This subject is not easy, but I think we're reaching a consensus (see my
3-steps proposal plan below). I answered your questions about the
(complex) interface I proposed, but we should focus on the first step
now (your initial proposal) and get back to the other steps later in
another email thread.
On 10/05/2023 21:21, Günther Noack wrote:
Hello Mickaël!
Sorry for the late reply. This was indeed a bit difficult for me to
understand; maybe it just needs more clarification.
Let me try to paraphrase your proposal (inline below) so we can
resolve the misunderstandings.
On Thu, May 04, 2023 at 11:12:00PM +0200, Mickaël Salaün wrote:
Thanks for this RFC, this is interesting!
I previously thought a lot about IOCTLs restrictions and here are some
notes:
IOCTLs are everywhere, from devices to filesystems (see fscrypt). Each
different file type may behave differently for the same IOCTL command/ID. It
is also worth noting that there are a lot of different IOCTLs, they are
growing over time, some might be dedicated to get some data (i.e. read) and
others to change the state of a device (i.e. write), some might be innocuous
(e.g. FIOCLEX, FIONCLEX) and others potentially harmful. _IOC_READ and
_IOC_WRITE can be useful but they are not mandatory, there are exceptions,
and it may be difficult to identify if a command pertains to one, the other,
or both kind of actions.
I then think it would be very useful to be able to tie file/device types to
a set of IOCTLs, letting user space libraries define and classify the IOCTL
semantic.
Instead of relying on a LANDLOCK_ACCESS_FS_IOCTL which would allow or deny
all IOCTLs, we can extend the path_beneath struct to manage IOCTLs in
addition to regular file accesses. Because dealing with a set of IOCTLs
would imply to deal with a lot of data and combinations, I thought about
creating groups of IOCTLs (defining access semantic) that could be matched
against file hierarchies. The composability nature of Landlock domains is
also another constraint to keep in mind.
// New rule type dedicated to define groups of IOCTLs.
struct landlock_ioctl_attr {
__u32 command; // IOCTL number/ID
dev_t device; // must be 0 for regular file and directory
__u8 file_type;
__u8 id_mask; // if 0, then applied globally
};
We could use landlock_add_rule(2) to fill a set of landlock_ioctl_attr into
a ruleset and use them with landlock_path_beneath_attr entries:
// LANDLOCK_RULE_PATH_BENEATH, leveraging the extensible design of
// landlock_path_beneath_attr, hence the same first fields.
struct landlock_path_beneath_attr {
__u64 allowed_access;
__s32 parent_fd;
__u16 allowed_ioctl_id_mask;
};
landlock_ioctl_attr includes a 8-bit mask for which each bit identifies a
set of allowed IOCTLs per device/file type. This mask is then tied to a
path_beneath_attr. We cannot use number IDs because of dev_t+IOCTL->ID
intersection conflicts. Using an id_mask enables to group (specific) IOCTLs
together, then creating synthetic access rights.
When merging a ruleset with a domain, each IOCTL ID mask is shifted and ORed
with the other layer ones to get a (8*16) 128-bit mask, stored in an
IOCTL/dev_t table and in the related landlock_object. When looking for an
IOCTL request, Landlock first looks into the IOCTL set ID table and get the
global set ID mask, which kind of translates to a composition of synthetic
access rights (stored with the landlock_layer.ioctl_access bitmask). We then
walk through all the inodes to match the whole mask.
I realize that this is complex and this explanation might be confusing
though. What do you think?
To be honest, I am not fully sure I understand the landlock_ioctl_attr
struct correctly.
My current guess is:
command: a single ioctl request number that should be permitted
device: if device != 0; require the dev_t (major+minor number)
of the file to match, before permitting the ioctl
file_type: if set (!= 0?), require the file type to match,
before permitting the ioctl
id_mask: Indicates the IDs of the groups where this ioctl
should be permitted(?)
So -- if we were to implement this without any optimizations -- the
logic of this is presumably something like this?:
If Landlock checks an ioctl with a request_id on a file f:
We look up the allowed_ioctl_id_mask and loop over the bits set in
that mask.
For every bit set in that mask, we look up the matching ioctl group.
Within the group, we look whether we can find a landlock_ioctl_attr
rule that belongs to the group which permits that ioctl request.
The request is granted by a landlock_ioctl_attr if:
attr.command == request_id
&& (attr.device == 0 || file_inode(file).i_rdev == attr.device)
&& (attr.file_type == 0 || landlock_get_file_type?(file))
Is this roughly what you imagined?
Yes :)
Some specific things I don't understand well are:
* How does id_mask identify the ioctl group? Do you envision an
interface where you can add a landlock_ioctl_attr to multiple groups
at once?
id_mask would be an arbitrary value picked by the user (library). I was
thinking about a bitmask (for allowed_ioctl_id_mask) because different
ioctl_attr could overlap (e.g., for different file hierarchy).
When the added landlock_ioctl_attr has id_mask=3, does it add that
attr to group 2 and 1?
No, it assigns this IOCTL attribute to the group 3, which can then be
matched against with allowed_ioctl_id_mask & (1 << 2).
* How does allowed_ioctl_id_mask match against the ioctl group IDs
(id_mask)? I'm particularly confused because the
allowed_ioctl_id_mask is __u16, whereas id_mask is __u8? Is this
intentional?
The idea was to assign each ioctl_attr to one group (ID) but enable to
match several groups with path_beneath_attr (allowed_ioctl_id_mask being
the only bitmask, whereas id_mask is a number). Yes, the naming is
confusing…
* What is file_type? Are those the file types as used in mknod(2),
so that you can distinguish between regular files, directories,
named pipes and the like?
Yes
* If I understand the proposal right, the .device and .file_type in
landlock_ioctl_attr are narrowing down the set of files that the
ioctl policy applies to. In all rules up until Landlock v3, which
we already have, the selection of the files which the rule applies
to is purely done based on file hierarchy.
Would it not be a more orthogonal API if the "file selection" part
of the Landlock API and the "policy adding" part for these selected
files were independent of each other? Then the .device and
.file_type selection scheme could be used for the existing policies
as well?
Both approaches have pros and cons. I propose a new incremental approach
below that starts with the simple case where there is no direct links
between different rule types (only the third step add that).
* When restricting by dev_t (major and minor number), aren't there use
cases where a system might have 256 CDROM drives, and you'd need to
allow-list all of these minor number combinations?
Indeed, we should be able to just ignore device minors.
* Aren't many ioctl use cases already usable with just the proposal I
made? If you add a rule to permit IOCTL for /dev/cdrom0, that
opened file will anyway only expose a small subset of the ioctls
that the kernel as a whole offers, no? Are there ioctls which are
offered independent of the file type which I'm overlooking? (Sorry,
this might be a naive question. :))
This is correct until users allow IOCTL for a directory (e.g. /etc). It
depends on use cases though.
On 02/05/2023 19:17, Günther Noack wrote:
Alternatives considered: Allow-listing specific ioctl requests
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
It would be technically possible to keep an allow-list of ioctl
requests and thereby control in more detail which ioctls should work.
I believe that this is not needed for the majority of use cases and
that it is a reasonable trade-off to make here (but I'm happy to hear
about counterexamples). The reasoning is:
* Many programs do not need ioctl at all,
and denying all ioctl(2) requests OK for these.
* Other programs need ioctl, but only for the terminal FDs.
This is supported because these file descriptors are usually
inherited from the parent process - so the parent process gets to
control the ioctl(2) policy for them.
* Some programs need ioctl on specific files that they are opening
themselves. They can allow-list these file paths for ioctl(2).
This makes the programs work, but it restricts a variety of other
ioctl requests which are otherwise possible through opening other
files.
Because the LANDLOCK_ACCESS_FS_IOCTL right is attached to the file
descriptor, programs have flexible options to control which ioctl
operations should work, without the implementation complexity of
additional ioctl allow-lists in the kernel.
Finally, the proposed approach is simpler in implementation and has
lower API complexity, but it does *not* preclude us from implementing
per-ioctl-request allow lists later, if that turns out to be necessary
at a later point.
I value this simplicity, but I'm also wondering about how much this
allow/deny all IOCTLs approach would be useful in real case scenarios. ;)
Mickaël, would you be open to gather some more data for this from
existing users, to understand better which use cases they have?
Of course, thanks!
I'm trying to project myself as an app developer, a sysadmin and a
desktop user. I know some sandboxed softwares, I sandboxed some, and I'm
looking into new ones. The main issue I see is (from the sysadmin and
user POV) when we want to manage a whole file hierarchy, which may
contain different file/device types.
(Looking in the direction of Jeff Xu, who has inquired about Landlock
for Chromium in the past -- do you happen to know in which ways you'd
want to restrict ioctls, if you have that need? :))
Related Work
~~~~~~~~~~~~
OpenBSD's pledge(2) [1] restricts ioctl(2) independent of the file
descriptor which is used. The implementers maintain multiple
allow-lists of predefined ioctl(2) operations required for different
application domains such as "audio", "bpf", "tty" and "inet".
OpenBSD does not guarantee ABI backwards compatibility to the same
extent as Linux does, so it's easier for them to update these lists in
later versions. It might not be a feasible approach for Linux though.
[1] https://man.openbsd.org/OpenBSD-7.3/pledge.2
Feedback I'm looking for
~~~~~~~~~~~~~~~~~~~~~~~~
Some specific points I would like to get your opinion on:
* Is this the right general approach to restricting ioctl(2)?
It will probably be possible to find counter-examples where the
alternative (see below) is better. I'd be interested in these, and
in how common they are, to understand whether we have picked the
right trade-off here.
In the long term, I'd like Landlock to be able to restrict a set of IOCTLs,
taking into account the type of file/device. Being able to deny all IOCTLs
might be useful and is much easier to implement though.
In my mind, it's a trade-off between implementation complexity and
flexibility.
My proposal is more simple-minded than what you explained, but might
solve the bulk of real-life use cases for a lower implementation
complexity. (With the caveat that I don't understand the real life use
cases well enough to know how far it really gets us, that's why this
is just an RFC.)
I'm not *just* saying this because I'm lazy ;), but I also feel that
we should be careful with the amount of complexity that we take on,
especially if there is a chance that it won't be needed in practice.
I agree :)
I think that it might be a feasible approach to start with the
LANDLOCK_ACCESS_FS_IOCTL approach and then look at its usage to
understand whether we see a significant number of programs whose
sandboxes are too coarse because of this.
If more fine-granular control is needed, we can still put the other
approach on top, and the additional complexity from
LANDLOCK_ACCESS_FS_IOCTL that we have to support is not that
dramatically high.
If more fine-granular control is not needed, we can skip the
implementation of the other approach and Landlock is simpler.
Then again, I'm somewhat new to kernel development still, I'm not sure
whether this is an approach that is deemed acceptable in this setting?
I agree that IOCTLs are a security risk and that we should propose a
simple solution short-term, and maybe a more complete one long-term.
The main issue with a unique IOCTL access right for a file hierarchy is
that we may not know which device/driver will be the target/server, and
hence if we need to allow some IOCTL for regular files (e.g., fscrypt),
we might end up allowing all IOCTLs.
Here is a plan to incrementally develop a fine-grained IOCTL access
control in 3 steps:
1/ Add a simple IOCTL access right for path_beneath: what you initially
proposed. For systems that already configure nodev mount points, it
could be even more useful (e.g., safely allow IOCTL on /home for
fscrypt, and specific /dev files otherwise).
2/ Create a new type of rule to identify file/device type:
struct landlock_inode_type_attr {
__u64 allowed_access; /* same as for path_beneath */
__u64 flags; /* bit 0: ignores device minor */
dev_t device; /* same as stat's st_rdev */
__u16 file_type; /* same as stat's st_mode & S_IFMT */
};
We'll probably need to differentiate the handled accesses for
path_beneath rules and those for inode_type rules to be more useful.
One issue with this type of rule is that it could be used as an oracle
to bypass stat restrictions. We could check if such (virtual) action is
allowed without the current domain though.
3/ Add a new type of rule to match IOCTL commands, with a mechanism to
tie this to inode_type rules (because a command ID is relative to a file
type/device), and potentially the same mechanism to tie inode_type rules
to path_beneath rules.
Each of this step can be implemented one after the other, and each one
is valuable. What do you think?
* Should we introduce a "landlock_fd_rights_limit()" syscall?
We could potentially implement a system call for dropping the
LANDLOCK_ACCESS_FS_IOCTL and LANDLOCK_ACCESS_FS_TRUNCATE rights from
existing file descriptors (independent of the type of file
descriptor, even).
Possible use cases would be to (a) restrict the rights on inherited
file descriptors like std{in,out,err} and to (b) restrict ioctl and
truncate operations on file descriptors that are not acquired
through open(2), such as network sockets.
This would be similar to the cap_rights_limit(2) system call in
FreeBSD's Capsicum.
This idea feels somewhat orthogonal to the ioctl patch, but it would
start to be more useful if the ioctl right exists.
This is indeed interesting, and that should be useful for the cases you
explained. I think supporting IOCTLs is more important for now, but a new
syscall to restrict FDs could be useful in the future.
Ack, OK. I agree, it's not that urgent yet.
We should also think about batch operations on FD (see the
close_range syscall), for instance to deny all IOCTLs on inherited
or received FDs.
Hm, you mean a landlock_fd_rights_limit_range() syscall to limit the
rights for an entire range of FDs?
I have to admit, I'm not familiar with the real-life use cases of
close_range(). In most programs I work with, it's difficult to reason
about their ordering once the program has really started to run. So I
imagine that close_range() is mostly used to "sanitize" the open file
descriptors at the start of main(), and you have a similar use case in
mind for this one as well?
If it's just about closing the range from 0 to 2, I'm not sure it's
worth adding a custom syscall. :)
The advantage of this kind of range is to efficiently manage all
potential FDs, and the main use case is to close (or change, see the
flags) everything *except" 0-2 (i.e. 3-~0), and then avoid a lot of
(potentially useless) syscalls.
The Landlock interface doesn't need to be a syscall. We could just add a
new rule type which could take a FD range and restrict them when calling
landlock_restrict_self(). Something like this:
struct landlock_fd_attr {
__u64 allowed_access;
__u32 first;
__u32 last;
}
Thanks for the review!
–Günther
