Setting NFS4 ACLs on exported zfs dataset fails with EIO on linux-5.12 #12076
Description
System information
| Type | Version/Name |
|---|---|
| Distribution Name | Gentoo |
| Distribution Version | latest |
| Linux Kernel | 5.12.4 |
| Architecture | x86_64 |
| ZFS Version | 2.1.0_rc5 |
| SPL Version | 2.1.0_rc5 |
Describe the problem you're observing
Setting an nfs4 acl on an nfs client fails with EIO if the server is exporting from ZFS.
This started happening after upgrading Linux from 5.10 to 5.12, and ZFS to 2.1.0_rc5 alongside it.
I downgraded Linux back to 5.10 afterwards, keeping ZFS 2.1.0_rc5, and the problem disappeared.
I also tested an ext4 based NFS export on Linux 5.12 and could not observe the issue either.
So this specifically happens with ZFS on Linux 5.12.
Looking at strace on failing application on the NFS client, it's specifically the setxattr() syscall that returns EIO.
Though on the NFS server, the file in question does not have any xattrs, since NFS probably is translating the NFS ACL to a posix one.
Describe how to reproduce the problem
Run nfs4_editfacl on any file on the export on the NFS client, and make any random change and save. It will fail with EIO.
Can also be reproduced with pure python:
python3 -c 'import os, shutil; print(repr(os.listxattr("testfile"))); shutil.copystat("testfile", "testfile.copy")'
Should first print at least the system.nfs4_acl xattr, and then fail in shutil with EIO.