On Tue, Mar 16, 2021 at 12:01:46PM -0400, Vivek Goyal wrote: > Hi Miklos, > > Please find attached a patch to fix the SGID clearing issue upon > ACL change. > > Luis reported that currently fstests generic/375 fails on virtiofs. And > reason being that we don't clear SGID when it should be. > > Setting ACL can lead to file mode change. And this in-turn also can > lead to clearing SGID bit if. > > - None of caller's groups match file owner group. > AND > - Caller does not have CAP_FSETID. > > Current implementation relies on server updating the mode. But file > server does not have enough information to do so. > > Initially I thought of sending CAP_FSETID information to server but > then I realized, it is just one of the pieces. What about all the > groups caller is a member of. If this has to work correctly, then > all the information will have to be sent to virtiofsd somehow. Just > sending CAP_FSETID information required adding V2 of fuse_setxattr_in > because we don't have any space for sending extra information. > > https://github.com/rhvgoyal/linux/commit/681cf5bdbba9c965c3dbd4337c16e9b17f27debe > > Also this approach will not work with idmapped mounts because server > does not have information about idmapped mappings. > > So I started to look at the approach of sending file mode updates > using SETATTR. As filesystems like 9pfs and ceph are doing. This > seems simpler approach. Though it has its issues too. > > - File mode update and setxattr(system.posix_acl_access) are not atomic. After reviewing (and testing) the patch, the only comment I have is that we should at least pr_warn() an eventual failure in setxattr(). But f that operation fails at that point, probably something went wrong on the other side and the kernel is unlikely to be able to revert the mode changes anyway. (And a nit: your patch seems to require some whitespaces clean-up.) Cheers, -- Luís > None of the approaches seem very clean to me. But sending SETATTR > explicitly seems to be lesser of two evils to me at this point of time. > Hence I am proposing this patch. > > I have run fstests acl tests and they pass. (./check -g acl). > > Corresponding virtiofsd patches are here. > > https://github.com/rhvgoyal/qemu/commits/acl-sgid-setattr > > What do you think. > > Vivek Goyal (1): > fuse: Add a mode where fuse client sends mode changes on ACL change > > fs/fuse/acl.c | 54 ++++++++++++++++++++++++++++++++++++--- > fs/fuse/dir.c | 11 ++++---- > fs/fuse/fuse_i.h | 9 ++++++- > fs/fuse/inode.c | 4 ++- > include/uapi/linux/fuse.h | 5 ++++ > 5 files changed, 71 insertions(+), 12 deletions(-) > > -- > 2.25.4 >
