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 // Target amount of disk space to use for the journal and cache files, in
43 #define DISK_CACHE_SIZE_TARGET (64 * 1024)
45 #define HEADER_MAGIC 0x676f4c0a
46 #define FOOTER_MAGIC 0x2e435243
49 uint32_t magic; // HEADER_MAGIC
50 uint64_t offset; // Starting byte offset of the log header
51 uint32_t size; // Size of the data item (bytes)
52 BlueSkyCloudID id; // Object identifier
53 } __attribute__((packed));
56 uint32_t magic; // FOOTER_MAGIC
57 uint32_t crc; // Computed from log_header to log_footer.magic
58 } __attribute__((packed));
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)
81 while (log->committed != NULL) {
82 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
83 g_mutex_lock(item->lock);
84 bluesky_cloudlog_stats_update(item, -1);
85 item->pending_write &= ~CLOUDLOG_JOURNAL;
86 item->location_flags |= CLOUDLOG_JOURNAL;
87 bluesky_cloudlog_stats_update(item, 1);
88 g_cond_signal(item->cond);
89 g_mutex_unlock(item->lock);
90 log->committed = g_slist_delete_link(log->committed, log->committed);
91 bluesky_cloudlog_unref(item);
95 if (bluesky_verbose && batchsize > 1)
96 g_print("Log batch size: %d\n", batchsize);
99 static gboolean log_open(BlueSkyLog *log)
110 if (log->current_log != NULL) {
111 bluesky_cachefile_unref(log->current_log);
112 log->current_log = NULL;
115 while (log->fd < 0) {
116 g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
117 log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
118 if (log->fd < 0 && errno == EEXIST) {
119 fprintf(stderr, "Log file %s already exists...\n", logname);
122 } else if (log->fd < 0) {
123 fprintf(stderr, "Error opening logfile %s: %m\n", logname);
128 log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
129 g_assert(log->current_log != NULL);
130 g_mutex_unlock(log->current_log->lock);
132 if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
133 fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
140 /* All log writes (at least for a single log) are made by one thread, so we
141 * don't need to worry about concurrent access to the log file. Log items to
142 * write are pulled off a queue (and so may be posted by any thread).
143 * fdatasync() is used to ensure the log items are stable on disk.
145 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
146 * each. If a log segment is not currently open (log->fd is negative), a new
147 * one is created. Log segment filenames are assigned sequentially.
149 * Log replay ought to be implemented later, and ought to set the initial
150 * sequence number appropriately.
152 static gpointer log_thread(gpointer d)
154 BlueSkyLog *log = (BlueSkyLog *)d;
158 if (!log_open(log)) {
163 BlueSkyCloudLog *item
164 = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
165 g_mutex_lock(item->lock);
166 g_assert(item->data != NULL);
168 /* The item may have already been written to the journal... */
169 if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
170 g_mutex_unlock(item->lock);
171 bluesky_cloudlog_unref(item);
172 g_atomic_int_add(&item->data_lock_count, -1);
176 bluesky_cloudlog_stats_update(item, -1);
177 item->pending_write |= CLOUDLOG_JOURNAL;
178 bluesky_cloudlog_stats_update(item, 1);
180 struct log_header header;
181 struct log_footer footer;
182 size_t size = sizeof(header) + sizeof(footer) + item->data->len;
185 offset = lseek(log->fd, 0, SEEK_CUR);
187 /* Check whether the item would overflow the allocated journal size.
188 * If so, start a new log segment. We only allow oversized log
189 * segments if they contain a single log entry. */
190 if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
195 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
196 header.offset = GUINT64_TO_LE(offset);
197 header.size = GUINT32_TO_LE(item->data->len);
198 header.id = item->id;
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 /* Since we have just written a new dirty object to the journal,
221 * increment the count of live dirty objects in that journal file. The
222 * count will be decremented when objects are deleted or written to the
224 if (!(item->location_flags & CLOUDLOG_CLOUD)) {
225 g_atomic_int_add(&log->current_log->dirty_refs, 1);
226 item->dirty_journal = log->current_log;
229 /* Replace the log item's string data with a memory-mapped copy of the
230 * data, now that it has been written to the log file. (Even if it
231 * isn't yet on disk, it should at least be in the page cache and so
232 * available to memory map.) */
233 bluesky_string_unref(item->data);
235 bluesky_cloudlog_fetch(item);
237 log->committed = g_slist_prepend(log->committed, item);
238 g_atomic_int_add(&item->data_lock_count, -1);
239 g_mutex_unlock(item->lock);
241 /* Force an if there are no other log items currently waiting to be
243 if (g_async_queue_length(log->queue) <= 0)
250 BlueSkyLog *bluesky_log_new(const char *log_directory)
252 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
254 log->log_directory = g_strdup(log_directory);
257 log->queue = g_async_queue_new();
258 log->mmap_lock = g_mutex_new();
259 log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
261 log->dirfd = open(log->log_directory, O_DIRECTORY);
262 if (log->dirfd < 0) {
263 fprintf(stderr, "Unable to open logging directory: %m\n");
267 g_thread_create(log_thread, log, FALSE, NULL);
272 void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
274 bluesky_cloudlog_ref(item);
275 g_atomic_int_add(&item->data_lock_count, 1);
276 g_async_queue_push(log->queue, item);
279 void bluesky_log_finish_all(GList *log_items)
281 while (log_items != NULL) {
282 BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
284 g_mutex_lock(item->lock);
285 while ((item->pending_write & CLOUDLOG_JOURNAL))
286 g_cond_wait(item->cond, item->lock);
287 g_mutex_unlock(item->lock);
288 bluesky_cloudlog_unref(item);
290 log_items = g_list_delete_link(log_items, log_items);
294 /* Memory-map the given log object into memory (read-only) and return a pointer
296 static int page_size = 0;
298 void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
300 g_atomic_int_add(&cachefile->refcount, -1);
303 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
304 BlueSkyCacheFile *cachefile);
306 static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
308 g_atomic_int_inc(&cachefile->refcount);
309 cachefile->fetching = TRUE;
310 g_print("Starting fetch of %s from cloud\n", cachefile->filename);
311 BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
312 async->op = STORE_OP_GET;
313 async->key = g_strdup(cachefile->filename);
314 bluesky_store_async_add_notifier(async,
315 (GFunc)cloudlog_fetch_complete,
317 bluesky_store_async_submit(async);
318 bluesky_store_async_unref(async);
321 static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
322 BlueSkyCacheFile *cachefile)
324 g_print("Fetch of %s from cloud complete, status = %d\n",
325 async->key, async->result);
327 g_mutex_lock(cachefile->lock);
328 if (async->result >= 0) {
329 char *pathname = g_strdup_printf("%s/%s",
330 cachefile->log->log_directory,
331 cachefile->filename);
332 if (!g_file_set_contents(pathname, async->data->data, async->data->len,
334 g_print("Error writing out fetched file to cache!\n");
337 cachefile->fetching = FALSE;
338 cachefile->ready = TRUE;
340 g_print("Error fetching from cloud, retrying...\n");
341 cloudlog_fetch_start(cachefile);
344 bluesky_cachefile_unref(cachefile);
345 g_cond_broadcast(cachefile->cond);
346 g_mutex_unlock(cachefile->lock);
349 /* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
350 * Returns the object in the locked state and with a reference taken. */
351 BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
352 int clouddir, int log_seq)
354 if (page_size == 0) {
355 page_size = getpagesize();
358 BlueSkyLog *log = fs->log;
364 // A request for a local log file
366 sprintf(logname, "journal-%08d", log_seq);
367 type = CLOUDLOG_JOURNAL;
369 sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
370 type = CLOUDLOG_CLOUD;
373 BlueSkyCacheFile *map;
374 g_mutex_lock(log->mmap_lock);
375 map = g_hash_table_lookup(log->mmap_cache, logname);
378 && type == CLOUDLOG_JOURNAL
379 && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
380 /* A stale reference to a journal file which doesn't exist any longer
381 * because it was reclaimed. Return NULL. */
382 } else if (map == NULL) {
383 g_print("Adding cache file %s\n", logname);
385 map = g_new0(BlueSkyCacheFile, 1);
388 map->lock = g_mutex_new();
390 g_mutex_lock(map->lock);
391 map->cond = g_cond_new();
392 map->filename = g_strdup(logname);
393 map->log_seq = log_seq;
395 g_atomic_int_set(&map->mapcount, 0);
396 g_atomic_int_set(&map->refcount, 0);
398 g_hash_table_insert(log->mmap_cache, map->filename, map);
400 // If the log file is stored in the cloud, we may need to fetch it
402 cloudlog_fetch_start(map);
404 g_mutex_lock(map->lock);
407 g_mutex_unlock(log->mmap_lock);
409 g_atomic_int_inc(&map->refcount);
413 BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir,
414 int log_seq, int log_offset, int log_size)
416 if (page_size == 0) {
417 page_size = getpagesize();
420 BlueSkyLog *log = fs->log;
421 BlueSkyCacheFile *map = bluesky_cachefile_lookup(fs, log_dir, log_seq);
427 if (map->addr == NULL) {
428 while (!map->ready && map->fetching) {
429 g_print("Waiting for log segment to be fetched from cloud...\n");
430 g_cond_wait(map->cond, map->lock);
433 int fd = openat(log->dirfd, map->filename, O_RDONLY);
436 fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
437 bluesky_cachefile_unref(map);
438 g_mutex_unlock(map->lock);
442 off_t length = lseek(fd, 0, SEEK_END);
443 map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
445 g_atomic_int_add(&log->disk_used, -(map->len / 1024));
447 g_atomic_int_add(&log->disk_used, map->len / 1024);
449 g_print("Re-mapped log segment %d...\n", log_seq);
450 g_atomic_int_inc(&map->refcount);
455 g_mutex_unlock(log->mmap_lock);
458 map->atime = bluesky_get_current_time();
459 str = bluesky_string_new_from_mmap(map, log_offset, log_size);
460 bluesky_cachefile_unref(map);
461 g_mutex_unlock(map->lock);
465 void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
470 if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
471 g_mutex_lock(mmap->lock);
472 if (g_atomic_int_get(&mmap->mapcount) == 0) {
473 g_print("Unmapped log segment %d...\n", mmap->log_seq);
474 munmap((void *)mmap->addr, mmap->len);
476 g_atomic_int_add(&mmap->refcount, -1);
478 g_mutex_unlock(mmap->lock);
482 /* Scan through all currently-stored files in the journal/cache and garbage
483 * collect old unused ones, if needed. */
484 static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
486 GList **files = (GList **)user_data;
487 *files = g_list_prepend(*files, value);
490 static gint compare_cachefiles(gconstpointer a, gconstpointer b)
494 ta = ((BlueSkyCacheFile *)a)->atime;
495 tb = ((BlueSkyCacheFile *)b)->atime;
504 void bluesky_cachefile_gc(BlueSkyFS *fs)
508 g_mutex_lock(fs->log->mmap_lock);
509 g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
511 /* Sort based on atime. The atime should be stable since it shouln't be
512 * updated except by threads which can grab the mmap_lock, which we already
514 files = g_list_sort(files, compare_cachefiles);
516 /* Walk the list of files, starting with the oldest, deleting files if
517 * possible until enough space has been reclaimed. */
518 g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
519 while (files != NULL) {
520 BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
521 /* Try to lock the structure, but if the lock is held by another thread
522 * then we'll just skip the file on this pass. */
523 if (g_mutex_trylock(cachefile->lock)) {
524 int64_t age = bluesky_get_current_time() - cachefile->atime;
525 g_print("%s addr=%p mapcount=%d refcount=%d dirty=%d atime_age=%f",
526 cachefile->filename, cachefile->addr, cachefile->mapcount,
527 cachefile->refcount, cachefile->dirty_refs, age / 1e6);
528 if (cachefile->fetching)
529 g_print(" (fetching)");
532 if (g_atomic_int_get(&fs->log->disk_used) > DISK_CACHE_SIZE_TARGET
533 && g_atomic_int_get(&cachefile->refcount) == 0
534 && g_atomic_int_get(&cachefile->mapcount) == 0
535 && g_atomic_int_get(&cachefile->dirty_refs) == 0)
537 g_print(" ...deleting\n");
538 if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
539 fprintf(stderr, "Unable to unlink journal %s: %m\n",
540 cachefile->filename);
543 g_atomic_int_add(&fs->log->disk_used, -(cachefile->len / 1024));
544 g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
545 g_mutex_unlock(cachefile->lock);
546 g_mutex_free(cachefile->lock);
547 g_cond_free(cachefile->cond);
548 g_free(cachefile->filename);
551 g_mutex_unlock(cachefile->lock);
554 files = g_list_delete_link(files, files);
558 g_mutex_unlock(fs->log->mmap_lock);