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)
83 BlueSkyCloudLog *log = g_new0(BlueSkyCloudLog, 1);
85 log->lock = g_mutex_new();
86 log->cond = g_cond_new();
88 log->type = LOGTYPE_UNKNOWN;
89 log->id = bluesky_cloudlog_new_id();
90 log->links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
91 g_atomic_int_set(&log->refcount, 1);
96 /* Helper function for updating memory usage statistics for a filesystem (the
97 * cache_log_* variables). This will increment (type=1) or decrement (type=-1)
98 * the counter associated with the current state of the cloud log item. The
99 * item should be locked or otherwise protected from concurrent access. */
100 void bluesky_cloudlog_stats_update(BlueSkyCloudLog *log, int type)
102 BlueSkyFS *fs = log->fs;
104 if (log->location_flags & CLOUDLOG_CLOUD) {
105 g_atomic_int_add(&fs->cache_log_cloud, type);
106 } else if (log->location_flags & CLOUDLOG_JOURNAL) {
107 g_atomic_int_add(&fs->cache_log_journal, type);
108 } else if (log->pending_write & CLOUDLOG_JOURNAL) {
109 g_atomic_int_add(&fs->cache_log_journal, type);
110 } else if (log->data != NULL) {
111 g_atomic_int_add(&fs->cache_log_dirty, type);
115 /* The reference held by the hash table does not count towards the reference
116 * count. When a new object is created, it initially has a reference count of
117 * 1 for the creator, and similarly fetching an item from the hash table will
118 * also create a reference. If the reference count drops to zero,
119 * bluesky_cloudlog_unref attempts to remove the object from the hash
120 * table--but there is a potential race since another thread might read the
121 * object from the hash table at the same time. So an object with a reference
122 * count of zero may still be resurrected, in which case we need to abort the
123 * destruction. Once the object is gone from the hash table, and if the
124 * reference count is still zero, it can actually be deleted. */
125 void bluesky_cloudlog_ref(BlueSkyCloudLog *log)
130 g_atomic_int_inc(&log->refcount);
133 void bluesky_cloudlog_unref(BlueSkyCloudLog *log)
138 if (g_atomic_int_dec_and_test(&log->refcount)) {
139 BlueSkyFS *fs = log->fs;
141 g_mutex_lock(fs->lock);
142 if (g_atomic_int_get(&log->refcount) > 0) {
143 g_mutex_unlock(fs->lock);
147 g_hash_table_remove(fs->locations, &log->id);
148 g_mutex_unlock(fs->lock);
150 bluesky_cloudlog_stats_update(log, -1);
151 log->type = LOGTYPE_INVALID;
152 g_mutex_free(log->lock);
153 g_cond_free(log->cond);
154 for (int i = 0; i < log->links->len; i++) {
155 BlueSkyCloudLog *c = g_array_index(log->links,
156 BlueSkyCloudLog *, i);
157 bluesky_cloudlog_unref(c);
159 g_array_unref(log->links);
160 bluesky_string_unref(log->data);
161 if (log->dirty_journal != NULL)
162 g_atomic_int_add(&log->dirty_journal->dirty_refs, -1);
167 /* Start a write of the object to the local log. */
168 void bluesky_cloudlog_sync(BlueSkyCloudLog *log)
170 bluesky_log_item_submit(log, log->fs->log);
173 /* Add the given entry to the global hash table containing cloud log entries.
174 * Takes ownership of the caller's reference. */
175 void bluesky_cloudlog_insert(BlueSkyCloudLog *log)
177 g_mutex_lock(log->fs->lock);
178 g_hash_table_insert(log->fs->locations, &log->id, log);
179 g_mutex_unlock(log->fs->lock);
186 uint32_t pointer_count;
187 } __attribute__((packed));
191 BlueSkyCloudPointer location;
192 } __attribute__((packed));
196 struct logref refs[0];
199 /* Ensure that a cloud log item is loaded in memory, and if not read it in.
200 * TODO: Make asynchronous, and make this also fetch from the cloud. Right now
201 * we only read from the log. Log item must be locked. */
202 void bluesky_cloudlog_fetch(BlueSkyCloudLog *log)
204 if (log->data != NULL)
207 if ((log->location_flags | log->pending_write) & CLOUDLOG_JOURNAL) {
208 bluesky_cloudlog_stats_update(log, -1);
209 log->data = bluesky_log_map_object(log->fs, -1, log->log_seq,
210 log->log_offset, log->log_size);
211 bluesky_cloudlog_stats_update(log, 1);
214 if (log->data == NULL && (log->location_flags & CLOUDLOG_CLOUD)) {
215 log->location_flags &= ~CLOUDLOG_JOURNAL;
216 bluesky_cloudlog_stats_update(log, -1);
217 log->data = bluesky_log_map_object(log->fs, log->location.directory,
218 log->location.sequence,
219 log->location.offset,
221 bluesky_cloudlog_stats_update(log, 1);
224 if (log->data == NULL) {
225 g_error("Unable to fetch cloudlog entry!");
228 g_cond_broadcast(log->cond);
231 BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
234 BlueSkyCloudLogState *state = fs->log_state;
236 if (log->location_flags & CLOUDLOG_CLOUD) {
237 return log->location;
240 for (int i = 0; i < log->links->len; i++) {
241 BlueSkyCloudLog *ref = g_array_index(log->links,
242 BlueSkyCloudLog *, i);
244 bluesky_cloudlog_serialize(ref, fs);
247 g_mutex_lock(log->lock);
248 bluesky_cloudlog_fetch(log);
249 g_assert(log->data != NULL);
251 bluesky_cloudlog_stats_update(log, -1);
253 /* TODO: Right now offset/size are set to the raw data, but we should add
254 * header parsing to the code which loads objects back in. */
255 log->location = state->location;
256 log->location.offset = state->data->len + sizeof(struct log_header);
257 log->location.size = log->data->len;
258 /* = sizeof(struct log_header) + sizeof(BlueSkyCloudID) * 0
261 struct log_header header;
262 memcpy(header.magic, "AgI ", 4);
263 header.size = GUINT32_TO_LE(log->location.size);
265 header.pointer_count = GUINT32_TO_LE(0);
267 g_string_append_len(state->data, (const char *)&header, sizeof(header));
268 g_string_append_len(state->data, log->data->data, log->data->len);
270 /* TODO: We should mark the objects as committed on the cloud until the
271 * data is flushed and acknowledged. */
272 log->pending_write |= CLOUDLOG_CLOUD;
273 bluesky_cloudlog_stats_update(log, 1);
274 state->writeback_list = g_slist_prepend(state->writeback_list, log);
275 bluesky_cloudlog_ref(log);
276 g_mutex_unlock(log->lock);
278 if (state->data->len > CLOUDLOG_SEGMENT_SIZE)
279 bluesky_cloudlog_flush(fs);
281 return log->location;
289 static void cloudlog_flush_complete(BlueSkyStoreAsync *async,
290 SerializedRecord *record)
292 g_print("Write of %s to cloud complete, status = %d\n",
293 async->key, async->result);
295 if (async->result >= 0) {
296 while (record->items != NULL) {
297 BlueSkyCloudLog *item = (BlueSkyCloudLog *)record->items->data;
298 g_mutex_lock(item->lock);
299 bluesky_cloudlog_stats_update(item, -1);
300 item->pending_write &= ~CLOUDLOG_CLOUD;
301 item->location_flags |= CLOUDLOG_CLOUD;
302 bluesky_cloudlog_stats_update(item, 1);
303 if (item->dirty_journal != NULL) {
304 g_atomic_int_add(&item->dirty_journal->dirty_refs, -1);
305 item->dirty_journal = NULL;
307 g_mutex_unlock(item->lock);
308 bluesky_cloudlog_unref(item);
310 record->items = g_slist_delete_link(record->items, record->items);
313 bluesky_string_unref(record->data);
314 g_slist_free(record->items);
317 g_print("Write should be resubmitted...\n");
319 BlueSkyStoreAsync *async2 = bluesky_store_async_new(async->store);
320 async2->op = STORE_OP_PUT;
321 async2->key = g_strdup(async->key);
322 async2->data = record->data;
323 bluesky_string_ref(record->data);
324 bluesky_store_async_submit(async2);
325 bluesky_store_async_add_notifier(async2,
326 (GFunc)cloudlog_flush_complete,
328 bluesky_store_async_unref(async2);
332 /* Finish up a partially-written cloud log segment and flush it to storage. */
333 void bluesky_cloudlog_flush(BlueSkyFS *fs)
335 BlueSkyCloudLogState *state = fs->log_state;
336 if (state->data == NULL || state->data->len == 0)
339 /* TODO: Append some type of commit record to the log segment? */
341 g_print("Serializing %zd bytes of data to cloud\n", state->data->len);
342 SerializedRecord *record = g_new0(SerializedRecord, 1);
343 record->data = bluesky_string_new_from_gstring(state->data);
344 record->items = state->writeback_list;
345 state->writeback_list = NULL;
347 BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
348 async->op = STORE_OP_PUT;
349 async->key = g_strdup_printf("log-%08d-%08d",
350 state->location.directory,
351 state->location.sequence);
352 async->data = record->data;
353 bluesky_string_ref(record->data);
354 bluesky_store_async_submit(async);
355 bluesky_store_async_add_notifier(async,
356 (GFunc)cloudlog_flush_complete,
358 bluesky_store_async_unref(async);
360 state->location.sequence++;
361 state->location.offset = 0;
362 state->data = g_string_new("");