* only incompletely written out before a crash, which should only happen for
* log records that were not considered committed). */
-// 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 << 22)
-
#define HEADER_MAGIC 0x676f4c0a
#define FOOTER_MAGIC 0x2e435243
+static size_t readbuf(int fd, char *buf, size_t len)
+{
+ size_t total_bytes = 0;
+ while (len > 0) {
+ ssize_t bytes = read(fd, buf, len);
+ if (bytes < 0 && errno == EINTR)
+ continue;
+ g_assert(bytes >= 0);
+ if (bytes == 0)
+ break;
+ buf += bytes;
+ len -= bytes;
+ }
+ return total_bytes;
+}
+
static void writebuf(int fd, const char *buf, size_t len)
{
while (len > 0) {
return;
fdatasync(log->fd);
+
+ /* Update disk-space usage statistics for the journal file. */
+ g_atomic_int_add(&log->disk_used, -log->current_log->disk_used);
+ struct stat statbuf;
+ if (fstat(log->fd, &statbuf) >= 0) {
+ /* Convert from 512-byte blocks to 1-kB units */
+ log->current_log->disk_used = (statbuf.st_blocks + 1) / 2;
+ }
+ g_atomic_int_add(&log->disk_used, log->current_log->disk_used);
+
while (log->committed != NULL) {
BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
g_mutex_lock(item->lock);
}
}
- log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
+ log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num,
+ FALSE);
g_assert(log->current_log != NULL);
g_mutex_unlock(log->current_log->lock);
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);
- async->data = bluesky_string_dup(async->data);
- bluesky_cloudlog_decrypt(async->data->data, async->data->len,
- cachefile->fs->keys, cachefile->items);
- 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);
-}
+static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile);
/* 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)
+ int clouddir, int log_seq,
+ gboolean start_fetch)
{
if (page_size == 0) {
page_size = getpagesize();
/* 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);
+ if (bluesky_verbose)
+ g_print("Adding cache file %s\n", logname);
map = g_new0(BlueSkyCacheFile, 1);
map->fs = fs;
g_mutex_lock(map->lock);
map->cond = g_cond_new();
map->filename = g_strdup(logname);
+ map->log_dir = clouddir;
map->log_seq = log_seq;
map->log = log;
g_atomic_int_set(&map->mapcount, 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);
+ int fd = openat(log->dirfd, logname, O_WRONLY | O_CREAT, 0600);
+ if (fd >= 0) {
+ ftruncate(fd, 5 << 20); // FIXME
+ close(fd);
+ }
} else {
g_mutex_lock(map->lock);
}
+
+ /* If the log file is stored in the cloud and has not been fully fetched,
+ * we may need to initiate a fetch now. */
+ if (clouddir >= 0 && start_fetch && !map->complete && !map->fetching)
+ cloudlog_fetch_start(map);
+
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)
+static void robust_pwrite(int fd, const char *buf, ssize_t count, off_t offset)
{
- if (page_size == 0) {
- page_size = getpagesize();
+ while (count > 0) {
+ ssize_t written = pwrite(fd, buf, count, offset);
+ if (written < 0) {
+ if (errno == EINTR)
+ continue;
+ g_warning("pwrite failure: %m");
+ return;
+ }
+ buf += written;
+ count -= written;
+ offset += written;
}
+}
- BlueSkyLog *log = fs->log;
- BlueSkyCacheFile *map = bluesky_cachefile_lookup(fs, log_dir, log_seq);
+static void cloudlog_partial_fetch_complete(BlueSkyStoreAsync *async,
+ BlueSkyCacheFile *cachefile);
+
+static void cloudlog_partial_fetch_start(BlueSkyCacheFile *cachefile,
+ size_t offset, size_t length)
+{
+ g_atomic_int_inc(&cachefile->refcount);
+ if (bluesky_verbose)
+ g_print("Starting partial fetch of %s from cloud (%zd + %zd)\n",
+ cachefile->filename, offset, length);
+ BlueSkyStoreAsync *async = bluesky_store_async_new(cachefile->fs->store);
+ async->op = STORE_OP_GET;
+ async->key = g_strdup(cachefile->filename);
+ async->start = offset;
+ async->len = length;
+ async->profile = bluesky_profile_get();
+ bluesky_store_async_add_notifier(async,
+ (GFunc)cloudlog_partial_fetch_complete,
+ cachefile);
+ bluesky_store_async_submit(async);
+ bluesky_store_async_unref(async);
+}
+
+static void cloudlog_partial_fetch_complete(BlueSkyStoreAsync *async,
+ BlueSkyCacheFile *cachefile)
+{
+ if (bluesky_verbose || async->result != 0)
+ g_print("Fetch of %s from cloud complete, status = %d\n",
+ async->key, async->result);
+
+ g_mutex_lock(cachefile->lock);
+ if (async->result >= 0) {
+ if (async->len == 0) {
+ if (bluesky_verbose)
+ g_print("Complete object was fetched.\n");
+ cachefile->complete = TRUE;
+ }
- if (map == NULL) {
+ /* Descrypt items fetched and write valid items out to the local log,
+ * but only if they do not overlap existing objects. This will protect
+ * against an attack by the cloud provider where one valid object is
+ * moved to another offset and used to overwrite data that we already
+ * have fetched. */
+ BlueSkyRangeset *items = bluesky_rangeset_new();
+ int fd = openat(cachefile->log->dirfd, cachefile->filename, O_WRONLY);
+ if (fd >= 0) {
+ gboolean allow_unauth;
+ async->data = bluesky_string_dup(async->data);
+ allow_unauth = cachefile->log_dir == BLUESKY_CLOUD_DIR_CLEANER;
+ bluesky_cloudlog_decrypt(async->data->data, async->data->len,
+ cachefile->fs->keys, items, allow_unauth);
+ uint64_t item_offset = 0;
+ while (TRUE) {
+ const BlueSkyRangesetItem *item;
+ item = bluesky_rangeset_lookup_next(items, item_offset);
+ if (item == NULL)
+ break;
+ if (bluesky_verbose) {
+ g_print(" item offset from range request: %d\n",
+ (int)(item->start + async->start));
+ }
+ if (bluesky_rangeset_insert(cachefile->items,
+ async->start + item->start,
+ item->length, item->data))
+ {
+ robust_pwrite(fd, async->data->data + item->start,
+ item->length, async->start + item->start);
+ } else {
+ g_print(" item overlaps existing data!\n");
+ }
+ item_offset = item->start + 1;
+ }
+ /* TODO: Iterate over items and merge into cached file. */
+ close(fd);
+ } else {
+ g_warning("Unable to open and write to cache file %s: %m",
+ cachefile->filename);
+ }
+
+ bluesky_rangeset_free(items);
+ } else {
+ g_print("Error fetching %s from cloud, retrying...\n", async->key);
+ cloudlog_partial_fetch_start(cachefile, async->start, async->len);
+ }
+
+ /* Update disk-space usage statistics, since the writes above may have
+ * consumed more space. */
+ g_atomic_int_add(&cachefile->log->disk_used, -cachefile->disk_used);
+ struct stat statbuf;
+ if (fstatat(cachefile->log->dirfd, cachefile->filename, &statbuf, 0) >= 0) {
+ /* Convert from 512-byte blocks to 1-kB units */
+ cachefile->disk_used = (statbuf.st_blocks + 1) / 2;
+ }
+ g_atomic_int_add(&cachefile->log->disk_used, cachefile->disk_used);
+
+ bluesky_cachefile_unref(cachefile);
+ g_cond_broadcast(cachefile->cond);
+ g_mutex_unlock(cachefile->lock);
+}
+
+static void cloudlog_fetch_start(BlueSkyCacheFile *cachefile)
+{
+ g_atomic_int_inc(&cachefile->refcount);
+ cachefile->fetching = TRUE;
+ if (bluesky_verbose)
+ 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);
+ async->profile = bluesky_profile_get();
+ bluesky_store_async_add_notifier(async,
+ (GFunc)cloudlog_partial_fetch_complete,
+ cachefile);
+ bluesky_store_async_submit(async);
+ bluesky_store_async_unref(async);
+}
+
+/* Map and return a read-only version of a byte range from a cached file. The
+ * CacheFile object must be locked. */
+BlueSkyRCStr *bluesky_cachefile_map_raw(BlueSkyCacheFile *cachefile,
+ off_t offset, size_t size)
+{
+ cachefile->atime = bluesky_get_current_time();
+
+ /* Easy case: the needed data is already in memory */
+ if (cachefile->addr != NULL && offset + size <= cachefile->len)
+ return bluesky_string_new_from_mmap(cachefile, offset, size);
+
+ int fd = openat(cachefile->log->dirfd, cachefile->filename, O_RDONLY);
+ if (fd < 0) {
+ fprintf(stderr, "Error opening logfile %s: %m\n",
+ cachefile->filename);
return NULL;
}
- /* Log segments fetched from the cloud might only be partially-fetched.
- * Check whether the object we are interested in is available. */
- if (log_dir >= 0) {
- if (!bluesky_rangeset_lookup(map->items, log_offset)) {
- g_warning("log-%d: Item at offset %d does not seem to be available\n", log_seq, log_offset);
- }
+ off_t length = lseek(fd, 0, SEEK_END);
+ if (offset + size > length) {
+ close(fd);
+ 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);
- }
+ /* File is not mapped in memory. Map the entire file in, then return a
+ * pointer to just the required data. */
+ if (cachefile->addr == NULL) {
+ cachefile->addr = (const char *)mmap(NULL, length, PROT_READ,
+ MAP_SHARED, fd, 0);
+ cachefile->len = length;
+ g_atomic_int_inc(&cachefile->refcount);
+
+ close(fd);
+ return bluesky_string_new_from_mmap(cachefile, offset, size);
+ }
+
+ /* Otherwise, the file was mapped in but doesn't cover the data we need.
+ * This shouldn't happen much, if at all, but if it does just read the data
+ * we need directly from the file. We lose memory-management benefits of
+ * using mmapped data, but otherwise this works. */
+ char *buf = g_malloc(size);
+ size_t actual_size = readbuf(fd, buf, size);
+ close(fd);
+ if (actual_size != size) {
+ g_free(buf);
+ return NULL;
+ } else {
+ return bluesky_string_new(buf, size);
+ }
+}
- int fd = openat(log->dirfd, map->filename, O_RDONLY);
+/* The arguments are mostly straightforward. log_dir is -1 for access from the
+ * journal, and non-negative for access to a cloud log segment. map_data
+ * should be TRUE for the case that are mapping just the data of an item where
+ * we have already parsed the item headers; this surpresses the error when the
+ * access is not to the first bytes of the item. */
+BlueSkyRCStr *bluesky_log_map_object(BlueSkyCloudLog *item, gboolean map_data)
+{
+ BlueSkyFS *fs = item->fs;
+ BlueSkyCacheFile *map = NULL;
+ BlueSkyRCStr *str = NULL;
+ int location = 0;
+ size_t file_offset = 0, file_size = 0;
+ gboolean range_request = bluesky_options.full_segment_fetches
+ ? FALSE : TRUE;
- if (fd < 0) {
- fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
- bluesky_cachefile_unref(map);
- g_mutex_unlock(map->lock);
- return NULL;
+ if (page_size == 0) {
+ page_size = getpagesize();
+ }
+
+ bluesky_cloudlog_stats_update(item, -1);
+
+ /* First, check to see if the journal still contains a copy of the item and
+ * if so use that. */
+ if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
+ map = bluesky_cachefile_lookup(fs, -1, item->log_seq, TRUE);
+ if (map != NULL) {
+ location = CLOUDLOG_JOURNAL;
+ file_offset = item->log_offset;
+ file_size = item->log_size;
}
+ }
- off_t length = lseek(fd, 0, SEEK_END);
- 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;
- g_atomic_int_add(&log->disk_used, map->len / 1024);
+ if (location == 0 && (item->location_flags & CLOUDLOG_CLOUD)) {
+ item->location_flags &= ~CLOUDLOG_JOURNAL;
+ map = bluesky_cachefile_lookup(fs,
+ item->location.directory,
+ item->location.sequence,
+ !range_request);
+ if (map == NULL) {
+ g_warning("Unable to remap cloud log segment!");
+ goto exit1;
+ }
+ location = CLOUDLOG_CLOUD;
+ file_offset = item->location.offset;
+ file_size = item->location.size;
+ }
- g_print("Re-mapped log segment %d...\n", log_seq);
- g_atomic_int_inc(&map->refcount);
+ /* Log segments fetched from the cloud might only be partially-fetched.
+ * Check whether the object we are interested in is available. */
+ if (location == CLOUDLOG_CLOUD) {
+ while (TRUE) {
+ const BlueSkyRangesetItem *rangeitem;
+ rangeitem = bluesky_rangeset_lookup(map->items, file_offset);
+ if (rangeitem != NULL && (rangeitem->start != file_offset
+ || rangeitem->length != file_size)) {
+ g_warning("log-%d: Item offset %zd seems to be invalid!",
+ (int)item->location.sequence, file_offset);
+ goto exit2;
+ }
+ if (rangeitem == NULL) {
+ if (bluesky_verbose) {
+ g_print("Item at offset 0x%zx not available, need to fetch.\n",
+ file_offset);
+ }
+ if (range_request) {
+ uint64_t start = file_offset, length = file_size, end;
+ if (map->prefetches != NULL)
+ bluesky_rangeset_get_extents(map->prefetches,
+ &start, &length);
+ start = MIN(start, file_offset);
+ end = MAX(start + length, file_offset + file_size);
+ length = end - start;
+ cloudlog_partial_fetch_start(map, start, length);
+ if (map->prefetches != NULL) {
+ bluesky_rangeset_free(map->prefetches);
+ map->prefetches = NULL;
+ }
+ }
+ g_cond_wait(map->cond, map->lock);
+ } else if (rangeitem->start == file_offset
+ && rangeitem->length == file_size) {
+ if (bluesky_verbose)
+ g_print("Item %zd now available.\n", file_offset);
+ break;
+ }
+ }
+ }
- close(fd);
+ if (map_data) {
+ if (location == CLOUDLOG_JOURNAL)
+ file_offset += sizeof(struct log_header);
+ else
+ file_offset += sizeof(struct cloudlog_header);
+
+ file_size = item->data_size;
}
+ str = bluesky_cachefile_map_raw(map, file_offset, file_size);
- BlueSkyRCStr *str;
- map->atime = bluesky_get_current_time();
- str = bluesky_string_new_from_mmap(map, log_offset, log_size);
+exit2:
bluesky_cachefile_unref(map);
g_mutex_unlock(map->lock);
+exit1:
+ bluesky_cloudlog_stats_update(item, 1);
return str;
}
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);
+ if (mmap->addr != NULL && g_atomic_int_get(&mmap->mapcount) == 0) {
+ if (bluesky_verbose)
+ 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);
}
+ g_assert(g_atomic_int_get(&mmap->mapcount) >= 0);
}
/******************************* JOURNAL REPLAY *******************************
projects / bluesky.git / blobdiff