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 << 24)
41 #define HEADER_MAGIC 0x676f4c0a
42 #define FOOTER_MAGIC 0x2e435243
45 uint32_t magic; // HEADER_MAGIC
46 uint64_t offset; // Starting byte offset of the log header
47 uint32_t size; // Size of the data item (bytes)
48 BlueSkyCloudID id; // Object identifier
49 } __attribute__((packed));
52 uint32_t magic; // FOOTER_MAGIC
53 uint32_t crc; // Computed from log_header to log_footer.magic
54 } __attribute__((packed));
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)
77 while (log->committed != NULL) {
78 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
79 g_mutex_lock(item->lock);
80 bluesky_cloudlog_stats_update(item, -1);
81 item->pending_write &= ~CLOUDLOG_JOURNAL;
82 item->location_flags |= CLOUDLOG_JOURNAL;
83 bluesky_cloudlog_stats_update(item, 1);
84 g_cond_signal(item->cond);
85 g_mutex_unlock(item->lock);
86 log->committed = g_slist_delete_link(log->committed, log->committed);
87 bluesky_cloudlog_unref(item);
91 if (bluesky_verbose && batchsize > 1)
92 g_print("Log batch size: %d\n", batchsize);
95 static gboolean log_open(BlueSkyLog *log)
106 if (log->current_log != NULL) {
107 bluesky_cachefile_unref(log->current_log);
108 log->current_log = NULL;
111 while (log->fd < 0) {
112 g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
113 log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
114 if (log->fd < 0 && errno == EEXIST) {
115 fprintf(stderr, "Log file %s already exists...\n", logname);
118 } else if (log->fd < 0) {
119 fprintf(stderr, "Error opening logfile %s: %m\n", logname);
124 log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
125 g_assert(log->current_log != NULL);
126 g_mutex_unlock(log->current_log->lock);
128 if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
129 fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
136 /* All log writes (at least for a single log) are made by one thread, so we
137 * don't need to worry about concurrent access to the log file. Log items to
138 * write are pulled off a queue (and so may be posted by any thread).
139 * fdatasync() is used to ensure the log items are stable on disk.
141 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
142 * each. If a log segment is not currently open (log->fd is negative), a new
143 * one is created. Log segment filenames are assigned sequentially.
145 * Log replay ought to be implemented later, and ought to set the initial
146 * sequence number appropriately.
148 static gpointer log_thread(gpointer d)
150 BlueSkyLog *log = (BlueSkyLog *)d;
154 if (!log_open(log)) {
159 BlueSkyCloudLog *item
160 = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
161 g_mutex_lock(item->lock);
162 g_assert(item->data != NULL);
164 /* The item may have already been written to the journal... */
165 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
166 g_mutex_unlock(item->lock);
167 bluesky_cloudlog_unref(item);
168 g_atomic_int_add(&item->data_lock_count, -1);
172 bluesky_cloudlog_stats_update(item, -1);
173 item->pending_write |= CLOUDLOG_JOURNAL;
174 bluesky_cloudlog_stats_update(item, 1);
176 struct log_header header;
177 struct log_footer footer;
178 size_t size = sizeof(header) + sizeof(footer) + item->data->len;
181 offset = lseek(log->fd, 0, SEEK_CUR);
183 /* Check whether the item would overflow the allocated journal size.
184 * If so, start a new log segment. We only allow oversized log
185 * segments if they contain a single log entry. */
186 if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
191 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
192 header.offset = GUINT64_TO_LE(offset);
193 header.size = GUINT32_TO_LE(item->data->len);
194 header.id = item->id;
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, item->data->data, item->data->len);
203 crc = crc32c(crc, item->data->data, item->data->len);
205 crc = crc32c(crc, (const char *)&footer,
206 sizeof(footer) - sizeof(uint32_t));
207 footer.crc = crc32c_finalize(crc);
208 writebuf(log->fd, (const char *)&footer, sizeof(footer));
210 item->log_seq = log->seq_num;
211 item->log_offset = offset + sizeof(header);
212 item->log_size = item->data->len;
214 offset += sizeof(header) + sizeof(footer) + item->data->len;
216 /* Since we have just written a new dirty object to the journal,
217 * increment the count of live dirty objects in that journal file. The
218 * count will be decremented when objects are deleted or written to the
220 g_atomic_int_add(&log->current_log->dirty_refs, 1);
221 item->dirty_journal = log->current_log;
223 /* Replace the log item's string data with a memory-mapped copy of the
224 * data, now that it has been written to the log file. (Even if it
225 * isn't yet on disk, it should at least be in the page cache and so
226 * available to memory map.) */
227 bluesky_string_unref(item->data);
229 bluesky_cloudlog_fetch(item);
231 log->committed = g_slist_prepend(log->committed, item);
232 g_atomic_int_add(&item->data_lock_count, -1);
233 g_mutex_unlock(item->lock);
235 /* Force an if there are no other log items currently waiting to be
237 if (g_async_queue_length(log->queue) <= 0)
244 BlueSkyLog *bluesky_log_new(const char *log_directory)
246 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
248 log->log_directory = g_strdup(log_directory);
251 log->queue = g_async_queue_new();
252 log->mmap_lock = g_mutex_new();
253 log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
255 log->dirfd = open(log->log_directory, O_DIRECTORY);
256 if (log->dirfd < 0) {
257 fprintf(stderr, "Unable to open logging directory: %m\n");
261 g_thread_create(log_thread, log, FALSE, NULL);
266 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
268 bluesky_cloudlog_ref(item);
269 g_atomic_int_add(&item->data_lock_count, 1);
270 g_async_queue_push(log->queue, item);
273 void bluesky_log_finish_all(GList *log_items)
275 while (log_items != NULL) {
276 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
278 g_mutex_lock(item->lock);
279 while ((item->pending_write & CLOUDLOG_JOURNAL))
280 g_cond_wait(item->cond, item->lock);
281 g_mutex_unlock(item->lock);
282 bluesky_cloudlog_unref(item);
284 log_items = g_list_delete_link(log_items, log_items);
288 /* Memory-map the given log object into memory (read-only) and return a pointer
290 static int page_size = 0;
292 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
294 g_atomic_int_add(&cachefile->refcount, -1);
297 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
298 BlueSkyCacheFile *cachefile)
300 g_print("Fetch of %s from cloud complete, status = %d\n",
301 async->key, async->result);
303 if (async->result < 0)
306 g_mutex_lock(cachefile->lock);
307 char *pathname = g_strdup_printf("%s/%s", cachefile->log->log_directory,
308 cachefile->filename);
309 if (!g_file_set_contents(pathname, async->data->data, async->data->len,
312 g_print("Error writing out fetched file to cache!\n");
315 cachefile->fetching = FALSE;
316 cachefile->ready = TRUE;
317 bluesky_cachefile_unref(cachefile);
318 g_cond_broadcast(cachefile->cond);
319 g_mutex_unlock(cachefile->lock);
322 /* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
323 * Returns the object in the locked state and with a reference taken. */
324 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
325 int clouddir, int log_seq)
327 if (page_size == 0) {
328 page_size = getpagesize();
331 BlueSkyLog *log = fs->log;
336 // A request for a local log file
338 sprintf(logname, "journal-%08d", log_seq);
339 type = CLOUDLOG_JOURNAL;
341 sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
342 type = CLOUDLOG_CLOUD;
345 BlueSkyCacheFile *map;
346 g_mutex_lock(log->mmap_lock);
347 map = g_hash_table_lookup(log->mmap_cache, logname);
350 /* TODO: stat() call */
351 map = g_new0(BlueSkyCacheFile, 1);
352 map->type = CLOUDLOG_JOURNAL;
353 map->lock = g_mutex_new();
355 g_mutex_lock(map->lock);
356 map->cond = g_cond_new();
357 map->filename = g_strdup(logname);
358 map->log_seq = log_seq;
360 g_atomic_int_set(&map->mapcount, 0);
361 g_atomic_int_set(&map->refcount, 0);
363 g_hash_table_insert(log->mmap_cache, map->filename, map);
365 // If the log file is stored in the cloud, we may need to fetch it
367 g_atomic_int_inc(&map->refcount);
368 map->fetching = TRUE;
369 g_print("Starting fetch of %s from cloud\n", logname);
370 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
371 async->op = STORE_OP_GET;
372 async->key = g_strdup(logname);
373 bluesky_store_async_add_notifier(async,
374 (GFunc)cloudlog_fetch_complete,
376 bluesky_store_async_submit(async);
377 bluesky_store_async_unref(async);
380 g_mutex_lock(map->lock);
383 g_mutex_unlock(log->mmap_lock);
384 g_atomic_int_inc(&map->refcount);
388 BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir,
389 int log_seq, int log_offset, int log_size)
391 if (page_size == 0) {
392 page_size = getpagesize();
395 BlueSkyLog *log = fs->log;
396 BlueSkyCacheFile *map = bluesky_cachefile_lookup(fs, log_dir, log_seq);
402 if (map->addr == NULL) {
403 while (!map->ready && map->fetching) {
404 g_print("Waiting for log segment to be fetched from cloud...\n");
405 g_cond_wait(map->cond, map->lock);
408 int fd = openat(log->dirfd, map->filename, O_RDONLY);
411 fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
412 bluesky_cachefile_unref(map);
413 g_mutex_unlock(map->lock);
417 off_t length = lseek(fd, 0, SEEK_END);
418 map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
420 g_atomic_int_add(&log->disk_used, -(map->len / 1024));
422 g_atomic_int_add(&log->disk_used, map->len / 1024);
424 g_print("Re-mapped log segment %d...\n", log_seq);
425 g_atomic_int_inc(&map->refcount);
430 g_mutex_unlock(log->mmap_lock);
433 map->atime = bluesky_get_current_time();
434 str = bluesky_string_new_from_mmap(map, log_offset, log_size);
435 bluesky_cachefile_unref(map);
436 g_mutex_unlock(map->lock);
440 void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
445 if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
446 g_mutex_lock(mmap->lock);
447 if (g_atomic_int_get(&mmap->mapcount) == 0) {
448 g_print("Unmapped log segment %d...\n", mmap->log_seq);
449 munmap((void *)mmap->addr, mmap->len);
451 g_atomic_int_add(&mmap->refcount, -1);
453 g_mutex_unlock(mmap->lock);
457 /* Scan through all currently-stored files in the journal/cache and garbage
458 * collect old unused ones, if needed. */
459 static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
461 GList **files = (GList **)user_data;
462 *files = g_list_prepend(*files, value);
465 static gint compare_cachefiles(gconstpointer a, gconstpointer b)
469 ta = ((BlueSkyCacheFile *)a)->atime;
470 tb = ((BlueSkyCacheFile *)b)->atime;
479 void bluesky_cachefile_gc(BlueSkyFS *fs)
483 g_mutex_lock(fs->log->mmap_lock);
484 g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
486 /* Sort based on atime. The atime should be stable since it shouln't be
487 * updated except by threads which can grab the mmap_lock, which we already
489 files = g_list_sort(files, compare_cachefiles);
491 /* Walk the list of files, starting with the oldest, deleting files if
492 * possible until enough space has been reclaimed. */
493 g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
494 while (files != NULL) {
495 BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
496 /* Try to lock the structure, but if the lock is held by another thread
497 * then we'll just skip the file on this pass. */
498 if (g_mutex_trylock(cachefile->lock)) {
499 int64_t age = bluesky_get_current_time() - cachefile->atime;
500 g_print("%s addr=%p mapcount=%d refcount=%d dirty=%d atime_age=%f",
501 cachefile->filename, cachefile->addr, cachefile->mapcount,
502 cachefile->refcount, cachefile->dirty_refs, age / 1e6);
503 if (cachefile->fetching)
504 g_print(" (fetching)");
507 if (g_atomic_int_get(&cachefile->refcount) == 0
508 && g_atomic_int_get(&cachefile->mapcount) == 0
509 && g_atomic_int_get(&cachefile->dirty_refs) == 0)
511 g_print(" ...deleting\n");
512 if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
513 fprintf(stderr, "Unable to unlink journal %s: %m\n",
514 cachefile->filename);
517 g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
518 g_mutex_unlock(cachefile->lock);
519 g_mutex_free(cachefile->lock);
520 g_cond_free(cachefile->cond);
521 g_free(cachefile->filename);
524 g_mutex_unlock(cachefile->lock);
527 files = g_list_delete_link(files, files);
531 g_mutex_unlock(fs->log->mmap_lock);