os/bluestore/BlueFS: log dirty files at sync time

Instead of logging an updated fnode every time we dirty
the metadata, log it only once, right before we flush.
This greatly reduces the amount of metadata we have to
write.

Signed-off-by: Sage Weil <sage@redhat.com>

diff --git a/src/os/bluestore/BlueFS.cc b/src/os/bluestore/BlueFS.cc
index 92987eec98769bdd8b5dfda1d3a0e158ef3bdabb..46eaae253e36309069a6acefc083da82046aee9a 100644 (file)
--- a/src/os/bluestore/BlueFS.cc
+++ b/src/os/bluestore/BlueFS.cc
@@ -1279,14 +1279,26 @@ int BlueFS::_flush_and_sync_log(std::unique_lock<std::mutex>& l,
             << log_seq_stable << ", done" << dendl;
     return 0;
   }
-  if (log_t.empty()) {
+  if (log_t.empty() && dirty_files.empty()) {
     dout(10) << __func__ << " want_seq " << want_seq
-            << " " << log_t << " not dirty, no-op" << dendl;
+            << " " << log_t << " not dirty, dirty_files empty, no-op" << dendl;
     return 0;
   }
+
   uint64_t seq = log_t.seq = ++log_seq;
   assert(want_seq == 0 || want_seq <= seq);
   log_t.uuid = super.uuid;
+
+  // log dirty files
+  auto lsi = dirty_files.find(seq);
+  if (lsi != dirty_files.end()) {
+    dout(20) << __func__ << " " << lsi->second.size() << " dirty_files" << dendl;
+    for (auto &f : lsi->second) {
+      dout(20) << __func__ << "   op_file_update " << f.fnode << dendl;
+      log_t.op_file_update(f.fnode);
+    }
+  }
+
   dout(10) << __func__ << " " << log_t << dendl;
   assert(!log_t.empty());
 
@@ -1432,7 +1444,6 @@ int BlueFS::_flush_range(FileWriter *h, uint64_t offset, uint64_t length)
   if (must_dirty) {
     h->file->fnode.mtime = ceph_clock_now(NULL);
     assert(h->file->fnode.ino >= 1);
-    log_t.op_file_update(h->file->fnode);
     if (h->file->dirty_seq == 0) {
       h->file->dirty_seq = log_seq + 1;
       dirty_files[h->file->dirty_seq].push_back(*h->file);