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