1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2009 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
14 #include "bluesky-private.h"
16 // Rough size limit for a log segment. This is not a firm limit and there are
17 // no absolute guarantees on the size of a log segment.
18 #define CLOUDLOG_SEGMENT_SIZE (4 << 20)
20 BlueSkyCloudID bluesky_cloudlog_new_id()
23 bluesky_crypt_random_bytes((uint8_t *)&id.bytes, sizeof(id));
27 gchar *bluesky_cloudlog_id_to_string(BlueSkyCloudID id)
29 char buf[sizeof(BlueSkyCloudID) * 2 + 1];
32 for (int i = 0; i < sizeof(BlueSkyCloudID); i++) {
33 sprintf(&buf[2*i], "%02x", (uint8_t)(id.bytes[i]));
39 BlueSkyCloudID bluesky_cloudlog_id_from_string(const gchar *idstr)
42 memset(&id, 0, sizeof(id));
43 for (int i = 0; i < 2*sizeof(BlueSkyCloudID); i++) {
46 g_warning("Short cloud id: %s\n", idstr);
50 if (c >= '0' && c <= '9')
52 else if (c >= 'a' && c <= 'f')
55 g_warning("Bad character in cloud id: %s\n", idstr);
56 id.bytes[i / 2] += val << (i % 2 ? 0 : 4);
61 gboolean bluesky_cloudlog_equal(gconstpointer a, gconstpointer b)
63 BlueSkyCloudID *id1 = (BlueSkyCloudID *)a, *id2 = (BlueSkyCloudID *)b;
65 return memcmp(id1, id2, sizeof(BlueSkyCloudID)) == 0;
68 guint bluesky_cloudlog_hash(gconstpointer a)
70 BlueSkyCloudID *id = (BlueSkyCloudID *)a;
72 // Assume that bits in the ID are randomly chosen so that any subset of the
73 // bits can be used as a hash key.
74 return *(guint *)(&id->bytes);
77 /* Formatting of cloud log segments. This handles grouping items together
78 * before writing a batch to the cloud, handling indirection through items like
79 * the inode map, etc. */
81 BlueSkyCloudLog *bluesky_cloudlog_new(BlueSkyFS *fs, const BlueSkyCloudID *id)
83 BlueSkyCloudLog *log = g_new0(BlueSkyCloudLog, 1);
85 log->lock = g_mutex_new();
86 log->cond = g_cond_new();
88 log->type = LOGTYPE_UNKNOWN;
90 memcpy(&log->id, id, sizeof(BlueSkyCloudID));
92 log->id = bluesky_cloudlog_new_id();
93 log->links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
94 g_atomic_int_set(&log->refcount, 1);
99 /* Helper function for updating memory usage statistics for a filesystem (the
100 * cache_log_* variables). This will increment (type=1) or decrement (type=-1)
101 * the counter associated with the current state of the cloud log item. The
102 * item should be locked or otherwise protected from concurrent access. */
103 void bluesky_cloudlog_stats_update(BlueSkyCloudLog *log, int type)
105 BlueSkyFS *fs = log->fs;
107 if (log->location_flags & CLOUDLOG_CLOUD) {
108 g_atomic_int_add(&fs->cache_log_cloud, type);
109 } else if (log->location_flags & CLOUDLOG_JOURNAL) {
110 g_atomic_int_add(&fs->cache_log_journal, type);
111 } else if (log->pending_write & CLOUDLOG_JOURNAL) {
112 g_atomic_int_add(&fs->cache_log_journal, type);
113 } else if (log->data != NULL) {
114 g_atomic_int_add(&fs->cache_log_dirty, type);
118 /* The reference held by the hash table does not count towards the reference
119 * count. When a new object is created, it initially has a reference count of
120 * 1 for the creator, and similarly fetching an item from the hash table will
121 * also create a reference. If the reference count drops to zero,
122 * bluesky_cloudlog_unref attempts to remove the object from the hash
123 * table--but there is a potential race since another thread might read the
124 * object from the hash table at the same time. So an object with a reference
125 * count of zero may still be resurrected, in which case we need to abort the
126 * destruction. Once the object is gone from the hash table, and if the
127 * reference count is still zero, it can actually be deleted. */
128 void bluesky_cloudlog_ref(BlueSkyCloudLog *log)
133 g_atomic_int_inc(&log->refcount);
136 void bluesky_cloudlog_unref(BlueSkyCloudLog *log)
141 if (g_atomic_int_dec_and_test(&log->refcount)) {
142 BlueSkyFS *fs = log->fs;
144 g_mutex_lock(fs->lock);
145 if (g_atomic_int_get(&log->refcount) > 0) {
146 g_mutex_unlock(fs->lock);
150 g_hash_table_remove(fs->locations, &log->id);
151 g_mutex_unlock(fs->lock);
153 bluesky_cloudlog_stats_update(log, -1);
154 log->type = LOGTYPE_INVALID;
155 g_mutex_free(log->lock);
156 g_cond_free(log->cond);
157 for (int i = 0; i < log->links->len; i++) {
158 BlueSkyCloudLog *c = g_array_index(log->links,
159 BlueSkyCloudLog *, i);
160 bluesky_cloudlog_unref(c);
162 g_array_unref(log->links);
163 bluesky_string_unref(log->data);
168 /* For locking reasons cloudlog unrefs may sometimes need to be performed in
169 * the future. We launch a thread for handling these delayed unreference
171 static gpointer cloudlog_unref_thread(gpointer q)
173 GAsyncQueue *queue = (GAsyncQueue *)q;
176 BlueSkyCloudLog *item = (BlueSkyCloudLog *)g_async_queue_pop(queue);
177 bluesky_cloudlog_unref(item);
183 void bluesky_cloudlog_unref_delayed(BlueSkyCloudLog *log)
186 g_async_queue_push(log->fs->unref_queue, log);
189 void bluesky_cloudlog_threads_init(BlueSkyFS *fs)
191 fs->unref_queue = g_async_queue_new();
192 g_thread_create(cloudlog_unref_thread, fs->unref_queue, FALSE, NULL);
195 /* Erase the information contained within the in-memory cloud log
196 * representation. This does not free up the item itself, but frees the data
197 * and references to other log items and resets the type back to unknown. If
198 * the object was written out to persistent storage, all state about it can be
199 * recovered by loading the object back in. The object must be locked before
200 * calling this function. */
201 void bluesky_cloudlog_erase(BlueSkyCloudLog *log)
203 g_assert(log->data_lock_count == 0);
205 log->type = LOGTYPE_UNKNOWN;
207 bluesky_string_unref(log->data);
209 log->data_lock_count = 0;
211 for (int i = 0; i < log->links->len; i++) {
212 BlueSkyCloudLog *c = g_array_index(log->links,
213 BlueSkyCloudLog *, i);
214 bluesky_cloudlog_unref(c);
216 g_array_unref(log->links);
217 log->links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
220 /* Start a write of the object to the local log. */
221 void bluesky_cloudlog_sync(BlueSkyCloudLog *log)
223 bluesky_log_item_submit(log, log->fs->log);
226 /* Add the given entry to the global hash table containing cloud log entries.
227 * Takes ownership of the caller's reference. */
228 void bluesky_cloudlog_insert_locked(BlueSkyCloudLog *log)
230 g_hash_table_insert(log->fs->locations, &log->id, log);
233 void bluesky_cloudlog_insert(BlueSkyCloudLog *log)
235 g_mutex_lock(log->fs->lock);
236 bluesky_cloudlog_insert_locked(log);
237 g_mutex_unlock(log->fs->lock);
240 /* Look up the cloud log entry for the given ID. If create is TRUE and the
241 * item does not exist, create a special pending entry that can later be filled
242 * in when the real item is loaded. The returned item has a reference held.
243 * As a special case, if a null ID is provided then NULL is returned. */
244 BlueSkyCloudLog *bluesky_cloudlog_get(BlueSkyFS *fs, BlueSkyCloudID id)
246 static BlueSkyCloudID id0 = {{0}};
248 if (memcmp(&id, &id0, sizeof(BlueSkyCloudID)) == 0)
251 g_mutex_lock(fs->lock);
252 BlueSkyCloudLog *item;
253 item = g_hash_table_lookup(fs->locations, &id);
255 item = bluesky_cloudlog_new(fs, &id);
256 g_hash_table_insert(fs->locations, &id, item);
258 bluesky_cloudlog_ref(item);
260 g_mutex_unlock(fs->lock);
264 /* Ensure that a cloud log item is loaded in memory, and if not read it in.
265 * TODO: Make asynchronous, and make this also fetch from the cloud. Right now
266 * we only read from the log. Log item must be locked. */
267 void bluesky_cloudlog_fetch(BlueSkyCloudLog *log)
269 if (log->data != NULL)
272 /* There are actually two cases: a full deserialization if we have not ever
273 * read the object before, and a partial deserialization where the metadata
274 * is already in memory and we just need to remap the data. If the object
275 * type has not yet been set, we'll need to read and parse the metadata.
276 * Once that is done, we can fall through the case of remapping the data
278 if (log->type == LOGTYPE_UNKNOWN) {
279 BlueSkyRCStr *raw = NULL;
280 if ((log->location_flags | log->pending_write) & CLOUDLOG_JOURNAL) {
281 raw = bluesky_log_map_object(log->fs, -1, log->log_seq,
282 log->log_offset, log->log_size);
285 if (raw == NULL && (log->location_flags & CLOUDLOG_CLOUD)) {
286 log->location_flags &= ~CLOUDLOG_JOURNAL;
287 raw = bluesky_log_map_object(log->fs,
288 log->location.directory,
289 log->location.sequence,
290 log->location.offset,
294 g_assert(raw != NULL);
295 bluesky_deserialize_cloudlog(log, raw->data, raw->len);
296 bluesky_string_unref(raw);
299 /* At this point all metadata should be available and we need only remap
300 * the object data. */
303 if ((log->location_flags | log->pending_write) & CLOUDLOG_JOURNAL) {
304 bluesky_cloudlog_stats_update(log, -1);
305 offset = log->log_offset + sizeof(struct log_header);
306 log->data = bluesky_log_map_object(log->fs, -1, log->log_seq,
307 offset, log->data_size);
308 bluesky_cloudlog_stats_update(log, 1);
311 if (log->data == NULL && (log->location_flags & CLOUDLOG_CLOUD)) {
312 log->location_flags &= ~CLOUDLOG_JOURNAL;
313 bluesky_cloudlog_stats_update(log, -1);
314 offset = log->location.offset + sizeof(struct cloudlog_header);
315 log->data = bluesky_log_map_object(log->fs, log->location.directory,
316 log->location.sequence,
317 offset, log->data_size);
318 bluesky_cloudlog_stats_update(log, 1);
321 if (log->data == NULL) {
322 g_error("Unable to fetch cloudlog entry!");
325 g_cond_broadcast(log->cond);
328 BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
331 BlueSkyCloudLogState *state = fs->log_state;
333 if ((log->location_flags | log->pending_write) & CLOUDLOG_CLOUD) {
334 return log->location;
337 for (int i = 0; i < log->links->len; i++) {
338 BlueSkyCloudLog *ref = g_array_index(log->links,
339 BlueSkyCloudLog *, i);
341 bluesky_cloudlog_serialize(ref, fs);
344 g_mutex_lock(log->lock);
345 bluesky_cloudlog_fetch(log);
346 g_assert(log->data != NULL);
348 bluesky_cloudlog_stats_update(log, -1);
350 GString *data1 = g_string_new("");
351 GString *data2 = g_string_new("");
352 GString *data3 = g_string_new("");
353 bluesky_serialize_cloudlog(log, data1, data2, data3);
355 log->location = state->location;
356 log->location.offset = state->data->len;
357 log->data_size = data1->len;
359 struct cloudlog_header header;
360 memcpy(header.magic, CLOUDLOG_MAGIC, 4);
361 memset(header.crypt_auth, sizeof(header.crypt_auth), 0);
362 memset(header.crypt_iv, sizeof(header.crypt_iv), 0);
363 header.type = log->type + '0';
364 header.size1 = GUINT32_TO_LE(data1->len);
365 header.size2 = GUINT32_TO_LE(data2->len);
366 header.size3 = GUINT32_TO_LE(data3->len);
368 header.inum = GUINT64_TO_LE(log->inum);
370 g_string_append_len(state->data, (const char *)&header, sizeof(header));
371 g_string_append_len(state->data, data1->str, data1->len);
372 g_string_append_len(state->data, data2->str, data2->len);
373 g_string_append_len(state->data, data3->str, data3->len);
375 log->location.size = state->data->len - log->location.offset;
377 g_string_free(data1, TRUE);
378 g_string_free(data2, TRUE);
379 g_string_free(data3, TRUE);
381 /* If the object we flushed was an inode, update the inode map. */
382 if (log->type == LOGTYPE_INODE) {
383 g_mutex_lock(fs->lock);
384 InodeMapEntry *entry = bluesky_inode_map_lookup(fs->inode_map,
387 entry->location = log->location;
389 bluesky_cloudlog_ref(entry->item);
390 g_mutex_unlock(fs->lock);
393 /* TODO: We should mark the objects as committed on the cloud until the
394 * data is flushed and acknowledged. */
395 log->pending_write |= CLOUDLOG_CLOUD;
396 bluesky_cloudlog_stats_update(log, 1);
397 state->writeback_list = g_slist_prepend(state->writeback_list, log);
398 bluesky_cloudlog_ref(log);
399 g_mutex_unlock(log->lock);
401 if (state->data->len > CLOUDLOG_SEGMENT_SIZE)
402 bluesky_cloudlog_flush(fs);
404 return log->location;
407 static void cloudlog_flush_complete(BlueSkyStoreAsync *async,
408 SerializedRecord *record)
410 g_print("Write of %s to cloud complete, status = %d\n",
411 async->key, async->result);
413 g_mutex_lock(record->lock);
414 if (async->result >= 0) {
415 while (record->items != NULL) {
416 BlueSkyCloudLog *item = (BlueSkyCloudLog *)record->items->data;
417 g_mutex_lock(item->lock);
418 bluesky_cloudlog_stats_update(item, -1);
419 item->pending_write &= ~CLOUDLOG_CLOUD;
420 item->location_flags |= CLOUDLOG_CLOUD;
421 bluesky_cloudlog_stats_update(item, 1);
422 g_mutex_unlock(item->lock);
423 bluesky_cloudlog_unref(item);
425 record->items = g_slist_delete_link(record->items, record->items);
428 bluesky_string_unref(record->data);
430 g_slist_free(record->items);
431 record->items = NULL;
432 record->complete = TRUE;
433 g_cond_broadcast(record->cond);
435 g_print("Write should be resubmitted...\n");
437 BlueSkyStoreAsync *async2 = bluesky_store_async_new(async->store);
438 async2->op = STORE_OP_PUT;
439 async2->key = g_strdup(async->key);
440 async2->data = record->data;
441 bluesky_string_ref(record->data);
442 bluesky_store_async_submit(async2);
443 bluesky_store_async_add_notifier(async2,
444 (GFunc)cloudlog_flush_complete,
446 bluesky_store_async_unref(async2);
448 g_mutex_unlock(record->lock);
451 /* Finish up a partially-written cloud log segment and flush it to storage. */
452 void bluesky_cloudlog_flush(BlueSkyFS *fs)
454 BlueSkyCloudLogState *state = fs->log_state;
455 if (state->data == NULL || state->data->len == 0)
458 /* TODO: Append some type of commit record to the log segment? */
460 g_print("Serializing %zd bytes of data to cloud\n", state->data->len);
461 SerializedRecord *record = g_new0(SerializedRecord, 1);
462 record->data = bluesky_string_new_from_gstring(state->data);
463 record->items = state->writeback_list;
464 record->lock = g_mutex_new();
465 record->cond = g_cond_new();
466 state->writeback_list = NULL;
468 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
469 async->op = STORE_OP_PUT;
470 async->key = g_strdup_printf("log-%08d-%08d",
471 state->location.directory,
472 state->location.sequence);
473 async->data = record->data;
474 bluesky_string_ref(record->data);
475 bluesky_store_async_submit(async);
476 bluesky_store_async_add_notifier(async,
477 (GFunc)cloudlog_flush_complete,
479 bluesky_store_async_unref(async);
481 state->pending_segments = g_list_prepend(state->pending_segments, record);
483 state->location.sequence++;
484 state->location.offset = 0;
485 state->data = g_string_new("");