#include "bluesky-private.h"
-#define WRITEBACK_DELAY (5 * 1000000)
-#define CACHE_CLEAN_DELAY (30 * 1000000)
+#define WRITEBACK_DELAY (20 * 1000000)
+#define CACHE_DROP_DELAY (20 * 1000000)
-/* Filesystem caching and cache coherency. */
+/* Filesystem caching and cache coherency. There are actually a couple of
+ * different tasks that are performed here:
+ * - Forcing data to the log if needed to reclaim memory or simply if the
+ * data has been dirty in memory long enough.
+ * - Writing batches of data to the cloud.
+ */
-static void writeback_complete(gpointer a, gpointer i)
+static void flushd_dirty_inode(BlueSkyInode *inode)
{
- BlueSkyInode *inode = (BlueSkyInode *)i;
+ BlueSkyFS *fs = inode->fs;
- g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
- "Writeback for inode %"PRIu64" complete", inode->inum);
+ g_mutex_lock(fs->lock);
+ bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
+ inode->unlogged_list = NULL;
+ g_mutex_unlock(fs->lock);
- g_mutex_lock(inode->lock);
+ /* Inode is clean; nothing to do. */
+ if (inode->change_count == inode->change_commit)
+ return;
- inode->change_commit = inode->change_pending;
- inode->change_pending = 0;
- if (inode->change_count == inode->change_commit) {
- /* If inode is no longer dirty... */
- inode->change_time = 0;
- g_mutex_lock(inode->fs->lock);
- bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
- inode->dirty_list = NULL;
- g_mutex_unlock(inode->fs->lock);
+ if (bluesky_verbose) {
+ g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+ "Starting flush of inode %"PRIu64, inode->inum);
}
- g_mutex_unlock(inode->lock);
+ bluesky_inode_start_sync(inode);
}
-/* Drop cached data for a given inode, if it is clean. inode must be locked. */
-static void drop_caches(BlueSkyInode *inode)
+/* Check whether memory usage may have dropped below critical thresholds for
+ * waking up waiting threads. */
+void flushd_check_wakeup(BlueSkyFS *fs)
{
- if (inode->type == BLUESKY_REGULAR)
- bluesky_file_drop_cached(inode);
+ int dirty = g_atomic_int_get(&fs->cache_dirty);
+ dirty += g_atomic_int_get(&fs->cache_log_dirty);
+
+ if (dirty <= bluesky_watermark_high_dirty)
+ g_cond_broadcast(fs->flushd_cond);
}
-static void flushd_inode(gpointer value, gpointer user_data)
+/* Try to flush dirty data to disk, either due to memory pressure or due to
+ * timeouts. */
+static void flushd_dirty(BlueSkyFS *fs)
{
- BlueSkyFS *fs = (BlueSkyFS *)user_data;
-
- BlueSkyInode *inode = (BlueSkyInode *)value;
-
- g_mutex_lock(inode->lock);
-
- if (inode->change_count == inode->change_commit) {
- uint64_t delay = bluesky_get_current_time() - inode->access_time;
- if (delay >= CACHE_CLEAN_DELAY) {
- drop_caches(inode);
-
- /* If the only references are the one we hold and the one in the
- * filesystem inum->inode hash table... First check the refcount
- * without the lock for speed, but if the check looks good verify
- * it after taking the filesystem lock. */
- if (inode->refcount == 2) {
- g_mutex_lock(fs->lock);
- if (inode->refcount == 2) {
- g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
- "Trying to drop inode %"PRIu64" from cache",
- inode->inum);
- if (g_hash_table_remove(fs->inodes, &inode->inum))
- bluesky_inode_unref(inode);
- }
- g_mutex_unlock(fs->lock);
- }
+ int64_t start_time = bluesky_get_current_time();
+ g_mutex_lock(fs->lock);
+
+ while (1) {
+ BlueSkyInode *inode;
+ if (fs->unlogged_list.prev == NULL)
+ break;
+ inode = fs->unlogged_list.prev->data;
+
+ if (bluesky_verbose) {
+ g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+ "Considering flushing inode %"PRIu64, inode->inum);
}
+ /* Stop processing dirty inodes if we both have enough memory available
+ * and the oldest inode is sufficiently new that it need not be flushed
+ * out. */
+ uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
+ if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_low_dirty
+ && elapsed < WRITEBACK_DELAY)
+ break;
+ if (inode->change_time > start_time)
+ break;
+
+ bluesky_inode_ref(inode);
+
+ g_mutex_unlock(fs->lock);
+
+ g_mutex_lock(inode->lock);
+ flushd_dirty_inode(inode);
g_mutex_unlock(inode->lock);
bluesky_inode_unref(inode);
- return;
+
+ g_mutex_lock(fs->lock);
+ flushd_check_wakeup(fs);
}
- if (inode->change_pending) {
- /* Waiting for an earlier writeback to finish, so don't start a new
- * writeback yet. */
+ g_cond_broadcast(fs->flushd_cond);
+
+ g_mutex_unlock(fs->lock);
+}
+
+/* Try to flush dirty data to the cloud. This will take a snapshot of the
+ * entire filesystem (though only point-in-time consistent for isolated inodes
+ * and not the filesystem as a whole) and ensure all data is written to the
+ * cloud. When the write completes, we will allow old journal segments (those
+ * that were fully written _before_ the snapshot process started) to be garbage
+ * collected. Newer journal segments can't be collected yet since they may
+ * still contain data which has not been written persistently to the cloud.
+ *
+ * Note that some of this code relies on the fact that only this thread of
+ * control (running flushd_cloud) is manipulating the inode map, and so
+ * concurrent updates to the inode map are prevented even without the
+ * filesystem lock held. Take great care if allowing multi-threaded access to
+ * the inode map... */
+static void flushd_cloud(BlueSkyFS *fs)
+{
+ g_mutex_lock(fs->lock);
+
+ /* TODO: Locking? Since we're reading a single variable this is probably
+ * atomic but a lock could be safer. */
+ BlueSkyCloudLog *marker = bluesky_log_get_commit_point(fs);
+ int journal_seq_start = fs->log->seq_num;
+
+ while (1) {
+ BlueSkyInode *inode;
+ if (fs->dirty_list.prev == NULL)
+ break;
+ inode = fs->dirty_list.prev->data;
+
+ if (bluesky_verbose) {
+ g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+ "Flushing inode %"PRIu64" to cloud", inode->inum);
+ }
+
+ bluesky_inode_ref(inode);
+
+ g_mutex_unlock(fs->lock);
+
+ g_mutex_lock(inode->lock);
+ g_assert(inode->change_cloud == inode->change_commit);
+ g_mutex_lock(fs->lock);
+ bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
+ inode->dirty_list = NULL;
+ g_mutex_unlock(fs->lock);
+
+ BlueSkyCloudLog *log = inode->committed_item;
+ inode->committed_item = NULL;
g_mutex_unlock(inode->lock);
+
+ if (log != NULL)
+ bluesky_cloudlog_serialize(log, fs);
bluesky_inode_unref(inode);
- return;
+ bluesky_cloudlog_unref(log);
+
+ g_mutex_lock(fs->lock);
}
+ g_mutex_unlock(fs->lock);
- uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
- if (elapsed < WRITEBACK_DELAY) {
- /* Give a bit more time before starting writeback. */
- g_mutex_unlock(inode->lock);
- bluesky_inode_unref(inode);
- return;
+ /* Write out any updated inode map entries, so that all inodes just written
+ * can be located, and then a final commit record. */
+ BlueSkyCloudLog *commit_record = bluesky_inode_map_serialize(fs);
+ if (commit_record != NULL) {
+ bluesky_cloudlog_serialize(commit_record, fs);
+ } else {
+ g_print("No need for a checkpoint record...\n");
}
- inode->change_pending = inode->change_count;
+ bluesky_cloudlog_flush(fs);
- g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
- "Starting flush of inode %"PRIu64, inode->inum);
+ /* Wait until all segments have been written to the cloud, so that it
+ * becomes safe to free up journal segments. */
+ while (fs->log_state->pending_segments != NULL) {
+ SerializedRecord *segment
+ = (SerializedRecord *)fs->log_state->pending_segments->data;
+ g_mutex_lock(segment->lock);
+ while (!segment->complete)
+ g_cond_wait(segment->cond, segment->lock);
+ g_mutex_unlock(segment->lock);
- /* Create a store barrier. All operations part of the writeback will be
- * added to this barrier, so when the barrier completes we know that the
- * writeback is finished. */
- BlueSkyStoreAsync *barrier = bluesky_store_async_new(fs->store);
- barrier->op = STORE_OP_BARRIER;
+ g_mutex_free(segment->lock);
+ g_cond_free(segment->cond);
+ g_free(segment);
- bluesky_inode_start_sync(inode, barrier);
+ fs->log_state->pending_segments
+ = g_list_delete_link(fs->log_state->pending_segments,
+ fs->log_state->pending_segments);
+ }
- bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
- bluesky_store_async_submit(barrier);
- bluesky_store_async_unref(barrier);
+ bluesky_log_write_commit_point(fs, marker);
+ bluesky_cloudlog_unref(commit_record);
- g_mutex_unlock(inode->lock);
- bluesky_inode_unref(inode);
+ g_print("All segments have been flushed, journal < %d is clean\n",
+ journal_seq_start);
+
+ fs->log->journal_watermark = journal_seq_start;
}
-/* Scan through the cache for dirty data and start flushing it to stable
- * storage. This does not guarantee that data is committed when it returns.
- * Instead, this can be called occasionally to ensure that dirty data is
- * gradually flushed.
- *
- * We do not want to hold the filesystem lock while flushing individual inodes,
- * a that could lead to deadlock. So first scan through the inode table to get
- * a reference to all inodes, then process that queue of inodes after dropping
- * the filesystem lock. */
-static void gather_inodes(gpointer key, gpointer value, gpointer user_data)
+/* Drop cached data for a given inode, if it is clean. inode must be locked. */
+static void drop_caches(BlueSkyInode *inode)
{
- GSList **list = (GSList **)user_data;
- *list = g_slist_prepend(*list, value);
- bluesky_inode_ref((BlueSkyInode *)value);
+ if (inode->type == BLUESKY_REGULAR)
+ bluesky_file_drop_cached(inode);
+
+ BlueSkyCloudLog *log = inode->committed_item;
+ if (log != NULL) {
+ g_mutex_lock(log->lock);
+ if (log->data != NULL
+ && g_atomic_int_get(&log->data_lock_count) == 0
+ && (log->location_flags != 0))
+ {
+ bluesky_cloudlog_stats_update(log, -1);
+ bluesky_string_unref(log->data);
+ log->data = NULL;
+ bluesky_cloudlog_stats_update(log, 1);
+ }
+ if (log->location_flags & CLOUDLOG_CLOUD) {
+ log->location_flags &= ~CLOUDLOG_JOURNAL;
+ }
+ g_mutex_unlock(log->lock);
+ }
}
-void bluesky_flushd_invoke(BlueSkyFS *fs)
+/* Drop clean data from the cache if needed. Clean data should generally be
+ * memory-mapped from log file or similar, so the kernel can drop this clean
+ * data from memory for us and hence memory management isn't too important.
+ * Mainly, we'll want to drop references to data that hasn't been accessed in a
+ * while so that it is possible to reclaim log segments on disk. */
+static void flushd_clean(BlueSkyFS *fs)
{
- GSList *list = NULL;
-
g_mutex_lock(fs->lock);
- g_hash_table_foreach(fs->inodes, gather_inodes, &list);
+
+ size_t inode_count = g_hash_table_size(fs->inodes);
+ if (!inode_count)
+ inode_count = 1;
+
+ while (inode_count-- > 0) {
+ BlueSkyInode *inode;
+ if (fs->accessed_list.prev == NULL)
+ break;
+ inode = fs->accessed_list.prev->data;
+
+ uint64_t elapsed = bluesky_get_current_time() - inode->access_time;
+ if (elapsed < CACHE_DROP_DELAY)
+ break;
+
+ if (bluesky_verbose) {
+ g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+ "Considering dropping cached data for inode %"PRIu64,
+ inode->inum);
+ }
+
+ bluesky_inode_ref(inode);
+
+ g_mutex_unlock(fs->lock);
+
+ g_mutex_lock(inode->lock);
+
+ g_mutex_lock(fs->lock);
+ bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
+ inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
+ g_mutex_unlock(fs->lock);
+
+ drop_caches(inode);
+
+ g_mutex_unlock(inode->lock);
+ bluesky_inode_unref(inode);
+
+ g_mutex_lock(fs->lock);
+ }
+
g_mutex_unlock(fs->lock);
+}
+
+/* Run the flush daemon for a single iteration, though if it is already
+ * executing returns immediately. */
+static gpointer flushd_task(BlueSkyFS *fs)
+{
+ if (!g_mutex_trylock(fs->flushd_lock))
+ return NULL;
+ flushd_dirty(fs);
+ flushd_cloud(fs);
+ flushd_clean(fs);
+ bluesky_cachefile_gc(fs);
+ g_mutex_unlock(fs->flushd_lock);
+
+ return NULL;
+}
+
+void bluesky_flushd_invoke(BlueSkyFS *fs)
+{
+ g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL);
+}
- list = g_slist_reverse(list);
- g_slist_foreach(list, flushd_inode, fs);
+void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
+{
+ if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty)
+ return;
- g_slist_free(list);
+ if (bluesky_verbose) {
+ g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+ "Too much data; invoking flushd: dirty=%d",
+ g_atomic_int_get(&fs->cache_dirty));
+ }
+
+ bluesky_flushd_invoke(fs);
+
+ /* If the system is under heavy memory pressure, actually delay execution
+ * so the flush daemon can catch up. */
+ while (g_atomic_int_get(&fs->cache_dirty)
+ + g_atomic_int_get(&fs->cache_log_dirty)
+ > bluesky_watermark_high_dirty) {
+ g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+ "Waiting due to memory pressure, dirty=%d + %d",
+ g_atomic_int_get(&fs->cache_dirty),
+ g_atomic_int_get(&fs->cache_log_dirty));
+ g_mutex_lock(fs->lock);
+ g_cond_wait(fs->flushd_cond, fs->lock);
+ g_mutex_unlock(fs->lock);
+ }
+}
+
+/* Start a perpetually-running thread that flushes the cache occasionally. */
+static gpointer flushd_thread(BlueSkyFS *fs)
+{
+ while (TRUE) {
+ bluesky_flushd_invoke(fs);
+ struct timespec delay;
+ delay.tv_sec = 2;
+ delay.tv_nsec = 0;
+ nanosleep(&delay, NULL);
+ }
+
+ return NULL;
+}
+
+void bluesky_flushd_thread_launch(BlueSkyFS *fs)
+{
+ g_thread_create((GThreadFunc)flushd_thread, fs, FALSE, NULL);
}