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 size_t readbuf(int fd, char *buf, size_t len)
46 size_t total_bytes = 0;
48 ssize_t bytes = read(fd, buf, len);
49 if (bytes < 0 && errno == EINTR)
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)
82 /* Update disk-space usage statistics for the journal file. */
83 g_atomic_int_add(&log->disk_used, -log->current_log->disk_used);
85 if (fstat(log->fd, &statbuf) >= 0) {
86 /* Convert from 512-byte blocks to 1-kB units */
87 log->current_log->disk_used = (statbuf.st_blocks + 1) / 2;
89 g_atomic_int_add(&log->disk_used, log->current_log->disk_used);
91 while (log->committed != NULL) {
92 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
93 g_mutex_lock(item->lock);
94 bluesky_cloudlog_stats_update(item, -1);
95 item->pending_write &= ~CLOUDLOG_JOURNAL;
97 = (item->location_flags & ~CLOUDLOG_UNCOMMITTED) | CLOUDLOG_JOURNAL;
98 bluesky_cloudlog_stats_update(item, 1);
99 g_cond_signal(item->cond);
100 g_mutex_unlock(item->lock);
101 log->committed = g_slist_delete_link(log->committed, log->committed);
102 bluesky_cloudlog_unref(item);
106 if (bluesky_verbose && batchsize > 1)
107 g_print("Log batch size: %d\n", batchsize);
110 static gboolean log_open(BlueSkyLog *log)
121 if (log->current_log != NULL) {
122 bluesky_cachefile_unref(log->current_log);
123 log->current_log = NULL;
126 while (log->fd < 0) {
127 g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
128 log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
129 if (log->fd < 0 && errno == EEXIST) {
130 fprintf(stderr, "Log file %s already exists...\n", logname);
133 } else if (log->fd < 0) {
134 fprintf(stderr, "Error opening logfile %s: %m\n", logname);
139 log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num,
141 g_assert(log->current_log != NULL);
142 g_mutex_unlock(log->current_log->lock);
144 if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
145 fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
152 /* All log writes (at least for a single log) are made by one thread, so we
153 * don't need to worry about concurrent access to the log file. Log items to
154 * write are pulled off a queue (and so may be posted by any thread).
155 * fdatasync() is used to ensure the log items are stable on disk.
157 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
158 * each. If a log segment is not currently open (log->fd is negative), a new
159 * one is created. Log segment filenames are assigned sequentially.
161 * Log replay ought to be implemented later, and ought to set the initial
162 * sequence number appropriately.
164 static gpointer log_thread(gpointer d)
166 BlueSkyLog *log = (BlueSkyLog *)d;
170 if (!log_open(log)) {
175 BlueSkyCloudLog *item
176 = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
177 g_mutex_lock(item->lock);
178 g_assert(item->data != NULL);
180 /* The item may have already been written to the journal... */
181 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
182 g_mutex_unlock(item->lock);
183 bluesky_cloudlog_unref(item);
184 g_atomic_int_add(&item->data_lock_count, -1);
188 bluesky_cloudlog_stats_update(item, -1);
189 item->pending_write |= CLOUDLOG_JOURNAL;
190 bluesky_cloudlog_stats_update(item, 1);
192 GString *data1 = g_string_new("");
193 GString *data2 = g_string_new("");
194 GString *data3 = g_string_new("");
195 bluesky_serialize_cloudlog(item, data1, data2, data3);
197 struct log_header header;
198 struct log_footer footer;
199 size_t size = sizeof(header) + sizeof(footer);
200 size += data1->len + data2->len + data3->len;
203 offset = lseek(log->fd, 0, SEEK_CUR);
205 /* Check whether the item would overflow the allocated journal size.
206 * If so, start a new log segment. We only allow oversized log
207 * segments if they contain a single log entry. */
208 if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
213 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
214 header.offset = GUINT32_TO_LE(offset);
215 header.size1 = GUINT32_TO_LE(data1->len);
216 header.size2 = GUINT32_TO_LE(data2->len);
217 header.size3 = GUINT32_TO_LE(data3->len);
218 header.type = item->type + '0';
219 header.id = item->id;
220 header.inum = GUINT64_TO_LE(item->inum);
221 footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
223 uint32_t crc = BLUESKY_CRC32C_SEED;
225 writebuf(log->fd, (const char *)&header, sizeof(header));
226 crc = crc32c(crc, (const char *)&header, sizeof(header));
228 writebuf(log->fd, data1->str, data1->len);
229 crc = crc32c(crc, data1->str, data1->len);
230 writebuf(log->fd, data2->str, data2->len);
231 crc = crc32c(crc, data2->str, data2->len);
232 writebuf(log->fd, data3->str, data3->len);
233 crc = crc32c(crc, data3->str, data3->len);
235 crc = crc32c(crc, (const char *)&footer,
236 sizeof(footer) - sizeof(uint32_t));
237 footer.crc = crc32c_finalize(crc);
238 writebuf(log->fd, (const char *)&footer, sizeof(footer));
240 item->log_seq = log->seq_num;
241 item->log_offset = offset;
242 item->log_size = size;
243 item->data_size = item->data->len;
247 g_string_free(data1, TRUE);
248 g_string_free(data2, TRUE);
249 g_string_free(data3, TRUE);
251 /* Replace the log item's string data with a memory-mapped copy of the
252 * data, now that it has been written to the log file. (Even if it
253 * isn't yet on disk, it should at least be in the page cache and so
254 * available to memory map.) */
255 bluesky_string_unref(item->data);
257 bluesky_cloudlog_fetch(item);
259 log->committed = g_slist_prepend(log->committed, item);
260 g_atomic_int_add(&item->data_lock_count, -1);
261 g_mutex_unlock(item->lock);
263 /* Force an if there are no other log items currently waiting to be
265 if (g_async_queue_length(log->queue) <= 0)
272 BlueSkyLog *bluesky_log_new(const char *log_directory)
274 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
276 log->log_directory = g_strdup(log_directory);
279 log->queue = g_async_queue_new();
280 log->mmap_lock = g_mutex_new();
281 log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
283 /* Determine the highest-numbered log file, so that we can start writing
284 * out new journal entries at the next sequence number. */
285 GDir *dir = g_dir_open(log_directory, 0, NULL);
288 while ((file = g_dir_read_name(dir)) != NULL) {
289 if (strncmp(file, "journal-", 8) == 0) {
290 log->seq_num = MAX(log->seq_num, atoi(&file[8]) + 1);
294 g_print("Starting journal at sequence number %d\n", log->seq_num);
297 log->dirfd = open(log->log_directory, O_DIRECTORY);
298 if (log->dirfd < 0) {
299 fprintf(stderr, "Unable to open logging directory: %m\n");
303 g_thread_create(log_thread, log, FALSE, NULL);
308 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
310 if (!(item->location_flags & CLOUDLOG_JOURNAL)) {
311 bluesky_cloudlog_ref(item);
312 item->location_flags |= CLOUDLOG_UNCOMMITTED;
313 g_atomic_int_add(&item->data_lock_count, 1);
314 g_async_queue_push(log->queue, item);
318 void bluesky_log_finish_all(GList *log_items)
320 while (log_items != NULL) {
321 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
323 g_mutex_lock(item->lock);
324 while ((item->location_flags & CLOUDLOG_UNCOMMITTED))
325 g_cond_wait(item->cond, item->lock);
326 g_mutex_unlock(item->lock);
327 bluesky_cloudlog_unref(item);
329 log_items = g_list_delete_link(log_items, log_items);
333 /* Return a committed cloud log record that can be used as a watermark for how
334 * much of the journal has been written. */
335 BlueSkyCloudLog *bluesky_log_get_commit_point(BlueSkyFS *fs)
337 BlueSkyCloudLog *marker = bluesky_cloudlog_new(fs, NULL);
338 marker->type = LOGTYPE_JOURNAL_MARKER;
339 marker->data = bluesky_string_new(g_strdup(""), 0);
340 bluesky_cloudlog_stats_update(marker, 1);
341 bluesky_cloudlog_sync(marker);
343 g_mutex_lock(marker->lock);
344 while ((marker->pending_write & CLOUDLOG_JOURNAL))
345 g_cond_wait(marker->cond, marker->lock);
346 g_mutex_unlock(marker->lock);
351 void bluesky_log_write_commit_point(BlueSkyFS *fs, BlueSkyCloudLog *marker)
353 BlueSkyCloudLog *commit = bluesky_cloudlog_new(fs, NULL);
354 commit->type = LOGTYPE_JOURNAL_CHECKPOINT;
356 uint32_t seq, offset;
357 seq = GUINT32_TO_LE(marker->log_seq);
358 offset = GUINT32_TO_LE(marker->log_offset);
359 GString *loc = g_string_new("");
360 g_string_append_len(loc, (const gchar *)&seq, sizeof(seq));
361 g_string_append_len(loc, (const gchar *)&offset, sizeof(offset));
362 commit->data = bluesky_string_new_from_gstring(loc);
363 bluesky_cloudlog_stats_update(commit, 1);
364 bluesky_cloudlog_sync(commit);
366 g_mutex_lock(commit->lock);
367 while ((commit->location_flags & CLOUDLOG_UNCOMMITTED))
368 g_cond_wait(commit->cond, commit->lock);
369 g_mutex_unlock(commit->lock);
371 bluesky_cloudlog_unref(marker);
372 bluesky_cloudlog_unref(commit);
375 /* Memory-map the given log object into memory (read-only) and return a pointer
377 static int page_size = 0;
379 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
381 g_atomic_int_add(&cachefile->refcount, -1);
384 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile);
386 /* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
387 * Returns the object in the locked state and with a reference taken. */
388 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
389 int clouddir, int log_seq,
390 gboolean start_fetch)
392 if (page_size == 0) {
393 page_size = getpagesize();
396 BlueSkyLog *log = fs->log;
402 // A request for a local log file
404 sprintf(logname, "journal-%08d", log_seq);
405 type = CLOUDLOG_JOURNAL;
407 sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
408 type = CLOUDLOG_CLOUD;
411 BlueSkyCacheFile *map;
412 g_mutex_lock(log->mmap_lock);
413 map = g_hash_table_lookup(log->mmap_cache, logname);
416 && type == CLOUDLOG_JOURNAL
417 && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
418 /* A stale reference to a journal file which doesn't exist any longer
419 * because it was reclaimed. Return NULL. */
420 } else if (map == NULL) {
422 g_print("Adding cache file %s\n", logname);
424 map = g_new0(BlueSkyCacheFile, 1);
427 map->lock = g_mutex_new();
429 g_mutex_lock(map->lock);
430 map->cond = g_cond_new();
431 map->filename = g_strdup(logname);
432 map->log_dir = clouddir;
433 map->log_seq = log_seq;
435 g_atomic_int_set(&map->mapcount, 0);
436 g_atomic_int_set(&map->refcount, 0);
437 map->items = bluesky_rangeset_new();
439 g_hash_table_insert(log->mmap_cache, map->filename, map);
441 int fd = openat(log->dirfd, logname, O_WRONLY | O_CREAT, 0600);
443 ftruncate(fd, 5 << 20); // FIXME
447 g_mutex_lock(map->lock);
451 /* If the log file is stored in the cloud and has not been fully fetched,
452 * we may need to initiate a fetch now. */
453 if (clouddir >= 0 && start_fetch && !map->complete && !map->fetching)
454 cloudlog_fetch_start(map);
456 g_mutex_unlock(log->mmap_lock);
458 g_atomic_int_inc(&map->refcount);
462 static void robust_pwrite(int fd, const char *buf, ssize_t count, off_t offset)
465 ssize_t written = pwrite(fd, buf, count, offset);
469 g_warning("pwrite failure: %m");
478 static void cloudlog_partial_fetch_complete(BlueSkyStoreAsync *async,
479 BlueSkyCacheFile *cachefile);
481 static void cloudlog_partial_fetch_start(BlueSkyCacheFile *cachefile,
482 size_t offset, size_t length)
484 g_atomic_int_inc(&cachefile->refcount);
486 g_print("Starting partial fetch of %s from cloud (%zd + %zd)\n",
487 cachefile->filename, offset, length);
488 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
489 async->op = STORE_OP_GET;
490 async->key = g_strdup(cachefile->filename);
491 async->start = offset;
493 async->profile = bluesky_profile_get();
494 bluesky_store_async_add_notifier(async,
495 (GFunc)cloudlog_partial_fetch_complete,
497 bluesky_store_async_submit(async);
498 bluesky_store_async_unref(async);
501 static void cloudlog_partial_fetch_complete(BlueSkyStoreAsync *async,
502 BlueSkyCacheFile *cachefile)
504 if (bluesky_verbose || async->result != 0)
505 g_print("Fetch of %s from cloud complete, status = %d\n",
506 async->key, async->result);
508 g_mutex_lock(cachefile->lock);
509 if (async->result >= 0) {
510 if (async->len == 0) {
512 g_print("Complete object was fetched.\n");
513 cachefile->complete = TRUE;
516 /* Descrypt items fetched and write valid items out to the local log,
517 * but only if they do not overlap existing objects. This will protect
518 * against an attack by the cloud provider where one valid object is
519 * moved to another offset and used to overwrite data that we already
521 BlueSkyRangeset *items = bluesky_rangeset_new();
522 int fd = openat(cachefile->log->dirfd, cachefile->filename, O_WRONLY);
524 gboolean allow_unauth;
525 async->data = bluesky_string_dup(async->data);
526 allow_unauth = cachefile->log_dir == BLUESKY_CLOUD_DIR_CLEANER;
527 bluesky_cloudlog_decrypt(async->data->data, async->data->len,
528 cachefile->fs->keys, items, allow_unauth);
529 uint64_t item_offset = 0;
531 const BlueSkyRangesetItem *item;
532 item = bluesky_rangeset_lookup_next(items, item_offset);
535 if (bluesky_verbose) {
536 g_print(" item offset from range request: %d\n",
537 (int)(item->start + async->start));
539 if (bluesky_rangeset_insert(cachefile->items,
540 async->start + item->start,
541 item->length, item->data))
543 robust_pwrite(fd, async->data->data + item->start,
544 item->length, async->start + item->start);
546 g_print(" item overlaps existing data!\n");
548 item_offset = item->start + 1;
550 /* TODO: Iterate over items and merge into cached file. */
553 g_warning("Unable to open and write to cache file %s: %m",
554 cachefile->filename);
557 bluesky_rangeset_free(items);
559 g_print("Error fetching %s from cloud, retrying...\n", async->key);
560 cloudlog_partial_fetch_start(cachefile, async->start, async->len);
563 /* Update disk-space usage statistics, since the writes above may have
564 * consumed more space. */
565 g_atomic_int_add(&cachefile->log->disk_used, -cachefile->disk_used);
567 if (fstatat(cachefile->log->dirfd, cachefile->filename, &statbuf, 0) >= 0) {
568 /* Convert from 512-byte blocks to 1-kB units */
569 cachefile->disk_used = (statbuf.st_blocks + 1) / 2;
571 g_atomic_int_add(&cachefile->log->disk_used, cachefile->disk_used);
573 bluesky_cachefile_unref(cachefile);
574 g_cond_broadcast(cachefile->cond);
575 g_mutex_unlock(cachefile->lock);
578 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
580 g_atomic_int_inc(&cachefile->refcount);
581 cachefile->fetching = TRUE;
583 g_print("Starting fetch of %s from cloud\n", cachefile->filename);
584 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
585 async->op = STORE_OP_GET;
586 async->key = g_strdup(cachefile->filename);
587 async->profile = bluesky_profile_get();
588 bluesky_store_async_add_notifier(async,
589 (GFunc)cloudlog_partial_fetch_complete,
591 bluesky_store_async_submit(async);
592 bluesky_store_async_unref(async);
595 /* Map and return a read-only version of a byte range from a cached file. The
596 * CacheFile object must be locked. */
597 BlueSkyRCStr *bluesky_cachefile_map_raw(BlueSkyCacheFile *cachefile,
598 off_t offset, size_t size)
600 cachefile->atime = bluesky_get_current_time();
602 /* Easy case: the needed data is already in memory */
603 if (cachefile->addr != NULL && offset + size <= cachefile->len)
604 return bluesky_string_new_from_mmap(cachefile, offset, size);
606 int fd = openat(cachefile->log->dirfd, cachefile->filename, O_RDONLY);
608 fprintf(stderr, "Error opening logfile %s: %m\n",
609 cachefile->filename);
613 off_t length = lseek(fd, 0, SEEK_END);
614 if (offset + size > length) {
619 /* File is not mapped in memory. Map the entire file in, then return a
620 * pointer to just the required data. */
621 if (cachefile->addr == NULL) {
622 cachefile->addr = (const char *)mmap(NULL, length, PROT_READ,
624 cachefile->len = length;
625 g_atomic_int_inc(&cachefile->refcount);
628 return bluesky_string_new_from_mmap(cachefile, offset, size);
631 /* Otherwise, the file was mapped in but doesn't cover the data we need.
632 * This shouldn't happen much, if at all, but if it does just read the data
633 * we need directly from the file. We lose memory-management benefits of
634 * using mmapped data, but otherwise this works. */
635 char *buf = g_malloc(size);
636 size_t actual_size = readbuf(fd, buf, size);
638 if (actual_size != size) {
642 return bluesky_string_new(buf, size);
646 /* The arguments are mostly straightforward. log_dir is -1 for access from the
647 * journal, and non-negative for access to a cloud log segment. map_data
648 * should be TRUE for the case that are mapping just the data of an item where
649 * we have already parsed the item headers; this surpresses the error when the
650 * access is not to the first bytes of the item. */
651 BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data)
653 BlueSkyFS *fs = item->fs;
654 BlueSkyCacheFile *map = NULL;
655 BlueSkyRCStr *str = NULL;
657 size_t file_offset = 0, file_size = 0;
658 gboolean range_request = bluesky_options.full_segment_fetches
661 if (page_size == 0) {
662 page_size = getpagesize();
665 bluesky_cloudlog_stats_update(item, -1);
667 /* First, check to see if the journal still contains a copy of the item and
669 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
670 map = bluesky_cachefile_lookup(fs, -1, item->log_seq, TRUE);
672 location = CLOUDLOG_JOURNAL;
673 file_offset = item->log_offset;
674 file_size = item->log_size;
678 if (location == 0 && (item->location_flags & CLOUDLOG_CLOUD)) {
679 item->location_flags &= ~CLOUDLOG_JOURNAL;
680 map = bluesky_cachefile_lookup(fs,
681 item->location.directory,
682 item->location.sequence,
685 g_warning("Unable to remap cloud log segment!");
688 location = CLOUDLOG_CLOUD;
689 file_offset = item->location.offset;
690 file_size = item->location.size;
693 /* Log segments fetched from the cloud might only be partially-fetched.
694 * Check whether the object we are interested in is available. */
695 if (location == CLOUDLOG_CLOUD) {
697 const BlueSkyRangesetItem *rangeitem;
698 rangeitem = bluesky_rangeset_lookup(map->items, file_offset);
699 if (rangeitem != NULL && (rangeitem->start != file_offset
700 || rangeitem->length != file_size)) {
701 g_warning("log-%d: Item offset %zd seems to be invalid!",
702 (int)item->location.sequence, file_offset);
705 if (rangeitem == NULL) {
706 if (bluesky_verbose) {
707 g_print("Item at offset 0x%zx not available, need to fetch.\n",
711 uint64_t start = file_offset, length = file_size, end;
712 if (map->prefetches != NULL)
713 bluesky_rangeset_get_extents(map->prefetches,
715 start = MIN(start, file_offset);
716 end = MAX(start + length, file_offset + file_size);
717 length = end - start;
718 cloudlog_partial_fetch_start(map, start, length);
719 if (map->prefetches != NULL) {
720 bluesky_rangeset_free(map->prefetches);
721 map->prefetches = NULL;
724 g_cond_wait(map->cond, map->lock);
725 } else if (rangeitem->start == file_offset
726 && rangeitem->length == file_size) {
728 g_print("Item %zd now available.\n", file_offset);
735 if (location == CLOUDLOG_JOURNAL)
736 file_offset += sizeof(struct log_header);
738 file_offset += sizeof(struct cloudlog_header);
740 file_size = item->data_size;
742 str = bluesky_cachefile_map_raw(map, file_offset, file_size);
745 bluesky_cachefile_unref(map);
746 g_mutex_unlock(map->lock);
748 bluesky_cloudlog_stats_update(item, 1);
752 void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
757 if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
758 g_mutex_lock(mmap->lock);
759 if (mmap->addr != NULL && g_atomic_int_get(&mmap->mapcount) == 0) {
761 g_print("Unmapped log segment %d...\n", mmap->log_seq);
762 munmap((void *)mmap->addr, mmap->len);
764 g_atomic_int_add(&mmap->refcount, -1);
766 g_mutex_unlock(mmap->lock);
768 g_assert(g_atomic_int_get(&mmap->mapcount) >= 0);
771 /******************************* JOURNAL REPLAY *******************************
772 * The journal replay code is used to recover filesystem state after a
773 * filesystem restart. We first look for the most recent commit record in the
774 * journal, which indicates the point before which all data in the journal has
775 * also been committed to the cloud. Then, we read in all data in the log past
778 static GList *directory_contents(const char *dirname)
780 GList *contents = NULL;
781 GDir *dir = g_dir_open(dirname, 0, NULL);
783 g_warning("Unable to open journal directory: %s", dirname);
788 while ((file = g_dir_read_name(dir)) != NULL) {
789 if (strncmp(file, "journal-", 8) == 0)
790 contents = g_list_prepend(contents, g_strdup(file));
794 contents = g_list_sort(contents, (GCompareFunc)strcmp);
799 static gboolean validate_journal_item(const char *buf, size_t len, off_t offset)
801 const struct log_header *header;
802 const struct log_footer *footer;
804 if (offset + sizeof(struct log_header) + sizeof(struct log_footer) > len)
807 header = (const struct log_header *)(buf + offset);
808 if (GUINT32_FROM_LE(header->magic) != HEADER_MAGIC)
810 if (GUINT32_FROM_LE(header->offset) != offset)
812 size_t size = GUINT32_FROM_LE(header->size1)
813 + GUINT32_FROM_LE(header->size2)
814 + GUINT32_FROM_LE(header->size3);
816 off_t footer_offset = offset + sizeof(struct log_header) + size;
817 if (footer_offset + sizeof(struct log_footer) > len)
819 footer = (const struct log_footer *)(buf + footer_offset);
821 if (GUINT32_FROM_LE(footer->magic) != FOOTER_MAGIC)
824 uint32_t crc = crc32c(BLUESKY_CRC32C_SEED, buf + offset,
825 sizeof(struct log_header) + sizeof(struct log_footer)
827 if (crc != BLUESKY_CRC32C_VALIDATOR) {
828 g_warning("Journal entry failed to validate: CRC %08x != %08x",
829 crc, BLUESKY_CRC32C_VALIDATOR);
836 /* Scan through a journal segment to extract correctly-written items (those
837 * that pass sanity checks and have a valid checksum). */
838 static void bluesky_replay_scan_journal(const char *buf, size_t len,
839 uint32_t *seq, uint32_t *start_offset)
841 const struct log_header *header;
844 while (validate_journal_item(buf, len, offset)) {
845 header = (const struct log_header *)(buf + offset);
846 size_t size = GUINT32_FROM_LE(header->size1)
847 + GUINT32_FROM_LE(header->size2)
848 + GUINT32_FROM_LE(header->size3);
850 if (header->type - '0' == LOGTYPE_JOURNAL_CHECKPOINT) {
851 const uint32_t *data = (const uint32_t *)((const char *)header + sizeof(struct log_header));
852 *seq = GUINT32_FROM_LE(data[0]);
853 *start_offset = GUINT32_FROM_LE(data[1]);
856 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
860 static void reload_item(BlueSkyCloudLog *log_item,
862 size_t len1, size_t len2, size_t len3)
864 BlueSkyFS *fs = log_item->fs;
865 /*const char *data1 = data;*/
866 const BlueSkyCloudID *data2
867 = (const BlueSkyCloudID *)(data + len1);
868 /*const BlueSkyCloudPointer *data3
869 = (const BlueSkyCloudPointer *)(data + len1 + len2);*/
871 bluesky_cloudlog_stats_update(log_item, -1);
872 bluesky_string_unref(log_item->data);
873 log_item->data = NULL;
874 log_item->location_flags = CLOUDLOG_JOURNAL;
875 bluesky_cloudlog_stats_update(log_item, 1);
878 memset(&id0, 0, sizeof(id0));
880 int link_count = len2 / sizeof(BlueSkyCloudID);
881 GArray *new_links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
882 for (int i = 0; i < link_count; i++) {
883 BlueSkyCloudID id = data2[i];
884 BlueSkyCloudLog *ref = NULL;
885 if (memcmp(&id, &id0, sizeof(BlueSkyCloudID)) != 0) {
886 g_mutex_lock(fs->lock);
887 ref = g_hash_table_lookup(fs->locations, &id);
889 bluesky_cloudlog_ref(ref);
891 g_mutex_unlock(fs->lock);
893 g_array_append_val(new_links, ref);
896 for (int i = 0; i < log_item->links->len; i++) {
897 BlueSkyCloudLog *c = g_array_index(log_item->links,
898 BlueSkyCloudLog *, i);
899 bluesky_cloudlog_unref(c);
901 g_array_unref(log_item->links);
902 log_item->links = new_links;
905 static void bluesky_replay_scan_journal2(BlueSkyFS *fs, GList **objects,
906 int log_seq, int start_offset,
907 const char *buf, size_t len)
909 const struct log_header *header;
910 off_t offset = start_offset;
912 while (validate_journal_item(buf, len, offset)) {
913 header = (const struct log_header *)(buf + offset);
914 g_print("In replay found valid item at offset %zd\n", offset);
915 size_t size = GUINT32_FROM_LE(header->size1)
916 + GUINT32_FROM_LE(header->size2)
917 + GUINT32_FROM_LE(header->size3);
919 BlueSkyCloudLog *log_item = bluesky_cloudlog_get(fs, header->id);
920 g_mutex_lock(log_item->lock);
921 *objects = g_list_prepend(*objects, log_item);
923 log_item->inum = GUINT64_FROM_LE(header->inum);
924 reload_item(log_item, buf + offset + sizeof(struct log_header),
925 GUINT32_FROM_LE(header->size1),
926 GUINT32_FROM_LE(header->size2),
927 GUINT32_FROM_LE(header->size3));
928 log_item->log_seq = log_seq;
929 log_item->log_offset = offset + sizeof(struct log_header);
930 log_item->log_size = header->size1;
932 bluesky_string_unref(log_item->data);
933 log_item->data = bluesky_string_new(g_memdup(buf + offset + sizeof(struct log_header), GUINT32_FROM_LE(header->size1)), GUINT32_FROM_LE(header->size1));
935 /* For any inodes which were read from the journal, deserialize the
936 * inode information, overwriting any old inode data. */
937 if (header->type - '0' == LOGTYPE_INODE) {
938 uint64_t inum = GUINT64_FROM_LE(header->inum);
940 g_mutex_lock(fs->lock);
941 inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
943 inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
944 inode->change_count = 0;
945 bluesky_insert_inode(fs, inode);
947 g_mutex_lock(inode->lock);
948 bluesky_inode_free_resources(inode);
949 if (!bluesky_deserialize_inode(inode, log_item))
950 g_print("Error deserializing inode %"PRIu64"\n", inum);
951 fs->next_inum = MAX(fs->next_inum, inum + 1);
952 bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
953 inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
954 bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
955 inode->dirty_list = bluesky_list_prepend(&fs->dirty_list, inode);
956 bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
957 inode->unlogged_list = NULL;
958 inode->change_cloud = inode->change_commit;
959 bluesky_cloudlog_ref(log_item);
960 bluesky_cloudlog_unref(inode->committed_item);
961 inode->committed_item = log_item;
962 g_mutex_unlock(inode->lock);
963 g_mutex_unlock(fs->lock);
965 bluesky_string_unref(log_item->data);
966 log_item->data = NULL;
967 g_mutex_unlock(log_item->lock);
969 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
973 void bluesky_replay(BlueSkyFS *fs)
975 BlueSkyLog *log = fs->log;
976 GList *logfiles = directory_contents(log->log_directory);
978 /* Scan through log files in reverse order to find the most recent commit
980 logfiles = g_list_reverse(logfiles);
981 uint32_t seq_num = 0, start_offset = 0;
982 while (logfiles != NULL) {
983 char *filename = g_strdup_printf("%s/%s", log->log_directory,
984 (char *)logfiles->data);
985 g_print("Scanning file %s\n", filename);
986 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
988 g_warning("Mapping logfile %s failed!\n", filename);
990 bluesky_replay_scan_journal(g_mapped_file_get_contents(map),
991 g_mapped_file_get_length(map),
992 &seq_num, &start_offset);
993 g_mapped_file_unref(map);
997 g_free(logfiles->data);
998 logfiles = g_list_delete_link(logfiles, logfiles);
999 if (seq_num != 0 || start_offset != 0)
1002 g_list_foreach(logfiles, (GFunc)g_free, NULL);
1003 g_list_free(logfiles);
1005 /* Now, scan forward starting from the given point in the log to
1006 * reconstruct all filesystem state. As we reload objects we hold a
1007 * reference to each loaded object. At the end we free all these
1008 * references, so that any objects which were not linked into persistent
1009 * filesystem data structures are freed. */
1010 GList *objects = NULL;
1012 char *filename = g_strdup_printf("%s/journal-%08d",
1013 log->log_directory, seq_num);
1014 g_print("Replaying file %s from offset %d\n", filename, start_offset);
1015 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
1018 g_warning("Mapping logfile failed, assuming end of journal\n");
1022 bluesky_replay_scan_journal2(fs, &objects, seq_num, start_offset,
1023 g_mapped_file_get_contents(map),
1024 g_mapped_file_get_length(map));
1025 g_mapped_file_unref(map);
1030 while (objects != NULL) {
1031 bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
1032 objects = g_list_delete_link(objects, objects);