Re: [PATCH bpf-next v2 3/3] bpf: Update LSM selftests for bpf_ima_inode_hash

From: Yonghong Song <yhs@fb.com>
To: KP Singh <kpsingh@chromium.org>, Mimi Zohar <zohar@linux.ibm.com>,
	Andrii Nakryiko <andrii.nakryiko@gmail.com>
Cc: James Morris <jmorris@namei.org>,
	open list <linux-kernel@vger.kernel.org>,
	bpf <bpf@vger.kernel.org>,
	Linux Security Module list 
	<linux-security-module@vger.kernel.org>,
	Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Florent Revest <revest@chromium.org>,
	Brendan Jackman <jackmanb@chromium.org>,
	Petr Vorel <pvorel@suse.cz>
Subject: Re: [PATCH bpf-next v2 3/3] bpf: Update LSM selftests for bpf_ima_inode_hash
Date: Mon, 23 Nov 2020 10:36:05 -0800	[thread overview]
Message-ID: <cf0d94ca-b6a0-1a1a-6cf2-a641002588bf@fb.com> (raw)
In-Reply-To: <CACYkzJ6VEKBJnJZ+CBvpF6C=Kft5A2O5f=Uu4rTMtUiRKN5S-g@mail.gmail.com>

On 11/23/20 10:27 AM, KP Singh wrote:
> [...]
> 
>>>>
>>>> Even if a custom policy has been loaded, potentially additional
>>>> measurements unrelated to this test would be included the measurement
>>>> list.  One way of limiting a rule to a specific test is by loopback
>>>> mounting a file system and defining a policy rule based on the loopback
>>>> mount unique uuid.
>>>
>>> Thanks Mimi!
>>>
>>> I wonder if we simply limit this to policy to /tmp and run an executable
>>> from /tmp (like test_local_storage.c does).
>>>
>>> The only side effect would be of extra hashes being calculated on
>>> binaries run from /tmp which is not too bad I guess?
>>
>> The builtin measurement policy (ima_policy=tcb") explicitly defines a
>> rule to not measure /tmp files.  Measuring /tmp results in a lot of
>> measurements.
>>
>> {.action = DONT_MEASURE, .fsmagic = TMPFS_MAGIC, .flags = IMA_FSMAGIC},
>>
>>>
>>> We could do the loop mount too, but I am guessing the most clean way
>>> would be to shell out to mount from the test? Are there some other examples
>>> of IMA we could look at?
>>
>> LTP loopback mounts a filesystem, since /tmp is not being measured with
>> the builtin "tcb" policy.  Defining new policy rules should be limited
>> to the loopback mount.  This would pave the way for defining IMA-
>> appraisal signature verification policy rules, without impacting the
>> running system.
> 
> +Andrii
> 
> Do you think we can split the IMA test out,
> have a little shell script that does the loopback mount, gets the
> FS UUID, updates the IMA policy and then runs a C program?
> 
> This would also allow "test_progs" to be independent of CONFIG_IMA.
> 
> I am guessing the structure would be something similar
> to test_xdp_redirect.sh

Look at sk_assign test.

sk_assign.c:    if (CHECK_FAIL(system("ip link set dev lo up")))
sk_assign.c:    if (CHECK_FAIL(system("ip route add local default dev lo")))
sk_assign.c:    if (CHECK_FAIL(system("ip -6 route add local default dev 
lo")))
sk_assign.c:    if (CHECK_FAIL(system("tc qdisc add dev lo clsact")))
sk_assign.c:    if (CHECK(system(tc_cmd), "BPF load failed;"

You can use "system" to invoke some bash commands to simulate a script
in the tests.

> 
> - KP
> 
>>
>> Mimi
>>