Btrfs: fix rare chances for data loss when doing a fast fsync

After the simplification of the fast fsync patch done recently by commit
b5e6c3e1 ("btrfs: always wait on ordered extents at fsync time") and
commit e7175a69 ("btrfs: remove the wait ordered logic in the
log_one_extent path"), we got a very short time window where we can get
extents logged without writeback completing first or extents logged
without logging the respective data checksums. Both issues can only happen
when doing a non-full (fast) fsync.

As soon as we enter btrfs_sync_file() we trigger writeback, then lock the
inode and then wait for the writeback to complete before starting to log
the inode. However before we acquire the inode's lock and after we started
writeback, it's possible that more writes happened and dirtied more pages.
If that happened and those pages get writeback triggered while we are
logging the inode (for example, the VM subsystem triggering it due to
memory pressure, or another concurrent fsync), we end up seeing the
respective extent maps in the inode's list of modified extents and will
log matching file extent items without waiting for the respective
ordered extents to complete, meaning that either of the following will
happen:

1) We log an extent after its writeback finishes but before its checksums
   are added to the csum tree, leading to -EIO errors when attempting to
   read the extent after a log replay.

2) We log an extent before its writeback finishes.
   Therefore after the log replay we will have a file extent item pointing
   to an unwritten extent (and without the respective data checksums as
   well).

This could not happen before the fast fsync patch simplification, because
for any extent we found in the list of modified extents, we would wait for
its respective ordered extent to finish writeback or collect its checksums
for logging if it did not complete yet.

Fix this by triggering writeback again after acquiring the inode's lock
and before waiting for ordered extents to complete.

Fixes: e7175a69 ("btrfs: remove the wait ordered logic in the log_one_extent path")
Fixes: b5e6c3e1

 ("btrfs: always wait on ordered extents at fsync time")
CC: stable@vger.kernel.org # 4.19+
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>