1 /* Cumulus: Smart Filesystem Backup to Dumb Servers
3 * Copyright (C) 2006-2008 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License along
17 * with this program; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
21 /* Main entry point for LBS. Contains logic for traversing the filesystem and
22 * constructing a backup. */
34 #include <sys/sysmacros.h>
35 #include <sys/types.h>
61 /* Version information. This will be filled in by the Makefile. */
62 #ifndef CUMULUS_VERSION
63 #define CUMULUS_VERSION Unknown
65 #define CUMULUS_STRINGIFY(s) CUMULUS_STRINGIFY2(s)
66 #define CUMULUS_STRINGIFY2(s) #s
67 static const char cumulus_version[] = CUMULUS_STRINGIFY(CUMULUS_VERSION);
69 static RemoteStore *remote = NULL;
70 static TarSegmentStore *tss = NULL;
71 static MetadataWriter *metawriter = NULL;
73 /* Buffer for holding a single block of data read from a file. */
74 static const size_t LBS_BLOCK_SIZE = 1024 * 1024;
75 static char *block_buf;
77 /* Local database, which tracks objects written in this and previous
78 * invocations to help in creating incremental snapshots. */
81 /* Keep track of all segments which are needed to reconstruct the snapshot. */
82 std::set<string> segment_list;
84 /* Snapshot intent: 1=daily, 7=weekly, etc. This is not used directly, but is
85 * stored in the local database and can help guide segment cleaning and
86 * snapshot expiration policies. */
87 double snapshot_intent = 1.0;
89 /* Selection of files to include/exclude in the snapshot. */
90 std::list<string> includes; // Paths in which files should be saved
91 std::list<string> excludes; // Paths which will not be saved
92 std::list<string> searches; // Directories we don't want to save, but
93 // do want to descend searching for data
96 bool relative_paths = true;
98 /* Whether verbose output is enabled. */
101 /* Ensure that the given segment is listed as a dependency of the current
103 void add_segment(const string& segment)
105 segment_list.insert(segment);
108 /* Read data from a file descriptor and return the amount of data read. A
109 * short read (less than the requested size) will only occur if end-of-file is
111 ssize_t file_read(int fd, char *buf, size_t maxlen)
113 size_t bytes_read = 0;
116 ssize_t res = read(fd, buf, maxlen);
120 fprintf(stderr, "error reading file: %m\n");
122 } else if (res == 0) {
134 /* Read the contents of a file (specified by an open file descriptor) and copy
135 * the data to the store. Returns the size of the file (number of bytes
136 * dumped), or -1 on error. */
137 int64_t dumpfile(int fd, dictionary &file_info, const string &path,
138 struct stat& stat_buf)
141 list<string> object_list;
142 const char *status = NULL; /* Status indicator printed out */
144 /* Look up this file in the old stat cache, if we can. If the stat
145 * information indicates that the file has not changed, do not bother
146 * re-reading the entire contents. Even if the information has been
147 * changed, we can use the list of old blocks in the search for a sub-block
148 * incremental representation. */
150 list<ObjectReference> old_blocks;
152 bool found = metawriter->find(path);
154 old_blocks = metawriter->get_blocks();
156 if (found && metawriter->is_unchanged(&stat_buf)) {
159 /* If any of the blocks in the object have been expired, then we should
160 * fall back to fully reading in the file. */
161 for (list<ObjectReference>::const_iterator i = old_blocks.begin();
162 i != old_blocks.end(); ++i) {
163 const ObjectReference &ref = *i;
164 if (!db->IsAvailable(ref)) {
171 /* If everything looks okay, use the cached information */
173 file_info["checksum"] = metawriter->get_checksum();
174 for (list<ObjectReference>::const_iterator i = old_blocks.begin();
175 i != old_blocks.end(); ++i) {
176 const ObjectReference &ref = *i;
177 object_list.push_back(ref.to_string());
179 add_segment(ref.get_segment());
182 size = stat_buf.st_size;
186 /* If the file is new or changed, we must read in the contents a block at a
191 subfile.load_old_blocks(old_blocks);
194 ssize_t bytes = file_read(fd, block_buf, LBS_BLOCK_SIZE);
198 fprintf(stderr, "Backup contents for %s may be incorrect\n",
203 hash.process(block_buf, bytes);
205 // Sparse file processing: if we read a block of all zeroes, encode
207 bool all_zero = true;
208 for (int i = 0; i < bytes; i++) {
209 if (block_buf[i] != 0) {
215 // Either find a copy of this block in an already-existing segment,
216 // or index it so it can be re-used in the future
217 double block_age = 0.0;
220 SHA1Checksum block_hash;
221 block_hash.process(block_buf, bytes);
222 string block_csum = block_hash.checksum_str();
225 ref = ObjectReference(ObjectReference::REF_ZERO);
226 ref.set_range(0, bytes);
228 ref = db->FindObject(block_csum, bytes);
231 list<ObjectReference> refs;
233 // Store a copy of the object if one does not yet exist
235 LbsObject *o = new LbsObject;
238 /* We might still have seen this checksum before, if the object
239 * was stored at some time in the past, but we have decided to
240 * clean the segment the object was originally stored in
241 * (FindObject will not return such objects). When rewriting
242 * the object contents, put it in a separate group, so that old
243 * objects get grouped together. The hope is that these old
244 * objects will continue to be used in the future, and we
245 * obtain segments which will continue to be well-utilized.
246 * Additionally, keep track of the age of the data by looking
247 * up the age of the block which was expired and using that
248 * instead of the current time. */
249 if (db->IsOldObject(block_csum, bytes,
250 &block_age, &object_group)) {
251 if (object_group == 0) {
252 o->set_group("data");
255 sprintf(group, "compacted-%d", object_group);
261 o->set_group("data");
265 subfile.analyze_new_block(block_buf, bytes);
266 refs = subfile.create_incremental(tss, o, block_age);
271 while (!refs.empty()) {
272 ref = refs.front(); refs.pop_front();
273 object_list.push_back(ref.to_string());
275 add_segment(ref.get_segment());
284 file_info["checksum"] = hash.checksum_str();
287 if (verbose && status != NULL)
288 printf(" [%s]\n", status);
290 string blocklist = "";
291 for (list<string>::iterator i = object_list.begin();
292 i != object_list.end(); ++i) {
293 if (i != object_list.begin())
297 file_info["data"] = blocklist;
302 /* Dump a specified filesystem object (file, directory, etc.) based on its
303 * inode information. If the object is a regular file, an open filehandle is
305 void dump_inode(const string& path, // Path within snapshot
306 const string& fullpath, // Path to object in filesystem
307 struct stat& stat_buf, // Results of stat() call
308 int fd) // Open filehandle if regular file
311 dictionary file_info;
316 printf("%s\n", path.c_str());
317 metawriter->find(path);
319 file_info["name"] = uri_encode(path);
320 file_info["mode"] = encode_int(stat_buf.st_mode & 07777, 8);
321 file_info["ctime"] = encode_int(stat_buf.st_ctime);
322 file_info["mtime"] = encode_int(stat_buf.st_mtime);
323 file_info["user"] = encode_int(stat_buf.st_uid);
324 file_info["group"] = encode_int(stat_buf.st_gid);
326 time_t now = time(NULL);
327 if (now - stat_buf.st_ctime < 30 || now - stat_buf.st_mtime < 30)
328 if ((stat_buf.st_mode & S_IFMT) != S_IFDIR)
329 file_info["volatile"] = "1";
331 struct passwd *pwd = getpwuid(stat_buf.st_uid);
333 file_info["user"] += " (" + uri_encode(pwd->pw_name) + ")";
336 struct group *grp = getgrgid(stat_buf.st_gid);
338 file_info["group"] += " (" + uri_encode(grp->gr_name) + ")";
341 if (stat_buf.st_nlink > 1 && (stat_buf.st_mode & S_IFMT) != S_IFDIR) {
342 file_info["links"] = encode_int(stat_buf.st_nlink);
345 file_info["inode"] = encode_int(major(stat_buf.st_dev))
346 + "/" + encode_int(minor(stat_buf.st_dev))
347 + "/" + encode_int(stat_buf.st_ino);
351 switch (stat_buf.st_mode & S_IFMT) {
360 inode_type = ((stat_buf.st_mode & S_IFMT) == S_IFBLK) ? 'b' : 'c';
361 file_info["device"] = encode_int(major(stat_buf.st_rdev))
362 + "/" + encode_int(minor(stat_buf.st_rdev));
367 /* Use the reported file size to allocate a buffer large enough to read
368 * the symlink. Allocate slightly more space, so that we ask for more
369 * bytes than we expect and so check for truncation. */
370 buf = new char[stat_buf.st_size + 2];
371 len = readlink(fullpath.c_str(), buf, stat_buf.st_size + 1);
373 fprintf(stderr, "error reading symlink: %m\n");
374 } else if (len <= stat_buf.st_size) {
376 file_info["target"] = uri_encode(buf);
377 } else if (len > stat_buf.st_size) {
378 fprintf(stderr, "error reading symlink: name truncated\n");
386 file_size = dumpfile(fd, file_info, path, stat_buf);
387 file_info["size"] = encode_int(file_size);
390 return; // error occurred; do not dump file
392 if (file_size != stat_buf.st_size) {
393 fprintf(stderr, "Warning: Size of %s changed during reading\n",
395 file_info["volatile"] = "1";
404 fprintf(stderr, "Unknown inode type: mode=%x\n", stat_buf.st_mode);
408 file_info["type"] = string(1, inode_type);
410 metawriter->add(file_info);
413 void scanfile(const string& path, bool include)
417 struct stat stat_buf;
424 true_path = "/" + path;
426 // Set to true if we should scan through the contents of this directory,
427 // but not actually back files up
428 bool scan_only = false;
430 // Check this file against the include/exclude list to see if it should be
432 for (list<string>::iterator i = includes.begin();
433 i != includes.end(); ++i) {
439 for (list<string>::iterator i = excludes.begin();
440 i != excludes.end(); ++i) {
446 for (list<string>::iterator i = searches.begin();
447 i != searches.end(); ++i) {
453 if (!include && !scan_only)
456 if (lstat(true_path.c_str(), &stat_buf) < 0) {
457 fprintf(stderr, "lstat(%s): %m\n", path.c_str());
461 if ((stat_buf.st_mode & S_IFMT) == S_IFREG) {
462 /* Be paranoid when opening the file. We have no guarantee that the
463 * file was not replaced between the stat() call above and the open()
464 * call below, so we might not even be opening a regular file. We
465 * supply flags to open to to guard against various conditions before
466 * we can perform an lstat to check that the file is still a regular
468 * - O_NOFOLLOW: in the event the file was replaced by a symlink
469 * - O_NONBLOCK: prevents open() from blocking if the file was
471 * We also add in O_NOATIME, since this may reduce disk writes (for
472 * inode updates). However, O_NOATIME may result in EPERM, so if the
473 * initial open fails, try again without O_NOATIME. */
474 fd = open(true_path.c_str(), O_RDONLY|O_NOATIME|O_NOFOLLOW|O_NONBLOCK);
476 fd = open(true_path.c_str(), O_RDONLY|O_NOFOLLOW|O_NONBLOCK);
479 fprintf(stderr, "Unable to open file %s: %m\n", path.c_str());
483 /* Drop the use of the O_NONBLOCK flag; we only wanted that for file
485 flags = fcntl(fd, F_GETFL);
486 fcntl(fd, F_SETFL, flags & ~O_NONBLOCK);
488 /* Perform the stat call again, and check that we still have a regular
490 if (fstat(fd, &stat_buf) < 0) {
491 fprintf(stderr, "fstat: %m\n");
496 if ((stat_buf.st_mode & S_IFMT) != S_IFREG) {
497 fprintf(stderr, "file is no longer a regular file!\n");
503 dump_inode(path, true_path, stat_buf, fd);
508 // If we hit a directory, now that we've written the directory itself,
509 // recursively scan the directory.
510 if ((stat_buf.st_mode & S_IFMT) == S_IFDIR) {
511 DIR *dir = opendir(true_path.c_str());
514 fprintf(stderr, "Error: %m\n");
519 vector<string> contents;
520 while ((ent = readdir(dir)) != NULL) {
521 string filename(ent->d_name);
522 if (filename == "." || filename == "..")
524 contents.push_back(filename);
529 sort(contents.begin(), contents.end());
531 for (vector<string>::iterator i = contents.begin();
532 i != contents.end(); ++i) {
533 const string& filename = *i;
535 scanfile(filename, include);
537 scanfile(path + "/" + filename, include);
542 /* Include the specified file path in the backups. Append the path to the
543 * includes list, and to ensure that we actually see the path when scanning the
544 * directory tree, add all the parent directories to the search list, which
545 * means we will scan through the directory listing even if the files
546 * themselves are excluded from being backed up. */
547 void add_include(const char *path)
549 /* Was an absolute path specified? If so, we'll need to start scanning
550 * from the root directory. Make sure that the user was consistent in
551 * providing either all relative paths or all absolute paths. */
552 if (path[0] == '/') {
553 if (includes.size() > 0 && relative_paths == true) {
555 "Error: Cannot mix relative and absolute paths!\n");
559 relative_paths = false;
561 // Skip over leading '/'
563 } else if (relative_paths == false && path[0] != '/') {
564 fprintf(stderr, "Error: Cannot mix relative and absolute paths!\n");
568 includes.push_back(path);
570 /* Split the specified path into directory components, and ensure that we
571 * descend into all the directories along the path. */
572 const char *slash = path;
577 while ((slash = strchr(slash + 1, '/')) != NULL) {
578 string component(path, slash - path);
579 searches.push_back(component);
583 void usage(const char *program)
588 "Usage: %s [OPTION]... --dest=DEST PATHS...\n"
589 "Produce backup snapshot of files in SOURCE and store to DEST.\n"
592 " --dest=PATH path where backup is to be written\n"
593 " --upload-script=COMMAND\n"
594 " program to invoke for each backup file generated\n"
595 " --exclude=PATH exclude files in PATH from snapshot\n"
596 " --localdb=PATH local backup metadata is stored in PATH\n"
597 " --tmpdir=PATH path for temporarily storing backup files\n"
598 " (defaults to TMPDIR environment variable or /tmp)\n"
599 " --filter=COMMAND program through which to filter segment data\n"
600 " (defaults to \"bzip2 -c\")\n"
601 " --filter-extension=EXT\n"
602 " string to append to segment files\n"
603 " (defaults to \".bz2\")\n"
604 " --signature-filter=COMMAND\n"
605 " program though which to filter descriptor\n"
606 " --scheme=NAME optional name for this snapshot\n"
607 " --intent=FLOAT intended backup type: 1=daily, 7=weekly, ...\n"
608 " (defaults to \"1\")\n"
609 " --full-metadata do not re-use metadata from previous backups\n"
610 " -v --verbose list files as they are backed up\n"
612 "Exactly one of --dest or --upload-script must be specified.\n",
613 cumulus_version, program
617 int main(int argc, char *argv[])
619 string backup_dest = "", backup_script = "";
620 string localdb_dir = "";
621 string backup_scheme = "";
622 string signature_filter = "";
624 string tmp_dir = "/tmp";
625 if (getenv("TMPDIR") != NULL)
626 tmp_dir = getenv("TMPDIR");
629 static struct option long_options[] = {
630 {"localdb", 1, 0, 0}, // 0
631 {"exclude", 1, 0, 0}, // 1
632 {"filter", 1, 0, 0}, // 2
633 {"filter-extension", 1, 0, 0}, // 3
634 {"dest", 1, 0, 0}, // 4
635 {"scheme", 1, 0, 0}, // 5
636 {"signature-filter", 1, 0, 0}, // 6
637 {"intent", 1, 0, 0}, // 7
638 {"full-metadata", 0, 0, 0}, // 8
639 {"tmpdir", 1, 0, 0}, // 9
640 {"upload-script", 1, 0, 0}, // 10
641 // Aliases for short options
642 {"verbose", 0, 0, 'v'},
647 int c = getopt_long(argc, argv, "v", long_options, &long_index);
653 switch (long_index) {
655 localdb_dir = optarg;
658 if (optarg[0] != '/')
659 excludes.push_back(optarg);
661 excludes.push_back(optarg + 1);
664 filter_program = optarg;
666 case 3: // --filter-extension
667 filter_extension = optarg;
670 backup_dest = optarg;
673 backup_scheme = optarg;
675 case 6: // --signature-filter
676 signature_filter = optarg;
679 snapshot_intent = atof(optarg);
680 if (snapshot_intent <= 0)
683 case 8: // --full-metadata
684 flag_full_metadata = true;
689 case 10: // --upload-script
690 backup_script = optarg;
693 fprintf(stderr, "Unhandled long option!\n");
708 if (optind == argc) {
713 searches.push_back(".");
714 for (int i = optind; i < argc; i++)
715 add_include(argv[i]);
717 if (backup_dest == "" && backup_script == "") {
719 "Error: Backup destination must be specified using --dest= or --upload-script=\n");
724 if (backup_dest != "" && backup_script != "") {
726 "Error: Cannot specify both --dest= and --upload-script=\n");
731 // Default for --localdb is the same as --dest
732 if (localdb_dir == "") {
733 localdb_dir = backup_dest;
735 if (localdb_dir == "") {
737 "Error: Must specify local database path with --localdb=\n");
742 block_buf = new char[LBS_BLOCK_SIZE];
744 /* Initialize the remote storage layer. If using an upload script, create
745 * a temporary directory for staging files. Otherwise, write backups
746 * directly to the destination directory. */
747 if (backup_script != "") {
748 tmp_dir = tmp_dir + "/lbs." + generate_uuid();
749 if (mkdir(tmp_dir.c_str(), 0700) < 0) {
750 fprintf(stderr, "Cannot create temporary directory %s: %m\n",
754 remote = new RemoteStore(tmp_dir);
755 remote->set_script(backup_script);
757 remote = new RemoteStore(backup_dest);
760 /* Store the time when the backup started, so it can be included in the
766 localtime_r(&now, &time_buf);
767 strftime(desc_buf, sizeof(desc_buf), "%Y%m%dT%H%M%S", &time_buf);
769 /* Open the local database which tracks all objects that are stored
770 * remotely, for efficient incrementals. Provide it with the name of this
772 string database_path = localdb_dir + "/localdb.sqlite";
774 db->Open(database_path.c_str(), desc_buf, backup_scheme.c_str(),
777 tss = new TarSegmentStore(remote, db);
779 /* Initialize the stat cache, for skipping over unchanged files. */
780 metawriter = new MetadataWriter(tss, localdb_dir.c_str(), desc_buf,
781 backup_scheme.c_str());
783 scanfile(".", false);
785 ObjectReference root_ref = metawriter->close();
786 add_segment(root_ref.get_segment());
787 string backup_root = root_ref.to_string();
795 /* Write out a checksums file which lists the checksums for all the
796 * segments included in this snapshot. The format is designed so that it
797 * may be easily verified using the sha1sums command. */
798 const char csum_type[] = "sha1";
799 string checksum_filename = "snapshot-";
800 if (backup_scheme.size() > 0)
801 checksum_filename += backup_scheme + "-";
802 checksum_filename = checksum_filename + desc_buf + "." + csum_type + "sums";
803 RemoteFile *checksum_file = remote->alloc_file(checksum_filename,
805 FILE *checksums = fdopen(checksum_file->get_fd(), "w");
807 for (std::set<string>::iterator i = segment_list.begin();
808 i != segment_list.end(); ++i) {
809 string seg_path, seg_csum;
810 if (db->GetSegmentChecksum(*i, &seg_path, &seg_csum)) {
811 const char *raw_checksum = NULL;
812 if (strncmp(seg_csum.c_str(), csum_type,
813 strlen(csum_type)) == 0) {
814 raw_checksum = seg_csum.c_str() + strlen(csum_type);
815 if (*raw_checksum == '=')
821 if (raw_checksum != NULL)
822 fprintf(checksums, "%s *%s\n",
823 raw_checksum, seg_path.c_str());
828 SHA1Checksum checksum_csum;
830 checksum_filename = checksum_file->get_local_path();
831 if (checksum_csum.process_file(checksum_filename.c_str())) {
832 csum = checksum_csum.checksum_str();
835 checksum_file->send();
839 /* All other files should be flushed to remote storage before writing the
840 * backup descriptor below, so that it is not possible to have a backup
841 * descriptor written out depending on non-existent (not yet written)
845 /* Write a backup descriptor file, which says which segments are needed and
846 * where to start to restore this snapshot. The filename is based on the
847 * current time. If a signature filter program was specified, filter the
848 * data through that to give a chance to sign the descriptor contents. */
849 string desc_filename = "snapshot-";
850 if (backup_scheme.size() > 0)
851 desc_filename += backup_scheme + "-";
852 desc_filename = desc_filename + desc_buf + ".lbs";
854 RemoteFile *descriptor_file = remote->alloc_file(desc_filename,
856 int descriptor_fd = descriptor_file->get_fd();
857 if (descriptor_fd < 0) {
858 fprintf(stderr, "Unable to open descriptor output file: %m\n");
861 pid_t signature_pid = 0;
862 if (signature_filter.size() > 0) {
863 int new_fd = spawn_filter(descriptor_fd, signature_filter.c_str(),
865 close(descriptor_fd);
866 descriptor_fd = new_fd;
868 FILE *descriptor = fdopen(descriptor_fd, "w");
870 fprintf(descriptor, "Format: LBS Snapshot v0.6\n");
871 fprintf(descriptor, "Producer: Cumulus %s\n", cumulus_version);
872 strftime(desc_buf, sizeof(desc_buf), "%Y-%m-%d %H:%M:%S %z", &time_buf);
873 fprintf(descriptor, "Date: %s\n", desc_buf);
874 if (backup_scheme.size() > 0)
875 fprintf(descriptor, "Scheme: %s\n", backup_scheme.c_str());
876 fprintf(descriptor, "Backup-Intent: %g\n", snapshot_intent);
877 fprintf(descriptor, "Root: %s\n", backup_root.c_str());
879 if (csum.size() > 0) {
880 fprintf(descriptor, "Checksums: %s\n", csum.c_str());
883 fprintf(descriptor, "Segments:\n");
884 for (std::set<string>::iterator i = segment_list.begin();
885 i != segment_list.end(); ++i) {
886 fprintf(descriptor, " %s\n", i->c_str());
893 waitpid(signature_pid, &status, 0);
895 if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
896 throw IOException("Signature filter process error");
900 descriptor_file->send();
905 if (backup_script != "") {
906 if (rmdir(tmp_dir.c_str()) < 0) {
908 "Warning: Cannot delete temporary directory %s: %m\n",