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 #define HEADER_MAGIC 0x676f4c0a
38 #define FOOTER_MAGIC 0x2e435243
40 static size_t readbuf(int fd, char *buf, size_t len)
42 size_t total_bytes = 0;
44 ssize_t bytes = read(fd, buf, len);
45 if (bytes < 0 && errno == EINTR)
56 static void writebuf(int fd, const char *buf, size_t len)
60 written = write(fd, buf, len);
61 if (written < 0 && errno == EINTR)
63 g_assert(written >= 0);
69 static void log_commit(BlueSkyLog *log)
78 /* Update disk-space usage statistics for the journal file. */
79 g_atomic_int_add(&log->disk_used, -log->current_log->disk_used);
81 if (fstat(log->fd, &statbuf) >= 0) {
82 /* Convert from 512-byte blocks to 1-kB units */
83 log->current_log->disk_used = (statbuf.st_blocks + 1) / 2;
85 g_atomic_int_add(&log->disk_used, log->current_log->disk_used);
87 while (log->committed != NULL) {
88 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
89 g_mutex_lock(item->lock);
90 bluesky_cloudlog_stats_update(item, -1);
91 item->pending_write &= ~CLOUDLOG_JOURNAL;
93 = (item->location_flags & ~CLOUDLOG_UNCOMMITTED) | CLOUDLOG_JOURNAL;
94 bluesky_cloudlog_stats_update(item, 1);
95 g_cond_signal(item->cond);
96 g_mutex_unlock(item->lock);
97 log->committed = g_slist_delete_link(log->committed, log->committed);
98 bluesky_cloudlog_unref(item);
102 if (bluesky_verbose && batchsize > 1)
103 g_print("Log batch size: %d\n", batchsize);
106 static gboolean log_open(BlueSkyLog *log)
117 if (log->current_log != NULL) {
118 bluesky_cachefile_unref(log->current_log);
119 log->current_log = NULL;
122 while (log->fd < 0) {
123 g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
124 log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
125 if (log->fd < 0 && errno == EEXIST) {
126 fprintf(stderr, "Log file %s already exists...\n", logname);
129 } else if (log->fd < 0) {
130 fprintf(stderr, "Error opening logfile %s: %m\n", logname);
135 log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num,
137 g_assert(log->current_log != NULL);
138 g_mutex_unlock(log->current_log->lock);
140 if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
141 fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
148 /* All log writes (at least for a single log) are made by one thread, so we
149 * don't need to worry about concurrent access to the log file. Log items to
150 * write are pulled off a queue (and so may be posted by any thread).
151 * fdatasync() is used to ensure the log items are stable on disk.
153 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
154 * each. If a log segment is not currently open (log->fd is negative), a new
155 * one is created. Log segment filenames are assigned sequentially.
157 * Log replay ought to be implemented later, and ought to set the initial
158 * sequence number appropriately.
160 static gpointer log_thread(gpointer d)
162 BlueSkyLog *log = (BlueSkyLog *)d;
166 if (!log_open(log)) {
171 BlueSkyCloudLog *item
172 = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
173 g_mutex_lock(item->lock);
174 g_assert(item->data != NULL);
176 /* The item may have already been written to the journal... */
177 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
178 g_mutex_unlock(item->lock);
179 bluesky_cloudlog_unref(item);
180 g_atomic_int_add(&item->data_lock_count, -1);
184 bluesky_cloudlog_stats_update(item, -1);
185 item->pending_write |= CLOUDLOG_JOURNAL;
186 bluesky_cloudlog_stats_update(item, 1);
188 GString *data1 = g_string_new("");
189 GString *data2 = g_string_new("");
190 GString *data3 = g_string_new("");
191 bluesky_serialize_cloudlog(item, data1, data2, data3);
193 struct log_header header;
194 struct log_footer footer;
195 size_t size = sizeof(header) + sizeof(footer);
196 size += data1->len + data2->len + data3->len;
199 offset = lseek(log->fd, 0, SEEK_CUR);
201 /* Check whether the item would overflow the allocated journal size.
202 * If so, start a new log segment. We only allow oversized log
203 * segments if they contain a single log entry. */
204 if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
209 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
210 header.offset = GUINT32_TO_LE(offset);
211 header.size1 = GUINT32_TO_LE(data1->len);
212 header.size2 = GUINT32_TO_LE(data2->len);
213 header.size3 = GUINT32_TO_LE(data3->len);
214 header.type = item->type + '0';
215 header.id = item->id;
216 header.inum = GUINT64_TO_LE(item->inum);
217 footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
219 uint32_t crc = BLUESKY_CRC32C_SEED;
221 writebuf(log->fd, (const char *)&header, sizeof(header));
222 crc = crc32c(crc, (const char *)&header, sizeof(header));
224 writebuf(log->fd, data1->str, data1->len);
225 crc = crc32c(crc, data1->str, data1->len);
226 writebuf(log->fd, data2->str, data2->len);
227 crc = crc32c(crc, data2->str, data2->len);
228 writebuf(log->fd, data3->str, data3->len);
229 crc = crc32c(crc, data3->str, data3->len);
231 crc = crc32c(crc, (const char *)&footer,
232 sizeof(footer) - sizeof(uint32_t));
233 footer.crc = crc32c_finalize(crc);
234 writebuf(log->fd, (const char *)&footer, sizeof(footer));
236 item->log_seq = log->seq_num;
237 item->log_offset = offset;
238 item->log_size = size;
239 item->data_size = item->data->len;
243 g_string_free(data1, TRUE);
244 g_string_free(data2, TRUE);
245 g_string_free(data3, TRUE);
247 /* Replace the log item's string data with a memory-mapped copy of the
248 * data, now that it has been written to the log file. (Even if it
249 * isn't yet on disk, it should at least be in the page cache and so
250 * available to memory map.) */
251 bluesky_string_unref(item->data);
253 bluesky_cloudlog_fetch(item);
255 log->committed = g_slist_prepend(log->committed, item);
256 g_atomic_int_add(&item->data_lock_count, -1);
257 g_mutex_unlock(item->lock);
259 /* Force an if there are no other log items currently waiting to be
261 if (g_async_queue_length(log->queue) <= 0)
268 BlueSkyLog *bluesky_log_new(const char *log_directory)
270 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
272 log->log_directory = g_strdup(log_directory);
275 log->queue = g_async_queue_new();
276 log->mmap_lock = g_mutex_new();
277 log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
279 /* Determine the highest-numbered log file, so that we can start writing
280 * out new journal entries at the next sequence number. */
281 GDir *dir = g_dir_open(log_directory, 0, NULL);
284 while ((file = g_dir_read_name(dir)) != NULL) {
285 if (strncmp(file, "journal-", 8) == 0) {
286 log->seq_num = MAX(log->seq_num, atoi(&file[8]) + 1);
290 g_print("Starting journal at sequence number %d\n", log->seq_num);
293 log->dirfd = open(log->log_directory, O_DIRECTORY);
294 if (log->dirfd < 0) {
295 fprintf(stderr, "Unable to open logging directory: %m\n");
299 g_thread_create(log_thread, log, FALSE, NULL);
304 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
306 if (!(item->location_flags & CLOUDLOG_JOURNAL)) {
307 bluesky_cloudlog_ref(item);
308 item->location_flags |= CLOUDLOG_UNCOMMITTED;
309 g_atomic_int_add(&item->data_lock_count, 1);
310 g_async_queue_push(log->queue, item);
314 void bluesky_log_finish_all(GList *log_items)
316 while (log_items != NULL) {
317 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
319 g_mutex_lock(item->lock);
320 while ((item->location_flags & CLOUDLOG_UNCOMMITTED))
321 g_cond_wait(item->cond, item->lock);
322 g_mutex_unlock(item->lock);
323 bluesky_cloudlog_unref(item);
325 log_items = g_list_delete_link(log_items, log_items);
329 /* Return a committed cloud log record that can be used as a watermark for how
330 * much of the journal has been written. */
331 BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs)
333 BlueSkyCloudLog *marker = bluesky_cloudlog_new(fs, NULL);
334 marker->type = LOGTYPE_JOURNAL_MARKER;
335 marker->data = bluesky_string_new(g_strdup(""), 0);
336 bluesky_cloudlog_stats_update(marker, 1);
337 bluesky_cloudlog_sync(marker);
339 g_mutex_lock(marker->lock);
340 while ((marker->pending_write & CLOUDLOG_JOURNAL))
341 g_cond_wait(marker->cond, marker->lock);
342 g_mutex_unlock(marker->lock);
347 void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker)
349 BlueSkyCloudLog *commit = bluesky_cloudlog_new(fs, NULL);
350 commit->type = LOGTYPE_JOURNAL_CHECKPOINT;
352 uint32_t seq, offset;
353 seq = GUINT32_TO_LE(marker->log_seq);
354 offset = GUINT32_TO_LE(marker->log_offset);
355 GString *loc = g_string_new("");
356 g_string_append_len(loc, (const gchar *)&seq, sizeof(seq));
357 g_string_append_len(loc, (const gchar *)&offset, sizeof(offset));
358 commit->data = bluesky_string_new_from_gstring(loc);
359 bluesky_cloudlog_stats_update(commit, 1);
360 bluesky_cloudlog_sync(commit);
362 g_mutex_lock(commit->lock);
363 while ((commit->location_flags & CLOUDLOG_UNCOMMITTED))
364 g_cond_wait(commit->cond, commit->lock);
365 g_mutex_unlock(commit->lock);
367 bluesky_cloudlog_unref(marker);
368 bluesky_cloudlog_unref(commit);
371 /* Memory-map the given log object into memory (read-only) and return a pointer
373 static int page_size = 0;
375 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
377 g_atomic_int_add(&cachefile->refcount, -1);
380 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile);
382 /* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
383 * Returns the object in the locked state and with a reference taken. */
384 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
385 int clouddir, int log_seq,
386 gboolean start_fetch)
388 if (page_size == 0) {
389 page_size = getpagesize();
392 BlueSkyLog *log = fs->log;
398 // A request for a local log file
400 sprintf(logname, "journal-%08d", log_seq);
401 type = CLOUDLOG_JOURNAL;
403 sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
404 type = CLOUDLOG_CLOUD;
407 BlueSkyCacheFile *map;
408 g_mutex_lock(log->mmap_lock);
409 map = g_hash_table_lookup(log->mmap_cache, logname);
412 && type == CLOUDLOG_JOURNAL
413 && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
414 /* A stale reference to a journal file which doesn't exist any longer
415 * because it was reclaimed. Return NULL. */
416 } else if (map == NULL) {
418 g_print("Adding cache file %s\n", logname);
420 map = g_new0(BlueSkyCacheFile, 1);
423 map->lock = g_mutex_new();
425 g_mutex_lock(map->lock);
426 map->cond = g_cond_new();
427 map->filename = g_strdup(logname);
428 map->log_dir = clouddir;
429 map->log_seq = log_seq;
431 g_atomic_int_set(&map->mapcount, 0);
432 g_atomic_int_set(&map->refcount, 0);
433 map->items = bluesky_rangeset_new();
435 g_hash_table_insert(log->mmap_cache, map->filename, map);
437 int fd = openat(log->dirfd, logname, O_WRONLY | O_CREAT, 0600);
439 ftruncate(fd, 5 << 20); // FIXME
443 g_mutex_lock(map->lock);
447 /* If the log file is stored in the cloud and has not been fully fetched,
448 * we may need to initiate a fetch now. */
449 if (clouddir >= 0 && start_fetch && !map->complete && !map->fetching)
450 cloudlog_fetch_start(map);
452 g_mutex_unlock(log->mmap_lock);
454 g_atomic_int_inc(&map->refcount);
458 static void robust_pwrite(int fd, const char *buf, ssize_t count, off_t offset)
461 ssize_t written = pwrite(fd, buf, count, offset);
465 g_warning("pwrite failure: %m");
474 static void cloudlog_partial_fetch_complete(BlueSkyStoreAsync *async,
475 BlueSkyCacheFile *cachefile);
477 static void cloudlog_partial_fetch_start(BlueSkyCacheFile *cachefile,
478 size_t offset, size_t length)
480 g_atomic_int_inc(&cachefile->refcount);
482 g_print("Starting partial fetch of %s from cloud (%zd + %zd)\n",
483 cachefile->filename, offset, length);
484 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
485 async->op = STORE_OP_GET;
486 async->key = g_strdup(cachefile->filename);
487 async->start = offset;
489 async->profile = bluesky_profile_get();
490 bluesky_store_async_add_notifier(async,
491 (GFunc)cloudlog_partial_fetch_complete,
493 bluesky_store_async_submit(async);
494 bluesky_store_async_unref(async);
497 static void cloudlog_partial_fetch_complete(BlueSkyStoreAsync *async,
498 BlueSkyCacheFile *cachefile)
500 if (bluesky_verbose || async->result != 0)
501 g_print("Fetch of %s from cloud complete, status = %d\n",
502 async->key, async->result);
504 g_mutex_lock(cachefile->lock);
505 if (async->result >= 0) {
506 if (async->len == 0) {
508 g_print("Complete object was fetched.\n");
509 cachefile->complete = TRUE;
512 /* Descrypt items fetched and write valid items out to the local log,
513 * but only if they do not overlap existing objects. This will protect
514 * against an attack by the cloud provider where one valid object is
515 * moved to another offset and used to overwrite data that we already
517 BlueSkyRangeset *items = bluesky_rangeset_new();
518 int fd = openat(cachefile->log->dirfd, cachefile->filename, O_WRONLY);
520 gboolean allow_unauth;
521 async->data = bluesky_string_dup(async->data);
522 allow_unauth = cachefile->log_dir == BLUESKY_CLOUD_DIR_CLEANER;
523 bluesky_cloudlog_decrypt(async->data->data, async->data->len,
524 cachefile->fs->keys, items, allow_unauth);
525 uint64_t item_offset = 0;
527 const BlueSkyRangesetItem *item;
528 item = bluesky_rangeset_lookup_next(items, item_offset);
531 if (bluesky_verbose) {
532 g_print(" item offset from range request: %d\n",
533 (int)(item->start + async->start));
535 if (bluesky_rangeset_insert(cachefile->items,
536 async->start + item->start,
537 item->length, item->data))
539 robust_pwrite(fd, async->data->data + item->start,
540 item->length, async->start + item->start);
542 g_print(" item overlaps existing data!\n");
544 item_offset = item->start + 1;
546 /* TODO: Iterate over items and merge into cached file. */
549 g_warning("Unable to open and write to cache file %s: %m",
550 cachefile->filename);
553 bluesky_rangeset_free(items);
555 g_print("Error fetching %s from cloud, retrying...\n", async->key);
556 cloudlog_partial_fetch_start(cachefile, async->start, async->len);
559 /* Update disk-space usage statistics, since the writes above may have
560 * consumed more space. */
561 g_atomic_int_add(&cachefile->log->disk_used, -cachefile->disk_used);
563 if (fstatat(cachefile->log->dirfd, cachefile->filename, &statbuf, 0) >= 0) {
564 /* Convert from 512-byte blocks to 1-kB units */
565 cachefile->disk_used = (statbuf.st_blocks + 1) / 2;
567 g_atomic_int_add(&cachefile->log->disk_used, cachefile->disk_used);
569 bluesky_cachefile_unref(cachefile);
570 g_cond_broadcast(cachefile->cond);
571 g_mutex_unlock(cachefile->lock);
574 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
576 g_atomic_int_inc(&cachefile->refcount);
577 cachefile->fetching = TRUE;
579 g_print("Starting fetch of %s from cloud\n", cachefile->filename);
580 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
581 async->op = STORE_OP_GET;
582 async->key = g_strdup(cachefile->filename);
583 async->profile = bluesky_profile_get();
584 bluesky_store_async_add_notifier(async,
585 (GFunc)cloudlog_partial_fetch_complete,
587 bluesky_store_async_submit(async);
588 bluesky_store_async_unref(async);
591 /* Map and return a read-only version of a byte range from a cached file. The
592 * CacheFile object must be locked. */
593 BlueSkyRCStr *bluesky_cachefile_map_raw(BlueSkyCacheFile *cachefile,
594 off_t offset, size_t size)
596 cachefile->atime = bluesky_get_current_time();
598 /* Easy case: the needed data is already in memory */
599 if (cachefile->addr != NULL && offset + size <= cachefile->len)
600 return bluesky_string_new_from_mmap(cachefile, offset, size);
602 int fd = openat(cachefile->log->dirfd, cachefile->filename, O_RDONLY);
604 fprintf(stderr, "Error opening logfile %s: %m\n",
605 cachefile->filename);
609 off_t length = lseek(fd, 0, SEEK_END);
610 if (offset + size > length) {
615 /* File is not mapped in memory. Map the entire file in, then return a
616 * pointer to just the required data. */
617 if (cachefile->addr == NULL) {
618 cachefile->addr = (const char *)mmap(NULL, length, PROT_READ,
620 cachefile->len = length;
621 g_atomic_int_inc(&cachefile->refcount);
624 return bluesky_string_new_from_mmap(cachefile, offset, size);
627 /* Otherwise, the file was mapped in but doesn't cover the data we need.
628 * This shouldn't happen much, if at all, but if it does just read the data
629 * we need directly from the file. We lose memory-management benefits of
630 * using mmapped data, but otherwise this works. */
631 char *buf = g_malloc(size);
632 size_t actual_size = readbuf(fd, buf, size);
634 if (actual_size != size) {
638 return bluesky_string_new(buf, size);
642 /* The arguments are mostly straightforward. log_dir is -1 for access from the
643 * journal, and non-negative for access to a cloud log segment. map_data
644 * should be TRUE for the case that are mapping just the data of an item where
645 * we have already parsed the item headers; this surpresses the error when the
646 * access is not to the first bytes of the item. */
647 BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data)
649 BlueSkyFS *fs = item->fs;
650 BlueSkyCacheFile *map = NULL;
651 BlueSkyRCStr *str = NULL;
653 size_t file_offset = 0, file_size = 0;
654 gboolean range_request = bluesky_options.full_segment_fetches
657 if (page_size == 0) {
658 page_size = getpagesize();
661 bluesky_cloudlog_stats_update(item, -1);
663 /* First, check to see if the journal still contains a copy of the item and
665 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
666 map = bluesky_cachefile_lookup(fs, -1, item->log_seq, TRUE);
668 location = CLOUDLOG_JOURNAL;
669 file_offset = item->log_offset;
670 file_size = item->log_size;
674 if (location == 0 && (item->location_flags & CLOUDLOG_CLOUD)) {
675 item->location_flags &= ~CLOUDLOG_JOURNAL;
676 map = bluesky_cachefile_lookup(fs,
677 item->location.directory,
678 item->location.sequence,
681 g_warning("Unable to remap cloud log segment!");
684 location = CLOUDLOG_CLOUD;
685 file_offset = item->location.offset;
686 file_size = item->location.size;
689 /* Log segments fetched from the cloud might only be partially-fetched.
690 * Check whether the object we are interested in is available. */
691 if (location == CLOUDLOG_CLOUD) {
693 const BlueSkyRangesetItem *rangeitem;
694 rangeitem = bluesky_rangeset_lookup(map->items, file_offset);
695 if (rangeitem != NULL && (rangeitem->start != file_offset
696 || rangeitem->length != file_size)) {
697 g_warning("log-%d: Item offset %zd seems to be invalid!",
698 (int)item->location.sequence, file_offset);
701 if (rangeitem == NULL) {
702 if (bluesky_verbose) {
703 g_print("Item at offset 0x%zx not available, need to fetch.\n",
707 uint64_t start = file_offset, length = file_size, end;
708 if (map->prefetches != NULL)
709 bluesky_rangeset_get_extents(map->prefetches,
711 start = MIN(start, file_offset);
712 end = MAX(start + length, file_offset + file_size);
713 length = end - start;
714 cloudlog_partial_fetch_start(map, start, length);
715 if (map->prefetches != NULL) {
716 bluesky_rangeset_free(map->prefetches);
717 map->prefetches = NULL;
720 g_cond_wait(map->cond, map->lock);
721 } else if (rangeitem->start == file_offset
722 && rangeitem->length == file_size) {
724 g_print("Item %zd now available.\n", file_offset);
731 if (location == CLOUDLOG_JOURNAL)
732 file_offset += sizeof(struct log_header);
734 file_offset += sizeof(struct cloudlog_header);
736 file_size = item->data_size;
738 str = bluesky_cachefile_map_raw(map, file_offset, file_size);
741 bluesky_cachefile_unref(map);
742 g_mutex_unlock(map->lock);
744 bluesky_cloudlog_stats_update(item, 1);
748 void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
753 if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
754 g_mutex_lock(mmap->lock);
755 if (mmap->addr != NULL && g_atomic_int_get(&mmap->mapcount) == 0) {
757 g_print("Unmapped log segment %d...\n", mmap->log_seq);
758 munmap((void *)mmap->addr, mmap->len);
760 g_atomic_int_add(&mmap->refcount, -1);
762 g_mutex_unlock(mmap->lock);
764 g_assert(g_atomic_int_get(&mmap->mapcount) >= 0);
767 /******************************* JOURNAL REPLAY *******************************
768 * The journal replay code is used to recover filesystem state after a
769 * filesystem restart. We first look for the most recent commit record in the
770 * journal, which indicates the point before which all data in the journal has
771 * also been committed to the cloud. Then, we read in all data in the log past
774 static GList *directory_contents(const char *dirname)
776 GList *contents = NULL;
777 GDir *dir = g_dir_open(dirname, 0, NULL);
779 g_warning("Unable to open journal directory: %s", dirname);
784 while ((file = g_dir_read_name(dir)) != NULL) {
785 if (strncmp(file, "journal-", 8) == 0)
786 contents = g_list_prepend(contents, g_strdup(file));
790 contents = g_list_sort(contents, (GCompareFunc)strcmp);
795 static gboolean validate_journal_item(const char *buf, size_t len, off_t offset)
797 const struct log_header *header;
798 const struct log_footer *footer;
800 if (offset + sizeof(struct log_header) + sizeof(struct log_footer) > len)
803 header = (const struct log_header *)(buf + offset);
804 if (GUINT32_FROM_LE(header->magic) != HEADER_MAGIC)
806 if (GUINT32_FROM_LE(header->offset) != offset)
808 size_t size = GUINT32_FROM_LE(header->size1)
809 + GUINT32_FROM_LE(header->size2)
810 + GUINT32_FROM_LE(header->size3);
812 off_t footer_offset = offset + sizeof(struct log_header) + size;
813 if (footer_offset + sizeof(struct log_footer) > len)
815 footer = (const struct log_footer *)(buf + footer_offset);
817 if (GUINT32_FROM_LE(footer->magic) != FOOTER_MAGIC)
820 uint32_t crc = crc32c(BLUESKY_CRC32C_SEED, buf + offset,
821 sizeof(struct log_header) + sizeof(struct log_footer)
823 if (crc != BLUESKY_CRC32C_VALIDATOR) {
824 g_warning("Journal entry failed to validate: CRC %08x != %08x",
825 crc, BLUESKY_CRC32C_VALIDATOR);
832 /* Scan through a journal segment to extract correctly-written items (those
833 * that pass sanity checks and have a valid checksum). */
834 static void bluesky_replay_scan_journal(const char *buf, size_t len,
835 uint32_t *seq, uint32_t *start_offset)
837 const struct log_header *header;
840 while (validate_journal_item(buf, len, offset)) {
841 header = (const struct log_header *)(buf + offset);
842 size_t size = GUINT32_FROM_LE(header->size1)
843 + GUINT32_FROM_LE(header->size2)
844 + GUINT32_FROM_LE(header->size3);
846 if (header->type - '0' == LOGTYPE_JOURNAL_CHECKPOINT) {
847 const uint32_t *data = (const uint32_t *)((const char *)header + sizeof(struct log_header));
848 *seq = GUINT32_FROM_LE(data[0]);
849 *start_offset = GUINT32_FROM_LE(data[1]);
852 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
856 static void reload_item(BlueSkyCloudLog *log_item,
858 size_t len1, size_t len2, size_t len3)
860 BlueSkyFS *fs = log_item->fs;
861 /*const char *data1 = data;*/
862 const BlueSkyCloudID *data2
863 = (const BlueSkyCloudID *)(data + len1);
864 /*const BlueSkyCloudPointer *data3
865 = (const BlueSkyCloudPointer *)(data + len1 + len2);*/
867 bluesky_cloudlog_stats_update(log_item, -1);
868 bluesky_string_unref(log_item->data);
869 log_item->data = NULL;
870 log_item->location_flags = CLOUDLOG_JOURNAL;
871 bluesky_cloudlog_stats_update(log_item, 1);
874 memset(&id0, 0, sizeof(id0));
876 int link_count = len2 / sizeof(BlueSkyCloudID);
877 GArray *new_links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
878 for (int i = 0; i < link_count; i++) {
879 BlueSkyCloudID id = data2[i];
880 BlueSkyCloudLog *ref = NULL;
881 if (memcmp(&id, &id0, sizeof(BlueSkyCloudID)) != 0) {
882 g_mutex_lock(fs->lock);
883 ref = g_hash_table_lookup(fs->locations, &id);
885 bluesky_cloudlog_ref(ref);
887 g_mutex_unlock(fs->lock);
889 g_array_append_val(new_links, ref);
892 for (int i = 0; i < log_item->links->len; i++) {
893 BlueSkyCloudLog *c = g_array_index(log_item->links,
894 BlueSkyCloudLog *, i);
895 bluesky_cloudlog_unref(c);
897 g_array_unref(log_item->links);
898 log_item->links = new_links;
901 static void bluesky_replay_scan_journal2(BlueSkyFS *fs, GList **objects,
902 int log_seq, int start_offset,
903 const char *buf, size_t len)
905 const struct log_header *header;
906 off_t offset = start_offset;
908 while (validate_journal_item(buf, len, offset)) {
909 header = (const struct log_header *)(buf + offset);
910 g_print("In replay found valid item at offset %zd\n", offset);
911 size_t size = GUINT32_FROM_LE(header->size1)
912 + GUINT32_FROM_LE(header->size2)
913 + GUINT32_FROM_LE(header->size3);
915 BlueSkyCloudLog *log_item = bluesky_cloudlog_get(fs, header->id);
916 g_mutex_lock(log_item->lock);
917 *objects = g_list_prepend(*objects, log_item);
919 log_item->inum = GUINT64_FROM_LE(header->inum);
920 reload_item(log_item, buf + offset + sizeof(struct log_header),
921 GUINT32_FROM_LE(header->size1),
922 GUINT32_FROM_LE(header->size2),
923 GUINT32_FROM_LE(header->size3));
924 log_item->log_seq = log_seq;
925 log_item->log_offset = offset + sizeof(struct log_header);
926 log_item->log_size = header->size1;
928 bluesky_string_unref(log_item->data);
929 log_item->data = bluesky_string_new(g_memdup(buf + offset + sizeof(struct log_header), GUINT32_FROM_LE(header->size1)), GUINT32_FROM_LE(header->size1));
931 /* For any inodes which were read from the journal, deserialize the
932 * inode information, overwriting any old inode data. */
933 if (header->type - '0' == LOGTYPE_INODE) {
934 uint64_t inum = GUINT64_FROM_LE(header->inum);
936 g_mutex_lock(fs->lock);
937 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
939 inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
940 inode->change_count = 0;
941 bluesky_insert_inode(fs, inode);
943 g_mutex_lock(inode->lock);
944 bluesky_inode_free_resources(inode);
945 if (!bluesky_deserialize_inode(inode, log_item))
946 g_print("Error deserializing inode %"PRIu64"\n", inum);
947 fs->next_inum = MAX(fs->next_inum, inum + 1);
948 bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
949 inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
950 bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
951 inode->dirty_list = bluesky_list_prepend(&fs->dirty_list, inode);
952 bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
953 inode->unlogged_list = NULL;
954 inode->change_cloud = inode->change_commit;
955 bluesky_cloudlog_ref(log_item);
956 bluesky_cloudlog_unref(inode->committed_item);
957 inode->committed_item = log_item;
958 g_mutex_unlock(inode->lock);
959 g_mutex_unlock(fs->lock);
961 bluesky_string_unref(log_item->data);
962 log_item->data = NULL;
963 g_mutex_unlock(log_item->lock);
965 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
969 void bluesky_replay(BlueSkyFS *fs)
971 BlueSkyLog *log = fs->log;
972 GList *logfiles = directory_contents(log->log_directory);
974 /* Scan through log files in reverse order to find the most recent commit
976 logfiles = g_list_reverse(logfiles);
977 uint32_t seq_num = 0, start_offset = 0;
978 while (logfiles != NULL) {
979 char *filename = g_strdup_printf("%s/%s", log->log_directory,
980 (char *)logfiles->data);
981 g_print("Scanning file %s\n", filename);
982 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
984 g_warning("Mapping logfile %s failed!\n", filename);
986 bluesky_replay_scan_journal(g_mapped_file_get_contents(map),
987 g_mapped_file_get_length(map),
988 &seq_num, &start_offset);
989 g_mapped_file_unref(map);
993 g_free(logfiles->data);
994 logfiles = g_list_delete_link(logfiles, logfiles);
995 if (seq_num != 0 || start_offset != 0)
998 g_list_foreach(logfiles, (GFunc)g_free, NULL);
999 g_list_free(logfiles);
1001 /* Now, scan forward starting from the given point in the log to
1002 * reconstruct all filesystem state. As we reload objects we hold a
1003 * reference to each loaded object. At the end we free all these
1004 * references, so that any objects which were not linked into persistent
1005 * filesystem data structures are freed. */
1006 GList *objects = NULL;
1008 char *filename = g_strdup_printf("%s/journal-%08d",
1009 log->log_directory, seq_num);
1010 g_print("Replaying file %s from offset %d\n", filename, start_offset);
1011 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
1014 g_warning("Mapping logfile failed, assuming end of journal\n");
1018 bluesky_replay_scan_journal2(fs, &objects, seq_num, start_offset,
1019 g_mapped_file_get_contents(map),
1020 g_mapped_file_get_length(map));
1021 g_mapped_file_unref(map);
1026 while (objects != NULL) {
1027 bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
1028 objects = g_list_delete_link(objects, objects);