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
44 static void writebuf(int fd, const char *buf, size_t len)
48 written = write(fd, buf, len);
49 if (written < 0 && errno == EINTR)
51 g_assert(written >= 0);
57 static void log_commit(BlueSkyLog *log)
65 while (log->committed != NULL) {
66 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
67 g_mutex_lock(item->lock);
68 bluesky_cloudlog_stats_update(item, -1);
69 item->pending_write &= ~CLOUDLOG_JOURNAL;
70 item->location_flags |= CLOUDLOG_JOURNAL;
71 bluesky_cloudlog_stats_update(item, 1);
72 g_cond_signal(item->cond);
73 g_mutex_unlock(item->lock);
74 log->committed = g_slist_delete_link(log->committed, log->committed);
75 bluesky_cloudlog_unref(item);
79 if (bluesky_verbose && batchsize > 1)
80 g_print("Log batch size: %d\n", batchsize);
83 static gboolean log_open(BlueSkyLog *log)
94 if (log->current_log != NULL) {
95 bluesky_cachefile_unref(log->current_log);
96 log->current_log = NULL;
100 g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
101 log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
102 if (log->fd < 0 && errno == EEXIST) {
103 fprintf(stderr, "Log file %s already exists...\n", logname);
106 } else if (log->fd < 0) {
107 fprintf(stderr, "Error opening logfile %s: %m\n", logname);
112 log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
113 g_assert(log->current_log != NULL);
114 g_mutex_unlock(log->current_log->lock);
116 if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
117 fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
124 /* All log writes (at least for a single log) are made by one thread, so we
125 * don't need to worry about concurrent access to the log file. Log items to
126 * write are pulled off a queue (and so may be posted by any thread).
127 * fdatasync() is used to ensure the log items are stable on disk.
129 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
130 * each. If a log segment is not currently open (log->fd is negative), a new
131 * one is created. Log segment filenames are assigned sequentially.
133 * Log replay ought to be implemented later, and ought to set the initial
134 * sequence number appropriately.
136 static gpointer log_thread(gpointer d)
138 BlueSkyLog *log = (BlueSkyLog *)d;
142 if (!log_open(log)) {
147 BlueSkyCloudLog *item
148 = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
149 g_mutex_lock(item->lock);
150 g_assert(item->data != NULL);
152 /* The item may have already been written to the journal... */
153 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
154 g_mutex_unlock(item->lock);
155 bluesky_cloudlog_unref(item);
156 g_atomic_int_add(&item->data_lock_count, -1);
160 bluesky_cloudlog_stats_update(item, -1);
161 item->pending_write |= CLOUDLOG_JOURNAL;
162 bluesky_cloudlog_stats_update(item, 1);
164 GString *data1 = g_string_new("");
165 GString *data2 = g_string_new("");
166 GString *data3 = g_string_new("");
167 bluesky_serialize_cloudlog(item, data1, data2, data3);
169 struct log_header header;
170 struct log_footer footer;
171 size_t size = sizeof(header) + sizeof(footer);
172 size += data1->len + data2->len + data3->len;
175 offset = lseek(log->fd, 0, SEEK_CUR);
177 /* Check whether the item would overflow the allocated journal size.
178 * If so, start a new log segment. We only allow oversized log
179 * segments if they contain a single log entry. */
180 if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
185 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
186 header.offset = GUINT32_TO_LE(offset);
187 header.size1 = GUINT32_TO_LE(data1->len);
188 header.size2 = GUINT32_TO_LE(data2->len);
189 header.size3 = GUINT32_TO_LE(data3->len);
190 header.type = item->type + '0';
191 header.id = item->id;
192 header.inum = GUINT64_TO_LE(item->inum);
193 footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
195 uint32_t crc = BLUESKY_CRC32C_SEED;
197 writebuf(log->fd, (const char *)&header, sizeof(header));
198 crc = crc32c(crc, (const char *)&header, sizeof(header));
200 writebuf(log->fd, data1->str, data1->len);
201 crc = crc32c(crc, data1->str, data1->len);
202 writebuf(log->fd, data2->str, data2->len);
203 crc = crc32c(crc, data2->str, data2->len);
204 writebuf(log->fd, data3->str, data3->len);
205 crc = crc32c(crc, data3->str, data3->len);
207 crc = crc32c(crc, (const char *)&footer,
208 sizeof(footer) - sizeof(uint32_t));
209 footer.crc = crc32c_finalize(crc);
210 writebuf(log->fd, (const char *)&footer, sizeof(footer));
212 item->log_seq = log->seq_num;
213 item->log_offset = offset;
214 item->log_size = size;
215 item->data_size = item->data->len;
219 g_string_free(data1, TRUE);
220 g_string_free(data2, TRUE);
221 g_string_free(data3, TRUE);
223 /* Replace the log item's string data with a memory-mapped copy of the
224 * data, now that it has been written to the log file. (Even if it
225 * isn't yet on disk, it should at least be in the page cache and so
226 * available to memory map.) */
227 bluesky_string_unref(item->data);
229 bluesky_cloudlog_fetch(item);
231 log->committed = g_slist_prepend(log->committed, item);
232 g_atomic_int_add(&item->data_lock_count, -1);
233 g_mutex_unlock(item->lock);
235 /* Force an if there are no other log items currently waiting to be
237 if (g_async_queue_length(log->queue) <= 0)
244 BlueSkyLog *bluesky_log_new(const char *log_directory)
246 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
248 log->log_directory = g_strdup(log_directory);
251 log->queue = g_async_queue_new();
252 log->mmap_lock = g_mutex_new();
253 log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
255 /* Determine the highest-numbered log file, so that we can start writing
256 * out new journal entries at the next sequence number. */
257 GDir *dir = g_dir_open(log_directory, 0, NULL);
260 while ((file = g_dir_read_name(dir)) != NULL) {
261 if (strncmp(file, "journal-", 8) == 0) {
262 log->seq_num = MAX(log->seq_num, atoi(&file[8]) + 1);
266 g_print("Starting journal at sequence number %d\n", log->seq_num);
269 log->dirfd = open(log->log_directory, O_DIRECTORY);
270 if (log->dirfd < 0) {
271 fprintf(stderr, "Unable to open logging directory: %m\n");
275 g_thread_create(log_thread, log, FALSE, NULL);
280 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
282 bluesky_cloudlog_ref(item);
283 g_atomic_int_add(&item->data_lock_count, 1);
284 g_async_queue_push(log->queue, item);
287 void bluesky_log_finish_all(GList *log_items)
289 while (log_items != NULL) {
290 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
292 g_mutex_lock(item->lock);
293 while ((item->pending_write & CLOUDLOG_JOURNAL))
294 g_cond_wait(item->cond, item->lock);
295 g_mutex_unlock(item->lock);
296 bluesky_cloudlog_unref(item);
298 log_items = g_list_delete_link(log_items, log_items);
302 /* Return a committed cloud log record that can be used as a watermark for how
303 * much of the journal has been written. */
304 BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs)
306 BlueSkyCloudLog *marker = bluesky_cloudlog_new(fs, NULL);
307 marker->type = LOGTYPE_JOURNAL_MARKER;
308 marker->data = bluesky_string_new(g_strdup(""), 0);
309 bluesky_cloudlog_sync(marker);
311 g_mutex_lock(marker->lock);
312 while ((marker->pending_write & CLOUDLOG_JOURNAL))
313 g_cond_wait(marker->cond, marker->lock);
314 g_mutex_unlock(marker->lock);
319 void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker)
321 BlueSkyCloudLog *commit = bluesky_cloudlog_new(fs, NULL);
322 commit->type = LOGTYPE_JOURNAL_CHECKPOINT;
324 uint32_t seq, offset;
325 seq = GUINT32_TO_LE(marker->log_seq);
326 offset = GUINT32_TO_LE(marker->log_offset);
327 GString *loc = g_string_new("");
328 g_string_append_len(loc, (const gchar *)&seq, sizeof(seq));
329 g_string_append_len(loc, (const gchar *)&offset, sizeof(offset));
330 commit->data = bluesky_string_new_from_gstring(loc);
331 bluesky_cloudlog_sync(commit);
333 g_mutex_lock(commit->lock);
334 while ((commit->pending_write & CLOUDLOG_JOURNAL))
335 g_cond_wait(commit->cond, commit->lock);
336 g_mutex_unlock(commit->lock);
338 bluesky_cloudlog_unref(marker);
339 bluesky_cloudlog_unref(commit);
342 /* Memory-map the given log object into memory (read-only) and return a pointer
344 static int page_size = 0;
346 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
348 g_atomic_int_add(&cachefile->refcount, -1);
351 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
352 BlueSkyCacheFile *cachefile);
354 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
356 g_atomic_int_inc(&cachefile->refcount);
357 cachefile->fetching = TRUE;
358 g_print("Starting fetch of %s from cloud\n", cachefile->filename);
359 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
360 async->op = STORE_OP_GET;
361 async->key = g_strdup(cachefile->filename);
362 bluesky_store_async_add_notifier(async,
363 (GFunc)cloudlog_fetch_complete,
365 bluesky_store_async_submit(async);
366 bluesky_store_async_unref(async);
369 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
370 BlueSkyCacheFile *cachefile)
372 g_print("Fetch of %s from cloud complete, status = %d\n",
373 async->key, async->result);
375 g_mutex_lock(cachefile->lock);
376 if (async->result >= 0) {
377 char *pathname = g_strdup_printf("%s/%s",
378 cachefile->log->log_directory,
379 cachefile->filename);
380 if (!g_file_set_contents(pathname, async->data->data, async->data->len,
382 g_print("Error writing out fetched file to cache!\n");
385 cachefile->fetching = FALSE;
386 cachefile->ready = TRUE;
388 g_print("Error fetching from cloud, retrying...\n");
389 cloudlog_fetch_start(cachefile);
392 bluesky_cachefile_unref(cachefile);
393 g_cond_broadcast(cachefile->cond);
394 g_mutex_unlock(cachefile->lock);
397 /* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
398 * Returns the object in the locked state and with a reference taken. */
399 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
400 int clouddir, int log_seq)
402 if (page_size == 0) {
403 page_size = getpagesize();
406 BlueSkyLog *log = fs->log;
412 // A request for a local log file
414 sprintf(logname, "journal-%08d", log_seq);
415 type = CLOUDLOG_JOURNAL;
417 sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
418 type = CLOUDLOG_CLOUD;
421 BlueSkyCacheFile *map;
422 g_mutex_lock(log->mmap_lock);
423 map = g_hash_table_lookup(log->mmap_cache, logname);
426 && type == CLOUDLOG_JOURNAL
427 && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
428 /* A stale reference to a journal file which doesn't exist any longer
429 * because it was reclaimed. Return NULL. */
430 } else if (map == NULL) {
431 g_print("Adding cache file %s\n", logname);
433 map = g_new0(BlueSkyCacheFile, 1);
436 map->lock = g_mutex_new();
438 g_mutex_lock(map->lock);
439 map->cond = g_cond_new();
440 map->filename = g_strdup(logname);
441 map->log_seq = log_seq;
443 g_atomic_int_set(&map->mapcount, 0);
444 g_atomic_int_set(&map->refcount, 0);
446 g_hash_table_insert(log->mmap_cache, map->filename, map);
448 // If the log file is stored in the cloud, we may need to fetch it
450 cloudlog_fetch_start(map);
452 g_mutex_lock(map->lock);
455 g_mutex_unlock(log->mmap_lock);
457 g_atomic_int_inc(&map->refcount);
461 BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir,
462 int log_seq, int log_offset, int log_size)
464 if (page_size == 0) {
465 page_size = getpagesize();
468 BlueSkyLog *log = fs->log;
469 BlueSkyCacheFile *map = bluesky_cachefile_lookup(fs, log_dir, log_seq);
475 if (map->addr == NULL) {
476 while (!map->ready && map->fetching) {
477 g_print("Waiting for log segment to be fetched from cloud...\n");
478 g_cond_wait(map->cond, map->lock);
481 int fd = openat(log->dirfd, map->filename, O_RDONLY);
484 fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
485 bluesky_cachefile_unref(map);
486 g_mutex_unlock(map->lock);
490 off_t length = lseek(fd, 0, SEEK_END);
491 map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
493 g_atomic_int_add(&log->disk_used, -(map->len / 1024));
495 g_atomic_int_add(&log->disk_used, map->len / 1024);
497 g_print("Re-mapped log segment %d...\n", log_seq);
498 g_atomic_int_inc(&map->refcount);
504 map->atime = bluesky_get_current_time();
505 str = bluesky_string_new_from_mmap(map, log_offset, log_size);
506 bluesky_cachefile_unref(map);
507 g_mutex_unlock(map->lock);
511 void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
516 if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
517 g_mutex_lock(mmap->lock);
518 if (g_atomic_int_get(&mmap->mapcount) == 0) {
519 g_print("Unmapped log segment %d...\n", mmap->log_seq);
520 munmap((void *)mmap->addr, mmap->len);
522 g_atomic_int_add(&mmap->refcount, -1);
524 g_mutex_unlock(mmap->lock);
528 /* Scan through all currently-stored files in the journal/cache and garbage
529 * collect old unused ones, if needed. */
530 static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
532 GList **files = (GList **)user_data;
533 *files = g_list_prepend(*files, value);
536 static gint compare_cachefiles(gconstpointer a, gconstpointer b)
540 ta = ((BlueSkyCacheFile *)a)->atime;
541 tb = ((BlueSkyCacheFile *)b)->atime;
550 void bluesky_cachefile_gc(BlueSkyFS *fs)
554 g_mutex_lock(fs->log->mmap_lock);
555 g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
557 /* Sort based on atime. The atime should be stable since it shouln't be
558 * updated except by threads which can grab the mmap_lock, which we already
560 files = g_list_sort(files, compare_cachefiles);
562 /* Walk the list of files, starting with the oldest, deleting files if
563 * possible until enough space has been reclaimed. */
564 g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
565 while (files != NULL) {
566 BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
567 /* Try to lock the structure, but if the lock is held by another thread
568 * then we'll just skip the file on this pass. */
569 if (g_mutex_trylock(cachefile->lock)) {
570 int64_t age = bluesky_get_current_time() - cachefile->atime;
571 g_print("%s addr=%p mapcount=%d refcount=%d atime_age=%f",
572 cachefile->filename, cachefile->addr, cachefile->mapcount,
573 cachefile->refcount, age / 1e6);
574 if (cachefile->fetching)
575 g_print(" (fetching)");
578 gboolean deletion_candidate = FALSE;
579 if (g_atomic_int_get(&fs->log->disk_used)
580 > bluesky_options.cache_size
581 && g_atomic_int_get(&cachefile->refcount) == 0
582 && g_atomic_int_get(&cachefile->mapcount) == 0)
584 deletion_candidate = TRUE;
587 /* Don't allow journal files to be reclaimed until all data is
588 * known to be durably stored in the cloud. */
589 if (cachefile->type == CLOUDLOG_JOURNAL
590 && cachefile->log_seq >= fs->log->journal_watermark)
592 deletion_candidate = FALSE;
595 if (deletion_candidate) {
596 g_print(" ...deleting\n");
597 if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
598 fprintf(stderr, "Unable to unlink journal %s: %m\n",
599 cachefile->filename);
602 g_atomic_int_add(&fs->log->disk_used, -(cachefile->len / 1024));
603 g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
604 g_mutex_unlock(cachefile->lock);
605 g_mutex_free(cachefile->lock);
606 g_cond_free(cachefile->cond);
607 g_free(cachefile->filename);
610 g_mutex_unlock(cachefile->lock);
613 files = g_list_delete_link(files, files);
617 g_mutex_unlock(fs->log->mmap_lock);
620 /******************************* JOURNAL REPLAY *******************************
621 * The journal replay code is used to recover filesystem state after a
622 * filesystem restart. We first look for the most recent commit record in the
623 * journal, which indicates the point before which all data in the journal has
624 * also been committed to the cloud. Then, we read in all data in the log past
627 static GList *directory_contents(const char *dirname)
629 GList *contents = NULL;
630 GDir *dir = g_dir_open(dirname, 0, NULL);
632 g_warning("Unable to open journal directory: %s", dirname);
637 while ((file = g_dir_read_name(dir)) != NULL) {
638 if (strncmp(file, "journal-", 8) == 0)
639 contents = g_list_prepend(contents, g_strdup(file));
643 contents = g_list_sort(contents, (GCompareFunc)strcmp);
648 static gboolean validate_journal_item(const char *buf, size_t len, off_t offset)
650 const struct log_header *header;
651 const struct log_footer *footer;
653 if (offset + sizeof(struct log_header) + sizeof(struct log_footer) > len)
656 header = (const struct log_header *)(buf + offset);
657 if (GUINT32_FROM_LE(header->magic) != HEADER_MAGIC)
659 if (GUINT32_FROM_LE(header->offset) != offset)
661 size_t size = GUINT32_FROM_LE(header->size1)
662 + GUINT32_FROM_LE(header->size2)
663 + GUINT32_FROM_LE(header->size3);
665 off_t footer_offset = offset + sizeof(struct log_header) + size;
666 if (footer_offset + sizeof(struct log_footer) > len)
668 footer = (const struct log_footer *)(buf + footer_offset);
670 if (GUINT32_FROM_LE(footer->magic) != FOOTER_MAGIC)
673 uint32_t crc = crc32c(BLUESKY_CRC32C_SEED, buf + offset,
674 sizeof(struct log_header) + sizeof(struct log_footer)
676 if (crc != BLUESKY_CRC32C_VALIDATOR) {
677 g_warning("Journal entry failed to validate: CRC %08x != %08x",
678 crc, BLUESKY_CRC32C_VALIDATOR);
685 /* Scan through a journal segment to extract correctly-written items (those
686 * that pass sanity checks and have a valid checksum). */
687 static void bluesky_replay_scan_journal(const char *buf, size_t len,
688 uint32_t *seq, uint32_t *start_offset)
690 const struct log_header *header;
693 while (validate_journal_item(buf, len, offset)) {
694 header = (const struct log_header *)(buf + offset);
695 size_t size = GUINT32_FROM_LE(header->size1)
696 + GUINT32_FROM_LE(header->size2)
697 + GUINT32_FROM_LE(header->size3);
699 if (header->type - '0' == LOGTYPE_JOURNAL_CHECKPOINT) {
700 const uint32_t *data = (const uint32_t *)((const char *)header + sizeof(struct log_header));
701 *seq = GUINT32_FROM_LE(data[0]);
702 *start_offset = GUINT32_FROM_LE(data[1]);
705 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
709 static void reload_item(BlueSkyCloudLog *log_item,
711 size_t len1, size_t len2, size_t len3)
713 BlueSkyFS *fs = log_item->fs;
714 /*const char *data1 = data;*/
715 const BlueSkyCloudID *data2
716 = (const BlueSkyCloudID *)(data + len1);
717 /*const BlueSkyCloudPointer *data3
718 = (const BlueSkyCloudPointer *)(data + len1 + len2);*/
720 bluesky_string_unref(log_item->data);
721 log_item->data = NULL;
722 log_item->location_flags = CLOUDLOG_JOURNAL;
725 memset(&id0, 0, sizeof(id0));
727 int link_count = len2 / sizeof(BlueSkyCloudID);
728 GArray *new_links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
729 for (int i = 0; i < link_count; i++) {
730 BlueSkyCloudID id = data2[i];
731 BlueSkyCloudLog *ref = NULL;
732 if (memcmp(&id, &id0, sizeof(BlueSkyCloudID)) != 0) {
733 g_mutex_lock(fs->lock);
734 ref = g_hash_table_lookup(fs->locations, &id);
736 bluesky_cloudlog_ref(ref);
738 g_mutex_unlock(fs->lock);
740 g_array_append_val(new_links, ref);
743 for (int i = 0; i < log_item->links->len; i++) {
744 BlueSkyCloudLog *c = g_array_index(log_item->links,
745 BlueSkyCloudLog *, i);
746 bluesky_cloudlog_unref(c);
748 g_array_unref(log_item->links);
749 log_item->links = new_links;
752 static void bluesky_replay_scan_journal2(BlueSkyFS *fs, GList **objects,
753 int log_seq, int start_offset,
754 const char *buf, size_t len)
756 const struct log_header *header;
757 off_t offset = start_offset;
759 while (validate_journal_item(buf, len, offset)) {
760 header = (const struct log_header *)(buf + offset);
761 g_print("In replay found valid item at offset %zd\n", offset);
762 size_t size = GUINT32_FROM_LE(header->size1)
763 + GUINT32_FROM_LE(header->size2)
764 + GUINT32_FROM_LE(header->size3);
766 BlueSkyCloudLog *log_item = bluesky_cloudlog_get(fs, header->id);
767 g_mutex_lock(log_item->lock);
768 *objects = g_list_prepend(*objects, log_item);
770 log_item->inum = GUINT64_FROM_LE(header->inum);
771 reload_item(log_item, buf + offset + sizeof(struct log_header),
772 GUINT32_FROM_LE(header->size1),
773 GUINT32_FROM_LE(header->size2),
774 GUINT32_FROM_LE(header->size3));
775 log_item->log_seq = log_seq;
776 log_item->log_offset = offset + sizeof(struct log_header);
777 log_item->log_size = header->size1;
779 bluesky_string_unref(log_item->data);
780 log_item->data = bluesky_string_new(g_memdup(buf + offset + sizeof(struct log_header), GUINT32_FROM_LE(header->size1)), GUINT32_FROM_LE(header->size1));
782 /* For any inodes which were read from the journal, deserialize the
783 * inode information, overwriting any old inode data. */
784 if (header->type - '0' == LOGTYPE_INODE) {
785 uint64_t inum = GUINT64_FROM_LE(header->inum);
787 g_mutex_lock(fs->lock);
788 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
790 inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
791 inode->change_count = 0;
792 bluesky_insert_inode(fs, inode);
794 g_mutex_lock(inode->lock);
795 bluesky_inode_free_resources(inode);
796 if (!bluesky_deserialize_inode(inode, log_item))
797 g_print("Error deserializing inode %"PRIu64"\n", inum);
798 fs->next_inum = MAX(fs->next_inum, inum + 1);
799 bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
800 inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
801 bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
802 inode->dirty_list = bluesky_list_prepend(&fs->dirty_list, inode);
803 bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
804 inode->unlogged_list = NULL;
805 inode->change_cloud = inode->change_commit;
806 bluesky_cloudlog_ref(log_item);
807 bluesky_cloudlog_unref(inode->committed_item);
808 inode->committed_item = log_item;
809 g_mutex_unlock(inode->lock);
810 g_mutex_unlock(fs->lock);
812 bluesky_string_unref(log_item->data);
813 log_item->data = NULL;
814 g_mutex_unlock(log_item->lock);
816 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
820 void bluesky_replay(BlueSkyFS *fs)
822 BlueSkyLog *log = fs->log;
823 GList *logfiles = directory_contents(log->log_directory);
825 /* Scan through log files in reverse order to find the most recent commit
827 logfiles = g_list_reverse(logfiles);
828 uint32_t seq_num = 0, start_offset = 0;
829 while (logfiles != NULL) {
830 char *filename = g_strdup_printf("%s/%s", log->log_directory,
831 (char *)logfiles->data);
832 g_print("Scanning file %s\n", filename);
833 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
835 g_warning("Mapping logfile %s failed!\n", filename);
837 bluesky_replay_scan_journal(g_mapped_file_get_contents(map),
838 g_mapped_file_get_length(map),
839 &seq_num, &start_offset);
840 g_mapped_file_unref(map);
844 g_free(logfiles->data);
845 logfiles = g_list_delete_link(logfiles, logfiles);
846 if (seq_num != 0 || start_offset != 0)
849 g_list_foreach(logfiles, (GFunc)g_free, NULL);
850 g_list_free(logfiles);
852 /* Now, scan forward starting from the given point in the log to
853 * reconstruct all filesystem state. As we reload objects we hold a
854 * reference to each loaded object. At the end we free all these
855 * references, so that any objects which were not linked into persistent
856 * filesystem data structures are freed. */
857 GList *objects = NULL;
859 char *filename = g_strdup_printf("%s/journal-%08d",
860 log->log_directory, seq_num);
861 g_print("Replaying file %s from offset %d\n", filename, start_offset);
862 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
865 g_warning("Mapping logfile failed, assuming end of journal\n");
869 bluesky_replay_scan_journal2(fs, &objects, seq_num, start_offset,
870 g_mapped_file_get_contents(map),
871 g_mapped_file_get_length(map));
872 g_mapped_file_unref(map);
877 while (objects != NULL) {
878 bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
879 objects = g_list_delete_link(objects, objects);