device_cgroup: rework device access check and exception checking

device_cgroup: rework device access check and exception checking

commit 79d719749d23234e9b725098aa49133f3ef7299d upstream.

Whenever a device file is opened and checked against current device
cgroup rules, it uses the same function (may_access()) as when a new
exception rule is added by writing devices.{allow,deny}. And in both
cases, the algorithm is the same, doesn't matter the behavior.

First problem is having device access to be considered the same as rule
checking. Consider the following structure:

	A	(default behavior: allow, exceptions disallow access)
	 \
	  B	(default behavior: allow, exceptions disallow access)

A new exception is added to B by writing devices.deny:

	c 12:34 rw

When checking if that exception is allowed in may_access():

	if (dev_cgroup->behavior == DEVCG_DEFAULT_ALLOW) {
		if (behavior == DEVCG_DEFAULT_ALLOW) {
			/* the exception will deny access to certain devices */
			return true;

Which is ok, since B is not getting more privileges than A, it doesn't
matter and the rule is accepted

Now, consider it's a device file open check and the process belongs to
cgroup B. The access will be generated as:

	behavior: allow
	exception: c 12:34 rw

The very same chunk of code will allow it, even if there's an explicit
exception telling to do otherwise.

A simple test case:

	# mkdir new_group
	# cd new_group
	# echo $$ >tasks
	# echo "c 1:3 w" >devices.deny
	# echo >/dev/null
	# echo $?
	0

This is a serious bug and was introduced on

	c39a2a3018f8 devcg: prepare may_access() for hierarchy support

To solve this problem, the device file open function was split from the
new exception check.

Second problem is how exceptions are processed by may_access(). The
first part of the said function tries to match fully with an existing
exception:

	list_for_each_entry_rcu(ex, &dev_cgroup->exceptions, list) {
		if ((refex->type & DEV_BLOCK) && !(ex->type & DEV_BLOCK))
			continue;
		if ((refex->type & DEV_CHAR) && !(ex->type & DEV_CHAR))
			continue;
		if (ex->major != ~0 && ex->major != refex->major)
			continue;
		if (ex->minor != ~0 && ex->minor != refex->minor)
			continue;
		if (refex->access & (~ex->access))
			continue;
		match = true;
		break;
	}

That means the new exception should be contained into an existing one to
be considered a match:

	New exception		Existing	match?	notes
	b 12:34 rwm		b 12:34 rwm	yes
	b 12:34 r		b *:34 rw	yes
	b 12:34 rw		b 12:34 w	no	extra "r"
	b *:34 rw		b 12:34 rw	no	too broad "*"
	b *:34 rw		b *:34 rwm	yes

Which is fine in some cases. Consider:

	A	(default behavior: deny, exceptions allow access)
	 \
	  B	(default behavior: deny, exceptions allow access)

In this case the full match makes sense, the new exception cannot add
more access than the parent allows

But this doesn't always work, consider:

	A	(default behavior: allow, exceptions disallow access)
	 \
	  B	(default behavior: deny, exceptions allow access)

In this case, a new exception in B shouldn't match any of the exceptions
in A, after all you can't allow something that was forbidden by A. But
consider this scenario:

	New exception	Existing in A	match?	outcome
	b 12:34 rw	b 12:34 r	no	exception is accepted

Because the new exception has "w" as extra, it doesn't match, so it'll
be added to B's exception list.

The same problem can happen during a file access check. Consider a
cgroup with allow as default behavior:

	Access		Exception	match?
	b 12:34 rw	b 12:34 r	no

In this case, the access didn't match any of the exceptions in the
cgroup, which is required since exceptions will disallow access.

To solve this problem, two new functions were created to match an
exception either fully or partially. In the example above, a partial
check will be performed and it'll produce a match since at least
"b 12:34 r" from "b 12:34 rw" access matches.

Cc: cgroups@vger.kernel.org
Cc: Tejun Heo <tj@kernel.org>
Cc: Serge Hallyn <serge.hallyn@canonical.com>
Cc: Li Zefan <lizefan@huawei.com>
Signed-off-by: Aristeu Rozanski <arozansk@redhat.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jiri Slaby <jslaby@suse.cz>