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 + 128];
 196 char rsync_ver[PATH_MAX * 2];
 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(const char *src, const 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,
 502                         "cannot replicate xattrs from '%s' to '%s': %s\n",
 503                                                 info->src, info->dest,
 504                                                 strerror(errno));
 505                                         errors++;
 506                                 }
 507                                 rc = 0;
 508                         }
 509                 }
 510                 start += strlen(info->xlist + start) + 1;
 511         }
 512
 513         lr_debug(DINFO, "setxattr: %s %s\n", info->src, info->dest);
 514
 515         return rc;
 516 }
 517
 518 /* Retrieve the filesystem path for a given FID and a given
 519    linkno. The path is returned in info->path */
 520 int lr_get_path_ln(struct lr_info *info, char *fidstr, int linkno)
 521 {
 522         long long recno = -1;
 523         int rc;
 524
 525         rc = llapi_fid2path(status->ls_source, fidstr, info->path,
 526                             PATH_MAX, &recno, &linkno);
 527         if (rc < 0 && rc != -ENOENT) {
 528                 fprintf(stderr, "fid2path error: (%s, %s) %d %s\n",
 529                         status->ls_source, fidstr, -rc, strerror(errno = -rc));
 530         }
 531
 532         return rc;
 533 }
 534
 535 /* Retrieve the filesystem path for a given FID. The path is returned
 536    in info->path */
 537 int lr_get_path(struct lr_info *info, char *fidstr)
 538 {
 539         return lr_get_path_ln(info, fidstr, 0);
 540 }
 541
 542 /* Generate the path for opening by FID */
 543 void lr_get_FID_PATH(char *mntpt, char *fidstr, char *buf, int bufsize)
 544 {
 545         /* Open-by-FID path is <mntpt>/.lustre/fid/[SEQ:OID:VER] */
 546         snprintf(buf, bufsize, "%s/%s/fid/%s", mntpt, dot_lustre_name,
 547                  fidstr);
 548         return;
 549 }
 550
 551 /* Read the symlink information into 'info->link' */
 552 int lr_get_symlink(struct lr_info *info)
 553 {
 554         int rc;
 555         char *link;
 556
 557         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
 558         rc = readlink(info->src, info->linktmp, PATH_MAX);
 559         if (rc > 0)
 560                 info->linktmp[rc] = '\0';
 561         else
 562                 return rc;
 563         lr_debug(DTRACE, "symlink: readlink returned %s\n", info->linktmp);
 564
 565         if (strncmp(info->linktmp, status->ls_source,
 566                     strlen(status->ls_source)) == 0) {
 567                 /* Strip source fs path and replace with target fs path. */
 568                 link = info->linktmp + strlen(status->ls_source);
 569                 snprintf(info->src, sizeof(info->src), "%s%s",
 570                          status->ls_targets[info->target_no], link);
 571                 link = info->src;
 572         } else {
 573                 link = info->linktmp;
 574         }
 575         rc = snprintf(info->link, sizeof(info->link), "%s", link);
 576         if (rc >= sizeof(info->link))
 577                 rc = -E2BIG;
 578         return rc;
 579 }
 580
 581 /* Create file/directory/device file/symlink. */
 582 int lr_mkfile(struct lr_info *info)
 583 {
 584         struct stat st;
 585         int rc = 0;
 586
 587         errno = 0;
 588         lr_debug(DINFO, "mkfile(%d) %s \n", info->type, info->dest);
 589         if (info->type == CL_MKDIR) {
 590                 rc = mkdir(info->dest, 0777);
 591         } else if (info->type == CL_SOFTLINK) {
 592                 lr_get_symlink(info);
 593                 rc = symlink(info->link, info->dest);
 594         } else if (info->type == CL_MKNOD) {
 595                 lr_get_FID_PATH(status->ls_source, info->tfid,
 596                                     info->src, PATH_MAX);
 597                 rc = stat(info->src, &st);
 598                 if (rc == -1) {
 599                         if (errno == ENOENT)
 600                                 return 0;
 601                         else
 602                                 return -errno;
 603                 }
 604                 rc = mknod(info->dest, st.st_mode, st.st_rdev);
 605         } else {
 606                 rc = mknod(info->dest, S_IFREG | 0777, 0);
 607         }
 608
 609         if (rc < 0) {
 610                 if (errno == EEXIST)
 611                         rc = 0;
 612                 else
 613                         return -errno;
 614         }
 615
 616         /* Sync data and attributes */
 617         if (info->type == CL_CREATE || info->type == CL_MKDIR) {
 618                 lr_debug(DTRACE, "Syncing data and attributes %s\n",
 619                          info->tfid);
 620                 (void) lr_copy_xattr(info);
 621                 if (info->type == CL_CREATE)
 622                         rc = lr_sync_data(info);
 623                 if (!rc)
 624                         rc = lr_copy_attr(info->src, info->dest);
 625
 626                 if (rc == -ENOENT)
 627                         /* Source file has disappeared. Not an error. */
 628                         rc = 0;
 629         } else {
 630                 lr_debug(DTRACE, "Not syncing data and attributes %s\n",
 631                          info->tfid);
 632         }
 633
 634         return rc;
 635 }
 636
 637 int lr_add_pc(const char *pfid, const char *tfid, const char *name)
 638 {
 639         struct lr_parent_child_list *p;
 640         size_t len;
 641
 642         p = calloc(1, sizeof(*p));
 643         if (p == NULL)
 644                 return -ENOMEM;
 645         len = snprintf(p->pc_log.pcl_pfid, sizeof(p->pc_log.pcl_pfid),
 646                        "%s", pfid);
 647         if (len >= sizeof(p->pc_log.pcl_pfid))
 648                 goto out_err;
 649         len = snprintf(p->pc_log.pcl_tfid, sizeof(p->pc_log.pcl_tfid),
 650                        "%s", tfid);
 651         if (len >= sizeof(p->pc_log.pcl_tfid))
 652                 goto out_err;
 653         len = snprintf(p->pc_log.pcl_name, sizeof(p->pc_log.pcl_name),
 654                        "%s", name);
 655         if (len >= sizeof(p->pc_log.pcl_name))
 656                 goto out_err;
 657
 658         p->pc_next = parents;
 659         parents = p;
 660         return 0;
 661
 662 out_err:
 663         free(p);
 664         return -E2BIG;
 665 }
 666
 667 void lr_cascade_move(const char *fid, const char *dest, struct lr_info *info)
 668 {
 669         struct lr_parent_child_list *curr, *prev;
 670         char d[4 * PATH_MAX + 1];
 671         int rc;
 672
 673         prev = curr = parents;
 674         while (curr) {
 675                 if (strcmp(curr->pc_log.pcl_pfid, fid) == 0) {
 676                         if (snprintf(d, sizeof(d), "%s/%s", dest,
 677                                      curr->pc_log.pcl_name) >= sizeof(d)) {
 678                                 fprintf(stderr, "Buffer truncated\n");
 679                                 return;
 680                         }
 681                         if (snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 682                                      status->ls_targets[info->target_no],
 683                                      SPECIAL_DIR, curr->pc_log.pcl_tfid) >=
 684                             sizeof(info->src))
 685                                 return;
 686                         rc = rename(info->src, d);
 687                         if (rc == -1) {
 688                                 fprintf(stderr, "Error renaming file "
 689                                         " %s to %s: %d\n",
 690                                         info->src, d, errno);
 691                                 errors++;
 692                         }
 693                         if (curr == parents)
 694                                 parents = curr->pc_next;
 695                         else
 696                                 prev->pc_next = curr->pc_next;
 697                         lr_cascade_move(curr->pc_log.pcl_tfid, d, info);
 698                         free(curr);
 699                         prev = curr = parents;
 700
 701                 } else {
 702                         prev = curr;
 703                         curr = curr->pc_next;
 704                 }
 705         }
 706 }
 707
 708 /* remove [info->spfid, info->sfid] from parents */
 709 int lr_remove_pc(const char *pfid, const char *tfid)
 710 {
 711         struct lr_parent_child_list *curr, *prev;
 712
 713         for (prev = curr = parents; curr; prev = curr, curr = curr->pc_next) {
 714                 if (strcmp(curr->pc_log.pcl_pfid, pfid) == 0 &&
 715                     strcmp(curr->pc_log.pcl_tfid, tfid) == 0) {
 716                         if (curr == parents)
 717                                 parents = curr->pc_next;
 718                         else
 719                                 prev->pc_next = curr->pc_next;
 720                         free(curr);
 721                         break;
 722                 }
 723         }
 724         return 0;
 725 }
 726
 727 /* Create file under SPECIAL_DIR with its tfid as its name. */
 728 int lr_mk_special(struct lr_info *info)
 729 {
 730         int rc;
 731
 732         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 733                 status->ls_targets[info->target_no], SPECIAL_DIR,
 734                 info->tfid);
 735
 736         rc = lr_mkfile(info);
 737         if (rc)
 738                 return rc;
 739
 740         rc = lr_add_pc(info->pfid, info->tfid, info->name);
 741         return rc;
 742 }
 743
 744 /* Remove a file or directory */
 745 int lr_rmfile(struct lr_info *info)
 746 {
 747         int rc;
 748
 749         if (info->type == CL_RMDIR)
 750                 rc = rmdir(info->dest);
 751         else
 752                 rc = unlink(info->dest);
 753         if (rc == -1)
 754                 rc = -errno;
 755         return rc;
 756 }
 757
 758 /* Recursively remove directory and its contents */
 759 int lr_rm_recursive(struct lr_info *info)
 760 {
 761         char *args[] = {
 762                 "rm",
 763                 "-rf",
 764                 "--",
 765                 info->dest,
 766                 NULL,
 767         };
 768         extern char **environ;
 769         int status;
 770         int rc;
 771
 772         status = callvpe("/bin/rm", args, environ);
 773         if (status < 0)
 774                 rc = -errno;
 775         else if (WIFEXITED(status))
 776                 rc = WEXITSTATUS(status) == 0 ? 0 : -EINVAL;
 777         else
 778                 rc = -EINTR;
 779
 780         return rc;
 781 }
 782
 783 /* Remove a file under SPECIAL_DIR with its tfid as its name. */
 784 int lr_rm_special(struct lr_info *info)
 785 {
 786         int rc;
 787
 788         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 789                  status->ls_targets[info->target_no], SPECIAL_DIR,
 790                  info->tfid);
 791         rc = lr_rmfile(info);
 792
 793         if (rc)
 794                 lr_debug(DINFO, "remove: %s; rc=%d, errno=%d\n",
 795                          info->dest, rc, errno);
 796         return rc;
 797 }
 798
 799 /* Replicate file and directory create events */
 800 int lr_create(struct lr_info *info)
 801 {
 802         int len;
 803         int rc1 = 0;
 804         int rc;
 805         int mkspecial = 0;
 806
 807         /* Is target FID present on the source? */
 808         rc = lr_get_path(info, info->tfid);
 809         if (rc == -ENOENT) {
 810                 /* Source file has disappeared. Not an error. */
 811                 lr_debug(DINFO, "create: tfid %s not found on"
 812                          "source-fs\n", info->tfid);
 813                 return 0;
 814         } else if (rc) {
 815                 return rc;
 816         }
 817         strcpy(info->savedpath, info->path);
 818
 819         /* Is parent FID present on the source */
 820         rc = lr_get_path(info, info->pfid);
 821         if (rc == -ENOENT) {
 822                 lr_debug(DINFO, "create: pfid %s not found on source-fs\n",
 823                          info->tfid);
 824                 mkspecial = 1;
 825         } else if (rc < 0) {
 826                 return rc;
 827         }
 828
 829         /* Is f2p(pfid)+name != f2p(tfid)? If not the file has moved. */
 830         len = strlen(info->path);
 831         if (len == 1 && info->path[0] == '/')
 832                 snprintf(info->dest, sizeof(info->dest), "%s", info->name);
 833         else if (len - 1 > 0 && info->path[len - 1] == '/')
 834                 snprintf(info->dest, sizeof(info->dest), "%s%s", info->path,
 835                          info->name);
 836         else
 837                 snprintf(info->dest, sizeof(info->dest), "%s/%s", info->path,
 838                          info->name);
 839
 840         lr_debug(DTRACE, "dest = %s; savedpath = %s\n", info->dest,
 841                  info->savedpath);
 842         if (strncmp(info->dest, info->savedpath, PATH_MAX) != 0) {
 843                 lr_debug(DTRACE, "create: file moved (%s). %s != %s\n",
 844                          info->tfid, info->dest, info->savedpath);
 845                 mkspecial = 1;
 846         }
 847
 848         /* Is f2p(pfid) present on the target? If not, the parent has
 849            moved */
 850         if (!mkspecial) {
 851                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
 852                          status->ls_targets[0], info->path);
 853                 if (access(info->dest, F_OK) != 0) {
 854                         lr_debug(DTRACE, "create: parent %s not found\n",
 855                                 info->dest);
 856                         mkspecial = 1;
 857                 }
 858         }
 859         for (info->target_no = 0; info->target_no < status->ls_num_targets;
 860              info->target_no++) {
 861                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
 862                         status->ls_targets[info->target_no], info->savedpath);
 863                 lr_get_FID_PATH(status->ls_source, info->tfid, info->src,
 864                                 PATH_MAX);
 865
 866                 if (!mkspecial)
 867                         rc1 = lr_mkfile(info);
 868                 if (mkspecial || rc1 == -ENOENT) {
 869                         rc1 = lr_mk_special(info);
 870                 }
 871                 if (rc1)
 872                         rc = rc1;
 873         }
 874         return rc;
 875 }
 876
 877 /* Replicate a file remove (rmdir/unlink) operation */
 878 int lr_remove(struct lr_info *info)
 879 {
 880         int rc = 0;
 881         int rc1;
 882
 883         for (info->target_no = 0; info->target_no < status->ls_num_targets;
 884              info->target_no++) {
 885
 886                 rc1 = lr_rm_special(info);
 887                 if (!rc1)
 888                         continue;
 889
 890                 rc1 = lr_get_path(info, info->pfid);
 891                 if (rc1 == -ENOENT) {
 892                         lr_debug(DINFO, "remove: pfid %s not found\n",
 893                                  info->pfid);
 894                         continue;
 895                 }
 896                 if (rc1) {
 897                         rc = rc1;
 898                         continue;
 899                 }
 900                 snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 901                         status->ls_targets[info->target_no], info->path,
 902                         info->name);
 903
 904                 rc1 = lr_rmfile(info);
 905                 lr_debug(DINFO, "remove: %s; rc1=%d, errno=%d\n",
 906                          info->dest, rc1, errno);
 907                 if (rc1 == -ENOTEMPTY)
 908                         rc1 = lr_rm_recursive(info);
 909
 910                 if (rc1) {
 911                         rc = rc1;
 912                         continue;
 913                 }
 914         }
 915         return rc;
 916 }
 917
 918 /* Replicate a rename/move operation. */
 919 int lr_move(struct lr_info *info)
 920 {
 921         int rc = 0;
 922         int rc1;
 923         int rc_dest, rc_src;
 924         int special_src = 0;
 925         int special_dest = 0;
 926         char srcpath[PATH_MAX + 1] = "";
 927
 928         assert(info->is_extended);
 929
 930         rc_src = lr_get_path(info, info->spfid);
 931         if (rc_src < 0 && rc_src != -ENOENT)
 932                 return rc_src;
 933         memcpy(srcpath, info->path, strlen(info->path));
 934
 935         rc_dest = lr_get_path(info, info->pfid);
 936         if (rc_dest < 0 && rc_dest != -ENOENT)
 937                 return rc_dest;
 938
 939         for (info->target_no = 0; info->target_no < status->ls_num_targets;
 940              info->target_no++) {
 941
 942                 if (!rc_dest) {
 943                         snprintf(info->dest, sizeof(info->dest), "%s/%s",
 944                                 status->ls_targets[info->target_no],
 945                                 info->path);
 946                         if (access(info->dest, F_OK) != 0) {
 947                                 rc_dest = -errno;
 948                         } else {
 949                                 snprintf(info->dest, sizeof(info->dest),
 950                                          "%s/%s/%s",
 951                                          status->ls_targets[info->target_no],
 952                                          info->path, info->name);
 953                         }
 954                         lr_debug(DINFO, "dest path %s rc_dest=%d\n", info->dest,
 955                                  rc_dest);
 956                 }
 957                 if (rc_dest == -ENOENT) {
 958                         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 959                                 status->ls_targets[info->target_no],
 960                                 SPECIAL_DIR, info->sfid);
 961                         special_dest = 1;
 962                         lr_debug(DINFO, "special dest %s\n", info->dest);
 963                 }
 964
 965                 if (!rc_src) {
 966                         snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 967                                 status->ls_targets[info->target_no],
 968                                 srcpath, info->sname);
 969                         lr_debug(DINFO, "src path %s rc_src=%d\n", info->src,
 970                                  rc_src);
 971                 }
 972                 if (rc_src == -ENOENT || (access(info->src, F_OK) != 0 &&
 973                                           errno == ENOENT)) {
 974                         snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 975                                 status->ls_targets[info->target_no],
 976                                 SPECIAL_DIR, info->sfid);
 977                         special_src = 1;
 978                         lr_debug(DINFO, "special src %s\n", info->src);
 979                 }
 980
 981                 rc1 = 0;
 982                 errno = 0;
 983                 if (strcmp(info->src, info->dest) != 0) {
 984                         rc1 = rename(info->src, info->dest);
 985                         if (rc1 == -1)
 986                                 rc1 = -errno;
 987                         lr_debug(DINFO, "rename returns %d\n", rc1);
 988                 }
 989
 990                 if (special_src)
 991                         rc1 = lr_remove_pc(info->spfid, info->sfid);
 992
 993                 if (!special_dest)
 994                         lr_cascade_move(info->sfid, info->dest, info);
 995                 else
 996                         rc1 = lr_add_pc(info->pfid, info->sfid, info->name);
 997
 998                 lr_debug(DINFO, "move: %s [to] %s rc1=%d, errno=%d\n",
 999                          info->src, info->dest, rc1, errno);
1000                 if (rc1)
1001                         rc = rc1;
1002         }
1003         return rc;
1004 }
1005
1006 /* Replicate a hard link */
1007 int lr_link(struct lr_info *info)
1008 {
1009         int i;
1010         int rc;
1011         int rc1;
1012         struct stat st;
1013
1014         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1015         rc = stat(info->src, &st);
1016         if (rc == -1)
1017                 return -errno;
1018
1019         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1020              info->target_no++) {
1021
1022                 info->src[0] = 0;
1023                 info->dest[0] = 0;
1024                 rc1 = 0;
1025
1026                 /*
1027                  * The changelog record has the new parent directory FID and
1028                  * name of the target file. So info->dest can be constructed
1029                  * by getting the path of the new parent directory and
1030                  * appending the target file name.
1031                  */
1032                 rc1 = lr_get_path(info, info->pfid);
1033                 lr_debug(rc1 ? 0 : DTRACE, "\tparent fid2path %s, %s, rc=%d\n",
1034                          info->path, info->name, rc1);
1035
1036                 if (rc1 == 0) {
1037                         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1038                                  status->ls_targets[info->target_no],
1039                                  info->path, info->name);
1040                         lr_debug(DINFO, "link destination is %s\n", info->dest);
1041                 }
1042
1043                 /* Search through the hardlinks to get the src */
1044                 for (i = 0; i < st.st_nlink && info->src[0] == 0; i++) {
1045                         size_t len;
1046
1047                         rc1 = lr_get_path_ln(info, info->tfid, i);
1048                         lr_debug(rc1 ? 0:DTRACE, "\tfid2path %s, %s, %d rc=%d\n",
1049                                  info->path, info->name, i, rc1);
1050                         if (rc1)
1051                                 break;
1052
1053                         /*
1054                          * Compare the path of target FID with info->dest
1055                          * to find out info->src.
1056                          */
1057                         len = sizeof(status->ls_targets[info->target_no]) +
1058                               sizeof(info->path);
1059                         char srcpath[len + 1];
1060
1061                         snprintf(srcpath, sizeof(srcpath), "%s/%s",
1062                                  status->ls_targets[info->target_no],
1063                                  info->path);
1064
1065                         if (strcmp(srcpath, info->dest) != 0) {
1066                                 snprintf(info->src, sizeof(info->src), "%s",
1067                                          srcpath);
1068                                 lr_debug(DINFO, "link source is %s\n",
1069                                          info->src);
1070                         }
1071                 }
1072
1073                 if (rc1) {
1074                         rc = rc1;
1075                         continue;
1076                 }
1077
1078                 if (info->src[0] == 0)
1079                         snprintf(info->src, sizeof(info->src), "%s/%s/%s",
1080                                 status->ls_targets[info->target_no],
1081                                 SPECIAL_DIR, info->tfid);
1082                 else if (info->dest[0] == 0)
1083                         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1084                                 status->ls_targets[info->target_no],
1085                                 SPECIAL_DIR, info->tfid);
1086
1087                 rc1 = link(info->src, info->dest);
1088                 lr_debug(DINFO, "link: %s [to] %s; rc1=%d %s\n",
1089                          info->src, info->dest, rc1,
1090                          strerror(rc1 ? errno : 0));
1091
1092                 if (rc1)
1093                         rc = rc1;
1094         }
1095         return rc;
1096 }
1097
1098 int lr_set_dest_for_attr(struct lr_info *info)
1099 {
1100         int rc;
1101
1102         snprintf(info->dest, sizeof(info->dest), "%s/%s",
1103                  status->ls_targets[info->target_no], info->path);
1104         rc = access(info->dest, F_OK);
1105         if (rc < 0)
1106                 rc = -errno;
1107
1108         if (rc != -ENOENT)
1109                 return rc;
1110
1111         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1112                  status->ls_targets[info->target_no], SPECIAL_DIR,
1113                  info->tfid);
1114
1115         rc = access(info->dest, F_OK);
1116         if (rc < 0)
1117                 return -errno;
1118
1119         return 0;
1120 }
1121
1122 /* Replicate file attributes */
1123 int lr_setattr(struct lr_info *info)
1124 {
1125         int rc1;
1126         int rc;
1127
1128         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1129
1130         rc = lr_get_path(info, info->tfid);
1131         if (rc == -ENOENT)
1132                 lr_debug(DINFO, "setattr: %s not present on source-fs\n",
1133                          info->src);
1134         if (rc)
1135                 return rc;
1136
1137         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1138              info->target_no++) {
1139                 rc = lr_set_dest_for_attr(info);
1140                 if (rc < 0)
1141                         continue;
1142
1143                 lr_debug(DINFO, "setattr: %s %s %s", info->src, info->dest,
1144                          info->tfid);
1145
1146                 rc1 = lr_sync_data(info);
1147                 if (!rc1)
1148                         rc1 = lr_copy_attr(info->src, info->dest);
1149                 if (rc1)
1150                         rc = rc1;
1151         }
1152         return rc;
1153 }
1154
1155 /* Replicate xattrs */
1156 int lr_setxattr(struct lr_info *info)
1157 {
1158         int rc, rc1;
1159
1160         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1161
1162         rc = lr_get_path(info, info->tfid);
1163         if (rc == -ENOENT)
1164                 lr_debug(DINFO, "setxattr: %s not present on source-fs\n",
1165                          info->src);
1166         if (rc)
1167                 return rc;
1168
1169         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1170              info->target_no++) {
1171                 rc = lr_set_dest_for_attr(info);
1172                 if (rc < 0)
1173                         continue;
1174
1175                 lr_debug(DINFO, "setxattr: %s %s %s\n", info->src, info->dest,
1176                          info->tfid);
1177
1178                 rc1 = lr_copy_xattr(info);
1179                 if (rc1)
1180                         rc = rc1;
1181         }
1182
1183         return rc;
1184 }
1185
1186 /* Parse a line of changelog entry */
1187 int lr_parse_line(void *priv, struct lr_info *info)
1188 {
1189         struct changelog_rec            *rec;
1190         struct changelog_ext_rename     *rnm;
1191         size_t                           namelen;
1192         size_t                           copylen = sizeof(info->name);
1193
1194         if (llapi_changelog_recv(priv, &rec) != 0)
1195                 return -1;
1196
1197         info->is_extended = !!(rec->cr_flags & CLF_RENAME);
1198         info->recno = rec->cr_index;
1199         info->type = rec->cr_type;
1200         snprintf(info->tfid, sizeof(info->tfid), DFID, PFID(&rec->cr_tfid));
1201         snprintf(info->pfid, sizeof(info->pfid), DFID, PFID(&rec->cr_pfid));
1202
1203         namelen = strnlen(changelog_rec_name(rec), rec->cr_namelen);
1204         if (copylen > namelen + 1)
1205                 copylen = namelen + 1;
1206         snprintf(info->name, copylen, "%s", changelog_rec_name(rec));
1207
1208         /* Don't use rnm if CLF_RENAME isn't set */
1209         rnm = changelog_rec_rename(rec);
1210         if (rec->cr_flags & CLF_RENAME && !fid_is_zero(&rnm->cr_sfid)) {
1211                 copylen = sizeof(info->sname);
1212
1213                 snprintf(info->sfid, sizeof(info->sfid), DFID,
1214                          PFID(&rnm->cr_sfid));
1215                 snprintf(info->spfid, sizeof(info->spfid), DFID,
1216                          PFID(&rnm->cr_spfid));
1217                 namelen = changelog_rec_snamelen(rec);
1218                 if (copylen > namelen + 1)
1219                         copylen = namelen + 1;
1220                 snprintf(info->sname, copylen, "%s", changelog_rec_sname(rec));
1221
1222                 if (verbose > 1)
1223                         printf("Rec %lld: %d %s %s\n", info->recno, info->type,
1224                                 info->name, info->sname);
1225         } else {
1226                 if (verbose > 1)
1227                         printf("Rec %lld: %d %s\n", info->recno, info->type,
1228                                 info->name);
1229         }
1230
1231         llapi_changelog_free(&rec);
1232
1233         rec_count++;
1234         return 0;
1235 }
1236
1237 /* Initialize the replication parameters */
1238 int lr_init_status()
1239 {
1240         size_t size = sizeof(struct lustre_rsync_status) + PATH_MAX + 1;
1241
1242         if (status != NULL)
1243                 return 0;
1244         status = calloc(size, 1);
1245         if (status == NULL)
1246                 return -ENOMEM;
1247         status->ls_version = REPLICATE_STATUS_VER;
1248         status->ls_size = size;
1249         status->ls_last_recno = -1;
1250         return 0;
1251 }
1252
1253 /* Make a backup of the statuslog */
1254 void lr_backup_log()
1255 {
1256         char backupfile[PATH_MAX];
1257
1258         if (logbackedup)
1259                 return;
1260         snprintf(backupfile, sizeof(backupfile), "%s.old", statuslog);
1261         (void) rename(statuslog, backupfile);
1262         logbackedup = 1;
1263
1264         return;
1265 }
1266
1267 /* Save replication parameters to a statuslog. */
1268 int lr_write_log()
1269 {
1270         int fd;
1271         size_t size;
1272         size_t write_size = status->ls_size;
1273         struct lr_parent_child_list *curr;
1274         int rc = 0;
1275
1276         if (statuslog == NULL)
1277                 return 0;
1278
1279         lr_backup_log();
1280
1281         fd = open(statuslog, O_WRONLY | O_CREAT | O_SYNC,
1282                              S_IRUSR | S_IWUSR);
1283         if (fd == -1) {
1284                 fprintf(stderr, "Error opening log file for writing (%s)\n",
1285                         statuslog);
1286                 return -1;
1287         }
1288         errno = 0;
1289         size = write(fd, status, write_size);
1290         if (size != write_size) {
1291                 fprintf(stderr, "Error writing to log file (%s) %d\n",
1292                         statuslog, errno);
1293                 close(fd);
1294                 return -1;
1295         }
1296
1297         for (curr = parents; curr; curr = curr->pc_next) {
1298                 size = write(fd, &curr->pc_log, sizeof(curr->pc_log));
1299                 if (size != sizeof(curr->pc_log)) {
1300                         fprintf(stderr, "Error writing to log file (%s) %d\n",
1301                                 statuslog, errno);
1302                         rc = -1;
1303                         break;
1304                 }
1305         }
1306         close(fd);
1307         return rc;
1308 }
1309
1310 /* Read statuslog and populate the replication parameters.  Command
1311  * line parameters take precedence over parameters in the log file.*/
1312 int lr_read_log()
1313 {
1314         struct lr_parent_child_list *tmp;
1315         struct lr_parent_child_log rec;
1316         struct lustre_rsync_status *s;
1317         int fd = -1;
1318         size_t size;
1319         size_t read_size = sizeof(struct lustre_rsync_status) + PATH_MAX + 1;
1320         int rc = 0;
1321
1322         if (statuslog == NULL)
1323                 return 0;
1324
1325         s = calloc(1, read_size);
1326         if (s == NULL) {
1327                 rc = -ENOMEM;
1328                 goto out;
1329         }
1330
1331         fd = open(statuslog, O_RDONLY);
1332         if (fd == -1) {
1333                 rc = -errno;
1334                 goto out;
1335         }
1336
1337         size = read(fd, s, read_size);
1338         if (size != read_size) {
1339                 rc = -EINVAL;
1340                 goto out;
1341         }
1342
1343         if (read_size < s->ls_size) {
1344                 read_size = s->ls_size;
1345                 s = lr_grow_buf(s, read_size);
1346                 if (s == NULL) {
1347                         rc = -ENOMEM;
1348                         goto out;
1349                 }
1350
1351                 if (lseek(fd, 0, SEEK_SET) == -1) {
1352                         rc = -ENOMEM;
1353                         goto out;
1354                 }
1355
1356                 size = read(fd, s, read_size);
1357                 if (size != read_size) {
1358                         rc = -EINVAL;
1359                         goto out;
1360                 }
1361         }
1362
1363         while (read(fd, &rec, sizeof(rec)) != 0) {
1364                 tmp = calloc(1, sizeof(*tmp));
1365                 if (!tmp) {
1366                         rc = -ENOMEM;
1367                         goto out;
1368                 }
1369
1370                 tmp->pc_log = rec;
1371                 tmp->pc_next = parents;
1372                 parents = tmp;
1373         }
1374
1375         /* copy uninitialized fields to status */
1376         if (status->ls_num_targets == 0) {
1377                 if (status->ls_size != s->ls_size) {
1378                         status = lr_grow_buf(status, s->ls_size);
1379                         if (status == NULL) {
1380                                 rc = -ENOMEM;
1381                                 goto out;
1382                         }
1383
1384                         status->ls_size = s->ls_size;
1385                 }
1386                 status->ls_num_targets = s->ls_num_targets;
1387                 memcpy(status->ls_targets, s->ls_targets,
1388                        (PATH_MAX + 1) * s->ls_num_targets);
1389         }
1390         if (status->ls_last_recno == -1)
1391                 status->ls_last_recno = s->ls_last_recno;
1392
1393         if (status->ls_registration[0] == '\0')
1394                 snprintf(status->ls_registration,
1395                          sizeof(status->ls_registration), "%s",
1396                          s->ls_registration);
1397
1398         if (status->ls_mdt_device[0] == '\0')
1399                 snprintf(status->ls_mdt_device,
1400                          sizeof(status->ls_mdt_device), "%s",
1401                          s->ls_mdt_device);
1402
1403         if (status->ls_source_fs[0] == '\0')
1404                 snprintf(status->ls_source_fs,
1405                          sizeof(status->ls_source_fs), "%s",
1406                          s->ls_source_fs);
1407
1408         if (status->ls_source[0] == '\0')
1409                 snprintf(status->ls_source,
1410                          sizeof(status->ls_source), "%s",
1411                          s->ls_source);
1412
1413  out:
1414         if (fd != -1)
1415                 close(fd);
1416         if (s)
1417                 free(s);
1418         return rc;
1419 }
1420
1421 /* Clear changelogs every CLEAR_INTERVAL records or at the end of
1422    processing. */
1423 int lr_clear_cl(struct lr_info *info, int force)
1424 {
1425         char            mdt_device[LR_NAME_MAXLEN + 1];
1426         int             rc = 0;
1427
1428         if (force || info->recno > status->ls_last_recno + CLEAR_INTERVAL) {
1429                 if (!noclear && !dryrun) {
1430                         /* llapi_changelog_clear modifies the mdt
1431                          * device name so make a copy of it until this
1432                          * is fixed.
1433                          */
1434                         snprintf(mdt_device, sizeof(mdt_device), "%s",
1435                                  status->ls_mdt_device);
1436                         rc = llapi_changelog_clear(mdt_device,
1437                                                    status->ls_registration,
1438                                                    info->recno);
1439                         if (rc)
1440                                 printf("Changelog clear (%s, %s, %lld) "
1441                                        "returned %d\n", status->ls_mdt_device,
1442                                        status->ls_registration, info->recno,
1443                                        rc);
1444                 }
1445
1446                 if (!rc && !dryrun) {
1447                         status->ls_last_recno = info->recno;
1448                         lr_write_log();
1449                 }
1450         }
1451
1452         return rc;
1453 }
1454
1455 /* Locate a usable version of rsync. At this point we'll use any
1456    version. */
1457 int lr_locate_rsync()
1458 {
1459         FILE *fp;
1460         int len;
1461
1462         /* Locate rsync */
1463         snprintf(rsync, sizeof(rsync), "%s -p %s", TYPE, RSYNC);
1464         fp = popen(rsync, "r");
1465         if (fp == NULL)
1466                 return -1;
1467
1468         if (fgets(rsync, sizeof(rsync), fp) == NULL) {
1469                 fclose(fp);
1470                 return -1;
1471         }
1472
1473         len = strlen(rsync);
1474         if (len > 0 && rsync[len - 1] == '\n')
1475                 rsync[len - 1] = '\0';
1476         fclose(fp);
1477
1478         /* Determine the version of rsync */
1479         snprintf(rsync_ver, sizeof(rsync_ver), "%s --version", rsync);
1480         fp = popen(rsync_ver, "r");
1481         if (fp == NULL)
1482                 return -1;
1483
1484         if (fgets(rsync_ver, sizeof(rsync_ver), fp) == NULL) {
1485                 fclose(fp);
1486                 return -1;
1487         }
1488         len = strlen(rsync_ver);
1489         if (len > 0 && rsync_ver[len - 1] == '\n')
1490                 rsync_ver[len - 1] = '\0';
1491         fclose(fp);
1492
1493         return 0;
1494
1495 }
1496
1497 /* Print the replication parameters */
1498 void lr_print_status(struct lr_info *info)
1499 {
1500         int i;
1501
1502         if (!verbose)
1503                 return;
1504
1505         printf("Lustre filesystem: %s\n", status->ls_source_fs);
1506         printf("MDT device: %s\n", status->ls_mdt_device);
1507         printf("Source: %s\n", status->ls_source);
1508         for (i = 0; i < status->ls_num_targets; i++)
1509                 printf("Target: %s\n", status->ls_targets[i]);
1510         if (statuslog != NULL)
1511                 printf("Statuslog: %s\n", statuslog);
1512         printf("Changelog registration: %s\n", status->ls_registration);
1513         printf("Starting changelog record: %jd\n",
1514                (uintmax_t)status->ls_last_recno);
1515         if (noxattr)
1516                 printf("Replicate xattrs: no\n");
1517         if (noclear)
1518                 printf("Clear changelog after use: no\n");
1519         if (use_rsync)
1520                 printf("Using rsync: %s (%s)\n", rsync, rsync_ver);
1521 }
1522
1523 void lr_print_failure(struct lr_info *info, int rc)
1524 {
1525         fprintf(stderr, "Replication of operation failed(%d):"
1526                 " %lld %s (%d) %s %s %s\n", rc, info->recno,
1527                 changelog_type2str(info->type), info->type, info->tfid,
1528                 info->pfid, info->name);
1529 }
1530
1531 /* Replicate filesystem operations from src_path to target_path */
1532 int lr_replicate()
1533 {
1534         void *changelog_priv;
1535         struct lr_info *info;
1536         struct lr_info *ext = NULL;
1537         time_t start;
1538         int xattr_not_supp;
1539         int i;
1540         int rc;
1541
1542         start = time(NULL);
1543
1544         info = calloc(1, sizeof(struct lr_info));
1545         if (info == NULL)
1546                 return -ENOMEM;
1547
1548         rc = llapi_search_fsname(status->ls_source, status->ls_source_fs);
1549         if (rc) {
1550                 fprintf(stderr, "Source path is not a valid Lustre client "
1551                         "mountpoint.\n");
1552                 goto out;
1553         }
1554
1555         if (status->ls_mdt_device[0] == '\0') {
1556                 int len;
1557
1558                 len = snprintf(status->ls_mdt_device,
1559                                sizeof(status->ls_mdt_device), "%s%s",
1560                                status->ls_source_fs, DEFAULT_MDT);
1561                 if (len >= sizeof(status->ls_mdt_device)) {
1562                         rc = -E2BIG;
1563                         goto out;
1564                 }
1565         }
1566
1567         ext = calloc(1, sizeof(struct lr_info));
1568         if (ext == NULL) {
1569                 rc = -ENOMEM;
1570                 goto out;
1571         }
1572
1573         for (i = 0, xattr_not_supp = 0; i < status->ls_num_targets; i++) {
1574                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
1575                          status->ls_targets[i], SPECIAL_DIR);
1576                 rc = mkdir(info->dest, 0777);
1577                 if (rc == -1 && errno != EEXIST) {
1578                         fprintf(stderr, "Error writing to target path %s.\n",
1579                                 status->ls_targets[i]);
1580                         rc = -errno;
1581                         goto out;
1582                 }
1583                 rc = llistxattr(info->src, info->xlist, info->xsize);
1584                 if (rc == -1 && errno == ENOTSUP) {
1585                         fprintf(stderr, "xattrs not supported on %s\n",
1586                                 status->ls_targets[i]);
1587                         xattr_not_supp++;
1588                 }
1589         }
1590         if (xattr_not_supp == status->ls_num_targets)
1591                 /* None of the targets support xattrs. */
1592                 noxattr = 1;
1593
1594         lr_print_status(info);
1595
1596         /* Open changelogs for consumption*/
1597         rc = llapi_changelog_start(&changelog_priv,
1598                                 CHANGELOG_FLAG_BLOCK |
1599                                 CHANGELOG_FLAG_JOBID |
1600                                 CHANGELOG_FLAG_EXTRA_FLAGS,
1601                                 status->ls_mdt_device, status->ls_last_recno);
1602         if (rc < 0) {
1603                 fprintf(stderr, "Error opening changelog file for fs %s.\n",
1604                         status->ls_source_fs);
1605                 goto out;
1606         }
1607
1608         rc = llapi_changelog_set_xflags(changelog_priv,
1609                                         CHANGELOG_EXTRA_FLAG_UIDGID |
1610                                         CHANGELOG_EXTRA_FLAG_NID |
1611                                         CHANGELOG_EXTRA_FLAG_OMODE |
1612                                         CHANGELOG_EXTRA_FLAG_XATTR);
1613         if (rc < 0) {
1614                 fprintf(stderr, "Error setting xflag in changelog for fs %s.\n",
1615                         status->ls_source_fs);
1616                 goto out;
1617         }
1618
1619         while (!quit && lr_parse_line(changelog_priv, info) == 0) {
1620                 rc = 0;
1621                 if (info->type == CL_RENAME && !info->is_extended) {
1622                         /* Newer rename operations extends changelog to store
1623                          * source file information, but old changelog has
1624                          * another record.
1625                          */
1626                         if (lr_parse_line(changelog_priv, ext) != 0)
1627                                 break;
1628                         memcpy(info->sfid, info->tfid, sizeof(info->sfid));
1629                         memcpy(info->spfid, info->pfid, sizeof(info->spfid));
1630                         memcpy(info->tfid, ext->tfid, sizeof(info->tfid));
1631                         memcpy(info->pfid, ext->pfid, sizeof(info->pfid));
1632                         snprintf(info->sname, sizeof(info->sname), "%s",
1633                                  info->name);
1634                         snprintf(info->name, sizeof(info->name), "%s",
1635                                  ext->name);
1636                         info->is_extended = 1;
1637                         info->recno = ext->recno; /* For lr_clear_cl(). */
1638                 }
1639
1640                 if (dryrun)
1641                         continue;
1642
1643                 lr_debug(DTRACE, "***** Start %lld %s (%d) %s %s %s *****\n",
1644                          info->recno, changelog_type2str(info->type),
1645                          info->type, info->tfid, info->pfid, info->name);
1646
1647                 switch(info->type) {
1648                 case CL_CREATE:
1649                 case CL_MKDIR:
1650                 case CL_MKNOD:
1651                 case CL_SOFTLINK:
1652                         rc = lr_create(info);
1653                         break;
1654                 case CL_RMDIR:
1655                 case CL_UNLINK:
1656                         rc = lr_remove(info);
1657                         break;
1658                 case CL_RENAME:
1659                         rc = lr_move(info);
1660                         break;
1661                 case CL_HARDLINK:
1662                         rc = lr_link(info);
1663                         break;
1664                 case CL_TRUNC:
1665                 case CL_SETATTR:
1666                         rc = lr_setattr(info);
1667                         break;
1668                 case CL_SETXATTR:
1669                         rc = lr_setxattr(info);
1670                         break;
1671                 case CL_CLOSE:
1672                 case CL_EXT:
1673                 case CL_OPEN:
1674                 case CL_GETXATTR:
1675                 case CL_DN_OPEN:
1676                 case CL_LAYOUT:
1677                 case CL_MARK:
1678                         /* Nothing needs to be done for these entries */
1679                         /* fallthrough */
1680                 default:
1681                         break;
1682                 }
1683
1684                 lr_debug(DTRACE, "##### End %lld %s (%d) %s %s %s rc=%d #####\n",
1685                          info->recno, changelog_type2str(info->type),
1686                          info->type, info->tfid, info->pfid, info->name, rc);
1687
1688                 if (rc && rc != -ENOENT) {
1689                         lr_print_failure(info, rc);
1690                         errors++;
1691                         if (abort_on_err)
1692                                 break;
1693                 }
1694                 lr_clear_cl(info, 0);
1695         }
1696
1697         llapi_changelog_fini(&changelog_priv);
1698
1699         if (errors || verbose)
1700                 printf("Errors: %d\n", errors);
1701
1702         /* Clear changelog records used so far */
1703         lr_clear_cl(info, 1);
1704
1705         if (verbose) {
1706                 printf("lustre_rsync took %ld seconds\n", time(NULL) - start);
1707                 printf("Changelog records consumed: %lld\n", rec_count);
1708         }
1709
1710         rc = 0;
1711
1712 out:
1713         if (info != NULL)
1714                 free(info);
1715         if (ext != NULL)
1716                 free(ext);
1717
1718         return rc;
1719 }
1720
1721 void
1722 termination_handler (int signum)
1723 {
1724         /* Set a flag for the replicator to gracefully shutdown */
1725         quit = 1;
1726         printf("lustre_rsync halting.\n");
1727 }
1728
1729 int main(int argc, char *argv[])
1730 {
1731         int newsize;
1732         int numtargets = 0;
1733         int rc = 0;
1734
1735         if ((rc = lr_init_status()) != 0)
1736                 return rc;
1737
1738         while ((rc = getopt_long(argc, argv, "as:t:m:u:l:vx:zc:ry:n:d:D:",
1739                                  long_opts, NULL)) >= 0) {
1740                 switch (rc) {
1741                 case 'a':
1742                         /* Assume absolute paths */
1743                         abort_on_err++;
1744                         break;
1745                 case 's':
1746                         /* Assume absolute paths */
1747                         snprintf(status->ls_source, sizeof(status->ls_source),
1748                                  "%s", optarg);
1749                         break;
1750                 case 't':
1751                         status->ls_num_targets++;
1752                         numtargets++;
1753                         if (numtargets != status->ls_num_targets) {
1754                                 /* Targets were read from a log
1755                                    file. The ones specified on the
1756                                    command line take precedence. The
1757                                    ones from the log file will be
1758                                    ignored. */
1759                                 status->ls_num_targets = numtargets;
1760                         }
1761                         newsize = sizeof (struct lustre_rsync_status) +
1762                                 (status->ls_num_targets * (PATH_MAX + 1));
1763                         if (status->ls_size != newsize) {
1764                                 status->ls_size = newsize;
1765                                 status = lr_grow_buf(status, newsize);
1766                                 if (status == NULL)
1767                                         return -ENOMEM;
1768                         }
1769                         snprintf(status->ls_targets[status->ls_num_targets - 1],
1770                                  sizeof(status->ls_targets[0]), "%s", optarg);
1771                         break;
1772                 case 'm':
1773                         snprintf(status->ls_mdt_device,
1774                                  sizeof(status->ls_mdt_device),
1775                                  "%s", optarg);
1776                         break;
1777                 case 'u':
1778                         snprintf(status->ls_registration,
1779                                  sizeof(status->ls_registration),
1780                                  "%s", optarg);
1781                         break;
1782                 case 'l':
1783                         statuslog = optarg;
1784                         (void) lr_read_log();
1785                         break;
1786                 case 'v':
1787                         verbose++;
1788                         break;
1789                 case 'x':
1790                         if (strcmp("no", optarg) == 0) {
1791                                 noxattr = 1;
1792                         } else if (strcmp("yes", optarg) != 0) {
1793                                 printf("Invalid parameter %s. "
1794                                        "Specify --xattr=no or --xattr=yes\n",
1795                                        optarg);
1796                                 return -1;
1797                         }
1798                         break;
1799                 case 'z':
1800                         dryrun = 1;
1801                         break;
1802                 case 'c':
1803                         /* Undocumented option cl-clear */
1804                         if (strcmp("no", optarg) == 0) {
1805                                 noclear = 1;
1806                         } else if (strcmp("yes", optarg) != 0) {
1807                                 printf("Invalid parameter %s. "
1808                                        "Specify --cl-clear=no "
1809                                        "or --cl-clear=yes\n",
1810                                        optarg);
1811                                 return -1;
1812                         }
1813                         break;
1814                 case 'r':
1815                         /* Undocumented option use-rsync */
1816                         use_rsync = 1;
1817                         break;
1818                 case 'y':
1819                         /* Undocumented option rsync-threshold */
1820                         rsync_threshold = atol(optarg);
1821                         break;
1822                 case 'n':
1823                         /* Undocumented option start-recno */
1824                         status->ls_last_recno = atol(optarg);
1825                         break;
1826                 case 'd':
1827                         /* Undocumented option debug */
1828                         debug = atoi(optarg);
1829                         if (debug < 0 || debug > 2)
1830                                 debug = 0;
1831                         break;
1832                 case 'D':
1833                         /* Undocumented option debug log file */
1834                         if (debug_log != NULL)
1835                                 fclose(debug_log);
1836                         debug_log = fopen(optarg, "a");
1837                         if (debug_log == NULL) {
1838                                 printf("Cannot open %s for debug log\n",
1839                                        optarg);
1840                                 return -1;
1841                         }
1842                         break;
1843                 default:
1844                         fprintf(stderr, "error: %s: option '%s' "
1845                                 "unrecognized.\n", argv[0], argv[optind - 1]);
1846                         lr_usage();
1847                         return -1;
1848                 }
1849         }
1850
1851         if (status->ls_last_recno == -1)
1852                 status->ls_last_recno = 0;
1853         if (strnlen(status->ls_registration, LR_NAME_MAXLEN) == 0) {
1854                 /* No registration ID was passed in. */
1855                 printf("Please specify changelog consumer registration id.\n");
1856                 lr_usage();
1857                 return -1;
1858         }
1859         if (strnlen(status->ls_source, PATH_MAX) == 0) {
1860                 fprintf(stderr, "Please specify the source path.\n");
1861                 lr_usage();
1862                 return -1;
1863         }
1864         if (strnlen(status->ls_targets[0], PATH_MAX) == 0) {
1865                 fprintf(stderr, "Please specify the target path.\n");
1866                 lr_usage();
1867                 return -1;
1868         }
1869
1870         rc = lr_locate_rsync();
1871         if (use_rsync && rc != 0) {
1872                 fprintf(stderr, "Error: unable to locate %s.\n", RSYNC);
1873                 exit(-1);
1874         }
1875
1876         signal(SIGINT, termination_handler);
1877         signal(SIGHUP, termination_handler);
1878         signal(SIGTERM, termination_handler);
1879
1880         rc = lr_replicate();
1881
1882         if (debug_log != NULL)
1883                 fclose(debug_log);
1884         return rc;
1885 }