X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Flog.c;h=66b4edb9131da2ad5151ae0af83158b49ca3aa64;hb=b8885660a453f0f1f30c9fed196b01a397482a19;hp=9dfaa5bfd61437505ca52ed7b339296a8804a6f0;hpb=fb14012672cefd58f8aa7587152f918824ad8ea8;p=bluesky.git diff --git a/bluesky/log.c b/bluesky/log.c index 9dfaa5b..66b4edb 100644 --- a/bluesky/log.c +++ b/bluesky/log.c @@ -77,8 +77,10 @@ static void log_commit(BlueSkyLog *log) while (log->committed != NULL) { BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data; g_mutex_lock(item->lock); + bluesky_cloudlog_stats_update(item, -1); item->pending_write &= ~CLOUDLOG_JOURNAL; item->location_flags |= CLOUDLOG_JOURNAL; + bluesky_cloudlog_stats_update(item, 1); g_cond_signal(item->cond); g_mutex_unlock(item->lock); log->committed = g_slist_delete_link(log->committed, log->committed); @@ -101,8 +103,13 @@ static gboolean log_open(BlueSkyLog *log) 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); @@ -114,6 +121,10 @@ static gboolean log_open(BlueSkyLog *log) } } + 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); } @@ -158,7 +169,9 @@ static gpointer log_thread(gpointer d) continue; } + bluesky_cloudlog_stats_update(item, -1); item->pending_write |= CLOUDLOG_JOURNAL; + bluesky_cloudlog_stats_update(item, 1); struct log_header header; struct log_footer footer; @@ -200,6 +213,21 @@ static gpointer log_thread(gpointer d) offset += sizeof(header) + sizeof(footer) + item->data->len; + /* Since we have just written a new dirty object to the journal, + * increment the count of live dirty objects in that journal file. The + * count will be decremented when objects are deleted or written to the + * cloud. */ + g_atomic_int_add(&log->current_log->dirty_refs, 1); + item->dirty_journal = log->current_log; + + /* Replace the log item's string data with a memory-mapped copy of the + * data, now that it has been written to the log file. (Even if it + * isn't yet on disk, it should at least be in the page cache and so + * available to memory map.) */ + bluesky_string_unref(item->data); + item->data = NULL; + bluesky_cloudlog_fetch(item); + log->committed = g_slist_prepend(log->committed, item); g_atomic_int_add(&item->data_lock_count, -1); g_mutex_unlock(item->lock); @@ -222,7 +250,7 @@ BlueSkyLog *bluesky_log_new(const char *log_directory) 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) { @@ -261,74 +289,244 @@ void bluesky_log_finish_all(GList *log_items) * 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) +{ + g_print("Fetch of %s from cloud complete, status = %d\n", + async->key, async->result); + + if (async->result < 0) + return; + + g_mutex_lock(cachefile->lock); + 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; + 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; + + 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) { - char logname[64]; - g_snprintf(logname, sizeof(logname), "log-%08d", log_seq); - int fd = openat(log->dirfd, logname, O_RDONLY); + /* TODO: stat() call */ + map = g_new0(BlueSkyCacheFile, 1); + map->type = CLOUDLOG_JOURNAL; + 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) { + g_atomic_int_inc(&map->refcount); + map->fetching = TRUE; + g_print("Starting fetch of %s from cloud\n", logname); + BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store); + async->op = STORE_OP_GET; + async->key = g_strdup(logname); + bluesky_store_async_add_notifier(async, + (GFunc)cloudlog_fetch_complete, + map); + bluesky_store_async_submit(async); + bluesky_store_async_unref(async); + } + } else { + g_mutex_lock(map->lock); + } + + g_mutex_unlock(log->mmap_lock); + 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) { + 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_hash_table_insert(log->mmap_cache, GINT_TO_POINTER(log_seq), map); + g_atomic_int_add(&log->disk_used, map->len / 1024); - 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 dirty=%d atime_age=%f", + cachefile->filename, cachefile->addr, cachefile->mapcount, + cachefile->refcount, cachefile->dirty_refs, age / 1e6); + if (cachefile->fetching) + g_print(" (fetching)"); + g_print("\n"); + + if (g_atomic_int_get(&cachefile->refcount) == 0 + && g_atomic_int_get(&cachefile->mapcount) == 0 + && g_atomic_int_get(&cachefile->dirty_refs) == 0) + { + 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_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); +}