* TODO: Licensing
*/
+#include <stdio.h>
#include <stdint.h>
#include <glib.h>
#include <string.h>
#include "bluesky-private.h"
-/* The locations hash table in the file system is used to map objects to their locations. Objects are named using 128- */
+// Rough size limit for a log segment. This is not a firm limit and there are
+// no absolute guarantees on the size of a log segment.
+#define CLOUDLOG_SEGMENT_SIZE (4 << 20)
-typedef struct {
+BlueSkyCloudID bluesky_cloudlog_new_id()
+{
BlueSkyCloudID id;
+ bluesky_crypt_random_bytes((uint8_t *)&id.bytes, sizeof(id));
+ return id;
+}
- BlueSkyCloudPointer *cloud_loc;
-} BlueSkyLocationEntry;
+gchar *bluesky_cloudlog_id_to_string(BlueSkyCloudID id)
+{
+ char buf[sizeof(BlueSkyCloudID) * 2 + 1];
+ buf[0] = '\0';
-BlueSkyCloudID bluesky_cloudlog_new_id()
+ for (int i = 0; i < sizeof(BlueSkyCloudID); i++) {
+ sprintf(&buf[2*i], "%02x", (uint8_t)(id.bytes[i]));
+ }
+
+ return g_strdup(buf);
+}
+
+BlueSkyCloudID bluesky_cloudlog_id_from_string(const gchar *idstr)
{
BlueSkyCloudID id;
- bluesky_crypt_random_bytes((uint8_t *)&id.bytes, sizeof(id));
+ memset(&id, 0, sizeof(id));
+ for (int i = 0; i < 2*sizeof(BlueSkyCloudID); i++) {
+ char c = idstr[i];
+ if (c == '\0') {
+ g_warning("Short cloud id: %s\n", idstr);
+ break;
+ }
+ int val = 0;
+ if (c >= '0' && c <= '9')
+ val = c - '0';
+ else if (c >= 'a' && c <= 'f')
+ val = c - 'a' + 10;
+ else
+ g_warning("Bad character in cloud id: %s\n", idstr);
+ id.bytes[i / 2] += val << (i % 2 ? 0 : 4);
+ }
return id;
}
{
BlueSkyCloudLog *log = g_new0(BlueSkyCloudLog, 1);
+ log->lock = g_mutex_new();
+ log->cond = g_cond_new();
log->fs = fs;
log->type = LOGTYPE_UNKNOWN;
log->id = bluesky_cloudlog_new_id();
+ log->links = g_array_new(FALSE, TRUE, sizeof(BlueSkyCloudLog *));
g_atomic_int_set(&log->refcount, 1);
return log;
}
+/* Helper function for updating memory usage statistics for a filesystem (the
+ * cache_log_* variables). This will increment (type=1) or decrement (type=-1)
+ * the counter associated with the current state of the cloud log item. The
+ * item should be locked or otherwise protected from concurrent access. */
+void bluesky_cloudlog_stats_update(BlueSkyCloudLog *log, int type)
+{
+ BlueSkyFS *fs = log->fs;
+
+ if (log->location_flags & CLOUDLOG_CLOUD) {
+ g_atomic_int_add(&fs->cache_log_cloud, type);
+ } else if (log->location_flags & CLOUDLOG_JOURNAL) {
+ g_atomic_int_add(&fs->cache_log_journal, type);
+ } else if (log->pending_write & CLOUDLOG_JOURNAL) {
+ g_atomic_int_add(&fs->cache_log_journal, type);
+ } else if (log->data != NULL) {
+ g_atomic_int_add(&fs->cache_log_dirty, type);
+ }
+}
+
+/* The reference held by the hash table does not count towards the reference
+ * count. When a new object is created, it initially has a reference count of
+ * 1 for the creator, and similarly fetching an item from the hash table will
+ * also create a reference. If the reference count drops to zero,
+ * bluesky_cloudlog_unref attempts to remove the object from the hash
+ * table--but there is a potential race since another thread might read the
+ * object from the hash table at the same time. So an object with a reference
+ * count of zero may still be resurrected, in which case we need to abort the
+ * destruction. Once the object is gone from the hash table, and if the
+ * reference count is still zero, it can actually be deleted. */
void bluesky_cloudlog_ref(BlueSkyCloudLog *log)
{
+ if (log == NULL)
+ return;
+
g_atomic_int_inc(&log->refcount);
}
void bluesky_cloudlog_unref(BlueSkyCloudLog *log)
{
+ if (log == NULL)
+ return;
+
if (g_atomic_int_dec_and_test(&log->refcount)) {
- g_print("Cloud log refcount dropped to zero.\n");
+ BlueSkyFS *fs = log->fs;
+
+ g_mutex_lock(fs->lock);
+ if (g_atomic_int_get(&log->refcount) > 0) {
+ g_mutex_unlock(fs->lock);
+ return;
+ }
+
+ g_hash_table_remove(fs->locations, &log->id);
+ g_mutex_unlock(fs->lock);
+
+ bluesky_cloudlog_stats_update(log, -1);
+ log->type = LOGTYPE_INVALID;
+ g_mutex_free(log->lock);
+ g_cond_free(log->cond);
+ for (int i = 0; i < log->links->len; i++) {
+ BlueSkyCloudLog *c = g_array_index(log->links,
+ BlueSkyCloudLog *, i);
+ bluesky_cloudlog_unref(c);
+ }
+ g_array_unref(log->links);
+ bluesky_string_unref(log->data);
+ g_free(log);
}
}
/* Start a write of the object to the local log. */
-BlueSkyLogItem *bluesky_cloudlog_sync(BlueSkyCloudLog *log)
+void bluesky_cloudlog_sync(BlueSkyCloudLog *log)
{
- BlueSkyLogItem *log_item = bluesky_log_item_new();
- log_item->key = g_strdup("cloudlog");
- log_item->data = log->data;
- bluesky_string_ref(log->data);
- bluesky_log_item_submit(log_item, log->fs->log);
- return log_item;
+ bluesky_log_item_submit(log, log->fs->log);
}
/* Add the given entry to the global hash table containing cloud log entries.
g_hash_table_insert(log->fs->locations, &log->id, log);
g_mutex_unlock(log->fs->lock);
}
+
+struct log_header {
+ char magic[4];
+ uint32_t size;
+ BlueSkyCloudID id;
+ uint32_t pointer_count;
+} __attribute__((packed));
+
+struct logref {
+ BlueSkyCloudID id;
+ BlueSkyCloudPointer location;
+} __attribute__((packed));
+
+struct log_footer {
+ char refmagic[4];
+ struct logref refs[0];
+};
+
+/* Ensure that a cloud log item is loaded in memory, and if not read it in.
+ * TODO: Make asynchronous, and make this also fetch from the cloud. Right now
+ * we only read from the log. Log item must be locked. */
+void bluesky_cloudlog_fetch(BlueSkyCloudLog *log)
+{
+ if (log->data != NULL)
+ return;
+
+ if ((log->location_flags | log->pending_write) & CLOUDLOG_JOURNAL) {
+ bluesky_cloudlog_stats_update(log, -1);
+ log->data = bluesky_log_map_object(log->fs, -1, log->log_seq,
+ log->log_offset, log->log_size);
+ bluesky_cloudlog_stats_update(log, 1);
+ }
+
+ if (log->data == NULL && (log->location_flags & CLOUDLOG_CLOUD)) {
+ log->location_flags &= ~CLOUDLOG_JOURNAL;
+ bluesky_cloudlog_stats_update(log, -1);
+ log->data = bluesky_log_map_object(log->fs, log->location.directory,
+ log->location.sequence,
+ log->location.offset,
+ log->location.size);
+ bluesky_cloudlog_stats_update(log, 1);
+ }
+
+ if (log->data == NULL) {
+ g_error("Unable to fetch cloudlog entry!");
+ }
+
+ g_cond_broadcast(log->cond);
+}
+
+BlueSkyCloudPointer bluesky_cloudlog_serialize(BlueSkyCloudLog *log,
+ BlueSkyFS *fs)
+{
+ BlueSkyCloudLogState *state = fs->log_state;
+
+ if (log->location_flags & CLOUDLOG_CLOUD) {
+ return log->location;
+ }
+
+ for (int i = 0; i < log->links->len; i++) {
+ BlueSkyCloudLog *ref = g_array_index(log->links,
+ BlueSkyCloudLog *, i);
+ if (ref != NULL)
+ bluesky_cloudlog_serialize(ref, fs);
+ }
+
+ g_mutex_lock(log->lock);
+ bluesky_cloudlog_fetch(log);
+ g_assert(log->data != NULL);
+
+ bluesky_cloudlog_stats_update(log, -1);
+
+ /* TODO: Right now offset/size are set to the raw data, but we should add
+ * header parsing to the code which loads objects back in. */
+ log->location = state->location;
+ log->location.offset = state->data->len + sizeof(struct log_header);
+ log->location.size = log->data->len;
+ /* = sizeof(struct log_header) + sizeof(BlueSkyCloudID) * 0
+ + log->data->len; */
+
+ struct log_header header;
+ memcpy(header.magic, "AgI ", 4);
+ header.size = GUINT32_TO_LE(log->location.size);
+ header.id = log->id;
+ header.pointer_count = GUINT32_TO_LE(0);
+
+ g_string_append_len(state->data, (const char *)&header, sizeof(header));
+ g_string_append_len(state->data, log->data->data, log->data->len);
+
+ /* TODO: We should mark the objects as committed on the cloud until the
+ * data is flushed and acknowledged. */
+ log->pending_write |= CLOUDLOG_CLOUD;
+ bluesky_cloudlog_stats_update(log, 1);
+ state->writeback_list = g_slist_prepend(state->writeback_list, log);
+ bluesky_cloudlog_ref(log);
+ g_mutex_unlock(log->lock);
+
+ if (state->data->len > CLOUDLOG_SEGMENT_SIZE)
+ bluesky_cloudlog_flush(fs);
+
+ return log->location;
+}
+
+typedef struct {
+ BlueSkyRCStr *data;
+ GSList *items;
+} SerializedRecord;
+
+static void cloudlog_flush_complete(BlueSkyStoreAsync *async,
+ SerializedRecord *record)
+{
+ g_print("Write of %s to cloud complete, status = %d\n",
+ async->key, async->result);
+
+ if (async->result >= 0) {
+ while (record->items != NULL) {
+ BlueSkyCloudLog *item = (BlueSkyCloudLog *)record->items->data;
+ g_mutex_lock(item->lock);
+ bluesky_cloudlog_stats_update(item, -1);
+ item->pending_write &= ~CLOUDLOG_CLOUD;
+ item->location_flags |= CLOUDLOG_CLOUD;
+ bluesky_cloudlog_stats_update(item, 1);
+ g_mutex_unlock(item->lock);
+ bluesky_cloudlog_unref(item);
+
+ record->items = g_slist_delete_link(record->items, record->items);
+ }
+
+ bluesky_string_unref(record->data);
+ g_slist_free(record->items);
+ g_free(record);
+ } else {
+ g_print("Write should be resubmitted...\n");
+
+ BlueSkyStoreAsync *async2 = bluesky_store_async_new(async->store);
+ async2->op = STORE_OP_PUT;
+ async2->key = g_strdup(async->key);
+ async2->data = record->data;
+ bluesky_string_ref(record->data);
+ bluesky_store_async_submit(async2);
+ bluesky_store_async_add_notifier(async2,
+ (GFunc)cloudlog_flush_complete,
+ record);
+ bluesky_store_async_unref(async2);
+ }
+}
+
+/* Finish up a partially-written cloud log segment and flush it to storage. */
+void bluesky_cloudlog_flush(BlueSkyFS *fs)
+{
+ BlueSkyCloudLogState *state = fs->log_state;
+ if (state->data == NULL || state->data->len == 0)
+ return;
+
+ /* TODO: Append some type of commit record to the log segment? */
+
+ g_print("Serializing %zd bytes of data to cloud\n", state->data->len);
+ SerializedRecord *record = g_new0(SerializedRecord, 1);
+ record->data = bluesky_string_new_from_gstring(state->data);
+ record->items = state->writeback_list;
+ state->writeback_list = NULL;
+
+ BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
+ async->op = STORE_OP_PUT;
+ async->key = g_strdup_printf("log-%08d-%08d",
+ state->location.directory,
+ state->location.sequence);
+ async->data = record->data;
+ bluesky_string_ref(record->data);
+ bluesky_store_async_submit(async);
+ bluesky_store_async_add_notifier(async,
+ (GFunc)cloudlog_flush_complete,
+ record);
+ bluesky_store_async_unref(async);
+
+ state->location.sequence++;
+ state->location.offset = 0;
+ state->data = g_string_new("");
+}
projects / bluesky.git / blobdiff