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  *
  31  * lustre/utils/lustre_rsync.c
  32  *
  33  * Author: Kalpak Shah <Kalpak.Shah@Sun.COM>
  34  * Author: Manoj Joseph <Manoj.Joseph@Sun.COM>
  35  */
  36
  37 /*
  38  * - lustre_rsync is a tool for replicating a lustre filesystem.
  39  *
  40  * - The source-fs is a live lustre filesystem. It is not a
  41  * snapshot. It is mounted and undergoing changes
  42  *
  43  * - The target-fs is a copy of the source-fs from the past. Let's
  44  * call this point, the 'sync point'.
  45  *
  46  * - There is a changelog of all metadata operations that happened on
  47  * the filesystem since the 'sync point'.
  48  *
  49  * - lustre_rsync replicates all the operations saved in the changelog
  50  * on to the target filesystem to make it identical to the source.
  51  *
  52  * To facilitate replication, the lustre filesystem provides
  53  *    a) a way to get the current filesystem path of a given FID
  54  *    b) a way to open files by specifying its FID
  55  *
  56  * The changelog only has a limited amount of information.
  57  *  tfid - The FID of the target file
  58  *  pfid - The FID of the parent of the target file (at the time of
  59  *         the operation)
  60  *  sfid - The FID of the source file
  61  *  spfid - The FID of the parent of the source file
  62  *  name - The name of the target file (at the time of the operation), the name
  63  *         of the source file is appended (delimited with '\0') if this
  64  *         operation involves a source
  65  *
  66  * With just this information, it is not alwasy possible to determine
  67  * the file paths for each operation. For instance, if pfid does not
  68  * exist on the source-fs (due to a subsequent deletion), its path
  69  * cannot be queried. In such cases, lustre_rsync keeps the files in a
  70  * special directory ("/.lustrerepl"). Once all the operations in a
  71  * changelog are replayed, all the files in this special directory
  72  * will get moved to the location as in the source-fs.
  73  *
  74  * Shorthand used: f2p(fid) = fid2path(fid)
  75  *
  76  * The following are the metadata operations of interest.
  77  * 1. creat
  78  *    If tfid is absent on the source-fs, ignore this operation
  79  *    If pfid is absent on the source-fs [or]
  80  *    if f2p(pfid) is not present on target-fs [or]
  81  *    if f2p(pfid)+name != f2p(tfid)
  82  *      creat .lustrerepl/tfid
  83  *      track [pfid,tfid,name]
  84  *    Else
  85  *      creat f2p[tfid]
  86  *
  87  * 2. remove
  88  *    If .lustrerepl/[tfid] is present on the target
  89  *      rm .lustrerepl/[tfid]
  90  *    Else if pfid is present on the source-fs,
  91  *      if f2p(pfid)+name is present,
  92  *        rm f2p(pfid)+name
  93  *
  94  * 3. move (spfid,sname) to (pfid,name)
  95  *    If pfid is present
  96  *      if spfid is also present, mv (spfid,sname) to (pfid,name)
  97  *      else mv .lustrerepl/[sfid] to (pfid,name)
  98  *    Else if pfid is not present,
  99  *      if spfid is present, mv (spfid,sname) .lustrerepl/[sfid]
 100  *    If moving out of .lustrerepl
 101  *      move out all its children in .lustrerepl.
 102  *      [pfid,tfid,name] tracked from (1) is used for this.
 103  */
 104
 105 #include <assert.h>
 106 #include <stdio.h>
 107 #include <stdlib.h>
 108 #include <string.h>
 109 #include <unistd.h>
 110 #include <getopt.h>
 111 #include <stdarg.h>
 112 #include <fcntl.h>
 113 #include <signal.h>
 114 #include <sys/stat.h>
 115 #include <sys/types.h>
 116 #include <errno.h>
 117 #include <limits.h>
 118 #include <utime.h>
 119 #include <time.h>
 120 #include <sys/xattr.h>
 121 #include <linux/types.h>
 122
 123 #include <libcfs/util/string.h>
 124 #include <lustre/lustreapi.h>
 125 #include "lustre_rsync.h"
 126 #include "callvpe.h"
 127
 128 #define REPLICATE_STATUS_VER 1
 129 #define CLEAR_INTERVAL 100
 130 #define DEFAULT_RSYNC_THRESHOLD 0xA00000 /* 10 MB */
 131
 132 #define TYPE_STR_LEN 16
 133
 134 #define DEFAULT_MDT "-MDT0000"
 135 #define SPECIAL_DIR ".lustrerepl"
 136 #define RSYNC "rsync"
 137 #define TYPE "type"
 138
 139 /* Debug flags */
 140 #define DINFO 1
 141 #define DTRACE 2
 142
 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  */
 148 struct lr_info {
 149         long long recno;
 150         int target_no;
 151         unsigned int is_extended:1;
 152         enum changelog_rec_type type;
 153         char tfid[LR_FID_STR_LEN];
 154         char pfid[LR_FID_STR_LEN];
 155         char sfid[LR_FID_STR_LEN];
 156         char spfid[LR_FID_STR_LEN];
 157         char sname[NAME_MAX + 1];
 158         char name[NAME_MAX + 1];
 159         char src[3 * PATH_MAX + 1];
 160         char dest[3 * PATH_MAX + 1];
 161         char path[PATH_MAX + 1];
 162         char savedpath[PATH_MAX + 1];
 163         char link[PATH_MAX + 1];
 164         char linktmp[PATH_MAX + 1];
 165         int bufsize;
 166         char *buf;
 167
 168         /* Variables for querying the xattributes */
 169         char *xlist;
 170         size_t xsize;
 171         char *xvalue;
 172         size_t xvsize;
 173 };
 174
 175 struct lr_parent_child_list {
 176         struct lr_parent_child_log pc_log;
 177         struct lr_parent_child_list *pc_next;
 178 };
 179
 180 struct lustre_rsync_status *status;
 181 char *statuslog;  /* Name of the status log file */
 182 int logbackedup;
 183 int noxattr;    /* Flag to turn off replicating xattrs */
 184 int noclear;    /* Flag to turn off clearing changelogs */
 185 int debug;      /* Flag to turn debugging information on and off */
 186 int verbose;    /* Verbose output */
 187 long long rec_count; /* No of changelog records that were processed */
 188 int errors;
 189 int dryrun;
 190 int use_rsync;  /* Flag to turn on use of rsync to copy data */
 191 long long rsync_threshold = DEFAULT_RSYNC_THRESHOLD;
 192 int quit;       /* Flag to stop processing the changelog; set on the
 193                  * receipt of a signal
 194                  */
 195 int abort_on_err;
 196
 197 char rsync[PATH_MAX + 128];
 198 char rsync_ver[PATH_MAX * 2];
 199 struct lr_parent_child_list *parents;
 200
 201 FILE *debug_log;
 202
 203 /* Command line options */
 204 struct option long_opts[] = {
 205         { .val = 'l',   .name = "statuslog",    .has_arg = required_argument },
 206         { .val = 'm',   .name = "mdt",          .has_arg = required_argument },
 207         { .val = 's',   .name = "source",       .has_arg = required_argument },
 208         { .val = 't',   .name = "target",       .has_arg = required_argument },
 209         { .val = 'u',   .name = "user",         .has_arg = required_argument },
 210         { .val = 'v',   .name = "verbose",      .has_arg = no_argument },
 211         { .val = 'x',   .name = "xattr",        .has_arg = required_argument },
 212         { .val = 'z',   .name = "dry-run",      .has_arg = no_argument },
 213         /* Undocumented options follow */
 214         { .val = 'a',   .name = "abort-on-err", .has_arg = no_argument },
 215         { .val = 'c',   .name = "cl-clear",     .has_arg = required_argument },
 216         { .val = 'd',   .name = "debug",        .has_arg = required_argument },
 217         { .val = 'D',   .name = "debuglog",     .has_arg = required_argument },
 218         { .val = 'n',   .name = "start-recno",  .has_arg = required_argument },
 219         { .val = 'r',   .name = "use-rsync",    .has_arg = no_argument },
 220         { .val = 'y',   .name = "rsync-threshold",
 221                                                 .has_arg = required_argument },
 222         { .name = NULL } };
 223
 224 /* Command line usage */
 225 void lr_usage(void)
 226 {
 227         fprintf(stderr, "\tlustre_rsync -s <lustre_root_path> -t <target_path> "
 228                 "-m <mdt> -r <user id> -l <status log>\n"
 229                 "lustre_rsync can also pick up parameters from a "
 230                 "status log created earlier.\n"
 231                 "\tlustre_rsync -l <log_file>\n"
 232                 "options:\n"
 233                 "\t--xattr <yes|no> replicate EAs\n"
 234                 "\t--abort-on-err   abort at first err\n"
 235                 "\t--verbose\n"
 236                 "\t--dry-run        don't write anything\n");
 237 }
 238
 239 /*
 240  * Print debug information. This is controlled by the value of the
 241  * global variable 'debug'
 242  */
 243 void lr_debug(int level, const char *fmt, ...)
 244 {
 245         va_list ap;
 246
 247         if (level > debug)
 248                 return;
 249
 250         va_start(ap, fmt);
 251         if (debug_log)
 252                 vfprintf(debug_log, fmt, ap);
 253         else
 254                 vfprintf(stdout, fmt, ap);
 255         va_end(ap);
 256 }
 257
 258 void *lr_grow_buf(void *buf, int size)
 259 {
 260         void *ptr;
 261
 262         ptr = realloc(buf, size);
 263         if (!ptr)
 264                 free(buf);
 265         return ptr;
 266 }
 267
 268 /* Use rsync to replicate file data */
 269 int lr_rsync_data(struct lr_info *info)
 270 {
 271         struct stat st_src, st_dest;
 272         int rc;
 273
 274         lr_debug(DTRACE, "Syncing data%s\n", info->tfid);
 275
 276         rc = stat(info->src, &st_src);
 277         if (rc == -1) {
 278                 fprintf(stderr, "Error: Unable to stat src=%s %s\n",
 279                         info->src, info->name);
 280                 if (errno == ENOENT)
 281                         return 0;
 282                 else
 283                         return -errno;
 284         }
 285         rc = stat(info->dest, &st_dest);
 286         if (rc == -1) {
 287                 fprintf(stderr, "Error: Unable to stat dest=%s\n",
 288                         info->dest);
 289                 return -errno;
 290         }
 291
 292         if (st_src.st_mtime != st_dest.st_mtime ||
 293             st_src.st_size != st_dest.st_size) {
 294                 /*
 295                  * XXX spawning off an rsync for every data sync and
 296                  * waiting synchronously is bad for performance.
 297                  * librsync could possibly used here. But it does not
 298                  * seem to be of production grade. Multi-threaded
 299                  * replication is also to be considered.
 300                  */
 301                 char *args[] = {
 302                         rsync,
 303                         "--inplace",
 304                         "--",
 305                         info->src,
 306                         info->dest,
 307                         NULL,
 308                 };
 309                 extern char **environ;
 310                 int status;
 311
 312                 lr_debug(DTRACE, "\t%s %s %s %s %s %s\n", args[0], args[1],
 313                          args[2], args[3], args[4], info->tfid);
 314
 315                 status = callvpe(rsync, args, environ);
 316                 if (status < 0) {
 317                         rc = -errno;
 318                 } else if (WIFEXITED(status)) {
 319                         status = WEXITSTATUS(status);
 320                         if (!status)
 321                                 rc = 0;
 322                         else if (status == 23 || status == 24)
 323                                 /* Error due to vanished source files;
 324                                  * Ignore this error
 325                                  */
 326                                 rc = 0;
 327                         else
 328                                 rc = -EINVAL;
 329                         if (status)
 330                                 lr_debug(DINFO, "rsync %s exited with %d %d\n",
 331                                          info->src, status, rc);
 332                 } else {
 333                         rc = -EINTR;
 334                 }
 335         } else {
 336                 lr_debug(DTRACE, "Not syncing %s and %s %s\n", info->src,
 337                          info->dest, info->tfid);
 338         }
 339
 340         return rc;
 341 }
 342
 343 int lr_copy_data(struct lr_info *info)
 344 {
 345         int fd_src = -1;
 346         int fd_dest = -1;
 347         int bufsize;
 348         int rsize;
 349         int rc = 0;
 350         struct stat st_src;
 351         struct stat st_dest;
 352
 353         fd_src = open(info->src, O_RDONLY);
 354         if (fd_src == -1)
 355                 return -errno;
 356         if (fstat(fd_src, &st_src) == -1 || stat(info->dest, &st_dest) == -1)
 357                 goto out;
 358
 359         if (st_src.st_mtime == st_dest.st_mtime &&
 360             st_src.st_size == st_dest.st_size)
 361                 goto out;
 362
 363         if (st_src.st_size > rsync_threshold && rsync[0] != '\0') {
 364                 /*
 365                  * It is more efficient to use rsync to replicate
 366                  * large files. Any file larger than rsync_threshold
 367                  * is handed off to rsync.
 368                  */
 369                 lr_debug(DTRACE, "Using rsync to replicate %s\n", info->tfid);
 370                 rc = lr_rsync_data(info);
 371                 goto out;
 372         }
 373
 374         fd_dest = open(info->dest, O_WRONLY | O_TRUNC, st_src.st_mode);
 375         if (fd_dest == -1) {
 376                 rc = -errno;
 377                 goto out;
 378         }
 379         bufsize = st_dest.st_blksize;
 380
 381         if (info->bufsize < bufsize) {
 382                 /* Grow buffer */
 383                 info->buf = lr_grow_buf(info->buf, bufsize);
 384                 if (!info->buf) {
 385                         rc = -ENOMEM;
 386                         goto out;
 387                 }
 388                 info->bufsize = bufsize;
 389         }
 390
 391         while (1) {
 392                 char *buf;
 393                 int wsize;
 394
 395                 buf = info->buf;
 396                 rsize = read(fd_src, buf, bufsize);
 397                 if (rsize == 0) {
 398                         rc = 0;
 399                         break;
 400                 }
 401                 if (rsize < 0) {
 402                         rc = -errno;
 403                         break;
 404                 }
 405                 do {
 406                         wsize = write(fd_dest, buf, rsize);
 407                         if (wsize <= 0) {
 408                                 rc = -errno;
 409                                 break;
 410                         }
 411                         rsize -= wsize;
 412                         buf += wsize;
 413                 } while (rsize > 0);
 414         }
 415         fsync(fd_dest);
 416
 417 out:
 418         if (fd_src != -1)
 419                 close(fd_src);
 420         if (fd_dest != -1)
 421                 close(fd_dest);
 422
 423         return rc;
 424 }
 425
 426 /* Copy data from source to destination */
 427 int lr_sync_data(struct lr_info *info)
 428 {
 429         if (use_rsync)
 430                 return lr_rsync_data(info);
 431         else
 432                 return lr_copy_data(info);
 433 }
 434
 435 /* Copy all attributes from file src to file dest */
 436 int lr_copy_attr(const char *src, const char *dest)
 437 {
 438         struct stat st;
 439         struct utimbuf time;
 440
 441         if (stat(src, &st) == -1 || chmod(dest, st.st_mode) == -1 ||
 442             chown(dest, st.st_uid, st.st_gid) == -1)
 443                 return -errno;
 444
 445         time.actime = st.st_atime;
 446         time.modtime = st.st_mtime;
 447         if (utime(dest, &time) == -1)
 448                 return -errno;
 449         return 0;
 450 }
 451
 452 /* Copy all xattrs from file info->src to info->dest */
 453 int lr_copy_xattr(struct lr_info *info)
 454 {
 455         size_t size = info->xsize;
 456         int start;
 457         int len;
 458         int rc;
 459
 460         if (noxattr)
 461                 return 0;
 462
 463         errno = 0;
 464         rc = llistxattr(info->src, info->xlist, size);
 465         lr_debug(DTRACE, "llistxattr(%s,%p) returned %d, errno=%d\n",
 466                  info->src, info->xlist, rc, errno);
 467         if ((rc > 0 && info->xlist == NULL) || errno == ERANGE) {
 468                 size = rc > PATH_MAX ? rc : PATH_MAX;
 469                 info->xlist = lr_grow_buf(info->xlist, size);
 470                 if (!info->xlist)
 471                         return -ENOMEM;
 472                 info->xsize = size;
 473                 rc = llistxattr(info->src, info->xlist, size);
 474                 lr_debug(DTRACE, "llistxattr %s returned %d, errno=%d\n",
 475                          info->src, rc, errno);
 476         }
 477         if (rc < 0)
 478                 return rc;
 479
 480         len = rc;
 481         start = 0;
 482         while (start < len) {
 483                 size = info->xvsize;
 484                 rc = lgetxattr(info->src, info->xlist + start,
 485                                info->xvalue, size);
 486                 if (!info->xvalue || errno == ERANGE) {
 487                         size = rc > PATH_MAX ? rc : PATH_MAX;
 488                         info->xvalue = lr_grow_buf(info->xvalue, size);
 489                         if (!info->xvalue)
 490                                 return -ENOMEM;
 491                         info->xvsize = size;
 492                         rc = lgetxattr(info->src, info->xlist + start,
 493                                        info->xvalue, size);
 494                 }
 495                 lr_debug(DTRACE, "\t(%s,%d) rc=%p\n", info->xlist + start,
 496                          info->xvalue, rc);
 497                 if (rc > 0) {
 498                         size = rc;
 499                         rc = lsetxattr(info->dest, info->xlist + start,
 500                                        info->xvalue, size, 0);
 501                         lr_debug(DTRACE, "\tlsetxattr(), rc=%d, errno=%d\n",
 502                                  rc, errno);
 503                         if (rc == -1) {
 504                                 if (errno != ENOTSUP) {
 505                                         fprintf(stderr, "cannot replicate xattrs from '%s' to '%s': %s\n",
 506                                                 info->src, info->dest,
 507                                                 strerror(errno));
 508                                         errors++;
 509                                 }
 510                                 rc = 0;
 511                         }
 512                 }
 513                 start += strlen(info->xlist + start) + 1;
 514         }
 515
 516         lr_debug(DINFO, "setxattr: %s %s\n", info->src, info->dest);
 517
 518         return rc;
 519 }
 520
 521 /*
 522  * Retrieve the filesystem path for a given FID and a given
 523  * linkno. The path is returned in info->path
 524  */
 525 int lr_get_path_ln(struct lr_info *info, char *fidstr, int linkno)
 526 {
 527         long long recno = -1;
 528         int rc;
 529
 530         rc = llapi_fid2path(status->ls_source, fidstr, info->path,
 531                             PATH_MAX, &recno, &linkno);
 532         if (rc < 0 && rc != -ENOENT) {
 533                 fprintf(stderr, "fid2path error: (%s, %s) %d %s\n",
 534                         status->ls_source, fidstr, -rc, strerror(errno = -rc));
 535         }
 536
 537         return rc;
 538 }
 539
 540 /*
 541  * Retrieve the filesystem path for a given FID. The path is returned
 542  * in info->path
 543  */
 544 int lr_get_path(struct lr_info *info, char *fidstr)
 545 {
 546         return lr_get_path_ln(info, fidstr, 0);
 547 }
 548
 549 /* Generate the path for opening by FID */
 550 void lr_get_FID_PATH(char *mntpt, char *fidstr, char *buf, int bufsize)
 551 {
 552         /* Open-by-FID path is <mntpt>/.lustre/fid/[SEQ:OID:VER] */
 553         snprintf(buf, bufsize, "%s/%s/fid/%s", mntpt, dot_lustre_name,
 554                  fidstr);
 555 }
 556
 557 /* Read the symlink information into 'info->link' */
 558 int lr_get_symlink(struct lr_info *info)
 559 {
 560         int rc;
 561         char *link;
 562
 563         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
 564         rc = readlink(info->src, info->linktmp, PATH_MAX);
 565         if (rc > 0)
 566                 info->linktmp[rc] = '\0';
 567         else
 568                 return rc;
 569         lr_debug(DTRACE, "symlink: readlink returned %s\n", info->linktmp);
 570
 571         if (strncmp(info->linktmp, status->ls_source,
 572                     strlen(status->ls_source)) == 0) {
 573                 /* Strip source fs path and replace with target fs path. */
 574                 link = info->linktmp + strlen(status->ls_source);
 575                 snprintf(info->src, sizeof(info->src), "%s%s",
 576                          status->ls_targets[info->target_no], link);
 577                 link = info->src;
 578         } else {
 579                 link = info->linktmp;
 580         }
 581         rc = snprintf(info->link, sizeof(info->link), "%s", link);
 582         if (rc >= sizeof(info->link))
 583                 rc = -E2BIG;
 584         return rc;
 585 }
 586
 587 /* Create file/directory/device file/symlink. */
 588 int lr_mkfile(struct lr_info *info)
 589 {
 590         struct stat st;
 591         int rc = 0;
 592
 593         errno = 0;
 594         lr_debug(DINFO, "mkfile(%d) %s\n", info->type, info->dest);
 595         if (info->type == CL_MKDIR) {
 596                 rc = mkdir(info->dest, 0777);
 597         } else if (info->type == CL_SOFTLINK) {
 598                 lr_get_symlink(info);
 599                 rc = symlink(info->link, info->dest);
 600         } else if (info->type == CL_MKNOD) {
 601                 lr_get_FID_PATH(status->ls_source, info->tfid,
 602                                 info->src, PATH_MAX);
 603                 rc = stat(info->src, &st);
 604                 if (rc == -1) {
 605                         if (errno == ENOENT)
 606                                 return 0;
 607                         else
 608                                 return -errno;
 609                 }
 610                 rc = mknod(info->dest, st.st_mode, st.st_rdev);
 611         } else {
 612                 rc = mknod(info->dest, S_IFREG | 0777, 0);
 613         }
 614
 615         if (rc < 0) {
 616                 if (errno == EEXIST)
 617                         rc = 0;
 618                 else
 619                         return -errno;
 620         }
 621
 622         /* Sync data and attributes */
 623         if (info->type == CL_CREATE || info->type == CL_MKDIR) {
 624                 lr_debug(DTRACE, "Syncing data and attributes %s\n",
 625                          info->tfid);
 626                 (void)lr_copy_xattr(info);
 627                 if (info->type == CL_CREATE)
 628                         rc = lr_sync_data(info);
 629                 if (!rc)
 630                         rc = lr_copy_attr(info->src, info->dest);
 631
 632                 if (rc == -ENOENT)
 633                         /* Source file has disappeared. Not an error. */
 634                         rc = 0;
 635         } else {
 636                 lr_debug(DTRACE, "Not syncing data and attributes %s\n",
 637                          info->tfid);
 638         }
 639
 640         return rc;
 641 }
 642
 643 int lr_add_pc(const char *pfid, const char *tfid, const char *name)
 644 {
 645         struct lr_parent_child_list *p;
 646         size_t len;
 647
 648         p = calloc(1, sizeof(*p));
 649         if (!p)
 650                 return -ENOMEM;
 651         len = snprintf(p->pc_log.pcl_pfid, sizeof(p->pc_log.pcl_pfid),
 652                        "%s", pfid);
 653         if (len >= sizeof(p->pc_log.pcl_pfid))
 654                 goto out_err;
 655         len = snprintf(p->pc_log.pcl_tfid, sizeof(p->pc_log.pcl_tfid),
 656                        "%s", tfid);
 657         if (len >= sizeof(p->pc_log.pcl_tfid))
 658                 goto out_err;
 659         len = snprintf(p->pc_log.pcl_name, sizeof(p->pc_log.pcl_name),
 660                        "%s", name);
 661         if (len >= sizeof(p->pc_log.pcl_name))
 662                 goto out_err;
 663
 664         p->pc_next = parents;
 665         parents = p;
 666         return 0;
 667
 668 out_err:
 669         free(p);
 670         return -E2BIG;
 671 }
 672
 673 void lr_cascade_move(const char *fid, const char *dest, struct lr_info *info)
 674 {
 675         struct lr_parent_child_list *curr, *prev;
 676         char d[4 * PATH_MAX + 1];
 677         int rc;
 678
 679         prev = curr = parents;
 680         while (curr) {
 681                 if (strcmp(curr->pc_log.pcl_pfid, fid) == 0) {
 682                         if (snprintf(d, sizeof(d), "%s/%s", dest,
 683                                      curr->pc_log.pcl_name) >= sizeof(d)) {
 684                                 fprintf(stderr, "Buffer truncated\n");
 685                                 return;
 686                         }
 687                         if (snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 688                                      status->ls_targets[info->target_no],
 689                                      SPECIAL_DIR, curr->pc_log.pcl_tfid) >=
 690                             sizeof(info->src))
 691                                 return;
 692                         rc = rename(info->src, d);
 693                         if (rc == -1) {
 694                                 fprintf(stderr, "Error renaming file %s to %s: %d\n",
 695                                         info->src, d, errno);
 696                                 errors++;
 697                         }
 698                         if (curr == parents)
 699                                 parents = curr->pc_next;
 700                         else
 701                                 prev->pc_next = curr->pc_next;
 702                         lr_cascade_move(curr->pc_log.pcl_tfid, d, info);
 703                         free(curr);
 704                         prev = curr = parents;
 705                 } else {
 706                         prev = curr;
 707                         curr = curr->pc_next;
 708                 }
 709         }
 710 }
 711
 712 /* remove [info->spfid, info->sfid] from parents */
 713 int lr_remove_pc(const char *pfid, const char *tfid)
 714 {
 715         struct lr_parent_child_list *curr, *prev;
 716
 717         for (prev = curr = parents; curr; prev = curr, curr = curr->pc_next) {
 718                 if (strcmp(curr->pc_log.pcl_pfid, pfid) == 0 &&
 719                     strcmp(curr->pc_log.pcl_tfid, tfid) == 0) {
 720                         if (curr == parents)
 721                                 parents = curr->pc_next;
 722                         else
 723                                 prev->pc_next = curr->pc_next;
 724                         free(curr);
 725                         break;
 726                 }
 727         }
 728         return 0;
 729 }
 730
 731 /* Create file under SPECIAL_DIR with its tfid as its name. */
 732 int lr_mk_special(struct lr_info *info)
 733 {
 734         int rc;
 735
 736         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 737                  status->ls_targets[info->target_no], SPECIAL_DIR, info->tfid);
 738
 739         rc = lr_mkfile(info);
 740         if (rc)
 741                 return rc;
 742
 743         rc = lr_add_pc(info->pfid, info->tfid, info->name);
 744         return rc;
 745 }
 746
 747 /* Remove a file or directory */
 748 int lr_rmfile(struct lr_info *info)
 749 {
 750         int rc;
 751
 752         if (info->type == CL_RMDIR)
 753                 rc = rmdir(info->dest);
 754         else
 755                 rc = unlink(info->dest);
 756         if (rc == -1)
 757                 rc = -errno;
 758         return rc;
 759 }
 760
 761 /* Recursively remove directory and its contents */
 762 int lr_rm_recursive(struct lr_info *info)
 763 {
 764         char *args[] = {
 765                 "rm",
 766                 "-rf",
 767                 "--",
 768                 info->dest,
 769                 NULL,
 770         };
 771         extern char **environ;
 772         int status;
 773         int rc;
 774
 775         status = callvpe("/bin/rm", args, environ);
 776         if (status < 0)
 777                 rc = -errno;
 778         else if (WIFEXITED(status))
 779                 rc = WEXITSTATUS(status) == 0 ? 0 : -EINVAL;
 780         else
 781                 rc = -EINTR;
 782
 783         return rc;
 784 }
 785
 786 /* Remove a file under SPECIAL_DIR with its tfid as its name. */
 787 int lr_rm_special(struct lr_info *info)
 788 {
 789         int rc;
 790
 791         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 792                  status->ls_targets[info->target_no], SPECIAL_DIR, info->tfid);
 793         rc = lr_rmfile(info);
 794
 795         if (rc)
 796                 lr_debug(DINFO, "remove: %s; rc=%d, errno=%d\n",
 797                          info->dest, rc, errno);
 798         return rc;
 799 }
 800
 801 /* Replicate file and directory create events */
 802 int lr_create(struct lr_info *info)
 803 {
 804         int len;
 805         int rc1 = 0;
 806         int rc;
 807         int mkspecial = 0;
 808
 809         /* Is target FID present on the source? */
 810         rc = lr_get_path(info, info->tfid);
 811         if (rc == -ENOENT) {
 812                 /* Source file has disappeared. Not an error. */
 813                 lr_debug(DINFO, "create: tfid %s not found on source-fs\n",
 814                          info->tfid);
 815                 return 0;
 816         } else if (rc) {
 817                 return rc;
 818         }
 819         strcpy(info->savedpath, info->path);
 820
 821         /* Is parent FID present on the source */
 822         rc = lr_get_path(info, info->pfid);
 823         if (rc == -ENOENT) {
 824                 lr_debug(DINFO, "create: pfid %s not found on source-fs\n",
 825                          info->tfid);
 826                 mkspecial = 1;
 827         } else if (rc < 0) {
 828                 return rc;
 829         }
 830
 831         /* Is f2p(pfid)+name != f2p(tfid)? If not the file has moved. */
 832         len = strlen(info->path);
 833         if (len == 1 && info->path[0] == '/')
 834                 snprintf(info->dest, sizeof(info->dest), "%s", info->name);
 835         else if (len - 1 > 0 && info->path[len - 1] == '/')
 836                 snprintf(info->dest, sizeof(info->dest), "%s%s", info->path,
 837                          info->name);
 838         else
 839                 snprintf(info->dest, sizeof(info->dest), "%s/%s", info->path,
 840                          info->name);
 841
 842         lr_debug(DTRACE, "dest = %s; savedpath = %s\n", info->dest,
 843                  info->savedpath);
 844         if (strncmp(info->dest, info->savedpath, PATH_MAX) != 0) {
 845                 lr_debug(DTRACE, "create: file moved (%s). %s != %s\n",
 846                          info->tfid, info->dest, info->savedpath);
 847                 mkspecial = 1;
 848         }
 849
 850         /* Is f2p(pfid) present on the target? If not, the parent has moved */
 851         if (!mkspecial) {
 852                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
 853                          status->ls_targets[0], info->path);
 854                 if (access(info->dest, F_OK) != 0) {
 855                         lr_debug(DTRACE, "create: parent %s not found\n",
 856                                  info->dest);
 857                         mkspecial = 1;
 858                 }
 859         }
 860         for (info->target_no = 0; info->target_no < status->ls_num_targets;
 861         info->target_no++) {
 862                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
 863                          status->ls_targets[info->target_no], info->savedpath);
 864                 lr_get_FID_PATH(status->ls_source, info->tfid, info->src,
 865                                 PATH_MAX);
 866
 867                 if (!mkspecial)
 868                         rc1 = lr_mkfile(info);
 869                 if (mkspecial || rc1 == -ENOENT)
 870                         rc1 = lr_mk_special(info);
 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                 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                 if (!rc_dest) {
 941                         snprintf(info->dest, sizeof(info->dest), "%s/%s",
 942                                  status->ls_targets[info->target_no],
 943                                  info->path);
 944                         if (access(info->dest, F_OK) != 0) {
 945                                 rc_dest = -errno;
 946                         } else {
 947                                 snprintf(info->dest, sizeof(info->dest),
 948                                          "%s/%s/%s",
 949                                          status->ls_targets[info->target_no],
 950                                          info->path, info->name);
 951                         }
 952                         lr_debug(DINFO, "dest path %s rc_dest=%d\n", info->dest,
 953                                  rc_dest);
 954                 }
 955                 if (rc_dest == -ENOENT) {
 956                         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
 957                                  status->ls_targets[info->target_no],
 958                                  SPECIAL_DIR, info->sfid);
 959                         special_dest = 1;
 960                         lr_debug(DINFO, "special dest %s\n", info->dest);
 961                 }
 962
 963                 if (!rc_src) {
 964                         snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 965                                  status->ls_targets[info->target_no],
 966                                  srcpath, info->sname);
 967                         lr_debug(DINFO, "src path %s rc_src=%d\n", info->src,
 968                                  rc_src);
 969                 }
 970                 if (rc_src == -ENOENT ||
 971                     (access(info->src, F_OK) != 0 && errno == ENOENT)) {
 972                         snprintf(info->src, sizeof(info->src), "%s/%s/%s",
 973                                  status->ls_targets[info->target_no],
 974                                  SPECIAL_DIR, info->sfid);
 975                         special_src = 1;
 976                         lr_debug(DINFO, "special src %s\n", info->src);
 977                 }
 978
 979                 rc1 = 0;
 980                 errno = 0;
 981                 if (strcmp(info->src, info->dest) != 0) {
 982                         rc1 = rename(info->src, info->dest);
 983                         if (rc1 == -1)
 984                                 rc1 = -errno;
 985                         lr_debug(DINFO, "rename returns %d\n", rc1);
 986                 }
 987
 988                 if (special_src)
 989                         rc1 = lr_remove_pc(info->spfid, info->sfid);
 990
 991                 if (!special_dest)
 992                         lr_cascade_move(info->sfid, info->dest, info);
 993                 else
 994                         rc1 = lr_add_pc(info->pfid, info->sfid, info->name);
 995
 996                 lr_debug(DINFO, "move: %s [to] %s rc1=%d, errno=%d\n",
 997                          info->src, info->dest, rc1, errno);
 998                 if (rc1)
 999                         rc = rc1;
1000         }
1001         return rc;
1002 }
1003
1004 /* Replicate a hard link */
1005 int lr_link(struct lr_info *info)
1006 {
1007         int i;
1008         int rc;
1009         int rc1;
1010         struct stat st;
1011
1012         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1013         rc = stat(info->src, &st);
1014         if (rc == -1)
1015                 return -errno;
1016
1017         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1018              info->target_no++) {
1019
1020                 info->src[0] = 0;
1021                 info->dest[0] = 0;
1022                 rc1 = 0;
1023
1024                 /*
1025                  * The changelog record has the new parent directory FID and
1026                  * name of the target file. So info->dest can be constructed
1027                  * by getting the path of the new parent directory and
1028                  * appending the target file name.
1029                  */
1030                 rc1 = lr_get_path(info, info->pfid);
1031                 lr_debug(rc1 ? 0 : DTRACE, "\tparent fid2path %s, %s, rc=%d\n",
1032                          info->path, info->name, rc1);
1033
1034                 if (rc1 == 0) {
1035                         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1036                                  status->ls_targets[info->target_no],
1037                                  info->path, info->name);
1038                         lr_debug(DINFO, "link destination is %s\n", info->dest);
1039                 }
1040
1041                 /* Search through the hardlinks to get the src */
1042                 for (i = 0; i < st.st_nlink && info->src[0] == 0; i++) {
1043                         size_t len;
1044
1045                         rc1 = lr_get_path_ln(info, info->tfid, i);
1046                         lr_debug(rc1 ? 0 : DTRACE,
1047                                  "\tfid2path %s, %s, %d rc=%d\n", info->path,
1048                                  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 int lr_set_dest_for_attr(struct lr_info *info)
1098 {
1099         int rc;
1100
1101         snprintf(info->dest, sizeof(info->dest), "%s/%s",
1102                  status->ls_targets[info->target_no], info->path);
1103         rc = access(info->dest, F_OK);
1104         if (rc < 0)
1105                 rc = -errno;
1106
1107         if (rc != -ENOENT)
1108                 return rc;
1109
1110         snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1111                  status->ls_targets[info->target_no], SPECIAL_DIR,
1112                  info->tfid);
1113
1114         rc = access(info->dest, F_OK);
1115         if (rc < 0)
1116                 return -errno;
1117
1118         return 0;
1119 }
1120
1121 /* Replicate file attributes */
1122 int lr_setattr(struct lr_info *info)
1123 {
1124         int rc1;
1125         int rc;
1126
1127         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1128
1129         rc = lr_get_path(info, info->tfid);
1130         if (rc == -ENOENT)
1131                 lr_debug(DINFO, "setattr: %s not present on source-fs\n",
1132                          info->src);
1133         if (rc)
1134                 return rc;
1135
1136         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1137              info->target_no++) {
1138                 rc = lr_set_dest_for_attr(info);
1139                 if (rc < 0)
1140                         continue;
1141
1142                 lr_debug(DINFO, "setattr: %s %s %s", info->src, info->dest,
1143                          info->tfid);
1144
1145                 rc1 = lr_sync_data(info);
1146                 if (!rc1)
1147                         rc1 = lr_copy_attr(info->src, info->dest);
1148                 if (rc1)
1149                         rc = rc1;
1150         }
1151         return rc;
1152 }
1153
1154 /* Replicate xattrs */
1155 int lr_setxattr(struct lr_info *info)
1156 {
1157         int rc, rc1;
1158
1159         lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1160
1161         rc = lr_get_path(info, info->tfid);
1162         if (rc == -ENOENT)
1163                 lr_debug(DINFO, "setxattr: %s not present on source-fs\n",
1164                          info->src);
1165         if (rc)
1166                 return rc;
1167
1168         for (info->target_no = 0; info->target_no < status->ls_num_targets;
1169              info->target_no++) {
1170                 rc = lr_set_dest_for_attr(info);
1171                 if (rc < 0)
1172                         continue;
1173
1174                 lr_debug(DINFO, "setxattr: %s %s %s\n", info->src, info->dest,
1175                          info->tfid);
1176
1177                 rc1 = lr_copy_xattr(info);
1178                 if (rc1)
1179                         rc = rc1;
1180         }
1181
1182         return rc;
1183 }
1184
1185 /* Parse a line of changelog entry */
1186 int lr_parse_line(void *priv, struct lr_info *info)
1187 {
1188         struct changelog_rec            *rec;
1189         struct changelog_ext_rename     *rnm;
1190         size_t                           namelen;
1191         size_t                           copylen = sizeof(info->name);
1192
1193         if (llapi_changelog_recv(priv, &rec) != 0)
1194                 return -1;
1195
1196         info->is_extended = !!(rec->cr_flags & CLF_RENAME);
1197         info->recno = rec->cr_index;
1198         info->type = rec->cr_type;
1199         snprintf(info->tfid, sizeof(info->tfid), DFID, PFID(&rec->cr_tfid));
1200         snprintf(info->pfid, sizeof(info->pfid), DFID, PFID(&rec->cr_pfid));
1201
1202         namelen = strnlen(changelog_rec_name(rec), rec->cr_namelen);
1203         if (copylen > namelen + 1)
1204                 copylen = namelen + 1;
1205         snprintf(info->name, copylen, "%s", changelog_rec_name(rec));
1206
1207         /* Don't use rnm if CLF_RENAME isn't set */
1208         rnm = changelog_rec_rename(rec);
1209         if (rec->cr_flags & CLF_RENAME && !fid_is_zero(&rnm->cr_sfid)) {
1210                 copylen = sizeof(info->sname);
1211
1212                 snprintf(info->sfid, sizeof(info->sfid), DFID,
1213                          PFID(&rnm->cr_sfid));
1214                 snprintf(info->spfid, sizeof(info->spfid), DFID,
1215                          PFID(&rnm->cr_spfid));
1216                 namelen = changelog_rec_snamelen(rec);
1217                 if (copylen > namelen + 1)
1218                         copylen = namelen + 1;
1219                 snprintf(info->sname, copylen, "%s", changelog_rec_sname(rec));
1220
1221                 if (verbose > 1)
1222                         printf("Rec %lld: %d %s %s\n", info->recno, info->type,
1223                                info->name, info->sname);
1224         } else {
1225                 if (verbose > 1)
1226                         printf("Rec %lld: %d %s\n", info->recno, info->type,
1227                                info->name);
1228         }
1229
1230         llapi_changelog_free(&rec);
1231
1232         rec_count++;
1233         return 0;
1234 }
1235
1236 /* Initialize the replication parameters */
1237 int lr_init_status(void)
1238 {
1239         size_t size = sizeof(struct lustre_rsync_status) + PATH_MAX + 1;
1240
1241         if (status)
1242                 return 0;
1243         status = calloc(size, 1);
1244         if (!status)
1245                 return -ENOMEM;
1246         status->ls_version = REPLICATE_STATUS_VER;
1247         status->ls_size = size;
1248         status->ls_last_recno = -1;
1249         return 0;
1250 }
1251
1252 /* Make a backup of the statuslog */
1253 void lr_backup_log(void)
1254 {
1255         char backupfile[PATH_MAX];
1256
1257         if (logbackedup)
1258                 return;
1259         snprintf(backupfile, sizeof(backupfile), "%s.old", statuslog);
1260         (void) rename(statuslog, backupfile);
1261         logbackedup = 1;
1262 }
1263
1264 /* Save replication parameters to a statuslog. */
1265 int lr_write_log(void)
1266 {
1267         int fd;
1268         size_t size;
1269         size_t write_size = status->ls_size;
1270         struct lr_parent_child_list *curr;
1271         int rc = 0;
1272
1273         if (!statuslog)
1274                 return 0;
1275
1276         lr_backup_log();
1277
1278         fd = open(statuslog, O_WRONLY | O_CREAT | O_SYNC, S_IRUSR | S_IWUSR);
1279         if (fd == -1) {
1280                 fprintf(stderr, "Error opening log file for writing (%s)\n",
1281                         statuslog);
1282                 return -1;
1283         }
1284         errno = 0;
1285         size = write(fd, status, write_size);
1286         if (size != write_size) {
1287                 fprintf(stderr, "Error writing to log file (%s) %d\n",
1288                         statuslog, errno);
1289                 close(fd);
1290                 return -1;
1291         }
1292
1293         for (curr = parents; curr; curr = curr->pc_next) {
1294                 size = write(fd, &curr->pc_log, sizeof(curr->pc_log));
1295                 if (size != sizeof(curr->pc_log)) {
1296                         fprintf(stderr, "Error writing to log file (%s) %d\n",
1297                                 statuslog, errno);
1298                         rc = -1;
1299                         break;
1300                 }
1301         }
1302         close(fd);
1303         return rc;
1304 }
1305
1306 /*
1307  * Read statuslog and populate the replication parameters.  Command
1308  * line parameters take precedence over parameters in the log file.
1309  */
1310 int lr_read_log(void)
1311 {
1312         struct lr_parent_child_list *tmp;
1313         struct lr_parent_child_log rec;
1314         struct lustre_rsync_status *s;
1315         int fd = -1;
1316         size_t size;
1317         size_t read_size = sizeof(struct lustre_rsync_status) + PATH_MAX + 1;
1318         int rc = 0;
1319
1320         if (!statuslog)
1321                 return 0;
1322
1323         s = calloc(1, read_size);
1324         if (!s) {
1325                 rc = -ENOMEM;
1326                 goto out;
1327         }
1328
1329         fd = open(statuslog, O_RDONLY);
1330         if (fd == -1) {
1331                 rc = -errno;
1332                 goto out;
1333         }
1334
1335         size = read(fd, s, read_size);
1336         if (size != read_size) {
1337                 rc = -EINVAL;
1338                 goto out;
1339         }
1340
1341         if (read_size < s->ls_size) {
1342                 read_size = s->ls_size;
1343                 s = lr_grow_buf(s, read_size);
1344                 if (!s) {
1345                         rc = -ENOMEM;
1346                         goto out;
1347                 }
1348
1349                 if (lseek(fd, 0, SEEK_SET) == -1) {
1350                         rc = -ENOMEM;
1351                         goto out;
1352                 }
1353
1354                 size = read(fd, s, read_size);
1355                 if (size != read_size) {
1356                         rc = -EINVAL;
1357                         goto out;
1358                 }
1359         }
1360
1361         while (read(fd, &rec, sizeof(rec)) != 0) {
1362                 tmp = calloc(1, sizeof(*tmp));
1363                 if (!tmp) {
1364                         rc = -ENOMEM;
1365                         goto out;
1366                 }
1367
1368                 tmp->pc_log = rec;
1369                 tmp->pc_next = parents;
1370                 parents = tmp;
1371         }
1372
1373         /* copy uninitialized fields to status */
1374         if (status->ls_num_targets == 0) {
1375                 if (status->ls_size != s->ls_size) {
1376                         status = lr_grow_buf(status, s->ls_size);
1377                         if (!status) {
1378                                 rc = -ENOMEM;
1379                                 goto out;
1380                         }
1381
1382                         status->ls_size = s->ls_size;
1383                 }
1384                 status->ls_num_targets = s->ls_num_targets;
1385                 memcpy(status->ls_targets, s->ls_targets,
1386                        (PATH_MAX + 1) * s->ls_num_targets);
1387         }
1388         if (status->ls_last_recno == -1)
1389                 status->ls_last_recno = s->ls_last_recno;
1390
1391         if (status->ls_registration[0] == '\0')
1392                 snprintf(status->ls_registration,
1393                          sizeof(status->ls_registration), "%s",
1394                          s->ls_registration);
1395
1396         if (status->ls_mdt_device[0] == '\0')
1397                 snprintf(status->ls_mdt_device,
1398                          sizeof(status->ls_mdt_device), "%s",
1399                          s->ls_mdt_device);
1400
1401         if (status->ls_source_fs[0] == '\0')
1402                 snprintf(status->ls_source_fs,
1403                          sizeof(status->ls_source_fs), "%s",
1404                          s->ls_source_fs);
1405
1406         if (status->ls_source[0] == '\0')
1407                 snprintf(status->ls_source,
1408                          sizeof(status->ls_source), "%s",
1409                          s->ls_source);
1410
1411 out:
1412         if (fd != -1)
1413                 close(fd);
1414         if (s)
1415                 free(s);
1416         return rc;
1417 }
1418
1419 /*
1420  * Clear changelogs every CLEAR_INTERVAL records or at the end of
1421  * processing.
1422  */
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                         /*
1431                          * llapi_changelog_clear modifies the mdt
1432                          * device name so make a copy of it until this
1433                          * is fixed.
1434                          */
1435                         snprintf(mdt_device, sizeof(mdt_device), "%s",
1436                                  status->ls_mdt_device);
1437                         rc = llapi_changelog_clear(mdt_device,
1438                                                    status->ls_registration,
1439                                                    info->recno);
1440                         if (rc)
1441                                 printf("Changelog clear (%s, %s, %lld) returned %d\n",
1442                                        status->ls_mdt_device,
1443                                        status->ls_registration, info->recno,
1444                                        rc);
1445                 }
1446
1447                 if (!rc && !dryrun) {
1448                         status->ls_last_recno = info->recno;
1449                         lr_write_log();
1450                 }
1451         }
1452
1453         return rc;
1454 }
1455
1456 /*
1457  * Locate a usable version of rsync. At this point we'll use any
1458  * version.
1459  */
1460 int lr_locate_rsync(void)
1461 {
1462         FILE *fp;
1463         int len;
1464
1465         /* Locate rsync */
1466         snprintf(rsync, sizeof(rsync), "%s -p %s", TYPE, RSYNC);
1467         fp = popen(rsync, "r");
1468         if (!fp)
1469                 return -1;
1470
1471         if (fgets(rsync, sizeof(rsync), fp) == NULL) {
1472                 fclose(fp);
1473                 return -1;
1474         }
1475
1476         len = strlen(rsync);
1477         if (len > 0 && rsync[len - 1] == '\n')
1478                 rsync[len - 1] = '\0';
1479         fclose(fp);
1480
1481         /* Determine the version of rsync */
1482         snprintf(rsync_ver, sizeof(rsync_ver), "%s --version", rsync);
1483         fp = popen(rsync_ver, "r");
1484         if (!fp)
1485                 return -1;
1486
1487         if (fgets(rsync_ver, sizeof(rsync_ver), fp) == NULL) {
1488                 fclose(fp);
1489                 return -1;
1490         }
1491         len = strlen(rsync_ver);
1492         if (len > 0 && rsync_ver[len - 1] == '\n')
1493                 rsync_ver[len - 1] = '\0';
1494         fclose(fp);
1495
1496         return 0;
1497 }
1498
1499 /* Print the replication parameters */
1500 void lr_print_status(struct lr_info *info)
1501 {
1502         int i;
1503
1504         if (!verbose)
1505                 return;
1506
1507         printf("Lustre filesystem: %s\n", status->ls_source_fs);
1508         printf("MDT device: %s\n", status->ls_mdt_device);
1509         printf("Source: %s\n", status->ls_source);
1510         for (i = 0; i < status->ls_num_targets; i++)
1511                 printf("Target: %s\n", status->ls_targets[i]);
1512         if (statuslog)
1513                 printf("Statuslog: %s\n", statuslog);
1514         printf("Changelog registration: %s\n", status->ls_registration);
1515         printf("Starting changelog record: %jd\n",
1516                (uintmax_t)status->ls_last_recno);
1517         if (noxattr)
1518                 printf("Replicate xattrs: no\n");
1519         if (noclear)
1520                 printf("Clear changelog after use: no\n");
1521         if (use_rsync)
1522                 printf("Using rsync: %s (%s)\n", rsync, rsync_ver);
1523 }
1524
1525 void lr_print_failure(struct lr_info *info, int rc)
1526 {
1527         fprintf(stderr,
1528                 "Replication of operation failed(%d): %lld %s (%d) %s %s %s\n",
1529                 rc, info->recno, changelog_type2str(info->type), info->type,
1530                 info->tfid, info->pfid, info->name);
1531 }
1532
1533 /* Replicate filesystem operations from src_path to target_path */
1534 int lr_replicate(void)
1535 {
1536         void *changelog_priv;
1537         struct lr_info *info;
1538         struct lr_info *ext = NULL;
1539         time_t start;
1540         int xattr_not_supp;
1541         int i;
1542         int rc;
1543
1544         start = time(NULL);
1545
1546         info = calloc(1, sizeof(struct lr_info));
1547         if (!info)
1548                 return -ENOMEM;
1549
1550         rc = llapi_search_fsname(status->ls_source, status->ls_source_fs);
1551         if (rc) {
1552                 fprintf(stderr, "Source path is not a valid Lustre client mountpoint.\n");
1553                 goto out;
1554         }
1555
1556         if (status->ls_mdt_device[0] == '\0') {
1557                 int len;
1558
1559                 len = snprintf(status->ls_mdt_device,
1560                                sizeof(status->ls_mdt_device), "%s%s",
1561                                status->ls_source_fs, DEFAULT_MDT);
1562                 if (len >= sizeof(status->ls_mdt_device)) {
1563                         rc = -E2BIG;
1564                         goto out;
1565                 }
1566         }
1567
1568         ext = calloc(1, sizeof(struct lr_info));
1569         if (!ext) {
1570                 rc = -ENOMEM;
1571                 goto out;
1572         }
1573
1574         for (i = 0, xattr_not_supp = 0; i < status->ls_num_targets; i++) {
1575                 snprintf(info->dest, sizeof(info->dest), "%s/%s",
1576                          status->ls_targets[i], SPECIAL_DIR);
1577                 rc = mkdir(info->dest, 0777);
1578                 if (rc == -1 && errno != EEXIST) {
1579                         fprintf(stderr, "Error writing to target path %s.\n",
1580                                 status->ls_targets[i]);
1581                         rc = -errno;
1582                         goto out;
1583                 }
1584                 rc = llistxattr(info->src, info->xlist, info->xsize);
1585                 if (rc == -1 && errno == ENOTSUP) {
1586                         fprintf(stderr, "xattrs not supported on %s\n",
1587                                 status->ls_targets[i]);
1588                         xattr_not_supp++;
1589                 }
1590         }
1591         if (xattr_not_supp == status->ls_num_targets)
1592                 /* None of the targets support xattrs. */
1593                 noxattr = 1;
1594
1595         lr_print_status(info);
1596
1597         /* Open changelogs for consumption*/
1598         rc = llapi_changelog_start(&changelog_priv,
1599                                    CHANGELOG_FLAG_BLOCK |
1600                                    CHANGELOG_FLAG_JOBID |
1601                                    CHANGELOG_FLAG_EXTRA_FLAGS,
1602                                    status->ls_mdt_device,
1603                                    status->ls_last_recno);
1604         if (rc < 0) {
1605                 fprintf(stderr, "Error opening changelog file for fs %s.\n",
1606                         status->ls_source_fs);
1607                 goto out;
1608         }
1609
1610         rc = llapi_changelog_set_xflags(changelog_priv,
1611                                         CHANGELOG_EXTRA_FLAG_UIDGID |
1612                                         CHANGELOG_EXTRA_FLAG_NID |
1613                                         CHANGELOG_EXTRA_FLAG_OMODE |
1614                                         CHANGELOG_EXTRA_FLAG_XATTR);
1615         if (rc < 0) {
1616                 fprintf(stderr, "Error setting xflag in changelog for fs %s.\n",
1617                         status->ls_source_fs);
1618                 goto out;
1619         }
1620
1621         while (!quit && lr_parse_line(changelog_priv, info) == 0) {
1622                 rc = 0;
1623
1624                 if (info->type == CL_RENAME && !info->is_extended) {
1625                         /*
1626                          * Newer rename operations extends changelog to store
1627                          * source file information, but old changelog has
1628                          * another record.
1629                          */
1630                         if (lr_parse_line(changelog_priv, ext) != 0)
1631                                 break;
1632                         memcpy(info->sfid, info->tfid, sizeof(info->sfid));
1633                         memcpy(info->spfid, info->pfid, sizeof(info->spfid));
1634                         memcpy(info->tfid, ext->tfid, sizeof(info->tfid));
1635                         memcpy(info->pfid, ext->pfid, sizeof(info->pfid));
1636                         snprintf(info->sname, sizeof(info->sname), "%s",
1637                                  info->name);
1638                         snprintf(info->name, sizeof(info->name), "%s",
1639                                  ext->name);
1640                         info->is_extended = 1;
1641                         info->recno = ext->recno; /* For lr_clear_cl(). */
1642                 }
1643
1644                 if (dryrun)
1645                         continue;
1646
1647                 lr_debug(DTRACE, "***** Start %lld %s (%d) %s %s %s *****\n",
1648                          info->recno, changelog_type2str(info->type),
1649                          info->type, info->tfid, info->pfid, info->name);
1650
1651                 switch (info->type) {
1652                 case CL_CREATE:
1653                 case CL_MKDIR:
1654                 case CL_MKNOD:
1655                 case CL_SOFTLINK:
1656                         rc = lr_create(info);
1657                         break;
1658                 case CL_RMDIR:
1659                 case CL_UNLINK:
1660                         rc = lr_remove(info);
1661                         break;
1662                 case CL_RENAME:
1663                         rc = lr_move(info);
1664                         break;
1665                 case CL_HARDLINK:
1666                         rc = lr_link(info);
1667                         break;
1668                 case CL_TRUNC:
1669                 case CL_SETATTR:
1670                         rc = lr_setattr(info);
1671                         break;
1672                 case CL_SETXATTR:
1673                         rc = lr_setxattr(info);
1674                         break;
1675                 case CL_CLOSE:
1676                 case CL_EXT:
1677                 case CL_OPEN:
1678                 case CL_GETXATTR:
1679                 case CL_DN_OPEN:
1680                 case CL_LAYOUT:
1681                 case CL_MARK:
1682                         /*
1683                          * Nothing needs to be done for these entries
1684                          * fallthrough
1685                          */
1686                 default:
1687                         break;
1688                 }
1689
1690                 lr_debug(DTRACE, "##### End %lld %s (%d) %s %s %s rc=%d #####\n",
1691                          info->recno, changelog_type2str(info->type),
1692                          info->type, info->tfid, info->pfid, info->name, rc);
1693
1694                 if (rc && rc != -ENOENT) {
1695                         lr_print_failure(info, rc);
1696                         errors++;
1697                         if (abort_on_err)
1698                                 break;
1699                 }
1700                 lr_clear_cl(info, 0);
1701         }
1702
1703         llapi_changelog_fini(&changelog_priv);
1704
1705         if (errors || verbose)
1706                 printf("Errors: %d\n", errors);
1707
1708         /* Clear changelog records used so far */
1709         lr_clear_cl(info, 1);
1710
1711         if (verbose) {
1712                 printf("lustre_rsync took %ld seconds\n", time(NULL) - start);
1713                 printf("Changelog records consumed: %lld\n", rec_count);
1714         }
1715
1716         rc = 0;
1717
1718 out:
1719         if (info)
1720                 free(info);
1721         if (ext)
1722                 free(ext);
1723
1724         return rc;
1725 }
1726
1727 void
1728 termination_handler (int signum)
1729 {
1730         /* Set a flag for the replicator to gracefully shutdown */
1731         quit = 1;
1732         printf("lustre_rsync halting.\n");
1733 }
1734
1735 int main(int argc, char *argv[])
1736 {
1737         int newsize;
1738         int numtargets = 0;
1739         int rc = 0;
1740
1741         if ((rc = lr_init_status()) != 0)
1742                 return rc;
1743
1744         while ((rc = getopt_long(argc, argv, "as:t:m:u:l:vx:zc:ry:n:d:D:",
1745                                  long_opts, NULL)) >= 0) {
1746                 switch (rc) {
1747                 case 'a':
1748                         /* Assume absolute paths */
1749                         abort_on_err++;
1750                         break;
1751                 case 's':
1752                         /* Assume absolute paths */
1753                         snprintf(status->ls_source, sizeof(status->ls_source),
1754                                  "%s", optarg);
1755                         break;
1756                 case 't':
1757                         status->ls_num_targets++;
1758                         numtargets++;
1759                         if (numtargets != status->ls_num_targets) {
1760                                 /*
1761                                  * Targets were read from a log file.
1762                                  * The ones specified on the command line
1763                                  * take precedence. The ones from the log
1764                                  * file will be ignored.
1765                                  */
1766                                 status->ls_num_targets = numtargets;
1767                         }
1768                         newsize = sizeof(struct lustre_rsync_status) +
1769                                 (status->ls_num_targets * (PATH_MAX + 1));
1770                         if (status->ls_size != newsize) {
1771                                 status->ls_size = newsize;
1772                                 status = lr_grow_buf(status, newsize);
1773                                 if (!status)
1774                                         return -ENOMEM;
1775                         }
1776                         snprintf(status->ls_targets[status->ls_num_targets - 1],
1777                                  sizeof(status->ls_targets[0]), "%s", optarg);
1778                         break;
1779                 case 'm':
1780                         snprintf(status->ls_mdt_device,
1781                                  sizeof(status->ls_mdt_device),
1782                                  "%s", optarg);
1783                         break;
1784                 case 'u':
1785                         snprintf(status->ls_registration,
1786                                  sizeof(status->ls_registration),
1787                                  "%s", optarg);
1788                         break;
1789                 case 'l':
1790                         statuslog = optarg;
1791                         (void)lr_read_log();
1792                         break;
1793                 case 'v':
1794                         verbose++;
1795                         break;
1796                 case 'x':
1797                         if (strcmp("no", optarg) == 0) {
1798                                 noxattr = 1;
1799                         } else if (strcmp("yes", optarg) != 0) {
1800                                 printf("Invalid parameter %s. Specify --xattr=no or --xattr=yes\n",
1801                                        optarg);
1802                                 return -1;
1803                         }
1804                         break;
1805                 case 'z':
1806                         dryrun = 1;
1807                         break;
1808                 case 'c':
1809                         /* Undocumented option cl-clear */
1810                         if (strcmp("no", optarg) == 0) {
1811                                 noclear = 1;
1812                         } else if (strcmp("yes", optarg) != 0) {
1813                                 printf("Invalid parameter %s. Specify --cl-clear=no or --cl-clear=yes\n",
1814                                        optarg);
1815                                 return -1;
1816                         }
1817                         break;
1818                 case 'r':
1819                         /* Undocumented option use-rsync */
1820                         use_rsync = 1;
1821                         break;
1822                 case 'y':
1823                         /* Undocumented option rsync-threshold */
1824                         rsync_threshold = atol(optarg);
1825                         break;
1826                 case 'n':
1827                         /* Undocumented option start-recno */
1828                         status->ls_last_recno = atol(optarg);
1829                         break;
1830                 case 'd':
1831                         /* Undocumented option debug */
1832                         debug = atoi(optarg);
1833                         if (debug < 0 || debug > 2)
1834                                 debug = 0;
1835                         break;
1836                 case 'D':
1837                         /* Undocumented option debug log file */
1838                         if (debug_log)
1839                                 fclose(debug_log);
1840                         debug_log = fopen(optarg, "a");
1841                         if (!debug_log) {
1842                                 printf("Cannot open %s for debug log\n",
1843                                        optarg);
1844                                 return -1;
1845                         }
1846                         break;
1847                 default:
1848                         fprintf(stderr, "error: %s: option '%s' unrecognized.\n",
1849                                 argv[0], argv[optind - 1]);
1850                         lr_usage();
1851                         return -1;
1852                 }
1853         }
1854
1855         if (status->ls_last_recno == -1)
1856                 status->ls_last_recno = 0;
1857         if (strnlen(status->ls_registration, LR_NAME_MAXLEN) == 0) {
1858                 /* No registration ID was passed in. */
1859                 printf("Please specify changelog consumer registration id.\n");
1860                 lr_usage();
1861                 return -1;
1862         }
1863         if (strnlen(status->ls_source, PATH_MAX) == 0) {
1864                 fprintf(stderr, "Please specify the source path.\n");
1865                 lr_usage();
1866                 return -1;
1867         }
1868         if (strnlen(status->ls_targets[0], PATH_MAX) == 0) {
1869                 fprintf(stderr, "Please specify the target path.\n");
1870                 lr_usage();
1871                 return -1;
1872         }
1873
1874         rc = lr_locate_rsync();
1875         if (use_rsync && rc != 0) {
1876                 fprintf(stderr, "Error: unable to locate %s.\n", RSYNC);
1877                 exit(-1);
1878         }
1879
1880         signal(SIGINT, termination_handler);
1881         signal(SIGHUP, termination_handler);
1882         signal(SIGTERM, termination_handler);
1883
1884         rc = lr_replicate();
1885
1886         if (debug_log)
1887                 fclose(debug_log);
1888         return rc;
1889 }