X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcache.c;h=2c1bc92c32a39f78ad639977c916923830caf613;hb=83fd6b61a6e092a22d4d5e59ed95f05f5e287f11;hp=d64cdbcb0710031ded621fdbd67614372f1ab0ba;hpb=ddaec40a37a5e65e53546b14632b1b0b35613264;p=bluesky.git

diff --git a/bluesky/cache.c b/bluesky/cache.c
index d64cdbc..2c1bc92 100644
--- a/bluesky/cache.c
+++ b/bluesky/cache.c
@@ -14,6 +14,7 @@
 #include "bluesky-private.h"
 
 #define WRITEBACK_DELAY (20 * 1000000)
+#define CACHE_DROP_DELAY (20 * 1000000)
 
 /* Filesystem caching and cache coherency.  There are actually a couple of
  * different tasks that are performed here:
@@ -22,33 +23,6 @@
  *   - Writing batches of data to the cloud.
  */
 
-#if 0
-static void writeback_complete(gpointer a, gpointer i)
-{
-    BlueSkyInode *inode = (BlueSkyInode *)i;
-
-    if (bluesky_verbose) {
-        g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
-              "Writeback for inode %"PRIu64" complete", inode->inum);
-    }
-
-    g_mutex_lock(inode->lock);
-
-    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);
-    }
-
-    g_mutex_unlock(inode->lock);
-}
-#endif
-
 static void flushd_dirty_inode(BlueSkyInode *inode)
 {
     BlueSkyFS *fs = inode->fs;
@@ -70,6 +44,17 @@ static void flushd_dirty_inode(BlueSkyInode *inode)
     bluesky_inode_start_sync(inode);
 }
 
+/* Check whether memory usage may have dropped below critical thresholds for
+ * waking up waiting threads. */
+void flushd_check_wakeup(BlueSkyFS *fs)
+{
+    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);
+}
+
 /* Try to flush dirty data to disk, either due to memory pressure or due to
  * timeouts. */
 static void flushd_dirty(BlueSkyFS *fs)
@@ -108,12 +93,18 @@ static void flushd_dirty(BlueSkyFS *fs)
         bluesky_inode_unref(inode);
 
         g_mutex_lock(fs->lock);
+        flushd_check_wakeup(fs);
     }
 
+    g_cond_broadcast(fs->flushd_cond);
+
     g_mutex_unlock(fs->lock);
 }
 
-/* Try to flush dirty data to the cloud. */
+/* 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();
@@ -158,7 +149,7 @@ static void flushd_cloud(BlueSkyFS *fs)
         if (log != NULL)
             bluesky_cloudlog_serialize(log, fs);
         bluesky_inode_unref(inode);
-        bluesky_cloudlog_ref(log);
+        bluesky_cloudlog_unref(log);
 
         g_mutex_lock(fs->lock);
     }
@@ -174,7 +165,11 @@ static void drop_caches(BlueSkyInode *inode)
         bluesky_file_drop_cached(inode);
 }
 
-/* Drop clean data from 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);
@@ -184,16 +179,15 @@ static void flushd_clean(BlueSkyFS *fs)
         inode_count = 1;
 
     while (inode_count-- > 0) {
-#if 0
-        if (g_atomic_int_get(&fs->cache_total) < bluesky_watermark_medium_total)
-            break;
-#endif
-
         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,
@@ -243,18 +237,30 @@ void bluesky_flushd_invoke(BlueSkyFS *fs)
 
 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;
 
     if (bluesky_verbose) {
         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));
+              "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. */