lustre/utils/lustre_rsync.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.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/utils/lustre_rsync.c
  33  *
  34  * Author: Kalpak Shah <Kalpak.Shah@Sun.COM>
  35  * Author: Manoj Joseph <Manoj.Joseph@Sun.COM>
  36  */
  37
  38 /*
  39  * - lustre_rsync is a tool for replicating a lustre filesystem.
  40  *
  41  * - The source-fs is a live lustre filesystem. It is not a
  42  * snapshot. It is mounted and undergoing changes
  43  *
  44  * - The target-fs is a copy of the source-fs from the past. Let's
  45  * call this point, the 'sync point'.
  46  *
  47  * - There is a changelog of all metadata operations that happened on
  48  * the filesystem since the 'sync point'.
  49  *
  50  * - lustre_rsync replicates all the operations saved in the changelog
  51  * on to the target filesystem to make it identical to the source.
  52  *
  53  * To facilitate replication, the lustre filesystem provides
  54  *    a) a way to get the current filesystem path of a given FID
  55  *    b) a way to open files by specifying its FID
  56  *
  57  * The changelog only has a limited amount of information.
  58  *  tfid - The FID of the target file
  59  *  pfid - The FID of the parent of the target file (at the time of
  60  *         the operation)
  61  *  sfid - The FID of the source file
  62  *  spfid - The FID of the parent of the source file
  63  *  name - The name of the target file (at the time of the operation), the name
  64  *         of the source file is appended (delimited with '\0') if this
  65  *         operation involves a source
  66  *
  67  * With just this information, it is not alwasy possible to determine
  68  * the file paths for each operation. For instance, if pfid does not
  69  * exist on the source-fs (due to a subsequent deletion), its path
  70  * cannot be queried. In such cases, lustre_rsync keeps the files in a
  71  * special directory ("/.lustrerepl"). Once all the operations in a
  72  * changelog are replayed, all the files in this special directory
  73  * will get moved to the location as in the source-fs.
  74  *
  75  * Shorthand used: f2p(fid) = fid2path(fid)
  76  *
  77  * The following are the metadata operations of interest.
  78  * 1. creat
  79  *    If tfid is absent on the source-fs, ignore this operation
  80  *    If pfid is absent on the source-fs [or]
  81  *    if f2p(pfid) is not present on target-fs [or]
  82  *    if f2p(pfid)+name != f2p(tfid)
  83  *      creat .lustrerepl/tfid
  84  *      track [pfid,tfid,name]
  85  *    Else
  86  *      creat f2p[tfid]
  87  *
  88  * 2. remove
  89  *    If .lustrerepl/[tfid] is present on the target
  90  *      rm .lustrerepl/[tfid]
  91  *    Else if pfid is present on the source-fs,
  92  *      if f2p(pfid)+name is present,
  93  *        rm f2p(pfid)+name
  94  *
  95  * 3. move (spfid,sname) to (pfid,name)
  96  *    If pfid is present
  97  *      if spfid is also present, mv (spfid,sname) to (pfid,name)
  98  *      else mv .lustrerepl/[sfid] to (pfid,name)
  99  *    Else if pfid is not present,
 100  *      if spfid is present, mv (spfid,sname) .lustrerepl/[sfid]
 101  *    If moving out of .lustrerepl
 102  *      move out all its children in .lustrerepl.
 103  *      [pfid,tfid,name] tracked from (1) is used for this.
 104  */
 105
 106 #include <assert.h>
 107 #include <stdio.h>
 108 #include <stdlib.h>
 109 #include <string.h>
 110 #include <unistd.h>
 111 #include <getopt.h>
 112 #include <stdarg.h>
 113 #include <fcntl.h>
 114 #include <signal.h>
 115 #include <sys/stat.h>
 116 #include <sys/types.h>
 117 #include <errno.h>
 118 #include <limits.h>
 119 #include <utime.h>
 120 #include <time.h>
 121 #include <sys/xattr.h>
 122 #include <linux/types.h>
 123
 124 #include <libcfs/util/string.h>
 125 #include <lustre/lustreapi.h>
 126 #include "lustre_rsync.h"
 127 #include "callvpe.h"
 128
 129 #define REPLICATE_STATUS_VER 1
 130 #define CLEAR_INTERVAL 100
 131 #define DEFAULT_RSYNC_THRESHOLD 0xA00000 /* 10 MB */
 132
 133 #define TYPE_STR_LEN 16
 134
 135 #define DEFAULT_MDT "-MDT0000"
 136 #define SPECIAL_DIR ".lustrerepl"
 137 #define RSYNC "rsync"
 138 #define TYPE "type"
 139
 140 /* Debug flags */
 141 #define DINFO 1
 142 #define DTRACE 2
 143
 144 /* Information for processing a changelog record. This structure is
 145    allocated on the heap instead of allocating large variables on the
 146    stack. */
 147 struct lr_info {
 148         long long recno;
 149         int target_no;
 150         unsigned int is_extended:1;
 151         enum changelog_rec_type type;
 152         char tfid[LR_FID_STR_LEN];
 153         char pfid[LR_FID_STR_LEN];
 154         char sfid[LR_FID_STR_LEN];
 155         char spfid[LR_FID_STR_LEN];
 156         char sname[NAME_MAX + 1];
 157         char name[NAME_MAX + 1];
 158         char src[3 * PATH_MAX + 1];
 159         char dest[3 * PATH_MAX + 1];
 160         char path[PATH_MAX + 1];
 161         char savedpath[PATH_MAX + 1];
 162         char link[PATH_MAX + 1];
 163         char linktmp[PATH_MAX + 1];
 164         int bufsize;
 165         char *buf;
 166
 167         /* Variables for querying the xattributes */
 168         char *xlist;
 169         size_t xsize;
 170         char *xvalue;
 171         size_t xvsize;
 172 };
 173
 174 struct lr_parent_child_list {
 175         struct lr_parent_child_log pc_log;
 176         struct lr_parent_child_list *pc_next;
 177 };
 178
 179 struct lustre_rsync_status *status;
 180 char *statuslog;  /* Name of the status log file */
 181 int logbackedup;
 182 int noxattr;    /* Flag to turn off replicating xattrs */
 183 int noclear;    /* Flag to turn off clearing changelogs */
 184 int debug;      /* Flag to turn debugging information on and off */
 185 int verbose;    /* Verbose output */
 186 long long rec_count; /* No of changelog records that were processed */
 187 int errors;
 188 int dryrun;
 189 int use_rsync;  /* Flag to turn on use of rsync to copy data */
 190 long long rsync_threshold = DEFAULT_RSYNC_THRESHOLD;
 191 int quit;       /* Flag to stop processing the changelog; set on the
 192                    receipt of a signal */
 193 int abort_on_err = 0;
 194
 195 char rsync[PATH_MAX];
 196 char rsync_ver[PATH_MAX];
 197 struct lr_parent_child_list *parents;
 198
 199 FILE *debug_log;
 200
 201 /* Command line options */
 202 struct option long_opts[] = {
 203         { .val = 'l',   .name = "statuslog",    .has_arg = required_argument },
 204         { .val = 'm',   .name = "mdt",          .has_arg = required_argument },
 205         { .val = 's',   .name = "source",       .has_arg = required_argument },
 206         { .val = 't',   .name = "target",       .has_arg = required_argument },
 207         { .val = 'u',   .name = "user",         .has_arg = required_argument },
 208         { .val = 'v',   .name = "verbose",      .has_arg = no_argument },
 209         { .val = 'x',   .name = "xattr",        .has_arg = required_argument },
 210         { .val = 'z',   .name = "dry-run",      .has_arg = no_argument },
 211         /* Undocumented options follow */
 212         { .val = 'a',   .name = "abort-on-err", .has_arg = no_argument },
 213         { .val = 'c',   .name = "cl-clear",     .has_arg = required_argument },
 214         { .val = 'd',   .name = "debug",        .has_arg = required_argument },
 215         { .val = 'D',   .name = "debuglog",     .has_arg = required_argument },
 216         { .val = 'n',   .name = "start-recno",  .has_arg = required_argument },
 217         { .val = 'r',   .name = "use-rsync",    .has_arg = no_argument },
 218         { .val = 'y',   .name = "rsync-threshold",
 219                                                 .has_arg = required_argument },
 220         { .name = NULL } };
 221
 222 /* Command line usage */
 223 void lr_usage()
 224 {
 225         fprintf(stderr, "\tlustre_rsync -s <lustre_root_path> -t <target_path> "
 226                 "-m <mdt> -r <user id> -l <status log>\n"
 227                 "lustre_rsync can also pick up parameters from a "
 228                 "status log created earlier.\n"
 229                 "\tlustre_rsync -l <log_file>\n"
 230                 "options:\n"
 231                 "\t--xattr <yes|no> replicate EAs\n"
 232                 "\t--abort-on-err   abort at first err\n"
 233                 "\t--verbose\n"
 234                 "\t--dry-run        don't write anything\n");
 235 }
 236
 237 /* Print debug information. This is controlled by the value of the
 238    global variable 'debug' */
 239 void lr_debug(int level, const char *fmt, ...)
 240 {
 241         va_list ap;
 242
 243         if (level > debug)
 244                 return;
 245
 246         va_start(ap, fmt);
 247         if (debug_log != NULL)
 248                 vfprintf(debug_log, fmt, ap);
 249         else
 250                 vfprintf(stdout, fmt, ap);
 251         va_end(ap);
 252 }
 253
 254
 255 void * lr_grow_buf(void *buf, int size)
 256 {
 257         void *ptr;
 258
 259         ptr = realloc(buf, size);
 260         if (ptr == NULL)
 261                 free(buf);
 262         return ptr;
 263 }
 264
 265
 266 /* Use rsync to replicate file data */
 267 int lr_rsync_data(struct lr_info *info)
 268 {
 269         struct stat st_src, st_dest;
 270         int rc;
 271
 272         lr_debug(DTRACE, "Syncing data%s\n", info->tfid);
 273
 274         rc = stat(info->src, &st_src);
 275         if (rc == -1) {
 276                 fprintf(stderr, "Error: Unable to stat src=%s %s\n",
 277                         info->src, info->name);
 278                 if (errno == ENOENT)
 279                         return 0;
 280                 else
 281                         return -errno;
 282         }
 283         rc = stat(info->dest, &st_dest);
 284         if (rc == -1) {
 285                 fprintf(stderr, "Error: Unable to stat dest=%s\n",
 286                         info->dest);
 287                 return -errno;
 288         }
 289
 290         if (st_src.st_mtime != st_dest.st_mtime ||
 291             st_src.st_size != st_dest.st_size) {
 292                 /* XXX spawning off an rsync for every data sync and
 293                  * waiting synchronously is bad for performance.
 294                  * librsync could possibly used here. But it does not
 295                  * seem to be of production grade. Multi-threaded
 296                  * replication is also to be considered.
 297                  */
 298                 char *args[] = {
 299                         rsync,
 300                         "--inplace",
 301                         "--",
 302                         info->src,
 303                         info->dest,
 304                         NULL,
 305                 };
 306                 extern char **environ;
 307                 int status;
 308
 309                 lr_debug(DTRACE, "\t%s %s %s %s %s %s\n", args[0], args[1],
 310                          args[2], args[3], args[4], info->tfid);
 311
 312                 status = callvpe(rsync, args, environ);
 313                 if (status < 0) {
 314                         rc = -errno;
 315                 } else if (WIFEXITED(status)) {
 316                         status = WEXITSTATUS(status);
 317                         if (!status)
 318                                 rc = 0;
 319                         else if (status == 23 || status == 24)
 320                                 /* Error due to vanished source files;
 321                                    Ignore this error*/
 322                                 rc = 0;
 323                         else
 324                                 rc = -EINVAL;
 325                         if (status)
 326                                 lr_debug(DINFO, "rsync %s exited with %d %d\n",
 327                                          info->src, status, rc);
 328                 } else {
 329                         rc = -EINTR;
 330                 }
 331         } else {
 332                 lr_debug(DTRACE, "Not syncing %s and %s %s\n", info->src,
 333                          info->dest, info->tfid);
 334         }
 335
 336         return rc;
 337 }
 338
 339 int lr_copy_data(struct lr_info *info)
 340 {
 341         int fd_src = -1;
 342         int fd_dest = -1;
 343         int bufsize;
 344         int rsize;
 345         int rc = 0;
 346         struct stat st_src;
 347         struct stat st_dest;
 348
 349         fd_src = open(info->src, O_RDONLY);
 350         if (fd_src == -1)
 351                 return -errno;
 352         if (fstat(fd_src, &st_src) == -1 ||
 353             stat(info->dest, &st_dest) == -1)
 354                 goto out;
 355
 356         if (st_src.st_mtime == st_dest.st_mtime &&
 357             st_src.st_size == st_dest.st_size)
 358                 goto out;
 359
 360         if (st_src.st_size > rsync_threshold && rsync[0] != '\0') {
 361                 /* It is more efficient to use rsync to replicate
 362                    large files. Any file larger than rsync_threshold
 363                    is handed off to rsync. */
 364                 lr_debug(DTRACE, "Using rsync to replicate %s\n", info->tfid);
 365                 rc = lr_rsync_data(info);
 366                 goto out;
 367         }
 368
 369         fd_dest = open(info->dest, O_WRONLY | O_TRUNC, st_src.st_mode);
 370         if (fd_dest == -1) {
 371                 rc = -errno;
 372                 goto out;
 373         }
 374         bufsize = st_dest.st_blksize;
 375
 376         if (info->bufsize < bufsize) {
 377                 /* Grow buffer */
 378                 info->buf = lr_grow_buf(info->buf, bufsize);
 379                 if (info->buf == NULL) {
 380                         rc = -ENOMEM;
 381                         goto out;
 382                 }
 383                 info->bufsize = bufsize;
 384         }
 385
 386         while (1) {
 387                 char *buf;
 388                 int wsize;
 389
 390                 buf = info->buf;
 391                 rsize = read(fd_src, buf, bufsize);
 392                 if (rsize == 0) {
 393                         rc = 0;
 394                         break;
 395                 }
 396                 if (rsize < 0) {
 397                         rc = -errno;
 398                         break;
 399                 }
 400                 do {
 401                         wsize = write(fd_dest, buf, rsize);
 402                         if (wsize <= 0) {
 403                                 rc = -errno;
 404                                 break;
 405                         }
 406                         rsize -= wsize;
 407                         buf += wsize;
 408                 } while (rsize > 0);
 409         }
 410         fsync(fd_dest);
 411
 412 out:
 413         if (fd_src != -1)
 414                 close(fd_src);
 415         if (fd_dest != -1)
 416                 close(fd_dest);
 417
 418         return rc;
 419 }
 420
 421 /* Copy data from source to destination */
 422 int lr_sync_data(struct lr_info *info)
 423 {
 424         if (use_rsync)
 425                 return lr_rsync_data(info);
 426         else
 427                 return lr_copy_data(info);
 428 }
 429
 430 /* Copy all attributes from file src to file dest */
 431 int lr_copy_attr(char *src, char *dest)
 432 {
 433         struct stat st;
 434         struct utimbuf time;
 435
 436         if (stat(src, &st) == -1 ||
 437             chmod(dest, st.st_mode) == -1 ||
 438             chown(dest, st.st_uid, st.st_gid) == -1)
 439                 return -errno;
 440
 441         time.actime = st.st_atime;
 442         time.modtime = st.st_mtime;
 443         if (utime(dest, &time) == -1)
 444                 return -errno;
 445         return 0;
 446 }
 447
 448 /* Copy all xattrs from file info->src to info->dest */
 449 int lr_copy_xattr(struct lr_info *info)
 450 {
 451         size_t size = info->xsize;
 452         int start;
 453         int len;
 454         int rc;
 455
 456         if (noxattr)
 457                 return 0;
 458
 459         errno = 0;
 460         rc = llistxattr(info->src, info->xlist, size);
 461         lr_debug(DTRACE, "llistxattr(%s,%p) returned %d, errno=%d\n",
 462                  info->src, info->xlist, rc, errno);
 463         if ((rc > 0 && info->xlist == NULL) || errno == ERANGE) {
 464                 size = rc > PATH_MAX ? rc : PATH_MAX;
 465                 info->xlist = lr_grow_buf(info->xlist, size);
 466                 if (info->xlist == NULL)
 467                         return -ENOMEM;
 468                 info->xsize = size;
 469                 rc = llistxattr(info->src, info->xlist, size);
 470                 lr_debug(DTRACE, "llistxattr %s returned %d, errno=%d\n",
 471                          info->src, rc, errno);
 472         }
 473         if (rc < 0)
 474                 return rc;
 475
 476         len = rc;
 477         start = 0;
 478         while (start < len) {
 479                 size = info->xvsize;
 480                 rc = lgetxattr(info->src, info->xlist + start,
 481                                info->xvalue, size);
 482                 if (info->xvalue == NULL || errno == ERANGE) {
 483                         size = rc > PATH_MAX ? rc : PATH_MAX;
 484                         info->xvalue = lr_grow_buf(info->xvalue, size);
 485                         if (info->xvalue == NULL)
 486                                 return -ENOMEM;
 487                         info->xvsize = size;
 488                         rc = lgetxattr(info->src, info->xlist + start,
 489                                        info->xvalue, size);
 490                 }
 491                 lr_debug(DTRACE, "\t(%s,%d) rc=%p\n", info->xlist + start,
 492                          info->xvalue, rc);
 493                 if (rc > 0) {
 494                         size = rc;
 495                         rc = lsetxattr(info->dest, info->xlist + start,
 496                                        info->xvalue, size, 0);
 497                         lr_debug(DTRACE, "\tlsetxattr(), rc=%d, errno=%d\n",
 498                                  rc, errno);
 499                         if (rc == -1) {
 500                                 if (errno != ENOTSUP) {
 501                                         fprintf(stderr, "Error replicating "
 502                                                 " xattr for %s: %d\n",
 503                                                 info->dest, errno);
 504                                         errors++;
 505                                 }
 506                                 rc = 0;
 507                         }
 508                 }
 509                 start += strlen(info->xlist + start) + 1;
 510         }
 511
 512         lr_debug(DINFO, "setxattr: %s %s\n", info->src, info->dest);
 513
 514         return rc;
 515 }
 516
 517 /* Retrieve the filesystem path for a given FID and a given
 518    linkno. The path is returned in info->path */
 519 int lr_get_path_ln(struct lr_info *info, char *fidstr, int linkno)
 520 {
 521         long long recno = -1;
 522         int rc;
 523
 524         rc = llapi_fid2path(status->ls_source, fidstr, info->path,
 525                             PATH_MAX, &recno, &linkno);
 526         if (rc < 0 && rc != -ENOENT) {
 527                 fprintf(stderr, "fid2path error: (%s, %s) %d %s\n",
 528                         status->ls_source, fidstr, -rc, strerror(errno = -rc));
 529         }
 530
 531         return rc;
 532 }
 533
 534 /* Retrieve the filesystem path for a given FID. The path is returned
 535    in info->path */
 536 int lr_get_path(struct lr_info *info, char *fidstr)
 537 {
 538         return lr_get_path_ln(info, fidstr, 0);
 539 }
 540
 541 /* Generate the path for opening by FID */
 542 void lr_get_FID_PATH(char *mntpt, char *fidstr, char *buf, int bufsize)
 543 {
 544         /* Open-by-FID path is <mntpt>/.lustre/fid/[SEQ:OID:VER] */
 545         snprintf(buf, bufsize, "%s/%s/fid/%s", mntpt, dot_lustre_name,
 546                  fidstr);
 547         return;
 548 }
 549
 550 /* Read the symlink information into 'info->link' */
 551 int lr_get_symlink(struct lr_info *info)
 552 {
 553         int rc;
 554         char *link;
 555
 556         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
 557         rc = readlink(info->src, info->linktmp, PATH_MAX);
 558         if (rc > 0)
 559                 info->linktmp[rc] = '\0';
 560         else
 561                 return rc;
 562         lr_debug(DTRACE, "symlink: readlink returned %s\n", info->linktmp);
 563
 564         if (strncmp(info->linktmp, status->ls_source,
 565                     strlen(status->ls_source)) == 0) {
 566                 /* Strip source fs path and replace with target fs path. */
 567                 link = info->linktmp + strlen(status->ls_source);
 568                 snprintf(info->src, sizeof(info->src), "%s%s",
 569                          status->ls_targets[info->target_no], link);
 570                 link = info->src;
 571         } else {
 572                 link = info->linktmp;
 573         }
 574         rc = snprintf(info->link, sizeof(info->link), "%s", link);
 575         if (rc >= sizeof(info->link))
 576                 rc = -E2BIG;
 577         return rc;
 578 }
 579
 580 /* Create file/directory/device file/symlink. */
 581 int lr_mkfile(struct lr_info *info)
 582 {
 583         struct stat st;
 584         int rc = 0;
 585
 586         errno = 0;
 587         lr_debug(DINFO, "mkfile(%d) %s \n", info->type, info->dest);
 588         if (info->type == CL_MKDIR) {
 589                 rc = mkdir(info->dest, 0777);
 590         } else if (info->type == CL_SOFTLINK) {
 591                 lr_get_symlink(info);
 592                 rc = symlink(info->link, info->dest);
 593         } else if (info->type == CL_MKNOD) {
 594                 lr_get_FID_PATH(status->ls_source, info->tfid,
 595                                     info->src, PATH_MAX);
 596                 rc = stat(info->src, &st);
 597                 if (rc == -1) {
 598                         if (errno == ENOENT)
 599                                 return 0;
 600                         else
 601                                 return -errno;
 602                 }
 603                 rc = mknod(info->dest, st.st_mode, st.st_rdev);
 604         } else {
 605                 rc = mknod(info->dest, S_IFREG | 0777, 0);
 606         }
 607
 608         if (rc < 0) {
 609                 if (errno == EEXIST)
 610                         rc = 0;
 611                 else
 612                         return -errno;
 613         }
 614
 615         /* Sync data and attributes */
 616         if (info->type == CL_CREATE || info->type == CL_MKDIR) {
 617                 lr_debug(DTRACE, "Syncing data and attributes %s\n",
 618                          info->tfid);
 619                 (void) lr_copy_xattr(info);
 620                 if (info->type == CL_CREATE)
 621                         rc = lr_sync_data(info);
 622                 if (!rc)
 623                         rc = lr_copy_attr(info->src, info->dest);
 624
 625                 if (rc == -ENOENT)
 626                         /* Source file has disappeared. Not an error. */
 627                         rc = 0;
 628         } else {
 629                 lr_debug(DTRACE, "Not syncing data and attributes %s\n",
 630                          info->tfid);
 631         }
 632
 633         return rc;
 634 }
 635
 636 int lr_add_pc(const char *pfid, const char *tfid, const char *name)
 637 {
 638         struct lr_parent_child_list *p;
 639         size_t len;
 640
 641         p = calloc(1, sizeof(*p));
 642         if (p == NULL)
 643                 return -ENOMEM;
 644         len = snprintf(p->pc_log.pcl_pfid, sizeof(p->pc_log.pcl_pfid),
 645                        "%s", pfid);
 646         if (len >= sizeof(p->pc_log.pcl_pfid))
 647                 goto out_err;
 648         len = snprintf(p->pc_log.pcl_tfid, sizeof(p->pc_log.pcl_tfid),
 649                        "%s", tfid);
 650         if (len >= sizeof(p->pc_log.pcl_tfid))
 651                 goto out_err;
 652         len = snprintf(p->pc_log.pcl_name, sizeof(p->pc_log.pcl_name),
 653                        "%s", name);
 654         if (len >= sizeof(p->pc_log.pcl_name))
 655                 goto out_err;
 656
 657         p->pc_next = parents;
 658         parents = p;
 659         return 0;
 660
 661 out_err:
 662         free(p);
 663         return -E2BIG;
 664 }
 665
 666 void lr_cascade_move(const char *fid, const char *dest, struct lr_info *info)
 667 {
 668         struct lr_parent_child_list *curr, *prev;
 669         char d[4 * PATH_MAX + 1];
 670         int rc;
 671
 672         prev = curr = parents;
 673         while (curr) {
 674                 if (strcmp(curr->pc_log.pcl_pfid, fid) == 0) {
 675                         if (snprintf(d, sizeof(d), "%s/%s", dest,
 676                                      curr->pc_log.pcl_name) >= sizeof(d)) {
 677                                 fprintf(stderr, "Buffer truncated\n");
 678                                 return;
 679                         }
 680                         if (snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 681                                      status->ls_targets[info->target_no],
 682                                      SPECIAL_DIR, curr->pc_log.pcl_tfid) >=
 683                             sizeof(info->src))
 684                                 return;
 685                         rc = rename(info->src, d);
 686                         if (rc == -1) {
 687                                 fprintf(stderr, "Error renaming file "
 688                                         " %s to %s: %d\n",
 689                                         info->src, d, errno);
 690                                 errors++;
 691                         }
 692                         if (curr == parents)
 693                                 parents = curr->pc_next;
 694                         else
 695                                 prev->pc_next = curr->pc_next;
 696                         lr_cascade_move(curr->pc_log.pcl_tfid, d, info);
 697                         free(curr);
 698                         prev = curr = parents;
 699
 700                 } else {
 701                         prev = curr;
 702                         curr = curr->pc_next;
 703                 }
 704         }
 705 }
 706
 707 /* remove [info->spfid, info->sfid] from parents */
 708 int lr_remove_pc(const char *pfid, const char *tfid)
 709 {
 710         struct lr_parent_child_list *curr, *prev;
 711
 712         for (prev = curr = parents; curr; prev = curr, curr = curr->pc_next) {
 713                 if (strcmp(curr->pc_log.pcl_pfid, pfid) == 0 &&
 714                     strcmp(curr->pc_log.pcl_tfid, tfid) == 0) {
 715                         if (curr == parents)
 716                                 parents = curr->pc_next;
 717                         else
 718                                 prev->pc_next = curr->pc_next;
 719                         free(curr);
 720                         break;
 721                 }
 722         }
 723         return 0;
 724 }
 725
 726 /* Create file under SPECIAL_DIR with its tfid as its name. */
 727 int lr_mk_special(struct lr_info *info)
 728 {
 729         int rc;
 730
 731         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 732                 status->ls_targets[info->target_no], SPECIAL_DIR,
 733                 info->tfid);
 734
 735         rc = lr_mkfile(info);
 736         if (rc)
 737                 return rc;
 738
 739         rc = lr_add_pc(info->pfid, info->tfid, info->name);
 740         return rc;
 741 }
 742
 743 /* Remove a file or directory */
 744 int lr_rmfile(struct lr_info *info)
 745 {
 746         int rc;
 747
 748         if (info->type == CL_RMDIR)
 749                 rc = rmdir(info->dest);
 750         else
 751                 rc = unlink(info->dest);
 752         if (rc == -1)
 753                 rc = -errno;
 754         return rc;
 755 }
 756
 757 /* Recursively remove directory and its contents */
 758 int lr_rm_recursive(struct lr_info *info)
 759 {
 760         char *args[] = {
 761                 "rm",
 762                 "-rf",
 763                 "--",
 764                 info->dest,
 765                 NULL,
 766         };
 767         extern char **environ;
 768         int status;
 769         int rc;
 770
 771         status = callvpe("/bin/rm", args, environ);
 772         if (status < 0)
 773                 rc = -errno;
 774         else if (WIFEXITED(status))
 775                 rc = WEXITSTATUS(status) == 0 ? 0 : -EINVAL;
 776         else
 777                 rc = -EINTR;
 778
 779         return rc;
 780 }
 781
 782 /* Remove a file under SPECIAL_DIR with its tfid as its name. */
 783 int lr_rm_special(struct lr_info *info)
 784 {
 785         int rc;
 786
 787         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 788                  status->ls_targets[info->target_no], SPECIAL_DIR,
 789                  info->tfid);
 790         rc = lr_rmfile(info);
 791
 792         if (rc)
 793                 lr_debug(DINFO, "remove: %s; rc=%d, errno=%d\n",
 794                          info->dest, rc, errno);
 795         return rc;
 796 }
 797
 798 /* Replicate file and directory create events */
 799 int lr_create(struct lr_info *info)
 800 {
 801         int len;
 802         int rc1 = 0;
 803         int rc;
 804         int mkspecial = 0;
 805
 806         /* Is target FID present on the source? */
 807         rc = lr_get_path(info, info->tfid);
 808         if (rc == -ENOENT) {
 809                 /* Source file has disappeared. Not an error. */
 810                 lr_debug(DINFO, "create: tfid %s not found on"
 811                          "source-fs\n", info->tfid);
 812                 return 0;
 813         } else if (rc) {
 814                 return rc;
 815         }
 816         strcpy(info->savedpath, info->path);
 817
 818         /* Is parent FID present on the source */
 819         rc = lr_get_path(info, info->pfid);
 820         if (rc == -ENOENT) {
 821                 lr_debug(DINFO, "create: pfid %s not found on source-fs\n",
 822                          info->tfid);
 823                 mkspecial = 1;
 824         } else if (rc < 0) {
 825                 return rc;
 826         }
 827
 828         /* Is f2p(pfid)+name != f2p(tfid)? If not the file has moved. */
 829         len = strlen(info->path);
 830         if (len == 1 && info->path[0] == '/')
 831                 snprintf(info->dest, sizeof(info->dest), "%s", info->name);
 832         else if (len - 1 > 0 && info->path[len - 1] == '/')
 833                 snprintf(info->dest, sizeof(info->dest), "%s%s", info->path,
 834                          info->name);
 835         else
 836                 snprintf(info->dest, sizeof(info->dest), "%s/%s", info->path,
 837                          info->name);
 838
 839         lr_debug(DTRACE, "dest = %s; savedpath = %s\n", info->dest,
 840                  info->savedpath);
 841         if (strncmp(info->dest, info->savedpath, PATH_MAX) != 0) {
 842                 lr_debug(DTRACE, "create: file moved (%s). %s != %s\n",
 843                          info->tfid, info->dest, info->savedpath);
 844                 mkspecial = 1;
 845         }
 846
 847         /* Is f2p(pfid) present on the target? If not, the parent has
 848            moved */
 849         if (!mkspecial) {
 850                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
 851                          status->ls_targets[0], info->path);
 852                 if (access(info->dest, F_OK) != 0) {
 853                         lr_debug(DTRACE, "create: parent %s not found\n",
 854                                 info->dest);
 855                         mkspecial = 1;
 856                 }
 857         }
 858         for (info->target_no = 0; info->target_no < status->ls_num_targets;
 859              info->target_no++) {
 860                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
 861                         status->ls_targets[info->target_no], info->savedpath);
 862                 lr_get_FID_PATH(status->ls_source, info->tfid, info->src,
 863                                 PATH_MAX);
 864
 865                 if (!mkspecial)
 866                         rc1 = lr_mkfile(info);
 867                 if (mkspecial || rc1 == -ENOENT) {
 868                         rc1 = lr_mk_special(info);
 869                 }
 870                 if (rc1)
 871                         rc = rc1;
 872         }
 873         return rc;
 874 }
 875
 876 /* Replicate a file remove (rmdir/unlink) operation */
 877 int lr_remove(struct lr_info *info)
 878 {
 879         int rc = 0;
 880         int rc1;
 881
 882         for (info->target_no = 0; info->target_no < status->ls_num_targets;
 883              info->target_no++) {
 884
 885                 rc1 = lr_rm_special(info);
 886                 if (!rc1)
 887                         continue;
 888
 889                 rc1 = lr_get_path(info, info->pfid);
 890                 if (rc1 == -ENOENT) {
 891                         lr_debug(DINFO, "remove: pfid %s not found\n",
 892                                  info->pfid);
 893                         continue;
 894                 }
 895                 if (rc1) {
 896                         rc = rc1;
 897                         continue;
 898                 }
 899                 snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 900                         status->ls_targets[info->target_no], info->path,
 901                         info->name);
 902
 903                 rc1 = lr_rmfile(info);
 904                 lr_debug(DINFO, "remove: %s; rc1=%d, errno=%d\n",
 905                          info->dest, rc1, errno);
 906                 if (rc1 == -ENOTEMPTY)
 907                         rc1 = lr_rm_recursive(info);
 908
 909                 if (rc1) {
 910                         rc = rc1;
 911                         continue;
 912                 }
 913         }
 914         return rc;
 915 }
 916
 917 /* Replicate a rename/move operation. */
 918 int lr_move(struct lr_info *info)
 919 {
 920         int rc = 0;
 921         int rc1;
 922         int rc_dest, rc_src;
 923         int special_src = 0;
 924         int special_dest = 0;
 925         char srcpath[PATH_MAX + 1] = "";
 926
 927         assert(info->is_extended);
 928
 929         rc_src = lr_get_path(info, info->spfid);
 930         if (rc_src < 0 && rc_src != -ENOENT)
 931                 return rc_src;
 932         memcpy(srcpath, info->path, strlen(info->path));
 933
 934         rc_dest = lr_get_path(info, info->pfid);
 935         if (rc_dest < 0 && rc_dest != -ENOENT)
 936                 return rc_dest;
 937
 938         for (info->target_no = 0; info->target_no < status->ls_num_targets;
 939              info->target_no++) {
 940
 941                 if (!rc_dest) {
 942                         snprintf(info->dest, sizeof(info->dest), "%s/%s",
 943                                 status->ls_targets[info->target_no],
 944                                 info->path);
 945                         if (access(info->dest, F_OK) != 0) {
 946                                 rc_dest = -errno;
 947                         } else {
 948                                 snprintf(info->dest, sizeof(info->dest),
 949                                          "%s/%s/%s",
 950                                          status->ls_targets[info->target_no],
 951                                          info->path, info->name);
 952                         }
 953                         lr_debug(DINFO, "dest path %s rc_dest=%d\n", info->dest,
 954                                  rc_dest);
 955                 }
 956                 if (rc_dest == -ENOENT) {
 957                         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 958                                 status->ls_targets[info->target_no],
 959                                 SPECIAL_DIR, info->sfid);
 960                         special_dest = 1;
 961                         lr_debug(DINFO, "special dest %s\n", info->dest);
 962                 }
 963
 964                 if (!rc_src) {
 965                         snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 966                                 status->ls_targets[info->target_no],
 967                                 srcpath, info->sname);
 968                         lr_debug(DINFO, "src path %s rc_src=%d\n", info->src,
 969                                  rc_src);
 970                 }
 971                 if (rc_src == -ENOENT || (access(info->src, F_OK) != 0 &&
 972                                           errno == ENOENT)) {
 973                         snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 974                                 status->ls_targets[info->target_no],
 975                                 SPECIAL_DIR, info->sfid);
 976                         special_src = 1;
 977                         lr_debug(DINFO, "special src %s\n", info->src);
 978                 }
 979
 980                 rc1 = 0;
 981                 errno = 0;
 982                 if (strcmp(info->src, info->dest) != 0) {
 983                         rc1 = rename(info->src, info->dest);
 984                         if (rc1 == -1)
 985                                 rc1 = -errno;
 986                         lr_debug(DINFO, "rename returns %d\n", rc1);
 987                 }
 988
 989                 if (special_src)
 990                         rc1 = lr_remove_pc(info->spfid, info->sfid);
 991
 992                 if (!special_dest)
 993                         lr_cascade_move(info->sfid, info->dest, info);
 994                 else
 995                         rc1 = lr_add_pc(info->pfid, info->sfid, info->name);
 996
 997                 lr_debug(DINFO, "move: %s [to] %s rc1=%d, errno=%d\n",
 998                          info->src, info->dest, rc1, errno);
 999                 if (rc1)
1000                         rc = rc1;
1001         }
1002         return rc;
1003 }
1004
1005 /* Replicate a hard link */
1006 int lr_link(struct lr_info *info)
1007 {
1008         int i;
1009         int rc;
1010         int rc1;
1011         struct stat st;
1012
1013         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1014         rc = stat(info->src, &st);
1015         if (rc == -1)
1016                 return -errno;
1017
1018         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1019              info->target_no++) {
1020
1021                 info->src[0] = 0;
1022                 info->dest[0] = 0;
1023                 rc1 = 0;
1024
1025                 /*
1026                  * The changelog record has the new parent directory FID and
1027                  * name of the target file. So info->dest can be constructed
1028                  * by getting the path of the new parent directory and
1029                  * appending the target file name.
1030                  */
1031                 rc1 = lr_get_path(info, info->pfid);
1032                 lr_debug(rc1 ? 0 : DTRACE, "\tparent fid2path %s, %s, rc=%d\n",
1033                          info->path, info->name, rc1);
1034
1035                 if (rc1 == 0) {
1036                         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1037                                  status->ls_targets[info->target_no],
1038                                  info->path, info->name);
1039                         lr_debug(DINFO, "link destination is %s\n", info->dest);
1040                 }
1041
1042                 /* Search through the hardlinks to get the src */
1043                 for (i = 0; i < st.st_nlink && info->src[0] == 0; i++) {
1044                         size_t len;
1045
1046                         rc1 = lr_get_path_ln(info, info->tfid, i);
1047                         lr_debug(rc1 ? 0:DTRACE, "\tfid2path %s, %s, %d rc=%d\n",
1048                                  info->path, info->name, i, rc1);
1049                         if (rc1)
1050                                 break;
1051
1052                         /*
1053                          * Compare the path of target FID with info->dest
1054                          * to find out info->src.
1055                          */
1056                         len = sizeof(status->ls_targets[info->target_no]) +
1057                               sizeof(info->path);
1058                         char srcpath[len + 1];
1059
1060                         snprintf(srcpath, sizeof(srcpath), "%s/%s",
1061                                  status->ls_targets[info->target_no],
1062                                  info->path);
1063
1064                         if (strcmp(srcpath, info->dest) != 0) {
1065                                 snprintf(info->src, sizeof(info->src), "%s",
1066                                          srcpath);
1067                                 lr_debug(DINFO, "link source is %s\n",
1068                                          info->src);
1069                         }
1070                 }
1071
1072                 if (rc1) {
1073                         rc = rc1;
1074                         continue;
1075                 }
1076
1077                 if (info->src[0] == 0)
1078                         snprintf(info->src, sizeof(info->src), "%s/%s/%s",
1079                                 status->ls_targets[info->target_no],
1080                                 SPECIAL_DIR, info->tfid);
1081                 else if (info->dest[0] == 0)
1082                         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1083                                 status->ls_targets[info->target_no],
1084                                 SPECIAL_DIR, info->tfid);
1085
1086                 rc1 = link(info->src, info->dest);
1087                 lr_debug(DINFO, "link: %s [to] %s; rc1=%d %s\n",
1088                          info->src, info->dest, rc1,
1089                          strerror(rc1 ? errno : 0));
1090
1091                 if (rc1)
1092                         rc = rc1;
1093         }
1094         return rc;
1095 }
1096
1097 /* Replicate file attributes */
1098 int lr_setattr(struct lr_info *info)
1099 {
1100         int rc1;
1101         int rc;
1102
1103         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1104
1105         rc = lr_get_path(info, info->tfid);
1106         if (rc == -ENOENT)
1107                 lr_debug(DINFO, "setattr: %s not present on source-fs\n",
1108                          info->src);
1109         if (rc)
1110                 return rc;
1111
1112         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1113              info->target_no++) {
1114
1115                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
1116                          status->ls_targets[info->target_no], info->path);
1117                 lr_debug(DINFO, "setattr: %s %s %s", info->src, info->dest,
1118                          info->tfid);
1119
1120                 rc1 = lr_sync_data(info);
1121                 if (!rc1)
1122                         rc1 = lr_copy_attr(info->src, info->dest);
1123                 if (rc1)
1124                         rc = rc1;
1125         }
1126         return rc;
1127 }
1128
1129 /* Replicate xattrs */
1130 int lr_setxattr(struct lr_info *info)
1131 {
1132         int rc, rc1;
1133
1134         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1135
1136         rc = lr_get_path(info, info->tfid);
1137         if (rc == -ENOENT)
1138                 lr_debug(DINFO, "setxattr: %s not present on source-fs\n",
1139                          info->src);
1140         if (rc)
1141                 return rc;
1142
1143         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1144              info->target_no++) {
1145
1146                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
1147                         status->ls_targets[info->target_no], info->path);
1148                 lr_debug(DINFO, "setxattr: %s %s %s\n", info->src, info->dest,
1149                          info->tfid);
1150
1151                 rc1 = lr_copy_xattr(info);
1152                 if (rc1)
1153                         rc = rc1;
1154         }
1155
1156         return rc;
1157 }
1158
1159 /* Parse a line of changelog entry */
1160 int lr_parse_line(void *priv, struct lr_info *info)
1161 {
1162         struct changelog_rec            *rec;
1163         struct changelog_ext_rename     *rnm;
1164         size_t                           namelen;
1165         size_t                           copylen = sizeof(info->name);
1166
1167         if (llapi_changelog_recv(priv, &rec) != 0)
1168                 return -1;
1169
1170         info->is_extended = !!(rec->cr_flags & CLF_RENAME);
1171         info->recno = rec->cr_index;
1172         info->type = rec->cr_type;
1173         snprintf(info->tfid, sizeof(info->tfid), DFID, PFID(&rec->cr_tfid));
1174         snprintf(info->pfid, sizeof(info->pfid), DFID, PFID(&rec->cr_pfid));
1175
1176         namelen = strnlen(changelog_rec_name(rec), rec->cr_namelen);
1177         if (copylen > namelen + 1)
1178                 copylen = namelen + 1;
1179         snprintf(info->name, copylen, "%s", changelog_rec_name(rec));
1180
1181         /* Don't use rnm if CLF_RENAME isn't set */
1182         rnm = changelog_rec_rename(rec);
1183         if (rec->cr_flags & CLF_RENAME && !fid_is_zero(&rnm->cr_sfid)) {
1184                 copylen = sizeof(info->sname);
1185
1186                 snprintf(info->sfid, sizeof(info->sfid), DFID,
1187                          PFID(&rnm->cr_sfid));
1188                 snprintf(info->spfid, sizeof(info->spfid), DFID,
1189                          PFID(&rnm->cr_spfid));
1190                 namelen = changelog_rec_snamelen(rec);
1191                 if (copylen > namelen + 1)
1192                         copylen = namelen + 1;
1193                 snprintf(info->sname, copylen, "%s", changelog_rec_sname(rec));
1194
1195                 if (verbose > 1)
1196                         printf("Rec %lld: %d %s %s\n", info->recno, info->type,
1197                                 info->name, info->sname);
1198         } else {
1199                 if (verbose > 1)
1200                         printf("Rec %lld: %d %s\n", info->recno, info->type,
1201                                 info->name);
1202         }
1203
1204         llapi_changelog_free(&rec);
1205
1206         rec_count++;
1207         return 0;
1208 }
1209
1210 /* Initialize the replication parameters */
1211 int lr_init_status()
1212 {
1213         size_t size = sizeof(struct lustre_rsync_status) + PATH_MAX + 1;
1214
1215         if (status != NULL)
1216                 return 0;
1217         status = calloc(size, 1);
1218         if (status == NULL)
1219                 return -ENOMEM;
1220         status->ls_version = REPLICATE_STATUS_VER;
1221         status->ls_size = size;
1222         status->ls_last_recno = -1;
1223         return 0;
1224 }
1225
1226 /* Make a backup of the statuslog */
1227 void lr_backup_log()
1228 {
1229         char backupfile[PATH_MAX];
1230
1231         if (logbackedup)
1232                 return;
1233         snprintf(backupfile, sizeof(backupfile), "%s.old", statuslog);
1234         (void) rename(statuslog, backupfile);
1235         logbackedup = 1;
1236
1237         return;
1238 }
1239
1240 /* Save replication parameters to a statuslog. */
1241 int lr_write_log()
1242 {
1243         int fd;
1244         size_t size;
1245         size_t write_size = status->ls_size;
1246         struct lr_parent_child_list *curr;
1247         int rc = 0;
1248
1249         if (statuslog == NULL)
1250                 return 0;
1251
1252         lr_backup_log();
1253
1254         fd = open(statuslog, O_WRONLY | O_CREAT | O_SYNC,
1255                              S_IRUSR | S_IWUSR);
1256         if (fd == -1) {
1257                 fprintf(stderr, "Error opening log file for writing (%s)\n",
1258                         statuslog);
1259                 return -1;
1260         }
1261         errno = 0;
1262         size = write(fd, status, write_size);
1263         if (size != write_size) {
1264                 fprintf(stderr, "Error writing to log file (%s) %d\n",
1265                         statuslog, errno);
1266                 close(fd);
1267                 return -1;
1268         }
1269
1270         for (curr = parents; curr; curr = curr->pc_next) {
1271                 size = write(fd, &curr->pc_log, sizeof(curr->pc_log));
1272                 if (size != sizeof(curr->pc_log)) {
1273                         fprintf(stderr, "Error writing to log file (%s) %d\n",
1274                                 statuslog, errno);
1275                         rc = -1;
1276                         break;
1277                 }
1278         }
1279         close(fd);
1280         return rc;
1281 }
1282
1283 /* Read statuslog and populate the replication parameters.  Command
1284  * line parameters take precedence over parameters in the log file.*/
1285 int lr_read_log()
1286 {
1287         struct lr_parent_child_list *tmp;
1288         struct lr_parent_child_log rec;
1289         struct lustre_rsync_status *s;
1290         int fd = -1;
1291         size_t size;
1292         size_t read_size = sizeof(struct lustre_rsync_status) + PATH_MAX + 1;
1293         int rc = 0;
1294
1295         if (statuslog == NULL)
1296                 return 0;
1297
1298         s = calloc(1, read_size);
1299         if (s == NULL) {
1300                 rc = -ENOMEM;
1301                 goto out;
1302         }
1303
1304         fd = open(statuslog, O_RDONLY);
1305         if (fd == -1) {
1306                 rc = -errno;
1307                 goto out;
1308         }
1309
1310         size = read(fd, s, read_size);
1311         if (size != read_size) {
1312                 rc = -EINVAL;
1313                 goto out;
1314         }
1315
1316         if (read_size < s->ls_size) {
1317                 read_size = s->ls_size;
1318                 s = lr_grow_buf(s, read_size);
1319                 if (s == NULL) {
1320                         rc = -ENOMEM;
1321                         goto out;
1322                 }
1323
1324                 if (lseek(fd, 0, SEEK_SET) == -1) {
1325                         rc = -ENOMEM;
1326                         goto out;
1327                 }
1328
1329                 size = read(fd, s, read_size);
1330                 if (size != read_size) {
1331                         rc = -EINVAL;
1332                         goto out;
1333                 }
1334         }
1335
1336         while (read(fd, &rec, sizeof(rec)) != 0) {
1337                 tmp = calloc(1, sizeof(*tmp));
1338                 if (!tmp) {
1339                         rc = -ENOMEM;
1340                         goto out;
1341                 }
1342
1343                 tmp->pc_log = rec;
1344                 tmp->pc_next = parents;
1345                 parents = tmp;
1346         }
1347
1348         /* copy uninitialized fields to status */
1349         if (status->ls_num_targets == 0) {
1350                 if (status->ls_size != s->ls_size) {
1351                         status = lr_grow_buf(status, s->ls_size);
1352                         if (status == NULL) {
1353                                 rc = -ENOMEM;
1354                                 goto out;
1355                         }
1356
1357                         status->ls_size = s->ls_size;
1358                 }
1359                 status->ls_num_targets = s->ls_num_targets;
1360                 memcpy(status->ls_targets, s->ls_targets,
1361                        (PATH_MAX + 1) * s->ls_num_targets);
1362         }
1363         if (status->ls_last_recno == -1)
1364                 status->ls_last_recno = s->ls_last_recno;
1365
1366         if (status->ls_registration[0] == '\0')
1367                 snprintf(status->ls_registration,
1368                          sizeof(status->ls_registration), "%s",
1369                          s->ls_registration);
1370
1371         if (status->ls_mdt_device[0] == '\0')
1372                 snprintf(status->ls_mdt_device,
1373                          sizeof(status->ls_mdt_device), "%s",
1374                          s->ls_mdt_device);
1375
1376         if (status->ls_source_fs[0] == '\0')
1377                 snprintf(status->ls_source_fs,
1378                          sizeof(status->ls_source_fs), "%s",
1379                          s->ls_source_fs);
1380
1381         if (status->ls_source[0] == '\0')
1382                 snprintf(status->ls_source,
1383                          sizeof(status->ls_source), "%s",
1384                          s->ls_source);
1385
1386  out:
1387         if (fd != -1)
1388                 close(fd);
1389         if (s)
1390                 free(s);
1391         return rc;
1392 }
1393
1394 /* Clear changelogs every CLEAR_INTERVAL records or at the end of
1395    processing. */
1396 int lr_clear_cl(struct lr_info *info, int force)
1397 {
1398         char            mdt_device[LR_NAME_MAXLEN + 1];
1399         long long       rec;
1400         int             rc = 0;
1401
1402         if (force || info->recno > status->ls_last_recno + CLEAR_INTERVAL) {
1403                 if (info->type == CL_RENAME)
1404                         rec = info->recno + 1;
1405                 else
1406                         rec = info->recno;
1407                 if (!noclear && !dryrun) {
1408                         /* llapi_changelog_clear modifies the mdt
1409                          * device name so make a copy of it until this
1410                          * is fixed.
1411                          */
1412                         snprintf(mdt_device, sizeof(mdt_device), "%s",
1413                                  status->ls_mdt_device);
1414                         rc = llapi_changelog_clear(mdt_device,
1415                                                    status->ls_registration,
1416                                                    rec);
1417                         if (rc)
1418                                 printf("Changelog clear (%s, %s, %lld) "
1419                                        "returned %d\n", status->ls_mdt_device,
1420                                        status->ls_registration, rec, rc);
1421                 }
1422                 if (!rc && !dryrun) {
1423                         status->ls_last_recno = rec;
1424                         lr_write_log();
1425
1426                 }
1427         }
1428
1429         return rc;
1430 }
1431
1432 /* Locate a usable version of rsync. At this point we'll use any
1433    version. */
1434 int lr_locate_rsync()
1435 {
1436         FILE *fp;
1437         int len;
1438
1439         /* Locate rsync */
1440         snprintf(rsync, sizeof(rsync), "%s -p %s", TYPE, RSYNC);
1441         fp = popen(rsync, "r");
1442         if (fp == NULL)
1443                 return -1;
1444
1445         if (fgets(rsync, sizeof(rsync), fp) == NULL) {
1446                 fclose(fp);
1447                 return -1;
1448         }
1449
1450         len = strlen(rsync);
1451         if (len > 0 && rsync[len - 1] == '\n')
1452                 rsync[len - 1] = '\0';
1453         fclose(fp);
1454
1455         /* Determine the version of rsync */
1456         snprintf(rsync_ver, sizeof(rsync_ver), "%s --version", rsync);
1457         fp = popen(rsync_ver, "r");
1458         if (fp == NULL)
1459                 return -1;
1460
1461         if (fgets(rsync_ver, sizeof(rsync_ver), fp) == NULL) {
1462                 fclose(fp);
1463                 return -1;
1464         }
1465         len = strlen(rsync_ver);
1466         if (len > 0 && rsync_ver[len - 1] == '\n')
1467                 rsync_ver[len - 1] = '\0';
1468         fclose(fp);
1469
1470         return 0;
1471
1472 }
1473
1474 /* Print the replication parameters */
1475 void lr_print_status(struct lr_info *info)
1476 {
1477         int i;
1478
1479         if (!verbose)
1480                 return;
1481
1482         printf("Lustre filesystem: %s\n", status->ls_source_fs);
1483         printf("MDT device: %s\n", status->ls_mdt_device);
1484         printf("Source: %s\n", status->ls_source);
1485         for (i = 0; i < status->ls_num_targets; i++)
1486                 printf("Target: %s\n", status->ls_targets[i]);
1487         if (statuslog != NULL)
1488                 printf("Statuslog: %s\n", statuslog);
1489         printf("Changelog registration: %s\n", status->ls_registration);
1490         printf("Starting changelog record: %jd\n",
1491                (uintmax_t)status->ls_last_recno);
1492         if (noxattr)
1493                 printf("Replicate xattrs: no\n");
1494         if (noclear)
1495                 printf("Clear changelog after use: no\n");
1496         if (use_rsync)
1497                 printf("Using rsync: %s (%s)\n", rsync, rsync_ver);
1498 }
1499
1500 void lr_print_failure(struct lr_info *info, int rc)
1501 {
1502         fprintf(stderr, "Replication of operation failed(%d):"
1503                 " %lld %s (%d) %s %s %s\n", rc, info->recno,
1504                 changelog_type2str(info->type), info->type, info->tfid,
1505                 info->pfid, info->name);
1506 }
1507
1508 /* Replicate filesystem operations from src_path to target_path */
1509 int lr_replicate()
1510 {
1511         void *changelog_priv;
1512         struct lr_info *info;
1513         struct lr_info *ext = NULL;
1514         time_t start;
1515         int xattr_not_supp;
1516         int i;
1517         int rc;
1518
1519         start = time(NULL);
1520
1521         info = calloc(1, sizeof(struct lr_info));
1522         if (info == NULL)
1523                 return -ENOMEM;
1524
1525         rc = llapi_search_fsname(status->ls_source, status->ls_source_fs);
1526         if (rc) {
1527                 fprintf(stderr, "Source path is not a valid Lustre client "
1528                         "mountpoint.\n");
1529                 goto out;
1530         }
1531
1532         if (status->ls_mdt_device[0] == '\0') {
1533                 int len;
1534
1535                 len = snprintf(status->ls_mdt_device,
1536                                sizeof(status->ls_mdt_device), "%s%s",
1537                                status->ls_source_fs, DEFAULT_MDT);
1538                 if (len >= sizeof(status->ls_mdt_device)) {
1539                         rc = -E2BIG;
1540                         goto out;
1541                 }
1542         }
1543
1544         ext = calloc(1, sizeof(struct lr_info));
1545         if (ext == NULL) {
1546                 rc = -ENOMEM;
1547                 goto out;
1548         }
1549
1550         for (i = 0, xattr_not_supp = 0; i < status->ls_num_targets; i++) {
1551                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
1552                          status->ls_targets[i], SPECIAL_DIR);
1553                 rc = mkdir(info->dest, 0777);
1554                 if (rc == -1 && errno != EEXIST) {
1555                         fprintf(stderr, "Error writing to target path %s.\n",
1556                                 status->ls_targets[i]);
1557                         rc = -errno;
1558                         goto out;
1559                 }
1560                 rc = llistxattr(info->src, info->xlist, info->xsize);
1561                 if (rc == -1 && errno == ENOTSUP) {
1562                         fprintf(stderr, "xattrs not supported on %s\n",
1563                                 status->ls_targets[i]);
1564                         xattr_not_supp++;
1565                 }
1566         }
1567         if (xattr_not_supp == status->ls_num_targets)
1568                 /* None of the targets support xattrs. */
1569                 noxattr = 1;
1570
1571         lr_print_status(info);
1572
1573         /* Open changelogs for consumption*/
1574         rc = llapi_changelog_start(&changelog_priv,
1575                                 CHANGELOG_FLAG_BLOCK |
1576                                 CHANGELOG_FLAG_JOBID |
1577                                 CHANGELOG_FLAG_EXTRA_FLAGS,
1578                                 status->ls_mdt_device, status->ls_last_recno);
1579         if (rc < 0) {
1580                 fprintf(stderr, "Error opening changelog file for fs %s.\n",
1581                         status->ls_source_fs);
1582                 goto out;
1583         }
1584
1585         rc = llapi_changelog_set_xflags(changelog_priv,
1586                                         CHANGELOG_EXTRA_FLAG_UIDGID |
1587                                         CHANGELOG_EXTRA_FLAG_NID |
1588                                         CHANGELOG_EXTRA_FLAG_OMODE |
1589                                         CHANGELOG_EXTRA_FLAG_XATTR);
1590         if (rc < 0) {
1591                 fprintf(stderr, "Error setting xflag in changelog for fs %s.\n",
1592                         status->ls_source_fs);
1593                 goto out;
1594         }
1595
1596         while (!quit && lr_parse_line(changelog_priv, info) == 0) {
1597                 rc = 0;
1598                 if (info->type == CL_RENAME && !info->is_extended) {
1599                         /* Newer rename operations extends changelog to store
1600                          * source file information, but old changelog has
1601                          * another record.
1602                          */
1603                         if (lr_parse_line(changelog_priv, ext) != 0)
1604                                 break;
1605                         memcpy(info->sfid, info->tfid, sizeof(info->sfid));
1606                         memcpy(info->spfid, info->pfid, sizeof(info->spfid));
1607                         memcpy(info->tfid, ext->tfid, sizeof(info->tfid));
1608                         memcpy(info->pfid, ext->pfid, sizeof(info->pfid));
1609                         snprintf(info->sname, sizeof(info->sname), "%s",
1610                                  info->name);
1611                         snprintf(info->name, sizeof(info->name), "%s",
1612                                  ext->name);
1613                         info->is_extended = 1;
1614                 }
1615
1616                 if (dryrun)
1617                         continue;
1618
1619                 lr_debug(DTRACE, "***** Start %lld %s (%d) %s %s %s *****\n",
1620                          info->recno, changelog_type2str(info->type),
1621                          info->type, info->tfid, info->pfid, info->name);
1622
1623                 switch(info->type) {
1624                 case CL_CREATE:
1625                 case CL_MKDIR:
1626                 case CL_MKNOD:
1627                 case CL_SOFTLINK:
1628                         rc = lr_create(info);
1629                         break;
1630                 case CL_RMDIR:
1631                 case CL_UNLINK:
1632                         rc = lr_remove(info);
1633                         break;
1634                 case CL_RENAME:
1635                         rc = lr_move(info);
1636                         break;
1637                 case CL_HARDLINK:
1638                         rc = lr_link(info);
1639                         break;
1640                 case CL_TRUNC:
1641                 case CL_SETATTR:
1642                         rc = lr_setattr(info);
1643                         break;
1644                 case CL_SETXATTR:
1645                         rc = lr_setxattr(info);
1646                         break;
1647                 case CL_CLOSE:
1648                 case CL_EXT:
1649                 case CL_OPEN:
1650                 case CL_GETXATTR:
1651                 case CL_DN_OPEN:
1652                 case CL_LAYOUT:
1653                 case CL_MARK:
1654                         /* Nothing needs to be done for these entries */
1655                         /* fallthrough */
1656                 default:
1657                         break;
1658                 }
1659
1660                 lr_debug(DTRACE, "##### End %lld %s (%d) %s %s %s rc=%d #####\n",
1661                          info->recno, changelog_type2str(info->type),
1662                          info->type, info->tfid, info->pfid, info->name, rc);
1663
1664                 if (rc && rc != -ENOENT) {
1665                         lr_print_failure(info, rc);
1666                         errors++;
1667                         if (abort_on_err)
1668                                 break;
1669                 }
1670                 lr_clear_cl(info, 0);
1671                 if (debug) {
1672                         bzero(info, sizeof(struct lr_info));
1673                         bzero(ext, sizeof(struct lr_info));
1674                 }
1675         }
1676
1677         llapi_changelog_fini(&changelog_priv);
1678
1679         if (errors || verbose)
1680                 printf("Errors: %d\n", errors);
1681
1682         /* Clear changelog records used so far */
1683         lr_clear_cl(info, 1);
1684
1685         if (verbose) {
1686                 printf("lustre_rsync took %ld seconds\n", time(NULL) - start);
1687                 printf("Changelog records consumed: %lld\n", rec_count);
1688         }
1689
1690         rc = 0;
1691
1692 out:
1693         if (info != NULL)
1694                 free(info);
1695         if (ext != NULL)
1696                 free(ext);
1697
1698         return rc;
1699 }
1700
1701 void
1702 termination_handler (int signum)
1703 {
1704         /* Set a flag for the replicator to gracefully shutdown */
1705         quit = 1;
1706         printf("lustre_rsync halting.\n");
1707 }
1708
1709 int main(int argc, char *argv[])
1710 {
1711         int newsize;
1712         int numtargets = 0;
1713         int rc = 0;
1714
1715         if ((rc = lr_init_status()) != 0)
1716                 return rc;
1717
1718         while ((rc = getopt_long(argc, argv, "as:t:m:u:l:vx:zc:ry:n:d:D:",
1719                                  long_opts, NULL)) >= 0) {
1720                 switch (rc) {
1721                 case 'a':
1722                         /* Assume absolute paths */
1723                         abort_on_err++;
1724                         break;
1725                 case 's':
1726                         /* Assume absolute paths */
1727                         snprintf(status->ls_source, sizeof(status->ls_source),
1728                                  "%s", optarg);
1729                         break;
1730                 case 't':
1731                         status->ls_num_targets++;
1732                         numtargets++;
1733                         if (numtargets != status->ls_num_targets) {
1734                                 /* Targets were read from a log
1735                                    file. The ones specified on the
1736                                    command line take precedence. The
1737                                    ones from the log file will be
1738                                    ignored. */
1739                                 status->ls_num_targets = numtargets;
1740                         }
1741                         newsize = sizeof (struct lustre_rsync_status) +
1742                                 (status->ls_num_targets * (PATH_MAX + 1));
1743                         if (status->ls_size != newsize) {
1744                                 status->ls_size = newsize;
1745                                 status = lr_grow_buf(status, newsize);
1746                                 if (status == NULL)
1747                                         return -ENOMEM;
1748                         }
1749                         snprintf(status->ls_targets[status->ls_num_targets - 1],
1750                                  sizeof(status->ls_targets[0]), "%s", optarg);
1751                         break;
1752                 case 'm':
1753                         snprintf(status->ls_mdt_device,
1754                                  sizeof(status->ls_mdt_device),
1755                                  "%s", optarg);
1756                         break;
1757                 case 'u':
1758                         snprintf(status->ls_registration,
1759                                  sizeof(status->ls_registration),
1760                                  "%s", optarg);
1761                         break;
1762                 case 'l':
1763                         statuslog = optarg;
1764                         (void) lr_read_log();
1765                         break;
1766                 case 'v':
1767                         verbose++;
1768                         break;
1769                 case 'x':
1770                         if (strcmp("no", optarg) == 0) {
1771                                 noxattr = 1;
1772                         } else if (strcmp("yes", optarg) != 0) {
1773                                 printf("Invalid parameter %s. "
1774                                        "Specify --xattr=no or --xattr=yes\n",
1775                                        optarg);
1776                                 return -1;
1777                         }
1778                         break;
1779                 case 'z':
1780                         dryrun = 1;
1781                         break;
1782                 case 'c':
1783                         /* Undocumented option cl-clear */
1784                         if (strcmp("no", optarg) == 0) {
1785                                 noclear = 1;
1786                         } else if (strcmp("yes", optarg) != 0) {
1787                                 printf("Invalid parameter %s. "
1788                                        "Specify --cl-clear=no "
1789                                        "or --cl-clear=yes\n",
1790                                        optarg);
1791                                 return -1;
1792                         }
1793                         break;
1794                 case 'r':
1795                         /* Undocumented option use-rsync */
1796                         use_rsync = 1;
1797                         break;
1798                 case 'y':
1799                         /* Undocumented option rsync-threshold */
1800                         rsync_threshold = atol(optarg);
1801                         break;
1802                 case 'n':
1803                         /* Undocumented option start-recno */
1804                         status->ls_last_recno = atol(optarg);
1805                         break;
1806                 case 'd':
1807                         /* Undocumented option debug */
1808                         debug = atoi(optarg);
1809                         if (debug < 0 || debug > 2)
1810                                 debug = 0;
1811                         break;
1812                 case 'D':
1813                         /* Undocumented option debug log file */
1814                         if (debug_log != NULL)
1815                                 fclose(debug_log);
1816                         debug_log = fopen(optarg, "a");
1817                         if (debug_log == NULL) {
1818                                 printf("Cannot open %s for debug log\n",
1819                                        optarg);
1820                                 return -1;
1821                         }
1822                         break;
1823                 default:
1824                         fprintf(stderr, "error: %s: option '%s' "
1825                                 "unrecognized.\n", argv[0], argv[optind - 1]);
1826                         lr_usage();
1827                         return -1;
1828                 }
1829         }
1830
1831         if (status->ls_last_recno == -1)
1832                 status->ls_last_recno = 0;
1833         if (strnlen(status->ls_registration, LR_NAME_MAXLEN) == 0) {
1834                 /* No registration ID was passed in. */
1835                 printf("Please specify changelog consumer registration id.\n");
1836                 lr_usage();
1837                 return -1;
1838         }
1839         if (strnlen(status->ls_source, PATH_MAX) == 0) {
1840                 fprintf(stderr, "Please specify the source path.\n");
1841                 lr_usage();
1842                 return -1;
1843         }
1844         if (strnlen(status->ls_targets[0], PATH_MAX) == 0) {
1845                 fprintf(stderr, "Please specify the target path.\n");
1846                 lr_usage();
1847                 return -1;
1848         }
1849
1850         rc = lr_locate_rsync();
1851         if (use_rsync && rc != 0) {
1852                 fprintf(stderr, "Error: unable to locate %s.\n", RSYNC);
1853                 exit(-1);
1854         }
1855
1856         signal(SIGINT, termination_handler);
1857         signal(SIGHUP, termination_handler);
1858         signal(SIGTERM, termination_handler);
1859
1860         rc = lr_replicate();
1861
1862         if (debug_log != NULL)
1863                 fclose(debug_log);
1864         return rc;
1865 }