// 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 LOG_SEGMENT_SIZE (1 << 24)
+#define LOG_SEGMENT_SIZE (1 << 22)
#define HEADER_MAGIC 0x676f4c0a
#define FOOTER_MAGIC 0x2e435243
log->fd = -1;
}
+ if (log->current_log != NULL) {
+ bluesky_cachefile_unref(log->current_log);
+ log->current_log = NULL;
+ }
+
while (log->fd < 0) {
- g_snprintf(logname, sizeof(logname), "log-%08d", log->seq_num);
+ g_snprintf(logname, sizeof(logname), "journal-%08d", log->seq_num);
log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
if (log->fd < 0 && errno == EEXIST) {
fprintf(stderr, "Log file %s already exists...\n", logname);
}
}
+ log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
+ g_assert(log->current_log != NULL);
+ g_mutex_unlock(log->current_log->lock);
+
if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
}
log->seq_num = 0;
log->queue = g_async_queue_new();
log->mmap_lock = g_mutex_new();
- log->mmap_cache = g_hash_table_new(NULL, NULL);
+ log->mmap_cache = g_hash_table_new(g_str_hash, g_str_equal);
log->dirfd = open(log->log_directory, O_DIRECTORY);
if (log->dirfd < 0) {
* to it. */
static int page_size = 0;
-BlueSkyRCStr *bluesky_log_map_object(BlueSkyLog *log,
- int log_seq, int log_offset, int log_size)
+void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
+{
+ g_atomic_int_add(&cachefile->refcount, -1);
+}
+
+static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
+ BlueSkyCacheFile *cachefile);
+
+static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
+{
+ g_atomic_int_inc(&cachefile->refcount);
+ cachefile->fetching = TRUE;
+ g_print("Starting fetch of %s from cloud\n", cachefile->filename);
+ BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
+ async->op = STORE_OP_GET;
+ async->key = g_strdup(cachefile->filename);
+ bluesky_store_async_add_notifier(async,
+ (GFunc)cloudlog_fetch_complete,
+ cachefile);
+ bluesky_store_async_submit(async);
+ bluesky_store_async_unref(async);
+}
+
+static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
+ BlueSkyCacheFile *cachefile)
+{
+ g_print("Fetch of %s from cloud complete, status = %d\n",
+ async->key, async->result);
+
+ g_mutex_lock(cachefile->lock);
+ if (async->result >= 0) {
+ char *pathname = g_strdup_printf("%s/%s",
+ cachefile->log->log_directory,
+ cachefile->filename);
+ if (!g_file_set_contents(pathname, async->data->data, async->data->len,
+ NULL))
+ g_print("Error writing out fetched file to cache!\n");
+ g_free(pathname);
+
+ cachefile->fetching = FALSE;
+ cachefile->ready = TRUE;
+ } else {
+ g_print("Error fetching from cloud, retrying...\n");
+ cloudlog_fetch_start(cachefile);
+ }
+
+ bluesky_cachefile_unref(cachefile);
+ g_cond_broadcast(cachefile->cond);
+ g_mutex_unlock(cachefile->lock);
+}
+
+/* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
+ * Returns the object in the locked state and with a reference taken. */
+BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
+ int clouddir, int log_seq)
{
if (page_size == 0) {
page_size = getpagesize();
}
- BlueSkyMmap *map;
+ BlueSkyLog *log = fs->log;
+
+ struct stat statbuf;
+ char logname[64];
+ int type;
+
+ // A request for a local log file
+ if (clouddir < 0) {
+ sprintf(logname, "journal-%08d", log_seq);
+ type = CLOUDLOG_JOURNAL;
+ } else {
+ sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
+ type = CLOUDLOG_CLOUD;
+ }
+
+ BlueSkyCacheFile *map;
g_mutex_lock(log->mmap_lock);
- map = g_hash_table_lookup(log->mmap_cache, GINT_TO_POINTER(log_seq));
+ map = g_hash_table_lookup(log->mmap_cache, logname);
+
+ if (map == NULL
+ && type == CLOUDLOG_JOURNAL
+ && fstatat(log->dirfd, logname, &statbuf, 0) < 0) {
+ /* A stale reference to a journal file which doesn't exist any longer
+ * because it was reclaimed. Return NULL. */
+ } else if (map == NULL) {
+ g_print("Adding cache file %s\n", logname);
+
+ map = g_new0(BlueSkyCacheFile, 1);
+ map->fs = fs;
+ map->type = type;
+ map->lock = g_mutex_new();
+ map->type = type;
+ g_mutex_lock(map->lock);
+ map->cond = g_cond_new();
+ map->filename = g_strdup(logname);
+ map->log_seq = log_seq;
+ map->log = log;
+ g_atomic_int_set(&map->mapcount, 0);
+ g_atomic_int_set(&map->refcount, 0);
+
+ g_hash_table_insert(log->mmap_cache, map->filename, map);
+
+ // If the log file is stored in the cloud, we may need to fetch it
+ if (clouddir >= 0)
+ cloudlog_fetch_start(map);
+ } else {
+ g_mutex_lock(map->lock);
+ }
+
+ g_mutex_unlock(log->mmap_lock);
+ if (map != NULL)
+ g_atomic_int_inc(&map->refcount);
+ return map;
+}
+
+BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir,
+ int log_seq, int log_offset, int log_size)
+{
+ if (page_size == 0) {
+ page_size = getpagesize();
+ }
+
+ BlueSkyLog *log = fs->log;
+ BlueSkyCacheFile *map = bluesky_cachefile_lookup(fs, log_dir, log_seq);
if (map == NULL) {
- char logname[64];
- g_snprintf(logname, sizeof(logname), "log-%08d", log_seq);
- int fd = openat(log->dirfd, logname, O_RDONLY);
+ return NULL;
+ }
+
+ if (map->addr == NULL) {
+ while (!map->ready && map->fetching) {
+ g_print("Waiting for log segment to be fetched from cloud...\n");
+ g_cond_wait(map->cond, map->lock);
+ }
+
+ int fd = openat(log->dirfd, map->filename, O_RDONLY);
if (fd < 0) {
- fprintf(stderr, "Error opening logfile %s: %m\n", logname);
- g_mutex_unlock(log->mmap_lock);
+ fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
+ bluesky_cachefile_unref(map);
+ g_mutex_unlock(map->lock);
return NULL;
}
- map = g_new0(BlueSkyMmap, 1);
-
off_t length = lseek(fd, 0, SEEK_END);
- map->log_seq = log_seq;
map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
fd, 0);
+ g_atomic_int_add(&log->disk_used, -(map->len / 1024));
map->len = length;
- map->log = log;
- g_atomic_int_set(&map->refcount, 0);
+ g_atomic_int_add(&log->disk_used, map->len / 1024);
- g_hash_table_insert(log->mmap_cache, GINT_TO_POINTER(log_seq), map);
-
- g_print("Mapped log segment %d...\n", log_seq);
+ g_print("Re-mapped log segment %d...\n", log_seq);
+ g_atomic_int_inc(&map->refcount);
close(fd);
}
g_mutex_unlock(log->mmap_lock);
- return bluesky_string_new_from_mmap(map, log_offset, log_size);
+ BlueSkyRCStr *str;
+ map->atime = bluesky_get_current_time();
+ str = bluesky_string_new_from_mmap(map, log_offset, log_size);
+ bluesky_cachefile_unref(map);
+ g_mutex_unlock(map->lock);
+ return str;
}
-void bluesky_mmap_unref(BlueSkyMmap *mmap)
+void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
{
if (mmap == NULL)
return;
- if (g_atomic_int_dec_and_test(&mmap->refcount)) {
- /* There is a potential race condition here: the BlueSkyLog contains a
- * hash table of currently-existing BlueSkyMmap objects, which does not
- * hold a reference. Some other thread might grab a new reference to
- * this object after reading it from the hash table. So, before
- * destruction we need to grab the lock for the hash table, then check
- * the reference count again. If it is still zero, we can proceed with
- * object destruction. */
- BlueSkyLog *log = mmap->log;
- g_mutex_lock(log->mmap_lock);
- if (g_atomic_int_get(&mmap->refcount) > 0) {
- g_mutex_unlock(log->mmap_lock);
- return;
+ if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
+ g_mutex_lock(mmap->lock);
+ if (g_atomic_int_get(&mmap->mapcount) == 0) {
+ g_print("Unmapped log segment %d...\n", mmap->log_seq);
+ munmap((void *)mmap->addr, mmap->len);
+ mmap->addr = NULL;
+ g_atomic_int_add(&mmap->refcount, -1);
}
-
- g_hash_table_remove(log->mmap_cache, GINT_TO_POINTER(mmap->log_seq));
- munmap((void *)mmap->addr, mmap->len);
- g_free(mmap);
- g_mutex_unlock(log->mmap_lock);
+ g_mutex_unlock(mmap->lock);
}
}
+/* Scan through all currently-stored files in the journal/cache and garbage
+ * collect old unused ones, if needed. */
+static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
+{
+ GList **files = (GList **)user_data;
+ *files = g_list_prepend(*files, value);
+}
+
+static gint compare_cachefiles(gconstpointer a, gconstpointer b)
+{
+ int64_t ta, tb;
+
+ ta = ((BlueSkyCacheFile *)a)->atime;
+ tb = ((BlueSkyCacheFile *)b)->atime;
+ if (ta < tb)
+ return -1;
+ else if (ta > tb)
+ return 1;
+ else
+ return 0;
+}
+
+void bluesky_cachefile_gc(BlueSkyFS *fs)
+{
+ GList *files = NULL;
+
+ g_mutex_lock(fs->log->mmap_lock);
+ g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
+
+ /* Sort based on atime. The atime should be stable since it shouln't be
+ * updated except by threads which can grab the mmap_lock, which we already
+ * hold. */
+ files = g_list_sort(files, compare_cachefiles);
+
+ /* Walk the list of files, starting with the oldest, deleting files if
+ * possible until enough space has been reclaimed. */
+ g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
+ while (files != NULL) {
+ BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
+ /* Try to lock the structure, but if the lock is held by another thread
+ * then we'll just skip the file on this pass. */
+ if (g_mutex_trylock(cachefile->lock)) {
+ int64_t age = bluesky_get_current_time() - cachefile->atime;
+ g_print("%s addr=%p mapcount=%d refcount=%d atime_age=%f",
+ cachefile->filename, cachefile->addr, cachefile->mapcount,
+ cachefile->refcount, age / 1e6);
+ if (cachefile->fetching)
+ g_print(" (fetching)");
+ g_print("\n");
+
+ gboolean deletion_candidate = FALSE;
+ if (g_atomic_int_get(&fs->log->disk_used)
+ > bluesky_options.cache_size
+ && g_atomic_int_get(&cachefile->refcount) == 0
+ && g_atomic_int_get(&cachefile->mapcount) == 0)
+ {
+ deletion_candidate = TRUE;
+ }
+
+ /* Don't allow journal files to be reclaimed until all data is
+ * known to be durably stored in the cloud. */
+ if (cachefile->type == CLOUDLOG_JOURNAL
+ && cachefile->log_seq >= fs->log->journal_watermark)
+ {
+ deletion_candidate = FALSE;
+ }
+
+ if (deletion_candidate) {
+ g_print(" ...deleting\n");
+ if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
+ fprintf(stderr, "Unable to unlink journal %s: %m\n",
+ cachefile->filename);
+ }
+
+ g_atomic_int_add(&fs->log->disk_used, -(cachefile->len / 1024));
+ g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
+ g_mutex_unlock(cachefile->lock);
+ g_mutex_free(cachefile->lock);
+ g_cond_free(cachefile->cond);
+ g_free(cachefile->filename);
+ g_free(cachefile);
+ } else {
+ g_mutex_unlock(cachefile->lock);
+ }
+ }
+ files = g_list_delete_link(files, files);
+ }
+ g_list_free(files);
+
+ g_mutex_unlock(fs->log->mmap_lock);
+}