original path */
1220                 char            fid[128];
1221                 char            path[PATH_MAX];
1222                 long long       recno = -1;
1223                 int             linkno = 0;
1224
1225                 sprintf(fid, DFID, PFID(&hai->hai_fid));
1226                 CT_TRACE("'%s' action %s reclen %d, cookie="LPX64,
1227                          fid, hsm_copytool_action2name(hai->hai_action),
1228                          hai->hai_len, hai->hai_cookie);
1229                 rc = llapi_fid2path(opt.o_mnt, fid, path,
1230                                     sizeof(path), &recno, &linkno);
1231                 if (rc < 0)
1232                         CT_ERROR(rc, "cannot get path of FID %s", fid);
1233                 else
1234                         CT_TRACE("processing file '%s'", path);
1235         }
1236
1237         switch (hai->hai_action) {
1238         /* set err_major, minor inside these functions */
1239         case HSMA_ARCHIVE:
1240                 rc = ct_archive(hai, hal_flags);
1241                 break;
1242         case HSMA_RESTORE:
1243                 rc = ct_restore(hai, hal_flags);
1244                 break;
1245         case HSMA_REMOVE:
1246                 rc = ct_remove(hai, hal_flags);
1247                 break;
1248         case HSMA_CANCEL:
1249                 CT_TRACE("cancel not implemented for file system '%s'",
1250                          opt.o_mnt);
1251                 /* Don't report progress to coordinator for this cookie:
1252                  * the copy function will get ECANCELED when reporting
1253                  * progress. */
1254                 err_minor++;
1255                 return 0;
1256                 break;
1257         default:
1258                 rc = -EINVAL;
1259                 CT_ERROR(rc, "unknown action %d, on '%s'", hai->hai_action,
1260                          opt.o_mnt);
1261                 err_minor++;
1262                 ct_fini(NULL, hai, 0, rc);
1263         }
1264
1265         return 0;
1266 }
1267
1268 struct ct_th_data {
1269         long                     hal_flags;
1270         struct hsm_action_item  *hai;
1271 };
1272
1273 static void *ct_thread(void *data)
1274 {
1275         struct ct_th_data *cttd = data;
1276         int rc;
1277
1278         rc = ct_process_item(cttd->hai, cttd->hal_flags);
1279
1280         free(cttd->hai);
1281         free(cttd);
1282         pthread_exit((void *)(intptr_t)rc);
1283 }
1284
1285 static int ct_process_item_async(const struct hsm_action_item *hai,
1286                                  long hal_flags)
1287 {
1288         pthread_attr_t           attr;
1289         pthread_t                thread;
1290         struct ct_th_data       *data;
1291         int                      rc;
1292
1293         data = malloc(sizeof(*data));
1294         if (data == NULL)
1295                 return -ENOMEM;
1296
1297         data->hai = malloc(hai->hai_len);
1298         if (data->hai == NULL) {
1299                 free(data);
1300                 return -ENOMEM;
1301         }
1302
1303         memcpy(data->hai, hai, hai->hai_len);
1304         data->hal_flags = hal_flags;
1305
1306         rc = pthread_attr_init(&attr);
1307         if (rc != 0) {
1308                 CT_ERROR(rc, "pthread_attr_init failed for '%s' service",
1309                          opt.o_mnt);
1310                 free(data->hai);
1311                 free(data);
1312                 return -rc;
1313         }
1314
1315         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
1316
1317         rc = pthread_create(&thread, &attr, ct_thread, data);
1318         if (rc != 0)
1319                 CT_ERROR(rc, "cannot create thread for '%s' service",
1320                          opt.o_mnt);
1321
1322         pthread_attr_destroy(&attr);
1323         return 0;
1324 }
1325
1326 static int ct_import_one(const char *src, const char *dst)
1327 {
1328         char            newarc[PATH_MAX];
1329         lustre_fid      fid;
1330         struct stat     st;
1331         int             rc;
1332
1333         CT_TRACE("importing '%s' from '%s'", dst, src);
1334
1335         if (stat(src, &st) < 0) {
1336                 rc = -errno;
1337                 CT_ERROR(rc, "cannot stat '%s'", src);
1338                 return rc;
1339         }
1340
1341         if (opt.o_dry_run)
1342                 return 0;
1343
1344         rc = llapi_hsm_import(dst,
1345                               opt.o_archive_cnt ? opt.o_archive_id[0] : 0,
1346                               &st, 0, 0, 0, 0, NULL, &fid);
1347         if (rc < 0) {
1348                 CT_ERROR(rc, "cannot import '%s' from '%s'", dst, src);
1349                 return rc;
1350         }
1351
1352         ct_path_archive(newarc, sizeof(newarc), opt.o_hsm_root, &fid);
1353
1354         rc = ct_mkdir_p(newarc);
1355         if (rc < 0) {
1356                 CT_ERROR(rc, "mkdir_p '%s' failed", newarc);
1357                 err_major++;
1358                 return rc;
1359
1360         }
1361
1362         /* Lots of choices now: mv, ln, ln -s ? */
1363         rc = link(src, newarc); /* hardlink */
1364         if (rc < 0) {
1365                 rc = -errno;
1366                 CT_ERROR(rc, "cannot link '%s' to '%s'", newarc, src);
1367                 err_major++;
1368                 return rc;
1369         }
1370         CT_TRACE("imported '%s' from '%s'=='%s'", dst, newarc, src);
1371
1372         return 0;
1373 }
1374
1375 static char *path_concat(const char *dirname, const char *basename)
1376 {
1377         char    *result;
1378         int      dirlen = strlen(dirname);
1379
1380         result = malloc(dirlen + strlen(basename) + 2);
1381         if (result == NULL)
1382                 return NULL;
1383
1384         memcpy(result, dirname, dirlen);
1385         result[dirlen] = '/';
1386         strcpy(result + dirlen + 1, basename);
1387
1388         return result;
1389 }
1390
1391 static int ct_import_recurse(const char *relpath)
1392 {
1393         DIR             *dir;
1394         struct dirent    ent, *cookie = NULL;
1395         char            *srcpath, *newpath;
1396         int              rc;
1397
1398         if (relpath == NULL)
1399                 return -EINVAL;
1400
1401         srcpath = path_concat(opt.o_hsm_root, relpath);
1402         if (srcpath == NULL) {
1403                 err_major++;
1404                 return -ENOMEM;
1405         }
1406
1407         dir = opendir(srcpath);
1408         if (dir == NULL) {
1409                 /* Not a dir, or error */
1410                 if (errno == ENOTDIR) {
1411                         /* Single regular file case, treat o_dst as absolute
1412                            final location. */
1413                         rc = ct_import_one(srcpath, opt.o_dst);
1414                 } else {
1415                         rc = -errno;
1416                         CT_ERROR(rc, "cannot opendir '%s'", srcpath);
1417                         err_major++;
1418                 }
1419                 free(srcpath);
1420                 return rc;
1421         }
1422         free(srcpath);
1423
1424         while (1) {
1425                 rc = readdir_r(dir, &ent, &cookie);
1426                 if (rc != 0) {
1427                         rc = -errno;
1428                         CT_ERROR(rc, "cannot readdir_r '%s'", relpath);
1429                         err_major++;
1430                         goto out;
1431                 } else if ((rc == 0) && (cookie == NULL)) {
1432                         /* end of directory */
1433                         break;
1434                 }
1435
1436                 if (!strcmp(ent.d_name, ".") ||
1437                     !strcmp(ent.d_name, ".."))
1438                         continue;
1439
1440                 /* New relative path */
1441                 newpath = path_concat(relpath, ent.d_name);
1442                 if (newpath == NULL) {
1443                         err_major++;
1444                         rc = -ENOMEM;
1445                         goto out;
1446                 }
1447
1448                 if (ent.d_type == DT_DIR) {
1449                         rc = ct_import_recurse(newpath);
1450                 } else {
1451                         char src[PATH_MAX];
1452                         char dst[PATH_MAX];
1453
1454                         sprintf(src, "%s/%s", opt.o_hsm_root, newpath);
1455                         sprintf(dst, "%s/%s", opt.o_dst, newpath);
1456                         /* Make the target dir in the Lustre fs */
1457                         rc = ct_mkdir_p(dst);
1458                         if (rc == 0) {
1459                                 /* Import the file */
1460                                 rc = ct_import_one(src, dst);
1461                         } else {
1462                                 CT_ERROR(rc, "ct_mkdir_p '%s' failed", dst);
1463                                 err_major++;
1464                         }
1465                 }
1466
1467                 if (rc != 0) {
1468                         CT_ERROR(rc, "cannot import '%s'", newpath);
1469                         if (err_major && opt.o_abort_on_error) {
1470                                 free(newpath);
1471                                 goto out;
1472                         }
1473                 }
1474                 free(newpath);
1475         }
1476
1477         rc = 0;
1478 out:
1479         closedir(dir);
1480         return rc;
1481 }
1482
1483 static int ct_rebind_one(const lustre_fid *old_fid, const lustre_fid *new_fid)
1484 {
1485         char    src[PATH_MAX];
1486         char    dst[PATH_MAX];
1487         int     rc;
1488
1489         CT_TRACE("rebind "DFID" to "DFID, PFID(old_fid), PFID(new_fid));
1490
1491         ct_path_archive(src, sizeof(src), opt.o_hsm_root, old_fid);
1492         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, new_fid);
1493
1494         if (!opt.o_dry_run) {
1495                 ct_mkdir_p(dst);
1496                 if (rename(src, dst)) {
1497                         rc = -errno;
1498                         CT_ERROR(rc, "cannot rename '%s' to '%s'", src, dst);
1499                         return -errno;
1500                 }
1501                 /* rename lov file */
1502                 strncat(src, ".lov", sizeof(src) - strlen(src) - 1);
1503                 strncat(dst, ".lov", sizeof(dst) - strlen(dst) - 1);
1504                 if (rename(src, dst))
1505                         CT_ERROR(errno, "cannot '%s' rename to '%s'", src, dst);
1506
1507         }
1508         return 0;
1509 }
1510
1511 static bool fid_is_file(lustre_fid *fid)
1512 {
1513         return fid_is_norm(fid) || fid_is_igif(fid);
1514 }
1515
1516 static bool should_ignore_line(const char *line)
1517 {
1518         int     i;
1519
1520         for (i = 0; line[i] != '\0'; i++) {
1521                 if (isspace(line[i]))
1522                         continue;
1523                 else if (line[i] == '#')
1524                         return true;
1525                 else
1526                         return false;
1527         }
1528
1529         return true;
1530 }
1531
1532 static int ct_rebind_list(const char *list)
1533 {
1534         int              rc;
1535         FILE            *filp;
1536         ssize_t          r;
1537         char            *line = NULL;
1538         size_t           line_size = 0;
1539         unsigned int     nl = 0;
1540         unsigned int     ok = 0;
1541
1542         filp = fopen(list, "r");
1543         if (filp == NULL) {
1544                 rc = -errno;
1545                 CT_ERROR(rc, "cannot open '%s'", list);
1546                 return rc;
1547         }
1548
1549         /* each line consists of 2 FID */
1550         while ((r = getline(&line, &line_size, filp)) != -1) {
1551                 lustre_fid      old_fid;
1552                 lustre_fid      new_fid;
1553
1554                 /* Ignore empty and commented out ('#...') lines. */
1555                 if (should_ignore_line(line))
1556                         continue;
1557
1558                 nl++;
1559
1560                 rc = sscanf(line, SFID" "SFID, RFID(&old_fid), RFID(&new_fid));
1561                 if (rc != 6 || !fid_is_file(&old_fid) ||
1562                     !fid_is_file(&new_fid)) {
1563                         CT_ERROR(EINVAL,
1564                                  "'%s' FID expected near '%s', line %u",
1565                                  list, line, nl);
1566                         err_major++;
1567                         continue;
1568                 }
1569
1570                 if (ct_rebind_one(&old_fid, &new_fid))
1571                         err_major++;
1572                 else
1573                         ok++;
1574         }
1575
1576         fclose(filp);
1577
1578         if (line)
1579                 free(line);
1580
1581         /* return 0 if all rebinds were sucessful */
1582         CT_TRACE("%u lines read from '%s', %u rebind successful", nl, list, ok);
1583
1584         return ok == nl ? 0 : -1;
1585 }
1586
1587 static int ct_rebind(void)
1588 {
1589         int     rc;
1590
1591         if (opt.o_dst) {
1592                 lustre_fid      old_fid;
1593                 lustre_fid      new_fid;
1594
1595                 if (sscanf(opt.o_src, SFID, RFID(&old_fid)) != 3 ||
1596                     !fid_is_file(&old_fid)) {
1597                         rc = -EINVAL;
1598                         CT_ERROR(rc, "'%s' invalid FID format", opt.o_src);
1599                         return rc;
1600                 }
1601
1602                 if (sscanf(opt.o_dst, SFID, RFID(&new_fid)) != 3 ||
1603                     !fid_is_file(&new_fid)) {
1604                         rc = -EINVAL;
1605                         CT_ERROR(rc, "'%s' invalid FID format", opt.o_dst);
1606                         return rc;
1607                 }
1608
1609                 rc = ct_rebind_one(&old_fid, &new_fid);
1610
1611                 return rc;
1612         }
1613
1614         /* o_src is a list file */
1615         rc = ct_rebind_list(opt.o_src);
1616
1617         return rc;
1618 }
1619
1620 static int ct_dir_level_max(const char *dirpath, __u16 *sub_seqmax)
1621 {
1622         DIR             *dir;
1623         int              rc;
1624         __u16            sub_seq;
1625         struct dirent    ent, *cookie = NULL;
1626
1627         *sub_seqmax = 0;
1628
1629         dir = opendir(dirpath);
1630         if (dir == NULL) {
1631                 rc = -errno;
1632                 CT_ERROR(rc, "cannot open directory '%s'", opt.o_hsm_root);
1633                 return rc;
1634         }
1635
1636         while ((rc = readdir_r(dir, &ent, &cookie)) == 0) {
1637                 if (cookie == NULL)
1638                         /* end of directory.
1639                          * rc is 0 and seqmax contains the max value. */
1640                         goto out;
1641
1642                 if (!strcmp(ent.d_name, ".") || !strcmp(ent.d_name, ".."))
1643                         continue;
1644
1645                 if (sscanf(ent.d_name, "%hx", &sub_seq) != 1) {
1646                         CT_TRACE("'%s' has an unexpected dirname format, "
1647                                  "skip entry", ent.d_name);
1648                         continue;
1649                 }
1650                 if (sub_seq > *sub_seqmax)
1651                         *sub_seqmax = sub_seq;
1652         }
1653         rc = -errno;
1654         CT_ERROR(rc, "cannot readdir_r '%s'", dirpath);
1655
1656 out:
1657         closedir(dir);
1658         return rc;
1659 }
1660
1661 static int ct_max_sequence(void)
1662 {
1663         int   rc, i;
1664         char  path[PATH_MAX];
1665         __u64 seq = 0;
1666         __u16 subseq;
1667
1668         strncpy(path, opt.o_hsm_root, sizeof(path));
1669         /* FID sequence is stored in top-level directory names:
1670          * hsm_root/16bits (high weight)/16 bits/16 bits/16 bits (low weight).
1671          */
1672         for (i = 0; i < 4; i++) {
1673                 rc = ct_dir_level_max(path, &subseq);
1674                 if (rc != 0)
1675                         return rc;
1676                 seq |= ((__u64)subseq << ((3 - i) * 16));
1677                 sprintf(path + strlen(path), "/%04x", subseq);
1678         }
1679
1680         printf("max_sequence: %016Lx\n", seq);
1681
1682         return 0;
1683 }
1684
1685 static void handler(int signal)
1686 {
1687         psignal(signal, "exiting");
1688         /* If we don't clean up upon interrupt, umount thinks there's a ref
1689          * and doesn't remove us from mtab (EINPROGRESS). The lustre client
1690          * does successfully unmount and the mount is actually gone, but the
1691          * mtab entry remains. So this just makes mtab happier. */
1692         llapi_hsm_copytool_unregister(&ctdata);
1693         _exit(1);
1694 }
1695
1696 /* Daemon waits for messages from the kernel; run it in the background. */
1697 static int ct_run(void)
1698 {
1699         int     rc;
1700
1701         if (opt.o_daemonize) {
1702                 rc = daemon(1, 1);
1703                 if (rc < 0) {
1704                         rc = -errno;
1705                         CT_ERROR(rc, "cannot daemonize");
1706                         return rc;
1707                 }
1708         }
1709
1710         setbuf(stdout, NULL);
1711
1712         if (opt.o_event_fifo != NULL) {
1713                 rc = llapi_hsm_register_event_fifo(opt.o_event_fifo);
1714                 if (rc < 0) {
1715                         CT_ERROR(rc, "failed to register event fifo");
1716                         return rc;
1717                 }
1718                 llapi_error_callback_set(llapi_hsm_log_error);
1719         }
1720
1721         rc = llapi_hsm_copytool_register(&ctdata, opt.o_mnt, 0,
1722                                          opt.o_archive_cnt, opt.o_archive_id);
1723         if (rc < 0) {
1724                 CT_ERROR(rc, "cannot start copytool interface");
1725                 return rc;
1726         }
1727
1728         signal(SIGINT, handler);
1729         signal(SIGTERM, handler);
1730
1731         while (1) {
1732                 struct hsm_action_list  *hal;
1733                 struct hsm_action_item  *hai;
1734                 int                      msgsize;
1735                 int                      i = 0;
1736
1737                 CT_TRACE("waiting for message from kernel");
1738
1739                 rc = llapi_hsm_copytool_recv(ctdata, &hal, &msgsize);
1740                 if (rc == -ESHUTDOWN) {
1741                         CT_TRACE("shutting down");
1742                         break;
1743                 } else if (rc == -EAGAIN) {
1744                         continue; /* msg not for us */
1745                 } else if (rc < 0) {
1746                         CT_WARN("cannot receive action list: %s",
1747                                 strerror(-rc));
1748                         err_major++;
1749                         if (opt.o_abort_on_error)
1750                                 break;
1751                         else
1752                                 continue;
1753                 }
1754
1755                 CT_TRACE("copytool fs=%s archive#=%d item_count=%d",
1756                          hal->hal_fsname, hal->hal_archive_id, hal->hal_count);
1757
1758                 if (strcmp(hal->hal_fsname, fs_name) != 0) {
1759                         CT_ERROR(EINVAL, "'%s' invalid fs name, expecting: %s",
1760                                  hal->hal_fsname, fs_name);
1761                         err_major++;
1762                         if (opt.o_abort_on_error)
1763                                 break;
1764                         else
1765                                 continue;
1766                 }
1767
1768                 hai = hai_first(hal);
1769                 while (++i <= hal->hal_count) {
1770                         if ((char *)hai - (char *)hal > msgsize) {
1771                                 rc = -EPROTO;
1772                                 CT_ERROR(rc,
1773                                          "'%s' item %d past end of message!",
1774                                          opt.o_mnt, i);
1775                                 err_major++;
1776                                 break;
1777                         }
1778                         rc = ct_process_item_async(hai, hal->hal_flags);
1779                         if (rc < 0)
1780                                 CT_ERROR(rc, "'%s' item %d process",
1781                                          opt.o_mnt, i);
1782                         if (opt.o_abort_on_error && err_major)
1783                                 break;
1784                         hai = hai_next(hai);
1785                 }
1786
1787                 llapi_hsm_action_list_free(&hal);
1788
1789                 if (opt.o_abort_on_error && err_major)
1790                         break;
1791         }
1792
1793         llapi_hsm_copytool_unregister(&ctdata);
1794         if (opt.o_event_fifo != NULL)
1795                 llapi_hsm_unregister_event_fifo(opt.o_event_fifo);
1796
1797         return rc;
1798 }
1799
1800 static int ct_setup(void)
1801 {
1802         int     rc;
1803
1804         /* set llapi message level */
1805         llapi_msg_set_level(opt.o_verbose);
1806
1807         arc_fd = open(opt.o_hsm_root, O_RDONLY);
1808         if (arc_fd < 0) {
1809                 rc = -errno;
1810                 CT_ERROR(rc, "cannot open archive at '%s'", opt.o_hsm_root);
1811                 return rc;
1812         }
1813
1814         rc = llapi_search_fsname(opt.o_mnt, fs_name);
1815         if (rc < 0) {
1816                 CT_ERROR(rc, "cannot find a Lustre filesystem mounted at '%s'",
1817                          opt.o_mnt);
1818                 return -rc;
1819         }
1820
1821         return rc;
1822 }
1823
1824 static int ct_cleanup(void)
1825 {
1826         int     rc;
1827
1828         if (arc_fd < 0)
1829                 return 0;
1830
1831         if (close(arc_fd) < 0) {
1832                 rc = -errno;
1833                 CT_ERROR(rc, "cannot close archive root directory");
1834                 return rc;
1835         }
1836
1837         return 0;
1838 }
1839
1840 int main(int argc, char **argv)
1841 {
1842         int     rc;
1843
1844         strncpy(cmd_name, basename(argv[0]), sizeof(cmd_name));
1845         rc = ct_parseopts(argc, argv);
1846         if (rc < 0) {
1847                 CT_WARN("try '%s --help' for more information", cmd_name);
1848                 return -rc;
1849         }
1850
1851         ct_setup();
1852
1853         switch (opt.o_action) {
1854         case CA_IMPORT:
1855                 rc = ct_import_recurse(opt.o_src);
1856                 break;
1857         case CA_REBIND:
1858                 rc = ct_rebind();
1859                 break;
1860         case CA_MAXSEQ:
1861                 rc = ct_max_sequence();
1862                 break;
1863         default:
1864                 rc = ct_run();
1865                 break;
1866         }
1867
1868         if (opt.o_action != CA_MAXSEQ)
1869                 CT_TRACE("process finished, errs: %d major, %d minor,"
1870                          " rc=%d (%s)", err_major, err_minor, rc,
1871                          strerror(-rc));
1872
1873         ct_cleanup();
1874
1875         return -rc;
1876 }
1877