> -----邮件原件----- > 发件人: Amir Goldstein [mailto:amir73il@xxxxxxxxx] > 发送时间: 2024年7月16日 16:19 > 收件人: Lv Fei（吕飞） <feilv@xxxxxxxxxxxx> > 抄送: miklos@xxxxxxxxxx; overlayfs <linux-unionfs@xxxxxxxxxxxxxxx> > 主题: Re: overlayfs issue: dir permission lost during overlayfs copy-up > > > And for version 5.4.276, I need to add fsync at the end of ovl_copy_up_inode (correspond to latest function ovl_copy_up_metadata), right? > > > > Sounds right, but I do not have time to examine out-of-tree backport effort. > First, you need to provide a working and tested patch for upstream. > If that gets accepted, we can discuss backporting efforts. > But the general idea is this: > ovl_do_rename() serves as the "atomic" copy up operation everything that happens before that will not be observed after crash if rename did not happen. > > The problem is that inode metadata changes that happened before rename are NOT guaranteed by POSIX to be observed after crash even by rename is observe unless fsync() is called on the source inode before rename. > > Thanks, > Amir. OK, I finished a patch based on 6.10 today, will post for review. Thanks, Fei
