X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcache.c;h=d94c18e9f62631de7350100ee603823b9b76e0dc;hb=a82b60b3b683840a7074110831bcbaa16a40f0eb;hp=58b7cd6f1ac79c6b67af07fc4ece5979f872312d;hpb=e49fed25d000a51b316b480c1bfb1b00c1f5d345;p=bluesky.git

diff --git a/bluesky/cache.c b/bluesky/cache.c
index 58b7cd6..d94c18e 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:
@@ -22,40 +23,13 @@
  *   - 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;
 
     g_mutex_lock(fs->lock);
-    bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
-    inode->dirty_list = NULL;
+    bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
+    inode->unlogged_list = NULL;
     g_mutex_unlock(fs->lock);
 
     /* Inode is clean; nothing to do. */
@@ -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)
@@ -79,9 +64,9 @@ static void flushd_dirty(BlueSkyFS *fs)
 
     while (1) {
         BlueSkyInode *inode;
-        if (fs->dirty_list.prev == NULL)
+        if (fs->unlogged_list.prev == NULL)
             break;
-        inode = fs->dirty_list.prev->data;
+        inode = fs->unlogged_list.prev->data;
 
         if (bluesky_verbose) {
             g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
@@ -107,10 +92,90 @@ static void flushd_dirty(BlueSkyFS *fs)
         g_mutex_unlock(inode->lock);
         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.  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. */
+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. */
+    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);
+        bluesky_cloudlog_unref(log);
+
         g_mutex_lock(fs->lock);
     }
 
     g_mutex_unlock(fs->lock);
+    bluesky_cloudlog_flush(fs);
+
+    /* 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);
+
+        g_mutex_free(segment->lock);
+        g_cond_free(segment->cond);
+        g_free(segment);
+
+        fs->log_state->pending_segments
+            = g_list_delete_link(fs->log_state->pending_segments,
+                                 fs->log_state->pending_segments);
+    }
+
+    g_print("All segments have been flushed, journal < %d is clean\n",
+            journal_seq_start);
+
+    fs->log->journal_watermark = journal_seq_start;
 }
 
 /* Drop cached data for a given inode, if it is clean.  inode must be locked. */
@@ -118,9 +183,31 @@ static void drop_caches(BlueSkyInode *inode)
 {
     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 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);
@@ -130,16 +217,15 @@ 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,
@@ -175,8 +261,9 @@ static gpointer flushd_task(BlueSkyFS *fs)
     if (!g_mutex_trylock(fs->flushd_lock))
         return NULL;
     flushd_dirty(fs);
-    bluesky_cloudlog_write_log(fs);
+    flushd_cloud(fs);
     flushd_clean(fs);
+    bluesky_cachefile_gc(fs);
     g_mutex_unlock(fs->flushd_lock);
 
     return NULL;
@@ -189,18 +276,30 @@ 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. */