lkml.org 
[lkml]   [2020]   [Sep]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v2 1/2] vfs: block chmod of symlinks
On Thu, Sep 17, 2020 at 05:15:03AM +0100, Al Viro wrote:
> On Thu, Sep 17, 2020 at 05:07:15AM +0100, Al Viro wrote:
> > On Wed, Sep 16, 2020 at 07:25:53AM +0100, Christoph Hellwig wrote:
> > > On Tue, Sep 15, 2020 at 08:22:54PM -0400, Rich Felker wrote:
> > > > It was discovered while implementing userspace emulation of fchmodat
> > > > AT_SYMLINK_NOFOLLOW (using O_PATH and procfs magic symlinks; otherwise
> > > > it's not possible to target symlinks with chmod operations) that some
> > > > filesystems erroneously allow access mode of symlinks to be changed,
> > > > but return failure with EOPNOTSUPP (see glibc issue #14578 and commit
> > > > a492b1e5ef). This inconsistency is non-conforming and wrong, and the
> > > > consensus seems to be that it was unintentional to allow link modes to
> > > > be changed in the first place.
> > > >
> > > > Signed-off-by: Rich Felker <dalias@libc.org>
> > > > ---
> > > > fs/open.c | 6 ++++++
> > > > 1 file changed, 6 insertions(+)
> > > >
> > > > diff --git a/fs/open.c b/fs/open.c
> > > > index 9af548fb841b..cdb7964aaa6e 100644
> > > > --- a/fs/open.c
> > > > +++ b/fs/open.c
> > > > @@ -570,6 +570,12 @@ int chmod_common(const struct path *path, umode_t mode)
> > > > struct iattr newattrs;
> > > > int error;
> > > >
> > > > + /* Block chmod from getting to fs layer. Ideally the fs would either
> > > > + * allow it or fail with EOPNOTSUPP, but some are buggy and return
> > > > + * an error but change the mode, which is non-conforming and wrong. */
> > > > + if (S_ISLNK(inode->i_mode))
> > > > + return -EOPNOTSUPP;
> > >
> > > Our usualy place for this would be setattr_prepare. Also the comment
> > > style is off, and I don't think we should talk about buggy file systems
> > > here, but a policy to not allow the chmod. I also suspect the right
> > > error value is EINVAL - EOPNOTSUPP isn't really used in normal posix
> > > file system interfaces.
> >
> > Er... Wasn't that an ACL-related crap? XFS calling posix_acl_chmod()
> > after it has committed to i_mode change, propagating the error to
> > caller of ->notify_change(), IIRC...
> >
> > Put it another way, why do we want
> > if (!inode->i_op->set_acl)
> > return -EOPNOTSUPP;
> > in posix_acl_chmod(), when we have
> > if (!IS_POSIXACL(inode))
> > return 0;
> > right next to it? If nothing else, make that
> > if (!IS_POSIXACL(inode) || !inode->i_op->get_acl)
> > return 0; // piss off - nothing to adjust here
>
> Arrgh... That'd break shmem and similar filesystems... Still, it
> feels like we should _not_ bother in cases when there's no ACL
> for that sucker; after all, if get_acl() returns NULL, we quietly
> return 0 and that's it.
>
> How about something like this instead?
>
> diff --git a/fs/posix_acl.c b/fs/posix_acl.c
> index 95882b3f5f62..2339160fabab 100644
> --- a/fs/posix_acl.c
> +++ b/fs/posix_acl.c
> @@ -559,8 +559,6 @@ posix_acl_chmod(struct inode *inode, umode_t mode)
>
> if (!IS_POSIXACL(inode))
> return 0;
> - if (!inode->i_op->set_acl)
> - return -EOPNOTSUPP;
>
> acl = get_acl(inode, ACL_TYPE_ACCESS);
> if (IS_ERR_OR_NULL(acl)) {
> @@ -569,6 +567,10 @@ posix_acl_chmod(struct inode *inode, umode_t mode)
> return PTR_ERR(acl);
> }
>
> + if (!inode->i_op->set_acl) {
> + posix_acl_release(acl);
> + return -EOPNOTSUPP;
> + }
> ret = __posix_acl_chmod(&acl, GFP_KERNEL, mode);
> if (ret)
> return ret;

Does this make chmod of links behave consistently (either succeeding
with return value 0, or returning -EOPNOTSUPP without doing anything)
for all filesystems? I'm fine with (and would probably prefer) this
fix if it's a complete one. If this goes in I think my patch 1/2 can
just be dropped and patch 2/2 behaves as intended; does that sound
correct to you?

Rich

\
 
 \ /
  Last update: 2020-09-17 21:09    [W:0.097 / U:3.656 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site