On Mon, May 27, 2024 at 07:51:39AM +0800, Ian Kent wrote: > Indeed, that's what I found when I had a quick look. > > > Maybe a dentry (since that's part of the subject of the path walk and inode > is readily > > accessible) flag could be used since there's opportunity to set it in vfs > callbacks that > > are done as a matter of course.
You might recheck ->d_seq after fetching ->get_link there; with XFS ->get_link() unconditionlly failing in RCU mode that would prevent this particular problem. But it would obviously have to be done in pick_link() itself (and I refuse to touch that area in 5.4 - carrying those changes across the e.g. 5.6 changes in pathwalk machinery is too much). And it's really just the tip of the iceberg - e.g. I'd expect a massive headache in ACL-related part of permission checks, etc.
