1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2010 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
10 #define _ATFILE_SOURCE
18 #include <sys/types.h>
24 #include "bluesky-private.h"
26 /* The logging layer for BlueSky. This is used to write filesystem changes
27 * durably to disk so that they can be recovered in the event of a system
30 /* The logging layer takes care out writing out a sequence of log records to
31 * disk. On disk, each record consists of a header, a data payload, and a
32 * footer. The footer contains a checksum of the record, meant to help with
33 * identifying corrupt log records (we would assume because the log record was
34 * only incompletely written out before a crash, which should only happen for
35 * log records that were not considered committed). */
37 // Rough size limit for a log segment. This is not a firm limit and there are
38 // no absolute guarantees on the size of a log segment.
39 #define LOG_SEGMENT_SIZE (1 << 22)
41 #define HEADER_MAGIC 0x676f4c0a
42 #define FOOTER_MAGIC 0x2e435243
45 uint32_t magic; // HEADER_MAGIC
46 uint8_t type; // Object type + '0'
47 uint32_t offset; // Starting byte offset of the log header
48 uint32_t size1; // Size of the data item (bytes)
51 uint64_t inum; // Inode which owns this data, if any
52 BlueSkyCloudID id; // Object identifier
53 } __attribute__((packed));
56 uint32_t magic; // FOOTER_MAGIC
57 uint32_t crc; // Computed from log_header to log_footer.magic
58 } __attribute__((packed));
60 static void writebuf(int fd, const char *buf, size_t len)
64 written = write(fd, buf, len);
65 if (written < 0 && errno == EINTR)
67 g_assert(written >= 0);
73 static void log_commit(BlueSkyLog *log)
81 while (log->committed != NULL) {
82 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
83 g_mutex_lock(item->lock);
84 bluesky_cloudlog_stats_update(item, -1);
85 item->pending_write &= ~CLOUDLOG_JOURNAL;
86 item->location_flags |= CLOUDLOG_JOURNAL;
87 bluesky_cloudlog_stats_update(item, 1);
88 g_cond_signal(item->cond);
89 g_mutex_unlock(item->lock);
90 log->committed = g_slist_delete_link(log->committed, log->committed);
91 bluesky_cloudlog_unref(item);
95 if (bluesky_verbose && batchsize > 1)
96 g_print("Log batch size: %d\n", batchsize);
99 static gboolean log_open(BlueSkyLog *log)
110 if (log->current_log != NULL) {
111 bluesky_cachefile_unref(log->current_log);
112 log->current_log = NULL;
115 while (log->fd < 0) {
116 g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
117 log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
118 if (log->fd < 0 && errno == EEXIST) {
119 fprintf(stderr, "Log file %s already exists...\n", logname);
122 } else if (log->fd < 0) {
123 fprintf(stderr, "Error opening logfile %s: %m\n", logname);
128 log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
129 g_assert(log->current_log != NULL);
130 g_mutex_unlock(log->current_log->lock);
132 if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
133 fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
140 /* All log writes (at least for a single log) are made by one thread, so we
141 * don't need to worry about concurrent access to the log file. Log items to
142 * write are pulled off a queue (and so may be posted by any thread).
143 * fdatasync() is used to ensure the log items are stable on disk.
145 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
146 * each. If a log segment is not currently open (log->fd is negative), a new
147 * one is created. Log segment filenames are assigned sequentially.
149 * Log replay ought to be implemented later, and ought to set the initial
150 * sequence number appropriately.
152 static gpointer log_thread(gpointer d)
154 BlueSkyLog *log = (BlueSkyLog *)d;
158 if (!log_open(log)) {
163 BlueSkyCloudLog *item
164 = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
165 g_mutex_lock(item->lock);
166 g_assert(item->data != NULL);
168 /* The item may have already been written to the journal... */
169 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
170 g_mutex_unlock(item->lock);
171 bluesky_cloudlog_unref(item);
172 g_atomic_int_add(&item->data_lock_count, -1);
176 bluesky_cloudlog_stats_update(item, -1);
177 item->pending_write |= CLOUDLOG_JOURNAL;
178 bluesky_cloudlog_stats_update(item, 1);
180 GString *data1 = g_string_new("");
181 GString *data2 = g_string_new("");
182 GString *data3 = g_string_new("");
183 bluesky_serialize_cloudlog(item, data1, data2, data3);
185 struct log_header header;
186 struct log_footer footer;
187 size_t size = sizeof(header) + sizeof(footer);
188 size += data1->len + data2->len + data3->len;
191 offset = lseek(log->fd, 0, SEEK_CUR);
193 /* Check whether the item would overflow the allocated journal size.
194 * If so, start a new log segment. We only allow oversized log
195 * segments if they contain a single log entry. */
196 if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
201 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
202 header.offset = GUINT32_TO_LE(offset);
203 header.size1 = GUINT32_TO_LE(data1->len);
204 header.size2 = GUINT32_TO_LE(data2->len);
205 header.size3 = GUINT32_TO_LE(data3->len);
206 header.type = item->type + '0';
207 header.id = item->id;
208 header.inum = GUINT64_TO_LE(item->inum);
209 footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
211 uint32_t crc = BLUESKY_CRC32C_SEED;
213 writebuf(log->fd, (const char *)&header, sizeof(header));
214 crc = crc32c(crc, (const char *)&header, sizeof(header));
216 writebuf(log->fd, data1->str, data1->len);
217 crc = crc32c(crc, data1->str, data1->len);
218 writebuf(log->fd, data2->str, data2->len);
219 crc = crc32c(crc, data2->str, data2->len);
220 writebuf(log->fd, data3->str, data3->len);
221 crc = crc32c(crc, data3->str, data3->len);
223 crc = crc32c(crc, (const char *)&footer,
224 sizeof(footer) - sizeof(uint32_t));
225 footer.crc = crc32c_finalize(crc);
226 writebuf(log->fd, (const char *)&footer, sizeof(footer));
228 item->log_seq = log->seq_num;
229 item->log_offset = offset + sizeof(header);
230 item->log_size = item->data->len;
232 offset += sizeof(header) + sizeof(footer) + item->data->len;
234 g_string_free(data1, TRUE);
235 g_string_free(data2, TRUE);
236 g_string_free(data3, TRUE);
238 /* Replace the log item's string data with a memory-mapped copy of the
239 * data, now that it has been written to the log file. (Even if it
240 * isn't yet on disk, it should at least be in the page cache and so
241 * available to memory map.) */
242 bluesky_string_unref(item->data);
244 bluesky_cloudlog_fetch(item);
246 log->committed = g_slist_prepend(log->committed, item);
247 g_atomic_int_add(&item->data_lock_count, -1);
248 g_mutex_unlock(item->lock);
250 /* Force an if there are no other log items currently waiting to be
252 if (g_async_queue_length(log->queue) <= 0)
259 BlueSkyLog *bluesky_log_new(const char *log_directory)
261 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
263 log->log_directory = g_strdup(log_directory);
266 log->queue = g_async_queue_new();
267 log->mmap_lock = g_mutex_new();
268 log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
270 log->dirfd = open(log->log_directory, O_DIRECTORY);
271 if (log->dirfd < 0) {
272 fprintf(stderr, "Unable to open logging directory: %m\n");
276 g_thread_create(log_thread, log, FALSE, NULL);
281 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
283 bluesky_cloudlog_ref(item);
284 g_atomic_int_add(&item->data_lock_count, 1);
285 g_async_queue_push(log->queue, item);
288 void bluesky_log_finish_all(GList *log_items)
290 while (log_items != NULL) {
291 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
293 g_mutex_lock(item->lock);
294 while ((item->pending_write & CLOUDLOG_JOURNAL))
295 g_cond_wait(item->cond, item->lock);
296 g_mutex_unlock(item->lock);
297 bluesky_cloudlog_unref(item);
299 log_items = g_list_delete_link(log_items, log_items);
303 /* Memory-map the given log object into memory (read-only) and return a pointer
305 static int page_size = 0;
307 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
309 g_atomic_int_add(&cachefile->refcount, -1);
312 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
313 BlueSkyCacheFile *cachefile);
315 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
317 g_atomic_int_inc(&cachefile->refcount);
318 cachefile->fetching = TRUE;
319 g_print("Starting fetch of %s from cloud\n", cachefile->filename);
320 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
321 async->op = STORE_OP_GET;
322 async->key = g_strdup(cachefile->filename);
323 bluesky_store_async_add_notifier(async,
324 (GFunc)cloudlog_fetch_complete,
326 bluesky_store_async_submit(async);
327 bluesky_store_async_unref(async);
330 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
331 BlueSkyCacheFile *cachefile)
333 g_print("Fetch of %s from cloud complete, status = %d\n",
334 async->key, async->result);
336 g_mutex_lock(cachefile->lock);
337 if (async->result >= 0) {
338 char *pathname = g_strdup_printf("%s/%s",
339 cachefile->log->log_directory,
340 cachefile->filename);
341 if (!g_file_set_contents(pathname, async->data->data, async->data->len,
343 g_print("Error writing out fetched file to cache!\n");
346 cachefile->fetching = FALSE;
347 cachefile->ready = TRUE;
349 g_print("Error fetching from cloud, retrying...\n");
350 cloudlog_fetch_start(cachefile);
353 bluesky_cachefile_unref(cachefile);
354 g_cond_broadcast(cachefile->cond);
355 g_mutex_unlock(cachefile->lock);
358 /* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
359 * Returns the object in the locked state and with a reference taken. */
360 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
361 int clouddir, int log_seq)
363 if (page_size == 0) {
364 page_size = getpagesize();
367 BlueSkyLog *log = fs->log;
373 // A request for a local log file
375 sprintf(logname, "journal-%08d", log_seq);
376 type = CLOUDLOG_JOURNAL;
378 sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
379 type = CLOUDLOG_CLOUD;
382 BlueSkyCacheFile *map;
383 g_mutex_lock(log->mmap_lock);
384 map = g_hash_table_lookup(log->mmap_cache, logname);
387 && type == CLOUDLOG_JOURNAL
388 && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
389 /* A stale reference to a journal file which doesn't exist any longer
390 * because it was reclaimed. Return NULL. */
391 } else if (map == NULL) {
392 g_print("Adding cache file %s\n", logname);
394 map = g_new0(BlueSkyCacheFile, 1);
397 map->lock = g_mutex_new();
399 g_mutex_lock(map->lock);
400 map->cond = g_cond_new();
401 map->filename = g_strdup(logname);
402 map->log_seq = log_seq;
404 g_atomic_int_set(&map->mapcount, 0);
405 g_atomic_int_set(&map->refcount, 0);
407 g_hash_table_insert(log->mmap_cache, map->filename, map);
409 // If the log file is stored in the cloud, we may need to fetch it
411 cloudlog_fetch_start(map);
413 g_mutex_lock(map->lock);
416 g_mutex_unlock(log->mmap_lock);
418 g_atomic_int_inc(&map->refcount);
422 BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir,
423 int log_seq, int log_offset, int log_size)
425 if (page_size == 0) {
426 page_size = getpagesize();
429 BlueSkyLog *log = fs->log;
430 BlueSkyCacheFile *map = bluesky_cachefile_lookup(fs, log_dir, log_seq);
436 if (map->addr == NULL) {
437 while (!map->ready && map->fetching) {
438 g_print("Waiting for log segment to be fetched from cloud...\n");
439 g_cond_wait(map->cond, map->lock);
442 int fd = openat(log->dirfd, map->filename, O_RDONLY);
445 fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
446 bluesky_cachefile_unref(map);
447 g_mutex_unlock(map->lock);
451 off_t length = lseek(fd, 0, SEEK_END);
452 map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
454 g_atomic_int_add(&log->disk_used, -(map->len / 1024));
456 g_atomic_int_add(&log->disk_used, map->len / 1024);
458 g_print("Re-mapped log segment %d...\n", log_seq);
459 g_atomic_int_inc(&map->refcount);
464 g_mutex_unlock(log->mmap_lock);
467 map->atime = bluesky_get_current_time();
468 str = bluesky_string_new_from_mmap(map, log_offset, log_size);
469 bluesky_cachefile_unref(map);
470 g_mutex_unlock(map->lock);
474 void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
479 if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
480 g_mutex_lock(mmap->lock);
481 if (g_atomic_int_get(&mmap->mapcount) == 0) {
482 g_print("Unmapped log segment %d...\n", mmap->log_seq);
483 munmap((void *)mmap->addr, mmap->len);
485 g_atomic_int_add(&mmap->refcount, -1);
487 g_mutex_unlock(mmap->lock);
491 /* Scan through all currently-stored files in the journal/cache and garbage
492 * collect old unused ones, if needed. */
493 static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
495 GList **files = (GList **)user_data;
496 *files = g_list_prepend(*files, value);
499 static gint compare_cachefiles(gconstpointer a, gconstpointer b)
503 ta = ((BlueSkyCacheFile *)a)->atime;
504 tb = ((BlueSkyCacheFile *)b)->atime;
513 void bluesky_cachefile_gc(BlueSkyFS *fs)
517 g_mutex_lock(fs->log->mmap_lock);
518 g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
520 /* Sort based on atime. The atime should be stable since it shouln't be
521 * updated except by threads which can grab the mmap_lock, which we already
523 files = g_list_sort(files, compare_cachefiles);
525 /* Walk the list of files, starting with the oldest, deleting files if
526 * possible until enough space has been reclaimed. */
527 g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
528 while (files != NULL) {
529 BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
530 /* Try to lock the structure, but if the lock is held by another thread
531 * then we'll just skip the file on this pass. */
532 if (g_mutex_trylock(cachefile->lock)) {
533 int64_t age = bluesky_get_current_time() - cachefile->atime;
534 g_print("%s addr=%p mapcount=%d refcount=%d atime_age=%f",
535 cachefile->filename, cachefile->addr, cachefile->mapcount,
536 cachefile->refcount, age / 1e6);
537 if (cachefile->fetching)
538 g_print(" (fetching)");
541 gboolean deletion_candidate = FALSE;
542 if (g_atomic_int_get(&fs->log->disk_used)
543 > bluesky_options.cache_size
544 && g_atomic_int_get(&cachefile->refcount) == 0
545 && g_atomic_int_get(&cachefile->mapcount) == 0)
547 deletion_candidate = TRUE;
550 /* Don't allow journal files to be reclaimed until all data is
551 * known to be durably stored in the cloud. */
552 if (cachefile->type == CLOUDLOG_JOURNAL
553 && cachefile->log_seq >= fs->log->journal_watermark)
555 deletion_candidate = FALSE;
558 if (deletion_candidate) {
559 g_print(" ...deleting\n");
560 if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
561 fprintf(stderr, "Unable to unlink journal %s: %m\n",
562 cachefile->filename);
565 g_atomic_int_add(&fs->log->disk_used, -(cachefile->len / 1024));
566 g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
567 g_mutex_unlock(cachefile->lock);
568 g_mutex_free(cachefile->lock);
569 g_cond_free(cachefile->cond);
570 g_free(cachefile->filename);
573 g_mutex_unlock(cachefile->lock);
576 files = g_list_delete_link(files, files);
580 g_mutex_unlock(fs->log->mmap_lock);
583 /******************************* JOURNAL REPLAY *******************************
584 * The journal replay code is used to recover filesystem state after a
585 * filesystem restart. We first look for the most recent commit record in the
586 * journal, which indicates the point before which all data in the journal has
587 * also been committed to the cloud. Then, we read in all data in the log past
590 static GList *directory_contents(const char *dirname)
592 GList *contents = NULL;
593 GDir *dir = g_dir_open(dirname, 0, NULL);
595 g_warning("Unable to open journal directory: %s", dirname);
600 while ((file = g_dir_read_name(dir)) != NULL) {
601 if (strncmp(file, "journal-", 8) == 0)
602 contents = g_list_prepend(contents, g_strdup(file));
606 contents = g_list_sort(contents, (GCompareFunc)strcmp);
611 static gboolean validate_journal_item(const char *buf, size_t len, off_t offset)
613 const struct log_header *header;
614 const struct log_footer *footer;
616 if (offset + sizeof(struct log_header) + sizeof(struct log_footer) > len)
619 header = (const struct log_header *)(buf + offset);
620 if (GUINT32_FROM_LE(header->magic) != HEADER_MAGIC)
622 if (GUINT32_FROM_LE(header->offset) != offset)
624 size_t size = GUINT32_FROM_LE(header->size1)
625 + GUINT32_FROM_LE(header->size2)
626 + GUINT32_FROM_LE(header->size3);
628 off_t footer_offset = offset + sizeof(struct log_header) + size;
629 if (footer_offset + sizeof(struct log_footer) > len)
631 footer = (const struct log_footer *)(buf + footer_offset);
633 if (GUINT32_FROM_LE(footer->magic) != FOOTER_MAGIC)
636 uint32_t crc = crc32c(BLUESKY_CRC32C_SEED, buf + offset,
637 sizeof(struct log_header) + sizeof(struct log_footer)
639 if (crc != BLUESKY_CRC32C_VALIDATOR) {
640 g_warning("Journal entry failed to validate: CRC %08x != %08x",
641 crc, BLUESKY_CRC32C_VALIDATOR);
648 /* Scan through a journal segment to extract correctly-written items (those
649 * that pass sanity checks and have a valid checksum). */
650 static void bluesky_replay_scan_journal(const char *buf, size_t len)
652 const struct log_header *header;
655 while (validate_journal_item(buf, len, offset)) {
656 header = (const struct log_header *)(buf + offset);
657 size_t size = GUINT32_FROM_LE(header->size1)
658 + GUINT32_FROM_LE(header->size2)
659 + GUINT32_FROM_LE(header->size3);
660 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
664 static void reload_item(BlueSkyCloudLog *log_item,
666 size_t len1, size_t len2, size_t len3)
668 BlueSkyFS *fs = log_item->fs;
669 /*const char *data1 = data;*/
670 const BlueSkyCloudID *data2
671 = (const BlueSkyCloudID *)(data + len1);
672 /*const BlueSkyCloudPointer *data3
673 = (const BlueSkyCloudPointer *)(data + len1 + len2);*/
675 bluesky_string_unref(log_item->data);
676 log_item->data = NULL;
677 log_item->location_flags = CLOUDLOG_JOURNAL;
680 memset(&id0, 0, sizeof(id0));
682 int link_count = len2 / sizeof(BlueSkyCloudID);
683 GArray *new_links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
684 for (int i = 0; i < link_count; i++) {
685 BlueSkyCloudID id = data2[i];
686 BlueSkyCloudLog *ref = NULL;
687 if (memcmp(&id, &id0, sizeof(BlueSkyCloudID)) != 0) {
688 g_mutex_lock(fs->lock);
689 ref = g_hash_table_lookup(fs->locations, &id);
691 bluesky_cloudlog_ref(ref);
693 g_mutex_unlock(fs->lock);
695 g_array_append_val(new_links, ref);
698 for (int i = 0; i < log_item->links->len; i++) {
699 BlueSkyCloudLog *c = g_array_index(log_item->links,
700 BlueSkyCloudLog *, i);
701 bluesky_cloudlog_unref(c);
703 g_array_unref(log_item->links);
704 log_item->links = new_links;
707 static void bluesky_replay_scan_journal2(BlueSkyFS *fs, GList **objects,
709 const char *buf, size_t len)
711 const struct log_header *header;
714 while (validate_journal_item(buf, len, offset)) {
715 header = (const struct log_header *)(buf + offset);
716 g_print("In replay found valid item at offset %zd\n", offset);
717 size_t size = GUINT32_FROM_LE(header->size1)
718 + GUINT32_FROM_LE(header->size2)
719 + GUINT32_FROM_LE(header->size3);
721 g_mutex_lock(fs->lock);
722 BlueSkyCloudLog *log_item;
723 log_item = g_hash_table_lookup(fs->locations, &header->id);
724 if (log_item == NULL) {
725 log_item = bluesky_cloudlog_new(fs, &header->id);
726 g_hash_table_insert(fs->locations, &log_item->id, log_item);
727 g_mutex_lock(log_item->lock);
729 bluesky_cloudlog_ref(log_item);
730 g_mutex_lock(log_item->lock);
732 g_mutex_unlock(fs->lock);
733 *objects = g_list_prepend(*objects, log_item);
735 log_item->inum = GUINT64_FROM_LE(header->inum);
736 reload_item(log_item, buf + offset + sizeof(struct log_header),
737 GUINT32_FROM_LE(header->size1),
738 GUINT32_FROM_LE(header->size2),
739 GUINT32_FROM_LE(header->size3));
740 log_item->log_seq = log_seq;
741 log_item->log_offset = offset + sizeof(struct log_header);
742 log_item->log_size = header->size1;
744 bluesky_string_unref(log_item->data);
745 log_item->data = bluesky_string_new(g_memdup(buf + offset + sizeof(struct log_header), GUINT32_FROM_LE(header->size1)), GUINT32_FROM_LE(header->size1));
747 /* For any inodes which were read from the journal, deserialize the
748 * inode information, overwriting any old inode data. */
749 if (header->type - '0' == LOGTYPE_INODE) {
750 uint64_t inum = GUINT64_FROM_LE(header->inum);
752 g_mutex_lock(fs->lock);
753 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
755 inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
756 inode->change_count = 0;
757 bluesky_insert_inode(fs, inode);
759 g_mutex_lock(inode->lock);
760 bluesky_inode_free_resources(inode);
761 if (!bluesky_deserialize_inode(inode, log_item))
762 g_print("Error deserializing inode %"PRIu64"\n", inum);
763 fs->next_inum = MAX(fs->next_inum, inum + 1);
764 g_mutex_unlock(inode->lock);
765 g_mutex_unlock(fs->lock);
767 bluesky_string_unref(log_item->data);
768 log_item->data = NULL;
769 g_mutex_unlock(log_item->lock);
771 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
775 void bluesky_replay(BlueSkyFS *fs)
777 BlueSkyLog *log = fs->log;
778 GList *logfiles = directory_contents(log->log_directory);
780 /* Scan through log files in reverse order to find the most recent commit
782 logfiles = g_list_reverse(logfiles);
783 while (logfiles != NULL) {
784 char *filename = g_strdup_printf("%s/%s", log->log_directory,
785 (char *)logfiles->data);
786 g_print("Scanning file %s\n", filename);
787 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
789 g_warning("Mapping logfile %s failed!\n", filename);
791 bluesky_replay_scan_journal(g_mapped_file_get_contents(map),
792 g_mapped_file_get_length(map));
793 g_mapped_file_unref(map);
797 g_free(logfiles->data);
798 logfiles = g_list_delete_link(logfiles, logfiles);
800 g_list_foreach(logfiles, (GFunc)g_free, NULL);
801 g_list_free(logfiles);
803 /* Now, scan forward starting from the given point in the log to
804 * reconstruct all filesystem state. As we reload objects we hold a
805 * reference to each loaded object. At the end we free all these
806 * references, so that any objects which were not linked into persistent
807 * filesystem data structures are freed. */
808 GList *objects = NULL;
811 char *filename = g_strdup_printf("%s/journal-%08d",
812 log->log_directory, seq_num);
813 g_print("Replaying file %s\n", filename);
814 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
817 g_warning("Mapping logfile failed, assuming end of journal\n");
821 bluesky_replay_scan_journal2(fs, &objects, seq_num,
822 g_mapped_file_get_contents(map),
823 g_mapped_file_get_length(map));
824 g_mapped_file_unref(map);
828 while (objects != NULL) {
829 bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
830 objects = g_list_delete_link(objects, objects);