ioctl CAP_LINUX_IMMUTABLE is checked in the wrong namespace

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Hello,
when using user namespaces I found a bug in the capability checks done by ioctl.

If someone tries to use chattr +i while in a different user namespace it will get the following:

ioctl(3, EXT2_IOC_SETFLAGS, 0x7fffa4fedacc) = -1 EPERM (Operation not permitted)

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Quoted 3 lines · Marian Marinov

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

{
	return ns_capable(&init_user_ns, cap);
}

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where using init_user_ns would ever make sense.

No? Otherwise, pretty much every single use of capable() would be broken, not just this once instances in ext4/ioctl.c.

- Ted
_______________________________________________
Containers mailing list Containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
https://lists.linuxfoundation.org/mailman/listinfo/containers

Quoted 18 lines · Theodore Ts'o

Quoted 3 lines · Marian Marinov

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

Quoted 8 lines

{
	return ns_capable(&init_user_ns, cap);
}

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be safe to check against the userns owning the inode?

Quoted 8 lines

No? Otherwise, pretty much every single use of capable() would be broken, not just this once instances in ext4/ioctl.c.

- Ted
_______________________________________________
Containers mailing list Containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
https://lists.linuxfoundation.org/mailman/listinfo/containers

Quoted 35 lines · Serge Hallyn

Quoted 18 lines · Theodore Ts'o

Quoted 3 lines · Marian Marinov

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

Quoted 8 lines

{
	return ns_capable(&init_user_ns, cap);
}

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' by definition.

Quoted 3 lines

So NACK to that particular patch. I'm not sure, but IIUC it should be safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 40 lines · Marian Marinov

Quoted 38 lines · Serge Hallyn

Quoted 18 lines · Theodore Ts'o

Quoted 3 lines · Marian Marinov

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

Quoted 8 lines

{
	return ns_capable(&init_user_ns, cap);
}

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be safe to check against the userns owning the inode?

Quoted 4 lines

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Right, and I think the two operations you're looking at seem sane to allow.

thanks,
-serge

Quoted 44 lines · Serge Hallyn

Quoted 43 lines · Marian Marinov

Quoted 31 lines · Serge Hallyn

Quoted 18 lines · Theodore Ts'o

Quoted 3 lines · Marian Marinov

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

Quoted 8 lines

{
	return ns_capable(&init_user_ns, cap);
}

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

Quoted 7 lines

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

Quoted 4 lines

Quoted 1 line

I agree that this is more sane.

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Signed-off-by Marian Marinov <[email protected]>

---
 fs/ext4/ioctl.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

fs/ext4/ioctl.c +1 -2

index d011b69..9418634 100644
@@ -265,7 +265,7 @@ long ext4_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
                 * This test looks nicer. Thanks to Pauline Middelink
                 */
                if ((flags ^ oldflags) & (EXT4_APPEND_FL | EXT4_IMMUTABLE_FL)) {
-                   if (!capable(CAP_LINUX_IMMUTABLE))
+                 if (!inode_capable(inode, CAP_LINUX_IMMUTABLE))
                                goto flags_out;
                }
 
---

1.8.4

Marian

Quoted 4 lines

thanks,
-serge

Signature · 5 lines

Marian Marinov
Founder & CEO of 1H Ltd.
Jabber/GTalk: hackman-/[email protected]
ICQ: 7556201
Mobile: +359 886 660 270

Quoted 50 lines · Marian Marinov

Quoted 48 lines · Serge Hallyn

Quoted 36 lines · Marian Marinov

Quoted 35 lines · Serge Hallyn

Quoted 18 lines · Theodore Ts'o

Quoted 3 lines · Marian Marinov

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

Quoted 8 lines

{
	return ns_capable(&init_user_ns, cap);
}

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' by definition.

Quoted 8 lines

Quoted 3 lines

So NACK to that particular patch. I'm not sure, but IIUC it should be safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Right, and I think the two operations you're looking at seem sane to allow.

Quoted 1 line

If you are ok with this patch, I will fix all file systems and send patches.

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

Nacked-by Andy Lutomirski <[email protected]>

Quoted 7 lines · Andy Lutomirski

Wait, what?

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Quoted 5 lines

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

Nacked-by Andy Lutomirski <[email protected]>

Right, but you can't set a mapping in a child namespace unless you have CAP_SETUID in the parent namespace, right?

Otherwise user
namespaces are completely broken from a security perspective, since inode_capable() could never do the right thing.

Personally, reading how user namespaces work, it makes the hair rise on the back of my neck. I'm not sure the concept works at all from a security perspective, but hey, I'm not using user namespaces, and some fool thought it was worth merging. :-)

Quoted 51 lines · Serge Hallyn

Quoted 50 lines · Marian Marinov

Quoted 48 lines · Serge Hallyn

Quoted 36 lines · Marian Marinov

Quoted 35 lines · Serge Hallyn

Quoted 18 lines · Theodore Ts'o

Quoted 3 lines · Marian Marinov

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

Quoted 8 lines

{
	return ns_capable(&init_user_ns, cap);
}

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' by definition.

Quoted 8 lines

Quoted 3 lines

So NACK to that particular patch. I'm not sure, but IIUC it should be safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Right, and I think the two operations you're looking at seem sane to allow.

Quoted 7 lines

Quoted 1 line

If you are ok with this patch, I will fix all file systems and send patches.

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

Nacked-by Andy Lutomirski <[email protected]>

Quoted 88 lines · Andy Lutomirski

Quoted 75 lines · Serge Hallyn

Quoted 65 lines · Marian Marinov

Quoted 63 lines · Serge Hallyn

Quoted 59 lines · Marian Marinov

Quoted 42 lines · Serge Hallyn

Quoted 29 lines · Theodore Ts'o

Quoted 6 lines · Marian Marinov

I'm proposing a fix to this, by replacing the
capable(CAP_LINUX_IMMUTABLE)
check with ns_capable(current_cred()->user_ns,
CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior
capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior
capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is
available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

{
        return ns_capable(&init_user_ns, cap);
}

Quoted 6 lines

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where
using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

Quoted 9 lines

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns,
CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be
safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 6 lines

Quoted 2 lines

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Quoted 7 lines

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by: Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Quoted 7 lines

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

The problem I'm trying to solve is this:

container with its own user namespace and CAP_SYS_IMMUTABLE should be able to use chattr on all files witch this container has access to.

Unfortunately with the capable(CAP_SYS_IMMUTABLE) check this is not working.

With the proposed two fixes CAP_SYS_IMMUTABLE started working in the container.

The first solution got its user namespace from the currently running process and the second gets its user namespace from the currently opened inode.

So what would be the best solution in this case?

Quoted 105 lines · Marian Marinov

Quoted 88 lines · Andy Lutomirski

Quoted 75 lines · Serge Hallyn

Quoted 65 lines · Marian Marinov

Quoted 63 lines · Serge Hallyn

Quoted 59 lines · Marian Marinov

Quoted 42 lines · Serge Hallyn

Quoted 29 lines · Theodore Ts'o

Quoted 6 lines · Marian Marinov

I'm proposing a fix to this, by replacing the
capable(CAP_LINUX_IMMUTABLE)
check with ns_capable(current_cred()->user_ns,
CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior
capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior
capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is
available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

{
        return ns_capable(&init_user_ns, cap);
}

Quoted 6 lines

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where
using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

Quoted 9 lines

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns,
CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be
safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 6 lines

Quoted 2 lines

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Quoted 7 lines

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by: Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Quoted 7 lines

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

The problem I'm trying to solve is this:

container with its own user namespace and CAP_SYS_IMMUTABLE should be able to use chattr on all files witch this container has access to.

Unfortunately with the capable(CAP_SYS_IMMUTABLE) check this is not working.

Quoted 7 lines

With the proposed two fixes CAP_SYS_IMMUTABLE started working in the container.

The first solution got its user namespace from the currently running process and the second gets its user namespace from the currently opened inode.

So what would be the best solution in this case?

I'd suggest adding a mount option like fs_owner_uid that names a uid that owns, in the sense of having unlimited access to, a filesystem. Then anyone with caps on a namespace owned by that uid could do whatever.

Eric?

--Andy

Quoted 121 lines · Andy Lutomirski

Quoted 105 lines · Marian Marinov

Quoted 88 lines · Andy Lutomirski

Quoted 75 lines · Serge Hallyn

Quoted 65 lines · Marian Marinov

Quoted 63 lines · Serge Hallyn

Quoted 59 lines · Marian Marinov

Quoted 42 lines · Serge Hallyn

Quoted 29 lines · Theodore Ts'o

Quoted 6 lines · Marian Marinov

I'm proposing a fix to this, by replacing the
capable(CAP_LINUX_IMMUTABLE)
check with ns_capable(current_cred()->user_ns,
CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior
capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior
capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is
available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

{
        return ns_capable(&init_user_ns, cap);
}

Quoted 6 lines

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where
using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

Quoted 9 lines

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns,
CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be
safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 6 lines

Quoted 2 lines

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Quoted 7 lines

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by: Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Quoted 7 lines

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

The problem I'm trying to solve is this:

container with its own user namespace and CAP_SYS_IMMUTABLE should be able to use chattr on all files witch this container has access to.

Unfortunately with the capable(CAP_SYS_IMMUTABLE) check this is not working.

Quoted 7 lines

With the proposed two fixes CAP_SYS_IMMUTABLE started working in the container.

The first solution got its user namespace from the currently running process and the second gets its user namespace from the currently opened inode.

So what would be the best solution in this case?

I'd suggest adding a mount option like fs_owner_uid that names a uid that owns, in the sense of having unlimited access to, a filesystem. Then anyone with caps on a namespace owned by that uid could do whatever.

Eric?

Quoted 1 line

--Andy

The most obvious problem I can think of with "do whatever" is that this will likely include mknod of char and block devices which you can then chown/chmod as you wish and use to access any devices on the system from an unprivileged container.
This can however be mitigated by using the devices cgroup controller.

You also probably wouldn't want any unprivileged user from the host to find a way to access that mounted filesytem but so long as you do the mount in a separate mountns and don't share uids between the host and the container, that should be fine too.

Quoted 122 lines · Stéphane Graber

Quoted 121 lines · Andy Lutomirski

Quoted 105 lines · Marian Marinov

Quoted 88 lines · Andy Lutomirski

Quoted 75 lines · Serge Hallyn

Quoted 65 lines · Marian Marinov

Quoted 63 lines · Serge Hallyn

Quoted 59 lines · Marian Marinov

Quoted 42 lines · Serge Hallyn

Quoted 29 lines · Theodore Ts'o

Quoted 6 lines · Marian Marinov

I'm proposing a fix to this, by replacing the
capable(CAP_LINUX_IMMUTABLE)
check with ns_capable(current_cred()->user_ns,
CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior
capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior
capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is
available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

{
        return ns_capable(&init_user_ns, cap);
}

Quoted 6 lines

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where
using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

Quoted 9 lines

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns,
CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be
safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 6 lines

Quoted 2 lines

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Quoted 7 lines

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by: Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Quoted 7 lines

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

The problem I'm trying to solve is this:

container with its own user namespace and CAP_SYS_IMMUTABLE should be able to use chattr on all files witch this container has access to.

Unfortunately with the capable(CAP_SYS_IMMUTABLE) check this is not working.

Quoted 7 lines

With the proposed two fixes CAP_SYS_IMMUTABLE started working in the container.

The first solution got its user namespace from the currently running process and the second gets its user namespace from the currently opened inode.

So what would be the best solution in this case?

I'd suggest adding a mount option like fs_owner_uid that names a uid that owns, in the sense of having unlimited access to, a filesystem. Then anyone with caps on a namespace owned by that uid could do whatever.

Eric?

Quoted 7 lines

Quoted 1 line

--Andy

The most obvious problem I can think of with "do whatever" is that this will likely include mknod of char and block devices which you can then chown/chmod as you wish and use to access any devices on the system from an unprivileged container.
This can however be mitigated by using the devices cgroup controller.

Or 'nodev'. setuid/setgid may have the same problem, too.

Implementing something like this would also make CAP_DAC_READ_SEARCH and CAP_DAC_OVERRIDE work.

Arguably it should be impossible to mount such a thing in the first place without global privilege.

Quoted 5 lines

You also probably wouldn't want any unprivileged user from the host to find a way to access that mounted filesytem but so long as you do the mount in a separate mountns and don't share uids between the host and the container, that should be fine too.

This part should be a nonissue -- an unprivileged user who has the right uid owns the namespace anyway, so this is the least of your worries.

--Andy

Quoted 145 lines · Andy Lutomirski

Quoted 122 lines · Stéphane Graber

Quoted 121 lines · Andy Lutomirski

Quoted 105 lines · Marian Marinov

Quoted 88 lines · Andy Lutomirski

Quoted 75 lines · Serge Hallyn

Quoted 65 lines · Marian Marinov

Quoted 63 lines · Serge Hallyn

Quoted 59 lines · Marian Marinov

Quoted 42 lines · Serge Hallyn

Quoted 29 lines · Theodore Ts'o

Quoted 6 lines · Marian Marinov

I'm proposing a fix to this, by replacing the
capable(CAP_LINUX_IMMUTABLE)
check with ns_capable(current_cred()->user_ns,
CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior
capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior
capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is
available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

{
         return ns_capable(&init_user_ns, cap);
}

Quoted 6 lines

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where
using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

Quoted 9 lines

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns,
CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be
safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 6 lines

Quoted 2 lines

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Quoted 7 lines

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by: Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Quoted 7 lines

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

The problem I'm trying to solve is this:

container with its own user namespace and CAP_SYS_IMMUTABLE should be able to use chattr on all files witch this container has access to.

Unfortunately with the capable(CAP_SYS_IMMUTABLE) check this is not working.

Quoted 7 lines

With the proposed two fixes CAP_SYS_IMMUTABLE started working in the container.

The first solution got its user namespace from the currently running process and the second gets its user namespace from the currently opened inode.

So what would be the best solution in this case?

I'd suggest adding a mount option like fs_owner_uid that names a uid that owns, in the sense of having unlimited access to, a filesystem. Then anyone with caps on a namespace owned by that uid could do whatever.

Eric?

Quoted 7 lines

Quoted 1 line

--Andy

The most obvious problem I can think of with "do whatever" is that this will likely include mknod of char and block devices which you can then chown/chmod as you wish and use to access any devices on the system from an unprivileged container.
This can however be mitigated by using the devices cgroup controller.

Or 'nodev'. setuid/setgid may have the same problem, too.

Implementing something like this would also make CAP_DAC_READ_SEARCH and CAP_DAC_OVERRIDE work.

Arguably it should be impossible to mount such a thing in the first place without global privilege.

Quoted 5 lines

You also probably wouldn't want any unprivileged user from the host to find a way to access that mounted filesytem but so long as you do the mount in a separate mountns and don't share uids between the host and the container, that should be fine too.

Quoted 5 lines

This part should be a nonissue -- an unprivileged user who has the right uid owns the namespace anyway, so this is the least of your worries.

--Andy

It should be a nonissue so long as we make sure that a file owned by a uid outside the scope of the container may not be changed even though fs_owner_uid is set. Otherwise, it's just a matter of chmod +S on say a shell and anyone who can see the fs from the host will be getting a root shell (assuming said file is owned by the host's uid 0).

So that's restricting slightly what "do whatever" would do in this case.

Quoted 190 lines · Stéphane Graber

Quoted 181 lines · Andy Lutomirski

Quoted 153 lines · Stéphane Graber

Quoted 151 lines · Andy Lutomirski

Quoted 134 lines · Marian Marinov

Quoted 113 lines · Andy Lutomirski

Quoted 92 lines · Serge Hallyn

Quoted 79 lines · Marian Marinov

Quoted 76 lines · Serge Hallyn

Quoted 70 lines · Marian Marinov

Quoted 56 lines · Serge Hallyn

Quoted 40 lines · Theodore Ts'o

On Tue, Apr 29, 2014 at 04:49:14PM +0300, Marian Marinov wrote:

I'm proposing a fix to this, by replacing the
capable(CAP_LINUX_IMMUTABLE)
check with ns_capable(current_cred()->user_ns,
CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior
capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior
capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is
available on the
* assumption that it's about to be used.
*/
bool capable(int cap)
{
return ns_capable(&init_user_ns, cap);
}
EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and
I
can't imagine any use case, where user namespaces are in
effect,
where
using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the
subject.

The patch by Marian is wrong. Anyone can do
'clone(CLONE_NEWUSER)',
setuid(0), execve, and end up satisfying
'ns_capable(current_cred()->userns,
CAP_SYS_IMMUTABLE)' by definition.

Quoted 5 lines

So NACK to that particular patch. I'm not sure, but IIUC it should
be
safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 7 lines

Quoted 2 lines

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Quoted 8 lines

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by: Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Inodes aren't owned by user namespaces; they're owned by users. And any
user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Quoted 4 lines

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this
gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

The problem I'm trying to solve is this:

container with its own user namespace and CAP_SYS_IMMUTABLE should be able
to use chattr on all files witch this container has access to.

Unfortunately with the capable(CAP_SYS_IMMUTABLE) check this is not working.

With the proposed two fixes CAP_SYS_IMMUTABLE started working in the container.

Quoted 6 lines

The first solution got its user namespace from the currently running process
and the second gets its user namespace from the currently opened inode.

So what would be the best solution in this case?

I'd suggest adding a mount option like fs_owner_uid that names a uid that owns, in the sense of having unlimited access to, a filesystem. Then anyone with caps on a namespace owned by that uid could do whatever.

Eric?

Quoted 8 lines

Quoted 1 line

--Andy

The most obvious problem I can think of with "do whatever" is that this will likely include mknod of char and block devices which you can then chown/chmod as you wish and use to access any devices on the system from an unprivileged container.
This can however be mitigated by using the devices cgroup controller.

Or 'nodev'. setuid/setgid may have the same problem, too.

Implementing something like this would also make CAP_DAC_READ_SEARCH and CAP_DAC_OVERRIDE work.

Arguably it should be impossible to mount such a thing in the first place without global privilege.

Quoted 5 lines

You also probably wouldn't want any unprivileged user from the host to find a way to access that mounted filesytem but so long as you do the mount in a separate mountns and don't share uids between the host and the container, that should be fine too.

Quoted 5 lines

This part should be a nonissue -- an unprivileged user who has the right uid owns the namespace anyway, so this is the least of your worries.

--Andy

Quoted 8 lines

It should be a nonissue so long as we make sure that a file owned by a uid outside the scope of the container may not be changed even though fs_owner_uid is set. Otherwise, it's just a matter of chmod +S on say a shell and anyone who can see the fs from the host will be getting a root shell (assuming said file is owned by the host's uid 0).

So that's restricting slightly what "do whatever" would do in this case.

In my case I give an LVM volume to each container and limit the container to only this block device using the devices cgroup.
So the inode_capable() fix worked like a charm for me.
The container can not see any filesystem other then its own.
And I have another patch for my kernel that prohibits setns from cgroup other then / which prevents programs from the container from getting out. clone() can be used to create new namespaces but can not be used to attach to already created namespaces.

Quoted 149 lines · Andy Lutomirski

Quoted 122 lines · Stéphane Graber

Quoted 121 lines · Andy Lutomirski

Quoted 105 lines · Marian Marinov

Quoted 88 lines · Andy Lutomirski

Quoted 75 lines · Serge Hallyn

Quoted 65 lines · Marian Marinov

Quoted 63 lines · Serge Hallyn

Quoted 59 lines · Marian Marinov

Quoted 42 lines · Serge Hallyn

Quoted 29 lines · Theodore Ts'o

Quoted 6 lines · Marian Marinov

I'm proposing a fix to this, by replacing the
capable(CAP_LINUX_IMMUTABLE)
check with ns_capable(current_cred()->user_ns,
CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior
capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior
capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is
available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

{
        return ns_capable(&init_user_ns, cap);
}

Quoted 6 lines

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where
using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

Quoted 9 lines

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns,
CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be
safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 6 lines

Quoted 2 lines

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Quoted 7 lines

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by: Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Quoted 7 lines

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

The problem I'm trying to solve is this:

container with its own user namespace and CAP_SYS_IMMUTABLE should be able to use chattr on all files witch this container has access to.

Unfortunately with the capable(CAP_SYS_IMMUTABLE) check this is not working.

Quoted 7 lines

With the proposed two fixes CAP_SYS_IMMUTABLE started working in the container.

The first solution got its user namespace from the currently running process and the second gets its user namespace from the currently opened inode.

So what would be the best solution in this case?

I'd suggest adding a mount option like fs_owner_uid that names a uid that owns, in the sense of having unlimited access to, a filesystem. Then anyone with caps on a namespace owned by that uid could do whatever.

Eric?

Quoted 7 lines

Quoted 1 line

--Andy

The most obvious problem I can think of with "do whatever" is that this will likely include mknod of char and block devices which you can then chown/chmod as you wish and use to access any devices on the system from an unprivileged container.
This can however be mitigated by using the devices cgroup controller.

Or 'nodev'. setuid/setgid may have the same problem, too.

Implementing something like this would also make CAP_DAC_READ_SEARCH and CAP_DAC_OVERRIDE work.

Arguably it should be impossible to mount such a thing in the first place without global privilege.

Quoted 5 lines

You also probably wouldn't want any unprivileged user from the host to find a way to access that mounted filesytem but so long as you do the mount in a separate mountns and don't share uids between the host and the container, that should be fine too.

This part should be a nonissue -- an unprivileged user who has the right uid owns the namespace anyway, so this is the least of your worries.

Quoted 1 line

--Andy

It should be a nonissue so long as we make sure that a file owned by a uid outside the scope of the container may not be changed even though fs_owner_uid is set. Otherwise, it's just a matter of chmod +S on say a shell and anyone who can see the fs from the host will be getting a root shell (assuming said file is owned by the host's uid 0).

Quoted 152 lines · Stéphane Graber

Quoted 150 lines · Andy Lutomirski

Quoted 122 lines · Stéphane Graber

Quoted 121 lines · Andy Lutomirski

Quoted 105 lines · Marian Marinov

Quoted 88 lines · Andy Lutomirski

Quoted 75 lines · Serge Hallyn

Quoted 65 lines · Marian Marinov

Quoted 63 lines · Serge Hallyn

Quoted 59 lines · Marian Marinov

Quoted 42 lines · Serge Hallyn

Quoted 35 lines · Theodore Ts'o

On Tue, Apr 29, 2014 at 04:49:14PM +0300, Marian Marinov wrote:

I'm proposing a fix to this, by replacing the
capable(CAP_LINUX_IMMUTABLE)
check with ns_capable(current_cred()->user_ns,
CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior
capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior
capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is
available on the
* assumption that it's about to be used.
*/
bool capable(int cap)
{
return ns_capable(&init_user_ns, cap);
}
EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I
can't imagine any use case, where user namespaces are in effect,
where
using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

Quoted 9 lines

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns,
CAP_SYS_IMMUTABLE)' by definition.

So NACK to that particular patch. I'm not sure, but IIUC it should be
safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Quoted 6 lines

Quoted 2 lines

Right, and I think the two operations you're looking at seem sane to allow.

If you are ok with this patch, I will fix all file systems and send patches.

Quoted 7 lines

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by: Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Quoted 7 lines

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

The problem I'm trying to solve is this:

container with its own user namespace and CAP_SYS_IMMUTABLE should be able to use chattr on all files witch this container has access to.

Unfortunately with the capable(CAP_SYS_IMMUTABLE) check this is not working.

Quoted 7 lines

With the proposed two fixes CAP_SYS_IMMUTABLE started working in the container.

The first solution got its user namespace from the currently running process and the second gets its user namespace from the currently opened inode.

So what would be the best solution in this case?

I'd suggest adding a mount option like fs_owner_uid that names a uid that owns, in the sense of having unlimited access to, a filesystem. Then anyone with caps on a namespace owned by that uid could do whatever.

Eric?

Quoted 7 lines

Quoted 1 line

--Andy

The most obvious problem I can think of with "do whatever" is that this will likely include mknod of char and block devices which you can then chown/chmod as you wish and use to access any devices on the system from an unprivileged container.
This can however be mitigated by using the devices cgroup controller.

Or 'nodev'. setuid/setgid may have the same problem, too.

Implementing something like this would also make CAP_DAC_READ_SEARCH and CAP_DAC_OVERRIDE work.

Arguably it should be impossible to mount such a thing in the first place without global privilege.

Quoted 5 lines

You also probably wouldn't want any unprivileged user from the host to find a way to access that mounted filesytem but so long as you do the mount in a separate mountns and don't share uids between the host and the container, that should be fine too.

This part should be a nonissue -- an unprivileged user who has the right uid owns the namespace anyway, so this is the least of your worries.

--Andy

Quoted 5 lines

It should be a nonissue so long as we make sure that a file owned by a uid outside the scope of the container may not be changed even though fs_owner_uid is set. Otherwise, it's just a matter of chmod +S on say a shell and anyone who can see the fs from the host will be getting a root shell (assuming said file is owned by the host's uid 0).

I feel like that's too fragile. I'd rather add a rule that one of

these filesystems always acts like it's nosuid unless you're inside a user namespace that matches fs_owner_uid.

Maybe even that is too weird. How about setuid, setgid, and fcaps only work on mounts that are in mount namespaces that are owned by the current user namespace or one of its parents? IOW, a struct mount is only trusted if mnt->mnt_ns->user_ns == current user ns or one of its parents?

Untrusted mounts would act like they are nosuid,nodev. Someone can try to figure out a safe way to relax nodev at some point.

--Andy
_______________________________________________
Containers mailing list Containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
https://lists.linuxfoundation.org/mailman/listinfo/containers

Quoted 6 lines · Andy Lutomirski

Quoted 5 lines

It should be a nonissue so long as we make sure that a file owned by a uid outside the scope of the container may not be changed even though fs_owner_uid is set. Otherwise, it's just a matter of chmod +S on say a shell and anyone who can see the fs from the host will be getting a root shell (assuming said file is owned by the host's uid 0).

Quoted 1 line

I feel like that's too fragile. I'd rather add a rule that one of

yeah I don't wnat to rush something like that. I'd rather stash the userns of the task which did the mounting and check against that. Note that would make it worthless unless and until we allowed mounting from non-init userns, but then we can only claim "our fs superblock readers suck and therefore containers can't mount an fs" so long before we start to feel some shame and audit them...

Quoted 11 lines

these filesystems always acts like it's nosuid unless you're inside a user namespace that matches fs_owner_uid.

Maybe even that is too weird. How about setuid, setgid, and fcaps only work on mounts that are in mount namespaces that are owned by the current user namespace or one of its parents? IOW, a struct mount is only trusted if mnt->mnt_ns->user_ns == current user ns or one of its parents?

Untrusted mounts would act like they are nosuid,nodev. Someone can try to figure out a safe way to relax nodev at some point.

Quoted 25 lines · Serge Hallyn

Quoted 7 lines · Andy Lutomirski

Quoted 5 lines

It should be a nonissue so long as we make sure that a file owned by a uid outside the scope of the container may not be changed even though fs_owner_uid is set. Otherwise, it's just a matter of chmod +S on say a shell and anyone who can see the fs from the host will be getting a root shell (assuming said file is owned by the host's uid 0).

I feel like that's too fragile. I'd rather add a rule that one of

yeah I don't wnat to rush something like that. I'd rather stash the userns of the task which did the mounting and check against that. Note that would make it worthless unless and until we allowed mounting from non-init userns, but then we can only claim "our fs superblock readers suck and therefore containers can't mount an fs" so long before we start to feel some shame and audit them...

Quoted 9 lines

these filesystems always acts like it's nosuid unless you're inside a user namespace that matches fs_owner_uid.

Maybe even that is too weird. How about setuid, setgid, and fcaps only work on mounts that are in mount namespaces that are owned by the current user namespace or one of its parents? IOW, a struct mount is only trusted if mnt->mnt_ns->user_ns == current user ns or one of its parents?

Quoted 2 lines

Quoted 2 lines

Untrusted mounts would act like they are nosuid,nodev. Someone can try to figure out a safe way to relax nodev at some point.

Do you like this variant? We could add a way for global root to mount an fs on behalf of a userns. I'd rather this be more explicit than just mounting it in a mount ns owned by the user namespace, though.

Quoted 29 lines · Andy Lutomirski

Quoted 27 lines · Serge Hallyn

Quoted 7 lines · Andy Lutomirski

Quoted 5 lines

It should be a nonissue so long as we make sure that a file owned by a uid outside the scope of the container may not be changed even though fs_owner_uid is set. Otherwise, it's just a matter of chmod +S on say a shell and anyone who can see the fs from the host will be getting a root shell (assuming said file is owned by the host's uid 0).

I feel like that's too fragile. I'd rather add a rule that one of

yeah I don't wnat to rush something like that. I'd rather stash the userns of the task which did the mounting and check against that. Note that would make it worthless unless and until we allowed mounting from non-init userns, but then we can only claim "our fs superblock readers suck and therefore containers can't mount an fs" so long before we start to feel some shame and audit them...

Quoted 11 lines

these filesystems always acts like it's nosuid unless you're inside a user namespace that matches fs_owner_uid.

Maybe even that is too weird. How about setuid, setgid, and fcaps only work on mounts that are in mount namespaces that are owned by the current user namespace or one of its parents? IOW, a struct mount is only trusted if mnt->mnt_ns->user_ns == current user ns or one of its parents?

Untrusted mounts would act like they are nosuid,nodev. Someone can try to figure out a safe way to relax nodev at some point.

Quoted 3 lines

Do you like this variant? We could add a way for global root to mount an fs on behalf of a userns. I'd rather this be more explicit than just mounting it in a mount ns owned by the user namespace, though.

I'm missing something. Which mnt are you talking about? A user can just clone a new userns and then clone(CLONE_NEWNS) to get a set of mounts owned by himself... We need to get a mnt (or a cred or straight to a userns) tied to the first mount of the superblock, istm.

Quoted 51 lines · Serge Hallyn

Quoted 50 lines · Marian Marinov

Quoted 48 lines · Serge Hallyn

Quoted 36 lines · Marian Marinov

Quoted 35 lines · Serge Hallyn

Quoted 18 lines · Theodore Ts'o

Quoted 3 lines · Marian Marinov

I'm proposing a fix to this, by replacing the capable(CAP_LINUX_IMMUTABLE) check with ns_capable(current_cred()->user_ns, CAP_LINUX_IMMUTABLE).

Um, wouldn't it be better to simply fix the capable() function?

/**
* capable - Determine if the current task has a superior capability in effect
* @cap: The capability to be tested for
*
* Return true if the current task has the given superior capability currently
* available for use, false if not.
*
* This sets PF_SUPERPRIV on the task if the capability is available on the
* assumption that it's about to be used.
*/
bool capable(int cap)

Quoted 8 lines

{
	return ns_capable(&init_user_ns, cap);
}

EXPORT_SYMBOL(capable);

The documentation states that it is for "the current task", and I can't imagine any use case, where user namespaces are in effect, where using init_user_ns would ever make sense.

the init_user_ns represents the user_ns owning the object, not the subject.

The patch by Marian is wrong. Anyone can do 'clone(CLONE_NEWUSER)', setuid(0), execve, and end up satisfying 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' by definition.

Quoted 8 lines

Quoted 3 lines

So NACK to that particular patch. I'm not sure, but IIUC it should be safe to check against the userns owning the inode?

So what you are proposing is to replace 'ns_capable(current_cred()->userns, CAP_SYS_IMMUTABLE)' with 'inode_capable(inode, CAP_SYS_IMMUTABLE)' ?

I agree that this is more sane.

Right, and I think the two operations you're looking at seem sane to allow.

Quoted 7 lines

Quoted 1 line

If you are ok with this patch, I will fix all file systems and send patches.

Sounds good, thanks.

Quoted 1 line

Signed-off-by Marian Marinov <[email protected]>

Acked-by Serge E. Hallyn <serge.hallyn-GeWIH/[email protected]>

Wait, what?

Inodes aren't owned by user namespaces; they're owned by users. And any user can arrange to have a user namespace in which they pass an inode_capable check on any inode that they own.

Presumably there's a reason that CAP_SYS_IMMUTABLE is needed. If this

gets merged, then it would be better to just drop CAP_SYS_IMMUTABLE entirely.

Nacked-by Andy Lutomirski <[email protected]>

I forget the details, but there was another case where I wanted to have the userns which 'owns' the whole fs available. I guess we'd have to check against that instead of using inode_capable.

I forget the details, but there was another case where I wanted to have the userns which 'owns' the whole fs available. I guess we'd have to check against that instead of using inode_capable.

Quoted 3 lines · Serge Hallyn

I forget the details, but there was another case where I wanted to have the userns which 'owns' the whole fs available. I guess we'd have to check against that instead of using inode_capable.

Yes, that sounds right.

And *please* tell me that that under no circumstances can anyone other than root@init_user_ns is allowed to use mknod....

Quoted 3 lines · Serge Hallyn

I forget the details, but there was another case where I wanted to have the userns which 'owns' the whole fs available. I guess we'd have to check against that instead of using inode_capable.

Yes, that sounds right.

And *please* tell me that that under no circumstances can anyone other than root@init_user_ns is allowed to use mknod....

Quoted 3 lines · Serge Hallyn

I forget the details, but there was another case where I wanted to have the userns which 'owns' the whole fs available. I guess we'd have to check against that instead of using inode_capable.

Yes, that sounds right.

And *please* tell me that that under no circumstances can anyone other than root@init_user_ns is allowed to use mknod....

Quoted 9 lines · Theodore Ts'o

Quoted 3 lines · Serge Hallyn

I forget the details, but there was another case where I wanted to have the userns which 'owns' the whole fs available. I guess we'd have to check against that instead of using inode_capable.

Yes, that sounds right.

And *please* tell me that that under no circumstances can anyone other than root@init_user_ns is allowed to use mknod....

Nope. mknod not allowed. capable(CAP_MKNOD) is required is required and I can't see any reason to change that.

As a rule of thumb, the only additional actions allowed in a user namespace above and beyond what an ordinary unpriviliged user would be allowed to do are those things which we only don't allow because they could confuse a setuid root executable.

If we ever allow the creation of immutable files by unprivileged users those files would at least have to be kept completely separate from the files the global root encounters (aka a disjoint mount namespace).

I do not currently see a path to safely using immutable files with just user namespace root permission.