Let ImportantFileWriter Use fdatasync

base/files/important_file_writer.cc

Index: base/files/important_file_writer.cc
diff --git a/base/files/important_file_writer.cc b/base/files/important_file_writer.cc
index 47b0b09e8938b1ef21ffe6e45e10a71aa97587e5..629ac948e38bf27129e120851e9db7de018d66c7 100644
--- a/base/files/important_file_writer.cc
+++ b/base/files/important_file_writer.cc
@@ -90,7 +90,7 @@ bool ImportantFileWriter::WriteFileAtomically(const FilePath& path,
   CHECK_LE(data.length(), static_cast<size_t>(kint32max));
   int bytes_written = tmp_file.Write(0, data.data(),
                                      static_cast<int>(data.length()));
-  bool flush_success = tmp_file.Flush();
+  bool flush_success = tmp_file.FlushData();

[rvargas (doing something else), 2015/03/20 18:56:05]

Incidentally, crbug.com/468046 is about doing _more_ work here instead of less.

[hashimoto, 2015/03/20 19:38:57]

Hmm, ImportantFileWriter's slowness is already a problem (crbug.com/469071).
Is it possible to do something to make it faster? (e.g. always keep 2 files to
avoid creating a temporary file every time, toggling between them when write
happens.)

[hashimoto, 2015/03/20 19:40:30]

Oops, wrong bug number.
crbug.com/418627 is the correct one.

[rvargas (doing something else), 2015/03/20 21:43:00]

Shouldn't we profile it from the field? The fact that it appears in a crash
(caused by a slow shutdown) doesn't really imply that the operation is too slow
(as in, having actual numbers will probably tell us more about what's going on).

On the other hand, while I really see no issue in ignoring some metadata (like
the times) when writing an "important" file, I see the issue in returning too
soon saying all's done only to find out after reboot that there's no file (or an
old one).

Before deciding to do something else to make this faster we need some data.

   tmp_file.Close();
 
   if (bytes_written < static_cast<int>(data.length())) {