Let ImportantFileWriter Use fdatasync

base/files/file.h

Index: base/files/file.h
diff --git a/base/files/file.h b/base/files/file.h
index 13c8a96fcefe29ad89112946e705ae86e07790bd..4dcad33af2ccb683bae98f8c430af9552a43cd14 100644
--- a/base/files/file.h
+++ b/base/files/file.h
@@ -253,10 +253,15 @@ class BASE_EXPORT File {
   // doesn't exist, |false| is returned.
   bool SetLength(int64 length);
 
-  // Instructs the filesystem to flush the file to disk. (POSIX: fsync, Windows:
-  // FlushFileBuffers).
+  // Instructs the filesystem to flush the file data and metadata to disk.
+  // (POSIX: fsync, Windows: FlushFileBuffers)
   bool Flush();
 
+  // Instructs the filesystem to flush the file data to disk. Not necessarily
+  // flushes metadata like modification time.
+  // (Linux: fdatasync, other platforms: same as Flush())

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

I don't think we want to end up with two methods that do almost the same thing,
especially if we are not keeping the same behavior across all platforms* (which
is one of the points of having base).

(*) as in both methods doing the same thing on Windows.

We should probably move to fdatasync and be done with it (I doubt
FlushFileBuffers guarantees that metadata is also flushed to disk). If someone
actually requires all metadata to be written then something must be done at the
directory level.

[hashimoto, 2015/03/20 19:38:56]

Good point, I thought someone might be relying on Flush() to flush metadata, but
if we've not been guaranteeing that after all, it should be OK.
Just switching to fdatasync sounds good. (maybe it might be better to make a PSA
on chromium-dev about the change?)

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

PSA sounds fine.

+  bool FlushData();
+
   // Updates the file times.
   bool SetTimes(Time last_access_time, Time last_modified_time);