1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2010 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
10 #define _ATFILE_SOURCE
18 #include <sys/types.h>
24 #include "bluesky-private.h"
26 /* The logging layer for BlueSky. This is used to write filesystem changes
27 * durably to disk so that they can be recovered in the event of a system
30 /* The logging layer takes care out writing out a sequence of log records to
31 * disk. On disk, each record consists of a header, a data payload, and a
32 * footer. The footer contains a checksum of the record, meant to help with
33 * identifying corrupt log records (we would assume because the log record was
34 * only incompletely written out before a crash, which should only happen for
35 * log records that were not considered committed). */
37 // Rough size limit for a log segment. This is not a firm limit and there are
38 // no absolute guarantees on the size of a log segment.
39 #define LOG_SEGMENT_SIZE (1 << 22)
41 #define HEADER_MAGIC 0x676f4c0a
42 #define FOOTER_MAGIC 0x2e435243
45 uint32_t magic; // HEADER_MAGIC
46 uint8_t type; // Object type + '0'
47 uint32_t offset; // Starting byte offset of the log header
48 uint32_t size; // Size of the data item (bytes)
49 uint64_t inum; // Inode which owns this data, if any
50 BlueSkyCloudID id; // Object identifier
51 } __attribute__((packed));
54 uint32_t magic; // FOOTER_MAGIC
55 uint32_t crc; // Computed from log_header to log_footer.magic
56 } __attribute__((packed));
58 static void writebuf(int fd, const char *buf, size_t len)
62 written = write(fd, buf, len);
63 if (written < 0 && errno == EINTR)
65 g_assert(written >= 0);
71 static void log_commit(BlueSkyLog *log)
79 while (log->committed != NULL) {
80 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
81 g_mutex_lock(item->lock);
82 bluesky_cloudlog_stats_update(item, -1);
83 item->pending_write &= ~CLOUDLOG_JOURNAL;
84 item->location_flags |= CLOUDLOG_JOURNAL;
85 bluesky_cloudlog_stats_update(item, 1);
86 g_cond_signal(item->cond);
87 g_mutex_unlock(item->lock);
88 log->committed = g_slist_delete_link(log->committed, log->committed);
89 bluesky_cloudlog_unref(item);
93 if (bluesky_verbose && batchsize > 1)
94 g_print("Log batch size: %d\n", batchsize);
97 static gboolean log_open(BlueSkyLog *log)
108 if (log->current_log != NULL) {
109 bluesky_cachefile_unref(log->current_log);
110 log->current_log = NULL;
113 while (log->fd < 0) {
114 g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
115 log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
116 if (log->fd < 0 && errno == EEXIST) {
117 fprintf(stderr, "Log file %s already exists...\n", logname);
120 } else if (log->fd < 0) {
121 fprintf(stderr, "Error opening logfile %s: %m\n", logname);
126 log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
127 g_assert(log->current_log != NULL);
128 g_mutex_unlock(log->current_log->lock);
130 if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
131 fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
138 /* All log writes (at least for a single log) are made by one thread, so we
139 * don't need to worry about concurrent access to the log file. Log items to
140 * write are pulled off a queue (and so may be posted by any thread).
141 * fdatasync() is used to ensure the log items are stable on disk.
143 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
144 * each. If a log segment is not currently open (log->fd is negative), a new
145 * one is created. Log segment filenames are assigned sequentially.
147 * Log replay ought to be implemented later, and ought to set the initial
148 * sequence number appropriately.
150 static gpointer log_thread(gpointer d)
152 BlueSkyLog *log = (BlueSkyLog *)d;
156 if (!log_open(log)) {
161 BlueSkyCloudLog *item
162 = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
163 g_mutex_lock(item->lock);
164 g_assert(item->data != NULL);
166 /* The item may have already been written to the journal... */
167 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
168 g_mutex_unlock(item->lock);
169 bluesky_cloudlog_unref(item);
170 g_atomic_int_add(&item->data_lock_count, -1);
174 bluesky_cloudlog_stats_update(item, -1);
175 item->pending_write |= CLOUDLOG_JOURNAL;
176 bluesky_cloudlog_stats_update(item, 1);
178 struct log_header header;
179 struct log_footer footer;
180 size_t size = sizeof(header) + sizeof(footer) + item->data->len;
183 offset = lseek(log->fd, 0, SEEK_CUR);
185 /* Check whether the item would overflow the allocated journal size.
186 * If so, start a new log segment. We only allow oversized log
187 * segments if they contain a single log entry. */
188 if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
193 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
194 header.offset = GUINT32_TO_LE(offset);
195 header.size = GUINT32_TO_LE(item->data->len);
196 header.type = item->type + '0';
197 header.id = item->id;
198 header.inum = GUINT64_TO_LE(item->inum);
199 footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
201 uint32_t crc = BLUESKY_CRC32C_SEED;
203 writebuf(log->fd, (const char *)&header, sizeof(header));
204 crc = crc32c(crc, (const char *)&header, sizeof(header));
206 writebuf(log->fd, item->data->data, item->data->len);
207 crc = crc32c(crc, item->data->data, item->data->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 + sizeof(header);
216 item->log_size = item->data->len;
218 offset += sizeof(header) + sizeof(footer) + item->data->len;
220 /* Replace the log item's string data with a memory-mapped copy of the
221 * data, now that it has been written to the log file. (Even if it
222 * isn't yet on disk, it should at least be in the page cache and so
223 * available to memory map.) */
224 bluesky_string_unref(item->data);
226 bluesky_cloudlog_fetch(item);
228 log->committed = g_slist_prepend(log->committed, item);
229 g_atomic_int_add(&item->data_lock_count, -1);
230 g_mutex_unlock(item->lock);
232 /* Force an if there are no other log items currently waiting to be
234 if (g_async_queue_length(log->queue) <= 0)
241 BlueSkyLog *bluesky_log_new(const char *log_directory)
243 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
245 log->log_directory = g_strdup(log_directory);
248 log->queue = g_async_queue_new();
249 log->mmap_lock = g_mutex_new();
250 log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
252 log->dirfd = open(log->log_directory, O_DIRECTORY);
253 if (log->dirfd < 0) {
254 fprintf(stderr, "Unable to open logging directory: %m\n");
258 g_thread_create(log_thread, log, FALSE, NULL);
263 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
265 bluesky_cloudlog_ref(item);
266 g_atomic_int_add(&item->data_lock_count, 1);
267 g_async_queue_push(log->queue, item);
270 void bluesky_log_finish_all(GList *log_items)
272 while (log_items != NULL) {
273 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
275 g_mutex_lock(item->lock);
276 while ((item->pending_write & CLOUDLOG_JOURNAL))
277 g_cond_wait(item->cond, item->lock);
278 g_mutex_unlock(item->lock);
279 bluesky_cloudlog_unref(item);
281 log_items = g_list_delete_link(log_items, log_items);
285 /* Memory-map the given log object into memory (read-only) and return a pointer
287 static int page_size = 0;
289 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
291 g_atomic_int_add(&cachefile->refcount, -1);
294 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
295 BlueSkyCacheFile *cachefile);
297 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
299 g_atomic_int_inc(&cachefile->refcount);
300 cachefile->fetching = TRUE;
301 g_print("Starting fetch of %s from cloud\n", cachefile->filename);
302 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
303 async->op = STORE_OP_GET;
304 async->key = g_strdup(cachefile->filename);
305 bluesky_store_async_add_notifier(async,
306 (GFunc)cloudlog_fetch_complete,
308 bluesky_store_async_submit(async);
309 bluesky_store_async_unref(async);
312 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
313 BlueSkyCacheFile *cachefile)
315 g_print("Fetch of %s from cloud complete, status = %d\n",
316 async->key, async->result);
318 g_mutex_lock(cachefile->lock);
319 if (async->result >= 0) {
320 char *pathname = g_strdup_printf("%s/%s",
321 cachefile->log->log_directory,
322 cachefile->filename);
323 if (!g_file_set_contents(pathname, async->data->data, async->data->len,
325 g_print("Error writing out fetched file to cache!\n");
328 cachefile->fetching = FALSE;
329 cachefile->ready = TRUE;
331 g_print("Error fetching from cloud, retrying...\n");
332 cloudlog_fetch_start(cachefile);
335 bluesky_cachefile_unref(cachefile);
336 g_cond_broadcast(cachefile->cond);
337 g_mutex_unlock(cachefile->lock);
340 /* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
341 * Returns the object in the locked state and with a reference taken. */
342 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
343 int clouddir, int log_seq)
345 if (page_size == 0) {
346 page_size = getpagesize();
349 BlueSkyLog *log = fs->log;
355 // A request for a local log file
357 sprintf(logname, "journal-%08d", log_seq);
358 type = CLOUDLOG_JOURNAL;
360 sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
361 type = CLOUDLOG_CLOUD;
364 BlueSkyCacheFile *map;
365 g_mutex_lock(log->mmap_lock);
366 map = g_hash_table_lookup(log->mmap_cache, logname);
369 && type == CLOUDLOG_JOURNAL
370 && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
371 /* A stale reference to a journal file which doesn't exist any longer
372 * because it was reclaimed. Return NULL. */
373 } else if (map == NULL) {
374 g_print("Adding cache file %s\n", logname);
376 map = g_new0(BlueSkyCacheFile, 1);
379 map->lock = g_mutex_new();
381 g_mutex_lock(map->lock);
382 map->cond = g_cond_new();
383 map->filename = g_strdup(logname);
384 map->log_seq = log_seq;
386 g_atomic_int_set(&map->mapcount, 0);
387 g_atomic_int_set(&map->refcount, 0);
389 g_hash_table_insert(log->mmap_cache, map->filename, map);
391 // If the log file is stored in the cloud, we may need to fetch it
393 cloudlog_fetch_start(map);
395 g_mutex_lock(map->lock);
398 g_mutex_unlock(log->mmap_lock);
400 g_atomic_int_inc(&map->refcount);
404 BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir,
405 int log_seq, int log_offset, int log_size)
407 if (page_size == 0) {
408 page_size = getpagesize();
411 BlueSkyLog *log = fs->log;
412 BlueSkyCacheFile *map = bluesky_cachefile_lookup(fs, log_dir, log_seq);
418 if (map->addr == NULL) {
419 while (!map->ready && map->fetching) {
420 g_print("Waiting for log segment to be fetched from cloud...\n");
421 g_cond_wait(map->cond, map->lock);
424 int fd = openat(log->dirfd, map->filename, O_RDONLY);
427 fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
428 bluesky_cachefile_unref(map);
429 g_mutex_unlock(map->lock);
433 off_t length = lseek(fd, 0, SEEK_END);
434 map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
436 g_atomic_int_add(&log->disk_used, -(map->len / 1024));
438 g_atomic_int_add(&log->disk_used, map->len / 1024);
440 g_print("Re-mapped log segment %d...\n", log_seq);
441 g_atomic_int_inc(&map->refcount);
446 g_mutex_unlock(log->mmap_lock);
449 map->atime = bluesky_get_current_time();
450 str = bluesky_string_new_from_mmap(map, log_offset, log_size);
451 bluesky_cachefile_unref(map);
452 g_mutex_unlock(map->lock);
456 void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
461 if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
462 g_mutex_lock(mmap->lock);
463 if (g_atomic_int_get(&mmap->mapcount) == 0) {
464 g_print("Unmapped log segment %d...\n", mmap->log_seq);
465 munmap((void *)mmap->addr, mmap->len);
467 g_atomic_int_add(&mmap->refcount, -1);
469 g_mutex_unlock(mmap->lock);
473 /* Scan through all currently-stored files in the journal/cache and garbage
474 * collect old unused ones, if needed. */
475 static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
477 GList **files = (GList **)user_data;
478 *files = g_list_prepend(*files, value);
481 static gint compare_cachefiles(gconstpointer a, gconstpointer b)
485 ta = ((BlueSkyCacheFile *)a)->atime;
486 tb = ((BlueSkyCacheFile *)b)->atime;
495 void bluesky_cachefile_gc(BlueSkyFS *fs)
499 g_mutex_lock(fs->log->mmap_lock);
500 g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
502 /* Sort based on atime. The atime should be stable since it shouln't be
503 * updated except by threads which can grab the mmap_lock, which we already
505 files = g_list_sort(files, compare_cachefiles);
507 /* Walk the list of files, starting with the oldest, deleting files if
508 * possible until enough space has been reclaimed. */
509 g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
510 while (files != NULL) {
511 BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
512 /* Try to lock the structure, but if the lock is held by another thread
513 * then we'll just skip the file on this pass. */
514 if (g_mutex_trylock(cachefile->lock)) {
515 int64_t age = bluesky_get_current_time() - cachefile->atime;
516 g_print("%s addr=%p mapcount=%d refcount=%d atime_age=%f",
517 cachefile->filename, cachefile->addr, cachefile->mapcount,
518 cachefile->refcount, age / 1e6);
519 if (cachefile->fetching)
520 g_print(" (fetching)");
523 gboolean deletion_candidate = FALSE;
524 if (g_atomic_int_get(&fs->log->disk_used)
525 > bluesky_options.cache_size
526 && g_atomic_int_get(&cachefile->refcount) == 0
527 && g_atomic_int_get(&cachefile->mapcount) == 0)
529 deletion_candidate = TRUE;
532 /* Don't allow journal files to be reclaimed until all data is
533 * known to be durably stored in the cloud. */
534 if (cachefile->type == CLOUDLOG_JOURNAL
535 && cachefile->log_seq >= fs->log->journal_watermark)
537 deletion_candidate = FALSE;
540 if (deletion_candidate) {
541 g_print(" ...deleting\n");
542 if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
543 fprintf(stderr, "Unable to unlink journal %s: %m\n",
544 cachefile->filename);
547 g_atomic_int_add(&fs->log->disk_used, -(cachefile->len / 1024));
548 g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
549 g_mutex_unlock(cachefile->lock);
550 g_mutex_free(cachefile->lock);
551 g_cond_free(cachefile->cond);
552 g_free(cachefile->filename);
555 g_mutex_unlock(cachefile->lock);
558 files = g_list_delete_link(files, files);
562 g_mutex_unlock(fs->log->mmap_lock);
565 /******************************* JOURNAL REPLAY *******************************
566 * The journal replay code is used to recover filesystem state after a
567 * filesystem restart. We first look for the most recent commit record in the
568 * journal, which indicates the point before which all data in the journal has
569 * also been committed to the cloud. Then, we read in all data in the log past
572 static GList *directory_contents(const char *dirname)
574 GList *contents = NULL;
575 GDir *dir = g_dir_open(dirname, 0, NULL);
577 g_warning("Unable to open journal directory: %s", dirname);
582 while ((file = g_dir_read_name(dir)) != NULL) {
583 if (strncmp(file, "journal-", 8) == 0)
584 contents = g_list_prepend(contents, g_strdup(file));
588 contents = g_list_sort(contents, (GCompareFunc)strcmp);
593 static gboolean validate_journal_item(const char *buf, size_t len, off_t offset)
595 const struct log_header *header;
596 const struct log_footer *footer;
598 if (offset + sizeof(struct log_header) + sizeof(struct log_footer) > len)
601 header = (const struct log_header *)(buf + offset);
602 if (GUINT32_FROM_LE(header->magic) != HEADER_MAGIC)
604 if (GUINT32_FROM_LE(header->offset) != offset)
606 size_t size = GUINT32_FROM_LE(header->size);
608 off_t footer_offset = offset + sizeof(struct log_header) + size;
609 if (footer_offset + sizeof(struct log_footer) > len)
611 footer = (const struct log_footer *)(buf + footer_offset);
613 if (GUINT32_FROM_LE(footer->magic) != FOOTER_MAGIC)
616 uint32_t crc = crc32c(BLUESKY_CRC32C_SEED, buf + offset,
617 sizeof(struct log_header) + sizeof(struct log_footer)
619 if (crc != BLUESKY_CRC32C_VALIDATOR) {
620 g_warning("Journal entry failed to validate: CRC %08x != %08x",
621 crc, BLUESKY_CRC32C_VALIDATOR);
628 /* Scan through a journal segment to extract correctly-written items (those
629 * that pass sanity checks and have a valid checksum). */
630 static void bluesky_replay_scan_journal(const char *buf, size_t len)
632 const struct log_header *header;
635 while (validate_journal_item(buf, len, offset)) {
636 header = (const struct log_header *)(buf + offset);
637 size_t size = GUINT32_FROM_LE(header->size);
638 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
642 static void bluesky_replay_scan_journal2(BlueSkyFS *fs, GList **objects,
644 const char *buf, size_t len)
646 const struct log_header *header;
649 while (validate_journal_item(buf, len, offset)) {
650 header = (const struct log_header *)(buf + offset);
651 g_print("In replay found valid item at offset %zd\n", offset);
652 size_t size = GUINT32_FROM_LE(header->size);
654 g_mutex_lock(fs->lock);
655 BlueSkyCloudLog *log_item;
656 log_item = g_hash_table_lookup(fs->locations, &header->id);
657 if (log_item == NULL) {
658 log_item = bluesky_cloudlog_new(fs, &header->id);
659 g_hash_table_insert(fs->locations, &log_item->id, log_item);
660 g_mutex_lock(log_item->lock);
662 bluesky_cloudlog_ref(log_item);
663 g_mutex_lock(log_item->lock);
665 g_mutex_unlock(fs->lock);
666 *objects = g_list_prepend(*objects, log_item);
668 bluesky_string_unref(log_item->data);
669 log_item->location_flags = CLOUDLOG_JOURNAL;
670 log_item->data = NULL;
671 log_item->log_seq = log_seq;
672 log_item->log_offset = offset + sizeof(struct log_header);
673 log_item->log_size = header->size;
674 g_mutex_unlock(log_item->lock);
676 offset += sizeof(struct log_header) + size + sizeof(struct log_footer);
680 void bluesky_replay(BlueSkyFS *fs)
682 BlueSkyLog *log = fs->log;
683 GList *logfiles = directory_contents(log->log_directory);
685 /* Scan through log files in reverse order to find the most recent commit
687 logfiles = g_list_reverse(logfiles);
688 while (logfiles != NULL) {
689 char *filename = g_strdup_printf("%s/%s", log->log_directory,
690 (char *)logfiles->data);
691 g_print("Scanning file %s\n", filename);
692 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
694 g_warning("Mapping logfile %s failed!\n", filename);
696 bluesky_replay_scan_journal(g_mapped_file_get_contents(map),
697 g_mapped_file_get_length(map));
698 g_mapped_file_unref(map);
702 g_free(logfiles->data);
703 logfiles = g_list_delete_link(logfiles, logfiles);
705 g_list_foreach(logfiles, (GFunc)g_free, NULL);
706 g_list_free(logfiles);
708 /* Now, scan forward starting from the given point in the log to
709 * reconstruct all filesystem state. As we reload objects we hold a
710 * reference to each loaded object. At the end we free all these
711 * references, so that any objects which were not linked into persistent
712 * filesystem data structures are freed. */
713 GList *objects = NULL;
716 char *filename = g_strdup_printf("%s/journal-%08d",
717 log->log_directory, seq_num);
718 g_print("Replaying file %s\n", filename);
719 GMappedFile *map = g_mapped_file_new(filename, FALSE, NULL);
722 g_warning("Mapping logfile failed, assuming end of journal\n");
726 bluesky_replay_scan_journal2(fs, &objects, seq_num,
727 g_mapped_file_get_contents(map),
728 g_mapped_file_get_length(map));
729 g_mapped_file_unref(map);
733 while (objects != NULL) {
734 bluesky_cloudlog_unref((BlueSkyCloudLog *)objects->data);
735 objects = g_list_delete_link(objects, objects);