* TODO: Licensing
*/
+#define _GNU_SOURCE
+#define _ATFILE_SOURCE
+
+#include <stdio.h>
#include <stdint.h>
+#include <stdlib.h>
#include <glib.h>
#include <string.h>
+#include <errno.h>
#include <inttypes.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <unistd.h>
#include "bluesky-private.h"
journal_seq_start);
fs->log->journal_watermark = journal_seq_start;
+
+ bluesky_inode_map_minimize(fs);
}
/* Drop cached data for a given inode, if it is clean. inode must be locked. */
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);
}
}
* 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)
+ * while so that it is possible to reclaim log segments on disk.
+ *
+ * If aggressive is set, try much harder to drop data from the caches to free
+ * up space. */
+static void flushd_clean(BlueSkyFS *fs, int aggressive)
{
g_mutex_lock(fs->lock);
inode = fs->accessed_list.prev->data;
uint64_t elapsed = bluesky_get_current_time() - inode->access_time;
- if (elapsed < CACHE_DROP_DELAY)
+ if (elapsed < CACHE_DROP_DELAY && !aggressive)
break;
if (bluesky_verbose) {
g_mutex_unlock(fs->lock);
}
+/* Scan through all currently-stored files in the journal/cache and garbage
+ * collect old unused ones, if needed. */
+static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
+{
+ GList **files = (GList **)user_data;
+ *files = g_list_prepend(*files, value);
+}
+
+static gint compare_cachefiles(gconstpointer a, gconstpointer b)
+{
+ int64_t ta, tb;
+
+ ta = ((BlueSkyCacheFile *)a)->atime;
+ tb = ((BlueSkyCacheFile *)b)->atime;
+ if (ta < tb)
+ return -1;
+ else if (ta > tb)
+ return 1;
+ else
+ return 0;
+}
+
+void bluesky_cachefile_gc(BlueSkyFS *fs)
+{
+ GList *files = NULL;
+
+ g_mutex_lock(fs->log->mmap_lock);
+ g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
+
+ /* Sort based on atime. The atime should be stable since it shouln't be
+ * updated except by threads which can grab the mmap_lock, which we already
+ * hold. */
+ files = g_list_sort(files, compare_cachefiles);
+
+ /* Walk the list of files, starting with the oldest, deleting files if
+ * possible until enough space has been reclaimed. */
+ g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
+ while (files != NULL) {
+ BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
+ /* Try to lock the structure, but if the lock is held by another thread
+ * then we'll just skip the file on this pass. */
+ if (g_mutex_trylock(cachefile->lock)) {
+ int64_t age = bluesky_get_current_time() - cachefile->atime;
+ if (bluesky_verbose) {
+ g_print("%s addr=%p mapcount=%d refcount=%d size=%d atime_age=%f",
+ cachefile->filename, cachefile->addr, cachefile->mapcount,
+ cachefile->refcount, cachefile->disk_used, age / 1e6);
+ if (cachefile->fetching)
+ g_print(" (fetching)");
+ g_print("\n");
+ }
+
+ gboolean deletion_candidate = FALSE;
+ if (g_atomic_int_get(&fs->log->disk_used)
+ > bluesky_options.cache_size
+ && g_atomic_int_get(&cachefile->refcount) == 0
+ && g_atomic_int_get(&cachefile->mapcount) == 0)
+ {
+ deletion_candidate = TRUE;
+ }
+
+ /* Don't allow journal files to be reclaimed until all data is
+ * known to be durably stored in the cloud. */
+ if (cachefile->type == CLOUDLOG_JOURNAL
+ && cachefile->log_seq >= fs->log->journal_watermark)
+ {
+ deletion_candidate = FALSE;
+ }
+
+ if (deletion_candidate) {
+ if (bluesky_verbose) {
+ g_print(" ...deleting\n");
+ }
+ if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
+ fprintf(stderr, "Unable to unlink journal %s: %m\n",
+ cachefile->filename);
+ }
+
+ g_atomic_int_add(&fs->log->disk_used, -cachefile->disk_used);
+ g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
+ bluesky_rangeset_free(cachefile->items);
+ if (cachefile->prefetches != NULL)
+ bluesky_rangeset_free(cachefile->prefetches);
+ g_mutex_unlock(cachefile->lock);
+ g_mutex_free(cachefile->lock);
+ g_cond_free(cachefile->cond);
+ g_free(cachefile->filename);
+ g_free(cachefile);
+ } else {
+ g_mutex_unlock(cachefile->lock);
+ }
+ }
+ files = g_list_delete_link(files, files);
+ }
+ g_list_free(files);
+ g_print("\nEnding cache size: %d kB\n", fs->log->disk_used);
+
+ g_mutex_unlock(fs->log->mmap_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;
+
+ g_print("\nCloudlog cache: %d dirty, %d writeback, %d journal, %d cloud\n",
+ g_atomic_int_get(&fs->cache_log_dirty),
+ g_atomic_int_get(&fs->cache_log_writeback),
+ g_atomic_int_get(&fs->cache_log_journal),
+ g_atomic_int_get(&fs->cache_log_cloud));
+
flushd_dirty(fs);
flushd_cloud(fs);
- flushd_clean(fs);
+ flushd_clean(fs, 0);
bluesky_cachefile_gc(fs);
+
+ /* If running out of disk cache space, make another more aggressive pass to
+ * free up space. */
+ if (g_atomic_int_get(&fs->log->disk_used) > bluesky_options.cache_size) {
+ g_print("Still short on disk space, trying again to free space...\n");
+ flushd_clean(fs, 1);
+ bluesky_cachefile_gc(fs);
+ }
+
g_mutex_unlock(fs->flushd_lock);
return NULL;
g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL);
}
-void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
+/* How urgent is flushing out data? Returns one of several values:
+ * 0 - memory state is fine
+ * 1 - should launch flushd if not already running
+ * 2 - should block writers until memory frees up
+ */
+static int compute_pressure(BlueSkyFS *fs)
{
+ /* LEVEL 2 */
+ /* Too much dirty data in memory? */
+ if (g_atomic_int_get(&fs->cache_dirty)
+ + g_atomic_int_get(&fs->cache_log_dirty)
+ > bluesky_watermark_high_dirty)
+ return 2;
+
+ /* Too much uncommitted data in the journal on disk, not yet flushed to the
+ * cloud? */
+ printf("Dirty journals: %d to %d\n",
+ fs->log->journal_watermark, fs->log->seq_num);
+ int dirty_limit;
+ dirty_limit = bluesky_options.cache_size / (LOG_SEGMENT_SIZE / 1024) / 2;
+ int dirty_journals = fs->log->seq_num - fs->log->journal_watermark + 1;
+ if (dirty_journals > 1 && dirty_journals >= dirty_limit) {
+ printf("Too many dirty journals (%d >= %d)\n",
+ dirty_journals, dirty_limit);
+ return 2;
+ }
+
+ /* LEVEL 1 */
if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty)
+ return 1;
+
+ if (dirty_journals > 1 && dirty_journals > dirty_limit / 2) {
+ printf("Many dirty journals (%d), should start writeback\n",
+ dirty_journals);
+ return 1;
+ }
+
+ return 0;
+}
+
+void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
+{
+ int pressure = compute_pressure(fs);
+ if (pressure == 0)
return;
if (bluesky_verbose) {
/* 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) {
+ while (pressure > 1) {
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);
+ pressure = compute_pressure(fs);
+ if (pressure > 1)
+ g_cond_wait(fs->flushd_cond, fs->lock);
g_mutex_unlock(fs->lock);
+ pressure = compute_pressure(fs);
}
}