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
  36 #ifndef _GNU_SOURCE
  37 #define _GNU_SOURCE
  38 #endif
  39 #include <stdio.h>
  40 #include <stdint.h>
  41 #include <stdlib.h>
  42 #include <dirent.h>
  43 #include <errno.h>
  44 #include <utime.h>
  45 #include <sys/xattr.h>
  46 #include <sys/syscall.h>
  47 #include <sys/types.h>
  48 #include <lustre/lustre_idl.h>
  49 #include <lustre/lustreapi.h>
  50
  51 /* Progress reporting period */
  52 #define REPORT_INTERVAL_DEFAULT 30
  53 /* HSM hash subdir permissions */
  54 #define DIR_PERM S_IRWXU
  55 /* HSM hash file permissions */
  56 #define FILE_PERM (S_IRUSR | S_IWUSR)
  57
  58 #define ONE_MB 0x100000
  59
  60 /* copytool uses a 32b bitmask field to register with kuc
  61  * archive num = 0 => all
  62  * archive num from 1 to 32
  63  */
  64 #define MAX_ARCHIVE_CNT (sizeof(__u32) * 8)
  65
  66 enum ct_action {
  67         CA_IMPORT = 1,
  68         CA_REBIND,
  69         CA_MAXSEQ,
  70 };
  71
  72 struct options {
  73         int                      o_copy_attrs;
  74         int                      o_daemonize;
  75         int                      o_dry_run;
  76         int                      o_abort_on_error;
  77         int                      o_shadow_tree;
  78         int                      o_verbose;
  79         int                      o_copy_xattrs;
  80         int                      o_archive_cnt;
  81         int                      o_archive_id[MAX_ARCHIVE_CNT];
  82         int                      o_report_int;
  83         unsigned long long       o_bandwidth;
  84         size_t                   o_chunk_size;
  85         enum ct_action           o_action;
  86         char                    *o_event_fifo;
  87         char                    *o_mnt;
  88         char                    *o_hsm_root;
  89         char                    *o_src; /* for import, or rebind */
  90         char                    *o_dst; /* for import, or rebind */
  91 };
  92
  93 /* everything else is zeroed */
  94 struct options opt = {
  95         .o_copy_attrs = 1,
  96         .o_shadow_tree = 1,
  97         .o_verbose = LLAPI_MSG_INFO,
  98         .o_copy_xattrs = 1,
  99         .o_report_int = REPORT_INTERVAL_DEFAULT,
 100         .o_chunk_size = ONE_MB,
 101 };
 102
 103 /* hsm_copytool_private will hold an open FD on the lustre mount point
 104  * for us. Additionally open one on the archive FS root to make sure
 105  * it doesn't drop out from under us (and remind the admin to shutdown
 106  * the copytool before unmounting). */
 107
 108 static int arc_fd = -1;
 109
 110 static int err_major;
 111 static int err_minor;
 112
 113 static char cmd_name[PATH_MAX];
 114 static char fs_name[MAX_OBD_NAME + 1];
 115
 116 static struct hsm_copytool_private *ctdata;
 117
 118
 119 #define CT_ERROR(_rc, _format, ...) \
 120         llapi_error(LLAPI_MSG_ERROR, _rc,                               \
 121                     "%s[%ld]: "_format,                                 \
 122                     cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 123 #define CT_DEBUG(_format, ...) \
 124         llapi_error(LLAPI_MSG_DEBUG | LLAPI_MSG_NO_ERRNO, 0,            \
 125                     "%s[%ld]: "_format,                                 \
 126                     cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 127 #define CT_WARN(_format, ...) \
 128         llapi_error(LLAPI_MSG_WARN | LLAPI_MSG_NO_ERRNO, 0,             \
 129                     "%s[%ld]: "_format,                                 \
 130                     cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 131 #define CT_TRACE(_format, ...) \
 132         llapi_error(LLAPI_MSG_INFO | LLAPI_MSG_NO_ERRNO, 0,             \
 133                     "%s[%ld]: "_format,                                 \
 134                     cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 135 #define CT_PRINTF(_format, ...) \
 136         llapi_printf(LLAPI_MSG_NORMAL,                                  \
 137                      "%s[%ld]: "_format,                                \
 138                      cmd_name, syscall(SYS_gettid), ## __VA_ARGS__)
 139
 140 static void usage(const char *name, int rc)
 141 {
 142         fprintf(stdout,
 143         " Usage: %s [options]... <mode> <lustre_mount_point>\n"
 144         "The Lustre HSM Posix copy tool can be used as a daemon or "
 145         "as a command line tool\n"
 146         "The Lustre HSM daemon acts on action requests from Lustre\n"
 147         "to copy files to and from an HSM archive system.\n"
 148         "This POSIX-flavored daemon makes regular POSIX filesystem calls\n"
 149         "to an HSM mounted at a given hsm_root.\n"
 150         "   --daemon            Daemon mode, run in background\n"
 151         " Options:\n"
 152         "   --no-attr           Don't copy file attributes\n"
 153         "   --no-shadow         Don't create shadow namespace in archive\n"
 154         "   --no-xattr          Don't copy file extended attributes\n"
 155         "The Lustre HSM tool performs administrator-type actions\n"
 156         "on a Lustre HSM archive.\n"
 157         "This POSIX-flavored tool can link an existing HSM namespace\n"
 158         "into a Lustre filesystem.\n"
 159         " Usage:\n"
 160         "   %s [options] --import <src> <dst> <lustre_mount_point>\n"
 161         "      import an archived subtree from\n"
 162         "       <src> (FID or relative path to hsm_root) into the Lustre\n"
 163         "             filesystem at\n"
 164         "       <dst> (absolute path)\n"
 165         "   %s [options] --rebind <old_FID> <new_FID> <lustre_mount_point>\n"
 166         "      rebind an entry in the HSM to a new FID\n"
 167         "       <old_FID> old FID the HSM entry is bound to\n"
 168         "       <new_FID> new FID to bind the HSM entry to\n"
 169         "   %s [options] --rebind <list_file> <lustre_mount_point>\n"
 170         "      perform the rebind operation for all FID in the list file\n"
 171         "       each line of <list_file> consists of <old_FID> <new_FID>\n"
 172         "   %s [options] --max-sequence <fsname>\n"
 173         "       return the max fid sequence of archived files\n"
 174         "   --abort-on-error          Abort operation on major error\n"
 175         "   -A, --archive <#>         Archive number (repeatable)\n"
 176         "   -b, --bandwidth <bw>      Limit I/O bandwidth (unit can be used\n,"
 177         "                             default is MB)\n"
 178         "   --dry-run                 Don't run, just show what would be done\n"
 179         "   -c, --chunk-size <sz>     I/O size used during data copy\n"
 180         "                             (unit can be used, default is MB)\n"
 181         "   -f, --event-fifo <path>   Write events stream to fifo\n"
 182         "   -p, --hsm-root <path>     Target HSM mount point\n"
 183         "   -q, --quiet               Produce less verbose output\n"
 184         "   -u, --update-interval <s> Interval between progress reports sent\n"
 185         "                             to Coordinator\n"
 186         "   -v, --verbose             Produce more verbose output\n",
 187         cmd_name, cmd_name, cmd_name, cmd_name, cmd_name);
 188
 189         exit(rc);
 190 }
 191
 192 static int ct_parseopts(int argc, char * const *argv)
 193 {
 194         struct option long_opts[] = {
 195                 {"abort-on-error", no_argument,       &opt.o_abort_on_error, 1},
 196                 {"abort_on_error", no_argument,       &opt.o_abort_on_error, 1},
 197                 {"archive",        required_argument, NULL,                'A'},
 198                 {"bandwidth",      required_argument, NULL,                'b'},
 199                 {"chunk-size",     required_argument, NULL,                'c'},
 200                 {"chunk_size",     required_argument, NULL,                'c'},
 201                 {"daemon",         no_argument,       &opt.o_daemonize,     1},
 202                 {"event-fifo",     required_argument, NULL,                'f'},
 203                 {"event_fifo",     required_argument, NULL,                'f'},
 204                 {"dry-run",        no_argument,       &opt.o_dry_run,       1},
 205                 {"help",           no_argument,       NULL,                'h'},
 206                 {"hsm-root",       required_argument, NULL,                'p'},
 207                 {"hsm_root",       required_argument, NULL,                'p'},
 208                 {"import",         no_argument,       NULL,                'i'},
 209                 {"max-sequence",   no_argument,       NULL,                'M'},
 210                 {"max_sequence",   no_argument,       NULL,                'M'},
 211                 {"no-attr",        no_argument,       &opt.o_copy_attrs,    0},
 212                 {"no_attr",        no_argument,       &opt.o_copy_attrs,    0},
 213                 {"no-shadow",      no_argument,       &opt.o_shadow_tree,   0},
 214                 {"no_shadow",      no_argument,       &opt.o_shadow_tree,   0},
 215                 {"no-xattr",       no_argument,       &opt.o_copy_xattrs,   0},
 216                 {"no_xattr",       no_argument,       &opt.o_copy_xattrs,   0},
 217                 {"quiet",          no_argument,       NULL,                'q'},
 218                 {"rebind",         no_argument,       NULL,                'r'},
 219                 {"update-interval", required_argument,  NULL,              'u'},
 220                 {"update_interval", required_argument,  NULL,              'u'},
 221                 {"verbose",        no_argument,       NULL,                'v'},
 222                 {0, 0, 0, 0}
 223         };
 224         int                      c, rc;
 225         unsigned long long       value;
 226         unsigned long long       unit;
 227
 228         optind = 0;
 229         while ((c = getopt_long(argc, argv, "A:b:c:f:hiMp:qru:v",
 230                                 long_opts, NULL)) != -1) {
 231                 switch (c) {
 232                 case 'A':
 233                         if ((opt.o_archive_cnt >= MAX_ARCHIVE_CNT) ||
 234                             (atoi(optarg) >= MAX_ARCHIVE_CNT)) {
 235                                 rc = -E2BIG;
 236                                 CT_ERROR(rc, "archive number must be less"
 237                                          "than %zu", MAX_ARCHIVE_CNT);
 238                                 return rc;
 239                         }
 240                         opt.o_archive_id[opt.o_archive_cnt] = atoi(optarg);
 241                         opt.o_archive_cnt++;
 242                         break;
 243                 case 'b': /* -b and -c have both a number with unit as arg */
 244                 case 'c':
 245                         unit = ONE_MB;
 246                         if (llapi_parse_size(optarg, &value, &unit, 0) < 0) {
 247                                 rc = -EINVAL;
 248                                 CT_ERROR(rc, "bad value for -%c '%s'", c,
 249                                          optarg);
 250                                 return rc;
 251                         }
 252                         if (c == 'c')
 253                                 opt.o_chunk_size = value;
 254                         else
 255                                 opt.o_bandwidth = value;
 256                         break;
 257                 case 'f':
 258                         opt.o_event_fifo = optarg;
 259                         break;
 260                 case 'h':
 261                         usage(argv[0], 0);
 262                 case 'i':
 263                         opt.o_action = CA_IMPORT;
 264                         break;
 265                 case 'M':
 266                         opt.o_action = CA_MAXSEQ;
 267                         break;
 268                 case 'p':
 269                         opt.o_hsm_root = optarg;
 270                         break;
 271                 case 'q':
 272                         opt.o_verbose--;
 273                         break;
 274                 case 'r':
 275                         opt.o_action = CA_REBIND;
 276                         break;
 277                 case 'u':
 278                         opt.o_report_int = atoi(optarg);
 279                         if (opt.o_report_int < 0) {
 280                                 rc = -EINVAL;
 281                                 CT_ERROR(rc, "bad value for -%c '%s'", c,
 282                                          optarg);
 283                                 return rc;
 284                         }
 285                         break;
 286                 case 'v':
 287                         opt.o_verbose++;
 288                         break;
 289                 case 0:
 290                         break;
 291                 default:
 292                         return -EINVAL;
 293                 }
 294         }
 295
 296         switch (opt.o_action) {
 297         case CA_IMPORT:
 298                 /* src dst mount_point */
 299                 if (argc != optind + 3) {
 300                         rc = -EINVAL;
 301                         CT_ERROR(rc, "--import requires 2 arguments");
 302                         return rc;
 303                 }
 304                 opt.o_src = argv[optind++];
 305                 opt.o_dst = argv[optind++];
 306                 break;
 307         case CA_REBIND:
 308                 /* FID1 FID2 mount_point or FILE mount_point */
 309                 if (argc == optind + 2) {
 310                         opt.o_src = argv[optind++];
 311                         opt.o_dst = NULL;
 312                 } else if (argc == optind + 3) {
 313                         opt.o_src = argv[optind++];
 314                         opt.o_dst = argv[optind++];
 315                 } else {
 316                         rc = -EINVAL;
 317                         CT_ERROR(rc, "--rebind requires 1 or 2 arguments");
 318                         return rc;
 319                 }
 320                 break;
 321         case CA_MAXSEQ:
 322         default:
 323                 /* just mount point */
 324                 break;
 325         }
 326
 327         if (argc != optind + 1) {
 328                 rc = -EINVAL;
 329                 CT_ERROR(rc, "no mount point specified");
 330                 return rc;
 331         }
 332
 333         opt.o_mnt = argv[optind];
 334
 335         CT_TRACE("action=%d src=%s dst=%s mount_point=%s",
 336                  opt.o_action, opt.o_src, opt.o_dst, opt.o_mnt);
 337
 338         if (!opt.o_dry_run && opt.o_hsm_root == NULL) {
 339                 rc = -EINVAL;
 340                 CT_ERROR(rc, "must specify a root directory for the backend");
 341                 return rc;
 342         }
 343
 344         if (opt.o_action == CA_IMPORT) {
 345                 if (opt.o_src && opt.o_src[0] == '/') {
 346                         rc = -EINVAL;
 347                         CT_ERROR(rc,
 348                                  "source path must be relative to HSM root");
 349                         return rc;
 350                 }
 351
 352                 if (opt.o_dst && opt.o_dst[0] != '/') {
 353                         rc = -EINVAL;
 354                         CT_ERROR(rc, "destination path must be absolute");
 355                         return rc;
 356                 }
 357         }
 358
 359         return 0;
 360 }
 361
 362 /* mkdir -p path */
 363 static int ct_mkdir_p(const char *path)
 364 {
 365         char    *saved, *ptr;
 366         int      rc;
 367
 368         ptr = strdup(path);
 369         if (ptr == NULL)
 370                 return -errno;
 371
 372         saved = ptr;
 373         while (*ptr == '/')
 374                 ptr++;
 375
 376         while ((ptr = strchr(ptr, '/')) != NULL) {
 377                 *ptr = '\0';
 378                 rc = mkdir(saved, DIR_PERM);
 379                 *ptr = '/';
 380                 if (rc < 0 && errno != EEXIST) {
 381                         rc = -errno;
 382                         CT_ERROR(rc, "cannot mkdir '%s'", path);
 383                         free(saved);
 384                         return rc;
 385                 }
 386                 ptr++;
 387         }
 388
 389         free(saved);
 390
 391         return 0;
 392 }
 393
 394 static int ct_save_stripe(int src_fd, const char *src, const char *dst)
 395 {
 396         char                     lov_file[PATH_MAX];
 397         char                     lov_buf[XATTR_SIZE_MAX];
 398         struct lov_user_md      *lum;
 399         int                      rc;
 400         int                      fd;
 401         ssize_t                  xattr_size;
 402
 403         snprintf(lov_file, sizeof(lov_file), "%s.lov", dst);
 404         CT_TRACE("saving stripe info of '%s' in %s", src, lov_file);
 405
 406         xattr_size = fgetxattr(src_fd, XATTR_LUSTRE_LOV, lov_buf,
 407                                sizeof(lov_buf));
 408         if (xattr_size < 0) {
 409                 rc = -errno;
 410                 CT_ERROR(rc, "cannot get stripe info on '%s'", src);
 411                 return rc;
 412         }
 413
 414         lum = (struct lov_user_md *)lov_buf;
 415
 416         if (lum->lmm_magic == LOV_USER_MAGIC_V1 ||
 417             lum->lmm_magic == LOV_USER_MAGIC_V3) {
 418                 /* Set stripe_offset to -1 so that it is not interpreted as a
 419                  * hint on restore. */
 420                 lum->lmm_stripe_offset = -1;
 421         }
 422
 423         fd = open(lov_file, O_TRUNC | O_CREAT | O_WRONLY, FILE_PERM);
 424         if (fd < 0) {
 425                 rc = -errno;
 426                 CT_ERROR(rc, "cannot open '%s'", lov_file);
 427                 goto err_cleanup;
 428         }
 429
 430         rc = write(fd, lum, xattr_size);
 431         if (rc < 0) {
 432                 rc = -errno;
 433                 CT_ERROR(rc, "cannot write %zd bytes to '%s'",
 434                          xattr_size, lov_file);
 435                 close(fd);
 436                 goto err_cleanup;
 437         }
 438
 439         rc = close(fd);
 440         if (rc < 0) {
 441                 rc = -errno;
 442                 CT_ERROR(rc, "cannot close '%s'", lov_file);
 443                 goto err_cleanup;
 444         }
 445
 446         return 0;
 447
 448 err_cleanup:
 449         unlink(lov_file);
 450
 451         return rc;
 452 }
 453
 454 static int ct_load_stripe(const char *src, void *lovea, size_t *lovea_size)
 455 {
 456         char     lov_file[PATH_MAX];
 457         int      rc;
 458         int      fd;
 459
 460         snprintf(lov_file, sizeof(lov_file), "%s.lov", src);
 461         CT_TRACE("reading stripe rules from '%s' for '%s'", lov_file, src);
 462
 463         fd = open(lov_file, O_RDONLY);
 464         if (fd < 0) {
 465                 CT_ERROR(errno, "cannot open '%s'", lov_file);
 466                 return -ENODATA;
 467         }
 468
 469         rc = read(fd, lovea, *lovea_size);
 470         if (rc < 0) {
 471                 CT_ERROR(errno, "cannot read %zu bytes from '%s'",
 472                          *lovea_size, lov_file);
 473                 close(fd);
 474                 return -ENODATA;
 475         }
 476
 477         *lovea_size = rc;
 478         close(fd);
 479
 480         return 0;
 481 }
 482
 483 static int ct_restore_stripe(const char *src, const char *dst, int dst_fd,
 484                              const void *lovea, size_t lovea_size)
 485 {
 486         int     rc;
 487
 488         rc = fsetxattr(dst_fd, XATTR_LUSTRE_LOV, lovea, lovea_size,
 489                        XATTR_CREATE);
 490         if (rc < 0) {
 491                 CT_ERROR(errno, "cannot set lov EA on '%s'", dst);
 492                 rc = -errno;
 493         }
 494
 495         return rc;
 496 }
 497
 498 static void bandwidth_ctl_delay(int wsize)
 499 {
 500         static unsigned long long       tot_bytes;
 501         static time_t                   start_time;
 502         static time_t                   last_time;
 503         time_t                          now = time(0);
 504         double                          tot_time;
 505         double                          excess;
 506         unsigned int                    sleep_time;
 507
 508         if (now > last_time + 5) {
 509                 tot_bytes = 0;
 510                 start_time = last_time = now;
 511         }
 512
 513         tot_bytes += wsize;
 514         tot_time = now - start_time;
 515         if (tot_time < 1)
 516                 tot_time = 1;
 517
 518         excess = tot_bytes - tot_time * opt.o_bandwidth;
 519         sleep_time = excess * 1000000 / opt.o_bandwidth;
 520         if ((now - start_time) % 10 == 1)
 521                 CT_TRACE("bandwith control: excess=%E sleep for %dus", excess,
 522                          sleep_time);
 523
 524         if (excess > 0)
 525                 usleep(sleep_time);
 526
 527         last_time = now;
 528 }
 529
 530 static int ct_copy_data(struct hsm_copyaction_private *hcp, const char *src,
 531                         const char *dst, int src_fd, int dst_fd,
 532                         const struct hsm_action_item *hai, long hal_flags)
 533 {
 534         struct hsm_extent        he;
 535         __u64                    offset = hai->hai_extent.offset;
 536         struct stat              src_st;
 537         struct stat              dst_st;
 538         char                    *buf = NULL;
 539         __u64                    write_total = 0;
 540         __u64                    length;
 541         time_t                   last_print_time = time(NULL);
 542         int                      rc = 0;
 543
 544         CT_TRACE("going to copy data from '%s' to '%s'", src, dst);
 545
 546         if (fstat(src_fd, &src_st) < 0) {
 547                 rc = -errno;
 548                 CT_ERROR(rc, "cannot stat '%s'", src);
 549                 return rc;
 550         }
 551
 552         if (!S_ISREG(src_st.st_mode)) {
 553                 rc = -EINVAL;
 554                 CT_ERROR(rc, "'%s' is not a regular file", src);
 555                 return rc;
 556         }
 557
 558         if (fstat(dst_fd, &dst_st) < 0) {
 559                 rc = -errno;
 560                 CT_ERROR(rc, "cannot stat '%s'", dst);
 561                 return rc;
 562         }
 563
 564         if (!S_ISREG(dst_st.st_mode)) {
 565                 rc = -EINVAL;
 566                 CT_ERROR(rc, "'%s' is not a regular file", dst);
 567                 return rc;
 568         }
 569
 570         rc = lseek(src_fd, hai->hai_extent.offset, SEEK_SET);
 571         if (rc < 0) {
 572                 rc = -errno;
 573                 CT_ERROR(rc,
 574                          "cannot seek for read to "LPU64" (len %jd) in '%s'",
 575                          hai->hai_extent.offset, (intmax_t)src_st.st_size, src);
 576                 return rc;
 577         }
 578
 579         /* Don't read beyond a given extent */
 580         length = min(hai->hai_extent.length, src_st.st_size);
 581
 582         he.offset = offset;
 583         he.length = 0;
 584         rc = llapi_hsm_action_progress(hcp, &he, length, 0);
 585         if (rc < 0) {
 586                 /* Action has been canceled or something wrong
 587                  * is happening. Stop copying data. */
 588                 CT_ERROR(rc, "progress ioctl for copy '%s'->'%s' failed",
 589                          src, dst);
 590                 goto out;
 591         }
 592
 593         errno = 0;
 594
 595         buf = malloc(opt.o_chunk_size);
 596         if (buf == NULL) {
 597                 rc = -ENOMEM;
 598                 goto out;
 599         }
 600
 601         CT_DEBUG("Going to copy "LPU64" bytes %s -> %s\n", length, src, dst);
 602
 603         while (write_total < length) {
 604                 ssize_t rsize;
 605                 ssize_t wsize;
 606                 int     chunk = (length - write_total > opt.o_chunk_size) ?
 607                                  opt.o_chunk_size : length - write_total;
 608
 609                 rsize = pread(src_fd, buf, chunk, offset);
 610                 if (rsize == 0)
 611                         /* EOF */
 612                         break;
 613
 614                 if (rsize < 0) {
 615                         rc = -errno;
 616                         CT_ERROR(rc, "cannot read from '%s'", src);
 617                         break;
 618                 }
 619
 620                 wsize = pwrite(dst_fd, buf, rsize, offset);
 621                 if (wsize < 0) {
 622                         rc = -errno;
 623                         CT_ERROR(rc, "cannot write to '%s'", dst);
 624                         break;
 625                 }
 626
 627                 write_total += wsize;
 628                 offset += wsize;
 629
 630                 if (opt.o_bandwidth != 0)
 631                         /* sleep if needed, to honor bandwidth limits */
 632                         bandwidth_ctl_delay(wsize);
 633
 634                 if (time(0) >= last_print_time + opt.o_report_int) {
 635                         last_print_time = time(0);
 636                         CT_TRACE("%%"LPU64" ", 100 * write_total / length);
 637                         he.length = write_total;
 638                         rc = llapi_hsm_action_progress(hcp, &he, length, 0);
 639                         if (rc < 0) {
 640                                 /* Action has been canceled or something wrong
 641                                  * is happening. Stop copying data. */
 642                                 CT_ERROR(rc, "progress ioctl for copy"
 643                                          " '%s'->'%s' failed", src, dst);
 644                                 goto out;
 645                         }
 646                 }
 647                 rc = 0;
 648         }
 649
 650 out:
 651         /*
 652          * truncate restored file
 653          * size is taken from the archive this is done to support
 654          * restore after a force release which leaves the file with the
 655          * wrong size (can big bigger than the new size)
 656          */
 657         if ((hai->hai_action == HSMA_RESTORE) &&
 658             (src_st.st_size < dst_st.st_size)) {
 659                 /*
 660                  * make sure the file is on disk before reporting success.
 661                  */
 662                 rc = ftruncate(dst_fd, src_st.st_size);
 663                 if (rc < 0) {
 664                         rc = -errno;
 665                         CT_ERROR(rc, "cannot truncate '%s' to size %jd",
 666                                  dst, (intmax_t)src_st.st_size);
 667                         err_major++;
 668                 }
 669         }
 670
 671         if (buf != NULL)
 672                 free(buf);
 673
 674         return rc;
 675 }
 676
 677 /* Copy file attributes from file src to file dest */
 678 static int ct_copy_attr(const char *src, const char *dst, int src_fd,
 679                         int dst_fd)
 680 {
 681         struct stat     st;
 682         struct timeval  times[2];
 683         int             rc;
 684
 685         if (fstat(src_fd, &st) < 0) {
 686                 rc = -errno;
 687                 CT_ERROR(rc, "cannot stat '%s'", src);
 688                 return rc;
 689         }
 690
 691         times[0].tv_sec = st.st_atime;
 692         times[0].tv_usec = 0;
 693         times[1].tv_sec = st.st_mtime;
 694         times[1].tv_usec = 0;
 695         if (fchmod(dst_fd, st.st_mode) < 0 ||
 696             fchown(dst_fd, st.st_uid, st.st_gid) < 0 ||
 697             futimes(dst_fd, times) < 0) {
 698                 rc = -errno;
 699                 CT_ERROR(rc, "cannot set attributes of '%s'", src);
 700                 return rc;
 701         }
 702
 703         return 0;
 704 }
 705
 706 static int ct_copy_xattr(const char *src, const char *dst, int src_fd,
 707                          int dst_fd, bool is_restore)
 708 {
 709         char     list[XATTR_LIST_MAX];
 710         char     value[XATTR_SIZE_MAX];
 711         char    *name;
 712         ssize_t  list_len;
 713         int      rc;
 714
 715         list_len = flistxattr(src_fd, list, sizeof(list));
 716         if (list_len < 0)
 717                 return -errno;
 718
 719         name = list;
 720         while (name < list + list_len) {
 721                 rc = fgetxattr(src_fd, name, value, sizeof(value));
 722                 if (rc < 0)
 723                         return -errno;
 724
 725                 /* when we restore, we do not restore lustre xattr */
 726                 if (!is_restore ||
 727                     (strncmp(XATTR_TRUSTED_PREFIX, name,
 728                              sizeof(XATTR_TRUSTED_PREFIX) - 1) != 0)) {
 729                         rc = fsetxattr(dst_fd, name, value, rc, 0);
 730                         CT_TRACE("fsetxattr of '%s' on '%s' rc=%d (%s)",
 731                                  name, dst, rc, strerror(errno));
 732                         /* lustre.* attrs aren't supported on other FS's */
 733                         if (rc < 0 && errno != EOPNOTSUPP) {
 734                                 rc = -errno;
 735                                 CT_ERROR(rc, "cannot set extended attribute"
 736                                          " '%s' of '%s'",
 737                                          name, dst);
 738                                 return rc;
 739                         }
 740                 }
 741                 name += strlen(name) + 1;
 742         }
 743
 744         return 0;
 745 }
 746
 747 static int ct_path_lustre(char *buf, int sz, const char *mnt,
 748                           const lustre_fid *fid)
 749 {
 750         return snprintf(buf, sz, "%s/%s/fid/"DFID_NOBRACE, mnt,
 751                         dot_lustre_name, PFID(fid));
 752 }
 753
 754 static int ct_path_archive(char *buf, int sz, const char *archive_dir,
 755                            const lustre_fid *fid)
 756 {
 757         return snprintf(buf, sz, "%s/%04x/%04x/%04x/%04x/%04x/%04x/"
 758                         DFID_NOBRACE, archive_dir,
 759                         (fid)->f_oid       & 0xFFFF,
 760                         (fid)->f_oid >> 16 & 0xFFFF,
 761                         (unsigned int)((fid)->f_seq       & 0xFFFF),
 762                         (unsigned int)((fid)->f_seq >> 16 & 0xFFFF),
 763                         (unsigned int)((fid)->f_seq >> 32 & 0xFFFF),
 764                         (unsigned int)((fid)->f_seq >> 48 & 0xFFFF),
 765                         PFID(fid));
 766 }
 767
 768 static bool ct_is_retryable(int err)
 769 {
 770         return err == -ETIMEDOUT;
 771 }
 772
 773 static int ct_begin_restore(struct hsm_copyaction_private **phcp,
 774                             const struct hsm_action_item *hai,
 775                             int mdt_index, int open_flags)
 776 {
 777         char     src[PATH_MAX];
 778         int      rc;
 779
 780         rc = llapi_hsm_action_begin(phcp, ctdata, hai, mdt_index, open_flags,
 781                                     false);
 782         if (rc < 0) {
 783                 ct_path_lustre(src, sizeof(src), opt.o_mnt, &hai->hai_fid);
 784                 CT_ERROR(rc, "llapi_hsm_action_begin() on '%s' failed", src);
 785         }
 786
 787         return rc;
 788 }
 789
 790 static int ct_begin(struct hsm_copyaction_private **phcp,
 791                     const struct hsm_action_item *hai)
 792 {
 793         /* Restore takes specific parameters. Call the same function w/ default
 794          * values for all other operations. */
 795         return ct_begin_restore(phcp, hai, -1, 0);
 796 }
 797
 798 static int ct_fini(struct hsm_copyaction_private **phcp,
 799                    const struct hsm_action_item *hai, int hp_flags, int ct_rc)
 800 {
 801         struct hsm_copyaction_private   *hcp;
 802         char                             lstr[PATH_MAX];
 803         int                              rc;
 804
 805         CT_TRACE("Action completed, notifying coordinator "
 806                  "cookie="LPX64", FID="DFID", hp_flags=%d err=%d",
 807                  hai->hai_cookie, PFID(&hai->hai_fid),
 808                  hp_flags, -ct_rc);
 809
 810         ct_path_lustre(lstr, sizeof(lstr), opt.o_mnt, &hai->hai_fid);
 811
 812         if (phcp == NULL || *phcp == NULL) {
 813                 rc = llapi_hsm_action_begin(&hcp, ctdata, hai, -1, 0, true);
 814                 if (rc < 0) {
 815                         CT_ERROR(rc, "llapi_hsm_action_begin() on '%s' failed",
 816                                  lstr);
 817                         return rc;
 818                 }
 819                 phcp = &hcp;
 820         }
 821
 822         rc = llapi_hsm_action_end(phcp, &hai->hai_extent, hp_flags, abs(ct_rc));
 823         if (rc == -ECANCELED)
 824                 CT_ERROR(rc, "completed action on '%s' has been canceled: "
 825                          "cookie="LPX64", FID="DFID, lstr, hai->hai_cookie,
 826                          PFID(&hai->hai_fid));
 827         else if (rc < 0)
 828                 CT_ERROR(rc, "llapi_hsm_action_end() on '%s' failed", lstr);
 829         else
 830                 CT_TRACE("llapi_hsm_action_end() on '%s' ok (rc=%d)",
 831                          lstr, rc);
 832
 833         return rc;
 834 }
 835
 836 static int ct_archive(const struct hsm_action_item *hai, const long hal_flags)
 837 {
 838         struct hsm_copyaction_private   *hcp = NULL;
 839         char                             src[PATH_MAX];
 840         char                             dst[PATH_MAX] = "";
 841         int                              rc;
 842         int                              rcf = 0;
 843         bool                             rename_needed = false;
 844         int                              hp_flags = 0;
 845         int                              open_flags;
 846         int                              src_fd = -1;
 847         int                              dst_fd = -1;
 848
 849         rc = ct_begin(&hcp, hai);
 850         if (rc < 0)
 851                 goto fini_major;
 852
 853         /* we fill archive so:
 854          * source = data FID
 855          * destination = lustre FID
 856          */
 857         ct_path_lustre(src, sizeof(src), opt.o_mnt, &hai->hai_dfid);
 858         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, &hai->hai_fid);
 859         if (hai->hai_extent.length == -1) {
 860                 /* whole file, write it to tmp location and atomically
 861                  * replace old archived file */
 862                 strlcat(dst, "_tmp", sizeof(dst));
 863                 /* we cannot rely on the same test because ct_copy_data()
 864                  * updates hai_extent.length */
 865                 rename_needed = true;
 866         }
 867
 868         CT_TRACE("archiving '%s' to '%s'", src, dst);
 869
 870         if (opt.o_dry_run) {
 871                 rc = 0;
 872                 goto fini_major;
 873         }
 874
 875         rc = ct_mkdir_p(dst);
 876         if (rc < 0) {
 877                 CT_ERROR(rc, "mkdir_p '%s' failed", dst);
 878                 goto fini_major;
 879         }
 880
 881         src_fd = llapi_hsm_action_get_fd(hcp);
 882         if (src_fd < 0) {
 883                 rc = -errno;
 884                 CT_ERROR(rc, "cannot open '%s' for read", src);
 885                 goto fini_major;
 886         }
 887
 888         open_flags = O_WRONLY | O_NOFOLLOW | O_NONBLOCK;
 889         /* If extent is specified, don't truncate an old archived copy */
 890         open_flags |= ((hai->hai_extent.length == -1) ? O_TRUNC : 0) | O_CREAT;
 891
 892         dst_fd = open(dst, open_flags, FILE_PERM);
 893         if (dst_fd == -1) {
 894                 rc = -errno;
 895                 CT_ERROR(rc, "cannot open '%s' for write", dst);
 896                 goto fini_major;
 897         }
 898
 899         /* saving stripe is not critical */
 900         rc = ct_save_stripe(src_fd, src, dst);
 901         if (rc < 0)
 902                 CT_ERROR(rc, "cannot save file striping info of '%s' in '%s'",
 903                          src, dst);
 904
 905         rc = ct_copy_data(hcp, src, dst, src_fd, dst_fd, hai, hal_flags);
 906         if (rc < 0) {
 907                 CT_ERROR(rc, "data copy failed from '%s' to '%s'", src, dst);
 908                 goto fini_major;
 909         }
 910
 911         rc = fsync(dst_fd);
 912         if (rc < 0) {
 913                 rc = -errno;
 914                 CT_ERROR(rc, "cannot flush '%s' archive file '%s'", src, dst);
 915                 goto fini_major;
 916         }
 917
 918         CT_TRACE("data archiving for '%s' to '%s' done", src, dst);
 919
 920         /* attrs will remain on the MDS; no need to copy them, except possibly
 921           for disaster recovery */
 922         if (opt.o_copy_attrs) {
 923                 rc = ct_copy_attr(src, dst, src_fd, dst_fd);
 924                 if (rc < 0) {
 925                         CT_ERROR(rc, "cannot copy attr of '%s' to '%s'",
 926                                  src, dst);
 927                         rcf = rc;
 928                 }
 929                 CT_TRACE("attr file for '%s' saved to archive '%s'",
 930                          src, dst);
 931         }
 932
 933         /* xattrs will remain on the MDS; no need to copy them, except possibly
 934          for disaster recovery */
 935         if (opt.o_copy_xattrs) {
 936                 rc = ct_copy_xattr(src, dst, src_fd, dst_fd, false);
 937                 if (rc < 0) {
 938                         CT_ERROR(rc, "cannot copy xattr of '%s' to '%s'",
 939                                  src, dst);
 940                         rcf = rcf ? rcf : rc;
 941                 }
 942                 CT_TRACE("xattr file for '%s' saved to archive '%s'",
 943                          src, dst);
 944         }
 945
 946         if (rename_needed == true) {
 947                 char     tmp_src[PATH_MAX];
 948                 char     tmp_dst[PATH_MAX];
 949
 950                 /* atomically replace old archived file */
 951                 ct_path_archive(src, sizeof(src), opt.o_hsm_root,
 952                                 &hai->hai_fid);
 953                 rc = rename(dst, src);
 954                 if (rc < 0) {
 955                         rc = -errno;
 956                         CT_ERROR(rc, "cannot rename '%s' to '%s'", dst, src);
 957                         goto fini_major;
 958                 }
 959                 /* rename lov file */
 960                 snprintf(tmp_src, sizeof(tmp_src), "%s.lov", src);
 961                 snprintf(tmp_dst, sizeof(tmp_dst), "%s.lov", dst);
 962                 rc = rename(tmp_dst, tmp_src);
 963                 if (rc < 0)
 964                         CT_ERROR(errno, "cannot rename '%s' to '%s'",
 965                                  tmp_dst, tmp_src);
 966         }
 967
 968         if (opt.o_shadow_tree) {
 969                 /* Create a namespace of softlinks that shadows the original
 970                  * Lustre namespace.  This will only be current at
 971                  * time-of-archive (won't follow renames).
 972                  * WARNING: release won't kill these links; a manual
 973                  * cleanup of dead links would be required.
 974                  */
 975                 char             buf[PATH_MAX];
 976                 long long        recno = -1;
 977                 int              linkno = 0;
 978                 char            *ptr;
 979                 int              depth = 0;
 980                 ssize_t          sz;
 981
 982                 sprintf(buf, DFID, PFID(&hai->hai_fid));
 983                 sprintf(src, "%s/shadow/", opt.o_hsm_root);
 984
 985                 ptr = opt.o_hsm_root;
 986                 while (*ptr)
 987                         (*ptr++ == '/') ? depth-- : 0;
 988
 989                 rc = llapi_fid2path(opt.o_mnt, buf, src + strlen(src),
 990                                     sizeof(src) - strlen(src), &recno, &linkno);
 991                 if (rc < 0) {
 992                         CT_ERROR(rc, "cannot get FID of '%s'", buf);
 993                         rcf = rcf ? rcf : rc;
 994                         goto fini_minor;
 995                 }
 996
 997                 /* Figure out how many parent dirs to symlink back */
 998                 ptr = src;
 999                 while (*ptr)
1000                         (*ptr++ == '/') ? depth++ : 0;
1001                 sprintf(buf, "..");
1002                 while (--depth > 1)
1003                         strcat(buf, "/..");
1004
1005                 ct_path_archive(dst, sizeof(dst), buf, &hai->hai_fid);
1006
1007                 if (ct_mkdir_p(src)) {
1008                         CT_ERROR(errno, "mkdir_p '%s' failed", src);
1009                         rcf = rcf ? rcf : -errno;
1010                         goto fini_minor;
1011                 }
1012                 /* symlink already exists ? */
1013                 sz = readlink(src, buf, sizeof(buf));
1014                 /* detect truncation */
1015                 if (sz == sizeof(buf)) {
1016                         rcf = rcf ? rcf : -E2BIG;
1017                         CT_ERROR(rcf, "readlink '%s' truncated", src);
1018                         goto fini_minor;
1019                 }
1020                 if (sz >= 0) {
1021                         buf[sz] = '\0';
1022                         if (sz == 0 || strncmp(buf, dst, sz) != 0) {
1023                                 if (unlink(src) && errno != ENOENT) {
1024                                         CT_ERROR(errno,
1025                                                  "cannot unlink symlink '%s'",
1026                                                  src);
1027                                         rcf = rcf ? rcf : -errno;
1028                                         goto fini_minor;
1029                                 /* unlink old symlink done */
1030                                 CT_TRACE("remove old symlink '%s' pointing"
1031                                          " to '%s'", src, buf);
1032                                 }
1033                         } else {
1034                                 /* symlink already ok */
1035                                 CT_TRACE("symlink '%s' already pointing"
1036                                          " to '%s'", src, dst);
1037                                 rcf = 0;
1038                                 goto fini_minor;
1039                         }
1040                 }
1041                 if (symlink(dst, src)) {
1042                         CT_ERROR(errno, "cannot symlink '%s' to '%s'",
1043                                  src, dst);
1044                         rcf = rcf ? rcf : -errno;
1045                         goto fini_minor;
1046                 }
1047                 CT_TRACE("symlink '%s' to '%s' done", src, dst);
1048         }
1049 fini_minor:
1050         if (rcf)
1051                 err_minor++;
1052         goto out;
1053
1054
1055 fini_major:
1056         err_major++;
1057
1058         unlink(dst);
1059         if (ct_is_retryable(rc))
1060                 hp_flags |= HP_FLAG_RETRY;
1061
1062         rcf = rc;
1063
1064 out:
1065         if (!(src_fd < 0))
1066                 close(src_fd);
1067
1068         if (!(dst_fd < 0))
1069                 close(dst_fd);
1070
1071         rc = ct_fini(&hcp, hai, hp_flags, rcf);
1072
1073         return rc;
1074 }
1075
1076 static int ct_restore(const struct hsm_action_item *hai, const long hal_flags)
1077 {
1078         struct hsm_copyaction_private   *hcp = NULL;
1079         char                             src[PATH_MAX];
1080         char                             dst[PATH_MAX];
1081         char                             lov_buf[XATTR_SIZE_MAX];
1082         size_t                           lov_size = sizeof(lov_buf);
1083         int                              rc;
1084         int                              hp_flags = 0;
1085         int                              src_fd = -1;
1086         int                              dst_fd = -1;
1087         int                              mdt_index = -1; /* Not implemented */
1088         int                              open_flags = 0;
1089         bool                             set_lovea;
1090         struct lu_fid                    dfid;
1091         /* we fill lustre so:
1092          * source = lustre FID in the backend
1093          * destination = data FID = volatile file
1094          */
1095
1096         /* build backend file name from released file FID */
1097         ct_path_archive(src, sizeof(src), opt.o_hsm_root, &hai->hai_fid);
1098
1099         /* restore loads and sets the LOVEA w/o interpreting it to avoid
1100          * dependency on the structure format. */
1101         rc = ct_load_stripe(src, lov_buf, &lov_size);
1102         if (rc < 0) {
1103                 CT_WARN("cannot get stripe rules for '%s' (%s), use default",
1104                         src, strerror(-rc));
1105                 set_lovea = false;
1106         } else {
1107                 open_flags |= O_LOV_DELAY_CREATE;
1108                 set_lovea = true;
1109         }
1110
1111         rc = ct_begin_restore(&hcp, hai, mdt_index, open_flags);
1112         if (rc < 0)
1113                 goto fini;
1114
1115         /* get the FID of the volatile file */
1116         rc = llapi_hsm_action_get_dfid(hcp, &dfid);
1117         if (rc < 0) {
1118                 CT_ERROR(rc, "restoring "DFID
1119                          ", cannot get FID of created volatile file",
1120                          PFID(&hai->hai_fid));
1121                 goto fini;
1122         }
1123
1124         /* build volatile "file name", for messages */
1125         snprintf(dst, sizeof(dst), "{VOLATILE}="DFID, PFID(&dfid));
1126
1127         CT_TRACE("restoring data from '%s' to '%s'", src, dst);
1128
1129         if (opt.o_dry_run) {
1130                 rc = 0;
1131                 goto fini;
1132         }
1133
1134         src_fd = open(src, O_RDONLY | O_NOATIME | O_NONBLOCK | O_NOFOLLOW);
1135         if (src_fd < 0) {
1136                 rc = -errno;
1137                 CT_ERROR(rc, "cannot open '%s' for read", src);
1138                 goto fini;
1139         }
1140
1141         dst_fd = llapi_hsm_action_get_fd(hcp);
1142
1143         if (set_lovea) {
1144                 /* the layout cannot be allocated through .fid so we have to
1145                  * restore a layout */
1146                 rc = ct_restore_stripe(src, dst, dst_fd, lov_buf, lov_size);
1147                 if (rc < 0) {
1148                         CT_ERROR(rc, "cannot restore file striping info"
1149                                  " for '%s' from '%s'", dst, src);
1150                         err_major++;
1151                         goto fini;
1152                 }
1153         }
1154
1155         rc = ct_copy_data(hcp, src, dst, src_fd, dst_fd, hai, hal_flags);
1156         if (rc < 0) {
1157                 CT_ERROR(rc, "cannot copy data from '%s' to '%s'",
1158                          src, dst);
1159                 err_major++;
1160                 if (ct_is_retryable(rc))
1161                         hp_flags |= HP_FLAG_RETRY;
1162                 goto fini;
1163         }
1164
1165         CT_TRACE("data restore from '%s' to '%s' done", src, dst);
1166
1167 fini:
1168         rc = ct_fini(&hcp, hai, hp_flags, rc);
1169
1170         /* object swaping is done by cdt at copy end, so close of volatile file
1171          * cannot be done before */
1172         if (!(src_fd < 0))
1173                 close(src_fd);
1174
1175         if (!(dst_fd < 0))
1176                 close(dst_fd);
1177
1178         return rc;
1179 }
1180
1181 static int ct_remove(const struct hsm_action_item *hai, const long hal_flags)
1182 {
1183         struct hsm_copyaction_private   *hcp = NULL;
1184         char                             dst[PATH_MAX];
1185         int                              rc;
1186
1187         rc = ct_begin(&hcp, hai);
1188         if (rc < 0)
1189                 goto fini;
1190
1191         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, &hai->hai_fid);
1192
1193         CT_TRACE("removing file '%s'", dst);
1194
1195         if (opt.o_dry_run) {
1196                 rc = 0;
1197                 goto fini;
1198         }
1199
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         strlcat(dst, ".lov", sizeof(dst));
1209         rc = unlink(dst);
1210         if (rc < 0) {
1211                 rc = -errno;
1212                 CT_ERROR(rc, "cannot unlink '%s'", dst);
1213                 err_minor++;
1214                 goto fini;
1215         }
1216
1217 fini:
1218         rc = ct_fini(&hcp, hai, 0, rc);
1219
1220         return rc;
1221 }
1222
1223 static int ct_process_item(struct hsm_action_item *hai, const long hal_flags)
1224 {
1225         int     rc = 0;
1226
1227         if (opt.o_verbose >= LLAPI_MSG_INFO || opt.o_dry_run) {
1228                 /* Print the original path */
1229                 char            fid[128];
1230                 char            path[PATH_MAX];
1231                 long long       recno = -1;
1232                 int             linkno = 0;
1233
1234                 sprintf(fid, DFID, PFID(&hai->hai_fid));
1235                 CT_TRACE("'%s' action %s reclen %d, cookie="LPX64,
1236                          fid, hsm_copytool_action2name(hai->hai_action),
1237                          hai->hai_len, hai->hai_cookie);
1238                 rc = llapi_fid2path(opt.o_mnt, fid, path,
1239                                     sizeof(path), &recno, &linkno);
1240                 if (rc < 0)
1241                         CT_ERROR(rc, "cannot get path of FID %s", fid);
1242                 else
1243                         CT_TRACE("processing file '%s'", path);
1244         }
1245
1246         switch (hai->hai_action) {
1247         /* set err_major, minor inside these functions */
1248         case HSMA_ARCHIVE:
1249                 rc = ct_archive(hai, hal_flags);
1250                 break;
1251         case HSMA_RESTORE:
1252                 rc = ct_restore(hai, hal_flags);
1253                 break;
1254         case HSMA_REMOVE:
1255                 rc = ct_remove(hai, hal_flags);
1256                 break;
1257         case HSMA_CANCEL:
1258                 CT_TRACE("cancel not implemented for file system '%s'",
1259                          opt.o_mnt);
1260                 /* Don't report progress to coordinator for this cookie:
1261                  * the copy function will get ECANCELED when reporting
1262                  * progress. */
1263                 err_minor++;
1264                 return 0;
1265                 break;
1266         default:
1267                 rc = -EINVAL;
1268                 CT_ERROR(rc, "unknown action %d, on '%s'", hai->hai_action,
1269                          opt.o_mnt);
1270                 err_minor++;
1271                 ct_fini(NULL, hai, 0, rc);
1272         }
1273
1274         return 0;
1275 }
1276
1277 struct ct_th_data {
1278         long                     hal_flags;
1279         struct hsm_action_item  *hai;
1280 };
1281
1282 static void *ct_thread(void *data)
1283 {
1284         struct ct_th_data *cttd = data;
1285         int rc;
1286
1287         rc = ct_process_item(cttd->hai, cttd->hal_flags);
1288
1289         free(cttd->hai);
1290         free(cttd);
1291         pthread_exit((void *)(intptr_t)rc);
1292 }
1293
1294 static int ct_process_item_async(const struct hsm_action_item *hai,
1295                                  long hal_flags)
1296 {
1297         pthread_attr_t           attr;
1298         pthread_t                thread;
1299         struct ct_th_data       *data;
1300         int                      rc;
1301
1302         data = malloc(sizeof(*data));
1303         if (data == NULL)
1304                 return -ENOMEM;
1305
1306         data->hai = malloc(hai->hai_len);
1307         if (data->hai == NULL) {
1308                 free(data);
1309                 return -ENOMEM;
1310         }
1311
1312         memcpy(data->hai, hai, hai->hai_len);
1313         data->hal_flags = hal_flags;
1314
1315         rc = pthread_attr_init(&attr);
1316         if (rc != 0) {
1317                 CT_ERROR(rc, "pthread_attr_init failed for '%s' service",
1318                          opt.o_mnt);
1319                 free(data->hai);
1320                 free(data);
1321                 return -rc;
1322         }
1323
1324         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
1325
1326         rc = pthread_create(&thread, &attr, ct_thread, data);
1327         if (rc != 0)
1328                 CT_ERROR(rc, "cannot create thread for '%s' service",
1329                          opt.o_mnt);
1330
1331         pthread_attr_destroy(&attr);
1332         return 0;
1333 }
1334
1335 static int ct_import_one(const char *src, const char *dst)
1336 {
1337         char            newarc[PATH_MAX];
1338         lustre_fid      fid;
1339         struct stat     st;
1340         int             rc;
1341
1342         CT_TRACE("importing '%s' from '%s'", dst, src);
1343
1344         if (stat(src, &st) < 0) {
1345                 rc = -errno;
1346                 CT_ERROR(rc, "cannot stat '%s'", src);
1347                 return rc;
1348         }
1349
1350         if (opt.o_dry_run)
1351                 return 0;
1352
1353         rc = llapi_hsm_import(dst,
1354                               opt.o_archive_cnt ? opt.o_archive_id[0] : 0,
1355                               &st, 0, 0, 0, 0, NULL, &fid);
1356         if (rc < 0) {
1357                 CT_ERROR(rc, "cannot import '%s' from '%s'", dst, src);
1358                 return rc;
1359         }
1360
1361         ct_path_archive(newarc, sizeof(newarc), opt.o_hsm_root, &fid);
1362
1363         rc = ct_mkdir_p(newarc);
1364         if (rc < 0) {
1365                 CT_ERROR(rc, "mkdir_p '%s' failed", newarc);
1366                 err_major++;
1367                 return rc;
1368
1369         }
1370
1371         /* Lots of choices now: mv, ln, ln -s ? */
1372         rc = link(src, newarc); /* hardlink */
1373         if (rc < 0) {
1374                 rc = -errno;
1375                 CT_ERROR(rc, "cannot link '%s' to '%s'", newarc, src);
1376                 err_major++;
1377                 return rc;
1378         }
1379         CT_TRACE("imported '%s' from '%s'=='%s'", dst, newarc, src);
1380
1381         return 0;
1382 }
1383
1384 static char *path_concat(const char *dirname, const char *basename)
1385 {
1386         char    *result;
1387         int      rc;
1388
1389         rc = asprintf(&result, "%s/%s", dirname, basename);
1390         if (rc < 0)
1391                 return NULL;
1392
1393         return result;
1394 }
1395
1396 static int ct_import_fid(const lustre_fid *import_fid)
1397 {
1398         char    fid_path[PATH_MAX];
1399         int     rc;
1400
1401         ct_path_lustre(fid_path, sizeof(fid_path), opt.o_mnt, import_fid);
1402         rc = access(fid_path, F_OK);
1403         if (rc == 0 || errno != ENOENT) {
1404                 rc = (errno == 0) ? -EEXIST : -errno;
1405                 CT_ERROR(rc, "cannot import '"DFID"'", PFID(import_fid));
1406                 return rc;
1407         }
1408
1409         ct_path_archive(fid_path, sizeof(fid_path), opt.o_hsm_root,
1410                         import_fid);
1411
1412         CT_TRACE("Resolving "DFID" to %s", PFID(import_fid), fid_path);
1413
1414         return ct_import_one(fid_path, opt.o_dst);
1415 }
1416
1417 static int ct_import_recurse(const char *relpath)
1418 {
1419         DIR             *dir;
1420         struct dirent    ent, *cookie = NULL;
1421         char            *srcpath, *newpath;
1422         lustre_fid       import_fid;
1423         int              rc;
1424
1425         if (relpath == NULL)
1426                 return -EINVAL;
1427
1428         /* Is relpath a FID? In which case SFID should expand to three
1429          * elements. */
1430         rc = sscanf(relpath, SFID, RFID(&import_fid));
1431         if (rc == 3)
1432                 return ct_import_fid(&import_fid);
1433
1434         srcpath = path_concat(opt.o_hsm_root, relpath);
1435         if (srcpath == NULL) {
1436                 err_major++;
1437                 return -ENOMEM;
1438         }
1439
1440         dir = opendir(srcpath);
1441         if (dir == NULL) {
1442                 /* Not a dir, or error */
1443                 if (errno == ENOTDIR) {
1444                         /* Single regular file case, treat o_dst as absolute
1445                            final location. */
1446                         rc = ct_import_one(srcpath, opt.o_dst);
1447                 } else {
1448                         rc = -errno;
1449                         CT_ERROR(rc, "cannot opendir '%s'", srcpath);
1450                         err_major++;
1451                 }
1452                 free(srcpath);
1453                 return rc;
1454         }
1455         free(srcpath);
1456
1457         while (1) {
1458                 rc = readdir_r(dir, &ent, &cookie);
1459                 if (rc != 0) {
1460                         rc = -errno;
1461                         CT_ERROR(rc, "cannot readdir_r '%s'", relpath);
1462                         err_major++;
1463                         goto out;
1464                 } else if ((rc == 0) && (cookie == NULL)) {
1465                         /* end of directory */
1466                         break;
1467                 }
1468
1469                 if (!strcmp(ent.d_name, ".") ||
1470                     !strcmp(ent.d_name, ".."))
1471                         continue;
1472
1473                 /* New relative path */
1474                 newpath = path_concat(relpath, ent.d_name);
1475                 if (newpath == NULL) {
1476                         err_major++;
1477                         rc = -ENOMEM;
1478                         goto out;
1479                 }
1480
1481                 if (ent.d_type == DT_DIR) {
1482                         rc = ct_import_recurse(newpath);
1483                 } else {
1484                         char src[PATH_MAX];
1485                         char dst[PATH_MAX];
1486
1487                         sprintf(src, "%s/%s", opt.o_hsm_root, newpath);
1488                         sprintf(dst, "%s/%s", opt.o_dst, newpath);
1489                         /* Make the target dir in the Lustre fs */
1490                         rc = ct_mkdir_p(dst);
1491                         if (rc == 0) {
1492                                 /* Import the file */
1493                                 rc = ct_import_one(src, dst);
1494                         } else {
1495                                 CT_ERROR(rc, "ct_mkdir_p '%s' failed", dst);
1496                                 err_major++;
1497                         }
1498                 }
1499
1500                 if (rc != 0) {
1501                         CT_ERROR(rc, "cannot import '%s'", newpath);
1502                         if (err_major && opt.o_abort_on_error) {
1503                                 free(newpath);
1504                                 goto out;
1505                         }
1506                 }
1507                 free(newpath);
1508         }
1509
1510         rc = 0;
1511 out:
1512         closedir(dir);
1513         return rc;
1514 }
1515
1516 static int ct_rebind_one(const lustre_fid *old_fid, const lustre_fid *new_fid)
1517 {
1518         char    src[PATH_MAX];
1519         char    dst[PATH_MAX];
1520         int     rc;
1521
1522         CT_TRACE("rebind "DFID" to "DFID, PFID(old_fid), PFID(new_fid));
1523
1524         ct_path_archive(src, sizeof(src), opt.o_hsm_root, old_fid);
1525         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, new_fid);
1526
1527         if (!opt.o_dry_run) {
1528                 ct_mkdir_p(dst);
1529                 if (rename(src, dst)) {
1530                         rc = -errno;
1531                         CT_ERROR(rc, "cannot rename '%s' to '%s'", src, dst);
1532                         return -errno;
1533                 }
1534                 /* rename lov file */
1535                 strlcat(src, ".lov", sizeof(src));
1536                 strlcat(dst, ".lov", sizeof(dst));
1537                 if (rename(src, dst))
1538                         CT_ERROR(errno, "cannot rename '%s' to '%s'", src, dst);
1539
1540         }
1541         return 0;
1542 }
1543
1544 static bool fid_is_file(lustre_fid *fid)
1545 {
1546         return fid_is_norm(fid) || fid_is_igif(fid);
1547 }
1548
1549 static bool should_ignore_line(const char *line)
1550 {
1551         int     i;
1552
1553         for (i = 0; line[i] != '\0'; i++) {
1554                 if (isspace(line[i]))
1555                         continue;
1556                 else if (line[i] == '#')
1557                         return true;
1558                 else
1559                         return false;
1560         }
1561
1562         return true;
1563 }
1564
1565 static int ct_rebind_list(const char *list)
1566 {
1567         int              rc;
1568         FILE            *filp;
1569         ssize_t          r;
1570         char            *line = NULL;
1571         size_t           line_size = 0;
1572         unsigned int     nl = 0;
1573         unsigned int     ok = 0;
1574
1575         filp = fopen(list, "r");
1576         if (filp == NULL) {
1577                 rc = -errno;
1578                 CT_ERROR(rc, "cannot open '%s'", list);
1579                 return rc;
1580         }
1581
1582         /* each line consists of 2 FID */
1583         while ((r = getline(&line, &line_size, filp)) != -1) {
1584                 lustre_fid      old_fid;
1585                 lustre_fid      new_fid;
1586
1587                 /* Ignore empty and commented out ('#...') lines. */
1588                 if (should_ignore_line(line))
1589                         continue;
1590
1591                 nl++;
1592
1593                 rc = sscanf(line, SFID" "SFID, RFID(&old_fid), RFID(&new_fid));
1594                 if (rc != 6 || !fid_is_file(&old_fid) ||
1595                     !fid_is_file(&new_fid)) {
1596                         CT_ERROR(EINVAL,
1597                                  "'%s' FID expected near '%s', line %u",
1598                                  list, line, nl);
1599                         err_major++;
1600                         continue;
1601                 }
1602
1603                 if (ct_rebind_one(&old_fid, &new_fid))
1604                         err_major++;
1605                 else
1606                         ok++;
1607         }
1608
1609         fclose(filp);
1610
1611         if (line)
1612                 free(line);
1613
1614         /* return 0 if all rebinds were sucessful */
1615         CT_TRACE("%u lines read from '%s', %u rebind successful", nl, list, ok);
1616
1617         return ok == nl ? 0 : -1;
1618 }
1619
1620 static int ct_rebind(void)
1621 {
1622         int     rc;
1623
1624         if (opt.o_dst) {
1625                 lustre_fid      old_fid;
1626                 lustre_fid      new_fid;
1627
1628                 if (sscanf(opt.o_src, SFID, RFID(&old_fid)) != 3 ||
1629                     !fid_is_file(&old_fid)) {
1630                         rc = -EINVAL;
1631                         CT_ERROR(rc, "'%s' invalid FID format", opt.o_src);
1632                         return rc;
1633                 }
1634
1635                 if (sscanf(opt.o_dst, SFID, RFID(&new_fid)) != 3 ||
1636                     !fid_is_file(&new_fid)) {
1637                         rc = -EINVAL;
1638                         CT_ERROR(rc, "'%s' invalid FID format", opt.o_dst);
1639                         return rc;
1640                 }
1641
1642                 rc = ct_rebind_one(&old_fid, &new_fid);
1643
1644                 return rc;
1645         }
1646
1647         /* o_src is a list file */
1648         rc = ct_rebind_list(opt.o_src);
1649
1650         return rc;
1651 }
1652
1653 static int ct_dir_level_max(const char *dirpath, __u16 *sub_seqmax)
1654 {
1655         DIR             *dir;
1656         int              rc;
1657         __u16            sub_seq;
1658         struct dirent    ent, *cookie = NULL;
1659
1660         *sub_seqmax = 0;
1661
1662         dir = opendir(dirpath);
1663         if (dir == NULL) {
1664                 rc = -errno;
1665                 CT_ERROR(rc, "cannot open directory '%s'", opt.o_hsm_root);
1666                 return rc;
1667         }
1668
1669         while ((rc = readdir_r(dir, &ent, &cookie)) == 0) {
1670                 if (cookie == NULL)
1671                         /* end of directory.
1672                          * rc is 0 and seqmax contains the max value. */
1673                         goto out;
1674
1675                 if (!strcmp(ent.d_name, ".") || !strcmp(ent.d_name, ".."))
1676                         continue;
1677
1678                 if (sscanf(ent.d_name, "%hx", &sub_seq) != 1) {
1679                         CT_TRACE("'%s' has an unexpected dirname format, "
1680                                  "skip entry", ent.d_name);
1681                         continue;
1682                 }
1683                 if (sub_seq > *sub_seqmax)
1684                         *sub_seqmax = sub_seq;
1685         }
1686         rc = -errno;
1687         CT_ERROR(rc, "cannot readdir_r '%s'", dirpath);
1688
1689 out:
1690         closedir(dir);
1691         return rc;
1692 }
1693
1694 static int ct_max_sequence(void)
1695 {
1696         int     rc, i;
1697         char    path[PATH_MAX];
1698         __u64   seq = 0;
1699         __u16   subseq;
1700
1701         strlcpy(path, opt.o_hsm_root, sizeof(path));
1702         /* FID sequence is stored in top-level directory names:
1703          * hsm_root/16bits (high weight)/16 bits/16 bits/16 bits (low weight).
1704          */
1705         for (i = 0; i < 4; i++) {
1706                 size_t path_len;
1707
1708                 rc = ct_dir_level_max(path, &subseq);
1709                 if (rc != 0)
1710                         return rc;
1711                 seq |= ((__u64)subseq << ((3 - i) * 16));
1712                 path_len = strlen(path);
1713                 rc = snprintf(path + path_len, sizeof(path) - path_len,
1714                               "/%04x", subseq);
1715                 if (rc >= (sizeof(path) - path_len))
1716                         return -E2BIG;
1717                 path[sizeof(path) - 1] = '\0';
1718         }
1719
1720         printf("max_sequence: "LPX64"\n", seq);
1721
1722         return 0;
1723 }
1724
1725 static void handler(int signal)
1726 {
1727         psignal(signal, "exiting");
1728         /* If we don't clean up upon interrupt, umount thinks there's a ref
1729          * and doesn't remove us from mtab (EINPROGRESS). The lustre client
1730          * does successfully unmount and the mount is actually gone, but the
1731          * mtab entry remains. So this just makes mtab happier. */
1732         llapi_hsm_copytool_unregister(&ctdata);
1733         _exit(1);
1734 }
1735
1736 /* Daemon waits for messages from the kernel; run it in the background. */
1737 static int ct_run(void)
1738 {
1739         int     rc;
1740
1741         if (opt.o_daemonize) {
1742                 rc = daemon(1, 1);
1743                 if (rc < 0) {
1744                         rc = -errno;
1745                         CT_ERROR(rc, "cannot daemonize");
1746                         return rc;
1747                 }
1748         }
1749
1750         setbuf(stdout, NULL);
1751
1752         if (opt.o_event_fifo != NULL) {
1753                 rc = llapi_hsm_register_event_fifo(opt.o_event_fifo);
1754                 if (rc < 0) {
1755                         CT_ERROR(rc, "failed to register event fifo");
1756                         return rc;
1757                 }
1758                 llapi_error_callback_set(llapi_hsm_log_error);
1759         }
1760
1761         rc = llapi_hsm_copytool_register(&ctdata, opt.o_mnt, 0,
1762                                          opt.o_archive_cnt, opt.o_archive_id);
1763         if (rc < 0) {
1764                 CT_ERROR(rc, "cannot start copytool interface");
1765                 return rc;
1766         }
1767
1768         signal(SIGINT, handler);
1769         signal(SIGTERM, handler);
1770
1771         while (1) {
1772                 struct hsm_action_list  *hal;
1773                 struct hsm_action_item  *hai;
1774                 int                      msgsize;
1775                 int                      i = 0;
1776
1777                 CT_TRACE("waiting for message from kernel");
1778
1779                 rc = llapi_hsm_copytool_recv(ctdata, &hal, &msgsize);
1780                 if (rc == -ESHUTDOWN) {
1781                         CT_TRACE("shutting down");
1782                         break;
1783                 } else if (rc == -EAGAIN) {
1784                         continue; /* msg not for us */
1785                 } else if (rc < 0) {
1786                         CT_WARN("cannot receive action list: %s",
1787                                 strerror(-rc));
1788                         err_major++;
1789                         if (opt.o_abort_on_error)
1790                                 break;
1791                         else
1792                                 continue;
1793                 }
1794
1795                 CT_TRACE("copytool fs=%s archive#=%d item_count=%d",
1796                          hal->hal_fsname, hal->hal_archive_id, hal->hal_count);
1797
1798                 if (strcmp(hal->hal_fsname, fs_name) != 0) {
1799                         rc = -EINVAL;
1800                         CT_ERROR(rc, "'%s' invalid fs name, expecting: %s",
1801                                  hal->hal_fsname, fs_name);
1802                         err_major++;
1803                         if (opt.o_abort_on_error)
1804                                 break;
1805                         else
1806                                 continue;
1807                 }
1808
1809                 hai = hai_first(hal);
1810                 while (++i <= hal->hal_count) {
1811                         if ((char *)hai - (char *)hal > msgsize) {
1812                                 rc = -EPROTO;
1813                                 CT_ERROR(rc,
1814                                          "'%s' item %d past end of message!",
1815                                          opt.o_mnt, i);
1816                                 err_major++;
1817                                 break;
1818                         }
1819                         rc = ct_process_item_async(hai, hal->hal_flags);
1820                         if (rc < 0)
1821                                 CT_ERROR(rc, "'%s' item %d process",
1822                                          opt.o_mnt, i);
1823                         if (opt.o_abort_on_error && err_major)
1824                                 break;
1825                         hai = hai_next(hai);
1826                 }
1827
1828                 if (opt.o_abort_on_error && err_major)
1829                         break;
1830         }
1831
1832         llapi_hsm_copytool_unregister(&ctdata);
1833         if (opt.o_event_fifo != NULL)
1834                 llapi_hsm_unregister_event_fifo(opt.o_event_fifo);
1835
1836         return rc;
1837 }
1838
1839 static int ct_setup(void)
1840 {
1841         int     rc;
1842
1843         /* set llapi message level */
1844         llapi_msg_set_level(opt.o_verbose);
1845
1846         arc_fd = open(opt.o_hsm_root, O_RDONLY);
1847         if (arc_fd < 0) {
1848                 rc = -errno;
1849                 CT_ERROR(rc, "cannot open archive at '%s'", opt.o_hsm_root);
1850                 return rc;
1851         }
1852
1853         rc = llapi_search_fsname(opt.o_mnt, fs_name);
1854         if (rc < 0) {
1855                 CT_ERROR(rc, "cannot find a Lustre filesystem mounted at '%s'",
1856                          opt.o_mnt);
1857                 return rc;
1858         }
1859
1860         return rc;
1861 }
1862
1863 static int ct_cleanup(void)
1864 {
1865         int     rc;
1866
1867         if (arc_fd < 0)
1868                 return 0;
1869
1870         if (close(arc_fd) < 0) {
1871                 rc = -errno;
1872                 CT_ERROR(rc, "cannot close archive root directory");
1873                 return rc;
1874         }
1875
1876         return 0;
1877 }
1878
1879 int main(int argc, char **argv)
1880 {
1881         int     rc;
1882
1883         strlcpy(cmd_name, basename(argv[0]), sizeof(cmd_name));
1884         rc = ct_parseopts(argc, argv);
1885         if (rc < 0) {
1886                 CT_WARN("try '%s --help' for more information", cmd_name);
1887                 return -rc;
1888         }
1889
1890         rc = ct_setup();
1891         if (rc < 0)
1892                 goto error_cleanup;
1893
1894         switch (opt.o_action) {
1895         case CA_IMPORT:
1896                 rc = ct_import_recurse(opt.o_src);
1897                 break;
1898         case CA_REBIND:
1899                 rc = ct_rebind();
1900                 break;
1901         case CA_MAXSEQ:
1902                 rc = ct_max_sequence();
1903                 break;
1904         default:
1905                 rc = ct_run();
1906                 break;
1907         }
1908
1909         if (opt.o_action != CA_MAXSEQ)
1910                 CT_TRACE("process finished, errs: %d major, %d minor,"
1911                          " rc=%d (%s)", err_major, err_minor, rc,
1912                          strerror(-rc));
1913
1914 error_cleanup:
1915         ct_cleanup();
1916
1917         return -rc;
1918 }
1919