Use fdatasync for commits

We can use fdatasync to save 1 extra write per call, for a total of 2 writes
per commit, since we do two sync, one for data block up to the header, then
another after the header.

As of OTP 25 (our oldest supported version):
  * On Linux/BSDs: fdatasync()
  * On Window: FlushFileBuffers() i.e. the same as for file:sync/1
  * On MacOS: fcntl(fd,F_FULLFSYNC/F_BARRIERFSYNC)

According to https://linux.die.net/man/2/fdatasync

 > fdatasync() is similar to fsync(), but does not flush modified metadata
 unless that metadata is needed in order to allow a subsequent data retrieval
 to be correctly handled. For example, changes to st_atime or
 st_mtime (respectively, time of last access and time of last modification; see
 stat(2)) do not require flushing because they are not necessary for a
 subsequent data read to be handled correctly. On the other hand, a change to
 the file size (st_size, as made by say ftruncate(2)), would require a metadata
 flush.

The key things for us are:

  * It updates the size (positions) correctly
  * We do not rely or care about atime/mtime for safety or correctness
  * Erlang VM does the right thing on all the supported OSes

Author: nickva

src/couch/src/couch_file.erl

@@ -599,7 +599,12 @@ format_status(_Opt, [PDict, #file{} = File]) ->
 
 fsync(Fd) ->
     T0 = erlang:monotonic_time(),
-    Res = file:sync(Fd),
+    % We do not rely on mtime/atime for our safety/consitency so we can use
+    % fdatasync. As of version 25 OTP will use:
+    %  - On Linux/BSDs: fdatasync()
+    %  - On Window: FlushFileBuffers() i.e. the same as for file:sync/1
+    %  - On MacOS: fcntl(fd,F_FULLFSYNC/F_BARRIERFSYNC)
+    Res = file:datasync(Fd),
     T1 = erlang:monotonic_time(),
     % Since histograms can consume floating point values we can measure in
     % nanoseconds, then turn it into floating point milliseconds