From: Michael Vrable <mvrable@cs.ucsd.edu>
Date: Thu, 12 Aug 2010 05:16:35 +0000 (-0700)
Subject: Attempt at limiting the rate at which memory is dirtied.
X-Git-Url: https://git.vrable.net/?a=commitdiff_plain;h=83fd6b61a6e092a22d4d5e59ed95f05f5e287f11;p=bluesky.git

Attempt at limiting the rate at which memory is dirtied.
---

diff --git a/bluesky/bluesky.h b/bluesky/bluesky.h
index 099aad4..820fa12 100644
--- a/bluesky/bluesky.h
+++ b/bluesky/bluesky.h
@@ -169,6 +169,9 @@ typedef struct {
     /* Mutex for the flush daemon, to prevent concurrent execution. */
     GMutex *flushd_lock;
 
+    /* Used to wait for the cache daemon to free up space */
+    GCond *flushd_cond;
+
     /* Mapping of object identifiers (blocks, inodes) to physical location (in
      * the local cache or in the logs in the cloud). */
     GHashTable *locations;
diff --git a/bluesky/cache.c b/bluesky/cache.c
index 001095b..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:
@@ -43,6 +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)
@@ -81,12 +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();
@@ -147,7 +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);
@@ -157,16 +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,
@@ -216,8 +237,7 @@ 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) {
@@ -227,6 +247,20 @@ void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
     }
 
     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. */
diff --git a/bluesky/inode.c b/bluesky/inode.c
index 39241ea..2696713 100644
--- a/bluesky/inode.c
+++ b/bluesky/inode.c
@@ -91,6 +91,7 @@ BlueSkyFS *bluesky_new_fs(gchar *name)
     fs->next_inum = BLUESKY_ROOT_INUM + 1;
     fs->store = bluesky_store_new("file");
     fs->flushd_lock = g_mutex_new();
+    fs->flushd_cond = g_cond_new();
     fs->locations = g_hash_table_new(bluesky_cloudlog_hash,
                                      bluesky_cloudlog_equal);