X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Flog.c;h=152f5980d604960c9afdb5c5abb870efd111c94e;hb=a8a9f2181e2e16d24d812ea1e7a7c8af42f0d2f1;hp=a4a81505d7565d522150e9be6df5bc79b3fecf99;hpb=0ba2a52ec1386c8928b397f775e4a4f97339fa2a;p=bluesky.git diff --git a/bluesky/log.c b/bluesky/log.c index a4a8150..152f598 100644 --- a/bluesky/log.c +++ b/bluesky/log.c @@ -19,6 +19,7 @@ #include #include #include +#include #include "bluesky-private.h" @@ -35,7 +36,22 @@ // 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 << 20) +#define LOG_SEGMENT_SIZE (1 << 23) + +#define HEADER_MAGIC 0x676f4c0a +#define FOOTER_MAGIC 0x2e435243 + +struct log_header { + uint32_t magic; // HEADER_MAGIC + uint64_t offset; // Starting byte offset of the log header + uint32_t size; // Size of the data item (bytes) + BlueSkyCloudID id; // Object identifier +} __attribute__((packed)); + +struct log_footer { + uint32_t magic; // FOOTER_MAGIC + uint32_t crc; // Computed from log_header to log_footer.magic +} __attribute__((packed)); static void writebuf(int fd, const char *buf, size_t len) { @@ -66,6 +82,12 @@ static gpointer log_thread(gpointer d) { BlueSkyLog *log = (BlueSkyLog *)d; + /* If there are multiple log items to write, we may write more than one + * before calling fsync(). The committed list is used to track all the + * items that should be marked as committed once that final fsync() is + * done. */ + GSList *committed = NULL; + int dirfd = open(log->log_directory, O_DIRECTORY); if (dirfd < 0) { fprintf(stderr, "Unable to open logging directory: %m\n"); @@ -89,16 +111,74 @@ static gpointer log_thread(gpointer d) fsync(dirfd); } - BlueSkyLogItem *item = (BlueSkyLogItem *)g_async_queue_pop(log->queue); + BlueSkyCloudLog *item + = (BlueSkyCloudLog *)g_async_queue_pop(log->queue); g_mutex_lock(item->lock); - writebuf(log->fd, item->key, strlen(item->key)); + g_assert(item->data != NULL); + + if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) { + g_mutex_unlock(item->lock); + bluesky_cloudlog_unref(item); + continue; + } + + item->pending_write |= CLOUDLOG_JOURNAL; + + off_t logsize = lseek(log->fd, 0, SEEK_CUR); + struct log_header header; + struct log_footer footer; + + header.magic = GUINT32_TO_LE(HEADER_MAGIC); + header.offset = GUINT64_TO_LE(logsize); + header.size = GUINT32_TO_LE(item->data->len); + header.id = item->id; + footer.magic = GUINT32_TO_LE(FOOTER_MAGIC); + + uint32_t crc = BLUESKY_CRC32C_SEED; + + writebuf(log->fd, (const char *)&header, sizeof(header)); + crc = crc32c(crc, (const char *)&header, sizeof(header)); + writebuf(log->fd, item->data->data, item->data->len); - fdatasync(log->fd); - item->committed = TRUE; - g_cond_signal(item->cond); + crc = crc32c(crc, item->data->data, item->data->len); + + crc = crc32c(crc, (const char *)&footer, + sizeof(footer) - sizeof(uint32_t)); + footer.crc = crc32c_finalize(crc); + writebuf(log->fd, (const char *)&footer, sizeof(footer)); + + item->log_seq = log->seq_num; + item->log_offset = logsize + sizeof(header); + item->log_size = item->data->len; + + logsize += sizeof(header) + sizeof(footer) + item->data->len; + + committed = g_slist_prepend(committed, item); g_mutex_unlock(item->lock); - off_t logsize = lseek(log->fd, 0, SEEK_CUR); + /* Force an fsync either if we will be closing this log segment and + * opening a new file, or if there are no other log items currently + * waiting to be written. */ + + if (logsize >= LOG_SEGMENT_SIZE + || g_async_queue_length(log->queue) <= 0) + { + int batchsize = 0; + fdatasync(log->fd); + while (committed != NULL) { + item = (BlueSkyCloudLog *)committed->data; + g_mutex_lock(item->lock); + item->pending_write &= ~CLOUDLOG_JOURNAL; + item->location_flags |= CLOUDLOG_JOURNAL; + g_cond_signal(item->cond); + g_mutex_unlock(item->lock); + committed = g_slist_delete_link(committed, committed); + batchsize++; + } + /* if (batchsize > 1) + g_print("Log batch size: %d\n", batchsize); */ + } + if (logsize < 0 || logsize >= LOG_SEGMENT_SIZE) { close(log->fd); log->fd = -1; @@ -123,36 +203,77 @@ BlueSkyLog *bluesky_log_new(const char *log_directory) return log; } -BlueSkyLogItem *bluesky_log_item_new() -{ - BlueSkyLogItem *item = g_new(BlueSkyLogItem, 1); - item->committed = FALSE; - item->lock = g_mutex_new(); - item->cond = g_cond_new(); - item->key = NULL; - item->data = NULL; - return item; -} - -void bluesky_log_item_submit(BlueSkyLogItem *item, BlueSkyLog *log) +void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log) { + bluesky_cloudlog_ref(item); g_async_queue_push(log->queue, item); } -static void bluesky_log_item_free(BlueSkyLogItem *item) +void bluesky_log_finish_all(GList *log_items) { - g_free(item->key); - bluesky_string_unref(item->data); - g_mutex_free(item->lock); - g_cond_free(item->cond); - g_free(item); + while (log_items != NULL) { + BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data; + + g_mutex_lock(item->lock); + while ((item->pending_write & CLOUDLOG_JOURNAL)) + g_cond_wait(item->cond, item->lock); + g_mutex_unlock(item->lock); + bluesky_cloudlog_unref(item); + + log_items = g_list_delete_link(log_items, log_items); + } } -void bluesky_log_item_finish(BlueSkyLogItem *item) +/* Memory-map the given log object into memory (read-only) and return a pointer + * to it. FIXME: Use some type of cache, map entire log segments, and use + * reference counting? */ +static int page_size = 0; + +BlueSkyRCStr *bluesky_log_map_object(BlueSkyLog *log, + int log_seq, int log_offset, int log_size) { - g_mutex_lock(item->lock); - while (!item->committed) - g_cond_wait(item->cond, item->lock); - g_mutex_unlock(item->lock); - bluesky_log_item_free(item); + if (page_size == 0) { + page_size = getpagesize(); + } + + int dirfd = open(log->log_directory, O_DIRECTORY); + if (dirfd < 0) { + fprintf(stderr, "Unable to open logging directory: %m\n"); + return NULL; + } + + char logfile[64]; + g_snprintf(logfile, sizeof(logfile), "log-%08d", log_seq); + int fd = openat(dirfd, logfile, O_RDONLY); + close(dirfd); + + if (fd < 0) { + fprintf(stderr, "Error opening logfile %s: %m\n", logfile); + return NULL; + } + + off_t off_start, off_end; + off_start = log_offset; + off_end = off_start + log_size; + off_start &= ~(page_size - 1); + off_end = (off_end + (page_size - 1)) & (page_size - 1); + + const char *ptr = (const char *)mmap(NULL, off_end - off_start, PROT_READ, + MAP_SHARED, fd, off_start); + + if (ptr == NULL) { + fprintf(stderr, "Error mapping logfile: %m\n"); + close(fd); + return NULL; + } + + close(fd); + + BlueSkyMmap *mmap = g_new0(BlueSkyMmap, 1); + mmap->addr = ptr; + mmap->len = off_end - off_start; + g_atomic_int_set(&mmap->refcount, 1); + + return bluesky_string_new_from_mmap(mmap, + log_offset - off_start, log_size); }