Re: [PATCH -next v2 3/6] landlock: add chmod and chown support

From: xiujianfeng
Date: Mon Nov 21 2022 - 08:49:16 EST


Hi,

在 2022/11/18 20:32, Mickaël Salaün 写道:

On 18/11/2022 10:03, xiujianfeng wrote:


在 2022/11/14 22:12, Mickaël Salaün 写道:

On 29/10/2022 10:33, xiujianfeng wrote:
Hi,

在 2022/9/2 1:34, Mickaël Salaün 写道:
CCing linux-fsdevel@xxxxxxxxxxxxxxx


On 01/09/2022 15:06, xiujianfeng wrote:
Hi,

在 2022/8/30 0:01, Mickaël Salaün 写道:

On 29/08/2022 03:17, xiujianfeng wrote:

Hi,

在 2022/8/28 3:30, Günther Noack 写道:
Hello!

the mapping between Landlock rights to LSM hooks is now as
follows in
your patch set:

* LANDLOCK_ACCESS_FS_CHMOD controls hook_path_chmod
* LANDLOCK_ACCESS_FS_CHGRP controls hook_path_chown
       (this hook can restrict both the chown(2) and chgrp(2)
syscalls)

Is this the desired mapping?

The previous discussion I found on the topic was in

[1]
https://lore.kernel.org/all/5873455f-fff9-618c-25b1-8b6a4ec94368@xxxxxxxxxxx/



[2]
https://lore.kernel.org/all/b1d69dfa-6d93-2034-7854-e2bc4017d20e@xxxxxxxxxxxxxxxx/



[3]
https://lore.kernel.org/all/c369c45d-5aa8-3e39-c7d6-b08b165495fd@xxxxxxxxxxx/




In my understanding the main arguments were the ones in [2] and [3].

There were no further responses to [3], so I was under the
impression
that we were gravitating towards an approach where the
file-metadata-modification operations were grouped more coarsely?

For example with the approach suggested in [3], which would be to
group the operations coarsely into (a) one Landlock right for
modifying file metadata that is used in security contexts, and
(b) one
Landlock right for modifying metadata that was used in non-security
contexts. That would mean that there would be:

(a) LANDLOCK_ACCESS_FS_MODIFY_SECURITY_ATTRIBUTES to control the
following operations:
       * chmod(2)-variants through hook_path_chmod,
       * chown(2)-variants and chgrp(2)-variants through
hook_path_chown,
       * setxattr(2)-variants and removexattr(2)-variants for
extended
         attributes that are not "user extended attributes" as
described in
         xattr(7) through hook_inode_setxattr and
hook_inode_removexattr

(b) LANDLOCK_ACCESS_FS_MODIFY_NON_SECURITY_ATTRIBUTES to control the
following operations:
       * utimes(2) and other operations for setting other
non-security
         sensitive attributes, probably through hook_inode_setattr(?)
       * xattr modifications like above, but for the "user extended
         attributes", though hook_inode_setxattr and
hook_inode_removexattr

In my mind, this would be a sensible grouping, and it would also
help
to decouple the userspace-exposed API from the underlying
implementation, as Casey suggested to do in [2].

Specifically for this patch set, if you want to use this
grouping, you
would only need to add one new Landlock right
(LANDLOCK_ACCESS_FS_MODIFY_SECURITY_ATTRIBUTES) as described above
under (a) (and maybe we can find a shorter name for it... :))?

Did I miss any operations here that would be necessary to restrict?

Would that make sense to you? Xiu, what is your opinion on how this
should be grouped? Do you have use cases in mind where a more
fine-grained grouping would be required?

I apologize I may missed that discussion when I prepared v2:(

Yes, agreed, this grouping is more sensible and resonnable. so in
this
patchset only one right will be added, and I suppose the first commit
which expand access_mask_t to u32 can be droped.


—Günther

P.S.: Regarding utimes: The hook_inode_setattr hook *also* gets
called
on a variety on attribute changes including file ownership, file
size
and file mode, so it might potentially interact with a bunch of
other
existing Landlock rights. Maybe that is not the right approach.
In any
case, it seems like it might require more thinking and it might be
sensible to do that in a separate patch set IMHO.

Thanks for you reminder, that seems it's more complicated to support
utimes, so I think we'd better not support it in this patchset.

The issue with this approach is that it makes it impossible to
properly
group such access rights. Indeed, to avoid inconsistencies and much
more
complexity, we cannot extend a Landlock access right once it is
defined.

We also need to consider that file ownership and permissions have a
default (e.g. umask), which is also a way to set them. How to
consistently manage that? What if the application wants to protect its
files with chmod 0400?

what do you mean by this? do you mean that we should have a set of
default permissions for files created by applications within the
sandbox, so that it can update metadata of its own file.

I mean that we need a consistent access control system, and for this we
need to consider all the ways an extended attribute can be set.

We can either extend the meaning of current access rights (controlled
with a ruleset flag for compatibility reasons), or create new access
rights. I think it would be better to add new dedicated rights to make
it more explicit and flexible.

I'm not sure about the right approach to properly control file
permission. We need to think about it. Do you have some ideas?

BTW, utimes can be controlled with the inode_setattr() LSM hook. Being
able to control arbitrary file time modification could be part of the
FS_WRITE_SAFE_METADATA, but modification and access time should always
be updated according to the file operation.




About the naming, I think we can start with:
- LANDLOCK_ACCESS_FS_READ_METADATA (read any file/dir metadata);
- LANDLOCK_ACCESS_FS_WRITE_SAFE_METADATA: change file times, user
xattr;

do you mean we should have permission controls on metadata level or
operation level? e.g. should we allow update on user xattr but deny
update on security xattr? or should we disallow update on any xattr?

- LANDLOCK_ACCESS_FS_WRITE_UNSAFE_METADATA: interpreted by the kernel
(could change non-Landlock DAC or MAC, which could be considered as a
policy bypass; or other various xattr that might be interpreted by
filesystems), this should be denied most of the time.

do you mean FS_WRITE_UNSAFE_METADATA is security-related? and
FS_WRITE_SAFE_METADATA is non-security-related?

Yes, FS_WRITE_UNSAFE_METADATA would be for security related
xattr/chmod/chown, and FS_WRITE_SAFE_METADATA for non-security xattr.
Both are mutually exclusive. This would involve the inode_setattr and
inode_setxattr LSM hooks. Looking at the calling sites, it seems
possible to replace all inode arguments with paths.

I though about differentiating user xattr, atime/mtime, DAC
(chown/chmod, posix ACLs), and other xattr, but it would be too complex
to get a consistent approach because of indirect consequences (e.g.
controlling umask, setegid, settimeofday…). Let's make it simple for now.

Here is an update on my previous proposal:

LANDLOCK_ACCESS_FS_READ_METADATA to read any file/dir metadata (i.e.
inode attr and xattr). In practice, for most use cases, this access
right should be granted whenever LANDLOCK_ACCESS_READ_DIR is allowed.

LANDLOCK_ACCESS_FS_WRITE_METADATA to *explicitly* write any inode attr
or xattr (i.e. chmod, chown, utime, and all xattr). It should be noted
that file modification time and access time should always be updated
according to the file operation (e.g. write, truncate) even when this
access is not explicitly allowed (according to vfs_utimes(),
ATTR_TIMES_SET and ATTR_TOUCH should enable to differentiate from
implicit time changes).

Thanks, I analyzed the relevant functions and the use of lsm hooks.
so I think what to do will be as follows:

LANDLOCK_ACCESS_FS_WRITE_METADATA controls the following hooks:
1.security_path_chmod
2.security_path_chown

These two chmod/chown hooks would be redundant with security_inode_setattr(). We then don't need to implement them.


3.security_inode_setattr
4.security_inode_setxattr
5.security_inode_removexattr > 6.security_inode_set_acl

Good catch. This new security_inode_set_acl hook is a good example of API refactoring. BTW, the related Cc list should be included in your next patch series.


LANDLOCK_ACCESS_FS_READ_METADATA controls the following hooks:
1.security_inode_getattr
2.security_inode_get_acl
3.security_inode_getxattr

Correct


and the following 7 hooks are using struct dentry * as parameter, should
be changed to struct path *, and also their callers.

security_inode_setattr
security_inode_setxattr
security_inode_removexattr
security_inode_set_acl
security_inode_getattr
security_inode_get_acl
security_inode_getxattr

Looks like it's a big change.

Your proposed approach looks good, and this will indeed touch a lot of files.

Because it interacts a lot with the filesystem subsystem, I propose to first write a set of patches that refactor the security_inode_*attr and security_inode_*_acl hooks to use struct file (or struct path when it makes sense) instead of struct dentry/inode (and to remove struct user_namespace as argument because it can be inferred thanks to file_mnt_user_ns). As for [1], using struct file only makes sense for a specific set of calls, and struct path should be used otherwise (e.g. syscalls dealing with file descriptors vs. with file paths).

You need to base this work on Christian's branch to be up-to-date with ongoing FS changes. I suggest to create one patch per function API change e.g., notify_change (merge the mnt_userns and dentry in a file argument), struct inode_operations.setattr (use a file argument instead of dentry)…


Thanks Mickaël, your advice is very clear, I will do it first.



Once this refactoring will be in -next, the landlock_file_security changes [1] will already be merged in master, and you will then be able to work on the Landlock specific parts with the new hooks.

[1] https://git.kernel.org/mic/c/b9f5ce27c8f8





Sorry for the late reply, I have problems with this work, for example,
before:
security_inode_setattr(struct user_namespace *mnt_userns,
                                            struct dentry *dentry,
                                            struct iattr *attr)
after:
security_inode_setattr(struct user_namespace *mnt_userns,
                                            struct path *path,
                                            struct iattr *attr)
then I change the second argument in notify_change() from struct *dentry
to struct path *, that makes this kind of changes in fs/overlayfs/
spread to lots of places because overlayfs basicly uses dentry instead
of path, the worst case may be here:

ovl_special_inode_operations.set_acl hook calls:
-->
ovl_set_acl(struct user_namespace *mnt_userns, struct dentry *dentry,
struct posix_acl *acl, int type)
-->
ovl_setattr(struct user_namespace *mnt_userns, struct dentry
*dentry,struct iattr *attr)
-->
ovl_do_notify_change(struct ovl_fs *ofs, struct dentry *upperdentry,
struct iattr *attr)

from the top of this callchain, I can not find a path to replace dentry,
did I miss something? or do you have better idea?

I think this can be solved thanks to the ovl_path_real() helper.
.
.