os/bluestore: Added checks for compacting BlueFS log for flush and fsync ops

This fixes https://tracker.ceph.com/issues/46194

Signed-off-by: Adam Kupczyk <akupczyk@redhat.com>
(cherry picked from commit fab0c7148e84b1612e8e422e1876643a7cd4070b)

Conflicts:
	src/os/bluestore/BlueFS.cc
	src/os/bluestore/BlueFS.h

diff --git a/src/os/bluestore/BlueFS.cc b/src/os/bluestore/BlueFS.cc
index 3575953a56a7229be0eb49106cf197362b15736c..61ece25b2a102a639494e6e6b7de40bd668576db 100644 (file)
--- a/src/os/bluestore/BlueFS.cc
+++ b/src/os/bluestore/BlueFS.cc
@@ -2548,11 +2548,24 @@ void BlueFS::wait_for_aio(FileWriter *h)
 }
 #endif
 
-int BlueFS::_flush(FileWriter *h, bool force)
+int BlueFS::_flush(FileWriter *h, bool force, std::unique_lock<ceph::mutex>& l)
+{
+  bool flushed = false;
+  int r = _flush(h, force, &flushed);
+  if (r == 0 && flushed) {
+    _maybe_compact_log(l);
+  }
+  return r;
+}
+
+int BlueFS::_flush(FileWriter *h, bool force, bool *flushed)
 {
   h->buffer_appender.flush();
   uint64_t length = h->buffer.length();
   uint64_t offset = h->pos;
+  if (flushed) {
+    *flushed = false;
+  }
   if (!force &&
       length < cct->_conf->bluefs_min_flush_size) {
     dout(10) << __func__ << " " << h << " ignoring, length " << length
@@ -2569,7 +2582,11 @@ int BlueFS::_flush(FileWriter *h, bool force)
            << std::hex << offset << "~" << length << std::dec
           << " to " << h->file->fnode << dendl;
   ceph_assert(h->pos <= h->file->fnode.size);
-  return _flush_range(h, offset, length);
+  int r = _flush_range(h, offset, length);
+  if (flushed) {
+    *flushed = true;
+  }
+  return r;
 }
 
 int BlueFS::_truncate(FileWriter *h, uint64_t offset)
@@ -2857,8 +2874,14 @@ void BlueFS::sync_metadata(bool avoid_compact)
     dout(10) << __func__ << " done in " << (ceph_clock_now() - start) << dendl;
   }
 
-  if (!avoid_compact &&
-      !cct->_conf->bluefs_replay_recovery_disable_compact &&
+  if (!avoid_compact) {
+    _maybe_compact_log(l);
+  }
+}
+
+void BlueFS::_maybe_compact_log(std::unique_lock<ceph::mutex>& l)
+{
+  if (!cct->_conf->bluefs_replay_recovery_disable_compact &&
       _should_compact_log()) {
     if (cct->_conf->bluefs_compact_log_sync) {
       _compact_log_sync();

diff --git a/src/os/bluestore/BlueFS.h b/src/os/bluestore/BlueFS.h
index e564e031334f87e0ba2ba0a0e90a1ac1096ca740..86748e4afb0964c59d61c2b512484171969c926c 100644 (file)
--- a/src/os/bluestore/BlueFS.h
+++ b/src/os/bluestore/BlueFS.h
@@ -362,7 +362,8 @@ private:
                                 PExtentVector* extents);
 
   int _flush_range(FileWriter *h, uint64_t offset, uint64_t length);
-  int _flush(FileWriter *h, bool force);
+  int _flush(FileWriter *h, bool focce, std::unique_lock<ceph::mutex>& l);
+  int _flush(FileWriter *h, bool force, bool *flushed = nullptr);
   int _fsync(FileWriter *h, std::unique_lock<ceph::mutex>& l);
 
 #ifdef HAVE_LIBAIO
@@ -513,6 +514,8 @@ public:
 
   /// sync any uncommitted state to disk
   void sync_metadata(bool avoid_compact);
+  /// test and compact log, if necessary
+  void _maybe_compact_log(std::unique_lock<ceph::mutex>& l);
 
   void set_slow_device_expander(BlueFSDeviceExpander* a) {
     slow_dev_expander = a;
@@ -549,9 +552,10 @@ public:
   // handler for discard event
   void handle_discard(unsigned dev, interval_set<uint64_t>& to_release);
 
-  void flush(FileWriter *h) {
-    std::lock_guard l(lock);
-    _flush(h, false);
+  void flush(FileWriter *h, bool force = false) {
+    std::unique_lock l(lock);
+    int r = _flush(h, force, l);
+    ceph_assert(r == 0);
   }
   void flush_range(FileWriter *h, uint64_t offset, uint64_t length) {
     std::lock_guard l(lock);
@@ -559,7 +563,9 @@ public:
   }
   int fsync(FileWriter *h) {
     std::unique_lock l(lock);
-    return _fsync(h, l);
+    int r = _fsync(h, l);
+    _maybe_compact_log(l);
+    return r;
   }
   int read(FileReader *h, FileReaderBuffer *buf, uint64_t offset, size_t len,
           bufferlist *outbl, char *out) {