#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
+#include <sys/mman.h>
#include "bluesky-private.h"
// no absolute guarantees on the size of a log segment.
#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)
{
while (len > 0) {
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);
+ 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);
/* 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. */
- off_t logsize = lseek(log->fd, 0, SEEK_CUR);
+
if (logsize >= LOG_SEGMENT_SIZE
|| g_async_queue_length(log->queue) <= 0)
{
int batchsize = 0;
fdatasync(log->fd);
while (committed != NULL) {
- item = (BlueSkyLogItem *)committed->data;
- item->committed = TRUE;
+ 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);
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);
}