On Thu, Mar 30, 2017 at 1:23 PM, Richard Weinberger <richard@xxxxxx> wrote: > Am 30.03.2017 um 11:49 schrieb Richard Weinberger: >> Am 30.03.2017 um 11:32 schrieb Adrian Hunter: >>>> diff --git a/fs/ubifs/dir.c b/fs/ubifs/dir.c >>>> index 0858213a4e63..0139155045fe 100644 >>>> --- a/fs/ubifs/dir.c >>>> +++ b/fs/ubifs/dir.c >>>> @@ -748,6 +748,11 @@ static int ubifs_link(struct dentry *old_dentry, struct inode *dir, >>>> goto out_fname; >>>> >>>> lock_2_inodes(dir, inode); >>>> + >>>> + /* Handle O_TMPFILE corner case, it is allowed to link a O_TMPFILE. */ >>>> + if (inode->i_nlink == 0) >>>> + ubifs_delete_orphan(c, inode->i_ino); >>> >>> Isn't there also a deletion inode in the journal? If the recovery sees that >>> won't it delete the file data? >> >> Yes, but ubifs_link() adds a new journal entry which revives the inode. >> This should cancel out the deletion, right? >> You know the UBIFS journal better than I do. :-) > > Reading deeper into the proved that I was wrong. > AFAIKT UBIFS' journal has currently no way to revive a deleted inode. > So, we have to think about a new solution. > Not that I know anything about ubifs, but why do you need the deleted inode record in the first place for an O_TMPFILE. vfs ensures you that you can only link back an O_TMPFILE, not a deleted inode. It does not appear to be the right thing to do to pass deletion=1 to ubifs_jnl_update(), but deletion=0 doesn't look right as well..