#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
+#include <sys/mman.h>
#include "bluesky-private.h"
// 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 << 23)
+#define LOG_SEGMENT_SIZE (1 << 24)
#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 keysize; // Size of the log key (bytes)
uint32_t size; // Size of the data item (bytes)
+ BlueSkyCloudID id; // Object identifier
} __attribute__((packed));
struct log_footer {
}
}
+static void log_commit(BlueSkyLog *log)
+{
+ int batchsize = 0;
+
+ if (log->fd < 0)
+ return;
+
+ fdatasync(log->fd);
+ 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);
+ bluesky_cloudlog_unref(item);
+ batchsize++;
+ }
+
+ if (bluesky_verbose && batchsize > 1)
+ g_print("Log batch size: %d\n", batchsize);
+}
+
+static gboolean log_open(BlueSkyLog *log)
+{
+ char logname[64];
+
+ if (log->fd >= 0) {
+ log_commit(log);
+ close(log->fd);
+ log->seq_num++;
+ log->fd = -1;
+ }
+
+ while (log->fd < 0) {
+ 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->seq_num++;
+ continue;
+ } else if (log->fd < 0) {
+ fprintf(stderr, "Error opening logfile %s: %m\n", logname);
+ return FALSE;
+ }
+ }
+
+ if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
+ fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
+ }
+ fsync(log->fd);
+ fsync(log->dirfd);
+ return TRUE;
+}
+
/* All log writes (at least for a single log) are made by one thread, so we
* don't need to worry about concurrent access to the log file. Log items to
* write are pulled off a queue (and so may be posted by any thread).
{
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");
- return NULL;
- }
-
while (TRUE) {
if (log->fd < 0) {
- char logfile[64];
- g_snprintf(logfile, sizeof(logfile), "log-%08d", log->seq_num);
- log->fd = openat(dirfd, logfile, O_CREAT|O_WRONLY|O_EXCL, 0600);
- if (log->fd < 0 && errno == EEXIST) {
- fprintf(stderr, "Log file %s already exists...\n", logfile);
- log->seq_num++;
- continue;
- } else if (log->fd < 0) {
- fprintf(stderr, "Error opening logfile %s: %m\n", logfile);
+ if (!log_open(log)) {
return NULL;
}
- fsync(log->fd);
- 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);
+ g_assert(item->data != NULL);
+
+ /* The item may have already been written to the journal... */
+ if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
+ g_mutex_unlock(item->lock);
+ bluesky_cloudlog_unref(item);
+ g_atomic_int_add(&item->data_lock_count, -1);
+ continue;
+ }
+
+ bluesky_cloudlog_stats_update(item, -1);
+ item->pending_write |= CLOUDLOG_JOURNAL;
+ bluesky_cloudlog_stats_update(item, 1);
- off_t logsize = lseek(log->fd, 0, SEEK_CUR);
struct log_header header;
struct log_footer footer;
+ size_t size = sizeof(header) + sizeof(footer) + item->data->len;
+ off_t offset = 0;
+ if (log->fd >= 0)
+ offset = lseek(log->fd, 0, SEEK_CUR);
+
+ /* Check whether the item would overflow the allocated journal size.
+ * If so, start a new log segment. We only allow oversized log
+ * segments if they contain a single log entry. */
+ if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
+ log_open(log);
+ offset = 0;
+ }
header.magic = GUINT32_TO_LE(HEADER_MAGIC);
- header.offset = GUINT64_TO_LE(logsize);
- header.keysize = GUINT32_TO_LE(strlen(item->key));
+ header.offset = GUINT64_TO_LE(offset);
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->key, strlen(item->key));
- crc = crc32c(crc, item->key, strlen(item->key));
-
writebuf(log->fd, item->data->data, item->data->len);
crc = crc32c(crc, item->data->data, item->data->len);
footer.crc = crc32c_finalize(crc);
writebuf(log->fd, (const char *)&footer, sizeof(footer));
- committed = g_slist_prepend(committed, item);
-
- /* 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. */
- logsize += strlen(item->key) + item->data->len + sizeof(header) + sizeof(footer);
-
- 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;
- 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); */
- }
+ item->log_seq = log->seq_num;
+ item->log_offset = offset + sizeof(header);
+ item->log_size = item->data->len;
- if (logsize < 0 || logsize >= LOG_SEGMENT_SIZE) {
- close(log->fd);
- log->fd = -1;
- log->seq_num++;
- }
+ offset += sizeof(header) + sizeof(footer) + item->data->len;
+
+ /* 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);
+
+ /* Force an if there are no other log items currently waiting to be
+ * written. */
+ if (g_async_queue_length(log->queue) <= 0)
+ log_commit(log);
}
return NULL;
log->fd = -1;
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->dirfd = open(log->log_directory, O_DIRECTORY);
+ if (log->dirfd < 0) {
+ fprintf(stderr, "Unable to open logging directory: %m\n");
+ return NULL;
+ }
g_thread_create(log_thread, log, FALSE, NULL);
return log;
}
-BlueSkyLogItem *bluesky_log_item_new()
+void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
{
- 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;
+ bluesky_cloudlog_ref(item);
+ g_atomic_int_add(&item->data_lock_count, 1);
+ g_async_queue_push(log->queue, item);
}
-void bluesky_log_item_submit(BlueSkyLogItem *item, BlueSkyLog *log)
+void bluesky_log_finish_all(GList *log_items)
{
- g_async_queue_push(log->queue, 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);
+ }
+}
+
+/* Memory-map the given log object into memory (read-only) and return a pointer
+ * to it. */
+static int page_size = 0;
+
+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);
}
-static void bluesky_log_item_free(BlueSkyLogItem *item)
+/* 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)
{
- g_free(item->key);
- bluesky_string_unref(item->data);
- g_mutex_free(item->lock);
- g_cond_free(item->cond);
- g_free(item);
+ if (page_size == 0) {
+ page_size = getpagesize();
+ }
+
+ BlueSkyLog *log = fs->log;
+
+ BlueSkyCacheFile *map;
+ g_mutex_lock(log->mmap_lock);
+ map = g_hash_table_lookup(log->mmap_cache, GINT_TO_POINTER(log_seq));
+
+ if (map == NULL) {
+ char *logname;
+
+ // A request for a local log file
+ if (clouddir < 0) {
+ logname = g_strdup_printf("journal-%08d", log_seq);
+ } else {
+ logname = g_strdup_printf("log-%08d-%08d", clouddir, log_seq);
+ }
+
+ /* TODO: stat() call */
+ map = g_new0(BlueSkyCacheFile, 1);
+ map->type = CLOUDLOG_JOURNAL;
+ map->lock = g_mutex_new();
+ g_mutex_lock(map->lock);
+ map->cond = g_cond_new();
+ map->filename = 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, GINT_TO_POINTER(log_seq), 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;
}
-void bluesky_log_item_finish(BlueSkyLogItem *item)
+BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir,
+ 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();
+ }
+
+ 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", map->filename);
+ bluesky_cachefile_unref(map);
+ g_mutex_unlock(map->lock);
+ return NULL;
+ }
+
+ off_t length = lseek(fd, 0, SEEK_END);
+ map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
+ fd, 0);
+ map->len = length;
+
+ g_print("Re-mapped log segment %d...\n", log_seq);
+ g_atomic_int_inc(&map->refcount);
+
+ close(fd);
+ }
+
+ g_mutex_unlock(log->mmap_lock);
+
+ 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(BlueSkyCacheFile *mmap)
+{
+ if (mmap == NULL)
+ 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_mutex_unlock(mmap->lock);
+ }
}