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         saved = ptr;
 370         while (*ptr == '/')
 371                 ptr++;
 372
 373         while ((ptr = strchr(ptr, '/')) != NULL) {
 374                 *ptr = '\0';
 375                 rc = mkdir(saved, DIR_PERM);
 376                 *ptr = '/';
 377                 if (rc < 0 && errno != EEXIST) {
 378                         rc = -errno;
 379                         CT_ERROR(rc, "cannot mkdir '%s'", path);
 380                         free(saved);
 381                         return rc;
 382                 }
 383                 ptr++;
 384         }
 385
 386         free(saved);
 387
 388         return 0;
 389 }
 390
 391 static int ct_save_stripe(int src_fd, const char *src, const char *dst)
 392 {
 393         char                     lov_file[PATH_MAX];
 394         char                     lov_buf[XATTR_SIZE_MAX];
 395         struct lov_user_md      *lum;
 396         int                      rc;
 397         int                      fd;
 398         ssize_t                  xattr_size;
 399
 400         snprintf(lov_file, sizeof(lov_file), "%s.lov", dst);
 401         CT_TRACE("saving stripe info of '%s' in %s", src, lov_file);
 402
 403         xattr_size = fgetxattr(src_fd, XATTR_LUSTRE_LOV, lov_buf,
 404                                sizeof(lov_buf));
 405         if (xattr_size < 0) {
 406                 rc = -errno;
 407                 CT_ERROR(rc, "cannot get stripe info on '%s'", src);
 408                 return rc;
 409         }
 410
 411         lum = (struct lov_user_md *)lov_buf;
 412
 413         if (lum->lmm_magic == LOV_USER_MAGIC_V1 ||
 414             lum->lmm_magic == LOV_USER_MAGIC_V3) {
 415                 /* Set stripe_offset to -1 so that it is not interpreted as a
 416                  * hint on restore. */
 417                 lum->lmm_stripe_offset = -1;
 418         }
 419
 420         fd = open(lov_file, O_TRUNC | O_CREAT | O_WRONLY, FILE_PERM);
 421         if (fd < 0) {
 422                 rc = -errno;
 423                 CT_ERROR(rc, "cannot open '%s'", lov_file);
 424                 goto err_cleanup;
 425         }
 426
 427         rc = write(fd, lum, xattr_size);
 428         if (rc < 0) {
 429                 rc = -errno;
 430                 CT_ERROR(rc, "cannot write %zd bytes to '%s'",
 431                          xattr_size, lov_file);
 432                 close(fd);
 433                 goto err_cleanup;
 434         }
 435
 436         rc = close(fd);
 437         if (rc < 0) {
 438                 rc = -errno;
 439                 CT_ERROR(rc, "cannot close '%s'", lov_file);
 440                 goto err_cleanup;
 441         }
 442
 443         return 0;
 444
 445 err_cleanup:
 446         unlink(lov_file);
 447
 448         return rc;
 449 }
 450
 451 static int ct_load_stripe(const char *src, void *lovea, size_t *lovea_size)
 452 {
 453         char     lov_file[PATH_MAX];
 454         int      rc;
 455         int      fd;
 456
 457         snprintf(lov_file, sizeof(lov_file), "%s.lov", src);
 458         CT_TRACE("reading stripe rules from '%s' for '%s'", lov_file, src);
 459
 460         fd = open(lov_file, O_RDONLY);
 461         if (fd < 0) {
 462                 CT_ERROR(errno, "cannot open '%s'", lov_file);
 463                 return -ENODATA;
 464         }
 465
 466         rc = read(fd, lovea, *lovea_size);
 467         if (rc < 0) {
 468                 CT_ERROR(errno, "cannot read %zu bytes from '%s'",
 469                          *lovea_size, lov_file);
 470                 close(fd);
 471                 return -ENODATA;
 472         }
 473
 474         *lovea_size = rc;
 475         close(fd);
 476
 477         return 0;
 478 }
 479
 480 static int ct_restore_stripe(const char *src, const char *dst, int dst_fd,
 481                              const void *lovea, size_t lovea_size)
 482 {
 483         int     rc;
 484
 485         rc = fsetxattr(dst_fd, XATTR_LUSTRE_LOV, lovea, lovea_size,
 486                        XATTR_CREATE);
 487         if (rc < 0) {
 488                 rc = -errno;
 489                 CT_ERROR(rc, "cannot set lov EA on '%s'", dst);
 490         }
 491
 492         return rc;
 493 }
 494
 495 static void bandwidth_ctl_delay(int wsize)
 496 {
 497         static unsigned long long       tot_bytes;
 498         static time_t                   start_time;
 499         static time_t                   last_time;
 500         time_t                          now = time(0);
 501         double                          tot_time;
 502         double                          excess;
 503         unsigned int                    sleep_time;
 504
 505         if (now > last_time + 5) {
 506                 tot_bytes = 0;
 507                 start_time = last_time = now;
 508         }
 509
 510         tot_bytes += wsize;
 511         tot_time = now - start_time;
 512         if (tot_time < 1)
 513                 tot_time = 1;
 514
 515         excess = tot_bytes - tot_time * opt.o_bandwidth;
 516         sleep_time = excess * 1000000 / opt.o_bandwidth;
 517         if ((now - start_time) % 10 == 1)
 518                 CT_TRACE("bandwith control: excess=%E sleep for %dus", excess,
 519                          sleep_time);
 520
 521         if (excess > 0)
 522                 usleep(sleep_time);
 523
 524         last_time = now;
 525 }
 526
 527 static int ct_copy_data(struct hsm_copyaction_private *hcp, const char *src,
 528                         const char *dst, int src_fd, int dst_fd,
 529                         const struct hsm_action_item *hai, long hal_flags)
 530 {
 531         struct hsm_extent        he;
 532         __u64                    offset = hai->hai_extent.offset;
 533         struct stat              src_st;
 534         struct stat              dst_st;
 535         char                    *buf;
 536         __u64                    write_total = 0;
 537         __u64                    length;
 538         time_t                   last_print_time = time(NULL);
 539         int                      rc = 0;
 540
 541         CT_TRACE("going to copy data from '%s' to '%s'", src, dst);
 542
 543         buf = malloc(opt.o_chunk_size);
 544         if (buf == NULL)
 545                 return -ENOMEM;
 546
 547         if (fstat(src_fd, &src_st) < 0) {
 548                 rc = -errno;
 549                 CT_ERROR(rc, "cannot stat '%s'", src);
 550                 return rc;
 551         }
 552
 553         if (!S_ISREG(src_st.st_mode)) {
 554                 rc = -EINVAL;
 555                 CT_ERROR(rc, "'%s' is not a regular file", src);
 556                 return rc;
 557         }
 558
 559         if (fstat(dst_fd, &dst_st) < 0) {
 560                 rc = -errno;
 561                 CT_ERROR(rc, "cannot stat '%s'", dst);
 562                 return rc;
 563         }
 564
 565         if (!S_ISREG(dst_st.st_mode)) {
 566                 rc = -EINVAL;
 567                 CT_ERROR(rc, "'%s' is not a regular file", dst);
 568                 return rc;
 569         }
 570
 571         rc = lseek(src_fd, hai->hai_extent.offset, SEEK_SET);
 572         if (rc < 0) {
 573                 CT_ERROR(errno,
 574                          "cannot seek for read to "LPU64" (len %jd) in '%s'",
 575                          hai->hai_extent.offset, (intmax_t)src_st.st_size, src);
 576                 rc = -errno;
 577                 goto out;
 578         }
 579
 580         /* Don't read beyond a given extent */
 581         length = min(hai->hai_extent.length, src_st.st_size);
 582
 583         he.offset = offset;
 584         he.length = 0;
 585         rc = llapi_hsm_action_progress(hcp, &he, length, 0);
 586         if (rc < 0) {
 587                 /* Action has been canceled or something wrong
 588                  * is happening. Stop copying data. */
 589                 CT_ERROR(rc, "progress ioctl for copy '%s'->'%s' failed",
 590                          src, dst);
 591                 goto out;
 592         }
 593
 594         errno = 0;
 595
 596         CT_DEBUG("Going to copy "LPU64" bytes %s -> %s\n", length, src, dst);
 597
 598         while (write_total < length) {
 599                 ssize_t rsize;
 600                 ssize_t wsize;
 601                 int     chunk = (length - write_total > opt.o_chunk_size) ?
 602                                  opt.o_chunk_size : length - write_total;
 603
 604                 rsize = pread(src_fd, buf, chunk, offset);
 605                 if (rsize == 0)
 606                         /* EOF */
 607                         break;
 608
 609                 if (rsize < 0) {
 610                         rc = -errno;
 611                         CT_ERROR(rc, "cannot read from '%s'", src);
 612                         break;
 613                 }
 614
 615                 wsize = pwrite(dst_fd, buf, rsize, offset);
 616                 if (wsize < 0) {
 617                         rc = -errno;
 618                         CT_ERROR(rc, "cannot write to '%s'", dst);
 619                         break;
 620                 }
 621
 622                 write_total += wsize;
 623                 offset += wsize;
 624
 625                 if (opt.o_bandwidth != 0)
 626                         /* sleep if needed, to honor bandwidth limits */
 627                         bandwidth_ctl_delay(wsize);
 628
 629                 if (time(0) >= last_print_time + opt.o_report_int) {
 630                         last_print_time = time(0);
 631                         CT_TRACE("%%"LPU64" ", 100 * write_total / length);
 632                         he.length = write_total;
 633                         rc = llapi_hsm_action_progress(hcp, &he, length, 0);
 634                         if (rc < 0) {
 635                                 /* Action has been canceled or something wrong
 636                                  * is happening. Stop copying data. */
 637                                 CT_ERROR(rc, "progress ioctl for copy"
 638                                          " '%s'->'%s' failed", src, dst);
 639                                 goto out;
 640                         }
 641                 }
 642                 rc = 0;
 643         }
 644
 645 out:
 646         /*
 647          * truncate restored file
 648          * size is taken from the archive this is done to support
 649          * restore after a force release which leaves the file with the
 650          * wrong size (can big bigger than the new size)
 651          */
 652         if ((hai->hai_action == HSMA_RESTORE) &&
 653             (src_st.st_size < dst_st.st_size)) {
 654                 /*
 655                  * make sure the file is on disk before reporting success.
 656                  */
 657                 rc = ftruncate(dst_fd, src_st.st_size);
 658                 if (rc < 0) {
 659                         rc = -errno;
 660                         CT_ERROR(rc,
 661                                  "cannot truncate '%s' to size %jd",
 662                                  dst, (intmax_t)src_st.st_size);
 663                         err_major++;
 664                 }
 665         }
 666
 667         free(buf);
 668
 669         return rc;
 670 }
 671
 672 /* Copy file attributes from file src to file dest */
 673 static int ct_copy_attr(const char *src, const char *dst, int src_fd,
 674                         int dst_fd)
 675 {
 676         struct stat     st;
 677         struct timeval  times[2];
 678         int             rc;
 679
 680         if (fstat(src_fd, &st) < 0) {
 681                 rc = -errno;
 682                 CT_ERROR(rc, "cannot stat '%s'", src);
 683                 return rc;
 684         }
 685
 686         times[0].tv_sec = st.st_atime;
 687         times[0].tv_usec = 0;
 688         times[1].tv_sec = st.st_mtime;
 689         times[1].tv_usec = 0;
 690         if (fchmod(dst_fd, st.st_mode) < 0 ||
 691             fchown(dst_fd, st.st_uid, st.st_gid) < 0 ||
 692             futimes(dst_fd, times) < 0) {
 693                 rc = -errno;
 694                 CT_ERROR(rc, "cannot set attributes of '%s'", src);
 695                 return rc;
 696         }
 697
 698         return 0;
 699 }
 700
 701 static int ct_copy_xattr(const char *src, const char *dst, int src_fd,
 702                          int dst_fd, bool is_restore)
 703 {
 704         char     list[XATTR_LIST_MAX];
 705         char     value[XATTR_SIZE_MAX];
 706         char    *name;
 707         ssize_t  list_len;
 708         int      rc;
 709
 710         list_len = flistxattr(src_fd, list, sizeof(list));
 711         if (list_len < 0)
 712                 return -errno;
 713
 714         name = list;
 715         while (name < list + list_len) {
 716                 rc = fgetxattr(src_fd, name, value, sizeof(value));
 717                 if (rc < 0)
 718                         return -errno;
 719
 720                 /* when we restore, we do not restore lustre xattr */
 721                 if (!is_restore ||
 722                     (strncmp(XATTR_TRUSTED_PREFIX, name,
 723                              sizeof(XATTR_TRUSTED_PREFIX) - 1) != 0)) {
 724                         rc = fsetxattr(dst_fd, name, value, rc, 0);
 725                         CT_TRACE("fsetxattr of '%s' on '%s' rc=%d (%s)",
 726                                  name, dst, rc, strerror(errno));
 727                         /* lustre.* attrs aren't supported on other FS's */
 728                         if (rc < 0 && errno != EOPNOTSUPP) {
 729                                 rc = -errno;
 730                                 CT_ERROR(rc, "cannot set extended attribute"
 731                                          " '%s' of '%s'",
 732                                          name, dst);
 733                                 return rc;
 734                         }
 735                 }
 736                 name += strlen(name) + 1;
 737         }
 738
 739         return 0;
 740 }
 741
 742 static int ct_path_lustre(char *buf, int sz, const char *mnt,
 743                           const lustre_fid *fid)
 744 {
 745         return snprintf(buf, sz, "%s/%s/fid/"DFID_NOBRACE, mnt,
 746                         dot_lustre_name, PFID(fid));
 747 }
 748
 749 static int ct_path_archive(char *buf, int sz, const char *archive_dir,
 750                            const lustre_fid *fid)
 751 {
 752         return snprintf(buf, sz, "%s/%04x/%04x/%04x/%04x/%04x/%04x/"
 753                         DFID_NOBRACE, archive_dir,
 754                         (fid)->f_oid       & 0xFFFF,
 755                         (fid)->f_oid >> 16 & 0xFFFF,
 756                         (unsigned int)((fid)->f_seq       & 0xFFFF),
 757                         (unsigned int)((fid)->f_seq >> 16 & 0xFFFF),
 758                         (unsigned int)((fid)->f_seq >> 32 & 0xFFFF),
 759                         (unsigned int)((fid)->f_seq >> 48 & 0xFFFF),
 760                         PFID(fid));
 761 }
 762
 763 static bool ct_is_retryable(int err)
 764 {
 765         return err == -ETIMEDOUT;
 766 }
 767
 768 static int ct_begin_restore(struct hsm_copyaction_private **phcp,
 769                             const struct hsm_action_item *hai,
 770                             int mdt_index, int open_flags)
 771 {
 772         char     src[PATH_MAX];
 773         int      rc;
 774
 775         rc = llapi_hsm_action_begin(phcp, ctdata, hai, mdt_index, open_flags,
 776                                     false);
 777         if (rc < 0) {
 778                 ct_path_lustre(src, sizeof(src), opt.o_mnt, &hai->hai_fid);
 779                 CT_ERROR(rc, "llapi_hsm_action_begin() on '%s' failed", src);
 780         }
 781
 782         return rc;
 783 }
 784
 785 static int ct_begin(struct hsm_copyaction_private **phcp,
 786                     const struct hsm_action_item *hai)
 787 {
 788         /* Restore takes specific parameters. Call the same function w/ default
 789          * values for all other operations. */
 790         return ct_begin_restore(phcp, hai, -1, 0);
 791 }
 792
 793 static int ct_fini(struct hsm_copyaction_private **phcp,
 794                    const struct hsm_action_item *hai, int hp_flags, int ct_rc)
 795 {
 796         struct hsm_copyaction_private   *hcp;
 797         char                             lstr[PATH_MAX];
 798         int                              rc;
 799
 800         CT_TRACE("Action completed, notifying coordinator "
 801                  "cookie="LPX64", FID="DFID", hp_flags=%d err=%d",
 802                  hai->hai_cookie, PFID(&hai->hai_fid),
 803                  hp_flags, -ct_rc);
 804
 805         ct_path_lustre(lstr, sizeof(lstr), opt.o_mnt, &hai->hai_fid);
 806
 807         if (phcp == NULL || *phcp == NULL) {
 808                 rc = llapi_hsm_action_begin(&hcp, ctdata, hai, -1, 0, true);
 809                 if (rc < 0) {
 810                         CT_ERROR(rc, "llapi_hsm_action_begin() on '%s' failed",
 811                                  lstr);
 812                         return rc;
 813                 }
 814                 phcp = &hcp;
 815         }
 816
 817         rc = llapi_hsm_action_end(phcp, &hai->hai_extent, hp_flags, abs(ct_rc));
 818         if (rc == -ECANCELED)
 819                 CT_ERROR(rc, "completed action on '%s' has been canceled: "
 820                          "cookie="LPX64", FID="DFID, lstr, hai->hai_cookie,
 821                          PFID(&hai->hai_fid));
 822         else if (rc < 0)
 823                 CT_ERROR(rc, "llapi_hsm_action_end() on '%s' failed", lstr);
 824         else
 825                 CT_TRACE("llapi_hsm_action_end() on '%s' ok (rc=%d)",
 826                          lstr, rc);
 827
 828         return rc;
 829 }
 830
 831 static int ct_archive(const struct hsm_action_item *hai, const long hal_flags)
 832 {
 833         struct hsm_copyaction_private   *hcp = NULL;
 834         char                             src[PATH_MAX];
 835         char                             dst[PATH_MAX] = "";
 836         int                              rc;
 837         int                              rcf = 0;
 838         bool                             rename_needed = false;
 839         int                              hp_flags = 0;
 840         int                              open_flags;
 841         int                              src_fd = -1;
 842         int                              dst_fd = -1;
 843
 844         rc = ct_begin(&hcp, hai);
 845         if (rc < 0)
 846                 goto fini_major;
 847
 848         /* we fill archive so:
 849          * source = data FID
 850          * destination = lustre FID
 851          */
 852         ct_path_lustre(src, sizeof(src), opt.o_mnt, &hai->hai_dfid);
 853         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, &hai->hai_fid);
 854         if (hai->hai_extent.length == -1) {
 855                 /* whole file, write it to tmp location and atomically
 856                  * replace old archived file */
 857                 strncat(dst, "_tmp", sizeof(dst) - strlen(dst) - 1);
 858                 /* we cannot rely on the same test because ct_copy_data()
 859                  * updates hai_extent.length */
 860                 rename_needed = true;
 861         }
 862
 863         CT_TRACE("archiving '%s' to '%s'", src, dst);
 864
 865         if (opt.o_dry_run) {
 866                 rc = 0;
 867                 goto fini_major;
 868         }
 869
 870         rc = ct_mkdir_p(dst);
 871         if (rc < 0) {
 872                 CT_ERROR(rc, "mkdir_p '%s' failed", dst);
 873                 goto fini_major;
 874         }
 875
 876         src_fd = llapi_hsm_action_get_fd(hcp);
 877         if (src_fd < 0) {
 878                 rc = src_fd;
 879                 CT_ERROR(rc, "cannot open '%s' for read", src);
 880                 goto fini_major;
 881         }
 882
 883         open_flags = O_WRONLY | O_NOFOLLOW | O_NONBLOCK;
 884         /* If extent is specified, don't truncate an old archived copy */
 885         open_flags |= ((hai->hai_extent.length == -1) ? O_TRUNC : 0) | O_CREAT;
 886
 887         dst_fd = open(dst, open_flags, FILE_PERM);
 888         if (dst_fd == -1) {
 889                 rc = -errno;
 890                 CT_ERROR(rc, "cannot open '%s' for write", dst);
 891                 goto fini_major;
 892         }
 893
 894         /* saving stripe is not critical */
 895         rc = ct_save_stripe(src_fd, src, dst);
 896         if (rc < 0)
 897                 CT_ERROR(rc, "cannot save file striping info of '%s' in '%s'",
 898                          src, dst);
 899
 900         rc = ct_copy_data(hcp, src, dst, src_fd, dst_fd, hai, hal_flags);
 901         if (rc < 0) {
 902                 CT_ERROR(rc, "data copy failed from '%s' to '%s'", src, dst);
 903                 goto fini_major;
 904         }
 905
 906         rc = fsync(dst_fd);
 907         if (rc < 0) {
 908                 rc = -errno;
 909                 CT_ERROR(rc, "cannot flush '%s' archive file '%s'", src, dst);
 910                 goto fini_major;
 911         }
 912
 913         CT_TRACE("data archiving for '%s' to '%s' done", src, dst);
 914
 915         /* attrs will remain on the MDS; no need to copy them, except possibly
 916           for disaster recovery */
 917         if (opt.o_copy_attrs) {
 918                 rc = ct_copy_attr(src, dst, src_fd, dst_fd);
 919                 if (rc < 0) {
 920                         CT_ERROR(rc, "cannot copy attr of '%s' to '%s'",
 921                                  src, dst);
 922                         rcf = rc;
 923                 }
 924                 CT_TRACE("attr file for '%s' saved to archive '%s'",
 925                          src, dst);
 926         }
 927
 928         /* xattrs will remain on the MDS; no need to copy them, except possibly
 929          for disaster recovery */
 930         if (opt.o_copy_xattrs) {
 931                 rc = ct_copy_xattr(src, dst, src_fd, dst_fd, false);
 932                 if (rc < 0) {
 933                         CT_ERROR(rc, "cannot copy xattr of '%s' to '%s'",
 934                                  src, dst);
 935                         rcf = rcf ? rcf : rc;
 936                 }
 937                 CT_TRACE("xattr file for '%s' saved to archive '%s'",
 938                          src, dst);
 939         }
 940
 941         if (rename_needed == true) {
 942                 char     tmp_src[PATH_MAX];
 943                 char     tmp_dst[PATH_MAX];
 944
 945                 /* atomically replace old archived file */
 946                 ct_path_archive(src, sizeof(src), opt.o_hsm_root,
 947                                 &hai->hai_fid);
 948                 rc = rename(dst, src);
 949                 if (rc < 0) {
 950                         rc = -errno;
 951                         CT_ERROR(rc, "cannot rename '%s' to '%s'", dst, src);
 952                         goto fini_major;
 953                 }
 954                 /* rename lov file */
 955                 snprintf(tmp_src, sizeof(tmp_src), "%s.lov", src);
 956                 snprintf(tmp_dst, sizeof(tmp_dst), "%s.lov", dst);
 957                 rc = rename(tmp_dst, tmp_src);
 958                 if (rc < 0)
 959                         CT_ERROR(errno, "cannot rename '%s' to '%s'",
 960                                  tmp_dst, tmp_src);
 961         }
 962
 963         if (opt.o_shadow_tree) {
 964                 /* Create a namespace of softlinks that shadows the original
 965                  * Lustre namespace.  This will only be current at
 966                  * time-of-archive (won't follow renames).
 967                  * WARNING: release won't kill these links; a manual
 968                  * cleanup of dead links would be required.
 969                  */
 970                 char             buf[PATH_MAX];
 971                 long long        recno = -1;
 972                 int              linkno = 0;
 973                 char            *ptr;
 974                 int              depth = 0;
 975                 int              sz;
 976
 977                 sprintf(buf, DFID, PFID(&hai->hai_fid));
 978                 sprintf(src, "%s/shadow/", opt.o_hsm_root);
 979
 980                 ptr = opt.o_hsm_root;
 981                 while (*ptr)
 982                         (*ptr++ == '/') ? depth-- : 0;
 983
 984                 rc = llapi_fid2path(opt.o_mnt, buf, src + strlen(src),
 985                                     sizeof(src) - strlen(src), &recno, &linkno);
 986                 if (rc < 0) {
 987                         CT_ERROR(rc, "cannot get FID of '%s'", buf);
 988                         rcf = rcf ? rcf : rc;
 989                         goto fini_minor;
 990                 }
 991
 992                 /* Figure out how many parent dirs to symlink back */
 993                 ptr = src;
 994                 while (*ptr)
 995                         (*ptr++ == '/') ? depth++ : 0;
 996                 sprintf(buf, "..");
 997                 while (--depth > 1)
 998                         strcat(buf, "/..");
 999
1000                 ct_path_archive(dst, sizeof(dst), buf, &hai->hai_fid);
1001
1002                 if (ct_mkdir_p(src)) {
1003                         CT_ERROR(errno, "mkdir_p '%s' failed", src);
1004                         rcf = rcf ? rcf : -errno;
1005                         goto fini_minor;
1006                 }
1007                 /* symlink already exists ? */
1008                 sz = readlink(src, buf, sizeof(buf));
1009                 if (sz >= 0) {
1010                         buf[sz] = '\0';
1011                         if (sz == 0 || strncmp(buf, dst, sz) != 0) {
1012                                 if (unlink(src) && errno != ENOENT) {
1013                                         CT_ERROR(errno,
1014                                                  "cannot unlink symlink '%s'",
1015                                                  src);
1016                                         rcf = rcf ? rcf : -errno;
1017                                         goto fini_minor;
1018                                 /* unlink old symlink done */
1019                                 CT_TRACE("remove old symlink '%s' pointing"
1020                                          " to '%s'", src, buf);
1021                                 }
1022                         } else {
1023                                 /* symlink already ok */
1024                                 CT_TRACE("symlink '%s' already pointing"
1025                                          " to '%s'", src, dst);
1026                                 rcf = 0;
1027                                 goto fini_minor;
1028                         }
1029                 }
1030                 if (symlink(dst, src)) {
1031                         CT_ERROR(errno, "cannot symlink '%s' to '%s'",
1032                                  src, dst);
1033                         rcf = rcf ? rcf : -errno;
1034                         goto fini_minor;
1035                 }
1036                 CT_TRACE("symlink '%s' to '%s' done", src, dst);
1037         }
1038 fini_minor:
1039         if (rcf)
1040                 err_minor++;
1041         goto out;
1042
1043
1044 fini_major:
1045         err_major++;
1046
1047         unlink(dst);
1048         if (ct_is_retryable(rc))
1049                 hp_flags |= HP_FLAG_RETRY;
1050
1051         rcf = rc;
1052
1053 out:
1054         if (!(src_fd < 0))
1055                 close(src_fd);
1056
1057         if (!(dst_fd < 0))
1058                 close(dst_fd);
1059
1060         rc = ct_fini(&hcp, hai, hp_flags, rcf);
1061
1062         return rc;
1063 }
1064
1065 static int ct_restore(const struct hsm_action_item *hai, const long hal_flags)
1066 {
1067         struct hsm_copyaction_private   *hcp = NULL;
1068         char                             src[PATH_MAX];
1069         char                             dst[PATH_MAX];
1070         char                             lov_buf[XATTR_SIZE_MAX];
1071         size_t                           lov_size = sizeof(lov_buf);
1072         int                              rc;
1073         int                              hp_flags = 0;
1074         int                              src_fd = -1;
1075         int                              dst_fd = -1;
1076         int                              mdt_index = -1; /* Not implemented */
1077         int                              open_flags = 0;
1078         bool                             set_lovea;
1079         struct lu_fid                    dfid;
1080         /* we fill lustre so:
1081          * source = lustre FID in the backend
1082          * destination = data FID = volatile file
1083          */
1084
1085         /* build backend file name from released file FID */
1086         ct_path_archive(src, sizeof(src), opt.o_hsm_root, &hai->hai_fid);
1087
1088         /* restore loads and sets the LOVEA w/o interpreting it to avoid
1089          * dependency on the structure format. */
1090         rc = ct_load_stripe(src, lov_buf, &lov_size);
1091         if (rc < 0) {
1092                 CT_WARN("cannot get stripe rules for '%s' (%s), use default",
1093                         src, strerror(-rc));
1094                 set_lovea = false;
1095         } else {
1096                 open_flags |= O_LOV_DELAY_CREATE;
1097                 set_lovea = true;
1098         }
1099
1100         rc = ct_begin_restore(&hcp, hai, mdt_index, open_flags);
1101         if (rc < 0)
1102                 goto fini;
1103
1104         /* get the FID of the volatile file */
1105         rc = llapi_hsm_action_get_dfid(hcp, &dfid);
1106         if (rc < 0) {
1107                 CT_ERROR(rc, "restoring "DFID
1108                          ", cannot get FID of created volatile file",
1109                          PFID(&hai->hai_fid));
1110                 goto fini;
1111         }
1112
1113         /* build volatile "file name", for messages */
1114         snprintf(dst, sizeof(dst), "{VOLATILE}="DFID, PFID(&dfid));
1115
1116         CT_TRACE("restoring data from '%s' to '%s'", src, dst);
1117
1118         if (opt.o_dry_run) {
1119                 rc = 0;
1120                 goto fini;
1121         }
1122
1123         src_fd = open(src, O_RDONLY | O_NOATIME | O_NONBLOCK | O_NOFOLLOW);
1124         if (src_fd < 0) {
1125                 rc = -errno;
1126                 CT_ERROR(rc, "cannot open '%s' for read", src);
1127                 goto fini;
1128         }
1129
1130         dst_fd = llapi_hsm_action_get_fd(hcp);
1131         if (dst_fd < 0) {
1132                 rc = dst_fd;
1133                 CT_ERROR(rc, "cannot open '%s' for write", dst);
1134                 goto fini;
1135         }
1136
1137         if (set_lovea) {
1138                 /* the layout cannot be allocated through .fid so we have to
1139                  * restore a layout */
1140                 rc = ct_restore_stripe(src, dst, dst_fd, lov_buf, lov_size);
1141                 if (rc < 0) {
1142                         CT_ERROR(rc, "cannot restore file striping info"
1143                                  " for '%s' from '%s'", dst, src);
1144                         err_major++;
1145                         goto fini;
1146                 }
1147         }
1148
1149         rc = ct_copy_data(hcp, src, dst, src_fd, dst_fd, hai, hal_flags);
1150         if (rc < 0) {
1151                 CT_ERROR(rc, "cannot copy data from '%s' to '%s'",
1152                          src, dst);
1153                 err_major++;
1154                 if (ct_is_retryable(rc))
1155                         hp_flags |= HP_FLAG_RETRY;
1156                 goto fini;
1157         }
1158
1159         CT_TRACE("data restore from '%s' to '%s' done", src, dst);
1160
1161 fini:
1162         rc = ct_fini(&hcp, hai, hp_flags, rc);
1163
1164         /* object swaping is done by cdt at copy end, so close of volatile file
1165          * cannot be done before */
1166         if (!(src_fd < 0))
1167                 close(src_fd);
1168
1169         if (!(dst_fd < 0))
1170                 close(dst_fd);
1171
1172         return rc;
1173 }
1174
1175 static int ct_remove(const struct hsm_action_item *hai, const long hal_flags)
1176 {
1177         struct hsm_copyaction_private   *hcp = NULL;
1178         char                             dst[PATH_MAX];
1179         int                              rc;
1180
1181         rc = ct_begin(&hcp, hai);
1182         if (rc < 0)
1183                 goto fini;
1184
1185         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, &hai->hai_fid);
1186
1187         CT_TRACE("removing file '%s'", dst);
1188
1189         if (opt.o_dry_run) {
1190                 rc = 0;
1191                 goto fini;
1192         }
1193
1194         rc = unlink(dst);
1195         if (rc < 0) {
1196                 rc = -errno;
1197                 CT_ERROR(rc, "cannot unlink '%s'", dst);
1198                 err_minor++;
1199                 goto fini;
1200         }
1201
1202         strncat(dst, ".lov", sizeof(dst) - strlen(dst) - 1);
1203         rc = unlink(dst);
1204         if (rc < 0) {
1205                 rc = -errno;
1206                 CT_ERROR(rc, "cannot unlink '%s'", dst);
1207                 err_minor++;
1208                 goto fini;
1209         }
1210
1211 fini:
1212         rc = ct_fini(&hcp, hai, 0, rc);
1213
1214         return rc;
1215 }
1216
1217 static int ct_process_item(struct hsm_action_item *hai, const long hal_flags)
1218 {
1219         int     rc = 0;
1220
1221         if (opt.o_verbose >= LLAPI_MSG_INFO || opt.o_dry_run) {
1222                 /* Print the original path */
1223                 char            fid[128];
1224                 char            path[PATH_MAX];
1225                 long long       recno = -1;
1226                 int             linkno = 0;
1227
1228                 sprintf(fid, DFID, PFID(&hai->hai_fid));
1229                 CT_TRACE("'%s' action %s reclen %d, cookie="LPX64,
1230                          fid, hsm_copytool_action2name(hai->hai_action),
1231                          hai->hai_len, hai->hai_cookie);
1232                 rc = llapi_fid2path(opt.o_mnt, fid, path,
1233                                     sizeof(path), &recno, &linkno);
1234                 if (rc < 0)
1235                         CT_ERROR(rc, "cannot get path of FID %s", fid);
1236                 else
1237                         CT_TRACE("processing file '%s'", path);
1238         }
1239
1240         switch (hai->hai_action) {
1241         /* set err_major, minor inside these functions */
1242         case HSMA_ARCHIVE:
1243                 rc = ct_archive(hai, hal_flags);
1244                 break;
1245         case HSMA_RESTORE:
1246                 rc = ct_restore(hai, hal_flags);
1247                 break;
1248         case HSMA_REMOVE:
1249                 rc = ct_remove(hai, hal_flags);
1250                 break;
1251         case HSMA_CANCEL:
1252                 CT_TRACE("cancel not implemented for file system '%s'",
1253                          opt.o_mnt);
1254                 /* Don't report progress to coordinator for this cookie:
1255                  * the copy function will get ECANCELED when reporting
1256                  * progress. */
1257                 err_minor++;
1258                 return 0;
1259                 break;
1260         default:
1261                 rc = -EINVAL;
1262                 CT_ERROR(rc, "unknown action %d, on '%s'", hai->hai_action,
1263                          opt.o_mnt);
1264                 err_minor++;
1265                 ct_fini(NULL, hai, 0, rc);
1266         }
1267
1268         return 0;
1269 }
1270
1271 struct ct_th_data {
1272         long                     hal_flags;
1273         struct hsm_action_item  *hai;
1274 };
1275
1276 static void *ct_thread(void *data)
1277 {
1278         struct ct_th_data *cttd = data;
1279         int rc;
1280
1281         rc = ct_process_item(cttd->hai, cttd->hal_flags);
1282
1283         free(cttd->hai);
1284         free(cttd);
1285         pthread_exit((void *)(intptr_t)rc);
1286 }
1287
1288 static int ct_process_item_async(const struct hsm_action_item *hai,
1289                                  long hal_flags)
1290 {
1291         pthread_attr_t           attr;
1292         pthread_t                thread;
1293         struct ct_th_data       *data;
1294         int                      rc;
1295
1296         data = malloc(sizeof(*data));
1297         if (data == NULL)
1298                 return -ENOMEM;
1299
1300         data->hai = malloc(hai->hai_len);
1301         if (data->hai == NULL) {
1302                 free(data);
1303                 return -ENOMEM;
1304         }
1305
1306         memcpy(data->hai, hai, hai->hai_len);
1307         data->hal_flags = hal_flags;
1308
1309         rc = pthread_attr_init(&attr);
1310         if (rc != 0) {
1311                 CT_ERROR(rc, "pthread_attr_init failed for '%s' service",
1312                          opt.o_mnt);
1313                 free(data->hai);
1314                 free(data);
1315                 return -rc;
1316         }
1317
1318         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
1319
1320         rc = pthread_create(&thread, &attr, ct_thread, data);
1321         if (rc != 0)
1322                 CT_ERROR(rc, "cannot create thread for '%s' service",
1323                          opt.o_mnt);
1324
1325         pthread_attr_destroy(&attr);
1326         return 0;
1327 }
1328
1329 static int ct_import_one(const char *src, const char *dst)
1330 {
1331         char            newarc[PATH_MAX];
1332         lustre_fid      fid;
1333         struct stat     st;
1334         int             rc;
1335
1336         CT_TRACE("importing '%s' from '%s'", dst, src);
1337
1338         if (stat(src, &st) < 0) {
1339                 rc = -errno;
1340                 CT_ERROR(rc, "cannot stat '%s'", src);
1341                 return rc;
1342         }
1343
1344         if (opt.o_dry_run)
1345                 return 0;
1346
1347         rc = llapi_hsm_import(dst,
1348                               opt.o_archive_cnt ? opt.o_archive_id[0] : 0,
1349                               &st, 0, 0, 0, 0, NULL, &fid);
1350         if (rc < 0) {
1351                 CT_ERROR(rc, "cannot import '%s' from '%s'", dst, src);
1352                 return rc;
1353         }
1354
1355         ct_path_archive(newarc, sizeof(newarc), opt.o_hsm_root, &fid);
1356
1357         rc = ct_mkdir_p(newarc);
1358         if (rc < 0) {
1359                 CT_ERROR(rc, "mkdir_p '%s' failed", newarc);
1360                 err_major++;
1361                 return rc;
1362
1363         }
1364
1365         /* Lots of choices now: mv, ln, ln -s ? */
1366         rc = link(src, newarc); /* hardlink */
1367         if (rc < 0) {
1368                 rc = -errno;
1369                 CT_ERROR(rc, "cannot link '%s' to '%s'", newarc, src);
1370                 err_major++;
1371                 return rc;
1372         }
1373         CT_TRACE("imported '%s' from '%s'=='%s'", dst, newarc, src);
1374
1375         return 0;
1376 }
1377
1378 static char *path_concat(const char *dirname, const char *basename)
1379 {
1380         char    *result;
1381         int      rc;
1382
1383         rc = asprintf(&result, "%s/%s", dirname, basename);
1384         if (rc < 0)
1385                 return NULL;
1386
1387         return result;
1388 }
1389
1390 static int ct_import_fid(const lustre_fid *import_fid)
1391 {
1392         char    fid_path[PATH_MAX];
1393         int     rc;
1394
1395         ct_path_lustre(fid_path, sizeof(fid_path), opt.o_mnt, import_fid);
1396         rc = access(fid_path, F_OK);
1397         if (rc == 0 || errno != ENOENT) {
1398                 rc = (errno == 0) ? -EEXIST : -errno;
1399                 CT_ERROR(rc, "cannot import '"DFID"'", PFID(import_fid));
1400                 return rc;
1401         }
1402
1403         ct_path_archive(fid_path, sizeof(fid_path), opt.o_hsm_root,
1404                         import_fid);
1405
1406         CT_TRACE("Resolving "DFID" to %s", PFID(import_fid), fid_path);
1407
1408         return ct_import_one(fid_path, opt.o_dst);
1409 }
1410
1411 static int ct_import_recurse(const char *relpath)
1412 {
1413         DIR             *dir;
1414         struct dirent    ent, *cookie = NULL;
1415         char            *srcpath, *newpath;
1416         lustre_fid       import_fid;
1417         int              rc;
1418
1419         if (relpath == NULL)
1420                 return -EINVAL;
1421
1422         /* Is relpath a FID? In which case SFID should expand to three
1423          * elements. */
1424         rc = sscanf(relpath, SFID, RFID(&import_fid));
1425         if (rc == 3)
1426                 return ct_import_fid(&import_fid);
1427
1428         srcpath = path_concat(opt.o_hsm_root, relpath);
1429         if (srcpath == NULL) {
1430                 err_major++;
1431                 return -ENOMEM;
1432         }
1433
1434         dir = opendir(srcpath);
1435         if (dir == NULL) {
1436                 /* Not a dir, or error */
1437                 if (errno == ENOTDIR) {
1438                         /* Single regular file case, treat o_dst as absolute
1439                            final location. */
1440                         rc = ct_import_one(srcpath, opt.o_dst);
1441                 } else {
1442                         rc = -errno;
1443                         CT_ERROR(rc, "cannot opendir '%s'", srcpath);
1444                         err_major++;
1445                 }
1446                 free(srcpath);
1447                 return rc;
1448         }
1449         free(srcpath);
1450
1451         while (1) {
1452                 rc = readdir_r(dir, &ent, &cookie);
1453                 if (rc != 0) {
1454                         rc = -errno;
1455                         CT_ERROR(rc, "cannot readdir_r '%s'", relpath);
1456                         err_major++;
1457                         goto out;
1458                 } else if ((rc == 0) && (cookie == NULL)) {
1459                         /* end of directory */
1460                         break;
1461                 }
1462
1463                 if (!strcmp(ent.d_name, ".") ||
1464                     !strcmp(ent.d_name, ".."))
1465                         continue;
1466
1467                 /* New relative path */
1468                 newpath = path_concat(relpath, ent.d_name);
1469                 if (newpath == NULL) {
1470                         err_major++;
1471                         rc = -ENOMEM;
1472                         goto out;
1473                 }
1474
1475                 if (ent.d_type == DT_DIR) {
1476                         rc = ct_import_recurse(newpath);
1477                 } else {
1478                         char src[PATH_MAX];
1479                         char dst[PATH_MAX];
1480
1481                         sprintf(src, "%s/%s", opt.o_hsm_root, newpath);
1482                         sprintf(dst, "%s/%s", opt.o_dst, newpath);
1483                         /* Make the target dir in the Lustre fs */
1484                         rc = ct_mkdir_p(dst);
1485                         if (rc == 0) {
1486                                 /* Import the file */
1487                                 rc = ct_import_one(src, dst);
1488                         } else {
1489                                 CT_ERROR(rc, "ct_mkdir_p '%s' failed", dst);
1490                                 err_major++;
1491                         }
1492                 }
1493
1494                 if (rc != 0) {
1495                         CT_ERROR(rc, "cannot import '%s'", newpath);
1496                         if (err_major && opt.o_abort_on_error) {
1497                                 free(newpath);
1498                                 goto out;
1499                         }
1500                 }
1501                 free(newpath);
1502         }
1503
1504         rc = 0;
1505 out:
1506         closedir(dir);
1507         return rc;
1508 }
1509
1510 static int ct_rebind_one(const lustre_fid *old_fid, const lustre_fid *new_fid)
1511 {
1512         char    src[PATH_MAX];
1513         char    dst[PATH_MAX];
1514         int     rc;
1515
1516         CT_TRACE("rebind "DFID" to "DFID, PFID(old_fid), PFID(new_fid));
1517
1518         ct_path_archive(src, sizeof(src), opt.o_hsm_root, old_fid);
1519         ct_path_archive(dst, sizeof(dst), opt.o_hsm_root, new_fid);
1520
1521         if (!opt.o_dry_run) {
1522                 ct_mkdir_p(dst);
1523                 if (rename(src, dst)) {
1524                         rc = -errno;
1525                         CT_ERROR(rc, "cannot rename '%s' to '%s'", src, dst);
1526                         return -errno;
1527                 }
1528                 /* rename lov file */
1529                 strncat(src, ".lov", sizeof(src) - strlen(src) - 1);
1530                 strncat(dst, ".lov", sizeof(dst) - strlen(dst) - 1);
1531                 if (rename(src, dst))
1532                         CT_ERROR(errno, "cannot rename '%s' to '%s'", src, dst);
1533
1534         }
1535         return 0;
1536 }
1537
1538 static bool fid_is_file(lustre_fid *fid)
1539 {
1540         return fid_is_norm(fid) || fid_is_igif(fid);
1541 }
1542
1543 static bool should_ignore_line(const char *line)
1544 {
1545         int     i;
1546
1547         for (i = 0; line[i] != '\0'; i++) {
1548                 if (isspace(line[i]))
1549                         continue;
1550                 else if (line[i] == '#')
1551                         return true;
1552                 else
1553                         return false;
1554         }
1555
1556         return true;
1557 }
1558
1559 static int ct_rebind_list(const char *list)
1560 {
1561         int              rc;
1562         FILE            *filp;
1563         ssize_t          r;
1564         char            *line = NULL;
1565         size_t           line_size = 0;
1566         unsigned int     nl = 0;
1567         unsigned int     ok = 0;
1568
1569         filp = fopen(list, "r");
1570         if (filp == NULL) {
1571                 rc = -errno;
1572                 CT_ERROR(rc, "cannot open '%s'", list);
1573                 return rc;
1574         }
1575
1576         /* each line consists of 2 FID */
1577         while ((r = getline(&line, &line_size, filp)) != -1) {
1578                 lustre_fid      old_fid;
1579                 lustre_fid      new_fid;
1580
1581                 /* Ignore empty and commented out ('#...') lines. */
1582                 if (should_ignore_line(line))
1583                         continue;
1584
1585                 nl++;
1586
1587                 rc = sscanf(line, SFID" "SFID, RFID(&old_fid), RFID(&new_fid));
1588                 if (rc != 6 || !fid_is_file(&old_fid) ||
1589                     !fid_is_file(&new_fid)) {
1590                         CT_ERROR(EINVAL,
1591                                  "'%s' FID expected near '%s', line %u",
1592                                  list, line, nl);
1593                         err_major++;
1594                         continue;
1595                 }
1596
1597                 if (ct_rebind_one(&old_fid, &new_fid))
1598                         err_major++;
1599                 else
1600                         ok++;
1601         }
1602
1603         fclose(filp);
1604
1605         if (line)
1606                 free(line);
1607
1608         /* return 0 if all rebinds were sucessful */
1609         CT_TRACE("%u lines read from '%s', %u rebind successful", nl, list, ok);
1610
1611         return ok == nl ? 0 : -1;
1612 }
1613
1614 static int ct_rebind(void)
1615 {
1616         int     rc;
1617
1618         if (opt.o_dst) {
1619                 lustre_fid      old_fid;
1620                 lustre_fid      new_fid;
1621
1622                 if (sscanf(opt.o_src, SFID, RFID(&old_fid)) != 3 ||
1623                     !fid_is_file(&old_fid)) {
1624                         rc = -EINVAL;
1625                         CT_ERROR(rc, "'%s' invalid FID format", opt.o_src);
1626                         return rc;
1627                 }
1628
1629                 if (sscanf(opt.o_dst, SFID, RFID(&new_fid)) != 3 ||
1630                     !fid_is_file(&new_fid)) {
1631                         rc = -EINVAL;
1632                         CT_ERROR(rc, "'%s' invalid FID format", opt.o_dst);
1633                         return rc;
1634                 }
1635
1636                 rc = ct_rebind_one(&old_fid, &new_fid);
1637
1638                 return rc;
1639         }
1640
1641         /* o_src is a list file */
1642         rc = ct_rebind_list(opt.o_src);
1643
1644         return rc;
1645 }
1646
1647 static int ct_dir_level_max(const char *dirpath, __u16 *sub_seqmax)
1648 {
1649         DIR             *dir;
1650         int              rc;
1651         __u16            sub_seq;
1652         struct dirent    ent, *cookie = NULL;
1653
1654         *sub_seqmax = 0;
1655
1656         dir = opendir(dirpath);
1657         if (dir == NULL) {
1658                 rc = -errno;
1659                 CT_ERROR(rc, "cannot open directory '%s'", opt.o_hsm_root);
1660                 return rc;
1661         }
1662
1663         while ((rc = readdir_r(dir, &ent, &cookie)) == 0) {
1664                 if (cookie == NULL)
1665                         /* end of directory.
1666                          * rc is 0 and seqmax contains the max value. */
1667                         goto out;
1668
1669                 if (!strcmp(ent.d_name, ".") || !strcmp(ent.d_name, ".."))
1670                         continue;
1671
1672                 if (sscanf(ent.d_name, "%hx", &sub_seq) != 1) {
1673                         CT_TRACE("'%s' has an unexpected dirname format, "
1674                                  "skip entry", ent.d_name);
1675                         continue;
1676                 }
1677                 if (sub_seq > *sub_seqmax)
1678                         *sub_seqmax = sub_seq;
1679         }
1680         rc = -errno;
1681         CT_ERROR(rc, "cannot readdir_r '%s'", dirpath);
1682
1683 out:
1684         closedir(dir);
1685         return rc;
1686 }
1687
1688 static int ct_max_sequence(void)
1689 {
1690         int   rc, i;
1691         char  path[PATH_MAX];
1692         __u64 seq = 0;
1693         __u16 subseq;
1694
1695         strncpy(path, opt.o_hsm_root, sizeof(path));
1696         /* FID sequence is stored in top-level directory names:
1697          * hsm_root/16bits (high weight)/16 bits/16 bits/16 bits (low weight).
1698          */
1699         for (i = 0; i < 4; i++) {
1700                 rc = ct_dir_level_max(path, &subseq);
1701                 if (rc != 0)
1702                         return rc;
1703                 seq |= ((__u64)subseq << ((3 - i) * 16));
1704                 sprintf(path + strlen(path), "/%04x", subseq);
1705         }
1706
1707         printf("max_sequence: "LPX64"\n", seq);
1708
1709         return 0;
1710 }
1711
1712 static void handler(int signal)
1713 {
1714         psignal(signal, "exiting");
1715         /* If we don't clean up upon interrupt, umount thinks there's a ref
1716          * and doesn't remove us from mtab (EINPROGRESS). The lustre client
1717          * does successfully unmount and the mount is actually gone, but the
1718          * mtab entry remains. So this just makes mtab happier. */
1719         llapi_hsm_copytool_unregister(&ctdata);
1720         _exit(1);
1721 }
1722
1723 /* Daemon waits for messages from the kernel; run it in the background. */
1724 static int ct_run(void)
1725 {
1726         int     rc;
1727
1728         if (opt.o_daemonize) {
1729                 rc = daemon(1, 1);
1730                 if (rc < 0) {
1731                         rc = -errno;
1732                         CT_ERROR(rc, "cannot daemonize");
1733                         return rc;
1734                 }
1735         }
1736
1737         setbuf(stdout, NULL);
1738
1739         if (opt.o_event_fifo != NULL) {
1740                 rc = llapi_hsm_register_event_fifo(opt.o_event_fifo);
1741                 if (rc < 0) {
1742                         CT_ERROR(rc, "failed to register event fifo");
1743                         return rc;
1744                 }
1745                 llapi_error_callback_set(llapi_hsm_log_error);
1746         }
1747
1748         rc = llapi_hsm_copytool_register(&ctdata, opt.o_mnt,
1749                                          opt.o_archive_cnt,
1750                                          opt.o_archive_id, 0);
1751         if (rc < 0) {
1752                 CT_ERROR(rc, "cannot start copytool interface");
1753                 return rc;
1754         }
1755
1756         signal(SIGINT, handler);
1757         signal(SIGTERM, handler);
1758
1759         while (1) {
1760                 struct hsm_action_list  *hal;
1761                 struct hsm_action_item  *hai;
1762                 int                      msgsize;
1763                 int                      i = 0;
1764
1765                 CT_TRACE("waiting for message from kernel");
1766
1767                 rc = llapi_hsm_copytool_recv(ctdata, &hal, &msgsize);
1768                 if (rc == -ESHUTDOWN) {
1769                         CT_TRACE("shutting down");
1770                         break;
1771                 } else if (rc < 0) {
1772                         CT_WARN("cannot receive action list: %s",
1773                                 strerror(-rc));
1774                         err_major++;
1775                         if (opt.o_abort_on_error)
1776                                 break;
1777                         else
1778                                 continue;
1779                 }
1780
1781                 CT_TRACE("copytool fs=%s archive#=%d item_count=%d",
1782                          hal->hal_fsname, hal->hal_archive_id, hal->hal_count);
1783
1784                 if (strcmp(hal->hal_fsname, fs_name) != 0) {
1785                         rc = -EINVAL;
1786                         CT_ERROR(rc, "'%s' invalid fs name, expecting: %s",
1787                                  hal->hal_fsname, fs_name);
1788                         err_major++;
1789                         if (opt.o_abort_on_error)
1790                                 break;
1791                         else
1792                                 continue;
1793                 }
1794
1795                 hai = hai_first(hal);
1796                 while (++i <= hal->hal_count) {
1797                         if ((char *)hai - (char *)hal > msgsize) {
1798                                 rc = -EPROTO;
1799                                 CT_ERROR(rc,
1800                                          "'%s' item %d past end of message!",
1801                                          opt.o_mnt, i);
1802                                 err_major++;
1803                                 break;
1804                         }
1805                         rc = ct_process_item_async(hai, hal->hal_flags);
1806                         if (rc < 0)
1807                                 CT_ERROR(rc, "'%s' item %d process",
1808                                          opt.o_mnt, i);
1809                         if (opt.o_abort_on_error && err_major)
1810                                 break;
1811                         hai = hai_next(hai);
1812                 }
1813
1814                 if (opt.o_abort_on_error && err_major)
1815                         break;
1816         }
1817
1818         llapi_hsm_copytool_unregister(&ctdata);
1819         if (opt.o_event_fifo != NULL)
1820                 llapi_hsm_unregister_event_fifo(opt.o_event_fifo);
1821
1822         return rc;
1823 }
1824
1825 static int ct_setup(void)
1826 {
1827         int     rc;
1828
1829         /* set llapi message level */
1830         llapi_msg_set_level(opt.o_verbose);
1831
1832         arc_fd = open(opt.o_hsm_root, O_RDONLY);
1833         if (arc_fd < 0) {
1834                 rc = -errno;
1835                 CT_ERROR(rc, "cannot open archive at '%s'", opt.o_hsm_root);
1836                 return rc;
1837         }
1838
1839         rc = llapi_search_fsname(opt.o_mnt, fs_name);
1840         if (rc < 0) {
1841                 CT_ERROR(rc, "cannot find a Lustre filesystem mounted at '%s'",
1842                          opt.o_mnt);
1843                 return rc;
1844         }
1845
1846         return rc;
1847 }
1848
1849 static int ct_cleanup(void)
1850 {
1851         int     rc;
1852
1853         if (arc_fd < 0)
1854                 return 0;
1855
1856         if (close(arc_fd) < 0) {
1857                 rc = -errno;
1858                 CT_ERROR(rc, "cannot close archive root directory");
1859                 return rc;
1860         }
1861
1862         return 0;
1863 }
1864
1865 int main(int argc, char **argv)
1866 {
1867         int     rc;
1868
1869         strncpy(cmd_name, basename(argv[0]), sizeof(cmd_name));
1870         rc = ct_parseopts(argc, argv);
1871         if (rc < 0) {
1872                 CT_WARN("try '%s --help' for more information", cmd_name);
1873                 return -rc;
1874         }
1875
1876         rc = ct_setup();
1877         if (rc < 0)
1878                 goto error_cleanup;
1879
1880         switch (opt.o_action) {
1881         case CA_IMPORT:
1882                 rc = ct_import_recurse(opt.o_src);
1883                 break;
1884         case CA_REBIND:
1885                 rc = ct_rebind();
1886                 break;
1887         case CA_MAXSEQ:
1888                 rc = ct_max_sequence();
1889                 break;
1890         default:
1891                 rc = ct_run();
1892                 break;
1893         }
1894
1895         if (opt.o_action != CA_MAXSEQ)
1896                 CT_TRACE("process finished, errs: %d major, %d minor,"
1897                          " rc=%d (%s)", err_major, err_minor, rc,
1898                          strerror(-rc));
1899
1900 error_cleanup:
1901         ct_cleanup();
1902
1903         return -rc;
1904 }
1905