#include "bluesky-private.h"
#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);
}
-#if 0
-static void flushd_inode(gpointer value, gpointer user_data)
+/* Check whether memory usage may have dropped below critical thresholds for
+ * waking up waiting threads. */
+void flushd_check_wakeup(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);
- }
- bluesky_list_unlink(&inode->fs->accessed_list,
- inode->accessed_list);
- inode->accessed_list = NULL;
- bluesky_list_unlink(&inode->fs->dirty_list,
- inode->dirty_list);
- inode->dirty_list = NULL;
- g_mutex_unlock(fs->lock);
- }
- }
-
- g_mutex_unlock(inode->lock);
- bluesky_inode_unref(inode);
- return;
- }
-
- if (inode->change_pending) {
- /* Waiting for an earlier writeback to finish, so don't start a new
- * writeback yet. */
- g_mutex_unlock(inode->lock);
- bluesky_inode_unref(inode);
- return;
- }
-
- 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;
- }
+ int dirty = g_atomic_int_get(&fs->cache_dirty);
+ dirty += g_atomic_int_get(&fs->cache_log_dirty);
- inode->change_pending = inode->change_count;
-
- g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
- "Starting flush of inode %"PRIu64, inode->inum);
+ if (dirty <= bluesky_watermark_high_dirty)
+ g_cond_broadcast(fs->flushd_cond);
+}
- /* 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;
+/* Try to flush dirty data to disk, either due to memory pressure or due to
+ * timeouts. */
+static void flushd_dirty(BlueSkyFS *fs)
+{
+ int64_t start_time = bluesky_get_current_time();
+ g_mutex_lock(fs->lock);
- bluesky_inode_start_sync(inode, barrier);
+ while (1) {
+ BlueSkyInode *inode;
+ if (fs->unlogged_list.prev == NULL)
+ break;
+ inode = fs->unlogged_list.prev->data;
- bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
- bluesky_store_async_submit(barrier);
- bluesky_store_async_unref(barrier);
+ if (bluesky_verbose) {
+ g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+ "Considering flushing inode %"PRIu64, inode->inum);
+ }
- g_mutex_unlock(inode->lock);
- bluesky_inode_unref(inode);
-}
-#endif
+ /* 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;
-static void flushd_dirty_inode(BlueSkyInode *inode)
-{
- BlueSkyFS *fs = inode->fs;
+ bluesky_inode_ref(inode);
- g_mutex_lock(fs->lock);
- bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
- inode->dirty_list = NULL;
- g_mutex_unlock(fs->lock);
+ g_mutex_unlock(fs->lock);
- /* Inode is clean; nothing to do. */
- if (inode->change_count == inode->change_commit)
- return;
+ g_mutex_lock(inode->lock);
+ flushd_dirty_inode(inode);
+ g_mutex_unlock(inode->lock);
+ bluesky_inode_unref(inode);
- /* Inode writeback is in progress; put back on the dirty list. */
- if (inode->change_pending) {
- /* Waiting for an earlier writeback to finish, so don't start a new
- * writeback yet. */
g_mutex_lock(fs->lock);
- inode->change_time = bluesky_get_current_time();
- bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
- inode->dirty_list = bluesky_list_prepend(&fs->dirty_list, inode);
- g_mutex_unlock(fs->lock);
- return;
+ flushd_check_wakeup(fs);
}
- g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
- "Starting flush of inode %"PRIu64, inode->inum);
- inode->change_pending = inode->change_count;
-
- /* 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;
-
- bluesky_inode_start_sync(inode, barrier);
+ g_cond_broadcast(fs->flushd_cond);
- bluesky_store_async_add_notifier(barrier, writeback_complete, inode);
- bluesky_store_async_submit(barrier);
- bluesky_store_async_unref(barrier);
+ g_mutex_unlock(fs->lock);
}
-/* Try to flush dirty data to disk, either due to memory pressure or due to
- * timeouts. */
-static void flushd_dirty(BlueSkyFS *fs)
+/* Try to flush dirty data to the cloud.
+ * TODO: Rewrite this to work on cloud log items rather than inodes, so we can
+ * better track which logs are fully synchronized to the cloud and can be
+ * garbage collected if needed? */
+static void flushd_cloud(BlueSkyFS *fs)
{
int64_t start_time = bluesky_get_current_time();
g_mutex_lock(fs->lock);
break;
inode = fs->dirty_list.prev->data;
- g_print("Considering flushing inode %"PRIu64"\n", inode->inum);
+ if (bluesky_verbose) {
+ g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
+ "Flushing inode %"PRIu64" to cloud", 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
g_mutex_lock(inode->lock);
flushd_dirty_inode(inode);
+ 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;
+ bluesky_cloudlog_ref(log);
g_mutex_unlock(inode->lock);
+
+ if (log != NULL)
+ bluesky_cloudlog_serialize(log, fs);
bluesky_inode_unref(inode);
+ bluesky_cloudlog_unref(log);
g_mutex_lock(fs->lock);
}
g_mutex_unlock(fs->lock);
+ bluesky_cloudlog_flush(fs);
}
/* Drop cached data for a given inode, if it is clean. inode must be locked. */
{
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);
+ }
}
-/* Drop clean data fromt the cache if needed due to memory pressure. */
+/* 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)
{
g_mutex_lock(fs->lock);
inode_count = 1;
while (inode_count-- > 0) {
- if (g_atomic_int_get(&fs->cache_total) < bluesky_watermark_medium_total)
- break;
-
BlueSkyInode *inode;
if (fs->accessed_list.prev == NULL)
break;
inode = fs->accessed_list.prev->data;
- g_print("Considering dropping cached data for inode %"PRIu64"\n",
- inode->inum);
+ 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);
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_conditional(BlueSkyFS *fs)
{
- if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_high_dirty
- && g_atomic_int_get(&fs->cache_total) < bluesky_watermark_high_total)
+ if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty)
return;
- g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
- "Too much data; invoking flushd: dirty=%d total=%d",
- g_atomic_int_get(&fs->cache_dirty),
- g_atomic_int_get(&fs->cache_total));
+ 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);
}