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;
71 = (item->location_flags & ~CLOUDLOG_UNCOMMITTED) | CLOUDLOG_JOURNAL;
72 bluesky_cloudlog_stats_update(item, 1);
73 g_cond_signal(item->cond);
74 g_mutex_unlock(item->lock);
75 log->committed = g_slist_delete_link(log->committed, log->committed);
76 bluesky_cloudlog_unref(item);
80 if (bluesky_verbose && batchsize > 1)
81 g_print("Log batch size: %d\n", batchsize);
84 static gboolean log_open(BlueSkyLog *log)
95 if (log->current_log != NULL) {
96 bluesky_cachefile_unref(log->current_log);
97 log->current_log = NULL;
100 while (log->fd < 0) {
101 g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
102 log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
103 if (log->fd < 0 && errno == EEXIST) {
104 fprintf(stderr, "Log file %s already exists...\n", logname);
107 } else if (log->fd < 0) {
108 fprintf(stderr, "Error opening logfile %s: %m\n", logname);
113 log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num,
115 g_assert(log->current_log != NULL);
116 g_mutex_unlock(log->current_log->lock);
118 if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
119 fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
126 /* All log writes (at least for a single log) are made by one thread, so we
127 * don't need to worry about concurrent access to the log file. Log items to
128 * write are pulled off a queue (and so may be posted by any thread).
129 * fdatasync() is used to ensure the log items are stable on disk.
131 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
132 * each. If a log segment is not currently open (log->fd is negative), a new
133 * one is created. Log segment filenames are assigned sequentially.
135 * Log replay ought to be implemented later, and ought to set the initial
136 * sequence number appropriately.
138 static gpointer log_thread(gpointer d)
140 BlueSkyLog *log = (BlueSkyLog *)d;
144 if (!log_open(log)) {
149 BlueSkyCloudLog *item
150 = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
151 g_mutex_lock(item->lock);
152 g_assert(item->data != NULL);
154 /* The item may have already been written to the journal... */
155 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
156 g_mutex_unlock(item->lock);
157 bluesky_cloudlog_unref(item);
158 g_atomic_int_add(&item->data_lock_count, -1);
162 bluesky_cloudlog_stats_update(item, -1);
163 item->pending_write |= CLOUDLOG_JOURNAL;
164 bluesky_cloudlog_stats_update(item, 1);
166 GString *data1 = g_string_new("");
167 GString *data2 = g_string_new("");
168 GString *data3 = g_string_new("");
169 bluesky_serialize_cloudlog(item, data1, data2, data3);
171 struct log_header header;
172 struct log_footer footer;
173 size_t size = sizeof(header) + sizeof(footer);
174 size += data1->len + data2->len + data3->len;
177 offset = lseek(log->fd, 0, SEEK_CUR);
179 /* Check whether the item would overflow the allocated journal size.
180 * If so, start a new log segment. We only allow oversized log
181 * segments if they contain a single log entry. */
182 if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
187 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
188 header.offset = GUINT32_TO_LE(offset);
189 header.size1 = GUINT32_TO_LE(data1->len);
190 header.size2 = GUINT32_TO_LE(data2->len);
191 header.size3 = GUINT32_TO_LE(data3->len);
192 header.type = item->type + '0';
193 header.id = item->id;
194 header.inum = GUINT64_TO_LE(item->inum);
195 footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
197 uint32_t crc = BLUESKY_CRC32C_SEED;
199 writebuf(log->fd, (const char *)&header, sizeof(header));
200 crc = crc32c(crc, (const char *)&header, sizeof(header));
202 writebuf(log->fd, data1->str, data1->len);
203 crc = crc32c(crc, data1->str, data1->len);
204 writebuf(log->fd, data2->str, data2->len);
205 crc = crc32c(crc, data2->str, data2->len);
206 writebuf(log->fd, data3->str, data3->len);
207 crc = crc32c(crc, data3->str, data3->len);
209 crc = crc32c(crc, (const char *)&footer,
210 sizeof(footer) - sizeof(uint32_t));
211 footer.crc = crc32c_finalize(crc);
212 writebuf(log->fd, (const char *)&footer, sizeof(footer));
214 item->log_seq = log->seq_num;
215 item->log_offset = offset;
216 item->log_size = size;
217 item->data_size = item->data->len;
221 g_string_free(data1, TRUE);
222 g_string_free(data2, TRUE);
223 g_string_free(data3, TRUE);
225 /* Replace the log item's string data with a memory-mapped copy of the
226 * data, now that it has been written to the log file. (Even if it
227 * isn't yet on disk, it should at least be in the page cache and so
228 * available to memory map.) */
229 bluesky_string_unref(item->data);
231 bluesky_cloudlog_fetch(item);
233 log->committed = g_slist_prepend(log->committed, item);
234 g_atomic_int_add(&item->data_lock_count, -1);
235 g_mutex_unlock(item->lock);
237 /* Force an if there are no other log items currently waiting to be
239 if (g_async_queue_length(log->queue) <= 0)
246 BlueSkyLog *bluesky_log_new(const char *log_directory)
248 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
250 log->log_directory = g_strdup(log_directory);
253 log->queue = g_async_queue_new();
254 log->mmap_lock = g_mutex_new();
255 log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
257 /* Determine the highest-numbered log file, so that we can start writing
258 * out new journal entries at the next sequence number. */
259 GDir *dir = g_dir_open(log_directory, 0, NULL);
262 while ((file = g_dir_read_name(dir)) != NULL) {
263 if (strncmp(file, "journal-", 8) == 0) {
264 log->seq_num = MAX(log->seq_num, atoi(&file[8]) + 1);
268 g_print("Starting journal at sequence number %d\n", log->seq_num);
271 log->dirfd = open(log->log_directory, O_DIRECTORY);
272 if (log->dirfd < 0) {
273 fprintf(stderr, "Unable to open logging directory: %m\n");
277 g_thread_create(log_thread, log, FALSE, NULL);
282 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
284 if (!(item->location_flags & CLOUDLOG_JOURNAL)) {
285 bluesky_cloudlog_ref(item);
286 item->location_flags |= CLOUDLOG_UNCOMMITTED;
287 g_atomic_int_add(&item->data_lock_count, 1);
288 g_async_queue_push(log->queue, item);
292 void bluesky_log_finish_all(GList *log_items)
294 while (log_items != NULL) {
295 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
297 g_mutex_lock(item->lock);
298 while ((item->location_flags & CLOUDLOG_UNCOMMITTED))
299 g_cond_wait(item->cond, item->lock);
300 g_mutex_unlock(item->lock);
301 bluesky_cloudlog_unref(item);
303 log_items = g_list_delete_link(log_items, log_items);
307 /* Return a committed cloud log record that can be used as a watermark for how
308 * much of the journal has been written. */
309 BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs)
311 BlueSkyCloudLog *marker = bluesky_cloudlog_new(fs, NULL);
312 marker->type = LOGTYPE_JOURNAL_MARKER;
313 marker->data = bluesky_string_new(g_strdup(""), 0);
314 bluesky_cloudlog_stats_update(marker, 1);
315 bluesky_cloudlog_sync(marker);
317 g_mutex_lock(marker->lock);
318 while ((marker->pending_write & CLOUDLOG_JOURNAL))
319 g_cond_wait(marker->cond, marker->lock);
320 g_mutex_unlock(marker->lock);
325 void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker)
327 BlueSkyCloudLog *commit = bluesky_cloudlog_new(fs, NULL);
328 commit->type = LOGTYPE_JOURNAL_CHECKPOINT;
330 uint32_t seq, offset;
331 seq = GUINT32_TO_LE(marker->log_seq);
332 offset = GUINT32_TO_LE(marker->log_offset);
333 GString *loc = g_string_new("");
334 g_string_append_len(loc, (const gchar *)&seq, sizeof(seq));
335 g_string_append_len(loc, (const gchar *)&offset, sizeof(offset));
336 commit->data = bluesky_string_new_from_gstring(loc);
337 bluesky_cloudlog_stats_update(commit, 1);
338 bluesky_cloudlog_sync(commit);
340 g_mutex_lock(commit->lock);
341 while ((commit->location_flags & CLOUDLOG_UNCOMMITTED))
342 g_cond_wait(commit->cond, commit->lock);
343 g_mutex_unlock(commit->lock);
345 bluesky_cloudlog_unref(marker);
346 bluesky_cloudlog_unref(commit);
349 /* Memory-map the given log object into memory (read-only) and return a pointer
351 static int page_size = 0;
353 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
355 g_atomic_int_add(&cachefile->refcount, -1);
358 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile);
360 /* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
361 * Returns the object in the locked state and with a reference taken. */
362 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
363 int clouddir, int log_seq,
364 gboolean start_fetch)
366 if (page_size == 0) {
367 page_size = getpagesize();
370 BlueSkyLog *log = fs->log;
376 // A request for a local log file
378 sprintf(logname, "journal-%08d", log_seq);
379 type = CLOUDLOG_JOURNAL;
381 sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
382 type = CLOUDLOG_CLOUD;
385 BlueSkyCacheFile *map;
386 g_mutex_lock(log->mmap_lock);
387 map = g_hash_table_lookup(log->mmap_cache, logname);
390 && type == CLOUDLOG_JOURNAL
391 && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
392 /* A stale reference to a journal file which doesn't exist any longer
393 * because it was reclaimed. Return NULL. */
394 } else if (map == NULL) {
395 g_print("Adding cache file %s\n", logname);
397 map = g_new0(BlueSkyCacheFile, 1);
400 map->lock = g_mutex_new();
402 g_mutex_lock(map->lock);
403 map->cond = g_cond_new();
404 map->filename = g_strdup(logname);
405 map->log_seq = log_seq;
407 g_atomic_int_set(&map->mapcount, 0);
408 g_atomic_int_set(&map->refcount, 0);
409 map->items = bluesky_rangeset_new();
411 g_hash_table_insert(log->mmap_cache, map->filename, map);
413 int fd = openat(log->dirfd, logname, O_WRONLY | O_CREAT, 0600);
415 ftruncate(fd, 5 << 20); // FIXME
419 // If the log file is stored in the cloud, we may need to fetch it
420 if (clouddir >= 0 && start_fetch)
421 cloudlog_fetch_start(map);
423 g_mutex_lock(map->lock);
426 g_mutex_unlock(log->mmap_lock);
428 g_atomic_int_inc(&map->refcount);
432 static void robust_pwrite(int fd, const char *buf, ssize_t count, off_t offset)
435 ssize_t written = pwrite(fd, buf, count, offset);
439 g_warning("pwrite failure: %m");
448 static void cloudlog_partial_fetch_complete(BlueSkyStoreAsync *async,
449 BlueSkyCacheFile *cachefile);
451 static void cloudlog_partial_fetch_start(BlueSkyCacheFile *cachefile,
452 size_t offset, size_t length)
454 g_atomic_int_inc(&cachefile->refcount);
455 g_print("Starting fetch of %s from cloud\n", cachefile->filename);
456 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
457 async->op = STORE_OP_GET;
458 async->key = g_strdup(cachefile->filename);
459 async->start = offset;
461 bluesky_store_async_add_notifier(async,
462 (GFunc)cloudlog_partial_fetch_complete,
464 bluesky_store_async_submit(async);
465 bluesky_store_async_unref(async);
468 static void cloudlog_partial_fetch_complete(BlueSkyStoreAsync *async,
469 BlueSkyCacheFile *cachefile)
471 g_print("Partial fetch of %s from cloud complete, status = %d\n",
472 async->key, async->result);
474 g_mutex_lock(cachefile->lock);
475 if (async->result >= 0) {
476 /* Descrypt items fetched and write valid items out to the local log,
477 * but only if they do not overlap existing objects. This will protect
478 * against an attack by the cloud provider where one valid object is
479 * moved to another offset and used to overwrite data that we already
481 BlueSkyRangeset *items = bluesky_rangeset_new();
482 int fd = openat(cachefile->log->dirfd, cachefile->filename, O_WRONLY);
484 async->data = bluesky_string_dup(async->data);
485 bluesky_cloudlog_decrypt(async->data->data, async->data->len,
486 cachefile->fs->keys, items);
487 uint64_t item_offset = 0;
489 const BlueSkyRangesetItem *item;
490 item = bluesky_rangeset_lookup_next(items, item_offset);
493 g_print(" item offset from range request: %d\n",
494 (int)(item->start + async->start));
495 if (bluesky_rangeset_insert(cachefile->items,
496 async->start + item->start,
497 item->length, item->data))
499 robust_pwrite(fd, async->data->data + item->start,
500 item->length, async->start + item->start);
502 g_print(" item overlaps existing data!\n");
504 item_offset = item->start + 1;
506 /* TODO: Iterate over items and merge into cached file. */
509 g_warning("Unable to open and write to cache file %s: %m",
510 cachefile->filename);
513 g_print("Error fetching from cloud, retrying...\n");
514 cloudlog_partial_fetch_start(cachefile, async->start, async->len);
517 bluesky_cachefile_unref(cachefile);
518 g_cond_broadcast(cachefile->cond);
519 g_mutex_unlock(cachefile->lock);
522 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
524 g_atomic_int_inc(&cachefile->refcount);
525 cachefile->fetching = TRUE;
526 g_print("Starting fetch of %s from cloud\n", cachefile->filename);
527 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
528 async->op = STORE_OP_GET;
529 async->key = g_strdup(cachefile->filename);
530 bluesky_store_async_add_notifier(async,
531 (GFunc)cloudlog_partial_fetch_complete,
533 bluesky_store_async_submit(async);
534 bluesky_store_async_unref(async);
537 /* The arguments are mostly straightforward. log_dir is -1 for access from the
538 * journal, and non-negative for access to a cloud log segment. map_data
539 * should be TRUE for the case that are mapping just the data of an item where
540 * we have already parsed the item headers; this surpresses the error when the
541 * access is not to the first bytes of the item. */
542 BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data)
544 BlueSkyFS *fs = item->fs;
545 BlueSkyLog *log = fs->log;
546 BlueSkyCacheFile *map = NULL;
547 BlueSkyRCStr *str = NULL;
549 size_t file_offset = 0, file_size = 0;
550 gboolean range_request = TRUE;
552 if (page_size == 0) {
553 page_size = getpagesize();
556 bluesky_cloudlog_stats_update(item, -1);
558 /* First, check to see if the journal still contains a copy of the item and
560 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
561 map = bluesky_cachefile_lookup(fs, -1, item->log_seq, TRUE);
563 location = CLOUDLOG_JOURNAL;
564 file_offset = item->log_offset;
565 file_size = item->log_size;
569 if (location == 0 && (item->location_flags & CLOUDLOG_CLOUD)) {
570 item->location_flags &= ~CLOUDLOG_JOURNAL;
571 map = bluesky_cachefile_lookup(fs,
572 item->location.directory,
573 item->location.sequence,
576 g_warning("Unable to remap cloud log segment!");
579 location = CLOUDLOG_CLOUD;
580 file_offset = item->location.offset;
581 file_size = item->location.size;
584 /* Log segments fetched from the cloud might only be partially-fetched.
585 * Check whether the object we are interested in is available. */
586 if (location == CLOUDLOG_CLOUD) {
588 const BlueSkyRangesetItem *rangeitem;
589 rangeitem = bluesky_rangeset_lookup(map->items, file_offset);
590 if (rangeitem != NULL && (rangeitem->start != file_offset
591 || rangeitem->length != file_size)) {
592 g_warning("log-%d: Item offset %zd seems to be invalid!",
593 (int)item->location.sequence, file_offset);
596 if (rangeitem == NULL) {
597 g_print("Item at offset 0x%zx not available, need to fetch.\n",
600 cloudlog_partial_fetch_start(map, file_offset, file_size);
601 g_cond_wait(map->cond, map->lock);
602 } else if (rangeitem->start == file_offset
603 && rangeitem->length == file_size) {
604 g_print("Item now available.\n");
610 if (map->addr == NULL) {
611 int fd = openat(log->dirfd, map->filename, O_RDONLY);
614 fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
618 off_t length = lseek(fd, 0, SEEK_END);
619 map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
621 g_atomic_int_add(&log->disk_used, -(map->len / 1024));
623 g_atomic_int_add(&log->disk_used, map->len / 1024);
625 g_atomic_int_inc(&map->refcount);
631 if (location == CLOUDLOG_JOURNAL)
632 file_offset += sizeof(struct log_header);
634 file_offset += sizeof(struct cloudlog_header);
636 file_size = item->data_size;
638 str = bluesky_string_new_from_mmap(map, file_offset, file_size);
639 map->atime = bluesky_get_current_time();
641 g_print("Returning item at offset 0x%zx.\n", file_offset);
644 bluesky_cachefile_unref(map);
645 g_mutex_unlock(map->lock);
647 bluesky_cloudlog_stats_update(item, 1);
651 void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
656 if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
657 g_mutex_lock(mmap->lock);
658 if (g_atomic_int_get(&mmap->mapcount) == 0) {
659 g_print("Unmapped log segment %d...\n", mmap->log_seq);
660 munmap((void *)mmap->addr, mmap->len);
662 g_atomic_int_add(&mmap->refcount, -1);
664 g_mutex_unlock(mmap->lock);
668 /******************************* JOURNAL REPLAY *******************************
669 * The journal replay code is used to recover filesystem state after a
670 * filesystem restart. We first look for the most recent commit record in the
671 * journal, which indicates the point before which all data in the journal has
672 * also been committed to the cloud. Then, we read in all data in the log past
675 static GList *directory_contents(const char *dirname)
677 GList *contents = NULL;
678 GDir *dir = g_dir_open(dirname, 0, NULL);
680 g_warning("Unable to open journal directory: %s", dirname);
685 while ((file = g_dir_read_name(dir)) != NULL) {
686 if (strncmp(file, "journal-", 8) == 0)
687 contents = g_list_prepend(contents, g_strdup(file));
691 contents = g_list_sort(contents, (GCompareFunc)strcmp);
696 static gboolean validate_journal_item(const char *buf, size_t len, off_t offset)
698 const struct log_header *header;
699 const struct log_footer *footer;
701 if (offset + sizeof(struct log_header) + sizeof(struct log_footer) > len)
704 header = (const struct log_header *)(buf + offset);
705 if (GUINT32_FROM_LE(header->magic) != HEADER_MAGIC)
707 if (GUINT32_FROM_LE(header->offset) != offset)
709 size_t size = GUINT32_FROM_LE(header->size1)
710 + GUINT32_FROM_LE(header->size2)
711 + GUINT32_FROM_LE(header->size3);
713 off_t footer_offset = offset + sizeof(struct log_header) + size;
714 if (footer_offset + sizeof(struct log_footer) > len)
716 footer = (const struct log_footer *)(buf + footer_offset);
718 if (GUINT32_FROM_LE(footer->magic) != FOOTER_MAGIC)
721 uint32_t crc = crc32c(BLUESKY_CRC32C_SEED, buf + offset,
722 sizeof(struct log_header) + sizeof(struct log_footer)
724 if (crc != BLUESKY_CRC32C_VALIDATOR) {
725 g_warning("Journal entry failed to validate: CRC %08x != %08x",
726 crc, BLUESKY_CRC32C_VALIDATOR);
733 /* Scan through a journal segment to extract correctly-written items (those
734 * that pass sanity checks and have a valid checksum). */
735 static void bluesky_replay_scan_journal(const char *buf, size_t len,
736 uint32_t *seq, uint32_t *start_offset)
738 const struct log_header *header;
741 while (validate_journal_item(buf, len, offset)) {
742 header = (const struct log_header *)(buf + offset);
743 size_t size = GUINT32_FROM_LE(header->size1)
744 + GUINT32_FROM_LE(header->size2)
745 + GUINT32_FROM_LE(header->size3);
747 if (header->type - '0' == LOGTYPE_JOURNAL_CHECKPOINT) {
748 const uint32_t *data = (const uint32_t *)((const char *)header + sizeof(struct log_header));
749 *seq = GUINT32_FROM_LE(data[0]);
750 *start_offset = GUINT32_FROM_LE(data[1]);
753 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
757 static void reload_item(BlueSkyCloudLog *log_item,
759 size_t len1, size_t len2, size_t len3)
761 BlueSkyFS *fs = log_item->fs;
762 /*const char *data1 = data;*/
763 const BlueSkyCloudID *data2
764 = (const BlueSkyCloudID *)(data + len1);
765 /*const BlueSkyCloudPointer *data3
766 = (const BlueSkyCloudPointer *)(data + len1 + len2);*/
768 bluesky_cloudlog_stats_update(log_item, -1);
769 bluesky_string_unref(log_item->data);
770 log_item->data = NULL;
771 log_item->location_flags = CLOUDLOG_JOURNAL;
772 bluesky_cloudlog_stats_update(log_item, 1);
775 memset(&id0, 0, sizeof(id0));
777 int link_count = len2 / sizeof(BlueSkyCloudID);
778 GArray *new_links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
779 for (int i = 0; i < link_count; i++) {
780 BlueSkyCloudID id = data2[i];
781 BlueSkyCloudLog *ref = NULL;
782 if (memcmp(&id, &id0, sizeof(BlueSkyCloudID)) != 0) {
783 g_mutex_lock(fs->lock);
784 ref = g_hash_table_lookup(fs->locations, &id);
786 bluesky_cloudlog_ref(ref);
788 g_mutex_unlock(fs->lock);
790 g_array_append_val(new_links, ref);
793 for (int i = 0; i < log_item->links->len; i++) {
794 BlueSkyCloudLog *c = g_array_index(log_item->links,
795 BlueSkyCloudLog *, i);
796 bluesky_cloudlog_unref(c);
798 g_array_unref(log_item->links);
799 log_item->links = new_links;
802 static void bluesky_replay_scan_journal2(BlueSkyFS *fs, GList **objects,
803 int log_seq, int start_offset,
804 const char *buf, size_t len)
806 const struct log_header *header;
807 off_t offset = start_offset;
809 while (validate_journal_item(buf, len, offset)) {
810 header = (const struct log_header *)(buf + offset);
811 g_print("In replay found valid item at offset %zd\n", offset);
812 size_t size = GUINT32_FROM_LE(header->size1)
813 + GUINT32_FROM_LE(header->size2)
814 + GUINT32_FROM_LE(header->size3);
816 BlueSkyCloudLog *log_item = bluesky_cloudlog_get(fs, header->id);
817 g_mutex_lock(log_item->lock);
818 *objects = g_list_prepend(*objects, log_item);
820 log_item->inum = GUINT64_FROM_LE(header->inum);
821 reload_item(log_item, buf + offset + sizeof(struct log_header),
822 GUINT32_FROM_LE(header->size1),
823 GUINT32_FROM_LE(header->size2),
824 GUINT32_FROM_LE(header->size3));
825 log_item->log_seq = log_seq;
826 log_item->log_offset = offset + sizeof(struct log_header);
827 log_item->log_size = header->size1;
829 bluesky_string_unref(log_item->data);
830 log_item->data = bluesky_string_new(g_memdup(buf + offset + sizeof(struct log_header), GUINT32_FROM_LE(header->size1)), GUINT32_FROM_LE(header->size1));
832 /* For any inodes which were read from the journal, deserialize the
833 * inode information, overwriting any old inode data. */
834 if (header->type - '0' == LOGTYPE_INODE) {
835 uint64_t inum = GUINT64_FROM_LE(header->inum);
837 g_mutex_lock(fs->lock);
838 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
840 inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
841 inode->change_count = 0;
842 bluesky_insert_inode(fs, inode);
844 g_mutex_lock(inode->lock);
845 bluesky_inode_free_resources(inode);
846 if (!bluesky_deserialize_inode(inode, log_item))
847 g_print("Error deserializing inode %"PRIu64"\n", inum);
848 fs->next_inum = MAX(fs->next_inum, inum + 1);
849 bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
850 inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
851 bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
852 inode->dirty_list = bluesky_list_prepend(&fs->dirty_list, inode);
853 bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
854 inode->unlogged_list = NULL;
855 inode->change_cloud = inode->change_commit;
856 bluesky_cloudlog_ref(log_item);
857 bluesky_cloudlog_unref(inode->committed_item);
858 inode->committed_item = log_item;
859 g_mutex_unlock(inode->lock);
860 g_mutex_unlock(fs->lock);
862 bluesky_string_unref(log_item->data);
863 log_item->data = NULL;
864 g_mutex_unlock(log_item->lock);
866 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
870 void bluesky_replay(BlueSkyFS *fs)
872 BlueSkyLog *log = fs->log;
873 GList *logfiles = directory_contents(log->log_directory);
875 /* Scan through log files in reverse order to find the most recent commit
877 logfiles = g_list_reverse(logfiles);
878 uint32_t seq_num = 0, start_offset = 0;
879 while (logfiles != NULL) {
880 char *filename = g_strdup_printf("%s/%s", log->log_directory,
881 (char *)logfiles->data);
882 g_print("Scanning file %s\n", filename);
883 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
885 g_warning("Mapping logfile %s failed!\n", filename);
887 bluesky_replay_scan_journal(g_mapped_file_get_contents(map),
888 g_mapped_file_get_length(map),
889 &seq_num, &start_offset);
890 g_mapped_file_unref(map);
894 g_free(logfiles->data);
895 logfiles = g_list_delete_link(logfiles, logfiles);
896 if (seq_num != 0 || start_offset != 0)
899 g_list_foreach(logfiles, (GFunc)g_free, NULL);
900 g_list_free(logfiles);
902 /* Now, scan forward starting from the given point in the log to
903 * reconstruct all filesystem state. As we reload objects we hold a
904 * reference to each loaded object. At the end we free all these
905 * references, so that any objects which were not linked into persistent
906 * filesystem data structures are freed. */
907 GList *objects = NULL;
909 char *filename = g_strdup_printf("%s/journal-%08d",
910 log->log_directory, seq_num);
911 g_print("Replaying file %s from offset %d\n", filename, start_offset);
912 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
915 g_warning("Mapping logfile failed, assuming end of journal\n");
919 bluesky_replay_scan_journal2(fs, &objects, seq_num, start_offset,
920 g_mapped_file_get_contents(map),
921 g_mapped_file_get_length(map));
922 g_mapped_file_unref(map);
927 while (objects != NULL) {
928 bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
929 objects = g_list_delete_link(objects, objects);